diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_tensor([728, 728])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_tensor([728, 728])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..af07a0fe9802d1b5bc1e2fbb31a03296c26d2775 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_tensor([728, 728])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a soft-focus close-up shot, we shoot from a low angle, the camera panning extremely slowly to the right, following the subject's movements. A massive yellow excavator dominates the frame, its weathered metal body mottled with dirt and scratches from work. In the warm afternoon sun, the fine dust particles lingering in the air shimmer golden against the backlight. The excavator's massive mechanical arm rises steadily and calmly, its massive bucket, laden with moist, dark-brown earth, moving precisely into position amid the deep, powerful hum of its hydraulic system. As the bucket tilts gently at a precisely controlled angle, the earth, mixed with rocks, cascades down like a heavy waterfall, merging into the massive mound gradually taking shape in the background. The entire picture is shrouded in a warm golden halo. The main colors are the yellow of the excavator and the brown of the soil. The light and shadow outline a rich texture on the rough surface, creating a unique beauty that combines industrial power with natural tranquility. The atmosphere is focused and powerful.", + "seed": 22358, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_728_seed_22358_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04856893820688128, + "vlm_logprob": -0.04856893820688128, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04856893820688128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025093400850892067 + ], + "opposite_logprob_tokens": [ + -4.4000935554504395 + ], + "logprob_reward": -0.025093400850892067, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.025093400850892067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17061491310596466, + 0.05110223591327667, + 0.15405702590942383, + 0.060192737728357315, + 0.15885628759860992, + 0.08313322812318802, + 0.07355444133281708, + 0.0509650744497776, + 0.1389726996421814, + 0.05855134129524231 + ], + "credit_map": [ + 0.014315550215542316, + 0.027515901252627373, + 0.020198345184326172, + 0.02469627372920513, + 0.02574784867465496, + 0.017759528011083603, + 0.015898441895842552, + 0.024483023211359978, + 0.005333086010068655, + 0.004974401090294123, + 0.00525980768725276, + 0.005964249838143587, + 0.004263814073055983, + 0.0067961630411446095, + 0.009043731726706028, + 0.009466982446610928, + 0.014223167672753334, + 0.027150867506861687, + 0.012230241671204567, + 0.026071790605783463, + 0.016446499153971672, + 0.025268349796533585, + 0.017994292080402374, + 0.014671818353235722, + 0.0069478279910981655, + 0.0070802075788378716, + 0.006214465014636517, + 0.013403313234448433, + 0.005136728752404451, + 0.006252143997699022, + 0.00813507754355669, + 0.007022972218692303, + 0.013495687395334244, + 0.0228157676756382, + 0.014567377977073193, + 0.025188790634274483, + 0.020460408180952072, + 0.020512530580163002, + 0.02509312517940998, + 0.016722597181797028, + 0.007238924037665129, + 0.010539472103118896, + 0.0072715706191957, + 0.01203940063714981, + 0.012040886096656322, + 0.016548097133636475, + 0.010085755959153175, + 0.00736911641433835, + 0.00625203549861908, + 0.010271096602082253, + 0.006331197917461395, + 0.009372810833156109, + 0.007747633382678032, + 0.012344866059720516, + 0.008542275056242943, + 0.012692523188889027, + 0.005081625655293465, + 0.006033375393599272, + 0.008811145089566708, + 0.006895861588418484, + 0.00588296027854085, + 0.006178127136081457, + 0.0061710975132882595, + 0.005910881329327822, + 0.013978362083435059, + 0.02236216887831688, + 0.01692276820540428, + 0.014745093882083893, + 0.018065540120005608, + 0.015704547986388206, + 0.019468341022729874, + 0.017725888639688492, + 0.009792688302695751, + 0.00603478541597724, + 0.006959320977330208, + 0.006784311030060053, + 0.0062883151695132256, + 0.00590279046446085, + 0.00880469847470522, + 0.007984431460499763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.025093400850892067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035185754299164 + ], + "opposite_logprob_tokens": [ + -1.9103518724441528 + ], + "logprob_reward": -0.16035185754299164, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035185754299164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05813896283507347, + 0.06426786631345749, + 0.14725318551063538, + 0.3207007348537445, + 0.09745083004236221, + 0.0817071869969368, + 0.048081375658512115, + 0.05877811461687088, + 0.07297682017087936, + 0.0506448820233345 + ], + "credit_map": [ + 0.007007230073213577, + 0.007185014430433512, + 0.009531842544674873, + 0.010646049864590168, + 0.006350729148834944, + 0.004903915338218212, + 0.007461087312549353, + 0.00505309272557497, + 0.009805052541196346, + 0.006649787537753582, + 0.008654257282614708, + 0.0074286251328885555, + 0.008012555539608002, + 0.007606170605868101, + 0.007747456897050142, + 0.008363969624042511, + 0.01588963158428669, + 0.02809099107980728, + 0.01257703173905611, + 0.024823566898703575, + 0.01746753603219986, + 0.02108798548579216, + 0.015830328688025475, + 0.011486112140119076, + 0.04547450691461563, + 0.039406124502420425, + 0.02718365378677845, + 0.06871969252824783, + 0.030497876927256584, + 0.0302798543125391, + 0.03191725164651871, + 0.04722178354859352, + 0.00868845172226429, + 0.008504210039973259, + 0.006284964736551046, + 0.01913296990096569, + 0.00900780688971281, + 0.01380015816539526, + 0.019173219799995422, + 0.012859039008617401, + 0.006367955356836319, + 0.010002292692661285, + 0.008534728549420834, + 0.010499877855181694, + 0.01279144175350666, + 0.017197901383042336, + 0.009189476259052753, + 0.00712351081892848, + 0.003678548615425825, + 0.009316038340330124, + 0.0037764105945825577, + 0.006749733351171017, + 0.006622037850320339, + 0.006510378327220678, + 0.006437984760850668, + 0.0049902452155947685, + 0.006145213730633259, + 0.007327908184379339, + 0.007142082788050175, + 0.009390580467879772, + 0.0065023023635149, + 0.006661179009824991, + 0.008987992070615292, + 0.00662085460498929, + 0.0065300678834319115, + 0.010538325645029545, + 0.009830751456320286, + 0.008131301961839199, + 0.009583335369825363, + 0.007638698443770409, + 0.0110768498852849, + 0.009647488594055176, + 0.0061852033250033855, + 0.004241522867232561, + 0.0043892874382436275, + 0.004316953010857105, + 0.008322285488247871, + 0.004028160125017166, + 0.011156782507896423, + 0.008004691451787949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.16035185754299164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606965810060501 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.01606965810060501, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606965810060501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07367278635501862, + 0.05560893192887306, + 0.3250584304332733, + 0.051508497446775436, + 0.11364585906267166, + 0.019915755838155746, + 0.08058365434408188, + 0.09534070640802383, + 0.10999953001737595, + 0.07466579973697662 + ], + "credit_map": [ + 0.008149140514433384, + 0.009915169328451157, + 0.00983591191470623, + 0.012010523118078709, + 0.010063153691589832, + 0.005146403331309557, + 0.011036987416446209, + 0.007515495177358389, + 0.0067351143807172775, + 0.005190360825508833, + 0.00817225780338049, + 0.007103846408426762, + 0.006134634837508202, + 0.007101365365087986, + 0.00779083464294672, + 0.0073805199936032295, + 0.0390968881547451, + 0.06683018803596497, + 0.030617525801062584, + 0.03981375694274902, + 0.03503291681408882, + 0.047088976949453354, + 0.037397801876068115, + 0.02918040007352829, + 0.004237265791743994, + 0.005273067858070135, + 0.0039986418560147285, + 0.01321584451943636, + 0.005444151349365711, + 0.006910294760018587, + 0.00598532147705555, + 0.006443910766392946, + 0.012314680963754654, + 0.015332750044763088, + 0.011547056958079338, + 0.014887591823935509, + 0.015110019594430923, + 0.022492073476314545, + 0.012196799740195274, + 0.009764901362359524, + 0.0018195438897237182, + 0.0037904828786849976, + 0.001909912913106382, + 0.0026798646431416273, + 0.0026551904156804085, + 0.003071236191317439, + 0.0021860417909920216, + 0.0018034838140010834, + 0.004928698763251305, + 0.014493845403194427, + 0.006467352155596018, + 0.01063243392854929, + 0.01199388038367033, + 0.011384657584130764, + 0.012586258351802826, + 0.008096531964838505, + 0.005540790967643261, + 0.009199457243084908, + 0.010676511563360691, + 0.017023364081978798, + 0.012376674450933933, + 0.01451099943369627, + 0.015415710397064686, + 0.010597199201583862, + 0.013035805895924568, + 0.016165096312761307, + 0.014127049595117569, + 0.013570886105298996, + 0.015542574226856232, + 0.011332887224853039, + 0.01267638336867094, + 0.013548859395086765, + 0.007637845817953348, + 0.007356078363955021, + 0.005043497774749994, + 0.007697252091020346, + 0.010400611907243729, + 0.006665632128715515, + 0.017091840505599976, + 0.012773046270012856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.01606965810060501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370248153805733 + ], + "opposite_logprob_tokens": [ + -3.4087023735046387 + ], + "logprob_reward": -0.03370248153805733, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370248153805733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09587342292070389, + 0.02329123020172119, + 0.21917173266410828, + 0.047460898756980896, + 0.12483305484056473, + 0.04978249594569206, + 0.09320401400327682, + 0.09156081080436707, + 0.1467258781194687, + 0.10809652507305145 + ], + "credit_map": [ + 0.006953727919608355, + 0.011616330593824387, + 0.009415267035365105, + 0.01554084662348032, + 0.016193298622965813, + 0.010501575656235218, + 0.011449739336967468, + 0.014202633872628212, + 0.0032059939112514257, + 0.0021111061796545982, + 0.0028082290664315224, + 0.0028235637582838535, + 0.0023898030631244183, + 0.003671515267342329, + 0.003402216359972954, + 0.0028788000345230103, + 0.02482660673558712, + 0.04863852635025978, + 0.020387621596455574, + 0.03217695653438568, + 0.025270448997616768, + 0.024746963754296303, + 0.025091778486967087, + 0.018032828345894814, + 0.006507151294499636, + 0.004689732566475868, + 0.0038327029906213284, + 0.011057992465794086, + 0.004470881074666977, + 0.006698407232761383, + 0.0035213911905884743, + 0.006682639010250568, + 0.010123164393007755, + 0.013291150331497192, + 0.013491258025169373, + 0.022388307377696037, + 0.015557448379695415, + 0.01925329491496086, + 0.017453737556934357, + 0.013274678029119968, + 0.0034922922495752573, + 0.006513318046927452, + 0.004816991742700338, + 0.006626880262047052, + 0.007558520417660475, + 0.0120094520971179, + 0.0044263191521167755, + 0.004338715225458145, + 0.007524216081947088, + 0.016779346391558647, + 0.008563226088881493, + 0.013126760721206665, + 0.014925996772944927, + 0.01283061970025301, + 0.011174619197845459, + 0.008279233239591122, + 0.005534930154681206, + 0.009871975518763065, + 0.010796474292874336, + 0.015457341447472572, + 0.010694376192986965, + 0.014064260758459568, + 0.015910113230347633, + 0.009231343865394592, + 0.015394126065075397, + 0.020923584699630737, + 0.015304184518754482, + 0.02303967997431755, + 0.018986979499459267, + 0.01652623526751995, + 0.020515112206339836, + 0.016035960987210274, + 0.010122002102434635, + 0.01267253514379263, + 0.007747603114694357, + 0.010146700777113438, + 0.017855260521173477, + 0.00895483698695898, + 0.02342940680682659, + 0.01716817170381546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.03370248153805733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17923252284526825, + 0.03525323048233986, + 0.0612061470746994, + 0.09006906300783157, + 0.057596951723098755, + 0.26976051926612854, + 0.06525487452745438, + 0.10897179692983627, + 0.06603492796421051, + 0.06661996245384216 + ], + "credit_map": [ + 0.01786057837307453, + 0.021480994299054146, + 0.02225765958428383, + 0.029988743364810944, + 0.020271962508559227, + 0.017384326085448265, + 0.02743937447667122, + 0.022548887878656387, + 0.003302308963611722, + 0.003905799938365817, + 0.00547692971304059, + 0.004472894128412008, + 0.003332545282319188, + 0.004381832666695118, + 0.005400226917117834, + 0.004980691242963076, + 0.007689015474170446, + 0.014430873095989227, + 0.005432448349893093, + 0.008234607055783272, + 0.007231307681649923, + 0.006687072571367025, + 0.006533867679536343, + 0.004966957028955221, + 0.006322750821709633, + 0.0070896758697927, + 0.0058084880001842976, + 0.026936249807476997, + 0.006655119359493256, + 0.010452432557940483, + 0.009127836674451828, + 0.01767650805413723, + 0.004772778134793043, + 0.0074643781408667564, + 0.004967492539435625, + 0.009469222277402878, + 0.007278651464730501, + 0.010780882090330124, + 0.006799621507525444, + 0.006063919514417648, + 0.019327333196997643, + 0.043473467230796814, + 0.031491801142692566, + 0.03514048084616661, + 0.0442286916077137, + 0.048106566071510315, + 0.02645750716328621, + 0.02153468318283558, + 0.004383245017379522, + 0.008093699812889099, + 0.008278219029307365, + 0.00875028595328331, + 0.011452335864305496, + 0.008816439658403397, + 0.00930272787809372, + 0.0061779203824698925, + 0.00880615133792162, + 0.010432671755552292, + 0.010455175302922726, + 0.018429892137646675, + 0.012564487755298615, + 0.015715956687927246, + 0.01998133584856987, + 0.012586130760610104, + 0.006561589427292347, + 0.01096813753247261, + 0.007206664886325598, + 0.007258598692715168, + 0.009109933860599995, + 0.007490403484553099, + 0.008660508319735527, + 0.008779092691838741, + 0.012761915102601051, + 0.0068536605685949326, + 0.007170453667640686, + 0.007408627774566412, + 0.008579418063163757, + 0.008013201877474785, + 0.009045413695275784, + 0.0067872703075408936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.00762729300186038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a soft-focus close-up shot, we shoot from a low angle, the camera panning extremely slowly to the right, following the subject's movements. A massive yellow excavator dominates the frame, its weathered metal body mottled with dirt and scratches from work. In the warm afternoon sun, the fine dust particles lingering in the air shimmer golden against the backlight. The excavator's massive mechanical arm rises steadily and calmly, its massive bucket, laden with moist, dark-brown earth, moving precisely into position amid the deep, powerful hum of its hydraulic system. As the bucket tilts gently at a precisely controlled angle, the earth, mixed with rocks, cascades down like a heavy waterfall, merging into the massive mound gradually taking shape in the background. The entire picture is shrouded in a warm golden halo. The main colors are the yellow of the excavator and the brown of the soil. The light and shadow outline a rich texture on the rough surface, creating a unique beauty that combines industrial power with natural tranquility. The atmosphere is focused and powerful.", + "seed": 22359, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_728_seed_22359_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04696515128016472, + "vlm_logprob": -0.04696515128016472, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.04696515128016472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02119101770222187 + ], + "opposite_logprob_tokens": [ + -4.396191120147705 + ], + "logprob_reward": -0.02119101770222187, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02119101770222187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35993844270706177, + 0.08659534901380539, + 0.048437006771564484, + 0.07837044447660446, + 0.07545576244592667, + 0.05675308778882027, + 0.05474390834569931, + 0.10065092146396637, + 0.09580662101507187, + 0.04324847832322121 + ], + "credit_map": [ + 0.03809061646461487, + 0.06474663317203522, + 0.02441694773733616, + 0.044926341623067856, + 0.059929441660642624, + 0.05655393749475479, + 0.043533358722925186, + 0.027741149067878723, + 0.01458529382944107, + 0.010281676426529884, + 0.009091491810977459, + 0.006910494063049555, + 0.010806430131196976, + 0.008921501226723194, + 0.009115805849432945, + 0.016882654279470444, + 0.004666921216994524, + 0.007010214496403933, + 0.008142520673573017, + 0.005050389561802149, + 0.004878697916865349, + 0.00708413403481245, + 0.004334563855081797, + 0.007269564084708691, + 0.007894215174019337, + 0.007798456121236086, + 0.01948714815080166, + 0.010330728255212307, + 0.008524546399712563, + 0.007722518872469664, + 0.009428680874407291, + 0.007184146903455257, + 0.00894835777580738, + 0.009531458839774132, + 0.016586007550358772, + 0.011531752534210682, + 0.0071565499529242516, + 0.008231316693127155, + 0.007788240909576416, + 0.005682076793164015, + 0.0035326622892171144, + 0.008241305127739906, + 0.006636269390583038, + 0.006221422925591469, + 0.009755865670740604, + 0.008866650052368641, + 0.0065179369412362576, + 0.006980976555496454, + 0.006614405196160078, + 0.007958493195474148, + 0.0079294228926301, + 0.005410294979810715, + 0.008819655515253544, + 0.005098643247038126, + 0.006398176308721304, + 0.006514817010611296, + 0.01632724702358246, + 0.010967472568154335, + 0.021889757364988327, + 0.00915451254695654, + 0.015638208016753197, + 0.009069656953215599, + 0.0077949571423232555, + 0.009809112176299095, + 0.010322253219783306, + 0.009981009177863598, + 0.012131121940910816, + 0.01729971542954445, + 0.016757044941186905, + 0.011741755530238152, + 0.009014440700411797, + 0.008559288457036018, + 0.0060832011513412, + 0.004397502169013023, + 0.006836558226495981, + 0.0057530696503818035, + 0.005653419066220522, + 0.004707853309810162, + 0.005411481484770775, + 0.004405395593494177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.02119101770222187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987322807312012 + ], + "opposite_logprob_tokens": [ + -1.8048732280731201 + ], + "logprob_reward": -0.17987322807312012, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987322807312012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05141739547252655, + 0.037229787558317184, + 0.020911194384098053, + 0.0419023372232914, + 0.4772275686264038, + 0.05975315719842911, + 0.0551321767270565, + 0.17337052524089813, + 0.07176786661148071, + 0.011288003996014595 + ], + "credit_map": [ + 0.004510894417762756, + 0.005045153666287661, + 0.006286263465881348, + 0.006506918463855982, + 0.00430901488289237, + 0.007450147997587919, + 0.011823873035609722, + 0.005485124886035919, + 0.0036923224106431007, + 0.006740037351846695, + 0.005374783184379339, + 0.005067481193691492, + 0.004224276635795832, + 0.004628387745469809, + 0.004189332015812397, + 0.0033131653908640146, + 0.002242389600723982, + 0.0034168814308941364, + 0.0025968977715820074, + 0.0029620290733873844, + 0.0020046604331582785, + 0.002913397504016757, + 0.002490580314770341, + 0.002284358022734523, + 0.0057970318011939526, + 0.0028580562211573124, + 0.013007136061787605, + 0.006061640102416277, + 0.0022908407263457775, + 0.003878789721056819, + 0.0038654180243611336, + 0.00414342200383544, + 0.09297406673431396, + 0.01702136918902397, + 0.19886820018291473, + 0.07945551723241806, + 0.014638730324804783, + 0.0273235235363245, + 0.024225754663348198, + 0.022720392793416977, + 0.00550526799634099, + 0.0041936771012842655, + 0.009672229178249836, + 0.008853444829583168, + 0.005946257151663303, + 0.00618830556049943, + 0.006659680977463722, + 0.012734288349747658, + 0.012371285818517208, + 0.0032770223915576935, + 0.02074817754328251, + 0.0054390281438827515, + 0.004979308228939772, + 0.0027576813008636236, + 0.002594041870906949, + 0.0029656318947672844, + 0.054631881415843964, + 0.007404010277241468, + 0.07209063321352005, + 0.009902042336761951, + 0.006469860207289457, + 0.00339430570602417, + 0.006959239952266216, + 0.01251853909343481, + 0.01253523025661707, + 0.006851378362625837, + 0.013564860448241234, + 0.00791581254452467, + 0.012791942805051804, + 0.008145639672875404, + 0.004422290716320276, + 0.005540704820305109, + 0.0011970906052738428, + 0.0016614283667877316, + 0.0014961222186684608, + 0.0009727143915370107, + 0.002080956706777215, + 0.0014722896739840508, + 0.0012652967125177383, + 0.0011421041563153267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.17987322807312012 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978152547031641 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.00978152547031641, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978152547031641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1514640599489212, + 0.10311051458120346, + 0.03950649872422218, + 0.051558718085289, + 0.3358157277107239, + 0.05336930602788925, + 0.05238428711891174, + 0.09015726298093796, + 0.03929564356803894, + 0.08333797752857208 + ], + "credit_map": [ + 0.0176943801343441, + 0.01917257159948349, + 0.017677785828709602, + 0.018217600882053375, + 0.017004158347845078, + 0.022581780329346657, + 0.021967874839901924, + 0.017147904261946678, + 0.016262326389551163, + 0.0119607700034976, + 0.011268818750977516, + 0.012101514264941216, + 0.009912187233567238, + 0.01083293929696083, + 0.012062924914062023, + 0.01870903931558132, + 0.005193948745727539, + 0.00371724646538496, + 0.005320938769727945, + 0.003684691619127989, + 0.0037637040950357914, + 0.00620309030637145, + 0.003669452155008912, + 0.007953428663313389, + 0.004631171002984047, + 0.00731581961736083, + 0.006617676932364702, + 0.005364271346479654, + 0.007205503527075052, + 0.005275953095406294, + 0.007554024923592806, + 0.007594297174364328, + 0.06574095785617828, + 0.010420035570859909, + 0.14179611206054688, + 0.056913089007139206, + 0.009129767306149006, + 0.018939943984150887, + 0.016665980219841003, + 0.016209831461310387, + 0.004562715068459511, + 0.010553020052611828, + 0.008109042420983315, + 0.007473289500921965, + 0.007791533600538969, + 0.0058196228928864, + 0.0053140148520469666, + 0.003746067639440298, + 0.011308345012366772, + 0.004110005684196949, + 0.018342165276408195, + 0.005511466413736343, + 0.004839540924876928, + 0.0025344162713736296, + 0.0028412924148142338, + 0.0028970541898161173, + 0.02676502615213394, + 0.0046117836609482765, + 0.035711295902729034, + 0.005529251415282488, + 0.004883019253611565, + 0.0025439783930778503, + 0.003957775421440601, + 0.00615513464435935, + 0.0032846136018633842, + 0.004085785709321499, + 0.0042417049407958984, + 0.00575433811172843, + 0.008598299697041512, + 0.005549964495003223, + 0.003546682884916663, + 0.004234254825860262, + 0.012232617475092411, + 0.010606523603200912, + 0.014913671649992466, + 0.009143944829702377, + 0.009312507696449757, + 0.006448248866945505, + 0.011088908649981022, + 0.00959155336022377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.00978152547031641 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201411046087742 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201411046087742, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201411046087742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14579813182353973, + 0.04155214503407478, + 0.029340475797653198, + 0.0815328061580658, + 0.2750234007835388, + 0.09363898634910583, + 0.02714742347598076, + 0.07359001785516739, + 0.09289749711751938, + 0.1394791603088379 + ], + "credit_map": [ + 0.012843922711908817, + 0.0162464901804924, + 0.019119461998343468, + 0.018227169290184975, + 0.015107200480997562, + 0.017035232856869698, + 0.02854841575026512, + 0.018670231103897095, + 0.005585091188549995, + 0.006927970331162214, + 0.00472885649651289, + 0.004745308309793472, + 0.005336780566722155, + 0.0047368016093969345, + 0.004786104895174503, + 0.004705230239778757, + 0.0032999301329255104, + 0.004520521964877844, + 0.002637788886204362, + 0.003167918184772134, + 0.0028051359113305807, + 0.0036580630112439394, + 0.003431156277656555, + 0.005819960031658411, + 0.007741757668554783, + 0.008533676154911518, + 0.014917641878128052, + 0.009928838349878788, + 0.009829054586589336, + 0.008754325099289417, + 0.011119937524199486, + 0.010707573033869267, + 0.05095279961824417, + 0.011522854678332806, + 0.10635451972484589, + 0.04367648810148239, + 0.011692609637975693, + 0.018878083676099777, + 0.01647552289068699, + 0.015470522455871105, + 0.008961131796240807, + 0.017126008868217468, + 0.014130903407931328, + 0.013170654885470867, + 0.013765777461230755, + 0.009677518159151077, + 0.010470512323081493, + 0.006336478982120752, + 0.0022824059706181288, + 0.0037463565822690725, + 0.003082543145865202, + 0.003630575491115451, + 0.0057716965675354, + 0.0028434316627681255, + 0.002508434234187007, + 0.00328198098577559, + 0.009112779051065445, + 0.012306904420256615, + 0.011793782003223896, + 0.008622352965176105, + 0.011066642589867115, + 0.0068648639135062695, + 0.007315163500607014, + 0.006507527083158493, + 0.010389688424766064, + 0.010001132264733315, + 0.012671004049479961, + 0.01273235585540533, + 0.018969925120472908, + 0.010980461724102497, + 0.008140245452523232, + 0.009012681432068348, + 0.019243789836764336, + 0.019409313797950745, + 0.024644983932375908, + 0.015897247940301895, + 0.015678441151976585, + 0.011687248945236206, + 0.016937119886279106, + 0.015981018543243408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.014201411046087742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778574109077454 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778574109077454, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778574109077454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16462711989879608, + 0.05573553591966629, + 0.03984346240758896, + 0.030810309574007988, + 0.29896730184555054, + 0.021101640537381172, + 0.035907723009586334, + 0.21083593368530273, + 0.059431493282318115, + 0.0827394649386406 + ], + "credit_map": [ + 0.015764860436320305, + 0.029273390769958496, + 0.02071031555533409, + 0.018272096291184425, + 0.027872933074831963, + 0.017901867628097534, + 0.018460730090737343, + 0.016370931640267372, + 0.011481580324470997, + 0.004870690871030092, + 0.006033214274793863, + 0.004111345391720533, + 0.0039037191309034824, + 0.005439621862024069, + 0.004712345544248819, + 0.015183014795184135, + 0.004980581346899271, + 0.0035049370490014553, + 0.0056275129318237305, + 0.00365452840924263, + 0.003355706576257944, + 0.006240803748369217, + 0.0037657232023775578, + 0.008713670074939728, + 0.0033815051428973675, + 0.0043951221741735935, + 0.006917943712323904, + 0.004213673993945122, + 0.0022194907069206238, + 0.003169378964230418, + 0.003091854276135564, + 0.0034213396720588207, + 0.05970736965537071, + 0.008240856230258942, + 0.12855398654937744, + 0.051602553576231, + 0.007074770051985979, + 0.015794333070516586, + 0.01413236279040575, + 0.013861036859452724, + 0.001984577625989914, + 0.002917627105489373, + 0.0030477780383080244, + 0.0026768341194838285, + 0.0028217362705618143, + 0.002271616365760565, + 0.0030464250594377518, + 0.002335046883672476, + 0.005540578160434961, + 0.0036937142722308636, + 0.007465646136552095, + 0.0033045399468392134, + 0.007079188711941242, + 0.003361723618581891, + 0.0025832364335656166, + 0.002879098756238818, + 0.0652768686413765, + 0.009071220643818378, + 0.08786030858755112, + 0.014235410839319229, + 0.007592950947582722, + 0.003950294107198715, + 0.007959850132465363, + 0.014889036305248737, + 0.00974504929035902, + 0.005444490350782871, + 0.010010641999542713, + 0.005252709612250328, + 0.01304282434284687, + 0.007477078586816788, + 0.0036171709652990103, + 0.004841526038944721, + 0.010121745988726616, + 0.008612954057753086, + 0.01578974351286888, + 0.012317351996898651, + 0.007597789634019136, + 0.007261419668793678, + 0.01270611584186554, + 0.008332343772053719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.009778574109077454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a soft-focus close-up shot, we shoot from a low angle, the camera panning extremely slowly to the right, following the subject's movements. A massive yellow excavator dominates the frame, its weathered metal body mottled with dirt and scratches from work. In the warm afternoon sun, the fine dust particles lingering in the air shimmer golden against the backlight. The excavator's massive mechanical arm rises steadily and calmly, its massive bucket, laden with moist, dark-brown earth, moving precisely into position amid the deep, powerful hum of its hydraulic system. As the bucket tilts gently at a precisely controlled angle, the earth, mixed with rocks, cascades down like a heavy waterfall, merging into the massive mound gradually taking shape in the background. The entire picture is shrouded in a warm golden halo. The main colors are the yellow of the excavator and the brown of the soil. The light and shadow outline a rich texture on the rough surface, creating a unique beauty that combines industrial power with natural tranquility. The atmosphere is focused and powerful.", + "seed": 22360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_728_seed_22360_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06167574347928166, + "vlm_logprob": -0.06167574347928166, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.06167574347928166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03024590201675892 + ], + "opposite_logprob_tokens": [ + -4.280245780944824 + ], + "logprob_reward": -0.03024590201675892, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03024590201675892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1519315391778946, + 0.04568935185670853, + 0.047920554876327515, + 0.09810379147529602, + 0.22631128132343292, + 0.0708559900522232, + 0.174962118268013, + 0.044256504625082016, + 0.06601198017597198, + 0.07395688444375992 + ], + "credit_map": [ + 0.014881081879138947, + 0.011652220971882343, + 0.02974020317196846, + 0.023115627467632294, + 0.020117150619626045, + 0.016231194138526917, + 0.01635994203388691, + 0.019834114238619804, + 0.005082756280899048, + 0.00747654726728797, + 0.004461231175810099, + 0.005166381597518921, + 0.006249322090297937, + 0.004635814111679792, + 0.007021077908575535, + 0.005596220027655363, + 0.004777946043759584, + 0.005018183961510658, + 0.008625976741313934, + 0.006232204381376505, + 0.00680640060454607, + 0.006070264615118504, + 0.004176365677267313, + 0.00621321564540267, + 0.011428902857005596, + 0.014311467297375202, + 0.011643493548035622, + 0.01864762231707573, + 0.013179288245737553, + 0.008563700132071972, + 0.008640720508992672, + 0.011688594706356525, + 0.029499631375074387, + 0.03662284091114998, + 0.038131069391965866, + 0.03759419545531273, + 0.016288496553897858, + 0.020007560029625893, + 0.022668736055493355, + 0.025498738512396812, + 0.012177672237157822, + 0.006898753810673952, + 0.012674050405621529, + 0.014295310713350773, + 0.0050959400832653046, + 0.007042398676276207, + 0.007825319655239582, + 0.004846544936299324, + 0.02048622816801071, + 0.018941370770335197, + 0.02622039057314396, + 0.02394893579185009, + 0.01953006163239479, + 0.026431242004036903, + 0.021728456020355225, + 0.017675448209047318, + 0.00672377273440361, + 0.004190399777144194, + 0.007208432536572218, + 0.005682099144905806, + 0.0051397038623690605, + 0.0063461014069616795, + 0.005131773184984922, + 0.003834223374724388, + 0.007884363643825054, + 0.009253307245671749, + 0.009365841746330261, + 0.0076757632195949554, + 0.005682526156306267, + 0.009880796074867249, + 0.009105585515499115, + 0.007163795176893473, + 0.009394173510372639, + 0.01005083043128252, + 0.011727939359843731, + 0.007942987605929375, + 0.007642810232937336, + 0.007203139364719391, + 0.011118155904114246, + 0.008876851759850979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.03024590201675892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252084881067276 + ], + "opposite_logprob_tokens": [ + -1.5020848512649536 + ], + "logprob_reward": -0.252084881067276, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252084881067276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07224766165018082, + 0.06010381132364273, + 0.030365388840436935, + 0.12044451385736465, + 0.23814761638641357, + 0.05689417943358421, + 0.14745672047138214, + 0.11619637161493301, + 0.14099687337875366, + 0.017146823927760124 + ], + "credit_map": [ + 0.005991433747112751, + 0.008101721294224262, + 0.008258817717432976, + 0.011319044977426529, + 0.011417343281209469, + 0.011409945785999298, + 0.008653662167489529, + 0.007095692679286003, + 0.004766376223415136, + 0.00733460346236825, + 0.008240297436714172, + 0.005949980113655329, + 0.009532000869512558, + 0.007452726364135742, + 0.010226413607597351, + 0.006601416505873203, + 0.003937344998121262, + 0.0040074498392641544, + 0.003416262101382017, + 0.003974065650254488, + 0.0035688963253051043, + 0.004517655819654465, + 0.0035045044496655464, + 0.0034392119850963354, + 0.013023294508457184, + 0.01282627135515213, + 0.015726514160633087, + 0.026821615174412727, + 0.01322170253843069, + 0.011779658496379852, + 0.00978092011064291, + 0.01726454123854637, + 0.022734276950359344, + 0.03651689738035202, + 0.031764235347509384, + 0.03501651808619499, + 0.02925211563706398, + 0.026856908574700356, + 0.03094971366226673, + 0.025056950747966766, + 0.00547895347699523, + 0.005033721681684256, + 0.01052961964160204, + 0.008712023496627808, + 0.005472877062857151, + 0.007800552062690258, + 0.00775616941973567, + 0.0061102621257305145, + 0.015828849747776985, + 0.014903312548995018, + 0.01984291523694992, + 0.0211646631360054, + 0.016449274495244026, + 0.022187966853380203, + 0.02183595485985279, + 0.01524379476904869, + 0.014786780811846256, + 0.01165178045630455, + 0.016156604513525963, + 0.012693257071077824, + 0.015736516565084457, + 0.01639348827302456, + 0.017953356727957726, + 0.010824589058756828, + 0.015417301096022129, + 0.017307789996266365, + 0.018955526873469353, + 0.012886679731309414, + 0.014546058140695095, + 0.029336491599678993, + 0.019389964640140533, + 0.013157072477042675, + 0.0022610710002481937, + 0.0019426029175519943, + 0.0027076941914856434, + 0.001849887426942587, + 0.001421369961462915, + 0.0018306951969861984, + 0.0022080799099057913, + 0.0029254243709146976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.252084881067276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1368832290172577, + 0.03873797133564949, + 0.033648304641246796, + 0.13973791897296906, + 0.14771273732185364, + 0.045317620038986206, + 0.3160570561885834, + 0.03908509016036987, + 0.05547579377889633, + 0.04734426364302635 + ], + "credit_map": [ + 0.012067427858710289, + 0.016803907230496407, + 0.018098006024956703, + 0.02246169000864029, + 0.02020200528204441, + 0.016298921778798103, + 0.01643788255751133, + 0.014513389207422733, + 0.002357012592256069, + 0.00583242392167449, + 0.006143194157630205, + 0.003910062834620476, + 0.005422521382570267, + 0.004389423877000809, + 0.006608267780393362, + 0.004075065720826387, + 0.0036047992762178183, + 0.0036224997602403164, + 0.005165614187717438, + 0.004762755241245031, + 0.0035720320884138346, + 0.005091293249279261, + 0.003623203607276082, + 0.004206106532365084, + 0.0150407450273633, + 0.025821002200245857, + 0.01995409093797207, + 0.030461829155683517, + 0.011459781788289547, + 0.009026146493852139, + 0.011555209755897522, + 0.01641911268234253, + 0.012774193659424782, + 0.014445831067860126, + 0.021675484254956245, + 0.016358567401766777, + 0.02461070753633976, + 0.024105461314320564, + 0.014820029027760029, + 0.018922464922070503, + 0.004744865465909243, + 0.0032030451111495495, + 0.007276580668985844, + 0.00798165611922741, + 0.003388151526451111, + 0.005552875809371471, + 0.006625097244977951, + 0.0065453494898974895, + 0.034958332777023315, + 0.034384991973638535, + 0.04235759750008583, + 0.04349621757864952, + 0.03472048044204712, + 0.05240871384739876, + 0.03566508740186691, + 0.03806564211845398, + 0.0051920474506914616, + 0.006140496116131544, + 0.007288500666618347, + 0.006097766105085611, + 0.0039046378806233406, + 0.0032431604340672493, + 0.004411722533404827, + 0.002806758973747492, + 0.007722686510533094, + 0.00986854825168848, + 0.008193557150661945, + 0.007622244767844677, + 0.0044148871675133705, + 0.0055837747640907764, + 0.00857397262006998, + 0.003496122546494007, + 0.006487169302999973, + 0.00551740312948823, + 0.0076841674745082855, + 0.0054761506617069244, + 0.004149605054408312, + 0.004129684064537287, + 0.007718557957559824, + 0.006181527394801378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.007631078828126192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641822263598442 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641822263598442, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641822263598442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07523258775472641, + 0.04714197665452957, + 0.051801349967718124, + 0.10522735118865967, + 0.1942777782678604, + 0.08772003650665283, + 0.23795071244239807, + 0.05617860332131386, + 0.07845553010702133, + 0.0660141110420227 + ], + "credit_map": [ + 0.008390245027840137, + 0.00756278820335865, + 0.010211903601884842, + 0.011562010273337364, + 0.011589162983000278, + 0.009160914458334446, + 0.008855503052473068, + 0.007900065742433071, + 0.004592420998960733, + 0.006016407161951065, + 0.007295786868780851, + 0.0035833409056067467, + 0.00762883061543107, + 0.005463677458465099, + 0.005664723459631205, + 0.006896790117025375, + 0.005221829283982515, + 0.005446788389235735, + 0.0073742433451116085, + 0.007669711019843817, + 0.008260278962552547, + 0.006450289860367775, + 0.006254489999264479, + 0.005123720038682222, + 0.012196163646876812, + 0.014610826037824154, + 0.013910340145230293, + 0.02058408223092556, + 0.010812298394739628, + 0.007990555837750435, + 0.009306281805038452, + 0.015816794708371162, + 0.027643432840704918, + 0.031864166259765625, + 0.03320756554603577, + 0.03483282029628754, + 0.013043388724327087, + 0.01567627675831318, + 0.019045278429985046, + 0.018964847549796104, + 0.008817078545689583, + 0.006037396378815174, + 0.012434126809239388, + 0.0139063885435462, + 0.010071132332086563, + 0.015013600699603558, + 0.009268434718251228, + 0.012171873822808266, + 0.03130058944225311, + 0.02828875556588173, + 0.03693092241883278, + 0.03216538578271866, + 0.024086711928248405, + 0.034263335168361664, + 0.027929112315177917, + 0.02298589237034321, + 0.006887849885970354, + 0.006424133665859699, + 0.008180239237844944, + 0.006963151507079601, + 0.005028441082686186, + 0.00790672842413187, + 0.008271799422800541, + 0.00651626056060195, + 0.009224634617567062, + 0.01341331098228693, + 0.010297866538167, + 0.008835029788315296, + 0.008066914975643158, + 0.011307635344564915, + 0.010840896517038345, + 0.006469245534390211, + 0.00864389818161726, + 0.009104159660637379, + 0.007665466982871294, + 0.007369159255176783, + 0.00850052759051323, + 0.007309679873287678, + 0.0078634237870574, + 0.009557791985571384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.008641822263598442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775033220648766 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775033220648766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775033220648766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12176765501499176, + 0.10823789983987808, + 0.022126706317067146, + 0.25710123777389526, + 0.1352875679731369, + 0.03354791924357414, + 0.19170379638671875, + 0.02274862304329872, + 0.055560026317834854, + 0.05191853642463684 + ], + "credit_map": [ + 0.011613154783844948, + 0.01138158980756998, + 0.01951807737350464, + 0.012478582561016083, + 0.017967985942959785, + 0.01603994332253933, + 0.01576157473027706, + 0.01700674183666706, + 0.008020513691008091, + 0.014296089299023151, + 0.016812747344374657, + 0.01027700025588274, + 0.01706668920814991, + 0.013004361651837826, + 0.01924598217010498, + 0.009514516219496727, + 0.0020368555560708046, + 0.002467554295435548, + 0.0028797825798392296, + 0.0038588130846619606, + 0.0024109352380037308, + 0.0030040533747524023, + 0.00251862034201622, + 0.002950092777609825, + 0.029589101672172546, + 0.04110350087285042, + 0.036244165152311325, + 0.04578346014022827, + 0.026727955788373947, + 0.020207470282912254, + 0.02087099477648735, + 0.0365745835006237, + 0.014060971327126026, + 0.01758114993572235, + 0.022864405065774918, + 0.017898907884955406, + 0.01784166321158409, + 0.01613042503595352, + 0.014503061771392822, + 0.01440698467195034, + 0.004567385651171207, + 0.0037165761459618807, + 0.004939503036439419, + 0.006509607657790184, + 0.0021341454703360796, + 0.0036608062218874693, + 0.004087710753083229, + 0.003932185471057892, + 0.02280779369175434, + 0.020575856789946556, + 0.027230961248278618, + 0.025242798030376434, + 0.02031460404396057, + 0.033000729978084564, + 0.02135557308793068, + 0.021175473928451538, + 0.002795916050672531, + 0.0026206637267023325, + 0.004561442881822586, + 0.003136883955448866, + 0.002034866251051426, + 0.0020640294533222914, + 0.002990022301673889, + 0.002544798655435443, + 0.00669697904959321, + 0.010427690111100674, + 0.00713816424831748, + 0.006820254493504763, + 0.004381174687296152, + 0.006478684023022652, + 0.008361468091607094, + 0.0052556139416992664, + 0.008637671358883381, + 0.008046754635870457, + 0.007967672310769558, + 0.006111491937190294, + 0.006129477173089981, + 0.005607724189758301, + 0.0057963221333920956, + 0.0036214229185134172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.009775033220648766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a soft-focus close-up shot, we shoot from a low angle, the camera panning extremely slowly to the right, following the subject's movements. A massive yellow excavator dominates the frame, its weathered metal body mottled with dirt and scratches from work. In the warm afternoon sun, the fine dust particles lingering in the air shimmer golden against the backlight. The excavator's massive mechanical arm rises steadily and calmly, its massive bucket, laden with moist, dark-brown earth, moving precisely into position amid the deep, powerful hum of its hydraulic system. As the bucket tilts gently at a precisely controlled angle, the earth, mixed with rocks, cascades down like a heavy waterfall, merging into the massive mound gradually taking shape in the background. The entire picture is shrouded in a warm golden halo. The main colors are the yellow of the excavator and the brown of the soil. The light and shadow outline a rich texture on the rough surface, creating a unique beauty that combines industrial power with natural tranquility. The atmosphere is focused and powerful.", + "seed": 22361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_0/index_728_seed_22361_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07407481772825122, + "vlm_logprob": -0.07407481772825122, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.07407481772825122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040564775466918945 + ], + "opposite_logprob_tokens": [ + -4.29056453704834 + ], + "logprob_reward": -0.040564775466918945, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.040564775466918945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1231672391295433, + 0.06038158759474754, + 0.03102767840027809, + 0.040556855499744415, + 0.12942343950271606, + 0.037529539316892624, + 0.02072897180914879, + 0.11224222928285599, + 0.04583938047289848, + 0.399103045463562 + ], + "credit_map": [ + 0.01448699552565813, + 0.012024948373436928, + 0.0126418462023139, + 0.010043123736977577, + 0.01792381890118122, + 0.027234509587287903, + 0.010203620418906212, + 0.018608378246426582, + 0.00603719986975193, + 0.0034218288492411375, + 0.007410390302538872, + 0.003385137068107724, + 0.007464919239282608, + 0.007041872013360262, + 0.014236816205084324, + 0.011383435688912868, + 0.004027875605970621, + 0.003455800237134099, + 0.003581518540158868, + 0.003012829925864935, + 0.004236759152263403, + 0.003438882064074278, + 0.0029137840028852224, + 0.006360230967402458, + 0.007442387752234936, + 0.003135201521217823, + 0.005737791769206524, + 0.0036371061578392982, + 0.004146867897361517, + 0.004072645679116249, + 0.0039185453206300735, + 0.008466310799121857, + 0.015670862048864365, + 0.01591687649488449, + 0.015304193831980228, + 0.014725672081112862, + 0.025024836882948875, + 0.015198378823697567, + 0.01072230190038681, + 0.01686033420264721, + 0.005334163550287485, + 0.004020100925117731, + 0.004513844382017851, + 0.004606723785400391, + 0.0036866064183413982, + 0.005375598557293415, + 0.004540935158729553, + 0.005451568402349949, + 0.002023946726694703, + 0.004136422649025917, + 0.002518864581361413, + 0.003233537310734391, + 0.00141227210406214, + 0.0021718137431889772, + 0.0031277602538466454, + 0.0021043564192950726, + 0.014274259097874165, + 0.018536701798439026, + 0.013276699930429459, + 0.02442272938787937, + 0.008455873467028141, + 0.011072494089603424, + 0.008652053773403168, + 0.013551434502005577, + 0.011813054792582989, + 0.004068473819643259, + 0.005026037339121103, + 0.006227309349924326, + 0.003902089549228549, + 0.005974362138658762, + 0.0037332265637815, + 0.005094831809401512, + 0.02156740054488182, + 0.014845783822238445, + 0.1468064934015274, + 0.014071183279156685, + 0.013472598977386951, + 0.1656339466571808, + 0.011895201168954372, + 0.010810445994138718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.040564775466918945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134394884109497 + ], + "opposite_logprob_tokens": [ + -1.3134394884109497 + ], + "logprob_reward": -0.3134394884109497, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134394884109497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05108275264501572, + 0.09214463084936142, + 0.074552983045578, + 0.035124361515045166, + 0.12742777168750763, + 0.08869124203920364, + 0.0271376334130764, + 0.06463022530078888, + 0.03087448887526989, + 0.408333957195282 + ], + "credit_map": [ + 0.006903826259076595, + 0.007757526356726885, + 0.0056711602956056595, + 0.0061730192974209785, + 0.006307689938694239, + 0.0065882569178938866, + 0.004730609245598316, + 0.006950664333999157, + 0.009349019266664982, + 0.009007307700812817, + 0.012908455915749073, + 0.006128818262368441, + 0.01398028340190649, + 0.012459546327590942, + 0.01734100468456745, + 0.01097018737345934, + 0.008301698602735996, + 0.007501206360757351, + 0.009334532544016838, + 0.013031582348048687, + 0.007815670222043991, + 0.006886153947561979, + 0.007843629457056522, + 0.013838502578437328, + 0.005251911468803883, + 0.003646396566182375, + 0.0055643292143940926, + 0.0030197894666343927, + 0.0037410324439406395, + 0.004735113121569157, + 0.003390378085896373, + 0.005775409284979105, + 0.020455239340662956, + 0.017696289345622063, + 0.016362471505999565, + 0.017038198187947273, + 0.019361088052392006, + 0.011425654403865337, + 0.010665892623364925, + 0.01442293357104063, + 0.007625808473676443, + 0.005230489186942577, + 0.012053587473928928, + 0.011001507751643658, + 0.008733509108424187, + 0.02211848273873329, + 0.010743881575763226, + 0.01118396408855915, + 0.001966496231034398, + 0.004567199852317572, + 0.0033866222947835922, + 0.003912007436156273, + 0.0018697238992899656, + 0.003723833244293928, + 0.005410011857748032, + 0.002301736967638135, + 0.012988208793103695, + 0.009957867674529552, + 0.0066503072157502174, + 0.009098431095480919, + 0.005506286397576332, + 0.008442128077149391, + 0.004203618038445711, + 0.007783372886478901, + 0.004615756683051586, + 0.004097620490938425, + 0.003496614284813404, + 0.004063527565449476, + 0.003791222581639886, + 0.004229316487908363, + 0.003062029369175434, + 0.003518400015309453, + 0.0212965440005064, + 0.016276424750685692, + 0.14796653389930725, + 0.01835566572844982, + 0.013047868385910988, + 0.16574160754680634, + 0.012844330631196499, + 0.012804991565644741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.3134394884109497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098943900316954 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004098943900316954, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098943900316954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14644281566143036, + 0.09757420420646667, + 0.06459202617406845, + 0.026967965066432953, + 0.08627621829509735, + 0.09948746114969254, + 0.09846552461385727, + 0.3237318694591522, + 0.022502444684505463, + 0.033959437161684036 + ], + "credit_map": [ + 0.01845567300915718, + 0.016437681391835213, + 0.016163505613803864, + 0.012352721765637398, + 0.023197675123810768, + 0.0276468675583601, + 0.012810335494577885, + 0.01937837339937687, + 0.010068071074783802, + 0.009434537030756474, + 0.010088033974170685, + 0.006648194044828415, + 0.012540396302938461, + 0.013574143871665001, + 0.015458662994205952, + 0.01976216956973076, + 0.0079690245911479, + 0.006407090928405523, + 0.008604936301708221, + 0.007400042377412319, + 0.007256816606968641, + 0.00693871034309268, + 0.00763052050024271, + 0.012384888716042042, + 0.0040158722549676895, + 0.002616914687678218, + 0.0039990246295928955, + 0.0023227306082844734, + 0.003996407613158226, + 0.0030289043206721544, + 0.002694629831239581, + 0.004293483681976795, + 0.013392031192779541, + 0.008370225317776203, + 0.010828346014022827, + 0.01088043488562107, + 0.008963162079453468, + 0.009989278391003609, + 0.0063956016674637794, + 0.01745714619755745, + 0.010560099966824055, + 0.009958636946976185, + 0.01211524847894907, + 0.00913233682513237, + 0.011441945098340511, + 0.00846037082374096, + 0.014236518181860447, + 0.02358231507241726, + 0.00920991599559784, + 0.00972449779510498, + 0.012761103920638561, + 0.02052653208374977, + 0.0074126385152339935, + 0.01485473569482565, + 0.010586971417069435, + 0.01338913757354021, + 0.058654434978961945, + 0.03602497652173042, + 0.038773782551288605, + 0.06476397067308426, + 0.028021518141031265, + 0.03770344331860542, + 0.023732034489512444, + 0.03605770692229271, + 0.0030288221314549446, + 0.0021271812729537487, + 0.002794741652905941, + 0.003918780013918877, + 0.002714532660320401, + 0.0027799508534371853, + 0.002126449253410101, + 0.003011986380442977, + 0.007139509078115225, + 0.004010546486824751, + 0.006597590632736683, + 0.003764975816011429, + 0.0032368989195674658, + 0.004383484832942486, + 0.002336606616154313, + 0.0024898278061300516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.004098943900316954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076345098204910755 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.0076345098204910755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076345098204910755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11264870315790176, + 0.0628272294998169, + 0.03832756355404854, + 0.04089309275150299, + 0.06055638939142227, + 0.1873970627784729, + 0.08641214668750763, + 0.1041119247674942, + 0.02974916435778141, + 0.27707669138908386 + ], + "credit_map": [ + 0.012096128426492214, + 0.016270924359560013, + 0.009918270632624626, + 0.008085716515779495, + 0.019023999571800232, + 0.016692418605089188, + 0.010375765152275562, + 0.020185478031635284, + 0.006581809837371111, + 0.004506666213274002, + 0.006704125087708235, + 0.0037503240164369345, + 0.010698135942220688, + 0.012094535864889622, + 0.008764063939452171, + 0.00972756464034319, + 0.004309662152081728, + 0.004673480987548828, + 0.005139369983226061, + 0.004491823725402355, + 0.005302272271364927, + 0.004322540480643511, + 0.0044934069737792015, + 0.005595010705292225, + 0.006314275786280632, + 0.0037025813944637775, + 0.005333741661161184, + 0.0038103926926851273, + 0.005425882525742054, + 0.005358709022402763, + 0.003880145261064172, + 0.007067367434501648, + 0.0058757225051522255, + 0.006283390801399946, + 0.007332661654800177, + 0.006158256437629461, + 0.00776447681710124, + 0.010409461334347725, + 0.005455913487821817, + 0.011276506818830967, + 0.012189137749373913, + 0.01955159567296505, + 0.021119290962815285, + 0.015368263237178326, + 0.03691760078072548, + 0.030482178553938866, + 0.02664848417043686, + 0.025120506063103676, + 0.007783305365592241, + 0.006078388076275587, + 0.011052471585571766, + 0.017934679985046387, + 0.005332919768989086, + 0.014085568487644196, + 0.012437967583537102, + 0.011706849560141563, + 0.019009824842214584, + 0.02016763761639595, + 0.012179439887404442, + 0.016927728429436684, + 0.008384112268686295, + 0.01136051770299673, + 0.0062830811366438866, + 0.009799587540328503, + 0.0047119976952672005, + 0.0024632110726088285, + 0.003244696883484721, + 0.004665331449359655, + 0.004593300633132458, + 0.0036081948783248663, + 0.002851317636668682, + 0.0036111152730882168, + 0.01599353924393654, + 0.013320318423211575, + 0.09719239175319672, + 0.013434866443276405, + 0.0082788011059165, + 0.11089072376489639, + 0.008923265151679516, + 0.009042778052389622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0076345098204910755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a soft-focus close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning extremely slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive yellow excavator with a weathered, dirt-covered metal body visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the excavator's mechanical arm rise steadily while its bucket tilts to cascade moist, dark-brown earth mixed with rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm afternoon sunlight with visible golden dust particles shimmering in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046363710425794125 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.0046363710425794125, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046363710425794125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18957611918449402, + 0.0836465060710907, + 0.04720376059412956, + 0.09661367535591125, + 0.03345348313450813, + 0.10622658580541611, + 0.047182776033878326, + 0.16374173760414124, + 0.048468317836523056, + 0.18388712406158447 + ], + "credit_map": [ + 0.017819173634052277, + 0.020403822883963585, + 0.01878974959254265, + 0.013872005976736546, + 0.030322935432195663, + 0.03958571329712868, + 0.019475098699331284, + 0.02930760383605957, + 0.00795744825154543, + 0.006298545747995377, + 0.007969938218593597, + 0.006119935307651758, + 0.012256358750164509, + 0.010555067099630833, + 0.01746666245162487, + 0.015022534877061844, + 0.0057019563391804695, + 0.006038249935954809, + 0.006075697019696236, + 0.0045182965695858, + 0.006405441556125879, + 0.006590587552636862, + 0.0034706578589975834, + 0.00840286910533905, + 0.01652754470705986, + 0.007609589956700802, + 0.012980695813894272, + 0.009149296209216118, + 0.007878356613218784, + 0.010551437735557556, + 0.008969918824732304, + 0.02294682338833809, + 0.004403033759444952, + 0.003889613552019, + 0.0038897236809134483, + 0.0033944197930395603, + 0.006148161366581917, + 0.004529865458607674, + 0.002812544582411647, + 0.004386117681860924, + 0.00623052753508091, + 0.00864415429532528, + 0.012640329077839851, + 0.007627780083566904, + 0.024713782593607903, + 0.022144880145788193, + 0.015075074508786201, + 0.009150046855211258, + 0.004142757039517164, + 0.0045008049346506596, + 0.005798555910587311, + 0.008989417925477028, + 0.0031944371294230223, + 0.008267839439213276, + 0.0055001238361001015, + 0.0067888349294662476, + 0.027949295938014984, + 0.02201744168996811, + 0.019587641581892967, + 0.03467477113008499, + 0.012943390756845474, + 0.01662316359579563, + 0.012374703772366047, + 0.017571313306689262, + 0.00824920553714037, + 0.004340867046266794, + 0.005126361269503832, + 0.007318650372326374, + 0.007479283958673477, + 0.0061895051039755344, + 0.00460236007347703, + 0.005162075161933899, + 0.011530257761478424, + 0.009578628465533257, + 0.06391675025224686, + 0.008509562350809574, + 0.007737595587968826, + 0.06969117373228073, + 0.006041374988853931, + 0.006881756242364645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0046363710425794125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_tensor([2225, 2225])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_tensor([2225, 2225])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f4015d00c5b21255a43c2c52b5eead5afad8255b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_tensor([2225, 2225])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, graceful 360-degree drone orbit begins, circling a lone, ancient olive tree with a steady, reverent motion. The camera begins at a slightly lower angle, emphasizing the tree's grandeur. The initial view focuses on the massive, gnarled trunk, its bark an intricate tapestry of deep cracks and twisted muscle structures in weathered grays and earthy browns. As the camera slides counterclockwise, the warm, golden light of the Mediterranean afternoon captures the intricate textures, revealing patches of moss in the cracks. The orbit gradually widens, revealing the tree's full form. Its dense, sinuous branches reach toward the sky like arthritic fingers, adorned with dense clusters of tiny, silvery-green leaves that shimmer and dance in the gentle, warm breeze. Light filters through the leaves, casting speckled, shifting patterns on the sun-drenched terracotta ground below. In the background, the camera's circular path reveals a landscape of rolling, sun-drenched hills, with the deep blue Aegean Sea shimmering in the hazy distance. The atmosphere is one of profound peace and timeless tranquility, a cinematic portrayal of nature's endurance and quiet majesty.", + "seed": 25760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_2225_seed_25760_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027981312992051244, + "vlm_logprob": -0.027981312992051244, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.027981312992051244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006865720264613628 + ], + "opposite_logprob_tokens": [ + -7.006865501403809 + ], + "logprob_reward": -0.006865720264613628, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.006865720264613628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0996546521782875, + 0.0398787185549736, + 0.029218226671218872, + 0.023777704685926437, + 0.0466897189617157, + 0.04116949811577797, + 0.45544594526290894, + 0.059941425919532776, + 0.057576984167099, + 0.14664718508720398 + ], + "credit_map": [ + 0.01201846357434988, + 0.014088696800172329, + 0.012025151401758194, + 0.011115814559161663, + 0.015540162101387978, + 0.010571770370006561, + 0.01293853484094143, + 0.011356048285961151, + 0.004004927817732096, + 0.007616931106895208, + 0.0036784661933779716, + 0.008104615844786167, + 0.00387712474912405, + 0.004659744910895824, + 0.004408624023199081, + 0.0035282792523503304, + 0.002009512623772025, + 0.003934471867978573, + 0.002625806024298072, + 0.0034800698049366474, + 0.003030484775081277, + 0.0057027400471270084, + 0.003537944285199046, + 0.0048971958458423615, + 0.002313825301826, + 0.0036483213771134615, + 0.0038215755484998226, + 0.003630066756159067, + 0.0022332745138555765, + 0.00288517982698977, + 0.0024525055196136236, + 0.002792955143377185, + 0.004156817682087421, + 0.008242393843829632, + 0.007933586835861206, + 0.005240523722022772, + 0.006371484138071537, + 0.005671099293977022, + 0.004710430279374123, + 0.004363376181572676, + 0.005336320493370295, + 0.004114104900509119, + 0.006850316654890776, + 0.0043062795884907246, + 0.004923421889543533, + 0.005886520259082317, + 0.0039458926767110825, + 0.005806636530905962, + 0.14658279716968536, + 0.052017468959093094, + 0.015519185923039913, + 0.014998754486441612, + 0.04421880468726158, + 0.1449420005083084, + 0.014352809637784958, + 0.022814083844423294, + 0.0055714864283800125, + 0.011653727851808071, + 0.005877777934074402, + 0.00881920661777258, + 0.008700343780219555, + 0.007760557811707258, + 0.005536447744816542, + 0.006021871231496334, + 0.004857752006500959, + 0.010514962486922741, + 0.006820682901889086, + 0.008355608209967613, + 0.006512469612061977, + 0.00952970515936613, + 0.005285400897264481, + 0.005700397305190563, + 0.020841730758547783, + 0.018369784578680992, + 0.020425621420145035, + 0.017533065751194954, + 0.023107388988137245, + 0.01886996068060398, + 0.009238678961992264, + 0.018260931596159935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.006865720264613628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495339259505272 + ], + "opposite_logprob_tokens": [ + -2.9299533367156982 + ], + "logprob_reward": -0.05495339259505272, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495339259505272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08659029752016068, + 0.02145427279174328, + 0.07371451705694199, + 0.06341587752103806, + 0.059232089668512344, + 0.05199059844017029, + 0.31765416264533997, + 0.06821636855602264, + 0.18124184012413025, + 0.07648999989032745 + ], + "credit_map": [ + 0.010805686935782433, + 0.011733504943549633, + 0.018299875780940056, + 0.00792632345110178, + 0.007853501476347446, + 0.006697745993733406, + 0.014090551063418388, + 0.00918310321867466, + 0.0024805841967463493, + 0.0023084678687155247, + 0.0019821918103843927, + 0.0026588537730276585, + 0.0029201230499893427, + 0.004369428381323814, + 0.0026973565109074116, + 0.0020372660364955664, + 0.004708508029580116, + 0.009574157185852528, + 0.006569560617208481, + 0.010957784950733185, + 0.009614153765141964, + 0.014401305466890335, + 0.009417771361768246, + 0.008471271954476833, + 0.007389678619801998, + 0.009149235673248768, + 0.0056547559797763824, + 0.012839009054005146, + 0.005369889549911022, + 0.009077919647097588, + 0.0067965262569487095, + 0.007138863205909729, + 0.006526407785713673, + 0.008976307697594166, + 0.005338979419320822, + 0.00849552545696497, + 0.007102756295353174, + 0.0068891616538167, + 0.008475622162222862, + 0.007427328731864691, + 0.005526352673768997, + 0.004313899669796228, + 0.006109900306910276, + 0.008404307998716831, + 0.006725159473717213, + 0.005390570964664221, + 0.008773140609264374, + 0.006747268605977297, + 0.09589380025863647, + 0.03820723295211792, + 0.01112457551062107, + 0.010659441351890564, + 0.03589404746890068, + 0.0992303267121315, + 0.010576190426945686, + 0.016068538650870323, + 0.006273018196225166, + 0.015309644863009453, + 0.004961673635989428, + 0.0064049349166452885, + 0.007255073171108961, + 0.014291442930698395, + 0.004844005219638348, + 0.008876573294401169, + 0.01782686449587345, + 0.03142116218805313, + 0.017966553568840027, + 0.02455294318497181, + 0.013957108370959759, + 0.029993414878845215, + 0.015169075690209866, + 0.03035472333431244, + 0.015473333187401295, + 0.010469547472894192, + 0.010530367493629456, + 0.009033082984387875, + 0.007988478988409042, + 0.011526806280016899, + 0.004074301104992628, + 0.0073940809816122055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.05495339259505272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028272203635424376 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028272203635424376, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028272203635424376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10590878129005432, + 0.028902459889650345, + 0.03950688615441322, + 0.018900714814662933, + 0.06368710845708847, + 0.022249232977628708, + 0.5465036630630493, + 0.059901751577854156, + 0.044487688690423965, + 0.06995165348052979 + ], + "credit_map": [ + 0.013733325526118279, + 0.01581921987235546, + 0.015230964869260788, + 0.012833223678171635, + 0.008957304060459137, + 0.014823236502707005, + 0.012195499613881111, + 0.012316000647842884, + 0.002903858432546258, + 0.004707658663392067, + 0.0024442512076348066, + 0.004426008090376854, + 0.003103633178398013, + 0.0035976578947156668, + 0.004440453834831715, + 0.0032789395190775394, + 0.0035203243605792522, + 0.008279153145849705, + 0.00356823718175292, + 0.004466746002435684, + 0.005345590878278017, + 0.003346499055624008, + 0.00637842109426856, + 0.004601913504302502, + 0.002022964647039771, + 0.0028116011526435614, + 0.002523166360333562, + 0.0030395113863050938, + 0.0025505181401968002, + 0.002183612436056137, + 0.001938110333867371, + 0.001831230241805315, + 0.006592855788767338, + 0.01016250066459179, + 0.014410738833248615, + 0.006415626034140587, + 0.007772715762257576, + 0.005442220717668533, + 0.006969546899199486, + 0.005920907016843557, + 0.002434307476505637, + 0.002915129764005542, + 0.0018872926011681557, + 0.002851012395694852, + 0.004845007788389921, + 0.002494861138984561, + 0.002697159070521593, + 0.002124462276697159, + 0.1872391700744629, + 0.05877741426229477, + 0.020604556426405907, + 0.010249750688672066, + 0.04824981093406677, + 0.1783875674009323, + 0.015250466763973236, + 0.027744915336370468, + 0.006685751955956221, + 0.010999157093465328, + 0.005632709711790085, + 0.0077378773130476475, + 0.009603927843272686, + 0.0074428655207157135, + 0.006064610555768013, + 0.005734850186854601, + 0.004662098828703165, + 0.007042858749628067, + 0.004790483508259058, + 0.006273019593209028, + 0.003913187421858311, + 0.006142020225524902, + 0.004236914683133364, + 0.007427106611430645, + 0.009655122645199299, + 0.008894138038158417, + 0.007174049969762564, + 0.005665669217705727, + 0.009507468901574612, + 0.009811270982027054, + 0.005121105816215277, + 0.01412283070385456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0028272203635424376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12080847471952438, + 0.03422943875193596, + 0.03275756537914276, + 0.041132159531116486, + 0.039668064564466476, + 0.03822501376271248, + 0.0890398919582367, + 0.12804989516735077, + 0.40182098746299744, + 0.07426852732896805 + ], + "credit_map": [ + 0.01302630826830864, + 0.018398772925138474, + 0.014361499808728695, + 0.01550816185772419, + 0.009748287498950958, + 0.01977381855249405, + 0.01461868267506361, + 0.01537294127047062, + 0.0033717688638716936, + 0.004518758971244097, + 0.0038378844037652016, + 0.00530356215313077, + 0.0030492476653307676, + 0.004302039742469788, + 0.004581674467772245, + 0.005264502950012684, + 0.002537595573812723, + 0.005238295532763004, + 0.0022186292335391045, + 0.005021366290748119, + 0.0036601892206817865, + 0.006206533871591091, + 0.004159665200859308, + 0.0037152920849621296, + 0.003907529637217522, + 0.006594102364033461, + 0.0039041724521666765, + 0.005864438600838184, + 0.005291962530463934, + 0.0049547431990504265, + 0.006789232138544321, + 0.0038259816356003284, + 0.004045023582875729, + 0.005330427549779415, + 0.0074861799366772175, + 0.0034297346137464046, + 0.0039043156430125237, + 0.006383138708770275, + 0.0035583123099058867, + 0.005530932918190956, + 0.004059045575559139, + 0.004309225827455521, + 0.004440347198396921, + 0.006018110550940037, + 0.0053536067716777325, + 0.004545454401522875, + 0.004202450159937143, + 0.005296777002513409, + 0.024524826556444168, + 0.009289350360631943, + 0.004374450072646141, + 0.004619548097252846, + 0.01008391659706831, + 0.02395441383123398, + 0.004022631328552961, + 0.008170755580067635, + 0.014909062534570694, + 0.018868904560804367, + 0.015725357457995415, + 0.010891527868807316, + 0.015348406508564949, + 0.021611664444208145, + 0.008901573717594147, + 0.02179340086877346, + 0.031734611839056015, + 0.06388701498508453, + 0.035001836717128754, + 0.03717326000332832, + 0.04608127102255821, + 0.05719926580786705, + 0.033601850271224976, + 0.09714183956384659, + 0.016564778983592987, + 0.011179842986166477, + 0.008181954734027386, + 0.0053500887006521225, + 0.011348015628755093, + 0.011419855989515781, + 0.004100222140550613, + 0.00612376956269145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.005249644163995981 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001058757305145 + ], + "opposite_logprob_tokens": [ + -2.6950106620788574 + ], + "logprob_reward": -0.07001058757305145, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001058757305145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10347171872854233, + 0.05284533277153969, + 0.03456679731607437, + 0.039156027138233185, + 0.03671741858124733, + 0.03810112178325653, + 0.490373432636261, + 0.04798087105154991, + 0.0940924733877182, + 0.06269485503435135 + ], + "credit_map": [ + 0.012344108894467354, + 0.010152922011911869, + 0.018469195812940598, + 0.009308128617703915, + 0.012464614585042, + 0.011289526708424091, + 0.017011679708957672, + 0.012431545183062553, + 0.004140450153499842, + 0.005597219802439213, + 0.006853730417788029, + 0.0063933879137039185, + 0.006439106538891792, + 0.009234894998371601, + 0.007756509352475405, + 0.006430032197386026, + 0.0025885417126119137, + 0.004950536880642176, + 0.0031000084709376097, + 0.0043897442519664764, + 0.0033711211290210485, + 0.0064187804237008095, + 0.004461288917809725, + 0.00528677087277174, + 0.004205821547657251, + 0.006427191663533449, + 0.0039653475396335125, + 0.009640964679419994, + 0.0033027080353349447, + 0.003348961006850004, + 0.004939809907227755, + 0.003325219266116619, + 0.00380416726693511, + 0.0033100994769483805, + 0.005272091366350651, + 0.0030151482205837965, + 0.0036620027385652065, + 0.007404678501188755, + 0.004388092085719109, + 0.005861136130988598, + 0.004617912229150534, + 0.0048562465235590935, + 0.004012908320873976, + 0.003560012672096491, + 0.009898661635816097, + 0.004367580637335777, + 0.003096431028097868, + 0.003691365709528327, + 0.1697932481765747, + 0.05201699212193489, + 0.01781260222196579, + 0.009125708602368832, + 0.042103178799152374, + 0.15998029708862305, + 0.014493513852357864, + 0.02504785545170307, + 0.004360755439847708, + 0.009850781410932541, + 0.004296779632568359, + 0.00601960439234972, + 0.007827739231288433, + 0.00560917379334569, + 0.004846643190830946, + 0.00516938790678978, + 0.008346989750862122, + 0.013221525587141514, + 0.011630350723862648, + 0.011303362436592579, + 0.008163080550730228, + 0.013472137041389942, + 0.006838895846158266, + 0.021116124466061592, + 0.008782128803431988, + 0.009264130145311356, + 0.007473795674741268, + 0.00793694332242012, + 0.007773588411509991, + 0.00895658042281866, + 0.004816926084458828, + 0.007690758910030127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.07001058757305145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, graceful 360-degree drone orbit begins, circling a lone, ancient olive tree with a steady, reverent motion. The camera begins at a slightly lower angle, emphasizing the tree's grandeur. The initial view focuses on the massive, gnarled trunk, its bark an intricate tapestry of deep cracks and twisted muscle structures in weathered grays and earthy browns. As the camera slides counterclockwise, the warm, golden light of the Mediterranean afternoon captures the intricate textures, revealing patches of moss in the cracks. The orbit gradually widens, revealing the tree's full form. Its dense, sinuous branches reach toward the sky like arthritic fingers, adorned with dense clusters of tiny, silvery-green leaves that shimmer and dance in the gentle, warm breeze. Light filters through the leaves, casting speckled, shifting patterns on the sun-drenched terracotta ground below. In the background, the camera's circular path reveals a landscape of rolling, sun-drenched hills, with the deep blue Aegean Sea shimmering in the hazy distance. The atmosphere is one of profound peace and timeless tranquility, a cinematic portrayal of nature's endurance and quiet majesty.", + "seed": 25761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_2225_seed_25761_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033934193197637796, + "vlm_logprob": -0.033934193197637796, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.033934193197637796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008439944125711918 + ], + "opposite_logprob_tokens": [ + -7.133440017700195 + ], + "logprob_reward": -0.008439944125711918, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.008439944125711918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1108952984213829, + 0.08470267802476883, + 0.11641988158226013, + 0.07188853621482849, + 0.18319286406040192, + 0.12813659012317657, + 0.06135479360818863, + 0.05908457934856415, + 0.04593644291162491, + 0.13838839530944824 + ], + "credit_map": [ + 0.007367555983364582, + 0.016043633222579956, + 0.011793417856097221, + 0.018261520192027092, + 0.01507831085473299, + 0.01626555807888508, + 0.012104985304176807, + 0.013980305753648281, + 0.008820761926472187, + 0.010987152345478535, + 0.006047992967069149, + 0.010473176836967468, + 0.009122490882873535, + 0.019192546606063843, + 0.012265138328075409, + 0.007793408818542957, + 0.005776199512183666, + 0.033231671899557114, + 0.004081250634044409, + 0.020878136157989502, + 0.007636573631316423, + 0.02883991412818432, + 0.005140300374478102, + 0.010835828259587288, + 0.007019814569503069, + 0.0113729452714324, + 0.006100844591856003, + 0.010444224812090397, + 0.015612398274242878, + 0.006251380313187838, + 0.009130221791565418, + 0.00595669262111187, + 0.010331123135983944, + 0.010559385642409325, + 0.008107014000415802, + 0.026602530851960182, + 0.012207197956740856, + 0.01306151319295168, + 0.07260756939649582, + 0.029716501012444496, + 0.009758146479725838, + 0.005854631774127483, + 0.006562442984431982, + 0.005955846048891544, + 0.008553571067750454, + 0.03547005355358124, + 0.007143398746848106, + 0.048838481307029724, + 0.00953458622097969, + 0.006311256438493729, + 0.009101933799684048, + 0.00901820044964552, + 0.008344741538167, + 0.009219481609761715, + 0.004869104828685522, + 0.004955483600497246, + 0.0093068303540349, + 0.0061182621866464615, + 0.005308229476213455, + 0.007765849120914936, + 0.0105854831635952, + 0.006296234671026468, + 0.004564119502902031, + 0.009139562025666237, + 0.00562002370133996, + 0.004752737004309893, + 0.005036405753344297, + 0.007035655900835991, + 0.005995756946504116, + 0.00437269126996398, + 0.004181584808975458, + 0.008941583335399628, + 0.01119259838014841, + 0.022841328755021095, + 0.011381248943507671, + 0.026491258293390274, + 0.018899910151958466, + 0.008253222331404686, + 0.020704450085759163, + 0.01862436532974243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.008439944125711918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865296185016632 + ], + "opposite_logprob_tokens": [ + -3.0486528873443604 + ], + "logprob_reward": -0.04865296185016632, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865296185016632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10142958909273148, + 0.05937328189611435, + 0.17864184081554413, + 0.19641166925430298, + 0.13732171058654785, + 0.061617348343133926, + 0.09738711267709732, + 0.05961824208498001, + 0.04367285221815109, + 0.06452637910842896 + ], + "credit_map": [ + 0.009901788085699081, + 0.014255781657993793, + 0.014710239134728909, + 0.013375145383179188, + 0.011340329423546791, + 0.012655580416321754, + 0.011798624880611897, + 0.013392104767262936, + 0.005864403676241636, + 0.009476316161453724, + 0.006917155347764492, + 0.008883995935320854, + 0.007310202810913324, + 0.007461932022124529, + 0.008974683471024036, + 0.004484596196562052, + 0.00843552965670824, + 0.05472457408905029, + 0.00583958625793457, + 0.034293804317712784, + 0.008571897633373737, + 0.04325136914849281, + 0.00728593161329627, + 0.01623915135860443, + 0.016503529623150826, + 0.019063828513026237, + 0.012630442157387733, + 0.029921505600214005, + 0.0670946016907692, + 0.011394781060516834, + 0.01658829301595688, + 0.023214703425765038, + 0.008303795009851456, + 0.009024196304380894, + 0.005464013200253248, + 0.020266467705368996, + 0.008567615412175655, + 0.008370460011065006, + 0.05320041999220848, + 0.024124741554260254, + 0.006516923196613789, + 0.007814246229827404, + 0.006091360468417406, + 0.009097050875425339, + 0.007680036127567291, + 0.005291529931128025, + 0.005775892175734043, + 0.01335031259804964, + 0.015449011698365211, + 0.007955634035170078, + 0.01396971009671688, + 0.012410945259034634, + 0.012254109606146812, + 0.018483754247426987, + 0.007738127838820219, + 0.009125827811658382, + 0.009748950600624084, + 0.007854561321437359, + 0.007796530611813068, + 0.007327181752771139, + 0.008429784327745438, + 0.006535937078297138, + 0.006129814777523279, + 0.005795489996671677, + 0.005298065021634102, + 0.005025253165513277, + 0.003941720351576805, + 0.006847218610346317, + 0.004327008500695229, + 0.005495642777532339, + 0.005271507427096367, + 0.0074664391577243805, + 0.00429287226870656, + 0.008213166147470474, + 0.0036474158987402916, + 0.016936523839831352, + 0.006585579831153154, + 0.006863223388791084, + 0.007856699638068676, + 0.01013090368360281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.04865296185016632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649542272090912 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649542272090912, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649542272090912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11661434918642044, + 0.015680143609642982, + 0.03497762978076935, + 0.4359745681285858, + 0.1687946915626526, + 0.022342504933476448, + 0.08411256968975067, + 0.015070595778524876, + 0.07090282440185547, + 0.035530172288417816 + ], + "credit_map": [ + 0.007097010035067797, + 0.026654819026589394, + 0.006070131901651621, + 0.019608095288276672, + 0.011839023791253567, + 0.015029863454401493, + 0.010128319263458252, + 0.020187076181173325, + 0.0017144294688478112, + 0.0021557549480348825, + 0.0017295668367296457, + 0.0023268910590559244, + 0.0024476603139191866, + 0.0018900949507951736, + 0.0017530412878841162, + 0.001662704860791564, + 0.0021631778217852116, + 0.009538996033370495, + 0.002042202278971672, + 0.0059684813022613525, + 0.0027842794079333544, + 0.004717180039733648, + 0.004087961744517088, + 0.0036753469612449408, + 0.034703902900218964, + 0.031290292739868164, + 0.011816572397947311, + 0.06757532805204391, + 0.1747494339942932, + 0.01905326545238495, + 0.03221110254526138, + 0.06457464396953583, + 0.01026882603764534, + 0.00965446513146162, + 0.008039048872888088, + 0.025408117100596428, + 0.01042422279715538, + 0.013190038502216339, + 0.06608263403177261, + 0.025727324187755585, + 0.0024074374232441187, + 0.0018664541421458125, + 0.0019745579920709133, + 0.0022199973464012146, + 0.0018096626736223698, + 0.004624133929610252, + 0.0027729268185794353, + 0.004667333327233791, + 0.009331343695521355, + 0.007525336928665638, + 0.007750739809125662, + 0.010114685632288456, + 0.013073251582682133, + 0.015601802617311478, + 0.0072339787147939205, + 0.01348142046481371, + 0.002090394264087081, + 0.0017049092566594481, + 0.0014385613612830639, + 0.0020606210455298424, + 0.0024569903034716845, + 0.0018381798872724175, + 0.0015495859552174807, + 0.0019313533557578921, + 0.007876790128648281, + 0.007224210072308779, + 0.007189499214291573, + 0.010802234522998333, + 0.007286873646080494, + 0.010005808435380459, + 0.009087883867323399, + 0.01142952125519514, + 0.0029641396831721067, + 0.00475967675447464, + 0.0034868656657636166, + 0.00773664191365242, + 0.003511491697281599, + 0.003544440260156989, + 0.004506202414631844, + 0.005020711570978165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.004649542272090912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076268198899924755 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.0076268198899924755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076268198899924755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16786408424377441, + 0.12692387402057648, + 0.034017257392406464, + 0.3110116720199585, + 0.09931515157222748, + 0.0871516689658165, + 0.06545683741569519, + 0.01802893541753292, + 0.02468552254140377, + 0.06554505974054337 + ], + "credit_map": [ + 0.009682671166956425, + 0.03372208774089813, + 0.014100980013608932, + 0.026953576132655144, + 0.02350618690252304, + 0.018082328140735626, + 0.018469378352165222, + 0.023346854373812675, + 0.01623472385108471, + 0.013596457429230213, + 0.01619502529501915, + 0.013053481467068195, + 0.011726456694304943, + 0.029654251411557198, + 0.014225320890545845, + 0.012238143011927605, + 0.0029785488732159138, + 0.004446541424840689, + 0.004554983228445053, + 0.003778028069064021, + 0.0036956716794520617, + 0.0037699320819228888, + 0.00435233861207962, + 0.006441210862249136, + 0.02459447644650936, + 0.02257687598466873, + 0.008190377615392208, + 0.051535964012145996, + 0.12622545659542084, + 0.014674384146928787, + 0.022245852276682854, + 0.040968239307403564, + 0.005499666556715965, + 0.005772541742771864, + 0.006049346178770065, + 0.016078662127256393, + 0.0062810382805764675, + 0.00819811224937439, + 0.03774476423859596, + 0.013691007159650326, + 0.00617418996989727, + 0.002817961387336254, + 0.002672328148037195, + 0.0038120471872389317, + 0.005299758166074753, + 0.02757386490702629, + 0.00341273145750165, + 0.035388778895139694, + 0.013160016387701035, + 0.007896460592746735, + 0.0056985970586538315, + 0.007848314940929413, + 0.010311618447303772, + 0.007211187854409218, + 0.005232159048318863, + 0.008098473772406578, + 0.0019096744945272803, + 0.002288163872435689, + 0.0014161652652546763, + 0.0023966971784830093, + 0.0035886322148144245, + 0.002428613603115082, + 0.001952688442543149, + 0.0020482975523918867, + 0.0031876685097813606, + 0.0027761971578001976, + 0.002716204384341836, + 0.003589235246181488, + 0.002510524820536375, + 0.0026210180949419737, + 0.0029969397000968456, + 0.004287732765078545, + 0.00454690121114254, + 0.008595761843025684, + 0.004369501955807209, + 0.01968613639473915, + 0.004265804309397936, + 0.005212221294641495, + 0.007837642915546894, + 0.011031079106032848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0076268198899924755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030169785022736 + ], + "opposite_logprob_tokens": [ + -2.350301742553711 + ], + "logprob_reward": -0.10030169785022736, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030169785022736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08321188390254974, + 0.05213005468249321, + 0.21271982789039612, + 0.25767582654953003, + 0.047874126583337784, + 0.09297425299882889, + 0.045772485435009, + 0.03886893764138222, + 0.038909297436475754, + 0.12986336648464203 + ], + "credit_map": [ + 0.005110179074108601, + 0.01593414880335331, + 0.009316202253103256, + 0.016846254467964172, + 0.010774784721434116, + 0.006663093343377113, + 0.010158932767808437, + 0.008408280089497566, + 0.004744295496493578, + 0.008960817009210587, + 0.0050694928504526615, + 0.009438036940991879, + 0.007607440929859877, + 0.005585830193012953, + 0.007197093218564987, + 0.003527044551447034, + 0.010118753649294376, + 0.068287193775177, + 0.005692233797162771, + 0.040360115468502045, + 0.010745400562882423, + 0.04662186652421951, + 0.011792376637458801, + 0.019101856276392937, + 0.020826546475291252, + 0.017512978985905647, + 0.008708001114428043, + 0.04122787341475487, + 0.10312892496585846, + 0.010466777719557285, + 0.017905540764331818, + 0.03789917752146721, + 0.0029565070290118456, + 0.004488151054829359, + 0.0028122260700911283, + 0.0056532034650444984, + 0.004808373283594847, + 0.0034320587292313576, + 0.01483433973044157, + 0.00888926163315773, + 0.00693737855181098, + 0.0036709257401525974, + 0.003845331259071827, + 0.006004732567816973, + 0.0059064836241304874, + 0.024566803127527237, + 0.004991038236767054, + 0.03705155476927757, + 0.005826318170875311, + 0.003521617967635393, + 0.005184393376111984, + 0.005555400159209967, + 0.00544034643098712, + 0.009781783446669579, + 0.0037693833000957966, + 0.006693234201520681, + 0.006816990207880735, + 0.004195937886834145, + 0.003643261967226863, + 0.0049603767693042755, + 0.005863551516085863, + 0.0035614650696516037, + 0.0045814430341124535, + 0.00524591188877821, + 0.003747473005205393, + 0.0031340874265879393, + 0.0036606122739613056, + 0.006334479432553053, + 0.006364426575601101, + 0.004219865892082453, + 0.005281939636915922, + 0.006166407838463783, + 0.007708475925028324, + 0.021117134019732475, + 0.008572553284466267, + 0.03191734477877617, + 0.013106037862598896, + 0.015227086842060089, + 0.013112393207848072, + 0.01910233497619629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.10030169785022736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, graceful 360-degree drone orbit begins, circling a lone, ancient olive tree with a steady, reverent motion. The camera begins at a slightly lower angle, emphasizing the tree's grandeur. The initial view focuses on the massive, gnarled trunk, its bark an intricate tapestry of deep cracks and twisted muscle structures in weathered grays and earthy browns. As the camera slides counterclockwise, the warm, golden light of the Mediterranean afternoon captures the intricate textures, revealing patches of moss in the cracks. The orbit gradually widens, revealing the tree's full form. Its dense, sinuous branches reach toward the sky like arthritic fingers, adorned with dense clusters of tiny, silvery-green leaves that shimmer and dance in the gentle, warm breeze. Light filters through the leaves, casting speckled, shifting patterns on the sun-drenched terracotta ground below. In the background, the camera's circular path reveals a landscape of rolling, sun-drenched hills, with the deep blue Aegean Sea shimmering in the hazy distance. The atmosphere is one of profound peace and timeless tranquility, a cinematic portrayal of nature's endurance and quiet majesty.", + "seed": 25762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_2225_seed_25762_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03159095114096999, + "vlm_logprob": -0.03159095114096999, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03159095114096999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007664203178137541 + ], + "opposite_logprob_tokens": [ + -6.382664203643799 + ], + "logprob_reward": -0.007664203178137541, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.007664203178137541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061053317040205, + 0.03680570051074028, + 0.05365964397788048, + 0.168390154838562, + 0.01848267763853073, + 0.027440138161182404, + 0.12360379844903946, + 0.46452346444129944, + 0.025146501138806343, + 0.020894577726721764 + ], + "credit_map": [ + 0.0074827098287642, + 0.00824330747127533, + 0.005411852616816759, + 0.013605195097625256, + 0.009006948210299015, + 0.005198301747441292, + 0.007078741677105427, + 0.005026267841458321, + 0.0050792996771633625, + 0.005569953937083483, + 0.004838569089770317, + 0.006146659143269062, + 0.003296105656772852, + 0.004127535969018936, + 0.0043995738960802555, + 0.0033480022102594376, + 0.006371332332491875, + 0.006496698595583439, + 0.008359433151781559, + 0.008337127976119518, + 0.006872597616165876, + 0.006212404929101467, + 0.005280231591314077, + 0.00572982057929039, + 0.010639695450663567, + 0.012381975539028645, + 0.01021327544003725, + 0.01204078458249569, + 0.01686478778719902, + 0.058074403554201126, + 0.019609814509749413, + 0.028565427288413048, + 0.001666335272602737, + 0.00255118147470057, + 0.0017326544038951397, + 0.0013627596199512482, + 0.003833665046840906, + 0.0030472620856016874, + 0.001963478047400713, + 0.0023253445979207754, + 0.002028076909482479, + 0.004876068327575922, + 0.0023599937558174133, + 0.0029925014823675156, + 0.0047415997833013535, + 0.006705477833747864, + 0.001963799586519599, + 0.0017726224614307284, + 0.019700193777680397, + 0.020860087126493454, + 0.0120509322732687, + 0.009847208857536316, + 0.016035493463277817, + 0.020199351012706757, + 0.00877942144870758, + 0.016131119802594185, + 0.03304392844438553, + 0.1276596039533615, + 0.0433695949614048, + 0.028549853712320328, + 0.15360119938850403, + 0.030533887445926666, + 0.024266911670565605, + 0.023498516529798508, + 0.002379724523052573, + 0.0033945348113775253, + 0.0029146834276616573, + 0.0034456464927643538, + 0.002362872241064906, + 0.0030977551359683275, + 0.004642567131668329, + 0.002908718539401889, + 0.0023953234776854515, + 0.002881868975237012, + 0.0018140504835173488, + 0.004385843873023987, + 0.0016748845810070634, + 0.0025294849183410406, + 0.002920011756941676, + 0.0022931115236133337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.007664203178137541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865852743387222 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865852743387222, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865852743387222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054645344614982605, + 0.029269319027662277, + 0.08056844025850296, + 0.05228361114859581, + 0.1636519730091095, + 0.047640882432460785, + 0.12441752851009369, + 0.4249124526977539, + 0.013321585021913052, + 0.009288880974054337 + ], + "credit_map": [ + 0.007432348094880581, + 0.008410329930484295, + 0.0036632781848311424, + 0.017048968002200127, + 0.0066145798191428185, + 0.0038143687415868044, + 0.004342003725469112, + 0.0033194622956216335, + 0.004888201132416725, + 0.004595345817506313, + 0.004276410676538944, + 0.0038964690174907446, + 0.002578360727056861, + 0.0036664463113993406, + 0.0025352302473038435, + 0.0028328520711511374, + 0.005932505242526531, + 0.01671990007162094, + 0.006073656026273966, + 0.00971569400280714, + 0.011639065109193325, + 0.016725456342101097, + 0.009078641422092915, + 0.004683513659983873, + 0.004433894529938698, + 0.005209051538258791, + 0.005900131072849035, + 0.0054517160169780254, + 0.007179887033998966, + 0.00878399983048439, + 0.009166660718619823, + 0.006158267147839069, + 0.014635591767728329, + 0.0169170293956995, + 0.0073963552713394165, + 0.007011266425251961, + 0.014045263640582561, + 0.06076885387301445, + 0.011178234592080116, + 0.03169937804341316, + 0.0038750949315726757, + 0.005902042146772146, + 0.0038931304588913918, + 0.003893696004524827, + 0.007300190627574921, + 0.013410978950560093, + 0.003666982753202319, + 0.005698759108781815, + 0.01785990223288536, + 0.021804407238960266, + 0.010875836946070194, + 0.012834961526095867, + 0.01557864062488079, + 0.0193015169352293, + 0.011126638390123844, + 0.015035617165267467, + 0.03330142796039581, + 0.12948471307754517, + 0.02526070550084114, + 0.0229055006057024, + 0.15392938256263733, + 0.027498019859194756, + 0.014738332480192184, + 0.01779433898627758, + 0.0015587675152346492, + 0.001751936855725944, + 0.0015094142872840166, + 0.0016418633749708533, + 0.001181006315164268, + 0.0016546619590371847, + 0.002723153680562973, + 0.001300780102610588, + 0.0010978113859891891, + 0.0012332294136285782, + 0.0012271965388208628, + 0.0010168403387069702, + 0.0011984056327491999, + 0.0013723435113206506, + 0.0011012338800355792, + 0.0010418198071420193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.04865852743387222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052642300724983215 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.0052642300724983215, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052642300724983215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06224464625120163, + 0.0356549508869648, + 0.02926860749721527, + 0.09698529541492462, + 0.05120113119482994, + 0.027480971068143845, + 0.1147124171257019, + 0.5187085270881653, + 0.038264527916908264, + 0.025478975847363472 + ], + "credit_map": [ + 0.006847721058875322, + 0.0076913852244615555, + 0.004075910896062851, + 0.009257870726287365, + 0.014870689250528812, + 0.007310094777494669, + 0.006740792188793421, + 0.005450183991342783, + 0.0047994269989430904, + 0.0041750050149858, + 0.0036222226917743683, + 0.0056740776635706425, + 0.0048104096204042435, + 0.0026713020633906126, + 0.006262655835598707, + 0.003639848669990897, + 0.002991028828546405, + 0.004922385327517986, + 0.0025072747375816107, + 0.004896472208201885, + 0.004009665455669165, + 0.003641503630205989, + 0.003173317527398467, + 0.003126960713416338, + 0.006357681471854448, + 0.008172222413122654, + 0.006917634047567844, + 0.008418682962656021, + 0.009497437626123428, + 0.02884657308459282, + 0.012907586060464382, + 0.015867479145526886, + 0.004404331557452679, + 0.005278308875858784, + 0.0025730817578732967, + 0.002517199609428644, + 0.005003544967621565, + 0.018112828955054283, + 0.0026734303683042526, + 0.010638408362865448, + 0.0027518447022885084, + 0.004941459745168686, + 0.0024559791199862957, + 0.002646822016686201, + 0.004503575619310141, + 0.0037759544793516397, + 0.0026392622385174036, + 0.003766070120036602, + 0.01374779175966978, + 0.022091316059231758, + 0.009303136728703976, + 0.013294991105794907, + 0.014921179041266441, + 0.014216858893632889, + 0.014114589430391788, + 0.01302255317568779, + 0.03659585490822792, + 0.15835438668727875, + 0.034348271787166595, + 0.027758775278925896, + 0.19195863604545593, + 0.02976156212389469, + 0.019308239221572876, + 0.02062278985977173, + 0.0033461572602391243, + 0.005234462209045887, + 0.0039039331022650003, + 0.0052474187687039375, + 0.0032684309408068657, + 0.003731377189978957, + 0.008722977712750435, + 0.00480977026745677, + 0.003625276265665889, + 0.00365090393461287, + 0.0044921343214809895, + 0.0015322413528338075, + 0.0035589684266597033, + 0.004449443891644478, + 0.0023245092015713453, + 0.0018454977544024587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.0052642300724983215 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348982006311417 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.026348982006311417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348982006311417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05378851667046547, + 0.0475696325302124, + 0.03660225495696068, + 0.1371537148952484, + 0.21712510287761688, + 0.07014992833137512, + 0.1363547146320343, + 0.20432254672050476, + 0.03358294442296028, + 0.06335064768791199 + ], + "credit_map": [ + 0.00997445173561573, + 0.007789321709424257, + 0.003961763344705105, + 0.00871617253869772, + 0.008396894671022892, + 0.007031884044408798, + 0.003870863001793623, + 0.004047165624797344, + 0.006151420529931784, + 0.0076202345080673695, + 0.005660071037709713, + 0.006910430733114481, + 0.00440357206389308, + 0.0053782095201313496, + 0.006114627234637737, + 0.005331062711775303, + 0.0034141449723392725, + 0.003803614294156432, + 0.004834437742829323, + 0.006309539545327425, + 0.007909012958407402, + 0.003563478821888566, + 0.004138398915529251, + 0.002629623981192708, + 0.008868650533258915, + 0.009423659183084965, + 0.0057542393915355206, + 0.011125314049422741, + 0.012459320947527885, + 0.0564630925655365, + 0.0056321159936487675, + 0.027427323162555695, + 0.020022274926304817, + 0.02465764619410038, + 0.015123311430215836, + 0.011561148799955845, + 0.025102224200963974, + 0.06830207258462906, + 0.017279023304581642, + 0.03507741913199425, + 0.00636313809081912, + 0.007911649532616138, + 0.004169785417616367, + 0.004423124250024557, + 0.012986399233341217, + 0.020990043878555298, + 0.004452476743608713, + 0.008853308856487274, + 0.019046492874622345, + 0.022349322214722633, + 0.012906491756439209, + 0.011352716013789177, + 0.017869943752884865, + 0.020466741174459457, + 0.012507329694926739, + 0.0198556799441576, + 0.01639283262193203, + 0.06735121458768845, + 0.007343626581132412, + 0.01061373669654131, + 0.07982319593429565, + 0.011991648003458977, + 0.004807465709745884, + 0.005998826585710049, + 0.0026858430355787277, + 0.004146084189414978, + 0.00406577717512846, + 0.004951802548021078, + 0.003032670821994543, + 0.004456651397049427, + 0.00689495075494051, + 0.0033491654321551323, + 0.00929114781320095, + 0.00912349671125412, + 0.009519511833786964, + 0.003990057855844498, + 0.00856893602758646, + 0.01350310631096363, + 0.004535953514277935, + 0.004818432964384556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.026348982006311417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001881301403046 + ], + "opposite_logprob_tokens": [ + -2.695018768310547 + ], + "logprob_reward": -0.07001881301403046, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001881301403046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09501120448112488, + 0.06572078168392181, + 0.0361967459321022, + 0.06881385296583176, + 0.15136732161045074, + 0.05585557967424393, + 0.1329338550567627, + 0.20118840038776398, + 0.16723692417144775, + 0.02567538060247898 + ], + "credit_map": [ + 0.01126146037131548, + 0.014737041667103767, + 0.0081420186907053, + 0.028308909386396408, + 0.010554649867117405, + 0.007422470487654209, + 0.007326567079871893, + 0.00725808460265398, + 0.01031350065022707, + 0.009376250207424164, + 0.005756216123700142, + 0.012699308805167675, + 0.007435536477714777, + 0.005028887651860714, + 0.009557892568409443, + 0.005553186871111393, + 0.004476973786950111, + 0.007152949459850788, + 0.003026680089533329, + 0.005219051148742437, + 0.004111611284315586, + 0.004553964361548424, + 0.004159965086728334, + 0.003495551412925124, + 0.004409770015627146, + 0.005324052181094885, + 0.003908148035407066, + 0.0058979555033147335, + 0.006364569067955017, + 0.02568657696247101, + 0.004812212195247412, + 0.01241057738661766, + 0.013582391664385796, + 0.01734416000545025, + 0.006135416682809591, + 0.006671166513115168, + 0.013671200722455978, + 0.05809914693236351, + 0.008943164721131325, + 0.026920665055513382, + 0.005098368972539902, + 0.011036437004804611, + 0.0038205008022487164, + 0.006791681982576847, + 0.006990847643464804, + 0.011685402132570744, + 0.004862688481807709, + 0.005569653119891882, + 0.020811637863516808, + 0.02384587563574314, + 0.01232975721359253, + 0.011397440917789936, + 0.017751893028616905, + 0.01972908154129982, + 0.010532164946198463, + 0.016536002978682518, + 0.016414836049079895, + 0.061859071254730225, + 0.009400615468621254, + 0.011140217073261738, + 0.07412140816450119, + 0.013157159090042114, + 0.005533043295145035, + 0.009562068618834019, + 0.012554592452943325, + 0.022984227165579796, + 0.017519371584057808, + 0.022041041404008865, + 0.013938488438725471, + 0.01667390950024128, + 0.04031756520271301, + 0.02120772935450077, + 0.003072209656238556, + 0.0036304087843745947, + 0.003393983468413353, + 0.0021872716024518013, + 0.003564610844478011, + 0.004389486741274595, + 0.002851439407095313, + 0.0025859710294753313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.07001881301403046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, graceful 360-degree drone orbit begins, circling a lone, ancient olive tree with a steady, reverent motion. The camera begins at a slightly lower angle, emphasizing the tree's grandeur. The initial view focuses on the massive, gnarled trunk, its bark an intricate tapestry of deep cracks and twisted muscle structures in weathered grays and earthy browns. As the camera slides counterclockwise, the warm, golden light of the Mediterranean afternoon captures the intricate textures, revealing patches of moss in the cracks. The orbit gradually widens, revealing the tree's full form. Its dense, sinuous branches reach toward the sky like arthritic fingers, adorned with dense clusters of tiny, silvery-green leaves that shimmer and dance in the gentle, warm breeze. Light filters through the leaves, casting speckled, shifting patterns on the sun-drenched terracotta ground below. In the background, the camera's circular path reveals a landscape of rolling, sun-drenched hills, with the deep blue Aegean Sea shimmering in the hazy distance. The atmosphere is one of profound peace and timeless tranquility, a cinematic portrayal of nature's endurance and quiet majesty.", + "seed": 25763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_1/index_2225_seed_25763_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03947165254503489, + "vlm_logprob": -0.03947165254503489, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03947165254503489, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007463902700692415 + ], + "opposite_logprob_tokens": [ + -6.5074639320373535 + ], + "logprob_reward": -0.007463902700692415, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.007463902700692415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08620449900627136, + 0.07613703608512878, + 0.10166297107934952, + 0.062195416539907455, + 0.11560939252376556, + 0.33425667881965637, + 0.04129301384091377, + 0.04965728893876076, + 0.07276008278131485, + 0.06022356450557709 + ], + "credit_map": [ + 0.008547932840883732, + 0.016304057091474533, + 0.013005449436604977, + 0.00941845215857029, + 0.006920570507645607, + 0.008466019295156002, + 0.009287919849157333, + 0.01425410620868206, + 0.01342033687978983, + 0.010264365933835506, + 0.008078308776021004, + 0.007143156602978706, + 0.009128154255449772, + 0.008867280557751656, + 0.007590982597321272, + 0.011644457466900349, + 0.011520306579768658, + 0.028449086472392082, + 0.006824446842074394, + 0.00636910879984498, + 0.009798718616366386, + 0.020346101373434067, + 0.007822110317647457, + 0.010533102788031101, + 0.006705771200358868, + 0.012796171009540558, + 0.004097753204405308, + 0.005281566176563501, + 0.01098205428570509, + 0.010901596397161484, + 0.0044820732437074184, + 0.00694844126701355, + 0.00856484193354845, + 0.02396024763584137, + 0.014875211752951145, + 0.016811495646834373, + 0.019099073484539986, + 0.010902881622314453, + 0.012721512466669083, + 0.008674140088260174, + 0.016692129895091057, + 0.1257815659046173, + 0.01465057022869587, + 0.09523552656173706, + 0.04048750549554825, + 0.00914683472365141, + 0.023932944983243942, + 0.008329587988555431, + 0.0027045956812798977, + 0.005814450327306986, + 0.005892827175557613, + 0.003112721024081111, + 0.004796018358319998, + 0.006141584366559982, + 0.0037410417571663857, + 0.009089780040085316, + 0.0035874671302735806, + 0.009096760302782059, + 0.0058120437897741795, + 0.010373646393418312, + 0.008167452178895473, + 0.0032470079604536295, + 0.003817677730694413, + 0.005555236712098122, + 0.007366809993982315, + 0.007777523249387741, + 0.009228533133864403, + 0.009526630863547325, + 0.012323613278567791, + 0.007462015375494957, + 0.012608734890818596, + 0.006466233637183905, + 0.005805607885122299, + 0.007467442192137241, + 0.005349814426153898, + 0.013131992891430855, + 0.006469338666647673, + 0.0038176816888153553, + 0.007749961223453283, + 0.010431733913719654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.007463902700692415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031118243932724 + ], + "opposite_logprob_tokens": [ + -2.350311279296875 + ], + "logprob_reward": -0.10031118243932724, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031118243932724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04224090650677681, + 0.016679532825946808, + 0.051364343613386154, + 0.065557561814785, + 0.2459198534488678, + 0.19077248871326447, + 0.28532981872558594, + 0.04008769243955612, + 0.04823636636137962, + 0.013811465352773666 + ], + "credit_map": [ + 0.0032658411655575037, + 0.006438316311687231, + 0.004700792953372002, + 0.005924609024077654, + 0.0039599318988621235, + 0.005053098779171705, + 0.005813081748783588, + 0.0070852357894182205, + 0.0022290621418505907, + 0.0024086590856313705, + 0.0024789466988295317, + 0.0012489776127040386, + 0.0020048674196004868, + 0.002236658241599798, + 0.0016991628799587488, + 0.0023731975816190243, + 0.0070493160746991634, + 0.010590657591819763, + 0.0036098130512982607, + 0.0030971632804721594, + 0.009615018032491207, + 0.006184905301779509, + 0.004189417697489262, + 0.007028053980320692, + 0.00754722161218524, + 0.014905543066561222, + 0.0037455144338309765, + 0.004213856067508459, + 0.009750190190970898, + 0.01277041807770729, + 0.004804422613233328, + 0.007820399478077888, + 0.020480459555983543, + 0.08974218368530273, + 0.011386873200535774, + 0.01566239632666111, + 0.02043008804321289, + 0.05646013841032982, + 0.01825845241546631, + 0.01349927019327879, + 0.008884337730705738, + 0.07107067108154297, + 0.009828224778175354, + 0.052975527942180634, + 0.022845881059765816, + 0.005232825875282288, + 0.014982683584094048, + 0.004952331073582172, + 0.03269096463918686, + 0.061221424490213394, + 0.024534905329346657, + 0.011927003972232342, + 0.02740904502570629, + 0.0809701606631279, + 0.01053451094776392, + 0.03604179620742798, + 0.0033958444837480783, + 0.003177163889631629, + 0.005248576868325472, + 0.007394207175821066, + 0.006672088988125324, + 0.004459385294467211, + 0.0041505214758217335, + 0.0055899047292768955, + 0.005493333097547293, + 0.00442439503967762, + 0.006203816272318363, + 0.004532002378255129, + 0.00957142561674118, + 0.0040011354722082615, + 0.009968835860490799, + 0.004041423089802265, + 0.0019005032954737544, + 0.0020857311319559813, + 0.0015369054162874818, + 0.0024605300277471542, + 0.001430888194590807, + 0.0009828947950154543, + 0.0017130044288933277, + 0.0017010075971484184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.10031118243932724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046491860412061214 + ], + "opposite_logprob_tokens": [ + -5.3796491622924805 + ], + "logprob_reward": -0.0046491860412061214, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046491860412061214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11685638874769211, + 0.09188110381364822, + 0.16886481642723083, + 0.043581534177064896, + 0.1836927980184555, + 0.09162310510873795, + 0.15076397359371185, + 0.051097411662340164, + 0.07224082201719284, + 0.029398113489151 + ], + "credit_map": [ + 0.008686146698892117, + 0.026893001049757004, + 0.010263064876198769, + 0.008374332450330257, + 0.006985331419855356, + 0.014303836971521378, + 0.008975505828857422, + 0.03237517550587654, + 0.014858686365187168, + 0.01278666965663433, + 0.011379729956388474, + 0.009207433089613914, + 0.012496720999479294, + 0.013804188929498196, + 0.006843441631644964, + 0.010504231788218021, + 0.02030332386493683, + 0.04065491259098053, + 0.013283352367579937, + 0.011796076782047749, + 0.02172158844769001, + 0.02503085508942604, + 0.01389309111982584, + 0.022181611508131027, + 0.0039566271007061005, + 0.009296431206166744, + 0.0044438824988901615, + 0.005156298168003559, + 0.0058691962622106075, + 0.005742543842643499, + 0.004071094561368227, + 0.005045461934059858, + 0.014783462509512901, + 0.06040534749627113, + 0.009702958166599274, + 0.01271012332290411, + 0.02223488874733448, + 0.03874168545007706, + 0.014961655251681805, + 0.010152669623494148, + 0.003809091169387102, + 0.03648437559604645, + 0.004376139957457781, + 0.01959422044456005, + 0.0156607236713171, + 0.004541255533695221, + 0.004669663030654192, + 0.002487633377313614, + 0.01318842638283968, + 0.050659626722335815, + 0.01108700130134821, + 0.007156419102102518, + 0.02135889232158661, + 0.0335683710873127, + 0.006755167618393898, + 0.006990065332502127, + 0.005472675431519747, + 0.005141869653016329, + 0.009542619809508324, + 0.008015153929591179, + 0.006623104680329561, + 0.006378115154802799, + 0.004551082383841276, + 0.005372791551053524, + 0.005718917585909367, + 0.005732337944209576, + 0.008540628477931023, + 0.013299618847668171, + 0.010649561882019043, + 0.0033278530463576317, + 0.012959347106516361, + 0.012012557126581669, + 0.0024733978789299726, + 0.005695766303688288, + 0.0037950214464217424, + 0.0042117564007639885, + 0.003154964419081807, + 0.002344959881156683, + 0.00403196457773447, + 0.003690282814204693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.0046491860412061214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947629921138287 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947629921138287, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947629921138287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02275567501783371, + 0.020682208240032196, + 0.018321657553315163, + 0.025691483169794083, + 0.21778737008571625, + 0.4429069459438324, + 0.0871102437376976, + 0.08288335800170898, + 0.04850536584854126, + 0.03335569426417351 + ], + "credit_map": [ + 0.0020119203254580498, + 0.003546783933416009, + 0.0026463435497134924, + 0.0018410843331366777, + 0.003187147667631507, + 0.004207218531519175, + 0.0024559232406318188, + 0.002859252505004406, + 0.0026967269368469715, + 0.0029765276703983545, + 0.002381175523623824, + 0.001719375723041594, + 0.00292926630936563, + 0.0031797662377357483, + 0.0013237256789579988, + 0.003475644625723362, + 0.003375845495611429, + 0.004061378072947264, + 0.0014706965303048491, + 0.0009892256930470467, + 0.0014167980989441276, + 0.003074066946282983, + 0.0017203246243298054, + 0.002213322790339589, + 0.0028983165975660086, + 0.004430314525961876, + 0.002721342956647277, + 0.003618289018049836, + 0.00360680790618062, + 0.0036683010403066874, + 0.0017530267359688878, + 0.0029950838070362806, + 0.018163230270147324, + 0.08116765320301056, + 0.01002571638673544, + 0.013056403025984764, + 0.01815936155617237, + 0.049754589796066284, + 0.016011813655495644, + 0.011448603123426437, + 0.015715111047029495, + 0.18944066762924194, + 0.01449254434555769, + 0.11613304167985916, + 0.06580659747123718, + 0.01707971841096878, + 0.016085512936115265, + 0.008153758011758327, + 0.010256832465529442, + 0.0222712941467762, + 0.0065750787034630775, + 0.003997887019068003, + 0.00981881469488144, + 0.023793013766407967, + 0.0034420692827552557, + 0.0069552515633404255, + 0.00777978403493762, + 0.0077101425267755985, + 0.012676707468926907, + 0.008275504224002361, + 0.020692598074674606, + 0.01029700692743063, + 0.004753184970468283, + 0.01069843489676714, + 0.004872697871178389, + 0.003568448359146714, + 0.0076944162137806416, + 0.005070933606475592, + 0.005181558895856142, + 0.005720129702240229, + 0.009757068008184433, + 0.006640113890171051, + 0.003367447294294834, + 0.004007679410278797, + 0.004087254870682955, + 0.006524852477014065, + 0.0029288027435541153, + 0.00333367264829576, + 0.0026286670472472906, + 0.006477316841483116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.005947629921138287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, ancient olive tree with a massive, gnarled trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tree's branches dense and sinuous with silvery-green leaves that shimmer in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a landscape of rolling hills with the deep blue Aegean Sea in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slightly lower angle and slide counterclockwise during the orbit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898636162281036 + ], + "opposite_logprob_tokens": [ + -2.578986406326294 + ], + "logprob_reward": -0.07898636162281036, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898636162281036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09959450364112854, + 0.06489109992980957, + 0.06965538114309311, + 0.030489742755889893, + 0.2260187715291977, + 0.2034825086593628, + 0.05931739881634712, + 0.07688944786787033, + 0.12709149718284607, + 0.04256967827677727 + ], + "credit_map": [ + 0.007989113219082355, + 0.014652480371296406, + 0.013793762773275375, + 0.010405749082565308, + 0.009642503224313259, + 0.013216651044785976, + 0.014499884098768234, + 0.015394353307783604, + 0.008247380144894123, + 0.012349314987659454, + 0.010435286909341812, + 0.005185603629797697, + 0.007739305030554533, + 0.00809208583086729, + 0.00364559143781662, + 0.009196530096232891, + 0.011499048210680485, + 0.013743182644248009, + 0.004959103185683489, + 0.005324262659996748, + 0.008071081712841988, + 0.011575225740671158, + 0.004611175507307053, + 0.00987230148166418, + 0.0025150019209831953, + 0.003960202448070049, + 0.004535239189863205, + 0.006003774236887693, + 0.003110091667622328, + 0.003495526500046253, + 0.0036845162976533175, + 0.003185391193255782, + 0.019124479964375496, + 0.0825471505522728, + 0.011087631806731224, + 0.013732164166867733, + 0.01983788050711155, + 0.05060025304555893, + 0.01676550693809986, + 0.01232371386140585, + 0.009380057454109192, + 0.07499362528324127, + 0.010963742621243, + 0.06053273379802704, + 0.021007537841796875, + 0.006936957594007254, + 0.01431087777018547, + 0.005356971640139818, + 0.007367844693362713, + 0.011946290731430054, + 0.005233310163021088, + 0.006332799792289734, + 0.0065192244946956635, + 0.010845528915524483, + 0.0038667209446430206, + 0.007205677684396505, + 0.004789009690284729, + 0.010157090611755848, + 0.009254864417016506, + 0.014573379419744015, + 0.020613398402929306, + 0.0056825862266123295, + 0.005599327385425568, + 0.006219788920134306, + 0.012131165713071823, + 0.012647115625441074, + 0.020806042477488518, + 0.014142417348921299, + 0.02118537761271, + 0.013265904039144516, + 0.022140435874462128, + 0.010773031041026115, + 0.0041931853629648685, + 0.007942154072225094, + 0.003279748372733593, + 0.007162987720221281, + 0.004449992440640926, + 0.003270459361374378, + 0.006941658444702625, + 0.005329493433237076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.07898636162281036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_tensor([4238, 4238])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_tensor([4238, 4238])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bf9b6b51bad3bba9ce2a91749f91e7617f6edc86 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_tensor([4238, 4238])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide shot, the camera slowly pans right, tracking a lone figure walking down a broad, rain-soaked street in front of the magnificent Arc de Triomphe. The figure is a sharp, resolute man, dressed in a long, dark trench coat, its damp surface reflecting chaotic light like polished obsidian. He walks with a confident, unhurried demeanor, his hands tucked deep into his pockets, his heavy boots deliberately splashing in shallow puddles. In the background, the monumental Arc de Triomphe looms, its classical form reimagined in a gritty, Grand Theft Auto art style; its stone surfaces are dark, weathered, and streaked with rain, while its intricate carvings are rendered into sharp, dramatic relief by harsh artificial light. The entire scene is saturated in the glow of a rain-soaked sky, with vibrant pinks, electric blues, and vivid greens casting long, distorted reflections across the smooth, inky black asphalt. A steady downpour catches the light, creating a shimmering, dynamic veil across the frame, creating a cool, moody and intense cinematic atmosphere that blends historical grandeur with a menacing, digitally rendered, and distinctly modern urban noir aesthetic.", + "seed": 23423, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_4238_seed_23423_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10158704929053783, + "vlm_logprob": -0.10158704929053783, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.10158704929053783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2106849104166031 + ], + "opposite_logprob_tokens": [ + -1.960684895515442 + ], + "logprob_reward": -0.2106849104166031, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.2106849104166031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07021263241767883, + 0.05326128751039505, + 0.05150872468948364, + 0.13205787539482117, + 0.15279236435890198, + 0.20645196735858917, + 0.1009555384516716, + 0.01979014277458191, + 0.1633346974849701, + 0.04963470995426178 + ], + "credit_map": [ + 0.006659114267677069, + 0.006610692013055086, + 0.011867916211485863, + 0.009725346229970455, + 0.01089855469763279, + 0.005659160204231739, + 0.008771266788244247, + 0.010020587593317032, + 0.006515196990221739, + 0.011408464051783085, + 0.009164705872535706, + 0.004834112245589495, + 0.005081263370811939, + 0.0036488634068518877, + 0.005687801167368889, + 0.006920885760337114, + 0.0059999339282512665, + 0.011524256318807602, + 0.007023615762591362, + 0.004319797735661268, + 0.005156803410500288, + 0.0033814560156315565, + 0.009785798378288746, + 0.004317068494856358, + 0.008276627399027348, + 0.022788459435105324, + 0.012498590163886547, + 0.027004562318325043, + 0.024655422195792198, + 0.00905737653374672, + 0.01757974363863468, + 0.01019709650427103, + 0.012209096923470497, + 0.023173369467258453, + 0.012580274604260921, + 0.01568399742245674, + 0.03547424077987671, + 0.012606697157025337, + 0.028941139578819275, + 0.012123562395572662, + 0.015359739772975445, + 0.025561003014445305, + 0.033421725034713745, + 0.031651176512241364, + 0.024508962407708168, + 0.012811461463570595, + 0.028736941516399384, + 0.034400977194309235, + 0.010174060240387917, + 0.0165831558406353, + 0.015811463817954063, + 0.019300775602459908, + 0.01022492628544569, + 0.0069650327786803246, + 0.012809548527002335, + 0.009086588397622108, + 0.002047070534899831, + 0.0028699550312012434, + 0.0026684391777962446, + 0.0034904645290225744, + 0.0020814682357013226, + 0.0013875989243388176, + 0.0023857851047068834, + 0.002859362168237567, + 0.01499887090176344, + 0.043112993240356445, + 0.012591727077960968, + 0.02544764056801796, + 0.01841479353606701, + 0.014455216005444527, + 0.022144800052046776, + 0.012168671004474163, + 0.005861308891326189, + 0.005663547199219465, + 0.005542196799069643, + 0.003542686812579632, + 0.0053178356029093266, + 0.011411051265895367, + 0.006016352213919163, + 0.006279736757278442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.2106849104166031 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813582420349121 + ], + "opposite_logprob_tokens": [ + -1.406358242034912 + ], + "logprob_reward": -0.2813582420349121, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813582420349121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061006102710962296, + 0.04671256244182587, + 0.062090713530778885, + 0.4096628427505493, + 0.1343102753162384, + 0.08197160810232162, + 0.04602469131350517, + 0.02761641889810562, + 0.1062377467751503, + 0.02436702325940132 + ], + "credit_map": [ + 0.005536292213946581, + 0.014127370901405811, + 0.008718767203390598, + 0.008876422420144081, + 0.007274297997355461, + 0.004608075600117445, + 0.007342671975493431, + 0.004522203467786312, + 0.005205737426877022, + 0.0050774808041751385, + 0.00771383848041296, + 0.005476726684719324, + 0.00642028171569109, + 0.003944753669202328, + 0.0063695101998746395, + 0.006504232995212078, + 0.007143906783312559, + 0.018055735155940056, + 0.007903666235506535, + 0.005065327510237694, + 0.0046477182768285275, + 0.004233384504914284, + 0.010510743595659733, + 0.0045302254147827625, + 0.02678670361638069, + 0.07746922224760056, + 0.045355867594480515, + 0.10953886061906815, + 0.04839281365275383, + 0.023594606667757034, + 0.053037919104099274, + 0.025486858561635017, + 0.010264984332025051, + 0.023098250851035118, + 0.011397677473723888, + 0.01823025569319725, + 0.03028513304889202, + 0.010664324276149273, + 0.022939937189221382, + 0.007429704535752535, + 0.0058104488998651505, + 0.010399769060313702, + 0.01339119765907526, + 0.012987568974494934, + 0.00998873170465231, + 0.004719451069831848, + 0.011154048144817352, + 0.013520396314561367, + 0.0035653056111186743, + 0.004100948106497526, + 0.00935580674558878, + 0.008474425412714481, + 0.0042511336505413055, + 0.004145721439272165, + 0.007344171870499849, + 0.00478717777878046, + 0.0021247072145342827, + 0.0038393796421587467, + 0.004419708624482155, + 0.0044165924191474915, + 0.0028750442434102297, + 0.00282864598557353, + 0.0036015829537063837, + 0.0035107582807540894, + 0.010211444459855556, + 0.0309181809425354, + 0.007791886571794748, + 0.01627558097243309, + 0.012022683396935463, + 0.007869018241763115, + 0.014230300672352314, + 0.006918651517480612, + 0.0030423274729400873, + 0.0025710745248943567, + 0.0020133901853114367, + 0.001745613873936236, + 0.00208190712146461, + 0.007643038406968117, + 0.002712489105761051, + 0.002557183150202036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.2813582420349121 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644795786589384 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004644795786589384, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644795786589384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1302497386932373, + 0.09574738889932632, + 0.1137685775756836, + 0.24052982032299042, + 0.1183587908744812, + 0.04587794095277786, + 0.11086660623550415, + 0.03211046755313873, + 0.08849547803401947, + 0.02399521693587303 + ], + "credit_map": [ + 0.014015207067131996, + 0.022923005744814873, + 0.016003208234906197, + 0.014708750881254673, + 0.017015652731060982, + 0.011673527769744396, + 0.016449864953756332, + 0.01746050827205181, + 0.01263165008276701, + 0.013409467414021492, + 0.012319356203079224, + 0.009415638633072376, + 0.017310718074440956, + 0.008528126403689384, + 0.015874246135354042, + 0.006258183158934116, + 0.011155742220580578, + 0.018662359565496445, + 0.01536009181290865, + 0.009026847779750824, + 0.02288021519780159, + 0.00960538350045681, + 0.01714351773262024, + 0.009934416972100735, + 0.013440962880849838, + 0.03749546408653259, + 0.028150996193289757, + 0.0634259507060051, + 0.032971274107694626, + 0.016783669590950012, + 0.029507625848054886, + 0.018753863871097565, + 0.01074764784425497, + 0.025994587689638138, + 0.009013043716549873, + 0.01706656627357006, + 0.022907190024852753, + 0.00863095372915268, + 0.017015669494867325, + 0.00698311859741807, + 0.003537710290402174, + 0.006025637499988079, + 0.00903397984802723, + 0.005136480089277029, + 0.004922458436340094, + 0.0051920171827077866, + 0.005504895467311144, + 0.0065247612074017525, + 0.009993668645620346, + 0.017909018322825432, + 0.017236093059182167, + 0.018452422693371773, + 0.01223810762166977, + 0.008635821752250195, + 0.01435257401317358, + 0.012048892676830292, + 0.0028973999433219433, + 0.002840151311829686, + 0.004540747497230768, + 0.005531100556254387, + 0.0037214644253253937, + 0.002697966992855072, + 0.005059834569692612, + 0.004821799695491791, + 0.008223423734307289, + 0.022485293447971344, + 0.007252659648656845, + 0.014288383536040783, + 0.010000329464673996, + 0.009228252805769444, + 0.010276168584823608, + 0.0067409551702439785, + 0.0018040869617834687, + 0.00254828785546124, + 0.0026260686572641134, + 0.005264580249786377, + 0.0029685196932405233, + 0.004213080275803804, + 0.00210630614310503, + 0.002464285120368004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004644795786589384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07812173664569855, + 0.08232630044221878, + 0.08149445056915283, + 0.4402454197406769, + 0.10515520721673965, + 0.03926081210374832, + 0.02887602336704731, + 0.05488912761211395, + 0.05227849632501602, + 0.03735243156552315 + ], + "credit_map": [ + 0.012379446998238564, + 0.008847005665302277, + 0.012973329052329063, + 0.01201450265944004, + 0.010433451272547245, + 0.006531653460115194, + 0.008521035313606262, + 0.006421311292797327, + 0.011180237866938114, + 0.020504403859376907, + 0.010542175732553005, + 0.010289164260029793, + 0.008032404817640781, + 0.0056506190448999405, + 0.008359002880752087, + 0.0077682966366410255, + 0.0072332522831857204, + 0.024730687960982323, + 0.00743597187101841, + 0.007298798765987158, + 0.008508812636137009, + 0.0048577068373560905, + 0.01586904749274254, + 0.0055601708590984344, + 0.02895570732653141, + 0.11056631058454514, + 0.0449543297290802, + 0.09248694032430649, + 0.0508384108543396, + 0.02608788199722767, + 0.05318921431899071, + 0.03316662833094597, + 0.010007425211369991, + 0.02121870033442974, + 0.011262430809438229, + 0.020145801827311516, + 0.013097810558974743, + 0.008009400218725204, + 0.012987450696527958, + 0.008426188491284847, + 0.0037752084899693727, + 0.0053453086875379086, + 0.007086018566042185, + 0.006670255213975906, + 0.004901960492134094, + 0.0027325195260345936, + 0.004199189133942127, + 0.004550353158265352, + 0.0028473997954279184, + 0.0030963981989771128, + 0.004993328358978033, + 0.004019397310912609, + 0.003292185254395008, + 0.0032635238021612167, + 0.002914925804361701, + 0.004448865540325642, + 0.004511474631726742, + 0.0075418101623654366, + 0.008278308436274529, + 0.008465058170258999, + 0.0058411117643117905, + 0.0054687573574483395, + 0.007487765047699213, + 0.007294843904674053, + 0.004373573232442141, + 0.00709215784445405, + 0.007731728721410036, + 0.00862205307930708, + 0.0058176578022539616, + 0.0058768619783222675, + 0.008062436245381832, + 0.004702028352767229, + 0.0035949358716607094, + 0.005609463434666395, + 0.004166523460298777, + 0.004053247161209583, + 0.00385713460855186, + 0.006486310623586178, + 0.004784945398569107, + 0.004799870308488607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615036141127348 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003615036141127348, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615036141127348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10511337965726852, + 0.08096741884946823, + 0.2932007908821106, + 0.09165089577436447, + 0.0571199506521225, + 0.05026812106370926, + 0.038448598235845566, + 0.18470947444438934, + 0.07037036120891571, + 0.02815101109445095 + ], + "credit_map": [ + 0.012799778021872044, + 0.014669408090412617, + 0.015593265183269978, + 0.016833893954753876, + 0.01563384011387825, + 0.006333189085125923, + 0.01194840483367443, + 0.011301600374281406, + 0.012966581620275974, + 0.0168917253613472, + 0.010762190446257591, + 0.008166592568159103, + 0.008269946090877056, + 0.0053300862200558186, + 0.00965744536370039, + 0.008922855369746685, + 0.031012363731861115, + 0.0826260894536972, + 0.03748687729239464, + 0.023831697180867195, + 0.0306838545948267, + 0.018969368189573288, + 0.045133884996175766, + 0.023456647992134094, + 0.007576714735478163, + 0.019286485388875008, + 0.008182425051927567, + 0.01978159323334694, + 0.010740268044173717, + 0.005074753426015377, + 0.012682035565376282, + 0.00832662358880043, + 0.0047219861298799515, + 0.01035634707659483, + 0.009355160407721996, + 0.009631933644413948, + 0.007764178328216076, + 0.0035725594498217106, + 0.007304983213543892, + 0.00441280473023653, + 0.0042565627954900265, + 0.005093814805150032, + 0.010257118381559849, + 0.007473401725292206, + 0.004270526114851236, + 0.0033773852046579123, + 0.006440372671931982, + 0.00909893587231636, + 0.00315587199293077, + 0.003955473192036152, + 0.006927936803549528, + 0.005523505620658398, + 0.004466764163225889, + 0.00419009430333972, + 0.005015397444367409, + 0.005213551223278046, + 0.017102312296628952, + 0.027817388996481895, + 0.024775398895144463, + 0.024546146392822266, + 0.02319713868200779, + 0.017931245267391205, + 0.02369798719882965, + 0.02564184181392193, + 0.007731974124908447, + 0.020231155678629875, + 0.005633697845041752, + 0.010694831609725952, + 0.0074432832188904285, + 0.005567236803472042, + 0.00795034971088171, + 0.005117830354720354, + 0.00306418864056468, + 0.0032165341544896364, + 0.002514933468773961, + 0.003984360955655575, + 0.0025646542198956013, + 0.006776282098144293, + 0.00322631630115211, + 0.0028037389274686575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.003615036141127348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot, the camera slowly pans right, tracking a lone figure walking down a broad, rain-soaked street in front of the magnificent Arc de Triomphe. The figure is a sharp, resolute man, dressed in a long, dark trench coat, its damp surface reflecting chaotic light like polished obsidian. He walks with a confident, unhurried demeanor, his hands tucked deep into his pockets, his heavy boots deliberately splashing in shallow puddles. In the background, the monumental Arc de Triomphe looms, its classical form reimagined in a gritty, Grand Theft Auto art style; its stone surfaces are dark, weathered, and streaked with rain, while its intricate carvings are rendered into sharp, dramatic relief by harsh artificial light. The entire scene is saturated in the glow of a rain-soaked sky, with vibrant pinks, electric blues, and vivid greens casting long, distorted reflections across the smooth, inky black asphalt. A steady downpour catches the light, creating a shimmering, dynamic veil across the frame, creating a cool, moody and intense cinematic atmosphere that blends historical grandeur with a menacing, digitally rendered, and distinctly modern urban noir aesthetic.", + "seed": 23424, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_4238_seed_23424_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1042376862373203, + "vlm_logprob": -0.1042376862373203, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.1042376862373203, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14874301850795746 + ], + "opposite_logprob_tokens": [ + -2.523742914199829 + ], + "logprob_reward": -0.14874301850795746, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.14874301850795746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20763437449932098, + 0.09064459055662155, + 0.10784953087568283, + 0.12796394526958466, + 0.0731852576136589, + 0.07953928411006927, + 0.07899057120084763, + 0.07127923518419266, + 0.09550795704126358, + 0.06740524619817734 + ], + "credit_map": [ + 0.01991981267929077, + 0.0252759400755167, + 0.02480531483888626, + 0.02750544808804989, + 0.02097218669950962, + 0.030056536197662354, + 0.025956859812140465, + 0.033142272382974625, + 0.009247473441064358, + 0.009918376803398132, + 0.01017890777438879, + 0.008715125732123852, + 0.01604805886745453, + 0.013228042982518673, + 0.011861038394272327, + 0.011447564698755741, + 0.010248202830553055, + 0.016369063407182693, + 0.011187546886503696, + 0.014239112846553326, + 0.019906843081116676, + 0.012560218572616577, + 0.012791621498763561, + 0.010546927340328693, + 0.01004259567707777, + 0.014156470075249672, + 0.013134176842868328, + 0.009141368791460991, + 0.025810586288571358, + 0.020267212763428688, + 0.01624751091003418, + 0.019164036959409714, + 0.005959549453109503, + 0.013110856525599957, + 0.007464124355465174, + 0.008384590037167072, + 0.009163271635770798, + 0.007679589558392763, + 0.009198062121868134, + 0.012225212529301643, + 0.006811079103499651, + 0.011007827706634998, + 0.010159127414226532, + 0.009623276069760323, + 0.010859396308660507, + 0.007805315311998129, + 0.011763084679841995, + 0.011510179378092289, + 0.007101124152541161, + 0.019391130656003952, + 0.006418291479349136, + 0.00875605084002018, + 0.01177358627319336, + 0.008509792387485504, + 0.00836276076734066, + 0.008677834644913673, + 0.007358210626989603, + 0.01638103649020195, + 0.005665896460413933, + 0.005527379456907511, + 0.011450613848865032, + 0.0073624830693006516, + 0.007574537303298712, + 0.009959072805941105, + 0.00920398160815239, + 0.015462055802345276, + 0.006391454953700304, + 0.010441934689879417, + 0.011865774169564247, + 0.01438073068857193, + 0.009802396409213543, + 0.017959628254175186, + 0.010046306997537613, + 0.007178111933171749, + 0.007756134960800409, + 0.011265615932643414, + 0.00648683262988925, + 0.007857641205191612, + 0.006683661136776209, + 0.010130949318408966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.14874301850795746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34867557883262634 + ], + "opposite_logprob_tokens": [ + -1.2236756086349487 + ], + "logprob_reward": -0.34867557883262634, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34867557883262634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2197878062725067, + 0.10854209214448929, + 0.07194241136312485, + 0.21282334625720978, + 0.015229876153171062, + 0.07746799290180206, + 0.0315103679895401, + 0.052050698548555374, + 0.08818097412586212, + 0.12246446311473846 + ], + "credit_map": [ + 0.025396760553121567, + 0.024515600875020027, + 0.03318222612142563, + 0.030229320749640465, + 0.021904857829213142, + 0.02704346552491188, + 0.025549864396452904, + 0.0319657139480114, + 0.016714604571461678, + 0.012043142691254616, + 0.009565531276166439, + 0.007506972178816795, + 0.012702780775725842, + 0.025304891169071198, + 0.009644090197980404, + 0.01506007369607687, + 0.006475904490798712, + 0.012326788157224655, + 0.009272675961256027, + 0.006903381086885929, + 0.01306846272200346, + 0.007486938498914242, + 0.007020830642431974, + 0.009387428872287273, + 0.015561181120574474, + 0.026659533381462097, + 0.018640469759702682, + 0.016139790415763855, + 0.05315132066607475, + 0.0299780685454607, + 0.03191465884447098, + 0.02077832631766796, + 0.0014223093166947365, + 0.0022884386125952005, + 0.0018654558807611465, + 0.0017604718450456858, + 0.0015883290907368064, + 0.0021069480571895838, + 0.0016742354491725564, + 0.0025236874353140593, + 0.006280955392867327, + 0.011459115892648697, + 0.009039017371833324, + 0.0129233468323946, + 0.010903134010732174, + 0.00711751077324152, + 0.011642069555819035, + 0.008102846331894398, + 0.0026181908324360847, + 0.00788190122693777, + 0.0027581152971833944, + 0.003708523465320468, + 0.0047705271281301975, + 0.003527261782437563, + 0.0029018137138336897, + 0.003344036405906081, + 0.004941439721733332, + 0.012960588559508324, + 0.0033940961584448814, + 0.008950556628406048, + 0.006734855938702822, + 0.0043337102979421616, + 0.0048818886280059814, + 0.005853564478456974, + 0.008616922423243523, + 0.014941908419132233, + 0.004857262130826712, + 0.009673863649368286, + 0.012867830693721771, + 0.012767709791660309, + 0.007896285504102707, + 0.016559185460209846, + 0.011697684414684772, + 0.014130108058452606, + 0.012458337470889091, + 0.029157554730772972, + 0.013929823413491249, + 0.011480848304927349, + 0.013865338638424873, + 0.0157447699457407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.34867557883262634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958176217973232 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958176217973232, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958176217973232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061486247926950455, + 0.1463349163532257, + 0.3780025541782379, + 0.036649368703365326, + 0.04974012076854706, + 0.05065697804093361, + 0.046082790940999985, + 0.03710479661822319, + 0.15592975914478302, + 0.038012467324733734 + ], + "credit_map": [ + 0.006776559166610241, + 0.00824800319969654, + 0.009861616417765617, + 0.006020338274538517, + 0.006399715784937143, + 0.0076184519566595554, + 0.005500770639628172, + 0.011060791090130806, + 0.026490526273846626, + 0.015114448964595795, + 0.011763456277549267, + 0.009048842824995518, + 0.012432044371962547, + 0.041586942970752716, + 0.009504019282758236, + 0.020394621416926384, + 0.03187460079789162, + 0.07000309973955154, + 0.03464971482753754, + 0.05013957992196083, + 0.06627503037452698, + 0.04775460809469223, + 0.04288960620760918, + 0.03441629186272621, + 0.0036435367073863745, + 0.005234088748693466, + 0.003982244059443474, + 0.0027712257578969, + 0.007785900495946407, + 0.004859373904764652, + 0.0042152851819992065, + 0.004157715477049351, + 0.004419416654855013, + 0.011686339974403381, + 0.0061786361038684845, + 0.0038868915289640427, + 0.0057527655735611916, + 0.005017323885113001, + 0.0038705195765942335, + 0.008928229101002216, + 0.003523764666169882, + 0.010360091924667358, + 0.004986194893717766, + 0.006209069862961769, + 0.007343617267906666, + 0.006563595030456781, + 0.005369230639189482, + 0.006301415152847767, + 0.0052827345207333565, + 0.011031496338546276, + 0.004381590988487005, + 0.005221008323132992, + 0.007128940895199776, + 0.0043433778919279575, + 0.0037978000473231077, + 0.004895840305835009, + 0.003376597538590431, + 0.009527381509542465, + 0.0028708097524940968, + 0.004763658158481121, + 0.004829241894185543, + 0.0032959093805402517, + 0.003741496242582798, + 0.004699700977653265, + 0.01744876056909561, + 0.02849716879427433, + 0.01072509866207838, + 0.01779188960790634, + 0.01961454190313816, + 0.020081207156181335, + 0.01503656804561615, + 0.026734527200460434, + 0.0033145607449114323, + 0.003499062266200781, + 0.005355433095246553, + 0.007695565931499004, + 0.004353807773441076, + 0.0042614745907485485, + 0.003941754810512066, + 0.005590806715190411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.005958176217973232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617767943069339 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617767943069339, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617767943069339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2560451924800873, + 0.10260201245546341, + 0.20068486034870148, + 0.09154962003231049, + 0.05859166756272316, + 0.11069547384977341, + 0.03481946140527725, + 0.026820635423064232, + 0.036707278341054916, + 0.08148381859064102 + ], + "credit_map": [ + 0.027700018137693405, + 0.03597376123070717, + 0.03766673803329468, + 0.03375079855322838, + 0.030740471556782722, + 0.03130405396223068, + 0.030073178932070732, + 0.028836166486144066, + 0.013328791595995426, + 0.01548803597688675, + 0.00932246819138527, + 0.00861494429409504, + 0.013302008621394634, + 0.019890056923031807, + 0.009289230220019817, + 0.013366460800170898, + 0.016375897452235222, + 0.039689503610134125, + 0.01915077492594719, + 0.026011500507593155, + 0.03424638882279396, + 0.024542689323425293, + 0.022070003673434258, + 0.01859808713197708, + 0.007012742105871439, + 0.011945953592658043, + 0.006573189981281757, + 0.007597445044666529, + 0.022833066061139107, + 0.013287506997585297, + 0.012783536687493324, + 0.009516162797808647, + 0.004688228480517864, + 0.007602168712764978, + 0.0034231843892484903, + 0.00447649834677577, + 0.010650601238012314, + 0.01038419920951128, + 0.006741080433130264, + 0.010625697672367096, + 0.009504193440079689, + 0.01866062916815281, + 0.016101261600852013, + 0.011410950683057308, + 0.013114971108734608, + 0.010287259705364704, + 0.013863223604857922, + 0.017752980813384056, + 0.0053574503399431705, + 0.006133458577096462, + 0.00417077774181962, + 0.0038633737713098526, + 0.004314040765166283, + 0.004513344261795282, + 0.002943581435829401, + 0.0035234319511801004, + 0.0040383958257734776, + 0.005575783550739288, + 0.0021500003058463335, + 0.0028138950001448393, + 0.0031656287610530853, + 0.0027854267973452806, + 0.0026657269336283207, + 0.00362577592022717, + 0.005471026059240103, + 0.005114453844726086, + 0.002823096700012684, + 0.003737503895536065, + 0.00465657701715827, + 0.0057985312305390835, + 0.0036312323063611984, + 0.005474853329360485, + 0.0058768135495483875, + 0.009180709719657898, + 0.007796335965394974, + 0.01940329372882843, + 0.008585001341998577, + 0.010136953555047512, + 0.009434772655367851, + 0.011069935746490955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.003617767943069339 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20048804581165314, + 0.1436053067445755, + 0.13066339492797852, + 0.12319567799568176, + 0.15099580585956573, + 0.07178516685962677, + 0.032014645636081696, + 0.023120947182178497, + 0.0991360992193222, + 0.02499493956565857 + ], + "credit_map": [ + 0.031380608677864075, + 0.015246910974383354, + 0.0482880100607872, + 0.017113802954554558, + 0.012665140442550182, + 0.024427136406302452, + 0.018684420734643936, + 0.03268202021718025, + 0.02889843098819256, + 0.011202001012861729, + 0.00931309163570404, + 0.0060619269497692585, + 0.012586865574121475, + 0.04571754112839699, + 0.009366347454488277, + 0.020459111779928207, + 0.010903723537921906, + 0.030490951612591743, + 0.009864420630037785, + 0.017001042142510414, + 0.02100285328924656, + 0.014030268415808678, + 0.014296218752861023, + 0.013073907233774662, + 0.012242168188095093, + 0.01615062914788723, + 0.009175475686788559, + 0.009474912658333778, + 0.028462260961532593, + 0.012943252921104431, + 0.01638151705265045, + 0.01836545206606388, + 0.012918959371745586, + 0.014580843970179558, + 0.010865012183785439, + 0.00847617443650961, + 0.021792879328131676, + 0.027522128075361252, + 0.01803651824593544, + 0.03680330142378807, + 0.008697476238012314, + 0.009429468773305416, + 0.008485476486384869, + 0.010457893833518028, + 0.009356075897812843, + 0.0064629665575921535, + 0.009607324376702309, + 0.0092884860932827, + 0.002996000461280346, + 0.006716455332934856, + 0.002822397742420435, + 0.003983048722147942, + 0.004070529714226723, + 0.003168549155816436, + 0.002741494681686163, + 0.005516169127076864, + 0.0021608939860016108, + 0.0057409158907830715, + 0.0009966730140149593, + 0.0018354776548221707, + 0.003509135916829109, + 0.002354423049837351, + 0.002124394988641143, + 0.004399032797664404, + 0.010502058081328869, + 0.01874658465385437, + 0.00659371679648757, + 0.011471319943666458, + 0.013271245174109936, + 0.012033693492412567, + 0.01008100900799036, + 0.016436470672488213, + 0.003377964487299323, + 0.003368320409208536, + 0.0026935278438031673, + 0.003984317649155855, + 0.0030467151664197445, + 0.0028756158426404, + 0.0024441417772322893, + 0.003204337554052472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014193889684975147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot, the camera slowly pans right, tracking a lone figure walking down a broad, rain-soaked street in front of the magnificent Arc de Triomphe. The figure is a sharp, resolute man, dressed in a long, dark trench coat, its damp surface reflecting chaotic light like polished obsidian. He walks with a confident, unhurried demeanor, his hands tucked deep into his pockets, his heavy boots deliberately splashing in shallow puddles. In the background, the monumental Arc de Triomphe looms, its classical form reimagined in a gritty, Grand Theft Auto art style; its stone surfaces are dark, weathered, and streaked with rain, while its intricate carvings are rendered into sharp, dramatic relief by harsh artificial light. The entire scene is saturated in the glow of a rain-soaked sky, with vibrant pinks, electric blues, and vivid greens casting long, distorted reflections across the smooth, inky black asphalt. A steady downpour catches the light, creating a shimmering, dynamic veil across the frame, creating a cool, moody and intense cinematic atmosphere that blends historical grandeur with a menacing, digitally rendered, and distinctly modern urban noir aesthetic.", + "seed": 23425, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_4238_seed_23425_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1807196382433176, + "vlm_logprob": -0.1807196382433176, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1807196382433176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2951511740684509 + ], + "opposite_logprob_tokens": [ + -1.5451512336730957 + ], + "logprob_reward": -0.2951511740684509, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2951511740684509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047830089926719666, + 0.06563513725996017, + 0.019986825063824654, + 0.017113879323005676, + 0.00925593450665474, + 0.05042476952075958, + 0.03255748376250267, + 0.03125661239027977, + 0.7174088358879089, + 0.008530490100383759 + ], + "credit_map": [ + 0.00575396278873086, + 0.0039189704693853855, + 0.0078429551795125, + 0.0038831252604722977, + 0.006944361142814159, + 0.004418456461280584, + 0.006507699843496084, + 0.008560555055737495, + 0.0062643662095069885, + 0.0047514596953988075, + 0.01841508224606514, + 0.006392274051904678, + 0.006492838729172945, + 0.010210787877440453, + 0.006623327732086182, + 0.006484991405159235, + 0.0017317961901426315, + 0.002871001372113824, + 0.003100875997915864, + 0.0035319116432219744, + 0.0025039042811840773, + 0.0018184970831498504, + 0.001968286233022809, + 0.002460550284013152, + 0.0018287166021764278, + 0.003479044185951352, + 0.0017981113633140922, + 0.0021044062450528145, + 0.0015303597319871187, + 0.0027423081919550896, + 0.0015730127925053239, + 0.0020579188130795956, + 0.0011013313196599483, + 0.0015334871131926775, + 0.0016434649005532265, + 0.0017800447531044483, + 0.0008989633060991764, + 0.000800090201664716, + 0.0005844391998834908, + 0.0009141129557974637, + 0.006701498292386532, + 0.004198278766125441, + 0.011725111864507198, + 0.004753692075610161, + 0.004346920643001795, + 0.004706026986241341, + 0.007927794009447098, + 0.006065443158149719, + 0.0030801713000983, + 0.003988567274063826, + 0.0046766288578510284, + 0.005301586352288723, + 0.004207829479128122, + 0.0035385421942919493, + 0.0035560012329369783, + 0.004208154045045376, + 0.00249268370680511, + 0.00356477708555758, + 0.0037154974415898323, + 0.005119979847222567, + 0.003072727005928755, + 0.004184733144938946, + 0.003377295332029462, + 0.005728916265070438, + 0.07142145931720734, + 0.05411815643310547, + 0.09597884863615036, + 0.0787307471036911, + 0.12134841084480286, + 0.05203421786427498, + 0.16261446475982666, + 0.08116248995065689, + 0.0009429458878003061, + 0.0005240715690888464, + 0.0017327265813946724, + 0.0010365069610998034, + 0.0011041739489883184, + 0.0013121945085003972, + 0.0010122342500835657, + 0.0008656355203129351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.2951511740684509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761765837669373 + ], + "opposite_logprob_tokens": [ + -0.8261765837669373 + ], + "logprob_reward": -0.5761765837669373, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761765837669373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15897074341773987, + 0.06759994477033615, + 0.05446444824337959, + 0.042409952729940414, + 0.07661616802215576, + 0.16939279437065125, + 0.2252132147550583, + 0.021314240992069244, + 0.15706326067447662, + 0.026955243200063705 + ], + "credit_map": [ + 0.016148453578352928, + 0.008704260922968388, + 0.05372843146324158, + 0.011725724674761295, + 0.013882583938539028, + 0.012742703780531883, + 0.02992270328104496, + 0.01211587805300951, + 0.006519259884953499, + 0.006892162840813398, + 0.018353275954723358, + 0.007746096700429916, + 0.007727249525487423, + 0.008004669100046158, + 0.007324820850044489, + 0.005032408982515335, + 0.006371740717440844, + 0.006749571301043034, + 0.010875885374844074, + 0.006509386468678713, + 0.006683175452053547, + 0.006254001520574093, + 0.005908793769776821, + 0.005111895501613617, + 0.004668677691370249, + 0.006705424748361111, + 0.006242720875889063, + 0.0045615304261446, + 0.003684613388031721, + 0.006717977579683065, + 0.003491309704259038, + 0.006337698549032211, + 0.009146638214588165, + 0.011520801112055779, + 0.004723485093563795, + 0.01575634256005287, + 0.008048389106988907, + 0.006579714361578226, + 0.009640811942517757, + 0.011199982836842537, + 0.01714586839079857, + 0.021769316866993904, + 0.0341360978782177, + 0.01730477809906006, + 0.0136587955057621, + 0.0184393972158432, + 0.023649893701076508, + 0.023288656026124954, + 0.02028699405491352, + 0.02302028238773346, + 0.021085919812321663, + 0.02778209000825882, + 0.03471656143665314, + 0.01830684021115303, + 0.052629873156547546, + 0.027384651824831963, + 0.0021387841552495956, + 0.002842791611328721, + 0.0025857749860733747, + 0.0035169022157788277, + 0.001954661449417472, + 0.002354940865188837, + 0.0029870334547013044, + 0.002933352952823043, + 0.016028927639126778, + 0.010904614813625813, + 0.021149106323719025, + 0.01873553916811943, + 0.025957180187106133, + 0.01003641914576292, + 0.036625079810619354, + 0.017626412212848663, + 0.003119339467957616, + 0.0014234361005946994, + 0.0055444627068936825, + 0.0031105156522244215, + 0.003512826282531023, + 0.0029403422959148884, + 0.00249794265255332, + 0.004806376062333584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.5761765837669373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650686591863632 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650686591863632, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650686591863632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11566197127103806, + 0.04936091974377632, + 0.08182034641504288, + 0.04048692807555199, + 0.04021138697862625, + 0.04725541174411774, + 0.08482304215431213, + 0.022643672302365303, + 0.5036135911941528, + 0.01412278600037098 + ], + "credit_map": [ + 0.010348772630095482, + 0.00946810096502304, + 0.03878983110189438, + 0.00843161903321743, + 0.010742496699094772, + 0.008374868892133236, + 0.020471055060625076, + 0.009035220369696617, + 0.005658383946865797, + 0.005622650496661663, + 0.007621106691658497, + 0.005054270848631859, + 0.009199528023600578, + 0.005480351857841015, + 0.006184740923345089, + 0.004539884626865387, + 0.008446861989796162, + 0.01588246412575245, + 0.012772071175277233, + 0.011234893463551998, + 0.007710645440965891, + 0.008849095553159714, + 0.008177668787539005, + 0.008746643550693989, + 0.00426878547295928, + 0.007271478418260813, + 0.0043281065300107, + 0.00551514932885766, + 0.0033878691028803587, + 0.006109801586717367, + 0.005330230575054884, + 0.004275504034012556, + 0.004749341867864132, + 0.0047403182834386826, + 0.003994478844106197, + 0.008898505009710789, + 0.004031592980027199, + 0.0035112849436700344, + 0.004432435147464275, + 0.005853421986103058, + 0.004656514152884483, + 0.005416343454271555, + 0.005983620882034302, + 0.006208072882145643, + 0.006374601740390062, + 0.005052546504884958, + 0.0072241234593093395, + 0.006339586805552244, + 0.007397746201604605, + 0.009845179505646229, + 0.010501831769943237, + 0.012084218673408031, + 0.008936939761042595, + 0.010607276111841202, + 0.007710555102676153, + 0.01773928664624691, + 0.0031644594855606556, + 0.0022331017535179853, + 0.004502084106206894, + 0.0027324650436639786, + 0.002023497596383095, + 0.0020010138396173716, + 0.002918015932664275, + 0.003069033846259117, + 0.04900198429822922, + 0.038791924715042114, + 0.06431858986616135, + 0.05609482526779175, + 0.08624979853630066, + 0.03714149072766304, + 0.11409474164247513, + 0.057920191437006, + 0.0015647237887606025, + 0.0009047748171724379, + 0.0018763643456622958, + 0.0016382650937885046, + 0.001987249357625842, + 0.0015810379991307855, + 0.0025689357426017523, + 0.0020014343317598104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.008650686591863632 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545594945549965 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545594945549965, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545594945549965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1766975075006485, + 0.07834439724683762, + 0.1385304033756256, + 0.0271396916359663, + 0.05101363733410835, + 0.07042980939149857, + 0.1342010200023651, + 0.07946008443832397, + 0.1409422606229782, + 0.1032412052154541 + ], + "credit_map": [ + 0.01566474325954914, + 0.010603506118059158, + 0.06623612344264984, + 0.014886924996972084, + 0.016624515876173973, + 0.008552159182727337, + 0.02132764272391796, + 0.022801904007792473, + 0.009334491565823555, + 0.00542418472468853, + 0.025105640292167664, + 0.009378230199217796, + 0.007699442096054554, + 0.006298614665865898, + 0.008669327944517136, + 0.006434471812099218, + 0.015886973589658737, + 0.016245050355792046, + 0.01848827488720417, + 0.02553657442331314, + 0.013790427707135677, + 0.020266203209757805, + 0.014546067453920841, + 0.013770841993391514, + 0.002273039659485221, + 0.005671937018632889, + 0.0037795624230057, + 0.003514113137498498, + 0.002607442671433091, + 0.0035183364525437355, + 0.0028912480920553207, + 0.0028840121813118458, + 0.005068901460617781, + 0.007596024312078953, + 0.006413572002202272, + 0.009671724401414394, + 0.005101176910102367, + 0.004448429681360722, + 0.0062750293873250484, + 0.006438783835619688, + 0.007448709104210138, + 0.008794198743999004, + 0.010654275305569172, + 0.007359718903899193, + 0.006081486586481333, + 0.013013786636292934, + 0.006572643294930458, + 0.010504995472729206, + 0.01410878635942936, + 0.012501493096351624, + 0.01834697090089321, + 0.01453377865254879, + 0.01702774502336979, + 0.011671962216496468, + 0.026912609115242958, + 0.019097689539194107, + 0.009940600953996181, + 0.0064244805835187435, + 0.014130685478448868, + 0.007001304067671299, + 0.013455750420689583, + 0.0064319949597120285, + 0.013370855711400509, + 0.008704413659870625, + 0.014015952125191689, + 0.010449322871863842, + 0.016215410083532333, + 0.016185011714696884, + 0.023532971739768982, + 0.010327639989554882, + 0.032499928027391434, + 0.01771603710949421, + 0.012218975462019444, + 0.006708871573209763, + 0.012729358859360218, + 0.007622881326824427, + 0.016177358105778694, + 0.012975522316992283, + 0.023498905822634697, + 0.011309335008263588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.012545594945549965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107415184378624 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.01107415184378624, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107415184378624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4922333061695099, + 0.06205607205629349, + 0.09130960702896118, + 0.04231330379843712, + 0.07828767597675323, + 0.03161008656024933, + 0.05160810798406601, + 0.04101694002747536, + 0.08182401210069656, + 0.02774093486368656 + ], + "credit_map": [ + 0.0435599721968174, + 0.04062272980809212, + 0.14561735093593597, + 0.037980224937200546, + 0.060290656983852386, + 0.03487247973680496, + 0.04844293370842934, + 0.08084693551063538, + 0.006071233190596104, + 0.007225916720926762, + 0.013569866307079792, + 0.007865533232688904, + 0.009160181507468224, + 0.004099844954907894, + 0.007812763564288616, + 0.0062507279217243195, + 0.014913053251802921, + 0.006254442501813173, + 0.007825000211596489, + 0.0074321855790913105, + 0.0133402980864048, + 0.021628055721521378, + 0.008607786148786545, + 0.011308783665299416, + 0.006210504099726677, + 0.005545958876609802, + 0.0055206106044352055, + 0.00449227774515748, + 0.0031889535021036863, + 0.008103903383016586, + 0.004509124904870987, + 0.004741972777992487, + 0.00900887232273817, + 0.009223888628184795, + 0.010029762983322144, + 0.014529643580317497, + 0.008142419159412384, + 0.005685874260962009, + 0.008625295013189316, + 0.013041919097304344, + 0.0038319651503115892, + 0.0037756101228296757, + 0.005366069730371237, + 0.004185996949672699, + 0.002889773575589061, + 0.004316392354667187, + 0.0028440169990062714, + 0.004400260746479034, + 0.00545941386371851, + 0.004626434296369553, + 0.006484428886324167, + 0.0056481207720935345, + 0.004417592193931341, + 0.0065950737334787846, + 0.008300417102873325, + 0.010076628066599369, + 0.003978631924837828, + 0.00301339547149837, + 0.005130823235958815, + 0.0062517873011529446, + 0.003537117736414075, + 0.004878802690654993, + 0.0034899881575256586, + 0.010736394673585892, + 0.00755339814350009, + 0.006481858901679516, + 0.009336880408227444, + 0.008933022618293762, + 0.014716198667883873, + 0.006207908038049936, + 0.0188145749270916, + 0.009780172258615494, + 0.0025086081586778164, + 0.0021017836406826973, + 0.0027532277163118124, + 0.0030222395434975624, + 0.0035341440234333277, + 0.0024008231703191996, + 0.007756604813039303, + 0.003663501236587763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01107415184378624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot, the camera slowly pans right, tracking a lone figure walking down a broad, rain-soaked street in front of the magnificent Arc de Triomphe. The figure is a sharp, resolute man, dressed in a long, dark trench coat, its damp surface reflecting chaotic light like polished obsidian. He walks with a confident, unhurried demeanor, his hands tucked deep into his pockets, his heavy boots deliberately splashing in shallow puddles. In the background, the monumental Arc de Triomphe looms, its classical form reimagined in a gritty, Grand Theft Auto art style; its stone surfaces are dark, weathered, and streaked with rain, while its intricate carvings are rendered into sharp, dramatic relief by harsh artificial light. The entire scene is saturated in the glow of a rain-soaked sky, with vibrant pinks, electric blues, and vivid greens casting long, distorted reflections across the smooth, inky black asphalt. A steady downpour catches the light, creating a shimmering, dynamic veil across the frame, creating a cool, moody and intense cinematic atmosphere that blends historical grandeur with a menacing, digitally rendered, and distinctly modern urban noir aesthetic.", + "seed": 23426, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_10/index_4238_seed_23426_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08035688144154847, + "vlm_logprob": -0.08035688144154847, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.08035688144154847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13555973768234253 + ], + "opposite_logprob_tokens": [ + -2.5105597972869873 + ], + "logprob_reward": -0.13555973768234253, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.13555973768234253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3103260099887848, + 0.09726738184690475, + 0.0720251277089119, + 0.05869491398334503, + 0.13273927569389343, + 0.0672977939248085, + 0.07012853771448135, + 0.05479293689131737, + 0.07001635432243347, + 0.06671169400215149 + ], + "credit_map": [ + 0.04288603737950325, + 0.04761672765016556, + 0.033175960183143616, + 0.036118973046541214, + 0.041659604758024216, + 0.034437187016010284, + 0.05325417220592499, + 0.021177342161536217, + 0.01364085916429758, + 0.01425432413816452, + 0.017368141561746597, + 0.009798035956919193, + 0.009531683288514614, + 0.009151781909167767, + 0.014754880219697952, + 0.00876767747104168, + 0.009333009831607342, + 0.015116746537387371, + 0.010256309062242508, + 0.005787055008113384, + 0.009623766876757145, + 0.007252299226820469, + 0.010395697318017483, + 0.0042602429166436195, + 0.00770274642854929, + 0.009879622608423233, + 0.005490536335855722, + 0.007698947098106146, + 0.008189179934561253, + 0.004935013595968485, + 0.009732577949762344, + 0.005066290032118559, + 0.02075275592505932, + 0.011761464178562164, + 0.02674294449388981, + 0.011117078363895416, + 0.009577970020473003, + 0.02400227263569832, + 0.010918033309280872, + 0.017866767942905426, + 0.008130364120006561, + 0.008735397830605507, + 0.007896329276263714, + 0.01425427570939064, + 0.006523103453218937, + 0.004235695116221905, + 0.009256240911781788, + 0.008266383782029152, + 0.013209984637796879, + 0.0067206635139882565, + 0.011516197584569454, + 0.0063735367730259895, + 0.005715276580303907, + 0.015348642133176327, + 0.005633198190480471, + 0.00561104342341423, + 0.009118050336837769, + 0.0049309441819787025, + 0.003666905453428626, + 0.0030248514376580715, + 0.004795494955033064, + 0.01800484210252762, + 0.004474732559174299, + 0.006777116563171148, + 0.007634191773831844, + 0.010416232980787754, + 0.00781303085386753, + 0.008716363459825516, + 0.00821383111178875, + 0.005103128030896187, + 0.016967415809631348, + 0.005152163095772266, + 0.008058623410761356, + 0.00849758181720972, + 0.007227516733109951, + 0.010923624970018864, + 0.00883003231137991, + 0.004774527624249458, + 0.010646861977875233, + 0.007752923294901848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.13555973768234253 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215205550193787 + ], + "opposite_logprob_tokens": [ + -1.5021520853042603 + ], + "logprob_reward": -0.25215205550193787, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215205550193787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15455961227416992, + 0.17043744027614594, + 0.032221321016550064, + 0.11028744280338287, + 0.08179710060358047, + 0.05842066928744316, + 0.1570514291524887, + 0.041802845895290375, + 0.03688640147447586, + 0.15653568506240845 + ], + "credit_map": [ + 0.020344356074929237, + 0.021895257756114006, + 0.019160641357302666, + 0.015325471758842468, + 0.022061768919229507, + 0.017964009195566177, + 0.026548460125923157, + 0.011259651742875576, + 0.021440710872411728, + 0.02285762131214142, + 0.02261231280863285, + 0.01988348551094532, + 0.01754874549806118, + 0.01680411957204342, + 0.03059488534927368, + 0.01869557984173298, + 0.004090690053999424, + 0.007116588298231363, + 0.0030480949208140373, + 0.0026749849785119295, + 0.005844959989190102, + 0.002594845602288842, + 0.004380867350846529, + 0.0024702921509742737, + 0.014536538161337376, + 0.015453655272722244, + 0.00919165089726448, + 0.012288657948374748, + 0.014898027293384075, + 0.015146031975746155, + 0.01976197399199009, + 0.009010916575789452, + 0.012405989691615105, + 0.009178644977509975, + 0.017273252829909325, + 0.010033530183136463, + 0.005818542558699846, + 0.006616942584514618, + 0.010789801366627216, + 0.009680401533842087, + 0.008204047568142414, + 0.008958443999290466, + 0.007910634391009808, + 0.01136033609509468, + 0.005684204865247011, + 0.005048719700425863, + 0.006571523379534483, + 0.004682765807956457, + 0.038766101002693176, + 0.0114483293145895, + 0.013687614351511002, + 0.007303598802536726, + 0.0121738500893116, + 0.05306672304868698, + 0.008235594257712364, + 0.01236963551491499, + 0.004618523642420769, + 0.004379427060484886, + 0.0033159018494188786, + 0.004101797007024288, + 0.005709794349968433, + 0.009766343981027603, + 0.00607002479955554, + 0.0038410364650189877, + 0.005396945867687464, + 0.006803148426115513, + 0.005230042152106762, + 0.003264744533225894, + 0.003981165122240782, + 0.0068662129342556, + 0.0021076721604913473, + 0.0032364726066589355, + 0.038624439388513565, + 0.010780245065689087, + 0.012104715220630169, + 0.010124868713319302, + 0.012421135790646076, + 0.04458717256784439, + 0.014096353203058243, + 0.013796773739159107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.25215205550193787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036221628542989492 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036221628542989492, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036221628542989492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12529675662517548, + 0.12142740190029144, + 0.06332859396934509, + 0.07557344436645508, + 0.10120755434036255, + 0.22937394678592682, + 0.020944418385624886, + 0.04250166565179825, + 0.10022054612636566, + 0.12012571096420288 + ], + "credit_map": [ + 0.016571728512644768, + 0.020643381401896477, + 0.010995857417583466, + 0.016703598201274872, + 0.016909662634134293, + 0.013671262189745903, + 0.016588222235441208, + 0.013213030993938446, + 0.01585390977561474, + 0.015647193416953087, + 0.01497092004865408, + 0.013802158646285534, + 0.012882568873465061, + 0.011151766404509544, + 0.02408624067902565, + 0.01303265430033207, + 0.011452240869402885, + 0.020215217024087906, + 0.009282945655286312, + 0.00446920795366168, + 0.005994834471493959, + 0.003974476829171181, + 0.005011498462408781, + 0.0029281675815582275, + 0.012848340906202793, + 0.020881598815321922, + 0.011240444146096706, + 0.007953139021992683, + 0.007009007502347231, + 0.004707442596554756, + 0.006593077443540096, + 0.004340394865721464, + 0.013454209081828594, + 0.012954818084836006, + 0.023004015907645226, + 0.011818591505289078, + 0.01166625041514635, + 0.009571113623678684, + 0.009203709661960602, + 0.009534846059978008, + 0.038966674357652664, + 0.034470297396183014, + 0.0458739697933197, + 0.02985718473792076, + 0.022767448797822, + 0.015757063403725624, + 0.02107950486242771, + 0.020601799711585045, + 0.0039416090585291386, + 0.002043782966211438, + 0.0027730390429496765, + 0.001212476403452456, + 0.0019256017403677106, + 0.004922572989016771, + 0.0015195017913356423, + 0.0026058354414999485, + 0.006558072753250599, + 0.004520669113844633, + 0.004631198011338711, + 0.0024234100710600615, + 0.004786960780620575, + 0.011167766526341438, + 0.004147706553339958, + 0.0042658839374780655, + 0.018911389634013176, + 0.01744779758155346, + 0.011338869109749794, + 0.009895429015159607, + 0.010497163981199265, + 0.008434087038040161, + 0.01481520477682352, + 0.008880609646439552, + 0.03070525825023651, + 0.00744926743209362, + 0.009497163817286491, + 0.006457463838160038, + 0.008233841508626938, + 0.0359189473092556, + 0.011775480583310127, + 0.010088291950523853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0036221628542989492 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10528185963630676, + 0.11146140098571777, + 0.245682492852211, + 0.1844293624162674, + 0.07386751472949982, + 0.06805845350027084, + 0.03136613219976425, + 0.022853296250104904, + 0.08572588115930557, + 0.071273572742939 + ], + "credit_map": [ + 0.013470228761434555, + 0.01870293915271759, + 0.008543912321329117, + 0.012221921235322952, + 0.015038605779409409, + 0.012088721618056297, + 0.017338106408715248, + 0.007877420634031296, + 0.014355012215673923, + 0.014732644893229008, + 0.011732716113328934, + 0.010019520297646523, + 0.013148163445293903, + 0.01419790182262659, + 0.021689115092158318, + 0.011586322449147701, + 0.04997161403298378, + 0.09386865794658661, + 0.013068574480712414, + 0.015841081738471985, + 0.02738310769200325, + 0.017923634499311447, + 0.017602410167455673, + 0.01002341229468584, + 0.023668237030506134, + 0.02681749500334263, + 0.020035773515701294, + 0.020548101514577866, + 0.022587383165955544, + 0.01655241847038269, + 0.03818649426102638, + 0.016033457592129707, + 0.011310228146612644, + 0.006348205730319023, + 0.01715155690908432, + 0.008728044107556343, + 0.004802707117050886, + 0.01007256843149662, + 0.006803431082516909, + 0.008650773204863071, + 0.009187517687678337, + 0.00795861054211855, + 0.01446785032749176, + 0.009416059590876102, + 0.006334115285426378, + 0.004805973265320063, + 0.00889633595943451, + 0.006991983391344547, + 0.00538483913987875, + 0.0029278811998665333, + 0.004575361963361502, + 0.0022152767051011324, + 0.0029218043200671673, + 0.006619239691644907, + 0.002612436655908823, + 0.004109290428459644, + 0.00271647609770298, + 0.0029709625523537397, + 0.002319660037755966, + 0.003013809211552143, + 0.003352673724293709, + 0.0024824070278555155, + 0.0039330958388745785, + 0.0020642122253775597, + 0.0132511081174016, + 0.01353617012500763, + 0.008968218229711056, + 0.007462217938154936, + 0.008360032923519611, + 0.016286304220557213, + 0.009840709157288074, + 0.008021116256713867, + 0.011099700815975666, + 0.009967271238565445, + 0.005767996422946453, + 0.008140881545841694, + 0.01005757786333561, + 0.011097774840891361, + 0.009517702274024487, + 0.005624670069664717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.007631551939994097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning right while tracking a lone man walking down a rain-soaked street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Arc de Triomphe visible in the background with a gritty, Grand Theft Auto art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man wearing a long, dark trench coat with a damp surface reflecting chaotic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene saturated with vibrant pinks, electric blues, and vivid greens casting reflections on the asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818899229168892 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.002818899229168892, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818899229168892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13320598006248474, + 0.07295314222574234, + 0.06174243241548538, + 0.1299493908882141, + 0.061334338039159775, + 0.08055952191352844, + 0.24060149490833282, + 0.04311806336045265, + 0.09756619483232498, + 0.07896941900253296 + ], + "credit_map": [ + 0.01807156391441822, + 0.021353479474782944, + 0.011148651130497456, + 0.017226586118340492, + 0.0182209350168705, + 0.010235500521957874, + 0.018929049372673035, + 0.018020212650299072, + 0.009656596928834915, + 0.0113884462043643, + 0.008771426975727081, + 0.00687272846698761, + 0.007877439260482788, + 0.008144402876496315, + 0.013097901828587055, + 0.00714420061558485, + 0.01063188724219799, + 0.015589624643325806, + 0.011589058674871922, + 0.006330628879368305, + 0.005513472948223352, + 0.0045063551515340805, + 0.0041290936060249805, + 0.0034523087088018656, + 0.022177018225193024, + 0.028681958094239235, + 0.013739115558564663, + 0.014672248624265194, + 0.012533641420304775, + 0.01421365886926651, + 0.012266346253454685, + 0.011665408499538898, + 0.009097278118133545, + 0.005889464635401964, + 0.007371488492935896, + 0.004714304581284523, + 0.0051176780834794044, + 0.015634527429938316, + 0.006045080721378326, + 0.007464515510946512, + 0.010676312260329723, + 0.009505124762654305, + 0.01010227669030428, + 0.01113745104521513, + 0.006878296844661236, + 0.007468166761100292, + 0.010395413264632225, + 0.01439647376537323, + 0.06408260017633438, + 0.015369485132396221, + 0.018602358177304268, + 0.006576885469257832, + 0.019543934613466263, + 0.08694380521774292, + 0.010337140411138535, + 0.01914527826011181, + 0.005336536094546318, + 0.007556953001767397, + 0.005809632129967213, + 0.00446805264800787, + 0.006864618044346571, + 0.004507852252572775, + 0.004693061113357544, + 0.0038813548162579536, + 0.017912128940224648, + 0.01442538108676672, + 0.01103015337139368, + 0.007345682475715876, + 0.010249100625514984, + 0.0213266983628273, + 0.006196252070367336, + 0.009080790914595127, + 0.019266800954937935, + 0.00478376355022192, + 0.007715074345469475, + 0.003299439325928688, + 0.004993173759430647, + 0.026525547727942467, + 0.004853667225688696, + 0.0075319563038647175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.002818899229168892 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_tensor([4971, 4971])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_tensor([4971, 4971])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4444ecf354cc239d1af760ae5474323ccca962ad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_tensor([4971, 4971])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning first-person close-up, the camera held perfectly still, immersing the viewer completely in the operator's field of vision. At the bottom of the frame, a pair of bony hands with meticulously manicured nails appeared steadily in view. In the right hand, a scalpel, gleaming with a dazzling metallic gleam, was gripped. Its sharp blade carefully and slowly sliced into the black plastic casing of an old-fashioned car CD player. The surface of the machine was covered in a crisscross of scratches, deep and shallow, and the marks of age. With steady, slow pressure from the wrist, the blade silently sliced through the casing, leaving a precise and deep incision. As the casing peeled away, piece by piece, revealing not the expected cold metal circuit boards and cables, but clumps and clumps of vibrant, moist, gray-white nerve tissue. Like countless tangled worms, they pulsed and wriggled in a slow, rhythmic rhythm. Fine veins of blood spread across the surface, and some nerve endings even twitched slightly at the contact of air. The entire picture adopts the artistic style of Junji Ito, presenting an extreme black and white high contrast. The only top light cuts the picture sharply, creating dazzling highlights on the metal surface of the scalpel and the slippery nerve tissue, while in the grooves of the scratches and the folds of the tissue, there is an impenetrable pure darkness. This extreme detail is combined with the biological grotesqueness to create a strong and disturbing atmosphere that is a mixture of surgical calmness and unspeakable physiological fear.", + "seed": 26686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_4971_seed_26686_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2072168545331806, + "vlm_logprob": -0.2072168545331806, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.2072168545331806, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9320031404495239 + ], + "opposite_logprob_tokens": [ + -0.5570031404495239 + ], + "logprob_reward": -0.9320031404495239, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9320031404495239, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2998793423175812, + 0.07578021287918091, + 0.0638018399477005, + 0.13645552098751068, + 0.055735569447278976, + 0.10543355345726013, + 0.032218996435403824, + 0.03923700377345085, + 0.09924487769603729, + 0.09221317619085312 + ], + "credit_map": [ + 0.021098840981721878, + 0.017822636291384697, + 0.07849008589982986, + 0.0365038737654686, + 0.02288384735584259, + 0.026348121464252472, + 0.06473439931869507, + 0.03199750557541847, + 0.00850539468228817, + 0.007032962515950203, + 0.010096716694533825, + 0.01196980755776167, + 0.009805149398744106, + 0.0074949138797819614, + 0.008235704153776169, + 0.012639557011425495, + 0.0076881819404661655, + 0.008340030908584595, + 0.00886127632111311, + 0.009480111300945282, + 0.0057094525545835495, + 0.004514500964432955, + 0.007348436862230301, + 0.011859847232699394, + 0.015245779417455196, + 0.021582908928394318, + 0.017058517783880234, + 0.02075977437198162, + 0.017501134425401688, + 0.009150427766144276, + 0.014364385046064854, + 0.020792577415704727, + 0.005417744163423777, + 0.005104007199406624, + 0.009349824860692024, + 0.00895819440484047, + 0.00603558961302042, + 0.0026204499881714582, + 0.0064765531569719315, + 0.011773197911679745, + 0.022732404991984367, + 0.0129068773239851, + 0.02099989354610443, + 0.01229874137789011, + 0.00887124240398407, + 0.006076831836253405, + 0.008870174176990986, + 0.012677383609116077, + 0.004095710813999176, + 0.005141749046742916, + 0.007033578585833311, + 0.0039421007968485355, + 0.0032942364923655987, + 0.0021697806660085917, + 0.003370530204847455, + 0.003171304939314723, + 0.0042336611077189445, + 0.005260113626718521, + 0.005706817843019962, + 0.005404276307672262, + 0.005808352027088404, + 0.003751672338694334, + 0.004452923312783241, + 0.004619182553142309, + 0.01665094681084156, + 0.013335798867046833, + 0.013821140863001347, + 0.014531138353049755, + 0.010904348455369473, + 0.004785929806530476, + 0.011951330117881298, + 0.013264230452477932, + 0.017345497384667397, + 0.008435788564383984, + 0.017533402889966965, + 0.01274698507040739, + 0.006599624175578356, + 0.00550922891125083, + 0.014247683808207512, + 0.009794955141842365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.9320031404495239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311535134911537 + ], + "opposite_logprob_tokens": [ + -3.1681153774261475 + ], + "logprob_reward": -0.04311535134911537, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311535134911537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4726185202598572, + 0.07018253952264786, + 0.13138128817081451, + 0.05693081393837929, + 0.01772758923470974, + 0.08171956241130829, + 0.05482141301035881, + 0.025729861110448837, + 0.021353742107748985, + 0.06753465533256531 + ], + "credit_map": [ + 0.030989184975624084, + 0.02971545234322548, + 0.13650156557559967, + 0.059200625866651535, + 0.028474697843194008, + 0.04899006336927414, + 0.09244287759065628, + 0.04630407318472862, + 0.013933442533016205, + 0.008200116455554962, + 0.006702242884784937, + 0.010225455276668072, + 0.00680495286360383, + 0.004761252086609602, + 0.011545316316187382, + 0.008009765297174454, + 0.014699183404445648, + 0.018345320597290993, + 0.017542414367198944, + 0.018332464620471, + 0.013885175809264183, + 0.009219417348504066, + 0.013109207153320312, + 0.02624809555709362, + 0.005262577906250954, + 0.010232411324977875, + 0.0077577196061611176, + 0.009164338000118732, + 0.007729910314083099, + 0.0038514805492013693, + 0.006044114474207163, + 0.006888259667903185, + 0.0015467520570382476, + 0.0024608620442450047, + 0.0022019643802195787, + 0.0025989303831011057, + 0.0024888492189347744, + 0.0013532823650166392, + 0.002127493266016245, + 0.00294945458881557, + 0.01568358764052391, + 0.009501206688582897, + 0.012796353548765182, + 0.014315806329250336, + 0.008017871528863907, + 0.004788348451256752, + 0.007570766843855381, + 0.009045615792274475, + 0.012035753577947617, + 0.0051753041334450245, + 0.009708395227789879, + 0.004632546100765467, + 0.005579998251050711, + 0.0038153165951371193, + 0.0056820944882929325, + 0.00819200649857521, + 0.0031855166889727116, + 0.004291660618036985, + 0.003230123547837138, + 0.0045665777288377285, + 0.0033975625410676003, + 0.0017379289492964745, + 0.0030590109527111053, + 0.002261479152366519, + 0.0027461256831884384, + 0.0032746659126132727, + 0.0031298391986638308, + 0.0032422684598714113, + 0.0025731902569532394, + 0.0014862908283248544, + 0.0023409933783113956, + 0.0025603678077459335, + 0.011377115733921528, + 0.006410543341189623, + 0.011682628653943539, + 0.009855044074356556, + 0.004603621084243059, + 0.0036319547798484564, + 0.012892640195786953, + 0.007081110030412674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.04311535134911537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980076752603054 + ], + "opposite_logprob_tokens": [ + -4.634800910949707 + ], + "logprob_reward": -0.00980076752603054, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980076752603054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2738870680332184, + 0.09909304976463318, + 0.06677330285310745, + 0.11015187203884125, + 0.07096808403730392, + 0.028452062979340553, + 0.127274289727211, + 0.05912993848323822, + 0.10742072016000748, + 0.05684959515929222 + ], + "credit_map": [ + 0.014720751903951168, + 0.02119017206132412, + 0.0710957869887352, + 0.0314074270427227, + 0.017422357574105263, + 0.0206908006221056, + 0.06836406141519547, + 0.028995713219046593, + 0.00865195132791996, + 0.010249009355902672, + 0.008413714356720448, + 0.015395402908325195, + 0.015014538541436195, + 0.011883163824677467, + 0.016851797699928284, + 0.012633469887077808, + 0.0076538873836398125, + 0.0067167081870138645, + 0.011568483896553516, + 0.010499340482056141, + 0.0074527026154100895, + 0.006699745077639818, + 0.008808542042970657, + 0.007373890373855829, + 0.013855059631168842, + 0.011658835224807262, + 0.014120290987193584, + 0.01658928021788597, + 0.01290417555719614, + 0.012685539200901985, + 0.010763987898826599, + 0.01757470890879631, + 0.007794382516294718, + 0.01065343338996172, + 0.008993404917418957, + 0.012710604816675186, + 0.009202069602906704, + 0.006163577549159527, + 0.007569012697786093, + 0.00788159016519785, + 0.004136423580348492, + 0.002776630222797394, + 0.0026812374126166105, + 0.005067878868430853, + 0.004248758777976036, + 0.0023886701092123985, + 0.0034173179883509874, + 0.0037351460196077824, + 0.020522432401776314, + 0.01740395464003086, + 0.028540680184960365, + 0.015627337619662285, + 0.01252761296927929, + 0.00944045651704073, + 0.011934342794120312, + 0.011277463287115097, + 0.009875296615064144, + 0.007465782575309277, + 0.010309023782610893, + 0.008570138365030289, + 0.007453872822225094, + 0.003876512637361884, + 0.006083874497562647, + 0.005495432764291763, + 0.017894692718982697, + 0.01481981948018074, + 0.012945917434990406, + 0.018458135426044464, + 0.012975050136446953, + 0.005021519027650356, + 0.01305532455444336, + 0.012250257655978203, + 0.0065622348338365555, + 0.005492503754794598, + 0.008671109564602375, + 0.00966566987335682, + 0.0065641929395496845, + 0.004068495240062475, + 0.008621739223599434, + 0.0072036501951515675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.00980076752603054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024921807926148176 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024921807926148176, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024921807926148176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719559282064438, + 0.1278219223022461, + 0.06648258864879608, + 0.09633062779903412, + 0.04498537629842758, + 0.08439909666776657, + 0.08788268268108368, + 0.07526879757642746, + 0.12012112140655518, + 0.12475181370973587 + ], + "credit_map": [ + 0.012024013325572014, + 0.011391195468604565, + 0.04132142663002014, + 0.02181067503988743, + 0.013701659627258778, + 0.014345613308250904, + 0.03581096604466438, + 0.021550379693508148, + 0.012951130978763103, + 0.01454995758831501, + 0.013362332247197628, + 0.018965905532240868, + 0.015210883691906929, + 0.012192447669804096, + 0.017736485227942467, + 0.022852787747979164, + 0.006659345235675573, + 0.010021244175732136, + 0.008655723184347153, + 0.008975498378276825, + 0.0075594400987029076, + 0.005207739770412445, + 0.007052905857563019, + 0.01235069427639246, + 0.011478018015623093, + 0.015550663694739342, + 0.011112350039184093, + 0.014654075726866722, + 0.009876348078250885, + 0.007064513396471739, + 0.009378104470670223, + 0.01721656695008278, + 0.005119175650179386, + 0.004714355804026127, + 0.00538703054189682, + 0.008498367853462696, + 0.0055544874630868435, + 0.003652219893410802, + 0.0049452148377895355, + 0.0071145291440188885, + 0.01484648510813713, + 0.00903385691344738, + 0.015633758157491684, + 0.015062009915709496, + 0.008573190309107304, + 0.005523747764527798, + 0.007008419372141361, + 0.008717640303075314, + 0.013881471939384937, + 0.01096371840685606, + 0.020452767610549927, + 0.011709139682352543, + 0.009040295146405697, + 0.006275271065533161, + 0.008073083125054836, + 0.007486939895898104, + 0.010754607617855072, + 0.012494376860558987, + 0.01573060266673565, + 0.010046685114502907, + 0.009798204526305199, + 0.004785156808793545, + 0.006140649318695068, + 0.005518518388271332, + 0.020837781950831413, + 0.015853263437747955, + 0.01429633516818285, + 0.0196464154869318, + 0.015577279031276703, + 0.006886015646159649, + 0.014380636624991894, + 0.012643400579690933, + 0.01729348860681057, + 0.012346956878900528, + 0.023327168077230453, + 0.023194724693894386, + 0.013098473660647869, + 0.007671205792576075, + 0.014308041892945766, + 0.013511753641068935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0024921807926148176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867283254861832 + ], + "opposite_logprob_tokens": [ + -3.048672914505005 + ], + "logprob_reward": -0.04867283254861832, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867283254861832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08876516669988632, + 0.08374923467636108, + 0.09773995727300644, + 0.06423641741275787, + 0.06705725938081741, + 0.11372644454240799, + 0.04244137555360794, + 0.0830836370587349, + 0.04052804782986641, + 0.3186725080013275 + ], + "credit_map": [ + 0.006789186503738165, + 0.006963636260479689, + 0.01625402271747589, + 0.02169235795736313, + 0.009264434687793255, + 0.0055227563716471195, + 0.012017332017421722, + 0.010261435993015766, + 0.01184450089931488, + 0.0074186609126627445, + 0.00977832730859518, + 0.017658710479736328, + 0.008730372413992882, + 0.006497707683593035, + 0.011154369451105595, + 0.010666579008102417, + 0.014351793564856052, + 0.011848784983158112, + 0.0120726702734828, + 0.013897458091378212, + 0.009064430370926857, + 0.007834136486053467, + 0.010181108489632607, + 0.018489575013518333, + 0.007656015455722809, + 0.01041395403444767, + 0.007932521402835846, + 0.008160924538969994, + 0.006763137876987457, + 0.0048112026415765285, + 0.0064535667188465595, + 0.01204509288072586, + 0.009109695442020893, + 0.010796982795000076, + 0.00961852166801691, + 0.008364344947040081, + 0.009198040701448917, + 0.0071563334204256535, + 0.005329013802111149, + 0.007484326604753733, + 0.017613422125577927, + 0.012216117233037949, + 0.021958038210868835, + 0.019945479929447174, + 0.010559072718024254, + 0.006868385244160891, + 0.009656096808612347, + 0.014909841120243073, + 0.004939853213727474, + 0.0058097634464502335, + 0.009384969249367714, + 0.005300583783537149, + 0.004922829102724791, + 0.002560388995334506, + 0.005853976588696241, + 0.003669015597552061, + 0.012379547581076622, + 0.010011580772697926, + 0.010426837019622326, + 0.01185606885701418, + 0.010910510085523129, + 0.008705779910087585, + 0.00947503186762333, + 0.009318283759057522, + 0.003857647767290473, + 0.0032651470974087715, + 0.00967509113252163, + 0.005044331308454275, + 0.0046987771056592464, + 0.0025319952983409166, + 0.005152476951479912, + 0.006302583031356335, + 0.03818324953317642, + 0.03346479684114456, + 0.051257915794849396, + 0.061439335346221924, + 0.026189254596829414, + 0.017933964729309082, + 0.058310747146606445, + 0.03189321607351303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.04867283254861832 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning first-person close-up, the camera held perfectly still, immersing the viewer completely in the operator's field of vision. At the bottom of the frame, a pair of bony hands with meticulously manicured nails appeared steadily in view. In the right hand, a scalpel, gleaming with a dazzling metallic gleam, was gripped. Its sharp blade carefully and slowly sliced into the black plastic casing of an old-fashioned car CD player. The surface of the machine was covered in a crisscross of scratches, deep and shallow, and the marks of age. With steady, slow pressure from the wrist, the blade silently sliced through the casing, leaving a precise and deep incision. As the casing peeled away, piece by piece, revealing not the expected cold metal circuit boards and cables, but clumps and clumps of vibrant, moist, gray-white nerve tissue. Like countless tangled worms, they pulsed and wriggled in a slow, rhythmic rhythm. Fine veins of blood spread across the surface, and some nerve endings even twitched slightly at the contact of air. The entire picture adopts the artistic style of Junji Ito, presenting an extreme black and white high contrast. The only top light cuts the picture sharply, creating dazzling highlights on the metal surface of the scalpel and the slippery nerve tissue, while in the grooves of the scratches and the folds of the tissue, there is an impenetrable pure darkness. This extreme detail is combined with the biological grotesqueness to create a strong and disturbing atmosphere that is a mixture of surgical calmness and unspeakable physiological fear.", + "seed": 26687, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_4971_seed_26687_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29939586501568555, + "vlm_logprob": -0.29939586501568555, + "vlm_yes_no_margin": 2.274999976158142, + "vlm_token_credit": -0.29939586501568555, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0703336000442505 + ], + "opposite_logprob_tokens": [ + -0.4453336000442505 + ], + "logprob_reward": -1.0703336000442505, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0703336000442505, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11421369761228561, + 0.055440980941057205, + 0.06290784478187561, + 0.12594284117221832, + 0.17724215984344482, + 0.03521355614066124, + 0.031657587736845016, + 0.1489236056804657, + 0.10737096518278122, + 0.14108678698539734 + ], + "credit_map": [ + 0.01873580552637577, + 0.014529293403029442, + 0.03350113332271576, + 0.01679803803563118, + 0.004201576113700867, + 0.006300783716142178, + 0.009231310337781906, + 0.010915755294263363, + 0.007138310465961695, + 0.00795545894652605, + 0.01130630262196064, + 0.007253143936395645, + 0.002730889245867729, + 0.005437776912003756, + 0.006121386773884296, + 0.007497709710150957, + 0.009269473142921925, + 0.007973844185471535, + 0.011440478265285492, + 0.007541681174188852, + 0.00459129735827446, + 0.007467169780284166, + 0.005759693682193756, + 0.008864209987223148, + 0.021621370688080788, + 0.01962982304394245, + 0.03435494378209114, + 0.013694253750145435, + 0.005684590898454189, + 0.009792483411729336, + 0.008652113378047943, + 0.012513261288404465, + 0.027455750852823257, + 0.01669314131140709, + 0.040124617516994476, + 0.022149384021759033, + 0.01008071843534708, + 0.020437950268387794, + 0.024281127378344536, + 0.01601947471499443, + 0.004791082348674536, + 0.00517592579126358, + 0.007762727327644825, + 0.003899014089256525, + 0.004032906144857407, + 0.004023095592856407, + 0.0031106891110539436, + 0.0024181173648685217, + 0.004754245281219482, + 0.0033690689597278833, + 0.005992450751364231, + 0.0034193838946521282, + 0.00309331645257771, + 0.003037887392565608, + 0.005043558776378632, + 0.0029476776253432035, + 0.018254511058330536, + 0.031636547297239304, + 0.022319037467241287, + 0.022100631147623062, + 0.020385367795825005, + 0.012460974976420403, + 0.014107851311564445, + 0.007658691611140966, + 0.012781632132828236, + 0.013563008047640324, + 0.014316529035568237, + 0.00862578023225069, + 0.012978561222553253, + 0.01683303341269493, + 0.01698996126651764, + 0.011282457038760185, + 0.029604965820908546, + 0.015998346731066704, + 0.02719452977180481, + 0.013852869160473347, + 0.005833007860928774, + 0.009686260484158993, + 0.006572170648723841, + 0.032344624400138855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -1.0703336000442505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521894872188568 + ], + "opposite_logprob_tokens": [ + -1.5021895170211792 + ], + "logprob_reward": -0.2521894872188568, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521894872188568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15211866796016693, + 0.18676303327083588, + 0.07592447102069855, + 0.030867232009768486, + 0.2281603217124939, + 0.038770146667957306, + 0.03165559843182564, + 0.058949537575244904, + 0.10933482646942139, + 0.08745618909597397 + ], + "credit_map": [ + 0.021118657663464546, + 0.021871699020266533, + 0.02272637188434601, + 0.017858315259218216, + 0.00881173275411129, + 0.013110856525599957, + 0.018586352467536926, + 0.028034674003720284, + 0.02690155990421772, + 0.025141365826129913, + 0.042245201766490936, + 0.022559938952326775, + 0.00973233301192522, + 0.017328830435872078, + 0.017433445900678635, + 0.025420352816581726, + 0.011837744154036045, + 0.009726847521960735, + 0.00949817430227995, + 0.006430872716009617, + 0.0075712986290454865, + 0.009291309863328934, + 0.008683775551617146, + 0.012884439900517464, + 0.004856019746512175, + 0.0044138869270682335, + 0.007611434441059828, + 0.004011212382465601, + 0.0015368014574050903, + 0.0023632231168448925, + 0.002963825361803174, + 0.0031108269467949867, + 0.031554531306028366, + 0.018625855445861816, + 0.046393152326345444, + 0.030924074351787567, + 0.013250662945210934, + 0.020316168665885925, + 0.045851439237594604, + 0.021244410425424576, + 0.004542197100818157, + 0.005239448975771666, + 0.005292179994285107, + 0.0026942219119518995, + 0.0050806086510419846, + 0.007268126122653484, + 0.004988889209926128, + 0.003664470976218581, + 0.004525556694716215, + 0.005444861017167568, + 0.005233196076005697, + 0.0035533851478248835, + 0.0027778868097811937, + 0.005028704646974802, + 0.0027535047847777605, + 0.0023385004606097937, + 0.007292105350643396, + 0.011567255482077599, + 0.009584092535078526, + 0.0074447644874453545, + 0.008281641639769077, + 0.005042524076998234, + 0.006233348045498133, + 0.0035038029309362173, + 0.011900984682142735, + 0.02265053428709507, + 0.013457662425935268, + 0.010234546847641468, + 0.009573202580213547, + 0.011844873428344727, + 0.01940908469259739, + 0.010263926349580288, + 0.01210649311542511, + 0.012579943984746933, + 0.021337606012821198, + 0.010575513355433941, + 0.005564198829233646, + 0.007164054550230503, + 0.007960094138979912, + 0.010168279521167278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.2521894872188568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370694860816002 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370694860816002, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370694860816002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12931600213050842, + 0.1575680375099182, + 0.06405463814735413, + 0.040190450847148895, + 0.09161366522312164, + 0.07981561124324799, + 0.04062933474779129, + 0.07314441353082657, + 0.16677802801132202, + 0.15688984096050262 + ], + "credit_map": [ + 0.018968302756547928, + 0.020271653309464455, + 0.020946258679032326, + 0.01326875202357769, + 0.00813189148902893, + 0.010505404323339462, + 0.014477165415883064, + 0.022746574133634567, + 0.022759104147553444, + 0.02120104618370533, + 0.03599165007472038, + 0.01894942671060562, + 0.00805894285440445, + 0.014894303865730762, + 0.015091040171682835, + 0.02062251977622509, + 0.010819472372531891, + 0.006191108841449022, + 0.011191337369382381, + 0.007209177128970623, + 0.005630721338093281, + 0.007081659976392984, + 0.006250170525163412, + 0.00968099758028984, + 0.006579873152077198, + 0.006499414797872305, + 0.008962202817201614, + 0.0038011870346963406, + 0.002539924345910549, + 0.0037631955929100513, + 0.003113438142463565, + 0.004931212402880192, + 0.011944839730858803, + 0.007379303220659494, + 0.012713066302239895, + 0.011936133727431297, + 0.0059511992149055, + 0.008955273777246475, + 0.023032521829009056, + 0.009701336733996868, + 0.009532490745186806, + 0.010005093179643154, + 0.010792774148285389, + 0.007135079242289066, + 0.012190500274300575, + 0.0116798160597682, + 0.011502253822982311, + 0.006977597251534462, + 0.005545313004404306, + 0.005357224494218826, + 0.010406081564724445, + 0.005277888849377632, + 0.0036771437153220177, + 0.0035867048427462578, + 0.004311000928282738, + 0.0024679789785295725, + 0.009862760081887245, + 0.016040438786149025, + 0.011179830878973007, + 0.009810087271034718, + 0.01003629807382822, + 0.005327387247234583, + 0.007333952467888594, + 0.0035536494106054306, + 0.01928076706826687, + 0.027061890810728073, + 0.018908130005002022, + 0.016080299392342567, + 0.013877016492187977, + 0.019359052181243896, + 0.02954004518687725, + 0.022670816630125046, + 0.024530885741114616, + 0.021784691140055656, + 0.033702317625284195, + 0.01614023558795452, + 0.00844863522797823, + 0.013523735105991364, + 0.013384465128183365, + 0.025374865159392357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.026370694860816002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11864811927080154, + 0.1104477271437645, + 0.07857766002416611, + 0.10430284589529037, + 0.05321947857737541, + 0.02575802430510521, + 0.0296417735517025, + 0.07191650569438934, + 0.16721166670322418, + 0.24027618765830994 + ], + "credit_map": [ + 0.0206749327480793, + 0.01936108060181141, + 0.031109336763620377, + 0.012695573270320892, + 0.006184358149766922, + 0.007575816009193659, + 0.009389991872012615, + 0.01165703684091568, + 0.01746279187500477, + 0.017403753474354744, + 0.03062368556857109, + 0.016186144202947617, + 0.003562086494639516, + 0.005617313552647829, + 0.010738070122897625, + 0.00885388720780611, + 0.01340185571461916, + 0.009829900227487087, + 0.02100462093949318, + 0.01103687658905983, + 0.004978396464139223, + 0.005223862361162901, + 0.0060129896737635136, + 0.007089159917086363, + 0.017898494377732277, + 0.01789012923836708, + 0.03297251835465431, + 0.013179963454604149, + 0.004382498096674681, + 0.006968876346945763, + 0.005090880673378706, + 0.00591948302462697, + 0.008810093626379967, + 0.005077499896287918, + 0.009806583635509014, + 0.005316616967320442, + 0.0031092502176761627, + 0.00802693422883749, + 0.004545019008219242, + 0.008527479134500027, + 0.0037721877451986074, + 0.003768302733078599, + 0.00635368749499321, + 0.002672078087925911, + 0.0022674808278679848, + 0.0031158567871898413, + 0.002101865131407976, + 0.00170656677801162, + 0.003866104409098625, + 0.004455086775124073, + 0.006895389873534441, + 0.0036334653850644827, + 0.0032694016117602587, + 0.0030484332237392664, + 0.0022520527709275484, + 0.002221839502453804, + 0.01084983628243208, + 0.016809554770588875, + 0.011831022799015045, + 0.01070910133421421, + 0.007776952814310789, + 0.004812086932361126, + 0.006195984315127134, + 0.002931963885203004, + 0.020848629996180534, + 0.023267250508069992, + 0.01890452578663826, + 0.016669543460011482, + 0.017403002828359604, + 0.021646039560437202, + 0.027656327933073044, + 0.020816350355744362, + 0.03956989571452141, + 0.03555937483906746, + 0.048146847635507584, + 0.026327790692448616, + 0.013591621071100235, + 0.01956450566649437, + 0.01535953301936388, + 0.0421566367149353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0052584195509552956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428271234035492 + ], + "opposite_logprob_tokens": [ + -2.017827033996582 + ], + "logprob_reward": -0.1428271234035492, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428271234035492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08859842270612717, + 0.04124030843377113, + 0.062498539686203, + 0.3608057498931885, + 0.17064382135868073, + 0.013812325894832611, + 0.03510701283812523, + 0.07471194863319397, + 0.03628699108958244, + 0.11629482358694077 + ], + "credit_map": [ + 0.016878265887498856, + 0.014718546532094479, + 0.024735841900110245, + 0.010636481456458569, + 0.0035582720302045345, + 0.004703741054981947, + 0.006383155472576618, + 0.006984129082411528, + 0.006696274038404226, + 0.008581263944506645, + 0.00902943778783083, + 0.004960709251463413, + 0.002487665042281151, + 0.002996667753905058, + 0.0034269546158611774, + 0.0030613441485911608, + 0.011594324372708797, + 0.012114981189370155, + 0.011727719567716122, + 0.005493147298693657, + 0.004279540851712227, + 0.004744735546410084, + 0.005017269402742386, + 0.007526826579123735, + 0.06377458572387695, + 0.0594140887260437, + 0.10686049610376358, + 0.04523257911205292, + 0.015511921606957912, + 0.023397918790578842, + 0.020393965765833855, + 0.026220226660370827, + 0.025829030200839043, + 0.01694219559431076, + 0.04181913286447525, + 0.018437165766954422, + 0.009855932556092739, + 0.019022012129426003, + 0.0162810068577528, + 0.022457348182797432, + 0.001725661801174283, + 0.0022399048320949078, + 0.0019914167933166027, + 0.0011674243723973632, + 0.0014052724000066519, + 0.002463500015437603, + 0.001589456689544022, + 0.0012296905042603612, + 0.004920371808111668, + 0.004023818764835596, + 0.009193985722959042, + 0.004681909456849098, + 0.002726114122197032, + 0.0027157210279256105, + 0.0049458895809948444, + 0.0018992068944498897, + 0.01181053277105093, + 0.020216161385178566, + 0.013524356298148632, + 0.007770419120788574, + 0.006586891133338213, + 0.005610784515738487, + 0.005625074729323387, + 0.003567721927538514, + 0.004459800664335489, + 0.005218114238232374, + 0.005420297384262085, + 0.004871533252298832, + 0.003641815623268485, + 0.0033136073034256697, + 0.005139568354934454, + 0.004222259856760502, + 0.01907748356461525, + 0.02061593160033226, + 0.020095106214284897, + 0.012712260708212852, + 0.005390399135649204, + 0.010770738124847412, + 0.008691329509019852, + 0.01894158311188221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.1428271234035492 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning first-person close-up, the camera held perfectly still, immersing the viewer completely in the operator's field of vision. At the bottom of the frame, a pair of bony hands with meticulously manicured nails appeared steadily in view. In the right hand, a scalpel, gleaming with a dazzling metallic gleam, was gripped. Its sharp blade carefully and slowly sliced into the black plastic casing of an old-fashioned car CD player. The surface of the machine was covered in a crisscross of scratches, deep and shallow, and the marks of age. With steady, slow pressure from the wrist, the blade silently sliced through the casing, leaving a precise and deep incision. As the casing peeled away, piece by piece, revealing not the expected cold metal circuit boards and cables, but clumps and clumps of vibrant, moist, gray-white nerve tissue. Like countless tangled worms, they pulsed and wriggled in a slow, rhythmic rhythm. Fine veins of blood spread across the surface, and some nerve endings even twitched slightly at the contact of air. The entire picture adopts the artistic style of Junji Ito, presenting an extreme black and white high contrast. The only top light cuts the picture sharply, creating dazzling highlights on the metal surface of the scalpel and the slippery nerve tissue, while in the grooves of the scratches and the folds of the tissue, there is an impenetrable pure darkness. This extreme detail is combined with the biological grotesqueness to create a strong and disturbing atmosphere that is a mixture of surgical calmness and unspeakable physiological fear.", + "seed": 26688, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_4971_seed_26688_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.30749457764904947, + "vlm_logprob": -0.30749457764904947, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.30749457764904947, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4211254119873047 + ], + "opposite_logprob_tokens": [ + -0.2961254417896271 + ], + "logprob_reward": -1.4211254119873047, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4211254119873047, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05259021371603012, + 0.02417607232928276, + 0.10148152709007263, + 0.05144305154681206, + 0.20291678607463837, + 0.13731807470321655, + 0.16913624107837677, + 0.06538597494363785, + 0.14507952332496643, + 0.050472524017095566 + ], + "credit_map": [ + 0.00779626052826643, + 0.0054887342266738415, + 0.005029178690165281, + 0.004751123953610659, + 0.0063020456582307816, + 0.005918713286519051, + 0.00922597199678421, + 0.008078183978796005, + 0.0030301243532449007, + 0.002333184238523245, + 0.0024957561399787664, + 0.003329312428832054, + 0.002840181114152074, + 0.002620271174237132, + 0.0029260835144668818, + 0.0046011595986783504, + 0.021099260076880455, + 0.007753252517431974, + 0.0061893765814602375, + 0.007805686444044113, + 0.021720154210925102, + 0.012425899505615234, + 0.010487360879778862, + 0.0140005424618721, + 0.008111231960356236, + 0.002690304536372423, + 0.003662948729470372, + 0.004938208963721991, + 0.0054574692621827126, + 0.010440073907375336, + 0.006245812401175499, + 0.009897001087665558, + 0.039919719099998474, + 0.011004705913364887, + 0.009459389373660088, + 0.006354811601340771, + 0.09104743599891663, + 0.009352009743452072, + 0.009492660872638226, + 0.026286058127880096, + 0.013035972602665424, + 0.008311080746352673, + 0.0060746511444449425, + 0.007845165207982063, + 0.02382311411201954, + 0.011578460223972797, + 0.022565316408872604, + 0.04408431053161621, + 0.020248902961611748, + 0.009067770093679428, + 0.02311503142118454, + 0.021592678502202034, + 0.011838367208838463, + 0.016496697440743446, + 0.028826890513300896, + 0.037949901074171066, + 0.005367098841816187, + 0.003728130366653204, + 0.0051780883222818375, + 0.007364882156252861, + 0.01246962696313858, + 0.011566560715436935, + 0.01144830510020256, + 0.00826328620314598, + 0.007986994460225105, + 0.0056407698430120945, + 0.0078615453094244, + 0.011464335024356842, + 0.014595159329473972, + 0.039898503571748734, + 0.01750507764518261, + 0.04012714698910713, + 0.009163849987089634, + 0.003839432029053569, + 0.004825047682970762, + 0.009601159021258354, + 0.004394305869936943, + 0.004484669771045446, + 0.006516013760119677, + 0.007648044265806675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -1.4211254119873047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062081895768642426 + ], + "opposite_logprob_tokens": [ + -2.812081813812256 + ], + "logprob_reward": -0.062081895768642426, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062081895768642426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05031430348753929, + 0.09434671700000763, + 0.12029407918453217, + 0.02233612909913063, + 0.49566927552223206, + 0.019204743206501007, + 0.05880028381943703, + 0.06671968847513199, + 0.039504099637269974, + 0.0328107550740242 + ], + "credit_map": [ + 0.00688348151743412, + 0.0045960163697600365, + 0.003777745645493269, + 0.004360800608992577, + 0.007181703578680754, + 0.00739474780857563, + 0.006846283096820116, + 0.009273527190089226, + 0.012056268751621246, + 0.006894290912896395, + 0.0072358595207333565, + 0.007132615428417921, + 0.01017963420599699, + 0.021401697769761086, + 0.01277739554643631, + 0.016668953001499176, + 0.019248899072408676, + 0.009573803283274174, + 0.0069666369818151, + 0.00708277290686965, + 0.022441023960709572, + 0.018976831808686256, + 0.014270669780671597, + 0.021733438596129417, + 0.0030569504015147686, + 0.0014000767841935158, + 0.0010726572945713997, + 0.0014281824696809053, + 0.0020403454545885324, + 0.0034819492138922215, + 0.0032949524465948343, + 0.006561013404279947, + 0.08445150405168533, + 0.02604507841169834, + 0.02211357280611992, + 0.013858415186405182, + 0.2237265706062317, + 0.020328020676970482, + 0.021385885775089264, + 0.08376021683216095, + 0.00303682591766119, + 0.0010774850379675627, + 0.0010936047183349729, + 0.0011194230755791068, + 0.002264254493638873, + 0.001639903406612575, + 0.0032567190937697887, + 0.005716528743505478, + 0.005462798289954662, + 0.0032591300550848246, + 0.008672785945236683, + 0.011920157819986343, + 0.004260248504579067, + 0.005399005487561226, + 0.009335055015981197, + 0.010491103865206242, + 0.005165742710232735, + 0.0045690638944506645, + 0.007475757040083408, + 0.007123010233044624, + 0.011850847862660885, + 0.00814164150506258, + 0.011741193942725658, + 0.010652424767613411, + 0.002400054130703211, + 0.0022755791433155537, + 0.0029135586228221655, + 0.004182230681180954, + 0.004723049234598875, + 0.007193890865892172, + 0.005456401035189629, + 0.01035933755338192, + 0.004767369944602251, + 0.0022037094458937645, + 0.0028867926448583603, + 0.006400523707270622, + 0.0033325455151498318, + 0.0025826606433838606, + 0.005087764468044043, + 0.005549389868974686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.062081895768642426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205407351255417 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205407351255417, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205407351255417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09886620193719864, + 0.08607707172632217, + 0.14391069114208221, + 0.06934065371751785, + 0.14131170511245728, + 0.08246498554944992, + 0.05664481595158577, + 0.07431308180093765, + 0.17972101271152496, + 0.06734976172447205 + ], + "credit_map": [ + 0.012379452586174011, + 0.012506424449384212, + 0.008348113857209682, + 0.008161729201674461, + 0.012003239244222641, + 0.01165475882589817, + 0.018438614904880524, + 0.015373866073787212, + 0.011226275935769081, + 0.008807679638266563, + 0.006073046941310167, + 0.006829659454524517, + 0.012003235518932343, + 0.013794108293950558, + 0.01411385927349329, + 0.013229209929704666, + 0.023899320513010025, + 0.01275905966758728, + 0.00818658247590065, + 0.00899310689419508, + 0.02773904986679554, + 0.020994389429688454, + 0.01721828617155552, + 0.02412090264260769, + 0.010644047521054745, + 0.00813056156039238, + 0.0044672624208033085, + 0.006330558098852634, + 0.010686587542295456, + 0.008070660755038261, + 0.010892239399254322, + 0.010118737816810608, + 0.019457237794995308, + 0.008356630802154541, + 0.00791954342275858, + 0.0077604567632079124, + 0.050393667072057724, + 0.013046876527369022, + 0.01025440264493227, + 0.02412289008498192, + 0.006659400649368763, + 0.005250925198197365, + 0.004122768063098192, + 0.004892469849437475, + 0.015939118340611458, + 0.007901713252067566, + 0.013799509033560753, + 0.023899080231785774, + 0.0062627424485981464, + 0.0034439810551702976, + 0.005504620727151632, + 0.007486134301871061, + 0.0052980114705860615, + 0.007896073162555695, + 0.007984021678566933, + 0.012769232504069805, + 0.00587266031652689, + 0.004424636252224445, + 0.00765233626589179, + 0.0056193131022155285, + 0.016284584999084473, + 0.010646379552781582, + 0.014561419375240803, + 0.009251754730939865, + 0.010628324002027512, + 0.009275113232433796, + 0.011464585550129414, + 0.01151695940643549, + 0.022365281358361244, + 0.04918346926569939, + 0.019832462072372437, + 0.04545480012893677, + 0.008460137993097305, + 0.005270275287330151, + 0.004933259915560484, + 0.00936307106167078, + 0.008172630332410336, + 0.007188810035586357, + 0.009577843360602856, + 0.014383736997842789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.014205407351255417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019432486733421683 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.0019432486733421683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019432486733421683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04521678015589714, + 0.017615655437111855, + 0.04934315010905266, + 0.027319790795445442, + 0.38934776186943054, + 0.028631074354052544, + 0.05662544444203377, + 0.10673163086175919, + 0.17416571080684662, + 0.10500301420688629 + ], + "credit_map": [ + 0.006859189365059137, + 0.004859688226133585, + 0.003602131735533476, + 0.0038371565751731396, + 0.004826853051781654, + 0.005880211945623159, + 0.007878188043832779, + 0.00747336121276021, + 0.0021949107758700848, + 0.0017701592296361923, + 0.0010821869364008307, + 0.001406508614309132, + 0.0022048221435397863, + 0.0028898781165480614, + 0.0030482634902000427, + 0.0030189240351319313, + 0.009368707425892353, + 0.0032206785399466753, + 0.003148286836221814, + 0.0029357934836298227, + 0.007817467674612999, + 0.010413909330964088, + 0.0047752102836966515, + 0.007663099095225334, + 0.004462437704205513, + 0.002596818609163165, + 0.00170571927446872, + 0.0018263502279296517, + 0.003567812964320183, + 0.005815871059894562, + 0.002221103757619858, + 0.005123679526150227, + 0.06280023604631424, + 0.020144525915384293, + 0.016894152387976646, + 0.009998452849686146, + 0.17549917101860046, + 0.01569945737719536, + 0.01631077006459236, + 0.07200100272893906, + 0.003380351234227419, + 0.0026244441978633404, + 0.0018258001655340195, + 0.0026791978161782026, + 0.006555960979312658, + 0.0032265319023281336, + 0.003907995764166117, + 0.004430792760103941, + 0.0066740564070641994, + 0.0036499742418527603, + 0.006028797477483749, + 0.009638319723308086, + 0.00472611840814352, + 0.006015597376972437, + 0.00733382860198617, + 0.012558752670884132, + 0.00823093019425869, + 0.006160295102745295, + 0.006270423065871, + 0.007870865054428577, + 0.020585445687174797, + 0.013498058542609215, + 0.021823294460773468, + 0.02229231595993042, + 0.012122424319386482, + 0.006446601822972298, + 0.009377801790833473, + 0.013749021105468273, + 0.017621012404561043, + 0.050604578107595444, + 0.018543675541877747, + 0.04570058733224869, + 0.02391844429075718, + 0.009575423784554005, + 0.007504418026655912, + 0.013433732092380524, + 0.012036746367812157, + 0.00861675851047039, + 0.014713536016643047, + 0.015203958377242088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.0019432486733421683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038116924464702606 + ], + "opposite_logprob_tokens": [ + -3.288116931915283 + ], + "logprob_reward": -0.038116924464702606, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038116924464702606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05708757042884827, + 0.03261403739452362, + 0.059770017862319946, + 0.027395889163017273, + 0.2789667248725891, + 0.13964585959911346, + 0.047952696681022644, + 0.069523386657238, + 0.12784259021282196, + 0.15920118987560272 + ], + "credit_map": [ + 0.007342614233493805, + 0.005283629521727562, + 0.005364814307540655, + 0.005404975265264511, + 0.009103963151574135, + 0.008222213946282864, + 0.008288309909403324, + 0.008077051490545273, + 0.004822578746825457, + 0.0024216272868216038, + 0.003123329486697912, + 0.0034100855700671673, + 0.004302099347114563, + 0.0037355225067585707, + 0.005418212153017521, + 0.005380580201745033, + 0.006128184497356415, + 0.005924316123127937, + 0.004405592568218708, + 0.004553226754069328, + 0.008345838636159897, + 0.012072852812707424, + 0.006500997114926577, + 0.011839013546705246, + 0.003953377716243267, + 0.002338005695492029, + 0.0015005548484623432, + 0.0018315137131139636, + 0.003808174282312393, + 0.005681873299181461, + 0.0026882195379585028, + 0.005594168324023485, + 0.043255988508462906, + 0.014026525430381298, + 0.01178655680269003, + 0.007288928609341383, + 0.12457752227783203, + 0.011103013530373573, + 0.012757355347275734, + 0.054170865565538406, + 0.026086023077368736, + 0.009183652698993683, + 0.011716787703335285, + 0.014479723758995533, + 0.021306199952960014, + 0.011688842438161373, + 0.01805589348077774, + 0.02712874673306942, + 0.0070435539819300175, + 0.0030910512432456017, + 0.004625310190021992, + 0.007169927936047316, + 0.003278519492596388, + 0.006155236624181271, + 0.006857350002974272, + 0.009731743484735489, + 0.004851974546909332, + 0.003383728675544262, + 0.005337551236152649, + 0.005583948455750942, + 0.010613645426928997, + 0.008481780998408794, + 0.016819216310977936, + 0.01445153821259737, + 0.007225696928799152, + 0.005249087233096361, + 0.005030748434364796, + 0.010623130016028881, + 0.012553294189274311, + 0.038267362862825394, + 0.012982398271560669, + 0.03591086342930794, + 0.02924525924026966, + 0.014570583589375019, + 0.012873442843556404, + 0.02277921512722969, + 0.01475356612354517, + 0.01943567395210266, + 0.02035273052752018, + 0.025190720334649086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.038116924464702606 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning first-person close-up, the camera held perfectly still, immersing the viewer completely in the operator's field of vision. At the bottom of the frame, a pair of bony hands with meticulously manicured nails appeared steadily in view. In the right hand, a scalpel, gleaming with a dazzling metallic gleam, was gripped. Its sharp blade carefully and slowly sliced into the black plastic casing of an old-fashioned car CD player. The surface of the machine was covered in a crisscross of scratches, deep and shallow, and the marks of age. With steady, slow pressure from the wrist, the blade silently sliced through the casing, leaving a precise and deep incision. As the casing peeled away, piece by piece, revealing not the expected cold metal circuit boards and cables, but clumps and clumps of vibrant, moist, gray-white nerve tissue. Like countless tangled worms, they pulsed and wriggled in a slow, rhythmic rhythm. Fine veins of blood spread across the surface, and some nerve endings even twitched slightly at the contact of air. The entire picture adopts the artistic style of Junji Ito, presenting an extreme black and white high contrast. The only top light cuts the picture sharply, creating dazzling highlights on the metal surface of the scalpel and the slippery nerve tissue, while in the grooves of the scratches and the folds of the tissue, there is an impenetrable pure darkness. This extreme detail is combined with the biological grotesqueness to create a strong and disturbing atmosphere that is a mixture of surgical calmness and unspeakable physiological fear.", + "seed": 26689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_11/index_4971_seed_26689_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2513602142920718, + "vlm_logprob": -0.2513602142920718, + "vlm_yes_no_margin": 3.449999988079071, + "vlm_token_credit": -0.2513602142920718, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.170600414276123 + ], + "opposite_logprob_tokens": [ + -0.42060035467147827 + ], + "logprob_reward": -1.170600414276123, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.170600414276123, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16840586066246033, + 0.0781177207827568, + 0.04886770620942116, + 0.05303528532385826, + 0.1274939328432083, + 0.1037001684308052, + 0.10566007345914841, + 0.13749952614307404, + 0.0632062703371048, + 0.11401345580816269 + ], + "credit_map": [ + 0.018930591642856598, + 0.028451744467020035, + 0.02323029935359955, + 0.02319873496890068, + 0.021483099088072777, + 0.015505760908126831, + 0.02122620679438114, + 0.01637941412627697, + 0.01865859143435955, + 0.008594689890742302, + 0.013379504904150963, + 0.011529934592545033, + 0.005759212654083967, + 0.005552385002374649, + 0.008703973144292831, + 0.005939430091530085, + 0.005794810596853495, + 0.005912924651056528, + 0.006528139114379883, + 0.007397727575153112, + 0.006459787953644991, + 0.00452777836471796, + 0.007479467894881964, + 0.0047670695930719376, + 0.008381719700992107, + 0.01092369481921196, + 0.0081661157310009, + 0.007431297563016415, + 0.006820931099355221, + 0.0037641306407749653, + 0.004664236214011908, + 0.002883154433220625, + 0.01838139444589615, + 0.018857652321457863, + 0.01783636212348938, + 0.01691281795501709, + 0.012068862095475197, + 0.009832588024437428, + 0.015941116958856583, + 0.017663130536675453, + 0.016345305368304253, + 0.018137512728571892, + 0.01657334342598915, + 0.012380514293909073, + 0.010470779612660408, + 0.007027314510196447, + 0.009569896385073662, + 0.013195495121181011, + 0.014865233562886715, + 0.014762251637876034, + 0.031088395044207573, + 0.011982124298810959, + 0.011700354516506195, + 0.0073739467188715935, + 0.00657205143943429, + 0.007315712049603462, + 0.014568067155778408, + 0.015969466418027878, + 0.02391073666512966, + 0.014048134908080101, + 0.018867507576942444, + 0.02435673214495182, + 0.012768973596394062, + 0.013009904883801937, + 0.006859459448605776, + 0.008354280143976212, + 0.01062728837132454, + 0.005989382974803448, + 0.011873745359480381, + 0.005234377458691597, + 0.007345882244408131, + 0.006921848747879267, + 0.0180242620408535, + 0.008471815846860409, + 0.018945250660181046, + 0.01128273457288742, + 0.00816368032246828, + 0.01152659673243761, + 0.009513184428215027, + 0.028085922822356224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -1.170600414276123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208895519375801 + ], + "opposite_logprob_tokens": [ + -2.812088966369629 + ], + "logprob_reward": -0.06208895519375801, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208895519375801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2479732781648636, + 0.044061481952667236, + 0.05096308887004852, + 0.06294972449541092, + 0.09988298267126083, + 0.10587354749441147, + 0.04694288969039917, + 0.05506701394915581, + 0.17508134245872498, + 0.11120463162660599 + ], + "credit_map": [ + 0.02860919013619423, + 0.038105763494968414, + 0.03506304323673248, + 0.03780796006321907, + 0.024564359337091446, + 0.01968720369040966, + 0.04115892946720123, + 0.022976834326982498, + 0.0064058140851557255, + 0.006076545920222998, + 0.005652765743434429, + 0.004949282389134169, + 0.007196567486971617, + 0.005762067157775164, + 0.004206432495266199, + 0.0038120094686746597, + 0.010169600136578083, + 0.004654929507523775, + 0.007136141415685415, + 0.0036101487930864096, + 0.00758549477905035, + 0.005441185086965561, + 0.005758920218795538, + 0.006606667768210173, + 0.006283901631832123, + 0.012411370873451233, + 0.0079652089625597, + 0.007673760876059532, + 0.012690220028162003, + 0.006774106062948704, + 0.004544231574982405, + 0.00460692448541522, + 0.012052865698933601, + 0.014556458219885826, + 0.010876741260290146, + 0.013924147002398968, + 0.010091720148921013, + 0.008239955641329288, + 0.01524914987385273, + 0.014891947619616985, + 0.019779140129685402, + 0.013741903938353062, + 0.018726227805018425, + 0.017889922484755516, + 0.007897847332060337, + 0.007976613007485867, + 0.00917774811387062, + 0.010684143751859665, + 0.00667925551533699, + 0.00718323327600956, + 0.010452384129166603, + 0.005155665334314108, + 0.004753099754452705, + 0.0027968869544565678, + 0.005696974694728851, + 0.004225386306643486, + 0.0072224074974656105, + 0.008975251577794552, + 0.010152063332498074, + 0.006224384065717459, + 0.006235478445887566, + 0.006801710929721594, + 0.0049952557310462, + 0.004460466094315052, + 0.017642192542552948, + 0.019375042989850044, + 0.027149748057127, + 0.017967464402318, + 0.03471324220299721, + 0.01952243410050869, + 0.0204007625579834, + 0.018310464918613434, + 0.01635851338505745, + 0.009999159723520279, + 0.018089571967720985, + 0.014856853522360325, + 0.011669103056192398, + 0.00961979292333126, + 0.00990709662437439, + 0.020704541355371475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.06208895519375801 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108794566243887 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.01108794566243887, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108794566243887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19466955959796906, + 0.04850219190120697, + 0.04848051443696022, + 0.05998844653367996, + 0.07640547305345535, + 0.04677353799343109, + 0.07962234318256378, + 0.08341941982507706, + 0.156483456492424, + 0.20565509796142578 + ], + "credit_map": [ + 0.027764389291405678, + 0.027431659400463104, + 0.02539011463522911, + 0.03124454990029335, + 0.01761234924197197, + 0.015381556935608387, + 0.029783230274915695, + 0.020061716437339783, + 0.0077102030627429485, + 0.009071632288396358, + 0.006841122638434172, + 0.005429754499346018, + 0.0068510607816278934, + 0.004180144984275103, + 0.0049822041764855385, + 0.003436069469898939, + 0.0075009483844041824, + 0.00449772272258997, + 0.007132758852094412, + 0.006405468564480543, + 0.006038688588887453, + 0.004740304313600063, + 0.006848820485174656, + 0.005315801594406366, + 0.012205039151012897, + 0.008344908244907856, + 0.009846712462604046, + 0.007722059730440378, + 0.006809884216636419, + 0.004303052555769682, + 0.005512367468327284, + 0.005244421772658825, + 0.009461549110710621, + 0.010404841974377632, + 0.008543864823877811, + 0.012175669893622398, + 0.008040281012654305, + 0.0053546433337032795, + 0.012037341482937336, + 0.010387290269136429, + 0.008311624638736248, + 0.0048349350690841675, + 0.008202648721635342, + 0.008215966634452343, + 0.004164405632764101, + 0.0040663727559149265, + 0.003910701721906662, + 0.005066883284598589, + 0.011542889289557934, + 0.012793123722076416, + 0.019268352538347244, + 0.007966941222548485, + 0.008563661947846413, + 0.005041261203587055, + 0.008275577798485756, + 0.006170538254082203, + 0.011309508234262466, + 0.016006100922822952, + 0.015038106590509415, + 0.010788719169795513, + 0.007652775850147009, + 0.008310715667903423, + 0.006722071208059788, + 0.007591420318931341, + 0.015175188891589642, + 0.01584988459944725, + 0.022954775020480156, + 0.01712675392627716, + 0.031331855803728104, + 0.020145941525697708, + 0.017699725925922394, + 0.016199324280023575, + 0.026453986763954163, + 0.01669851876795292, + 0.04839647561311722, + 0.019394734874367714, + 0.0174961406737566, + 0.018166709691286087, + 0.01458964217454195, + 0.04445890709757805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01108794566243887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942058908753097 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.001942058908753097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942058908753097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19373252987861633, + 0.06795003265142441, + 0.06527942419052124, + 0.03913043439388275, + 0.07146721333265305, + 0.08062448352575302, + 0.04549547657370567, + 0.0766109824180603, + 0.14378613233566284, + 0.2159232646226883 + ], + "credit_map": [ + 0.022679593414068222, + 0.030728403478860855, + 0.02635183185338974, + 0.026094356551766396, + 0.02137354016304016, + 0.01898430846631527, + 0.029037458822131157, + 0.01848301850259304, + 0.00906967744231224, + 0.011726174503564835, + 0.009609282948076725, + 0.007825258187949657, + 0.009541243314743042, + 0.006753707770258188, + 0.006337180733680725, + 0.007087500765919685, + 0.011174752376973629, + 0.006450395565479994, + 0.009124214760959148, + 0.011035185307264328, + 0.006374710239470005, + 0.0054392460733652115, + 0.00914585031569004, + 0.0065350658260285854, + 0.005774770397692919, + 0.005425132345408201, + 0.005149872042238712, + 0.00374217564240098, + 0.006699581630527973, + 0.0038288242649286985, + 0.0038377034943550825, + 0.004672367591410875, + 0.007890827022492886, + 0.013519159518182278, + 0.007567314431071281, + 0.009853668510913849, + 0.008367492817342281, + 0.005174812860786915, + 0.009044491685926914, + 0.010049437172710896, + 0.01781476102769375, + 0.008482061326503754, + 0.019286202266812325, + 0.007149843964725733, + 0.006223304662853479, + 0.005845226347446442, + 0.0069100940600037575, + 0.008912983350455761, + 0.007374411914497614, + 0.005296255927532911, + 0.013978006318211555, + 0.004891315475106239, + 0.003753473749384284, + 0.0033565012272447348, + 0.0032120966352522373, + 0.0036334071774035692, + 0.009725074283778667, + 0.012635256163775921, + 0.014560261741280556, + 0.008440542034804821, + 0.00865191500633955, + 0.00953372847288847, + 0.005445472430437803, + 0.007618727162480354, + 0.01157466135919094, + 0.014396517537534237, + 0.021601779386401176, + 0.016692189499735832, + 0.026228491216897964, + 0.021436378359794617, + 0.014400631189346313, + 0.017455466091632843, + 0.030659111216664314, + 0.017929287627339363, + 0.04982423037290573, + 0.01898815482854843, + 0.018639808520674706, + 0.01849249191582203, + 0.016248969361186028, + 0.04514118283987045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.001942058908753097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective with the camera held perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands with meticulously manicured nails visible at the bottom of the frame holding a scalpel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scalpel slicing slowly into a black plastic casing of an old-fashioned car CD player covered in scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the casing reveal vibrant, moist, gray-white nerve tissue that pulses and wriggles instead of metal circuit boards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered in an extreme black and white high contrast style resembling the artistic style of Junji Ito?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081697419285774 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081697419285774, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081697419285774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17857736349105835, + 0.05285276100039482, + 0.058389149606227875, + 0.026548970490694046, + 0.04729831963777542, + 0.08947378396987915, + 0.032301440834999084, + 0.04604203253984451, + 0.18372701108455658, + 0.28478917479515076 + ], + "credit_map": [ + 0.018634909763932228, + 0.02920619212090969, + 0.022933144122362137, + 0.024997597560286522, + 0.01843639276921749, + 0.015887180343270302, + 0.027669623494148254, + 0.02081233821809292, + 0.006613482255488634, + 0.007064612582325935, + 0.006312173791229725, + 0.006413523573428392, + 0.008268116973340511, + 0.006182699464261532, + 0.006571135949343443, + 0.005427017342299223, + 0.009732521139085293, + 0.005933321546763182, + 0.007984652183949947, + 0.00934694241732359, + 0.005731198005378246, + 0.004934539552778006, + 0.007575238589197397, + 0.007150733843445778, + 0.004072506446391344, + 0.002496611326932907, + 0.004631119780242443, + 0.004494090098887682, + 0.002162744989618659, + 0.0017064216081053019, + 0.0040065268985927105, + 0.0029789486434310675, + 0.005442305002361536, + 0.007225559093058109, + 0.0059400456957519054, + 0.0074274600483477116, + 0.005336106289178133, + 0.003348731901496649, + 0.006589051336050034, + 0.005989063065499067, + 0.01820298284292221, + 0.009431888349354267, + 0.01754925027489662, + 0.01175959873944521, + 0.006689307279884815, + 0.006916423793882132, + 0.008771265856921673, + 0.010153072886168957, + 0.004725131206214428, + 0.004817407112568617, + 0.007367612794041634, + 0.0036424167919903994, + 0.0027921246364712715, + 0.0032809781841933727, + 0.0029764221981167793, + 0.002699346514418721, + 0.005368825513869524, + 0.005801908206194639, + 0.007214764133095741, + 0.005484134424477816, + 0.005931421648710966, + 0.005918446462601423, + 0.005346617661416531, + 0.004975910298526287, + 0.020801497623324394, + 0.021508537232875824, + 0.037671852856874466, + 0.021107973530888557, + 0.028138667345046997, + 0.01682930439710617, + 0.019307967275381088, + 0.018361201509833336, + 0.060174643993377686, + 0.02099861577153206, + 0.05826480686664581, + 0.018384957686066628, + 0.022814204916357994, + 0.02330717071890831, + 0.023608317598700523, + 0.05723647028207779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.011081697419285774 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_tensor([5367, 5367])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_tensor([5367, 5367])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d9d126c2ddb203483dedeb3727f136957b82875a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_tensor([5367, 5367])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The high-contrast macro shot begins, the camera tracking uncomfortably slowly from left to right, maintaining a tight profile. The subject is a dead black ant, its black exoskeleton rendered with exquisite clarity. Its jaws make a final, desperate grip on the vibrant green edge of a waxy leaf, its stiff legs curled tightly around its dark, segmented body. As the camera continues its deliberate crawl, it passes over the ant's thorax and peaks at its head, from which a sinister fungal stalk erupts vertically. The eerie, pale orange plant is fibrous and strangely textured, its surface covered in a layer of glistening, globular morning dew that catches the sharp, directed light like tiny jewels. The intense, high-contrast lighting casts deep, dramatic shadows, sculpting every gruesome detail of the stem's texture and the ant's lifeless form, creating an eerie atmosphere that is both clinically beautiful and deeply unsettling against the backdrop, which dissolves into a soft, out-of-focus blur of dark green and earthy brown.", + "seed": 7434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_5367_seed_7434_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03725875774398446, + "vlm_logprob": -0.03725875774398446, + "vlm_yes_no_margin": 3.724999952316284, + "vlm_token_credit": -0.03725875774398446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08067944645881653 + ], + "opposite_logprob_tokens": [ + -2.580679416656494 + ], + "logprob_reward": -0.08067944645881653, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08067944645881653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05396266281604767, + 0.021273108199238777, + 0.1502905935049057, + 0.11783929169178009, + 0.272662878036499, + 0.02879137173295021, + 0.01637791283428669, + 0.03716051205992699, + 0.1799442172050476, + 0.12169747054576874 + ], + "credit_map": [ + 0.007556611206382513, + 0.005658864974975586, + 0.005810630042105913, + 0.005909897852689028, + 0.006762155797332525, + 0.008114810101687908, + 0.00942474976181984, + 0.004724941682070494, + 0.0024754025507718325, + 0.0037759854458272457, + 0.0019398992881178856, + 0.0020100914407521486, + 0.0030038829427212477, + 0.0031631619203835726, + 0.002616592450067401, + 0.002288091229274869, + 0.013810615986585617, + 0.02904665097594261, + 0.010431349277496338, + 0.019647756591439247, + 0.02895340323448181, + 0.02109261229634285, + 0.016225088387727737, + 0.011083117686212063, + 0.03351079300045967, + 0.012057443149387836, + 0.006804892793297768, + 0.006950351409614086, + 0.014187535271048546, + 0.015751786530017853, + 0.020043831318616867, + 0.008532664738595486, + 0.01832849532365799, + 0.07641379535198212, + 0.00897421408444643, + 0.006424727849662304, + 0.10965679585933685, + 0.032445523887872696, + 0.01092622708529234, + 0.009493094868957996, + 0.0026084769051522017, + 0.0033294258173555136, + 0.0020117524545639753, + 0.002952579641714692, + 0.00721570011228323, + 0.003879101015627384, + 0.003396518062800169, + 0.0033978158608078957, + 0.0018900401191785932, + 0.0020964224822819233, + 0.0018732757307589054, + 0.003047296544536948, + 0.0018456964753568172, + 0.0021982896141707897, + 0.0016993340104818344, + 0.0017275590216740966, + 0.005494648590683937, + 0.0028756705578416586, + 0.005185029469430447, + 0.004738890565931797, + 0.008399567566812038, + 0.0048148054629564285, + 0.003092866623774171, + 0.0025590346194803715, + 0.04944019019603729, + 0.016231369227170944, + 0.007692402228713036, + 0.011279502883553505, + 0.018526414409279823, + 0.051039211452007294, + 0.014794283546507359, + 0.010940859094262123, + 0.016400853171944618, + 0.015375649556517601, + 0.01415723841637373, + 0.016495464369654655, + 0.014690608717501163, + 0.015098928473889828, + 0.018810119479894638, + 0.010668598115444183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.08067944645881653 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209662288427353 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.018209662288427353, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018209662288427353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05641613528132439, + 0.01858798786997795, + 0.1002664789557457, + 0.06238460913300514, + 0.15624931454658508, + 0.10307851433753967, + 0.024791374802589417, + 0.1334839016199112, + 0.2637035548686981, + 0.08103813230991364 + ], + "credit_map": [ + 0.006376967299729586, + 0.004415578208863735, + 0.004556239116936922, + 0.007819710299372673, + 0.010486053302884102, + 0.009010640904307365, + 0.0076686604879796505, + 0.006082287523895502, + 0.001455274410545826, + 0.003174412529915571, + 0.001885782228782773, + 0.00185667653568089, + 0.0026116331573575735, + 0.002554075326770544, + 0.002695173490792513, + 0.002354961121454835, + 0.010253611020743847, + 0.01480159256607294, + 0.008323603309690952, + 0.014263355173170567, + 0.02062034234404564, + 0.01430946309119463, + 0.010739235207438469, + 0.006955282762646675, + 0.009591566398739815, + 0.007166152819991112, + 0.003940993454307318, + 0.00460128765553236, + 0.011777359992265701, + 0.009640904143452644, + 0.010841994546353817, + 0.004824348259717226, + 0.021062538027763367, + 0.024756334722042084, + 0.009568399749696255, + 0.008917237631976604, + 0.03282087668776512, + 0.03166145831346512, + 0.011305824853479862, + 0.016156641766428947, + 0.008174720220267773, + 0.010461730882525444, + 0.007999081164598465, + 0.009775656275451183, + 0.02682465873658657, + 0.01636149361729622, + 0.011967787519097328, + 0.011513391509652138, + 0.00316072185523808, + 0.0021283829119056463, + 0.004282286390662193, + 0.0035065675619989634, + 0.0027147354558110237, + 0.003551878733560443, + 0.002895812038332224, + 0.002550989855080843, + 0.017834464088082314, + 0.008489278145134449, + 0.014937344007194042, + 0.022841308265924454, + 0.03448951244354248, + 0.019327715039253235, + 0.010065894573926926, + 0.005498382728546858, + 0.0779329165816307, + 0.01969374157488346, + 0.012177047319710255, + 0.013637089170515537, + 0.027677835896611214, + 0.07928543537855148, + 0.015515144914388657, + 0.017784345895051956, + 0.012493696995079517, + 0.00944685097783804, + 0.009548937901854515, + 0.012636666186153889, + 0.008529937826097012, + 0.009495673701167107, + 0.010309550911188126, + 0.008576816879212856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.018209662288427353 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626228500157595 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007626228500157595, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626228500157595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043704818934202194, + 0.03791561350226402, + 0.08219590038061142, + 0.05880703777074814, + 0.3172937333583832, + 0.020172186195850372, + 0.024435507133603096, + 0.07136169075965881, + 0.2154187709093094, + 0.12869471311569214 + ], + "credit_map": [ + 0.0068430243991315365, + 0.005689883138984442, + 0.004131526220589876, + 0.004821179434657097, + 0.005999667104333639, + 0.006943272426724434, + 0.005998184438794851, + 0.0032780857291072607, + 0.004849578253924847, + 0.007187844719737768, + 0.003152349963784218, + 0.003616661997511983, + 0.005796852987259626, + 0.005875483620911837, + 0.004064618609845638, + 0.0033722249791026115, + 0.00827071163803339, + 0.014186657033860683, + 0.007178021594882011, + 0.011492710560560226, + 0.010512467473745346, + 0.01209699735045433, + 0.008780683390796185, + 0.009677647612988949, + 0.015541257336735725, + 0.007079137954860926, + 0.004000456538051367, + 0.0037617164198309183, + 0.007740098983049393, + 0.00668077589944005, + 0.009918888099491596, + 0.004084705375134945, + 0.021708190441131592, + 0.09012266993522644, + 0.011830446310341358, + 0.009235012345016003, + 0.12669603526592255, + 0.032947931438684464, + 0.014064664952456951, + 0.010688787326216698, + 0.0018846201710402966, + 0.002870506839826703, + 0.0022510020062327385, + 0.0027909018099308014, + 0.00319284712895751, + 0.002766106743365526, + 0.0018089639488607645, + 0.002607238944619894, + 0.002939395373687148, + 0.002510985592380166, + 0.00497809611260891, + 0.0030699647031724453, + 0.0024935409892350435, + 0.00360863353125751, + 0.002804742194712162, + 0.0020301472395658493, + 0.010344256646931171, + 0.005873727612197399, + 0.010240737348794937, + 0.00953458622097969, + 0.015438657253980637, + 0.010088595561683178, + 0.006017749663442373, + 0.0038233862724155188, + 0.06523197889328003, + 0.012193122878670692, + 0.008894486352801323, + 0.010519319213926792, + 0.02352495677769184, + 0.06730430573225021, + 0.01490433793514967, + 0.012846251018345356, + 0.020083965733647346, + 0.015688393265008926, + 0.0122830206528306, + 0.01685318909585476, + 0.016579627990722656, + 0.017617160454392433, + 0.016178689897060394, + 0.013410673476755619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.007626228500157595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05829935893416405, + 0.018851738423109055, + 0.0940278023481369, + 0.013170886784791946, + 0.33160826563835144, + 0.02363080531358719, + 0.032586924731731415, + 0.07604260742664337, + 0.24455437064170837, + 0.10722718387842178 + ], + "credit_map": [ + 0.007067778147757053, + 0.006731236353516579, + 0.004380021244287491, + 0.006471268367022276, + 0.011113572865724564, + 0.010184329003095627, + 0.006152183283120394, + 0.006198971066623926, + 0.0023558223620057106, + 0.004151294939219952, + 0.0019194617634639144, + 0.0023134334478527308, + 0.0021237600594758987, + 0.0030744816176593304, + 0.0016803550533950329, + 0.0012331283651292324, + 0.01150942500680685, + 0.014194702729582787, + 0.006887514144182205, + 0.011343435384333134, + 0.014037278480827808, + 0.018067345023155212, + 0.009128976613283157, + 0.008859124034643173, + 0.001639154739677906, + 0.0015009413473308086, + 0.0021147625520825386, + 0.0007799159502610564, + 0.002445781836286187, + 0.002050400711596012, + 0.0013879492180421948, + 0.0012519798474386334, + 0.018441202118992805, + 0.13488131761550903, + 0.007646474987268448, + 0.009081944823265076, + 0.1276303082704544, + 0.016621183604002, + 0.009649968706071377, + 0.007655860390514135, + 0.0018511372618377209, + 0.0036822629626840353, + 0.0023274035193026066, + 0.002654950600117445, + 0.0034818982239812613, + 0.005537099204957485, + 0.001999043859541416, + 0.0020970089826732874, + 0.003367400262504816, + 0.004090520087629557, + 0.005332662723958492, + 0.006120501086115837, + 0.0025758780539035797, + 0.0040580532513558865, + 0.0037769151385873556, + 0.0032649931963533163, + 0.010147565975785255, + 0.005886796861886978, + 0.010186483152210712, + 0.01157449558377266, + 0.017912836745381355, + 0.010681702755391598, + 0.005984213203191757, + 0.0036685140803456306, + 0.0737152099609375, + 0.014244304038584232, + 0.01030141580849886, + 0.012259300798177719, + 0.026434611529111862, + 0.07601945102214813, + 0.016864988952875137, + 0.014715097844600677, + 0.014942100271582603, + 0.014986088499426842, + 0.010954434983432293, + 0.017410654574632645, + 0.014331253245472908, + 0.010025030933320522, + 0.015177100896835327, + 0.009400524199008942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.009779755026102066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999869644641876 + ], + "opposite_logprob_tokens": [ + -2.6949987411499023 + ], + "logprob_reward": -0.06999869644641876, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999869644641876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026469510048627853, + 0.022196857258677483, + 0.052092134952545166, + 0.02147245593369007, + 0.7588337063789368, + 0.046751245856285095, + 0.011969906277954578, + 0.01847449317574501, + 0.025344740599393845, + 0.0163949653506279 + ], + "credit_map": [ + 0.00410199211910367, + 0.004277269821614027, + 0.002252675360068679, + 0.002861062064766884, + 0.004838688299059868, + 0.003983301110565662, + 0.002306169830262661, + 0.0018483499297872186, + 0.002475825371220708, + 0.005109185352921486, + 0.001730453222990036, + 0.0029005983378738165, + 0.002275422913953662, + 0.004088160581886768, + 0.0019989246502518654, + 0.0016182841500267386, + 0.006419106852263212, + 0.009587491862475872, + 0.003977817948907614, + 0.004963106941431761, + 0.009774954989552498, + 0.009068076498806477, + 0.004237892106175423, + 0.004063683561980724, + 0.006384498905390501, + 0.0024785182904452085, + 0.0012936050770804286, + 0.0011353259906172752, + 0.002476633759215474, + 0.002856443403288722, + 0.003247442888095975, + 0.0015999855240806937, + 0.04254314303398132, + 0.27283021807670593, + 0.01842697151005268, + 0.018806086853146553, + 0.31144171953201294, + 0.05181087926030159, + 0.02552858740091324, + 0.01744605228304863, + 0.002758399583399296, + 0.00470318365842104, + 0.0036766277626156807, + 0.004514876287430525, + 0.01142027322202921, + 0.007730417884886265, + 0.005755330435931683, + 0.0061921305023133755, + 0.0011151350336149335, + 0.0014349413104355335, + 0.0018049563514068723, + 0.0026306905783712864, + 0.0012851088540628552, + 0.0014393384335562587, + 0.0012978024315088987, + 0.0009619328775443137, + 0.0037800567224621773, + 0.0009013327653519809, + 0.0015754111809656024, + 0.0013401326723396778, + 0.005876234732568264, + 0.0028911586850881577, + 0.0010359674924984574, + 0.0010741985170170665, + 0.00488536711782217, + 0.0025151490699499846, + 0.0024414437357336283, + 0.002754686400294304, + 0.0031236521899700165, + 0.004473837558180094, + 0.0036414607893675566, + 0.0015091426903381944, + 0.0030935341492295265, + 0.0017084268620237708, + 0.0015309829032048583, + 0.0022565368562936783, + 0.0022762054577469826, + 0.002127778483554721, + 0.0017657957505434752, + 0.0016357051208615303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.06999869644641876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The high-contrast macro shot begins, the camera tracking uncomfortably slowly from left to right, maintaining a tight profile. The subject is a dead black ant, its black exoskeleton rendered with exquisite clarity. Its jaws make a final, desperate grip on the vibrant green edge of a waxy leaf, its stiff legs curled tightly around its dark, segmented body. As the camera continues its deliberate crawl, it passes over the ant's thorax and peaks at its head, from which a sinister fungal stalk erupts vertically. The eerie, pale orange plant is fibrous and strangely textured, its surface covered in a layer of glistening, globular morning dew that catches the sharp, directed light like tiny jewels. The intense, high-contrast lighting casts deep, dramatic shadows, sculpting every gruesome detail of the stem's texture and the ant's lifeless form, creating an eerie atmosphere that is both clinically beautiful and deeply unsettling against the backdrop, which dissolves into a soft, out-of-focus blur of dark green and earthy brown.", + "seed": 7435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_5367_seed_7435_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026448822766542434, + "vlm_logprob": -0.026448822766542434, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.026448822766542434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03592231497168541 + ], + "opposite_logprob_tokens": [ + -3.4109222888946533 + ], + "logprob_reward": -0.03592231497168541, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03592231497168541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052118562161922455, + 0.016949784010648727, + 0.0232841819524765, + 0.05193180590867996, + 0.035684119910001755, + 0.28653499484062195, + 0.04842135310173035, + 0.015606543980538845, + 0.10605821013450623, + 0.36341044306755066 + ], + "credit_map": [ + 0.004291038494557142, + 0.0066562131978571415, + 0.007415789179503918, + 0.008794410154223442, + 0.007190318312495947, + 0.006941599305719137, + 0.006692823953926563, + 0.004136370960623026, + 0.0019405712373554707, + 0.002487336052581668, + 0.0025984139647334814, + 0.001856703427620232, + 0.002422942314296961, + 0.0018573745619505644, + 0.00213914574123919, + 0.00164729671087116, + 0.003257663222029805, + 0.0018219857010990381, + 0.0031281234696507454, + 0.0052509065717458725, + 0.002791324397549033, + 0.0023983544670045376, + 0.0028474177233874798, + 0.0017884059343487024, + 0.003594141686335206, + 0.003290328662842512, + 0.004622738342732191, + 0.0031147615518420935, + 0.004536957014352083, + 0.004070895258337259, + 0.01820223778486252, + 0.01049974374473095, + 0.002526918426156044, + 0.012120923958718777, + 0.0014832895249128342, + 0.001422553090378642, + 0.0023279048036783934, + 0.006946402136236429, + 0.002261882182210684, + 0.006594245322048664, + 0.016326000913977623, + 0.06598080694675446, + 0.022173192352056503, + 0.024457845836877823, + 0.10582520812749863, + 0.01896083727478981, + 0.016216089949011803, + 0.01659501902759075, + 0.007142988499253988, + 0.007697346620261669, + 0.005610105115920305, + 0.004164160694926977, + 0.008522218093276024, + 0.00842920783907175, + 0.0029482899699360132, + 0.003907035570591688, + 0.002298397710546851, + 0.0013249734183773398, + 0.0017548722680658102, + 0.0012375738006085157, + 0.002868474228307605, + 0.0028875621501356363, + 0.002355567179620266, + 0.0008791226428002119, + 0.018682798370718956, + 0.020566297695040703, + 0.008881684392690659, + 0.01974276453256607, + 0.013291272334754467, + 0.008651325479149818, + 0.008765860460698605, + 0.00747620640322566, + 0.06268192827701569, + 0.03947339579463005, + 0.06367774307727814, + 0.032897140830755234, + 0.03354330733418465, + 0.03943588212132454, + 0.03765999898314476, + 0.05404103174805641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.03592231497168541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983800880610943 + ], + "opposite_logprob_tokens": [ + -3.5298380851745605 + ], + "logprob_reward": -0.02983800880610943, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983800880610943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05447778478264809, + 0.07178016006946564, + 0.048095837235450745, + 0.047910742461681366, + 0.20859071612358093, + 0.060361750423908234, + 0.04654589667916298, + 0.03530295193195343, + 0.21514582633972168, + 0.2117883861064911 + ], + "credit_map": [ + 0.004528828430920839, + 0.005757597275078297, + 0.007176679093390703, + 0.007223189342767, + 0.006122392602264881, + 0.009168942458927631, + 0.008912218734622002, + 0.005587934050709009, + 0.007820434868335724, + 0.023312987759709358, + 0.006125179119408131, + 0.00797312892973423, + 0.009227454662322998, + 0.005734396167099476, + 0.005823536776006222, + 0.005763040855526924, + 0.003643697826191783, + 0.004002640023827553, + 0.006461364682763815, + 0.00930836983025074, + 0.005528127308934927, + 0.005699438508599997, + 0.007396034896373749, + 0.006056163460016251, + 0.005454080644994974, + 0.006550312042236328, + 0.005305400118231773, + 0.003468472044914961, + 0.005644715391099453, + 0.004615932237356901, + 0.009702460840344429, + 0.007169369142502546, + 0.013693335466086864, + 0.0707891434431076, + 0.008273953571915627, + 0.009103312157094479, + 0.014994209632277489, + 0.03969086334109306, + 0.012161118909716606, + 0.0398847796022892, + 0.004397199023514986, + 0.00902312621474266, + 0.004734373185783625, + 0.012550520710647106, + 0.01543815154582262, + 0.004015997517853975, + 0.00387818762101233, + 0.006324198562651873, + 0.006789706647396088, + 0.005008337553590536, + 0.0045895944349467754, + 0.0061577968299388885, + 0.007544552441686392, + 0.009537257254123688, + 0.003179179737344384, + 0.0037394738756120205, + 0.005157701205462217, + 0.004553377628326416, + 0.005102319177240133, + 0.0032188119366765022, + 0.004764148034155369, + 0.004676396958529949, + 0.005091703962534666, + 0.002738493261858821, + 0.03829178214073181, + 0.03869974985718727, + 0.020092438906431198, + 0.04283977672457695, + 0.026265345513820648, + 0.015148931182920933, + 0.016512643545866013, + 0.017295166850090027, + 0.03744408115744591, + 0.027619322761893272, + 0.03407767042517662, + 0.020532213151454926, + 0.022879499942064285, + 0.021141381934285164, + 0.02070542611181736, + 0.027388788759708405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.02983800880610943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251897498965263 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251897498965263, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251897498965263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057255424559116364, + 0.0226332638412714, + 0.04750968888401985, + 0.07022317498922348, + 0.11836572736501694, + 0.027171144261956215, + 0.0820743590593338, + 0.011484868824481964, + 0.06553010642528534, + 0.49775224924087524 + ], + "credit_map": [ + 0.004946104716509581, + 0.006718132644891739, + 0.008457540534436703, + 0.012396235950291157, + 0.008465313352644444, + 0.005834092386066914, + 0.006238540168851614, + 0.004199464805424213, + 0.001885000616312027, + 0.004925298038870096, + 0.002860401989892125, + 0.0036310823634266853, + 0.002812287537381053, + 0.002249073004350066, + 0.002335221040993929, + 0.0019348993664607406, + 0.004209463484585285, + 0.004458926152437925, + 0.007871047593653202, + 0.006407595705240965, + 0.005997099447995424, + 0.0067722853273153305, + 0.004670111462473869, + 0.007123163435608149, + 0.006258158944547176, + 0.012450932525098324, + 0.006041958462446928, + 0.004832427017390728, + 0.00668635219335556, + 0.006345394998788834, + 0.012058229185640812, + 0.015549720264971256, + 0.008787579834461212, + 0.024536415934562683, + 0.010039200074970722, + 0.007324507925659418, + 0.023044947534799576, + 0.0178226251155138, + 0.010609647259116173, + 0.016200803220272064, + 0.002263034228235483, + 0.00854178424924612, + 0.002171941101551056, + 0.0026995702646672726, + 0.006057595834136009, + 0.0022064067889004946, + 0.0018485031323507428, + 0.0013823084300383925, + 0.010929432697594166, + 0.011775245890021324, + 0.0063500008545815945, + 0.008103552274405956, + 0.02062586322426796, + 0.012529207393527031, + 0.005284937564283609, + 0.006476124282926321, + 0.0015795323997735977, + 0.0011955255176872015, + 0.0017595430836081505, + 0.0012768525630235672, + 0.0016922056674957275, + 0.0014711454277858138, + 0.001768971444107592, + 0.0007410921389237046, + 0.010553506202995777, + 0.013907980173826218, + 0.005622428376227617, + 0.007817468605935574, + 0.01191349234431982, + 0.00674879364669323, + 0.004708243068307638, + 0.004258195403963327, + 0.08735573291778564, + 0.053039081394672394, + 0.08662063628435135, + 0.0459778793156147, + 0.04886611923575401, + 0.04976135864853859, + 0.055099040269851685, + 0.07103241235017776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.005251897498965263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183909356594086 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183909356594086, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018183909356594086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0261667612940073, + 0.01569414511322975, + 0.03492666780948639, + 0.018741469830274582, + 0.0846734344959259, + 0.01978987082839012, + 0.01645010896027088, + 0.006545811891555786, + 0.03183550015091896, + 0.7451761960983276 + ], + "credit_map": [ + 0.002492512110620737, + 0.0027164698112756014, + 0.0036798380315303802, + 0.003736987477168441, + 0.0034277010709047318, + 0.00396033376455307, + 0.0033650740515440702, + 0.0027878452092409134, + 0.0016377391293644905, + 0.004728606902062893, + 0.0017716005677357316, + 0.001563241588883102, + 0.0019726466853171587, + 0.0014679711312055588, + 0.0015257903141900897, + 0.001026549725793302, + 0.006443425081670284, + 0.0029680475126951933, + 0.006310335826128721, + 0.004503617528825998, + 0.0031261048279702663, + 0.0037681548856198788, + 0.0030182027257978916, + 0.004788780119270086, + 0.002531277947127819, + 0.003042333060875535, + 0.0027188207022845745, + 0.001658270019106567, + 0.0016244201688095927, + 0.0019073367584496737, + 0.0017714118584990501, + 0.0034875976853072643, + 0.005549794528633356, + 0.02700074203312397, + 0.004971574991941452, + 0.003693057456985116, + 0.006501417141407728, + 0.015264004468917847, + 0.00545780174434185, + 0.016235044226050377, + 0.0017264083726331592, + 0.005541983060538769, + 0.0017449745209887624, + 0.0012898067943751812, + 0.005863412283360958, + 0.001428001094609499, + 0.0011290655238553882, + 0.0010662191780284047, + 0.0022473614662885666, + 0.0020314068533480167, + 0.0012835059314966202, + 0.0017950317123904824, + 0.004488622769713402, + 0.002017377410084009, + 0.0014368107076734304, + 0.0011499939719215035, + 0.0011645016493275762, + 0.0007155959610827267, + 0.0010652999626472592, + 0.0007577127544209361, + 0.0007751649827696383, + 0.0006895502447150648, + 0.0008912396151572466, + 0.0004867468378506601, + 0.005540874786674976, + 0.007528271991759539, + 0.0029932584147900343, + 0.007420710753649473, + 0.0024230987764894962, + 0.0016557940980419517, + 0.002272882964462042, + 0.002000608015805483, + 0.126268208026886, + 0.07746957987546921, + 0.13259071111679077, + 0.06537485122680664, + 0.07138892263174057, + 0.08056645095348358, + 0.07663893699645996, + 0.11487851291894913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.018183909356594086 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043047983199357986 + ], + "opposite_logprob_tokens": [ + -3.1680479049682617 + ], + "logprob_reward": -0.043047983199357986, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043047983199357986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1825946420431137, + 0.06349876523017883, + 0.13032588362693787, + 0.05526018887758255, + 0.10343749821186066, + 0.08747872710227966, + 0.1337163895368576, + 0.035097237676382065, + 0.09544982761144638, + 0.11314085870981216 + ], + "credit_map": [ + 0.017485061660408974, + 0.024464331567287445, + 0.024017170071601868, + 0.020476365461945534, + 0.021256843581795692, + 0.026959847658872604, + 0.02983219549059868, + 0.018102839589118958, + 0.005664427299052477, + 0.01639571040868759, + 0.006774096749722958, + 0.008752184920012951, + 0.008683959022164345, + 0.005094528663903475, + 0.00683183828368783, + 0.005302025470882654, + 0.02206406183540821, + 0.011213058605790138, + 0.01978207379579544, + 0.024924153462052345, + 0.010066080838441849, + 0.012850549072027206, + 0.017153622582554817, + 0.012272278778254986, + 0.0064758858643472195, + 0.005587111692875624, + 0.007941960357129574, + 0.004101581405848265, + 0.00911625288426876, + 0.005394298117607832, + 0.005366010125726461, + 0.011277088895440102, + 0.0065355985425412655, + 0.03188982233405113, + 0.006390597205609083, + 0.004904300905764103, + 0.009367156773805618, + 0.01862006075680256, + 0.006003242451697588, + 0.01972673088312149, + 0.0050882138311862946, + 0.013789288699626923, + 0.007225971668958664, + 0.01040943618863821, + 0.032494205981492996, + 0.0054274871945381165, + 0.006970149464905262, + 0.006073975469917059, + 0.019159680232405663, + 0.017050720751285553, + 0.02042747102677822, + 0.017264055088162422, + 0.014852769672870636, + 0.027516886591911316, + 0.008279106579720974, + 0.009165704250335693, + 0.005746818613260984, + 0.0032029764261096716, + 0.004707204643636942, + 0.0024651032872498035, + 0.007578634191304445, + 0.005658928770571947, + 0.003856887109577656, + 0.001880687428638339, + 0.018031656742095947, + 0.011698886752128601, + 0.007358735892921686, + 0.013391906395554543, + 0.014479024335741997, + 0.011813797056674957, + 0.0106803635135293, + 0.00799546018242836, + 0.019098874181509018, + 0.012855738401412964, + 0.01736719161272049, + 0.01101458165794611, + 0.009957296773791313, + 0.013743313029408455, + 0.011592413298785686, + 0.017511451616883278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.043047983199357986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The high-contrast macro shot begins, the camera tracking uncomfortably slowly from left to right, maintaining a tight profile. The subject is a dead black ant, its black exoskeleton rendered with exquisite clarity. Its jaws make a final, desperate grip on the vibrant green edge of a waxy leaf, its stiff legs curled tightly around its dark, segmented body. As the camera continues its deliberate crawl, it passes over the ant's thorax and peaks at its head, from which a sinister fungal stalk erupts vertically. The eerie, pale orange plant is fibrous and strangely textured, its surface covered in a layer of glistening, globular morning dew that catches the sharp, directed light like tiny jewels. The intense, high-contrast lighting casts deep, dramatic shadows, sculpting every gruesome detail of the stem's texture and the ant's lifeless form, creating an eerie atmosphere that is both clinically beautiful and deeply unsettling against the backdrop, which dissolves into a soft, out-of-focus blur of dark green and earthy brown.", + "seed": 7436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_5367_seed_7436_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09655256252735853, + "vlm_logprob": -0.09655256252735853, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.09655256252735853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38892555236816406 + ], + "opposite_logprob_tokens": [ + -1.138925552368164 + ], + "logprob_reward": -0.38892555236816406, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38892555236816406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16730394959449768, + 0.06856834143400192, + 0.057529352605342865, + 0.34513822197914124, + 0.12298593670129776, + 0.02760176733136177, + 0.04114814102649689, + 0.05214804410934448, + 0.06123529374599457, + 0.05634097382426262 + ], + "credit_map": [ + 0.013475187122821808, + 0.012141004204750061, + 0.021930303424596786, + 0.03927437216043472, + 0.011729518882930279, + 0.022302139550447464, + 0.027574127539992332, + 0.018877290189266205, + 0.0051981317810714245, + 0.00520796375349164, + 0.009352445602416992, + 0.01288533117622137, + 0.00918307900428772, + 0.009661505930125713, + 0.01085632760077715, + 0.0062235621735453606, + 0.004901152569800615, + 0.004240692127496004, + 0.0074933553114533424, + 0.012756467796862125, + 0.0042965104803442955, + 0.009251989424228668, + 0.007202909328043461, + 0.007386272773146629, + 0.024081895127892494, + 0.025316376239061356, + 0.06774194538593292, + 0.05135486274957657, + 0.03146137297153473, + 0.05274403467774391, + 0.046623874455690384, + 0.04581384360790253, + 0.01282363012433052, + 0.013638078235089779, + 0.012623701244592667, + 0.015857292339205742, + 0.015947481617331505, + 0.018120620399713516, + 0.01982743665575981, + 0.014147705398499966, + 0.002664222614839673, + 0.0025832848623394966, + 0.0046676816418766975, + 0.004502490162849426, + 0.002553887665271759, + 0.003668947611004114, + 0.0038586570881307125, + 0.003102595452219248, + 0.004224620293825865, + 0.0067969379015266895, + 0.0037334063090384007, + 0.005142007488757372, + 0.006030323915183544, + 0.006347537040710449, + 0.003924784250557423, + 0.004948524758219719, + 0.004900835454463959, + 0.010382656008005142, + 0.0048594847321510315, + 0.00790382456034422, + 0.0066347322426736355, + 0.007785362657159567, + 0.005137999076396227, + 0.004543147515505552, + 0.010524454526603222, + 0.010546913370490074, + 0.004465282429009676, + 0.007338056806474924, + 0.00698216212913394, + 0.007945970632135868, + 0.00646858150139451, + 0.006963876076042652, + 0.006193163804709911, + 0.010791689157485962, + 0.004664980340749025, + 0.007087311241775751, + 0.0090158861130476, + 0.007755706086754799, + 0.005771527532488108, + 0.005060707218945026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.38892555236816406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060963399708271 + ], + "opposite_logprob_tokens": [ + -3.8956096172332764 + ], + "logprob_reward": -0.02060963399708271, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060963399708271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08101421594619751, + 0.0815393254160881, + 0.05608063563704491, + 0.22119581699371338, + 0.10593457520008087, + 0.07217099517583847, + 0.09738969057798386, + 0.15723031759262085, + 0.07218486815690994, + 0.05525960400700569 + ], + "credit_map": [ + 0.008263468742370605, + 0.0062501514330506325, + 0.008426367305219173, + 0.015861036255955696, + 0.007718199398368597, + 0.011296667158603668, + 0.012012925930321217, + 0.01118539273738861, + 0.006249870173633099, + 0.005419572349637747, + 0.010421082377433777, + 0.013872424140572548, + 0.010776505805552006, + 0.011697845533490181, + 0.012205710634589195, + 0.010896299965679646, + 0.0055628190748393536, + 0.0034406159538775682, + 0.006067827809602022, + 0.008145724423229694, + 0.004070192109793425, + 0.009367674589157104, + 0.0075384266674518585, + 0.01188734918832779, + 0.016221903264522552, + 0.016434114426374435, + 0.04206596314907074, + 0.0320361852645874, + 0.02135395258665085, + 0.03458515927195549, + 0.030600808560848236, + 0.027897708117961884, + 0.009915439411997795, + 0.009067896753549576, + 0.010225697420537472, + 0.010413120500743389, + 0.018176255747675896, + 0.01755133643746376, + 0.016517668962478638, + 0.014067151583731174, + 0.006791914347559214, + 0.006989751942455769, + 0.012286996468901634, + 0.009100418537855148, + 0.0067626070231199265, + 0.009287232533097267, + 0.010844028554856777, + 0.01010803785175085, + 0.00909724272787571, + 0.007225333247333765, + 0.008089895360171795, + 0.009921660646796227, + 0.012884704396128654, + 0.015048697590827942, + 0.010176599957048893, + 0.024945542216300964, + 0.010854045860469341, + 0.023070083931088448, + 0.013755462132394314, + 0.014875659719109535, + 0.031114967539906502, + 0.027431154623627663, + 0.021157143637537956, + 0.014971780590713024, + 0.011402592994272709, + 0.013081561774015427, + 0.00472975242882967, + 0.009292635135352612, + 0.007375806104391813, + 0.00924040749669075, + 0.008238736540079117, + 0.008823365904390812, + 0.006858514156192541, + 0.008296441286802292, + 0.003797115758061409, + 0.006371051073074341, + 0.010229945182800293, + 0.008423768915235996, + 0.00568712642416358, + 0.005595636088401079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.02060963399708271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094076342880726 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094076342880726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094076342880726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1911005973815918, + 0.050331469625234604, + 0.11535783112049103, + 0.18002969026565552, + 0.07755130529403687, + 0.07215142995119095, + 0.04101256653666496, + 0.1132940873503685, + 0.10327456891536713, + 0.05589653179049492 + ], + "credit_map": [ + 0.018940415233373642, + 0.018153294920921326, + 0.02139795571565628, + 0.04074002802371979, + 0.01632399670779705, + 0.02961915172636509, + 0.022734304890036583, + 0.023191440850496292, + 0.003928654361516237, + 0.0040327017195522785, + 0.006205258425325155, + 0.008592575788497925, + 0.008049244061112404, + 0.0071586319245398045, + 0.007610213477164507, + 0.004754188004881144, + 0.008789537474513054, + 0.008394453674554825, + 0.013492916710674763, + 0.02920863777399063, + 0.009836439043283463, + 0.019293488934636116, + 0.012838758528232574, + 0.013503600843250751, + 0.01335885189473629, + 0.014583243057131767, + 0.03443574905395508, + 0.025240899994969368, + 0.01704915426671505, + 0.027474474161863327, + 0.023941440507769585, + 0.023945879191160202, + 0.007709191180765629, + 0.008758732117712498, + 0.007215564139187336, + 0.00876213051378727, + 0.00926679652184248, + 0.011517845094203949, + 0.013320820406079292, + 0.011000222526490688, + 0.0072205523028969765, + 0.007232321426272392, + 0.009524316526949406, + 0.014773937873542309, + 0.006024070084095001, + 0.008623290807008743, + 0.01115961279720068, + 0.007593333255499601, + 0.0038738357834517956, + 0.0051834462210536, + 0.004650170914828777, + 0.005931368097662926, + 0.005039371084421873, + 0.005839964374899864, + 0.005003726109862328, + 0.005490684416145086, + 0.010882235132157803, + 0.027992116287350655, + 0.009956862777471542, + 0.01717521622776985, + 0.01243229117244482, + 0.015160447917878628, + 0.010795561596751213, + 0.008899353444576263, + 0.020602816715836525, + 0.014802586287260056, + 0.007573129143565893, + 0.011437851004302502, + 0.01579827256500721, + 0.014338597655296326, + 0.009924473240971565, + 0.008796840906143188, + 0.004641604144126177, + 0.006611510645598173, + 0.005863754544407129, + 0.00790507160127163, + 0.00804069358855486, + 0.007449871860444546, + 0.008008863776922226, + 0.00737515976652503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.004094076342880726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191891998052597 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014191891998052597, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191891998052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16284491121768951, + 0.045494843274354935, + 0.0991169661283493, + 0.1477023959159851, + 0.13504457473754883, + 0.050602495670318604, + 0.0586325004696846, + 0.11318212747573853, + 0.1240914836525917, + 0.06328769028186798 + ], + "credit_map": [ + 0.01356927864253521, + 0.01654263772070408, + 0.01604616269469261, + 0.02656165324151516, + 0.01762096770107746, + 0.025166919454932213, + 0.019668128341436386, + 0.027669165283441544, + 0.003160925814881921, + 0.0048005408607423306, + 0.005733944475650787, + 0.008716695941984653, + 0.005444326438009739, + 0.0057610715739429, + 0.006179001182317734, + 0.005698340944945812, + 0.005796970333904028, + 0.00701965531334281, + 0.011437928304076195, + 0.026534711942076683, + 0.008604816161096096, + 0.0169775802642107, + 0.013196701183915138, + 0.009548609144985676, + 0.011087128892540932, + 0.011288679204881191, + 0.02867182344198227, + 0.020038019865751266, + 0.01605149172246456, + 0.022404776886105537, + 0.022174986079335213, + 0.015985479578375816, + 0.010161619633436203, + 0.010478482581675053, + 0.01998848095536232, + 0.015562511049211025, + 0.01622653752565384, + 0.018878262490034103, + 0.021501770243048668, + 0.022246917709708214, + 0.005582758691161871, + 0.006205734796822071, + 0.00828430987894535, + 0.0070219519548118114, + 0.004694266244769096, + 0.00621282123029232, + 0.00826180074363947, + 0.004338855389505625, + 0.006021172273904085, + 0.011160864494740963, + 0.005982971750199795, + 0.008795754984021187, + 0.006278321146965027, + 0.00841896515339613, + 0.00670783594250679, + 0.005266620311886072, + 0.007672539446502924, + 0.017023548483848572, + 0.011226624250411987, + 0.012666121125221252, + 0.020976299419999123, + 0.018331291154026985, + 0.014470139518380165, + 0.010815564543008804, + 0.02429242990911007, + 0.027814671397209167, + 0.007549309637397528, + 0.012688112445175648, + 0.014960164204239845, + 0.016251850873231888, + 0.010867134667932987, + 0.009667816571891308, + 0.00706021161749959, + 0.008879459463059902, + 0.005248624831438065, + 0.009872839786112309, + 0.010602308437228203, + 0.0089970538392663, + 0.006176313851028681, + 0.006450881250202656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.014191891998052597 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941657930612564 + ], + "opposite_logprob_tokens": [ + -2.9299416542053223 + ], + "logprob_reward": -0.054941657930612564, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941657930612564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1006951630115509, + 0.09965157508850098, + 0.07412798702716827, + 0.17226405441761017, + 0.105217345058918, + 0.13323672115802765, + 0.1252557635307312, + 0.04588386043906212, + 0.06195912882685661, + 0.08170835673809052 + ], + "credit_map": [ + 0.009280476719141006, + 0.010355217382311821, + 0.018678773194551468, + 0.018406786024570465, + 0.009353573434054852, + 0.01355623360723257, + 0.012876246124505997, + 0.008187852799892426, + 0.006908906623721123, + 0.009156479500234127, + 0.008267412893474102, + 0.011940966360270977, + 0.015528285875916481, + 0.0152092594653368, + 0.013979438692331314, + 0.01866082102060318, + 0.006042998284101486, + 0.005060162860900164, + 0.009146155789494514, + 0.0195920467376709, + 0.007212040014564991, + 0.013111745938658714, + 0.008219561539590359, + 0.005743276793509722, + 0.011290651746094227, + 0.014114643447101116, + 0.03242572024464607, + 0.023761751130223274, + 0.016405519098043442, + 0.026175176724791527, + 0.02463321015238762, + 0.023457391187548637, + 0.009575417265295982, + 0.007947537116706371, + 0.01054784283041954, + 0.0175833310931921, + 0.012668703682720661, + 0.016829997301101685, + 0.015596477314829826, + 0.01446803379803896, + 0.012256709858775139, + 0.011601647362112999, + 0.023558836430311203, + 0.02321731485426426, + 0.01256741676479578, + 0.01854039914906025, + 0.01897665672004223, + 0.012517746537923813, + 0.0179118774831295, + 0.01746889017522335, + 0.010519608855247498, + 0.024279814213514328, + 0.013513712212443352, + 0.016125524416565895, + 0.012134765274822712, + 0.013301570899784565, + 0.00474320724606514, + 0.005783685017377138, + 0.004091331735253334, + 0.007517693564295769, + 0.0062125264666974545, + 0.0065017021261155605, + 0.006640045903623104, + 0.004393666051328182, + 0.009985314682126045, + 0.007711729500442743, + 0.004292227327823639, + 0.006715002004057169, + 0.007787609938532114, + 0.008488056249916553, + 0.008202696219086647, + 0.008776495233178139, + 0.008280765265226364, + 0.011519496329128742, + 0.004659425467252731, + 0.007221318781375885, + 0.015619669109582901, + 0.013835644349455833, + 0.00925036333501339, + 0.01132166851311922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.054941657930612564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The high-contrast macro shot begins, the camera tracking uncomfortably slowly from left to right, maintaining a tight profile. The subject is a dead black ant, its black exoskeleton rendered with exquisite clarity. Its jaws make a final, desperate grip on the vibrant green edge of a waxy leaf, its stiff legs curled tightly around its dark, segmented body. As the camera continues its deliberate crawl, it passes over the ant's thorax and peaks at its head, from which a sinister fungal stalk erupts vertically. The eerie, pale orange plant is fibrous and strangely textured, its surface covered in a layer of glistening, globular morning dew that catches the sharp, directed light like tiny jewels. The intense, high-contrast lighting casts deep, dramatic shadows, sculpting every gruesome detail of the stem's texture and the ant's lifeless form, creating an eerie atmosphere that is both clinically beautiful and deeply unsettling against the backdrop, which dissolves into a soft, out-of-focus blur of dark green and earthy brown.", + "seed": 7437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_12/index_5367_seed_7437_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02513131988234818, + "vlm_logprob": -0.02513131988234818, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.02513131988234818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05690838024020195 + ], + "opposite_logprob_tokens": [ + -2.931908369064331 + ], + "logprob_reward": -0.05690838024020195, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05690838024020195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054730791598558426, + 0.224067822098732, + 0.08277854323387146, + 0.043613482266664505, + 0.2551029324531555, + 0.055594924837350845, + 0.03931393101811409, + 0.14765197038650513, + 0.06751823425292969, + 0.029627347365021706 + ], + "credit_map": [ + 0.009340783581137657, + 0.006101681850850582, + 0.00601777620613575, + 0.005652515683323145, + 0.00920735765248537, + 0.007761875167489052, + 0.004236957058310509, + 0.0064118411391973495, + 0.06980327516794205, + 0.019609585404396057, + 0.005480979569256306, + 0.004238998983055353, + 0.09559979289770126, + 0.01857619918882847, + 0.005379630718380213, + 0.005379378329962492, + 0.012394942343235016, + 0.009095790795981884, + 0.0066545577719807625, + 0.0044992100447416306, + 0.017754944041371346, + 0.0068616727367043495, + 0.021124374121427536, + 0.004393044393509626, + 0.008894470520317554, + 0.005083885043859482, + 0.0051924739964306355, + 0.004869387485086918, + 0.005414925515651703, + 0.00531212380155921, + 0.005526395048946142, + 0.0033198187593370676, + 0.0310395248234272, + 0.028123870491981506, + 0.016048410907387733, + 0.010979886166751385, + 0.03232517093420029, + 0.05244401469826698, + 0.013339311815798283, + 0.07080274820327759, + 0.005008947569876909, + 0.003957944456487894, + 0.0018070511287078261, + 0.002746706595644355, + 0.0027793706394732, + 0.01705004833638668, + 0.002693173009902239, + 0.019551685079932213, + 0.00299825775437057, + 0.0035848841071128845, + 0.002445216290652752, + 0.002997303381562233, + 0.004628144204616547, + 0.0026124860160052776, + 0.011183572001755238, + 0.008864069357514381, + 0.009829324670135975, + 0.004419196862727404, + 0.0033935788087546825, + 0.004553031176328659, + 0.05606953427195549, + 0.0072097331285476685, + 0.05740157514810562, + 0.004776000045239925, + 0.008388050831854343, + 0.013173172250390053, + 0.010480314493179321, + 0.010372249409556389, + 0.011224968358874321, + 0.0044846865348517895, + 0.005647092591971159, + 0.003747703740373254, + 0.003648522309958935, + 0.004573635291308165, + 0.004527125041931868, + 0.0041853198781609535, + 0.003305539023131132, + 0.00291896378621459, + 0.0031628403812646866, + 0.0033054000232368708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.05690838024020195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018216334283351898 + ], + "opposite_logprob_tokens": [ + -4.018216133117676 + ], + "logprob_reward": -0.018216334283351898, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018216334283351898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060648441314697266, + 0.09238744527101517, + 0.05047580599784851, + 0.015709329396486282, + 0.17380495369434357, + 0.03232390806078911, + 0.03018348664045334, + 0.32652178406715393, + 0.1918586641550064, + 0.026086196303367615 + ], + "credit_map": [ + 0.01143753994256258, + 0.0060096606612205505, + 0.006491278763860464, + 0.00540177570655942, + 0.015081536956131458, + 0.008088264614343643, + 0.0033964558970183134, + 0.004741924349218607, + 0.02239210158586502, + 0.009305950254201889, + 0.0033768347930163145, + 0.0032780696637928486, + 0.035070739686489105, + 0.011245091445744038, + 0.0032398987095803022, + 0.004478746559470892, + 0.008601199835538864, + 0.006063707172870636, + 0.005812422838062048, + 0.004865119233727455, + 0.007477253675460815, + 0.005476261023432016, + 0.007429507561028004, + 0.00475033326074481, + 0.0034469314850866795, + 0.0020191248040646315, + 0.0013071607099846005, + 0.0015227760886773467, + 0.002289260271936655, + 0.00210108351893723, + 0.0018061698647215962, + 0.0012168220710009336, + 0.018292663618922234, + 0.012646260671317577, + 0.008182823657989502, + 0.006711574736982584, + 0.01453760638833046, + 0.04559147730469704, + 0.008126368746161461, + 0.05971616879105568, + 0.0036943331360816956, + 0.002300286665558815, + 0.001455930294468999, + 0.0026517375372350216, + 0.0019952424336224794, + 0.008542838506400585, + 0.002277828985825181, + 0.009405707009136677, + 0.0038101798854768276, + 0.0038225071039050817, + 0.0042466712184250355, + 0.003089628182351589, + 0.004956175107508898, + 0.003958497662097216, + 0.00303592998534441, + 0.0032638940028846264, + 0.022018827497959137, + 0.014561503194272518, + 0.009815005585551262, + 0.011137192137539387, + 0.12126141041517258, + 0.01972266472876072, + 0.11686881631612778, + 0.011136339046061039, + 0.020333822816610336, + 0.03761603310704231, + 0.03516867756843567, + 0.032625455409288406, + 0.02678404003381729, + 0.011532370001077652, + 0.01646297052502632, + 0.011335288174450397, + 0.0020795108284801245, + 0.002371870679780841, + 0.0012956102145835757, + 0.0028194962069392204, + 0.003996203187853098, + 0.006235585082322359, + 0.0030783163383603096, + 0.004209599457681179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.018216334283351898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036143234465271235 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036143234465271235, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036143234465271235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03239922225475311, + 0.04102462902665138, + 0.029174285009503365, + 0.029482709243893623, + 0.02584097348153591, + 0.03268881514668465, + 0.13389423489570618, + 0.6048327684402466, + 0.04787464812397957, + 0.022787680849432945 + ], + "credit_map": [ + 0.005854626651853323, + 0.005094442516565323, + 0.00180231430567801, + 0.0029361266642808914, + 0.008203077130019665, + 0.0033202441409230232, + 0.0028413732070475817, + 0.0023470157757401466, + 0.012034866027534008, + 0.003345484146848321, + 0.0012958364095538855, + 0.0010604256531223655, + 0.01711081899702549, + 0.003786134533584118, + 0.0011437240755185485, + 0.0012473373208194971, + 0.006156301125884056, + 0.003389344085007906, + 0.0022503279615193605, + 0.0015874388627707958, + 0.004393119364976883, + 0.004089145455509424, + 0.005275449249893427, + 0.002033158205449581, + 0.005284040234982967, + 0.0030856523662805557, + 0.004696221556514502, + 0.0035031670704483986, + 0.0032008264679461718, + 0.004265021998435259, + 0.0030307353008538485, + 0.002417044248431921, + 0.003508720314130187, + 0.002912329277023673, + 0.002309513743966818, + 0.0028796757105737925, + 0.003199129831045866, + 0.004441028460860252, + 0.002684391802176833, + 0.003906183410435915, + 0.0026168751064687967, + 0.0020015842746943235, + 0.0008709827088750899, + 0.001014676527120173, + 0.001510513131506741, + 0.010897633619606495, + 0.0013685968006029725, + 0.012407952919602394, + 0.009862533770501614, + 0.005492943339049816, + 0.003696201601997018, + 0.0031591402366757393, + 0.015699421986937523, + 0.006009005010128021, + 0.053960006684064865, + 0.03601498529314995, + 0.04044302552938461, + 0.02177542634308338, + 0.015085149556398392, + 0.018488869071006775, + 0.22944417595863342, + 0.03153858333826065, + 0.22860556840896606, + 0.019452018663287163, + 0.005181385204195976, + 0.01253245584666729, + 0.0046716853976249695, + 0.007048088125884533, + 0.008218069560825825, + 0.003646447556093335, + 0.0035774349234998226, + 0.0029990815091878176, + 0.002543508540838957, + 0.003883322700858116, + 0.0025836662389338017, + 0.003671183716505766, + 0.0025585596449673176, + 0.002504161559045315, + 0.0026920465752482414, + 0.002351231873035431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.0036143234465271235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057752013206482 + ], + "opposite_logprob_tokens": [ + -3.8955774307250977 + ], + "logprob_reward": -0.02057752013206482, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057752013206482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009436560794711113, + 0.020612899214029312, + 0.013432358391582966, + 0.004246418364346027, + 0.031019072979688644, + 0.009352174587547779, + 0.06341882050037384, + 0.8099263310432434, + 0.03118380345404148, + 0.00737154483795166 + ], + "credit_map": [ + 0.0022193787153810263, + 0.0013624347047880292, + 0.000619536149315536, + 0.0008426283020526171, + 0.001653190003708005, + 0.0011663594050332904, + 0.0006351027986966074, + 0.0009379326947964728, + 0.006285864394158125, + 0.001679360750131309, + 0.0005023121484555304, + 0.00036251323763281107, + 0.008991621434688568, + 0.0018419702537357807, + 0.0004786728823091835, + 0.00047058859490789473, + 0.0034745035227388144, + 0.0010404326021671295, + 0.0008548159385100007, + 0.0005810819566249847, + 0.0016188084846362472, + 0.0012946584029123187, + 0.0037730110343545675, + 0.0007950480794534087, + 0.00070340366801247, + 0.00042724935337901115, + 0.0004636033554561436, + 0.0004025974194519222, + 0.0004559381341096014, + 0.0007571203168481588, + 0.0006419564015232027, + 0.0003945501521229744, + 0.003304343670606613, + 0.0022840192541480064, + 0.0017191920196637511, + 0.002094371011480689, + 0.002455933718010783, + 0.007331205066293478, + 0.002323226537555456, + 0.009506786242127419, + 0.0010030291741713881, + 0.0008075885125435889, + 0.0009528882801532745, + 0.0007746725459583104, + 0.001020519994199276, + 0.002027129987254739, + 0.0013688921462744474, + 0.0013974554603919387, + 0.00442013144493103, + 0.002783268690109253, + 0.0012206071987748146, + 0.0015894616954028606, + 0.007961455732584, + 0.002702937461435795, + 0.025859348475933075, + 0.016881626099348068, + 0.05419925972819328, + 0.03225700929760933, + 0.02289368212223053, + 0.026434430852532387, + 0.3027516007423401, + 0.04611599072813988, + 0.29869648814201355, + 0.026577945798635483, + 0.0038897707127034664, + 0.0064272694289684296, + 0.005382149945944548, + 0.0047874473966658115, + 0.004312756936997175, + 0.0021030998323112726, + 0.0025050535332411528, + 0.001776259858161211, + 0.0007631159387528896, + 0.0008076074300333858, + 0.0011321251513436437, + 0.0013776980340480804, + 0.0007139996741898358, + 0.0007658065296709538, + 0.0009378824033774436, + 0.0008733104332350194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.02057752013206482 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a dead black ant with a visible black exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant gripping the edge of a vibrant green waxy leaf with its jaws and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister pale orange fungal stalk erupt vertically from the ant's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fungal stalk covered in glistening globular dew that catches sharp directed light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track slowly from left to right, passing over the ant's thorax and peaking at its head?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340041309595108 + ], + "opposite_logprob_tokens": [ + -3.6513400077819824 + ], + "logprob_reward": -0.026340041309595108, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340041309595108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09850063920021057, + 0.03555900231003761, + 0.06888336688280106, + 0.03986363112926483, + 0.055584002286195755, + 0.05272452533245087, + 0.11666641384363174, + 0.4118466079235077, + 0.08718428015708923, + 0.033187489956617355 + ], + "credit_map": [ + 0.016274530440568924, + 0.013700375333428383, + 0.005955213215202093, + 0.006095180753618479, + 0.03219149261713028, + 0.009450179524719715, + 0.005674109794199467, + 0.009159567765891552, + 0.009481963701546192, + 0.0036611855030059814, + 0.0015402185963466763, + 0.0015893985982984304, + 0.012667345814406872, + 0.0035149240866303444, + 0.0014543512370437384, + 0.0016496217576786876, + 0.022180100902915, + 0.005464534275233746, + 0.0036766426637768745, + 0.0028819693252444267, + 0.012702103704214096, + 0.007567540276795626, + 0.009607533924281597, + 0.004802947863936424, + 0.008813624270260334, + 0.0032226217444986105, + 0.003199011320248246, + 0.0034838891588151455, + 0.005137443542480469, + 0.007414575200527906, + 0.00507417693734169, + 0.0035182926803827286, + 0.0050620404072105885, + 0.003753253258764744, + 0.0022538357879966497, + 0.002356467302888632, + 0.004898310638964176, + 0.014917396940290928, + 0.002731171902269125, + 0.019611535593867302, + 0.004328710027039051, + 0.0026281645987182856, + 0.002222204115241766, + 0.0023553227074444294, + 0.0025044833309948444, + 0.017264124006032944, + 0.0027970171067863703, + 0.018624506890773773, + 0.008033089339733124, + 0.004884188994765282, + 0.003126615658402443, + 0.002890627132728696, + 0.012497806921601295, + 0.004858897067606449, + 0.04668614640831947, + 0.03368905931711197, + 0.027529563754796982, + 0.01756172440946102, + 0.01246528048068285, + 0.014269420877099037, + 0.15303431451320648, + 0.025023315101861954, + 0.1478087604045868, + 0.01415424793958664, + 0.00952758826315403, + 0.015518254600465298, + 0.013569117523729801, + 0.012289301492273808, + 0.012898648157715797, + 0.007010466419160366, + 0.009688882157206535, + 0.00668202992528677, + 0.003978101536631584, + 0.0036165243946015835, + 0.005715136416256428, + 0.005277394317090511, + 0.0038601828273385763, + 0.004488985985517502, + 0.003378442022949457, + 0.002872728044167161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.026340041309595108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_tensor([2917, 2917])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_tensor([2917, 2917])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..93d35bef9efad00bb79c0747a9a740117ea87f17 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_tensor([2917, 2917])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera executes dynamic rotating shots, spiraling 360-degree barrel rolls, perfectly reflecting the exhilarating and carefree aerial acrobatics of a magnificent avian creature soaring through the cosmos. Its body is composed not of flesh but of shimmering, sun-filled crystal feathers; each feather is a perfectly faceted, translucent prism, capturing and refracting ambient light, spreading a kaleidoscope of color across its body while emitting a soft, inner golden glow. As it joyfully swoops and dives into a vast, vibrant cloud of swirling fuchsia, deep purple, and glowing cyan gas, it playfully chases its tail—a trailing ribbon of ethereal energy woven from sapphire blue and molten gold light particles, leaving a luminous wake. With effortless grace, it treats the floating crystal asteroids like toys, quickly nudging the smaller ones with its beak to send them tumbling away, and weaving in a tight, joyful circle around the larger metallic structure. The entire scene is a dreamlike symphony of light, with the clouds providing a soft, multi-hued glow that is amplified by the countless prismatic points of light shimmering in the creature’s own luminescent tail and crystalline feathers, creating an atmosphere of pure, unadulterated cosmic joy and infinite freedom.", + "seed": 14635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_2917_seed_14635_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08224710933864117, + "vlm_logprob": -0.08224710933864117, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.08224710933864117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01371978223323822 + ], + "opposite_logprob_tokens": [ + -4.38871955871582 + ], + "logprob_reward": -0.01371978223323822, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01371978223323822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19308561086654663, + 0.07037890702486038, + 0.11406825482845306, + 0.2890469431877136, + 0.06814716011285782, + 0.06484551727771759, + 0.0770857036113739, + 0.048622649163007736, + 0.043529052287340164, + 0.03119022399187088 + ], + "credit_map": [ + 0.02516848035156727, + 0.01428272482007742, + 0.011293846182525158, + 0.011013472452759743, + 0.022455329075455666, + 0.067299984395504, + 0.016237696632742882, + 0.025334075093269348, + 0.006297239568084478, + 0.011867466382682323, + 0.009066524915397167, + 0.007781396619975567, + 0.007816568948328495, + 0.009937954135239124, + 0.009540696628391743, + 0.00807106401771307, + 0.011469813995063305, + 0.014938603155314922, + 0.011031114496290684, + 0.01390534732490778, + 0.015805287286639214, + 0.01593240723013878, + 0.017131205648183823, + 0.013854472897946835, + 0.05658731237053871, + 0.019090307876467705, + 0.10307205468416214, + 0.024999359622597694, + 0.013315129093825817, + 0.011401105672121048, + 0.012922583147883415, + 0.04765910282731056, + 0.00671406788751483, + 0.010411791503429413, + 0.0105409175157547, + 0.0053157806396484375, + 0.006845396943390369, + 0.014937514439225197, + 0.006318494211882353, + 0.007063195575028658, + 0.005770872812718153, + 0.005328021012246609, + 0.00943120289593935, + 0.0059328386560082436, + 0.007366948761045933, + 0.009561456739902496, + 0.008809381164610386, + 0.012644796632230282, + 0.00922058243304491, + 0.0067780716344714165, + 0.008952426724135876, + 0.011884105391800404, + 0.01217306312173605, + 0.007187884766608477, + 0.012814601883292198, + 0.00807497464120388, + 0.01034663524478674, + 0.00437601562589407, + 0.009282933548092842, + 0.005861552432179451, + 0.0035932082682847977, + 0.00362485577352345, + 0.005986149422824383, + 0.005551294889301062, + 0.0037590544670820236, + 0.0049355775117874146, + 0.004849718883633614, + 0.006812115199863911, + 0.0037530427798628807, + 0.006927893031388521, + 0.007317360490560532, + 0.005174290854483843, + 0.0019316773395985365, + 0.0027197592426091433, + 0.0032260871957987547, + 0.005442702677100897, + 0.002398888813331723, + 0.003446705173701048, + 0.006157826166599989, + 0.005866574589163065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01371978223323822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058791182935238 + ], + "opposite_logprob_tokens": [ + -3.895587921142578 + ], + "logprob_reward": -0.02058791182935238, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058791182935238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43611854314804077, + 0.0692363902926445, + 0.036503925919532776, + 0.1523975431919098, + 0.065652035176754, + 0.043424032628536224, + 0.08985307812690735, + 0.06372801214456558, + 0.016943497583270073, + 0.026142923161387444 + ], + "credit_map": [ + 0.05708634853363037, + 0.033750735223293304, + 0.022859672084450722, + 0.02102207951247692, + 0.0485486164689064, + 0.1552690714597702, + 0.03659738227725029, + 0.06098467856645584, + 0.004703629296272993, + 0.0123039735481143, + 0.006235715001821518, + 0.006485272198915482, + 0.009818553924560547, + 0.010290342383086681, + 0.010551405139267445, + 0.008847497403621674, + 0.003393050516024232, + 0.004669893067330122, + 0.003699991386383772, + 0.006701834499835968, + 0.00488493824377656, + 0.004726656712591648, + 0.004235616885125637, + 0.004191948566585779, + 0.01604105532169342, + 0.011902370490133762, + 0.04019064083695412, + 0.013836577534675598, + 0.011080928146839142, + 0.012941335327923298, + 0.008099411614239216, + 0.038305237889289856, + 0.006790136452764273, + 0.00939722079783678, + 0.0082792853936553, + 0.006574299652129412, + 0.00868018064647913, + 0.010685366578400135, + 0.008322279900312424, + 0.00692326994612813, + 0.00347482250072062, + 0.003465864807367325, + 0.004434347618371248, + 0.0072303772903978825, + 0.005921465810388327, + 0.008274251595139503, + 0.0059709493070840836, + 0.004651955794543028, + 0.00979736540466547, + 0.00841598305851221, + 0.009959383867681026, + 0.01471302006393671, + 0.014829848892986774, + 0.008446617051959038, + 0.0146415950730443, + 0.00904927495867014, + 0.012719091027975082, + 0.005728481337428093, + 0.012960602529346943, + 0.008161695674061775, + 0.004611966200172901, + 0.003945691045373678, + 0.008032017387449741, + 0.007568475790321827, + 0.0017688211519271135, + 0.0014997613616287708, + 0.0023341805208474398, + 0.002742823213338852, + 0.0019719444680958986, + 0.002285656286403537, + 0.002475154120475054, + 0.0018651564605534077, + 0.0014507409650832415, + 0.002087640343233943, + 0.003152333665639162, + 0.005415025167167187, + 0.0017679096199572086, + 0.0033487759064882994, + 0.005142465233802795, + 0.0037780331913381815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.02058791182935238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988835275173187 + ], + "opposite_logprob_tokens": [ + -1.804888367652893 + ], + "logprob_reward": -0.17988835275173187, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988835275173187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5896130204200745, + 0.05176010727882385, + 0.034764423966407776, + 0.10104474425315857, + 0.034694403409957886, + 0.04565652459859848, + 0.047077663242816925, + 0.048463597893714905, + 0.030227530747652054, + 0.016697954386472702 + ], + "credit_map": [ + 0.08057687431573868, + 0.039328884333372116, + 0.03157242760062218, + 0.024175887927412987, + 0.06915347278118134, + 0.21918970346450806, + 0.048849090933799744, + 0.07676670700311661, + 0.003016469767317176, + 0.007044529542326927, + 0.004396150354295969, + 0.003950298763811588, + 0.00953766517341137, + 0.008698178455233574, + 0.008122877217829227, + 0.006993935443460941, + 0.00324532319791615, + 0.005888570100069046, + 0.003264073049649596, + 0.005073740147054195, + 0.00492791598662734, + 0.005762284155935049, + 0.003015488851815462, + 0.0035870291758328676, + 0.020322022959589958, + 0.0052386196330189705, + 0.037409648299217224, + 0.007085976656526327, + 0.005243760999292135, + 0.005119863897562027, + 0.0041552213951945305, + 0.016469623893499374, + 0.004526157397776842, + 0.004744570702314377, + 0.003696552012115717, + 0.004097213968634605, + 0.005577536299824715, + 0.00404772162437439, + 0.004658524412661791, + 0.003346128622069955, + 0.003964207600802183, + 0.004757437855005264, + 0.0036397636868059635, + 0.005051434971392155, + 0.005433123558759689, + 0.010511782020330429, + 0.0057060387916862965, + 0.006592736579477787, + 0.004727809689939022, + 0.004586522933095694, + 0.006748088635504246, + 0.007479856256395578, + 0.007017773110419512, + 0.0045922789722681046, + 0.007279009558260441, + 0.004646324086934328, + 0.008195293135941029, + 0.004466507583856583, + 0.007228069473057985, + 0.006981634069234133, + 0.004117722623050213, + 0.004003515467047691, + 0.006673482712358236, + 0.006797377485781908, + 0.0036897645331919193, + 0.0028078784234821796, + 0.0027793841436505318, + 0.004349739756435156, + 0.004588440991938114, + 0.004573827609419823, + 0.0035536668729037046, + 0.0038848291151225567, + 0.0013477664906531572, + 0.0011070958571508527, + 0.0017569024348631501, + 0.0027667642571032047, + 0.0016463600331917405, + 0.002479535760357976, + 0.003558265045285225, + 0.0020352634601294994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.17988835275173187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270465850830078 + ], + "opposite_logprob_tokens": [ + -2.127046585083008 + ], + "logprob_reward": -0.1270465850830078, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270465850830078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5826351046562195, + 0.04871015623211861, + 0.03374192491173744, + 0.06208403408527374, + 0.04067908599972725, + 0.06575572490692139, + 0.051794618368148804, + 0.05426950752735138, + 0.030619606375694275, + 0.02971024625003338 + ], + "credit_map": [ + 0.08007975667715073, + 0.040276575833559036, + 0.031013375148177147, + 0.023523464798927307, + 0.0668068677186966, + 0.21636711061000824, + 0.04814271256327629, + 0.07642525434494019, + 0.003628736361861229, + 0.009488841518759727, + 0.005853339564055204, + 0.0042328257113695145, + 0.005831147078424692, + 0.0070027620531618595, + 0.006948307156562805, + 0.0057242028415203094, + 0.003427570452913642, + 0.004882364999502897, + 0.0034721973352134228, + 0.0049390229396522045, + 0.005097291432321072, + 0.004791310057044029, + 0.003595186397433281, + 0.003536986419931054, + 0.007098040077835321, + 0.0036585009656846523, + 0.01944032311439514, + 0.004078825935721397, + 0.00450911745429039, + 0.004018487874418497, + 0.00397914182394743, + 0.015301601961255074, + 0.005159637425094843, + 0.004997001029551029, + 0.004696081858128309, + 0.004524098709225655, + 0.005040613003075123, + 0.006401319056749344, + 0.005082848947495222, + 0.004777488298714161, + 0.004850676283240318, + 0.006508667021989822, + 0.009071581065654755, + 0.007010088302195072, + 0.007306873798370361, + 0.011176219210028648, + 0.009360571391880512, + 0.01047105249017477, + 0.005679558031260967, + 0.004745880141854286, + 0.006183318328112364, + 0.007350396830588579, + 0.008265024051070213, + 0.004803408868610859, + 0.008992644026875496, + 0.005774395074695349, + 0.008440402336418629, + 0.005325092934072018, + 0.005876055918633938, + 0.007846948690712452, + 0.006925193592905998, + 0.0046306028962135315, + 0.00831181276589632, + 0.006913404446095228, + 0.003292914479970932, + 0.0028106020763516426, + 0.002890221541747451, + 0.0041180821135640144, + 0.0029318423476070166, + 0.004831099417060614, + 0.005214047618210316, + 0.0045308000408113, + 0.0019071069546043873, + 0.0028772838413715363, + 0.003353593870997429, + 0.003935099579393864, + 0.0020940955728292465, + 0.0037155048921704292, + 0.006013751495629549, + 0.005813813302665949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.1270465850830078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999291479587555 + ], + "opposite_logprob_tokens": [ + -2.694993019104004 + ], + "logprob_reward": -0.06999291479587555, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999291479587555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7088983654975891, + 0.06199418380856514, + 0.019091537222266197, + 0.04492762312293053, + 0.04496712237596512, + 0.015779845416545868, + 0.03941723331809044, + 0.027198340743780136, + 0.016043199226260185, + 0.021682489663362503 + ], + "credit_map": [ + 0.09780378639698029, + 0.04811961203813553, + 0.038140859454870224, + 0.028030071407556534, + 0.08253654837608337, + 0.26282191276550293, + 0.058970678597688675, + 0.09247495234012604, + 0.005111952777951956, + 0.010516385547816753, + 0.007134241983294487, + 0.005636796820908785, + 0.006984305568039417, + 0.009806822054088116, + 0.00894636195152998, + 0.007857314310967922, + 0.0021942919120192528, + 0.0027554468251764774, + 0.002130489330738783, + 0.0025558259803801775, + 0.003202896798029542, + 0.0025871521793305874, + 0.0019459492759779096, + 0.0017194850370287895, + 0.0073849973268806934, + 0.0057699671015143394, + 0.00438203290104866, + 0.005954108200967312, + 0.005970541387796402, + 0.0034810814540833235, + 0.006170723121613264, + 0.0058141727931797504, + 0.004915502388030291, + 0.006537429988384247, + 0.005608975887298584, + 0.005041459109634161, + 0.006443662103265524, + 0.005908965598791838, + 0.005821868311613798, + 0.004689259920269251, + 0.0016093294834718108, + 0.0019031668780371547, + 0.0016598242800682783, + 0.0021168971434235573, + 0.0021176941227167845, + 0.00225646048784256, + 0.001783276442438364, + 0.002333197044208646, + 0.0042981733568012714, + 0.004403839353471994, + 0.006765362340956926, + 0.0037701737601310015, + 0.005810778588056564, + 0.0032577826641499996, + 0.0065242755226790905, + 0.004586843773722649, + 0.005109194666147232, + 0.0027089433278888464, + 0.0035774558782577515, + 0.003601835109293461, + 0.002370729111135006, + 0.0029139539692550898, + 0.002653045579791069, + 0.004263181705027819, + 0.0012124216882511973, + 0.0016447146190330386, + 0.0019914309959858656, + 0.0022571601439267397, + 0.00190593465231359, + 0.002375146606937051, + 0.002814551582559943, + 0.0018418364925310016, + 0.001380900270305574, + 0.0016532891895622015, + 0.001838955795392394, + 0.0027676820755004883, + 0.0016872802516445518, + 0.004455495160073042, + 0.00316798803396523, + 0.00473089748993516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.06999291479587555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera executes dynamic rotating shots, spiraling 360-degree barrel rolls, perfectly reflecting the exhilarating and carefree aerial acrobatics of a magnificent avian creature soaring through the cosmos. Its body is composed not of flesh but of shimmering, sun-filled crystal feathers; each feather is a perfectly faceted, translucent prism, capturing and refracting ambient light, spreading a kaleidoscope of color across its body while emitting a soft, inner golden glow. As it joyfully swoops and dives into a vast, vibrant cloud of swirling fuchsia, deep purple, and glowing cyan gas, it playfully chases its tail—a trailing ribbon of ethereal energy woven from sapphire blue and molten gold light particles, leaving a luminous wake. With effortless grace, it treats the floating crystal asteroids like toys, quickly nudging the smaller ones with its beak to send them tumbling away, and weaving in a tight, joyful circle around the larger metallic structure. The entire scene is a dreamlike symphony of light, with the clouds providing a soft, multi-hued glow that is amplified by the countless prismatic points of light shimmering in the creature’s own luminescent tail and crystalline feathers, creating an atmosphere of pure, unadulterated cosmic joy and infinite freedom.", + "seed": 14636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_2917_seed_14636_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0542815001681447, + "vlm_logprob": -0.0542815001681447, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.0542815001681447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016630472615361214 + ], + "opposite_logprob_tokens": [ + -4.14163064956665 + ], + "logprob_reward": -0.016630472615361214, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016630472615361214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44079843163490295, + 0.048134513199329376, + 0.1733628809452057, + 0.03081638179719448, + 0.07583054155111313, + 0.0797487124800682, + 0.06570426374673843, + 0.022053668275475502, + 0.03179525211453438, + 0.031755343079566956 + ], + "credit_map": [ + 0.043038755655288696, + 0.07683154940605164, + 0.05347658693790436, + 0.06471797823905945, + 0.05715250223875046, + 0.03919186815619469, + 0.07509241253137589, + 0.03129676729440689, + 0.0043823374435305595, + 0.005113853141665459, + 0.005991324782371521, + 0.005689315497875214, + 0.008869582787156105, + 0.007089212536811829, + 0.005629600491374731, + 0.005369286052882671, + 0.05906527116894722, + 0.006477560847997665, + 0.07287272065877914, + 0.00833952333778143, + 0.004744121339172125, + 0.011009675450623035, + 0.006389103829860687, + 0.0044649154879152775, + 0.004514514002948999, + 0.003990020137280226, + 0.006276720203459263, + 0.0028571870643645525, + 0.0026805659290403128, + 0.003384706564247608, + 0.0031317693646997213, + 0.00398089736700058, + 0.007621727418154478, + 0.017310569062829018, + 0.00963661726564169, + 0.01328529603779316, + 0.008305535651743412, + 0.0073104300536215305, + 0.007716839667409658, + 0.0046435254625976086, + 0.023895980790257454, + 0.005501292645931244, + 0.025125961750745773, + 0.0068030161783099174, + 0.005653536878526211, + 0.003770123003050685, + 0.004528242629021406, + 0.004470553714782, + 0.014794239774346352, + 0.004747324623167515, + 0.0192987322807312, + 0.006035270635038614, + 0.009161839261651039, + 0.0038636229000985622, + 0.0051161679439246655, + 0.0026870674919337034, + 0.0018384562572464347, + 0.002071538008749485, + 0.004050134215503931, + 0.003006191458553076, + 0.004673784598708153, + 0.00200286484323442, + 0.0024332012981176376, + 0.0019774967804551125, + 0.004129080101847649, + 0.003892485750839114, + 0.004909795708954334, + 0.0039271400310099125, + 0.005051206331700087, + 0.002689073560759425, + 0.0042344434186816216, + 0.0029620290733873844, + 0.003958947025239468, + 0.003617895534262061, + 0.0055749015882611275, + 0.004339310340583324, + 0.003701339475810528, + 0.00293638464063406, + 0.004014917649328709, + 0.0036116482224315405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.016630472615361214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791205637156963 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791205637156963, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791205637156963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253107875585556, + 0.06655872613191605, + 0.33647286891937256, + 0.03407884016633034, + 0.03764808550477028, + 0.06284106522798538, + 0.025312794372439384, + 0.02632957696914673, + 0.14479006826877594, + 0.0406571701169014 + ], + "credit_map": [ + 0.02419370226562023, + 0.03719932213425636, + 0.02918650023639202, + 0.03387489914894104, + 0.030328931286931038, + 0.019016310572624207, + 0.03077712096273899, + 0.020734012126922607, + 0.007445594295859337, + 0.005861697252839804, + 0.011859031394124031, + 0.011033521965146065, + 0.010853647254407406, + 0.00838753767311573, + 0.006512259598821402, + 0.004605431109666824, + 0.11192657053470612, + 0.012591994367539883, + 0.14282190799713135, + 0.01648969016969204, + 0.009770028293132782, + 0.022044943645596504, + 0.012446979992091656, + 0.008380735293030739, + 0.00598150584846735, + 0.0030485661700367928, + 0.004708973225206137, + 0.0023433323949575424, + 0.0027607628144323826, + 0.008070492185652256, + 0.0032057426869869232, + 0.00395946204662323, + 0.0033855638466775417, + 0.006172467488795519, + 0.0037871464155614376, + 0.006165966857224703, + 0.004577373154461384, + 0.0050695184618234634, + 0.005167549476027489, + 0.0033224995713680983, + 0.009496573358774185, + 0.008269931189715862, + 0.013009232468903065, + 0.00582217937335372, + 0.010582908056676388, + 0.00734517490491271, + 0.004231002647429705, + 0.004084066487848759, + 0.0028998174238950014, + 0.0032249274663627148, + 0.0031299153342843056, + 0.0028285514563322067, + 0.00425399374216795, + 0.0029862693045288324, + 0.0037684100680053234, + 0.002220909111201763, + 0.002163466066122055, + 0.0035965112037956715, + 0.004135284572839737, + 0.0036608290392905474, + 0.005031953100115061, + 0.0026995434891432524, + 0.002457902068272233, + 0.0025840881280601025, + 0.019326478242874146, + 0.01791105791926384, + 0.026342879980802536, + 0.023964744061231613, + 0.016368113458156586, + 0.008808530867099762, + 0.020297028124332428, + 0.011771218851208687, + 0.004708289634436369, + 0.004762112163007259, + 0.010632187128067017, + 0.0057584685273468494, + 0.004168628249317408, + 0.003092881292104721, + 0.003266085870563984, + 0.0042685153894126415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.009791205637156963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292202025651932 + ], + "opposite_logprob_tokens": [ + -2.237921953201294 + ], + "logprob_reward": -0.11292202025651932, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292202025651932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05677139759063721, + 0.06485267728567123, + 0.1386653631925583, + 0.05504215881228447, + 0.055225204676389694, + 0.14805760979652405, + 0.05333905667066574, + 0.07341914623975754, + 0.3063226044178009, + 0.048304807394742966 + ], + "credit_map": [ + 0.004788793157786131, + 0.007187290582805872, + 0.006811289582401514, + 0.009136049076914787, + 0.006737911608070135, + 0.0036568453069776297, + 0.014257547445595264, + 0.004195669200271368, + 0.006821416784077883, + 0.006601459346711636, + 0.009979686699807644, + 0.008682915940880775, + 0.011201394721865654, + 0.0075159575790166855, + 0.007719402201473713, + 0.006330439355224371, + 0.045386072248220444, + 0.006021769251674414, + 0.057812243700027466, + 0.006520635914057493, + 0.00515203271061182, + 0.009188578464090824, + 0.004803952295333147, + 0.003780070459470153, + 0.007172842975705862, + 0.007470262236893177, + 0.010034075938165188, + 0.0053726183250546455, + 0.005301373545080423, + 0.0076211243867874146, + 0.006074849050492048, + 0.005995013751089573, + 0.00443330779671669, + 0.008977966383099556, + 0.006878524087369442, + 0.008051112294197083, + 0.00574777415022254, + 0.008300386369228363, + 0.008278470486402512, + 0.0045576621778309345, + 0.03652770072221756, + 0.014013847336173058, + 0.040115147829055786, + 0.014553080312907696, + 0.012071669101715088, + 0.007685375399887562, + 0.012807455845177174, + 0.010283323004841805, + 0.013381216675043106, + 0.004541563801467419, + 0.016540231183171272, + 0.0057444204576313496, + 0.0038703379686921835, + 0.0028023889753967524, + 0.0040593743324279785, + 0.002399514662101865, + 0.007187223061919212, + 0.014109047129750252, + 0.00843055546283722, + 0.00876687653362751, + 0.010773915797472, + 0.008753236383199692, + 0.006622734013944864, + 0.00877555925399065, + 0.04254842922091484, + 0.03116788901388645, + 0.0548347644507885, + 0.05602796748280525, + 0.03291007876396179, + 0.016864323988556862, + 0.047784000635147095, + 0.02418516017496586, + 0.006286518648266792, + 0.0066722482442855835, + 0.005454543046653271, + 0.007919400930404663, + 0.0073833586648106575, + 0.004808656871318817, + 0.00664989510551095, + 0.0031301877461373806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.11292202025651932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305814579129219 + ], + "opposite_logprob_tokens": [ + -3.168058156967163 + ], + "logprob_reward": -0.04305814579129219, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305814579129219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3883950114250183, + 0.06820600479841232, + 0.12258424609899521, + 0.07001860439777374, + 0.0441705547273159, + 0.051902733743190765, + 0.06308296322822571, + 0.05534360557794571, + 0.10946625471115112, + 0.02682999148964882 + ], + "credit_map": [ + 0.037973057478666306, + 0.07090707868337631, + 0.051199667155742645, + 0.05566820129752159, + 0.05580886825919151, + 0.03167348355054855, + 0.0579913891851902, + 0.027173249050974846, + 0.005176968406885862, + 0.005808403715491295, + 0.013996768742799759, + 0.00794970989227295, + 0.011029488407075405, + 0.010315795429050922, + 0.007601769175380468, + 0.0063270991668105125, + 0.02428566664457321, + 0.011884965933859348, + 0.032863110303878784, + 0.01137703750282526, + 0.016093973070383072, + 0.008944975212216377, + 0.010382052510976791, + 0.006752463057637215, + 0.009093042463064194, + 0.009030523709952831, + 0.010757101699709892, + 0.006098415702581406, + 0.0053887697868049145, + 0.011299963109195232, + 0.006933523807674646, + 0.01141726691275835, + 0.004134890157729387, + 0.006605171132832766, + 0.004945147782564163, + 0.005185076501220465, + 0.004762774333357811, + 0.006223584059625864, + 0.007736652623862028, + 0.004577257204800844, + 0.006424915976822376, + 0.007232367061078548, + 0.006212267559021711, + 0.00865172315388918, + 0.007074805907905102, + 0.005263730883598328, + 0.005043631885200739, + 0.005999288521707058, + 0.016237817704677582, + 0.00436586095020175, + 0.01972266286611557, + 0.005859164521098137, + 0.005514783784747124, + 0.004705472383648157, + 0.0036818955559283495, + 0.0029953073244541883, + 0.005881136748939753, + 0.008731523528695107, + 0.007153669837862253, + 0.008035956881940365, + 0.008368016220629215, + 0.005769029259681702, + 0.004673670511692762, + 0.00673060305416584, + 0.013501371257007122, + 0.012651506811380386, + 0.02110329270362854, + 0.015104923397302628, + 0.011793836019933224, + 0.007373538333922625, + 0.016578437760472298, + 0.011359344236552715, + 0.0026326440274715424, + 0.0032639573328197002, + 0.005823804996907711, + 0.004346977453678846, + 0.0025494061410427094, + 0.0028618124779313803, + 0.0025670966133475304, + 0.002784292446449399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.04305814579129219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900565654039383 + ], + "opposite_logprob_tokens": [ + -2.464005708694458 + ], + "logprob_reward": -0.08900565654039383, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900565654039383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1467859148979187, + 0.09010472148656845, + 0.3652923107147217, + 0.06057000160217285, + 0.0495070181787014, + 0.04609707370400429, + 0.044954147189855576, + 0.04780060052871704, + 0.12900222837924957, + 0.0198859591037035 + ], + "credit_map": [ + 0.015127426944673061, + 0.021568220108747482, + 0.015024470165371895, + 0.021046658977866173, + 0.017779890447854996, + 0.013812040910124779, + 0.029819093644618988, + 0.012608109042048454, + 0.00871058739721775, + 0.00837993249297142, + 0.02135499194264412, + 0.01338651031255722, + 0.012071644887328148, + 0.012520489282906055, + 0.006958784535527229, + 0.0067217848263680935, + 0.10032720118761063, + 0.021584104746580124, + 0.1409044712781906, + 0.02117322012782097, + 0.02290584333240986, + 0.02457118220627308, + 0.0189631637185812, + 0.014863109216094017, + 0.006769814062863588, + 0.007474022451788187, + 0.012433861382305622, + 0.0054571861401200294, + 0.005471919197589159, + 0.008745217695832253, + 0.006854135077446699, + 0.007363846059888601, + 0.004884439054876566, + 0.00871530082076788, + 0.005494854412972927, + 0.007194481324404478, + 0.006468100938946009, + 0.006026707123965025, + 0.006223790813237429, + 0.004499341361224651, + 0.006798481103032827, + 0.00704052671790123, + 0.006327594164758921, + 0.004994134418666363, + 0.009261787869036198, + 0.003584392834454775, + 0.003730194643139839, + 0.004359961021691561, + 0.009583364240825176, + 0.003622087650001049, + 0.010977048426866531, + 0.004982581827789545, + 0.005341500975191593, + 0.0037072168197482824, + 0.004531385377049446, + 0.0022089630365371704, + 0.004891494754701853, + 0.007302736863493919, + 0.005636002402752638, + 0.006713148672133684, + 0.007318262942135334, + 0.00657130591571331, + 0.004160507116466761, + 0.005207139998674393, + 0.015025453642010689, + 0.016266770660877228, + 0.018115591257810593, + 0.025849148631095886, + 0.015252589248120785, + 0.0069191125221550465, + 0.015172632411122322, + 0.016400933265686035, + 0.0022924577351659536, + 0.0024072271771728992, + 0.005743912421166897, + 0.002685623476281762, + 0.0017609045607969165, + 0.0013590463204309344, + 0.0018225809326395392, + 0.0018142055487260222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.08900565654039383 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera executes dynamic rotating shots, spiraling 360-degree barrel rolls, perfectly reflecting the exhilarating and carefree aerial acrobatics of a magnificent avian creature soaring through the cosmos. Its body is composed not of flesh but of shimmering, sun-filled crystal feathers; each feather is a perfectly faceted, translucent prism, capturing and refracting ambient light, spreading a kaleidoscope of color across its body while emitting a soft, inner golden glow. As it joyfully swoops and dives into a vast, vibrant cloud of swirling fuchsia, deep purple, and glowing cyan gas, it playfully chases its tail—a trailing ribbon of ethereal energy woven from sapphire blue and molten gold light particles, leaving a luminous wake. With effortless grace, it treats the floating crystal asteroids like toys, quickly nudging the smaller ones with its beak to send them tumbling away, and weaving in a tight, joyful circle around the larger metallic structure. The entire scene is a dreamlike symphony of light, with the clouds providing a soft, multi-hued glow that is amplified by the countless prismatic points of light shimmering in the creature’s own luminescent tail and crystalline feathers, creating an atmosphere of pure, unadulterated cosmic joy and infinite freedom.", + "seed": 14637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_2917_seed_14637_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027454950474202634, + "vlm_logprob": -0.027454950474202634, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.027454950474202634, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009449399076402187 + ], + "opposite_logprob_tokens": [ + -4.759449481964111 + ], + "logprob_reward": -0.009449399076402187, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009449399076402187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17409737408161163, + 0.09950575977563858, + 0.10797975957393646, + 0.0580904558300972, + 0.0748601034283638, + 0.10560894757509232, + 0.1427762806415558, + 0.05566029995679855, + 0.07021105289459229, + 0.11120988428592682 + ], + "credit_map": [ + 0.013285066932439804, + 0.017094776034355164, + 0.021666716784238815, + 0.020796310156583786, + 0.02895178087055683, + 0.03185220807790756, + 0.015293658711016178, + 0.025156885385513306, + 0.014382435940206051, + 0.01830129697918892, + 0.011716199107468128, + 0.013607937842607498, + 0.010130563750863075, + 0.009352365508675575, + 0.013165667653083801, + 0.008849307894706726, + 0.011539234779775143, + 0.014550278894603252, + 0.010552620515227318, + 0.027716562151908875, + 0.0075593325309455395, + 0.006766036152839661, + 0.021602360531687737, + 0.007693346589803696, + 0.008120808750391006, + 0.008239329792559147, + 0.008252478204667568, + 0.0077683692798018456, + 0.005001054145395756, + 0.007774678990244865, + 0.006815133150666952, + 0.006118610966950655, + 0.005089556332677603, + 0.023176243528723717, + 0.005089535843580961, + 0.004438401199877262, + 0.010648231022059917, + 0.009708852507174015, + 0.003917934838682413, + 0.012791362591087818, + 0.00907003041356802, + 0.023671090602874756, + 0.014128477312624454, + 0.01724199950695038, + 0.013399179093539715, + 0.009677558206021786, + 0.011667070910334587, + 0.006753557361662388, + 0.021160820499062538, + 0.01768077351152897, + 0.02293097972869873, + 0.02608465403318405, + 0.015231119468808174, + 0.018006231635808945, + 0.01354780700057745, + 0.008133922703564167, + 0.010690748691558838, + 0.005759231746196747, + 0.00934480782598257, + 0.004554929211735725, + 0.004792791325598955, + 0.0084138885140419, + 0.006360230501741171, + 0.005743682850152254, + 0.008428229950368404, + 0.009681960567831993, + 0.015517557971179485, + 0.00889180600643158, + 0.005693167448043823, + 0.0055536231957376, + 0.008760555647313595, + 0.007684166077524424, + 0.010341793298721313, + 0.008439050987362862, + 0.0276823490858078, + 0.00780215160921216, + 0.01037373673170805, + 0.012653078883886337, + 0.02032538130879402, + 0.013592368923127651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.009449399076402187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956161767244339 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956161767244339, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956161767244339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16395069658756256, + 0.07539065927267075, + 0.22149118781089783, + 0.10151766985654831, + 0.06274092197418213, + 0.07400986552238464, + 0.10853452980518341, + 0.054939430207014084, + 0.07225945591926575, + 0.06516558676958084 + ], + "credit_map": [ + 0.013347983360290527, + 0.01914558745920658, + 0.018362995237112045, + 0.017064347863197327, + 0.01740540936589241, + 0.03631279990077019, + 0.014012999832630157, + 0.028298575431108475, + 0.007232754956930876, + 0.01453449111431837, + 0.00883376132696867, + 0.011411732994019985, + 0.007965290918946266, + 0.008057079277932644, + 0.010095546022057533, + 0.007259999867528677, + 0.023843882605433464, + 0.031235970556735992, + 0.025682387873530388, + 0.034415461122989655, + 0.02171543799340725, + 0.026264239102602005, + 0.03501235321164131, + 0.023321455344557762, + 0.01195879839360714, + 0.01386856660246849, + 0.012981223873794079, + 0.013597127981483936, + 0.014881907030940056, + 0.013517936691641808, + 0.01297690998762846, + 0.0077352020889520645, + 0.004411756526678801, + 0.005429974291473627, + 0.013360973447561264, + 0.005214966367930174, + 0.00880118552595377, + 0.006929713301360607, + 0.005140512250363827, + 0.013451838865876198, + 0.005277298390865326, + 0.008623787201941013, + 0.01234910823404789, + 0.012392984703183174, + 0.007785668130964041, + 0.01225034985691309, + 0.010719153098762035, + 0.004611516371369362, + 0.016406815499067307, + 0.012582262046635151, + 0.01588413305580616, + 0.020018422976136208, + 0.012265423312783241, + 0.014237700961530209, + 0.010062343440949917, + 0.007077431306242943, + 0.006744323298335075, + 0.007682384457439184, + 0.006218825466930866, + 0.00527707114815712, + 0.00502801826223731, + 0.0104777030646801, + 0.0075459592044353485, + 0.005965142510831356, + 0.008760057389736176, + 0.008089677430689335, + 0.014037557877600193, + 0.00887215044349432, + 0.006322514731436968, + 0.005833677481859922, + 0.010283241979777813, + 0.010060581378638744, + 0.006875714287161827, + 0.0062265293672680855, + 0.013194656930863857, + 0.006804618984460831, + 0.007371855899691582, + 0.008343055844306946, + 0.008721544407308102, + 0.007627605460584164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.005956161767244339 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000625878572464 + ], + "opposite_logprob_tokens": [ + -2.6950063705444336 + ], + "logprob_reward": -0.07000625878572464, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000625878572464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08387864381074905, + 0.06774922460317612, + 0.10371340066194534, + 0.16325825452804565, + 0.35478147864341736, + 0.05370020866394043, + 0.04972201958298683, + 0.048416975885629654, + 0.03505343198776245, + 0.03972635045647621 + ], + "credit_map": [ + 0.0067281718365848064, + 0.008541202172636986, + 0.006560413166880608, + 0.009785725735127926, + 0.007418500259518623, + 0.017929300665855408, + 0.005752106197178364, + 0.021163227036595345, + 0.007945209741592407, + 0.011212015524506569, + 0.006912237033247948, + 0.009330317378044128, + 0.008154012262821198, + 0.0073928809724748135, + 0.009074448607861996, + 0.007728102616965771, + 0.016820114105939865, + 0.013680749572813511, + 0.011967910453677177, + 0.01728268712759018, + 0.006810205057263374, + 0.013883095234632492, + 0.01329501997679472, + 0.00997361820191145, + 0.017188258469104767, + 0.022230317816138268, + 0.021191980689764023, + 0.023565446957945824, + 0.017320696264505386, + 0.023603405803442, + 0.02056461200118065, + 0.017593542113900185, + 0.01739511825144291, + 0.138553187251091, + 0.06236397102475166, + 0.011094759218394756, + 0.03621186688542366, + 0.02556828409433365, + 0.01142709981650114, + 0.052167169749736786, + 0.005036225542426109, + 0.007432338781654835, + 0.007503744680434465, + 0.0074387588538229465, + 0.006505684461444616, + 0.00893656350672245, + 0.006200036965310574, + 0.004646853543817997, + 0.0066946716979146, + 0.0075744157657027245, + 0.00637574028223753, + 0.00847515370696783, + 0.0062393005937337875, + 0.004941073711961508, + 0.006506328005343676, + 0.0029153365176171064, + 0.00653179083019495, + 0.003955396357923746, + 0.0061906627379357815, + 0.004561095964163542, + 0.004686858505010605, + 0.009441309608519077, + 0.007583939470350742, + 0.005465919617563486, + 0.004827181808650494, + 0.003901168005540967, + 0.005327711813151836, + 0.0038530926685780287, + 0.004102613776922226, + 0.00382793671451509, + 0.004535243380814791, + 0.004678485449403524, + 0.004001503344625235, + 0.003673865692690015, + 0.007226516492664814, + 0.0038730567321181297, + 0.00500040128827095, + 0.004790524020791054, + 0.007117954548448324, + 0.004042529966682196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.07000625878572464 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818004623055458 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.01818004623055458, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818004623055458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057713788002729416, + 0.07028590142726898, + 0.04210042953491211, + 0.09863253682851791, + 0.13540300726890564, + 0.30571839213371277, + 0.043086010962724686, + 0.10034135729074478, + 0.09424440562725067, + 0.05247419327497482 + ], + "credit_map": [ + 0.004948201589286327, + 0.0043303100392222404, + 0.006972029339522123, + 0.0076818340457975864, + 0.004774417728185654, + 0.011383617296814919, + 0.006186113692820072, + 0.011437265202403069, + 0.008527430705726147, + 0.014571374282240868, + 0.0069210464134812355, + 0.00887896865606308, + 0.00916218850761652, + 0.007584125269204378, + 0.00988488458096981, + 0.004755886271595955, + 0.004864428658038378, + 0.005621243268251419, + 0.003207117784768343, + 0.005899179261177778, + 0.005298526491969824, + 0.005806335248053074, + 0.0054174549877643585, + 0.005986144300550222, + 0.009890452958643436, + 0.013600029982626438, + 0.011204791255295277, + 0.013173168525099754, + 0.011833703145384789, + 0.014414701610803604, + 0.011816589161753654, + 0.01269910205155611, + 0.006767553742974997, + 0.04635913297533989, + 0.04042317718267441, + 0.008905555121600628, + 0.011842897161841393, + 0.01043631974607706, + 0.004720922559499741, + 0.00594745110720396, + 0.02228669822216034, + 0.058050576597452164, + 0.04401533305644989, + 0.04484851285815239, + 0.045552387833595276, + 0.02397659607231617, + 0.04981371387839317, + 0.01717459224164486, + 0.007688011042773724, + 0.00550429429858923, + 0.006082572042942047, + 0.007686225231736898, + 0.0050413706339895725, + 0.004432392306625843, + 0.004305721260607243, + 0.002345423214137554, + 0.023751119151711464, + 0.006794637069106102, + 0.024105463176965714, + 0.00874978955835104, + 0.006082062143832445, + 0.011651811189949512, + 0.009809404611587524, + 0.009397076442837715, + 0.01063322089612484, + 0.01299192477017641, + 0.01769286021590233, + 0.010319017805159092, + 0.007705194875597954, + 0.00892698671668768, + 0.01229268778115511, + 0.013682510703802109, + 0.0049703302793204784, + 0.004474487155675888, + 0.013638251461088657, + 0.004488927312195301, + 0.004437214694917202, + 0.00607072189450264, + 0.00925402157008648, + 0.005140235647559166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.01818004623055458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368288651108742 + ], + "opposite_logprob_tokens": [ + -3.4086828231811523 + ], + "logprob_reward": -0.03368288651108742, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368288651108742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16951531171798706, + 0.16520649194717407, + 0.06237345561385155, + 0.08942651003599167, + 0.13031835854053497, + 0.14145229756832123, + 0.06054812669754028, + 0.0683702901005745, + 0.07987295836210251, + 0.03291620314121246 + ], + "credit_map": [ + 0.008704938925802708, + 0.014905673451721668, + 0.0124735776335001, + 0.02704554609954357, + 0.013902499340474606, + 0.039094142615795135, + 0.012290781363844872, + 0.041098155081272125, + 0.03221617266535759, + 0.03195798024535179, + 0.016972998157143593, + 0.012787437066435814, + 0.02166886068880558, + 0.015552773140370846, + 0.018895620480179787, + 0.01515466719865799, + 0.011845205910503864, + 0.008137647062540054, + 0.007433658465743065, + 0.010839893482625484, + 0.003831758862361312, + 0.007037242408841848, + 0.006879409309476614, + 0.006368639413267374, + 0.00975011382251978, + 0.014551668427884579, + 0.007806586567312479, + 0.012427088804543018, + 0.011490702629089355, + 0.010403037071228027, + 0.012279323302209377, + 0.010717993602156639, + 0.006454229820519686, + 0.051260050386190414, + 0.016406647861003876, + 0.00497840391471982, + 0.01318595465272665, + 0.011510932818055153, + 0.0043817488476634026, + 0.022140389308333397, + 0.009309370070695877, + 0.01784665137529373, + 0.021712826564908028, + 0.026174763217568398, + 0.01734483242034912, + 0.015570498071610928, + 0.025674954056739807, + 0.007818399928510189, + 0.0074319541454315186, + 0.008909310214221478, + 0.007290924433618784, + 0.0071167731657624245, + 0.009612232446670532, + 0.00966458860784769, + 0.006169179454445839, + 0.004353162832558155, + 0.01553240418434143, + 0.006039424799382687, + 0.016342606395483017, + 0.006523433141410351, + 0.0043996842578053474, + 0.0070023140870034695, + 0.006600255146622658, + 0.005930165760219097, + 0.009782384149730206, + 0.009673149324953556, + 0.011227617971599102, + 0.00885799340903759, + 0.008599274791777134, + 0.007431816775351763, + 0.011591845192015171, + 0.012708870694041252, + 0.003869847860187292, + 0.00357263907790184, + 0.006955383345484734, + 0.003384452313184738, + 0.003818114520981908, + 0.003894014051184058, + 0.004826327785849571, + 0.002595426980406046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03368288651108742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera executes dynamic rotating shots, spiraling 360-degree barrel rolls, perfectly reflecting the exhilarating and carefree aerial acrobatics of a magnificent avian creature soaring through the cosmos. Its body is composed not of flesh but of shimmering, sun-filled crystal feathers; each feather is a perfectly faceted, translucent prism, capturing and refracting ambient light, spreading a kaleidoscope of color across its body while emitting a soft, inner golden glow. As it joyfully swoops and dives into a vast, vibrant cloud of swirling fuchsia, deep purple, and glowing cyan gas, it playfully chases its tail—a trailing ribbon of ethereal energy woven from sapphire blue and molten gold light particles, leaving a luminous wake. With effortless grace, it treats the floating crystal asteroids like toys, quickly nudging the smaller ones with its beak to send them tumbling away, and weaving in a tight, joyful circle around the larger metallic structure. The entire scene is a dreamlike symphony of light, with the clouds providing a soft, multi-hued glow that is amplified by the countless prismatic points of light shimmering in the creature’s own luminescent tail and crystalline feathers, creating an atmosphere of pure, unadulterated cosmic joy and infinite freedom.", + "seed": 14638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_13/index_2917_seed_14638_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08124116323888302, + "vlm_logprob": -0.08124116323888302, + "vlm_yes_no_margin": 2.899999976158142, + "vlm_token_credit": -0.08124116323888302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01699257083237171 + ], + "opposite_logprob_tokens": [ + -4.141992568969727 + ], + "logprob_reward": -0.01699257083237171, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01699257083237171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1032416969537735, + 0.13345924019813538, + 0.10290273278951645, + 0.08221612125635147, + 0.04708045721054077, + 0.054539941251277924, + 0.07908753305673599, + 0.1590234786272049, + 0.16059261560440063, + 0.07785621285438538 + ], + "credit_map": [ + 0.008296037092804909, + 0.017281023785471916, + 0.01853480562567711, + 0.02857927605509758, + 0.00969779770821333, + 0.005462823435664177, + 0.008133720606565475, + 0.007256207522004843, + 0.014529737643897533, + 0.026094917207956314, + 0.020173287019133568, + 0.03145718201994896, + 0.01189903263002634, + 0.005880239885300398, + 0.01858869194984436, + 0.004836142063140869, + 0.011808821931481361, + 0.014823660254478455, + 0.023037420585751534, + 0.011385363526642323, + 0.007245357614010572, + 0.006595279090106487, + 0.00925390049815178, + 0.018752917647361755, + 0.010643109679222107, + 0.007242611143738031, + 0.010449664667248726, + 0.01839880459010601, + 0.006531490944325924, + 0.004867883864790201, + 0.013083897531032562, + 0.010998650453984737, + 0.00640261871740222, + 0.006407186388969421, + 0.004104936961084604, + 0.007157163228839636, + 0.00321853905916214, + 0.003146317321807146, + 0.008540164679288864, + 0.008103527128696442, + 0.005989440716803074, + 0.009315532632172108, + 0.004457773175090551, + 0.009411844424903393, + 0.004040972795337439, + 0.006382120307534933, + 0.007667974103242159, + 0.007274279836565256, + 0.0066525898873806, + 0.013597109355032444, + 0.004971944261342287, + 0.008786768652498722, + 0.009127200581133366, + 0.008182663470506668, + 0.017556341364979744, + 0.01021291222423315, + 0.012782176956534386, + 0.018021997064352036, + 0.017443837597966194, + 0.022981582209467888, + 0.019495433196425438, + 0.02310274913907051, + 0.019287047907710075, + 0.02590864524245262, + 0.021591048687696457, + 0.022016262635588646, + 0.02021273598074913, + 0.02035355754196644, + 0.016212433576583862, + 0.01521549466997385, + 0.021358627825975418, + 0.02363244630396366, + 0.008117051795125008, + 0.01639782264828682, + 0.009927213191986084, + 0.01145900972187519, + 0.011695273220539093, + 0.0053365807980299, + 0.009564852342009544, + 0.0053584035485982895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.01699257083237171 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078222543001175 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078222543001175, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078222543001175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2184731662273407, + 0.06756556779146194, + 0.07926250249147415, + 0.1251092404127121, + 0.025632651522755623, + 0.21450118720531464, + 0.036855194717645645, + 0.06975676119327545, + 0.06463230401277542, + 0.09821142256259918 + ], + "credit_map": [ + 0.01707099750638008, + 0.029284700751304626, + 0.042465828359127045, + 0.07129470258951187, + 0.01726052723824978, + 0.011259907856583595, + 0.016668086871504784, + 0.013168406672775745, + 0.006239356938749552, + 0.01251520961523056, + 0.007503565866500139, + 0.011735723353922367, + 0.012772207148373127, + 0.005729686934500933, + 0.006369166076183319, + 0.004700649995356798, + 0.011831585317850113, + 0.00850760005414486, + 0.012450985610485077, + 0.00854155607521534, + 0.007423899136483669, + 0.015177104622125626, + 0.007000321988016367, + 0.008329449221491814, + 0.012507405132055283, + 0.009524394758045673, + 0.017790060490369797, + 0.03341832011938095, + 0.006701350212097168, + 0.006658648140728474, + 0.020410973578691483, + 0.01809808239340782, + 0.003382447175681591, + 0.0036997543647885323, + 0.0023101407568901777, + 0.003497879486531019, + 0.0021288562566041946, + 0.002047001849859953, + 0.004381190985441208, + 0.004185377154499292, + 0.024617735296487808, + 0.04150184616446495, + 0.013488381169736385, + 0.02846457250416279, + 0.01689169742166996, + 0.03182080015540123, + 0.02567201480269432, + 0.03204411640763283, + 0.0045368061400949955, + 0.0063957166858017445, + 0.0022727488540112972, + 0.003526023356243968, + 0.005325373262166977, + 0.00532491272315383, + 0.0061214398592710495, + 0.003352169878780842, + 0.00602735998108983, + 0.008433523587882519, + 0.008941627107560635, + 0.009094004519283772, + 0.01293724961578846, + 0.008356421254575253, + 0.00959873665124178, + 0.006367833819240332, + 0.006103255320340395, + 0.010594567283987999, + 0.0072579653933644295, + 0.010577242821455002, + 0.008157143369317055, + 0.005570834036916494, + 0.009686330333352089, + 0.0066849603317677975, + 0.01031988300383091, + 0.021609032526612282, + 0.014352209866046906, + 0.013741603121161461, + 0.014315000735223293, + 0.0064038061536848545, + 0.01024635974317789, + 0.007223521359264851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.016078222543001175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279374480247498 + ], + "opposite_logprob_tokens": [ + -2.017793655395508 + ], + "logprob_reward": -0.14279374480247498, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279374480247498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11014586687088013, + 0.15069599449634552, + 0.16120634973049164, + 0.1796826720237732, + 0.05831228569149971, + 0.08385470509529114, + 0.04326832294464111, + 0.08726242929697037, + 0.05963689461350441, + 0.06593447923660278 + ], + "credit_map": [ + 0.012349224649369717, + 0.013594433665275574, + 0.018802054226398468, + 0.025123078376054764, + 0.009153034538030624, + 0.0066190557554364204, + 0.012490666471421719, + 0.012014318257570267, + 0.014591668732464314, + 0.022832516580820084, + 0.02396550588309765, + 0.02365330420434475, + 0.023477351292967796, + 0.009054573252797127, + 0.02389281615614891, + 0.009228263050317764, + 0.022252732887864113, + 0.022849449887871742, + 0.029894964769482613, + 0.016577119007706642, + 0.012945718131959438, + 0.010986973531544209, + 0.016135400161147118, + 0.029564008116722107, + 0.009282022714614868, + 0.00816367007791996, + 0.018886040896177292, + 0.03439651057124138, + 0.00713866762816906, + 0.007833577692508698, + 0.051169756799936295, + 0.04281243309378624, + 0.005726254545152187, + 0.00862745102494955, + 0.004842032678425312, + 0.010697976686060429, + 0.005014646332710981, + 0.0048798141069710255, + 0.010201376862823963, + 0.008322729729115963, + 0.0068655842915177345, + 0.01172496099025011, + 0.007297626230865717, + 0.013287152163684368, + 0.008029965683817863, + 0.01799989491701126, + 0.009275496006011963, + 0.009374028071761131, + 0.004787232261151075, + 0.006179070565849543, + 0.002064301399514079, + 0.005172320641577244, + 0.008457609452307224, + 0.005680516362190247, + 0.006896561942994595, + 0.004030712880194187, + 0.0075703151524066925, + 0.01059223897755146, + 0.014006967656314373, + 0.011696027591824532, + 0.011598184704780579, + 0.0077138966880738735, + 0.014098867774009705, + 0.00998593494296074, + 0.005725888069719076, + 0.01082839909940958, + 0.006909301038831472, + 0.011075512506067753, + 0.008182100020349026, + 0.0045265923254191875, + 0.00828709825873375, + 0.004101999569684267, + 0.0070618074387311935, + 0.015709813684225082, + 0.010003037750720978, + 0.008945932611823082, + 0.008836043067276478, + 0.004632030613720417, + 0.005854696501046419, + 0.004891117103397846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.14279374480247498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002181559801102 + ], + "opposite_logprob_tokens": [ + -2.695021867752075 + ], + "logprob_reward": -0.07002181559801102, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002181559801102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06143263354897499, + 0.054887380450963974, + 0.09929835796356201, + 0.275472491979599, + 0.0585455447435379, + 0.08289345353841782, + 0.03626243770122528, + 0.0569169819355011, + 0.19234956800937653, + 0.08194112777709961 + ], + "credit_map": [ + 0.006753884255886078, + 0.011348569765686989, + 0.00909966230392456, + 0.012342195026576519, + 0.007743818685412407, + 0.0039040888659656048, + 0.0050056492909789085, + 0.005234770942479372, + 0.00552900368347764, + 0.008023402653634548, + 0.00825971644371748, + 0.007648042403161526, + 0.00930248387157917, + 0.003914023749530315, + 0.008219875395298004, + 0.003990837372839451, + 0.013414910063147545, + 0.015154128894209862, + 0.01908174343407154, + 0.012048008851706982, + 0.007334832567721605, + 0.005153101868927479, + 0.010707269422709942, + 0.01640436239540577, + 0.016804799437522888, + 0.013972393237054348, + 0.028033066540956497, + 0.044743120670318604, + 0.010229705832898617, + 0.01275000348687172, + 0.08283720165491104, + 0.06610222160816193, + 0.006029055919498205, + 0.007555287331342697, + 0.005723609123378992, + 0.013519398868083954, + 0.0035092122852802277, + 0.005048464518040419, + 0.009399550966918468, + 0.0077609708532691, + 0.009106956422328949, + 0.014102296903729439, + 0.004935532342642546, + 0.012427298352122307, + 0.005472227931022644, + 0.01333009172230959, + 0.010367942973971367, + 0.01315111294388771, + 0.00628158962354064, + 0.005019949283450842, + 0.0020408914424479008, + 0.003016327740624547, + 0.005553234368562698, + 0.005064063705503941, + 0.004725288134068251, + 0.004561097361147404, + 0.005182656925171614, + 0.007844032719731331, + 0.008031659759581089, + 0.008493494242429733, + 0.010704440996050835, + 0.003563250182196498, + 0.007541429251432419, + 0.005556023214012384, + 0.02491043694317341, + 0.025393368676304817, + 0.030311977490782738, + 0.023740829899907112, + 0.02271726168692112, + 0.016842829063534737, + 0.020384280011057854, + 0.028048604726791382, + 0.008041813969612122, + 0.01920313760638237, + 0.01261120568960905, + 0.011271788738667965, + 0.010670166462659836, + 0.005612649954855442, + 0.008146529085934162, + 0.006383844185620546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.07002181559801102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an avian creature with shimmering, faceted crystal feathers instead of flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature emit a soft inner golden glow and refract light into a kaleidoscope of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a trailing ribbon of ethereal energy composed of sapphire blue and molten gold light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature interact with floating crystal asteroids by nudging smaller ones with its beak and circling a larger metallic structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute dynamic rotating shots including 360-degree barrel rolls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603194624185562 + ], + "opposite_logprob_tokens": [ + -1.910319447517395 + ], + "logprob_reward": -0.1603194624185562, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603194624185562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15375970304012299, + 0.1646398901939392, + 0.12096398323774338, + 0.09080582857131958, + 0.057194892317056656, + 0.1403137892484665, + 0.09759578853845596, + 0.06613428145647049, + 0.043809011578559875, + 0.06478285789489746 + ], + "credit_map": [ + 0.01794389821588993, + 0.017611831426620483, + 0.02839754894375801, + 0.04252535477280617, + 0.010689370334148407, + 0.007681556511670351, + 0.016327178105711937, + 0.01258294377475977, + 0.01912883296608925, + 0.025106074288487434, + 0.02867130935192108, + 0.032370805740356445, + 0.014468936249613762, + 0.007413272280246019, + 0.03100845217704773, + 0.006472194567322731, + 0.016007810831069946, + 0.011572317220270634, + 0.033941157162189484, + 0.013465670868754387, + 0.007441788911819458, + 0.0074594407342374325, + 0.012884366326034069, + 0.018191423267126083, + 0.007134391926229, + 0.007903234101831913, + 0.011791497468948364, + 0.020890967920422554, + 0.004692973103374243, + 0.005113521590828896, + 0.017389681190252304, + 0.015889544039964676, + 0.00753402803093195, + 0.009316613897681236, + 0.005487796850502491, + 0.011905880644917488, + 0.0040002139285206795, + 0.0050391522236168385, + 0.00708762789145112, + 0.006823573261499405, + 0.012141243554651737, + 0.015989629551768303, + 0.014545026235282421, + 0.036140501499176025, + 0.010054437443614006, + 0.016304079443216324, + 0.016597174108028412, + 0.018541691824793816, + 0.011746070347726345, + 0.017652105540037155, + 0.005144171416759491, + 0.010253467597067356, + 0.010742652229964733, + 0.011244160123169422, + 0.02167474664747715, + 0.009138409048318863, + 0.005936379078775644, + 0.009353948757052422, + 0.007309128064662218, + 0.009477624669671059, + 0.011919844895601273, + 0.0059945411048829556, + 0.009198489598929882, + 0.0069443173706531525, + 0.005258719436824322, + 0.006780615076422691, + 0.004487781319767237, + 0.005756455007940531, + 0.005338903982192278, + 0.0042365663684904575, + 0.005311035551130772, + 0.006638928782194853, + 0.005930144339799881, + 0.012863164767622948, + 0.00724734365940094, + 0.010849479585886002, + 0.008873696438968182, + 0.005904304329305887, + 0.0063537610694766045, + 0.006760964170098305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.1603194624185562 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_tensor([7439, 7439])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_tensor([7439, 7439])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3c99e665f096dcc07c9ac8922c6ec9de68c06019 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_tensor([7439, 7439])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, shot from a level angle, slowly pans forward, focusing on an elderly man with a face deeply wrinkled and furrowed. He wears a thick, dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar. His wispy silver hair gleams softly in the warm light. He sits in an old wooden armchair, a rough stone fireplace behind him. The crackling, orange-red flames in the hearth provide the scene's sole source of light. The flickering flames cast soft, dynamic shadows across his deep wrinkles, accentuating the depths of his dark brown eyes, which shimmer with wisdom and opacity. He leans forward slightly, one hand, covered in age spots, moving slowly and expressively in the air. His lips move slightly as he speaks, leisurely unwinding a forgotten story to an audience whose outlines are blurred and out of focus. The entire picture is shrouded in warm amber light, with the main colors being dark brown and warm orange, creating an intimate atmosphere of extreme tranquility, nostalgia, and wisdom, as if time has frozen at this moment for an ancient story.", + "seed": 2579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_7439_seed_2579_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08148931292816997, + "vlm_logprob": -0.08148931292816997, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.08148931292816997, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011264890432357788 + ], + "opposite_logprob_tokens": [ + -5.636264801025391 + ], + "logprob_reward": -0.011264890432357788, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.011264890432357788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14221201837062836, + 0.165474072098732, + 0.06421281397342682, + 0.11581012606620789, + 0.028324363753199577, + 0.16364789009094238, + 0.05291905626654625, + 0.09509489685297012, + 0.0692855715751648, + 0.10301919281482697 + ], + "credit_map": [ + 0.017190005630254745, + 0.02493865415453911, + 0.010858711786568165, + 0.022253841161727905, + 0.018612604588270187, + 0.018386133015155792, + 0.014366421848535538, + 0.015605645254254341, + 0.020626196637749672, + 0.02427145652472973, + 0.015652915462851524, + 0.033394955098629, + 0.019436370581388474, + 0.02262970246374607, + 0.015337773598730564, + 0.014124706387519836, + 0.007884466089308262, + 0.008640280924737453, + 0.006840397138148546, + 0.007038459647446871, + 0.007075809873640537, + 0.008020637556910515, + 0.00730229914188385, + 0.011410458013415337, + 0.008458628319203854, + 0.016308806836605072, + 0.019926242530345917, + 0.016337092965841293, + 0.015254836529493332, + 0.021642006933689117, + 0.009576956741511822, + 0.00830555148422718, + 0.0022075038868933916, + 0.004822708200663328, + 0.004062148276716471, + 0.004558865446597338, + 0.0033331962767988443, + 0.0047080023214221, + 0.002347489818930626, + 0.002284449990838766, + 0.008319266140460968, + 0.007556131575256586, + 0.043369993567466736, + 0.013653267174959183, + 0.010942663066089153, + 0.061775173991918564, + 0.010014520958065987, + 0.008016876876354218, + 0.003571620909497142, + 0.011627052910625935, + 0.005797555670142174, + 0.011891392059624195, + 0.004255176987498999, + 0.0069575197994709015, + 0.004896783269941807, + 0.003921953029930592, + 0.008154096081852913, + 0.016193458810448647, + 0.0157126747071743, + 0.0155483428388834, + 0.007582085207104683, + 0.015536152757704258, + 0.009740117937326431, + 0.006627965718507767, + 0.0050296662375330925, + 0.01470638532191515, + 0.0077931578271090984, + 0.020317915827035904, + 0.004471194464713335, + 0.007467823568731546, + 0.006519320886582136, + 0.0029800983611494303, + 0.005580094177275896, + 0.008525078184902668, + 0.028526021167635918, + 0.008815648034214973, + 0.005166836082935333, + 0.03298768028616905, + 0.009311016649007797, + 0.004106828477233648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.011264890432357788 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026382071897387505 + ], + "opposite_logprob_tokens": [ + -3.6513819694519043 + ], + "logprob_reward": -0.026382071897387505, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026382071897387505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1345430165529251, + 0.11810210347175598, + 0.07323134690523148, + 0.23937474191188812, + 0.08864544332027435, + 0.19203872978687286, + 0.03150179982185364, + 0.054210834205150604, + 0.04214636981487274, + 0.026205625385046005 + ], + "credit_map": [ + 0.010601168498396873, + 0.02646569348871708, + 0.011899534612894058, + 0.0379968136548996, + 0.013988455757498741, + 0.015402295626699924, + 0.009879331104457378, + 0.008309727534651756, + 0.02117532677948475, + 0.021779930219054222, + 0.011450100690126419, + 0.017735673114657402, + 0.011581096798181534, + 0.011070084758102894, + 0.011424202471971512, + 0.01188568864017725, + 0.008025179617106915, + 0.011411682702600956, + 0.00749792717397213, + 0.010728186927735806, + 0.005349322687834501, + 0.009980606846511364, + 0.0063584670424461365, + 0.01387997530400753, + 0.011473202146589756, + 0.009159417822957039, + 0.0758720114827156, + 0.01639731228351593, + 0.010291339829564095, + 0.09613169729709625, + 0.012198061682283878, + 0.007851717993617058, + 0.006990846246480942, + 0.007447489071637392, + 0.012821867130696774, + 0.009377478621900082, + 0.009313063696026802, + 0.017552968114614487, + 0.008310355246067047, + 0.016831373795866966, + 0.009908145293593407, + 0.009525305591523647, + 0.05110493302345276, + 0.01403980702161789, + 0.011088572442531586, + 0.0724693164229393, + 0.011332862079143524, + 0.012569784186780453, + 0.0024371047038584948, + 0.005836966447532177, + 0.003636930836364627, + 0.00498845474794507, + 0.003109982702881098, + 0.003958286717534065, + 0.003988728858530521, + 0.0035453441087156534, + 0.0040526119992136955, + 0.00966663472354412, + 0.0069809709675610065, + 0.012224821373820305, + 0.0037098575849086046, + 0.006638087332248688, + 0.007193003315478563, + 0.0037448445800691843, + 0.001792127382941544, + 0.003242251230403781, + 0.011289744637906551, + 0.0042862012051045895, + 0.0025058079045265913, + 0.014136402867734432, + 0.002917674370110035, + 0.001976157072931528, + 0.0021776161156594753, + 0.004839009139686823, + 0.0036710083950310946, + 0.004442793317139149, + 0.0023346352390944958, + 0.004090568050742149, + 0.0030824022833257914, + 0.0015675933100283146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.026382071897387505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04872199520468712 + ], + "opposite_logprob_tokens": [ + -3.0487220287323 + ], + "logprob_reward": -0.04872199520468712, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04872199520468712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06819264590740204, + 0.07082698494195938, + 0.12577219307422638, + 0.08204402774572372, + 0.15790586173534393, + 0.19041979312896729, + 0.0894966796040535, + 0.07591290026903152, + 0.04169267788529396, + 0.09773627668619156 + ], + "credit_map": [ + 0.008843367919325829, + 0.01336775440722704, + 0.004910669755190611, + 0.012728470377624035, + 0.006669649388641119, + 0.008126167580485344, + 0.006578044034540653, + 0.0069685280323028564, + 0.014538359828293324, + 0.012112902477383614, + 0.007827946916222572, + 0.008777287788689137, + 0.005179817322641611, + 0.009821916930377483, + 0.006412213202565908, + 0.00615653907880187, + 0.02292700670659542, + 0.01311518158763647, + 0.0166026521474123, + 0.015087404288351536, + 0.0067936694249510765, + 0.019002174958586693, + 0.01018836535513401, + 0.022055748850107193, + 0.004737484268844128, + 0.006423345301300287, + 0.01933247037231922, + 0.0098942955955863, + 0.007165692280977964, + 0.023871809244155884, + 0.004849081858992577, + 0.005769846960902214, + 0.010978542268276215, + 0.012302952818572521, + 0.02617485448718071, + 0.01310963649302721, + 0.015257343649864197, + 0.03830428421497345, + 0.01355606596916914, + 0.028222177177667618, + 0.010154780931770802, + 0.012737243436276913, + 0.046095848083496094, + 0.013632173649966717, + 0.013502448797225952, + 0.06115904077887535, + 0.012469397857785225, + 0.02066885121166706, + 0.006710744462907314, + 0.014927767217159271, + 0.010548960417509079, + 0.014688287861645222, + 0.011261578649282455, + 0.01117286179214716, + 0.011549776419997215, + 0.008636697195470333, + 0.005767231807112694, + 0.011950492858886719, + 0.00921657495200634, + 0.013640720397233963, + 0.0062766121700406075, + 0.014566815458238125, + 0.0069000134244561195, + 0.007594444788992405, + 0.0018993355333805084, + 0.0048330421559512615, + 0.00828141625970602, + 0.007615474984049797, + 0.0032878455240279436, + 0.011050727218389511, + 0.002653077943250537, + 0.002071755938231945, + 0.005963386967778206, + 0.009202338755130768, + 0.023205794394016266, + 0.00851633120328188, + 0.008052517659962177, + 0.02933499775826931, + 0.007669149432331324, + 0.005791760049760342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.04872199520468712 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19147104024887085, + 0.07278314977884293, + 0.051598671823740005, + 0.15797866880893707, + 0.11674211919307709, + 0.12846098840236664, + 0.08945466578006744, + 0.09299030900001526, + 0.058156900107860565, + 0.04036343842744827 + ], + "credit_map": [ + 0.015068099834024906, + 0.0218118354678154, + 0.020970119163393974, + 0.027299536392092705, + 0.02730286680161953, + 0.03138836845755577, + 0.01937108114361763, + 0.02825913019478321, + 0.008889545686542988, + 0.011592783033847809, + 0.0072531611658632755, + 0.010016721673309803, + 0.008025548420846462, + 0.012851768173277378, + 0.00630003260448575, + 0.007853591814637184, + 0.005500652361661196, + 0.00775259081274271, + 0.005508454516530037, + 0.008148682303726673, + 0.007808319758623838, + 0.005432036705315113, + 0.006046587135642767, + 0.005401345435529947, + 0.007754893973469734, + 0.009509062394499779, + 0.04681321978569031, + 0.013624885119497776, + 0.007303390186280012, + 0.05708223953843117, + 0.008614327758550644, + 0.007276658900082111, + 0.009000122547149658, + 0.009496956132352352, + 0.01894921064376831, + 0.011862726882100105, + 0.012111871503293514, + 0.024323759600520134, + 0.011588658206164837, + 0.01940881460905075, + 0.00932283140718937, + 0.010303515940904617, + 0.031171470880508423, + 0.009466860443353653, + 0.0076752458699047565, + 0.04626872390508652, + 0.007185232359915972, + 0.007067110389471054, + 0.007076534908264875, + 0.016966266557574272, + 0.012489707209169865, + 0.015344938263297081, + 0.008472747169435024, + 0.011977316811680794, + 0.008936534635722637, + 0.008190624415874481, + 0.008530751802027225, + 0.010147574357688427, + 0.015317676588892937, + 0.014172650873661041, + 0.011745136231184006, + 0.012832225300371647, + 0.010651613585650921, + 0.009592678397893906, + 0.0031673195771872997, + 0.0069752163253724575, + 0.012881956063210964, + 0.007872120477259159, + 0.003430197248235345, + 0.01641845516860485, + 0.004629846196621656, + 0.002781788818538189, + 0.0037778925616294146, + 0.005901775322854519, + 0.0057770004495978355, + 0.004751675296574831, + 0.003772812895476818, + 0.007964171469211578, + 0.005014070309698582, + 0.003404041286557913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.007630250882357359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344735622406006 + ], + "opposite_logprob_tokens": [ + -1.31344735622406 + ], + "logprob_reward": -0.31344735622406006, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344735622406006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09042411297559738, + 0.047044120728969574, + 0.04176364839076996, + 0.2907607853412628, + 0.041486699134111404, + 0.21887119114398956, + 0.031907759606838226, + 0.09235260635614395, + 0.07952496409416199, + 0.06586415320634842 + ], + "credit_map": [ + 0.009826866909861565, + 0.018054857850074768, + 0.008395583368837833, + 0.015586109831929207, + 0.010088907554745674, + 0.009893998503684998, + 0.006253790110349655, + 0.01232399232685566, + 0.00992803368717432, + 0.007237787824124098, + 0.005182572640478611, + 0.0055335452780127525, + 0.0039521073922514915, + 0.00540147302672267, + 0.00442151166498661, + 0.005387085024267435, + 0.004480297677218914, + 0.005824424792081118, + 0.00462327990680933, + 0.006497771944850683, + 0.004687202628701925, + 0.00525752454996109, + 0.004610420670360327, + 0.00578272296115756, + 0.013160023838281631, + 0.00962629821151495, + 0.09482935816049576, + 0.02061435952782631, + 0.011977216228842735, + 0.11733829230070114, + 0.014671625569462776, + 0.008543591946363449, + 0.00355616700835526, + 0.004466358106583357, + 0.0061409310437738895, + 0.0041176509112119675, + 0.0037363823503255844, + 0.008560522459447384, + 0.0035549351014196873, + 0.007353748194873333, + 0.011243454180657864, + 0.009277584962546825, + 0.06138857081532478, + 0.01593421772122383, + 0.011428957805037498, + 0.08796049654483795, + 0.013222151435911655, + 0.008415739051997662, + 0.0024176822043955326, + 0.00457148440182209, + 0.0037489701062440872, + 0.005101629067212343, + 0.0029080586973577738, + 0.004633489530533552, + 0.004107736516743898, + 0.004418706055730581, + 0.007665736600756645, + 0.018308961763978004, + 0.011501014232635498, + 0.019979745149612427, + 0.006989378947764635, + 0.009816017933189869, + 0.010699446313083172, + 0.007392298895865679, + 0.004155130125582218, + 0.012607564218342304, + 0.015429709106683731, + 0.015506772324442863, + 0.003932258114218712, + 0.01995275169610977, + 0.005629265680909157, + 0.002311506075784564, + 0.0044883619993925095, + 0.00836880225688219, + 0.016781868413090706, + 0.006355559919029474, + 0.003367673372849822, + 0.018460668623447418, + 0.005674517247825861, + 0.0023666962515562773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.31344735622406006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, slowly pans forward, focusing on an elderly man with a face deeply wrinkled and furrowed. He wears a thick, dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar. His wispy silver hair gleams softly in the warm light. He sits in an old wooden armchair, a rough stone fireplace behind him. The crackling, orange-red flames in the hearth provide the scene's sole source of light. The flickering flames cast soft, dynamic shadows across his deep wrinkles, accentuating the depths of his dark brown eyes, which shimmer with wisdom and opacity. He leans forward slightly, one hand, covered in age spots, moving slowly and expressively in the air. His lips move slightly as he speaks, leisurely unwinding a forgotten story to an audience whose outlines are blurred and out of focus. The entire picture is shrouded in warm amber light, with the main colors being dark brown and warm orange, creating an intimate atmosphere of extreme tranquility, nostalgia, and wisdom, as if time has frozen at this moment for an ancient story.", + "seed": 2580, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_7439_seed_2580_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12542441487312317, + "vlm_logprob": -0.12542441487312317, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.12542441487312317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020408054813742638 + ], + "opposite_logprob_tokens": [ + -6.1454081535339355 + ], + "logprob_reward": -0.020408054813742638, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.020408054813742638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15040281414985657, + 0.23588787019252777, + 0.08882912993431091, + 0.06245024502277374, + 0.05543816089630127, + 0.04338379576802254, + 0.16652721166610718, + 0.05859106034040451, + 0.0729934498667717, + 0.06549627333879471 + ], + "credit_map": [ + 0.025678906589746475, + 0.020418792963027954, + 0.011418689973652363, + 0.015063314698636532, + 0.02703624963760376, + 0.011681215837597847, + 0.01721077226102352, + 0.021894874051213264, + 0.06348061561584473, + 0.02963283471763134, + 0.017362890765070915, + 0.025075124576687813, + 0.03224068135023117, + 0.015527535229921341, + 0.013960989192128181, + 0.03860720619559288, + 0.018636899068951607, + 0.009663252159953117, + 0.007403839845210314, + 0.011543278582394123, + 0.014559597708284855, + 0.006184637080878019, + 0.007712685503065586, + 0.013124937191605568, + 0.012949327006936073, + 0.007619002368301153, + 0.005634467117488384, + 0.006432219874113798, + 0.010672047734260559, + 0.0045182243920862675, + 0.005289109423756599, + 0.009335847571492195, + 0.007942362688481808, + 0.006020694971084595, + 0.008921122178435326, + 0.00622705789282918, + 0.009224743582308292, + 0.0038918082136660814, + 0.004520723130553961, + 0.00868965033441782, + 0.00644676573574543, + 0.005715403705835342, + 0.006379321683198214, + 0.004703725688159466, + 0.005727317184209824, + 0.003849233966320753, + 0.004240988288074732, + 0.006321039516478777, + 0.019546352326869965, + 0.02794366516172886, + 0.0180653128772974, + 0.022303052246570587, + 0.021376140415668488, + 0.018347321078181267, + 0.015002978965640068, + 0.023942390456795692, + 0.0064178574830293655, + 0.006213405169546604, + 0.005839410703629255, + 0.006209393031895161, + 0.011006850749254227, + 0.005141582805663347, + 0.010588333010673523, + 0.007174231112003326, + 0.011166688054800034, + 0.007597971707582474, + 0.008197395130991936, + 0.008258750662207603, + 0.009377136826515198, + 0.009088109247386456, + 0.006613883655518293, + 0.012693511322140694, + 0.01388544961810112, + 0.008327960036695004, + 0.005251442082226276, + 0.004696782678365707, + 0.011877442710101604, + 0.00540611008182168, + 0.004910331219434738, + 0.011140749789774418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.020408054813742638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023317189887166023 + ], + "opposite_logprob_tokens": [ + -3.7733170986175537 + ], + "logprob_reward": -0.023317189887166023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317189887166023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20766322314739227, + 0.1791699379682541, + 0.11320295184850693, + 0.1266278624534607, + 0.08469430357217789, + 0.07505211979150772, + 0.0546305850148201, + 0.07104406505823135, + 0.0449756495654583, + 0.04293927177786827 + ], + "credit_map": [ + 0.04468419775366783, + 0.02270115725696087, + 0.01006112340837717, + 0.0239003486931324, + 0.03268095478415489, + 0.013734282925724983, + 0.02945639006793499, + 0.030444767326116562, + 0.04123922437429428, + 0.019338136538863182, + 0.009615100920200348, + 0.018819456920027733, + 0.02800179459154606, + 0.016448289155960083, + 0.017338218167424202, + 0.0283697247505188, + 0.020255859941244125, + 0.011900315061211586, + 0.006882618647068739, + 0.014002852141857147, + 0.019725173711776733, + 0.011520223692059517, + 0.012216894887387753, + 0.01669902913272381, + 0.02615197002887726, + 0.0168538149446249, + 0.01616942696273327, + 0.015005098655819893, + 0.014144943095743656, + 0.01001138985157013, + 0.012528006918728352, + 0.01576322317123413, + 0.012778599746525288, + 0.009239135310053825, + 0.01052333228290081, + 0.008864129893481731, + 0.013732745312154293, + 0.009776311926543713, + 0.006973337847739458, + 0.012806713581085205, + 0.008679731748998165, + 0.014228523708879948, + 0.007074630819261074, + 0.008183978497982025, + 0.013615983538329601, + 0.006291666533797979, + 0.006997654214501381, + 0.009979949332773685, + 0.006230036728084087, + 0.006822542753070593, + 0.008669650182127953, + 0.008686174638569355, + 0.006944123189896345, + 0.004757129121571779, + 0.004740770440548658, + 0.007780161686241627, + 0.012583053670823574, + 0.01001383364200592, + 0.008240770548582077, + 0.006869399454444647, + 0.010609950870275497, + 0.00571027584373951, + 0.006489006336778402, + 0.010527776554226875, + 0.00790302362293005, + 0.006931455805897713, + 0.004702783655375242, + 0.004610849544405937, + 0.006028508301824331, + 0.004336495418101549, + 0.004664779175072908, + 0.005797753110527992, + 0.0060443272814154625, + 0.0042903427965939045, + 0.005738517735153437, + 0.0035091531462967396, + 0.009794574230909348, + 0.0029579014517366886, + 0.005149769596755505, + 0.005454687401652336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023317189887166023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357226073741913 + ], + "opposite_logprob_tokens": [ + -3.6513571739196777 + ], + "logprob_reward": -0.026357226073741913, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357226073741913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24860136210918427, + 0.2042454034090042, + 0.07667122781276703, + 0.07550041377544403, + 0.08496925234794617, + 0.053009048104286194, + 0.06310921162366867, + 0.10140717029571533, + 0.05167483910918236, + 0.040812041610479355 + ], + "credit_map": [ + 0.03978089988231659, + 0.024162715300917625, + 0.027898060157895088, + 0.03228276968002319, + 0.03553454205393791, + 0.01744949445128441, + 0.03288920596241951, + 0.038603655993938446, + 0.028651731088757515, + 0.022562598809599876, + 0.027468020096421242, + 0.022299867123365402, + 0.027511216700077057, + 0.022424446418881416, + 0.02000252716243267, + 0.033324986696243286, + 0.014093706384301186, + 0.011096256785094738, + 0.00548831932246685, + 0.009231014177203178, + 0.012383661232888699, + 0.008336012251675129, + 0.005559142678976059, + 0.01048310473561287, + 0.010915170423686504, + 0.007675413507968187, + 0.012136670760810375, + 0.010770399123430252, + 0.008545341901481152, + 0.0060806432738900185, + 0.007283944170922041, + 0.012092825025320053, + 0.009779486805200577, + 0.009219115599989891, + 0.01699632965028286, + 0.008934910409152508, + 0.013386181555688381, + 0.006558467634022236, + 0.007329824846237898, + 0.01276493165642023, + 0.006589096505194902, + 0.007101552560925484, + 0.008942323736846447, + 0.00645345076918602, + 0.008406859822571278, + 0.003621389390900731, + 0.005078440997749567, + 0.00681593082845211, + 0.006788260769098997, + 0.010157570242881775, + 0.010699987411499023, + 0.007748651783913374, + 0.009244703687727451, + 0.005525392945855856, + 0.005384603515267372, + 0.007560044061392546, + 0.016692029312253, + 0.012079164385795593, + 0.012022477574646473, + 0.011975826695561409, + 0.011860178783535957, + 0.013551213778555393, + 0.010508256033062935, + 0.012718028388917446, + 0.00651798490434885, + 0.004124687053263187, + 0.007444489281624556, + 0.005073672626167536, + 0.007402338087558746, + 0.008870964869856834, + 0.004948044195771217, + 0.007292656227946281, + 0.007686462718993425, + 0.004109558649361134, + 0.004863295238465071, + 0.003689523320645094, + 0.005214447621256113, + 0.0046429201029241085, + 0.0034137889742851257, + 0.007192046847194433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.026357226073741913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370259702205658 + ], + "opposite_logprob_tokens": [ + -3.4087026119232178 + ], + "logprob_reward": -0.03370259702205658, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370259702205658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20259739458560944, + 0.07321484386920929, + 0.04855592921376228, + 0.34689566493034363, + 0.036702580749988556, + 0.08393531292676926, + 0.05654945597052574, + 0.06345494836568832, + 0.03312598541378975, + 0.054967861622571945 + ], + "credit_map": [ + 0.05144418403506279, + 0.028844688087701797, + 0.01552187092602253, + 0.023517638444900513, + 0.02083723433315754, + 0.01230440940707922, + 0.019574599340558052, + 0.03055277094244957, + 0.017111551016569138, + 0.009629609063267708, + 0.005704062525182962, + 0.0070610372349619865, + 0.0095768291503191, + 0.0068037305027246475, + 0.005717010702937841, + 0.011611011810600758, + 0.007240243721753359, + 0.006870356388390064, + 0.004231816623359919, + 0.0076215374283492565, + 0.0053428360261023045, + 0.005726528819650412, + 0.004454912152141333, + 0.007067699916660786, + 0.08143395930528641, + 0.050397325307130814, + 0.04720180109143257, + 0.03980925679206848, + 0.03275182470679283, + 0.02175748720765114, + 0.025970948860049248, + 0.04757304862141609, + 0.006063146982342005, + 0.0032266522757709026, + 0.008209739811718464, + 0.0034729703329503536, + 0.005142358597368002, + 0.0029462028760463, + 0.0029659178107976913, + 0.004675596486777067, + 0.010712615214288235, + 0.00871449988335371, + 0.014700403437018394, + 0.010675765573978424, + 0.011123168282210827, + 0.007916470058262348, + 0.007129015866667032, + 0.012963383458554745, + 0.006272328086197376, + 0.005169597454369068, + 0.01239047758281231, + 0.007458970416337252, + 0.0057042669504880905, + 0.0061906492337584496, + 0.004732953384518623, + 0.008630212396383286, + 0.011699449270963669, + 0.010070730932056904, + 0.006477073300629854, + 0.0060775382444262505, + 0.00978382769972086, + 0.005352240987122059, + 0.006417065858840942, + 0.007577022537589073, + 0.0051828972063958645, + 0.0029764093924313784, + 0.0035287130158394575, + 0.003400317393243313, + 0.003884346457198262, + 0.004709156695753336, + 0.0023925136774778366, + 0.007051633205264807, + 0.010299693793058395, + 0.007391240447759628, + 0.006079305429011583, + 0.0045780944637954235, + 0.00844691600650549, + 0.004914434161037207, + 0.006720609497278929, + 0.0065375701524317265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.03370259702205658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233370065689087 + ], + "opposite_logprob_tokens": [ + -0.8983370065689087 + ], + "logprob_reward": -0.5233370065689087, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233370065689087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1331554353237152, + 0.10121355205774307, + 0.08077869564294815, + 0.05422721058130264, + 0.1265711784362793, + 0.0724448636174202, + 0.19759997725486755, + 0.0720084011554718, + 0.11481592804193497, + 0.047184817492961884 + ], + "credit_map": [ + 0.018748115748167038, + 0.014474845491349697, + 0.01323662605136633, + 0.016681013628840446, + 0.0238112173974514, + 0.013805964961647987, + 0.015242251567542553, + 0.017155399546027184, + 0.012932298704981804, + 0.011372137814760208, + 0.00747652305290103, + 0.011734124273061752, + 0.019647885113954544, + 0.010004722513258457, + 0.01314610242843628, + 0.014899753965437412, + 0.010536963120102882, + 0.009017080999910831, + 0.004265488591045141, + 0.009050711058080196, + 0.014695530757308006, + 0.013702948577702045, + 0.008450398221611977, + 0.011059576645493507, + 0.006638922728598118, + 0.006479267962276936, + 0.005231386050581932, + 0.008053974248468876, + 0.010238444432616234, + 0.004616459831595421, + 0.004330319352447987, + 0.00863843597471714, + 0.01862887106835842, + 0.011212441138923168, + 0.0319807343184948, + 0.011217127554118633, + 0.01691906899213791, + 0.010541893541812897, + 0.009869758039712906, + 0.016201278194785118, + 0.01044766791164875, + 0.009923547506332397, + 0.011433851905167103, + 0.008230977691709995, + 0.01161312498152256, + 0.0051556546241045, + 0.006737628020346165, + 0.008902407251298428, + 0.024911409243941307, + 0.024846406653523445, + 0.027869008481502533, + 0.023235175758600235, + 0.022047095000743866, + 0.024768803268671036, + 0.013726236298680305, + 0.03619584068655968, + 0.004887137562036514, + 0.008583209477365017, + 0.005737314932048321, + 0.009955228306353092, + 0.013113423250615597, + 0.005082103423774242, + 0.014900190755724907, + 0.009749793447554111, + 0.0144773218780756, + 0.01568908430635929, + 0.01364112738519907, + 0.013193383812904358, + 0.019575657323002815, + 0.013232694938778877, + 0.010233397595584393, + 0.014773250557482243, + 0.0076766544952988625, + 0.005763248540461063, + 0.005616871640086174, + 0.0035628259647637606, + 0.00862184539437294, + 0.004064235370606184, + 0.005693928338587284, + 0.0061852093786001205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.5233370065689087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, slowly pans forward, focusing on an elderly man with a face deeply wrinkled and furrowed. He wears a thick, dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar. His wispy silver hair gleams softly in the warm light. He sits in an old wooden armchair, a rough stone fireplace behind him. The crackling, orange-red flames in the hearth provide the scene's sole source of light. The flickering flames cast soft, dynamic shadows across his deep wrinkles, accentuating the depths of his dark brown eyes, which shimmer with wisdom and opacity. He leans forward slightly, one hand, covered in age spots, moving slowly and expressively in the air. His lips move slightly as he speaks, leisurely unwinding a forgotten story to an audience whose outlines are blurred and out of focus. The entire picture is shrouded in warm amber light, with the main colors being dark brown and warm orange, creating an intimate atmosphere of extreme tranquility, nostalgia, and wisdom, as if time has frozen at this moment for an ancient story.", + "seed": 2581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_7439_seed_2581_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1429013816639781, + "vlm_logprob": -0.1429013816639781, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.1429013816639781, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011754064820706844 + ], + "opposite_logprob_tokens": [ + -5.511754035949707 + ], + "logprob_reward": -0.011754064820706844, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.011754064820706844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30721402168273926, + 0.07687170803546906, + 0.061789460480213165, + 0.08649888634681702, + 0.08127082884311676, + 0.08576873689889908, + 0.0818149521946907, + 0.057695258408784866, + 0.054710645228624344, + 0.10636546462774277 + ], + "credit_map": [ + 0.030502639710903168, + 0.05736450105905533, + 0.047912076115608215, + 0.047611746937036514, + 0.03879770264029503, + 0.0332324281334877, + 0.028344178572297096, + 0.0234487596899271, + 0.007967975921928883, + 0.01690981723368168, + 0.010354140773415565, + 0.00934799388051033, + 0.00906192883849144, + 0.009514207020401955, + 0.007452274672687054, + 0.006263371556997299, + 0.006023171357810497, + 0.011929702013731003, + 0.009781279601156712, + 0.005914000794291496, + 0.004828039091080427, + 0.011747080832719803, + 0.005057739559561014, + 0.006508448626846075, + 0.011996856890618801, + 0.010398435406386852, + 0.013395717367529869, + 0.009298204444348812, + 0.005749035160988569, + 0.017643198370933533, + 0.007331711705774069, + 0.010685727000236511, + 0.008399421349167824, + 0.013435082510113716, + 0.01299263071268797, + 0.00899647455662489, + 0.006202362943440676, + 0.016226235777139664, + 0.007027328945696354, + 0.007991296239197254, + 0.007260379381477833, + 0.01299209427088499, + 0.013051589950919151, + 0.011635378003120422, + 0.00880788266658783, + 0.016498716548085213, + 0.008365017361938953, + 0.007157677318900824, + 0.00639967480674386, + 0.011608729138970375, + 0.011077129282057285, + 0.013430507853627205, + 0.007795893587172031, + 0.013792607001960278, + 0.01004558801651001, + 0.0076648215763270855, + 0.005576537922024727, + 0.010330503806471825, + 0.007459959480911493, + 0.0147840715944767, + 0.004231944214552641, + 0.004448033403605223, + 0.006467488594353199, + 0.004396721720695496, + 0.0048751928843557835, + 0.006629562005400658, + 0.009978401474654675, + 0.005201546475291252, + 0.004034105688333511, + 0.008683599531650543, + 0.007661682553589344, + 0.0076465546153485775, + 0.009764105081558228, + 0.0232248492538929, + 0.016962658613920212, + 0.020811211317777634, + 0.009737798012793064, + 0.0064703552052378654, + 0.0145412627607584, + 0.0048532248474657536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.011754064820706844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331520989537239 + ], + "opposite_logprob_tokens": [ + -3.773315191268921 + ], + "logprob_reward": -0.02331520989537239, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331520989537239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21842002868652344, + 0.18505248427391052, + 0.05056368187069893, + 0.06523800641298294, + 0.16996228694915771, + 0.13428549468517303, + 0.02327939309179783, + 0.05073367804288864, + 0.041116517037153244, + 0.06134846433997154 + ], + "credit_map": [ + 0.023778272792696953, + 0.0330621637403965, + 0.0321064218878746, + 0.03200497105717659, + 0.01994919404387474, + 0.04279540851712227, + 0.016267409548163414, + 0.01845618709921837, + 0.018487688153982162, + 0.03334267437458038, + 0.02543586678802967, + 0.027558185160160065, + 0.018808461725711823, + 0.02529679425060749, + 0.017552325502038002, + 0.018570492044091225, + 0.0056238919496536255, + 0.007487723138183355, + 0.006596547085791826, + 0.0057899910025298595, + 0.004871025215834379, + 0.011188691481947899, + 0.005224908702075481, + 0.00378090376034379, + 0.0064092716202139854, + 0.0084719592705369, + 0.009852016344666481, + 0.006308870390057564, + 0.006574674509465694, + 0.016913019120693207, + 0.003973048180341721, + 0.006735140923410654, + 0.0159876961261034, + 0.03384348005056381, + 0.023375049233436584, + 0.03731474652886391, + 0.012462880462408066, + 0.01843421347439289, + 0.016986696049571037, + 0.011557532474398613, + 0.014226271770894527, + 0.009574680589139462, + 0.021503455936908722, + 0.00918433628976345, + 0.009825223125517368, + 0.04042347893118858, + 0.012041767127811909, + 0.01750629022717476, + 0.0023202933371067047, + 0.0027643071953207254, + 0.0037240132223814726, + 0.0020768989343196154, + 0.0021120349410921335, + 0.005982865579426289, + 0.0023929609451442957, + 0.0019060190534219146, + 0.004972438793629408, + 0.006409613881260157, + 0.007782839238643646, + 0.0047900062054395676, + 0.00632515037432313, + 0.008668648079037666, + 0.00772741949185729, + 0.004057561047375202, + 0.0032672688830643892, + 0.006116311065852642, + 0.005667793098837137, + 0.005229185335338116, + 0.0044374181888997555, + 0.008072230033576488, + 0.004121128004044294, + 0.004205185454338789, + 0.0052295564673841, + 0.009951234795153141, + 0.012396401725709438, + 0.009159451350569725, + 0.005310154519975185, + 0.0048127383925020695, + 0.010765910148620605, + 0.0037230183370411396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02331520989537239 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233097355812788 + ], + "opposite_logprob_tokens": [ + -3.7733097076416016 + ], + "logprob_reward": -0.0233097355812788, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233097355812788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16672083735466003, + 0.12313933670520782, + 0.0913073793053627, + 0.2099766731262207, + 0.11344562470912933, + 0.11476648598909378, + 0.06867215782403946, + 0.029650578275322914, + 0.03586304932832718, + 0.04645790159702301 + ], + "credit_map": [ + 0.015231489203870296, + 0.02500869520008564, + 0.025878477841615677, + 0.02055959962308407, + 0.01533294003456831, + 0.033004384487867355, + 0.01758434809744358, + 0.014120873995125294, + 0.011627805419266224, + 0.016831012442708015, + 0.019607597962021828, + 0.01003472600132227, + 0.013232607394456863, + 0.03062259778380394, + 0.009559051133692265, + 0.011623935773968697, + 0.009890761226415634, + 0.013273297809064388, + 0.015649108216166496, + 0.008999302983283997, + 0.00858609564602375, + 0.02103997953236103, + 0.0069833253510296345, + 0.006885501090437174, + 0.023748859763145447, + 0.016363728791475296, + 0.03654882684350014, + 0.013519247993826866, + 0.017286088317632675, + 0.07132437080144882, + 0.01344028115272522, + 0.017745260149240494, + 0.010153641924262047, + 0.009217733517289162, + 0.019099490717053413, + 0.00820700079202652, + 0.009120403788983822, + 0.03718607500195503, + 0.008913076482713223, + 0.011548195034265518, + 0.009240388870239258, + 0.016452018171548843, + 0.017583105713129044, + 0.013093370944261551, + 0.009804174304008484, + 0.016297241672873497, + 0.015279693529009819, + 0.01701648347079754, + 0.006887795869261026, + 0.006044262554496527, + 0.011566609144210815, + 0.003969099838286638, + 0.005234861746430397, + 0.022227613255381584, + 0.005201392807066441, + 0.007540519814938307, + 0.0035148542374372482, + 0.004802650306373835, + 0.0041396599262952805, + 0.004190016072243452, + 0.0027829590253531933, + 0.004267185460776091, + 0.0035484149120748043, + 0.0024048348423093557, + 0.0028540119528770447, + 0.005392180755734444, + 0.005047653801739216, + 0.0034054871648550034, + 0.00274331821128726, + 0.00744620943441987, + 0.004841573070734739, + 0.004132612142711878, + 0.004293234087526798, + 0.00596580421552062, + 0.010205804370343685, + 0.005798203870654106, + 0.003311825217679143, + 0.004753930494189262, + 0.009550518356263638, + 0.0025785784237086773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0233097355812788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294596001505852 + ], + "opposite_logprob_tokens": [ + -3.773294687271118 + ], + "logprob_reward": -0.023294596001505852, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294596001505852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23413768410682678, + 0.08163657784461975, + 0.06020805239677429, + 0.15335677564144135, + 0.04454929009079933, + 0.06255090981721878, + 0.14557388424873352, + 0.02648012526333332, + 0.11094220727682114, + 0.0805644616484642 + ], + "credit_map": [ + 0.021010033786296844, + 0.028044313192367554, + 0.030735976994037628, + 0.023334577679634094, + 0.02489645779132843, + 0.04946718364953995, + 0.02333281747996807, + 0.03331629931926727, + 0.0077883400954306126, + 0.013318709097802639, + 0.011009958572685719, + 0.007421665359288454, + 0.007832213304936886, + 0.016047196462750435, + 0.0074135009199380875, + 0.01080499030649662, + 0.00727803772315383, + 0.007592464331537485, + 0.008018523454666138, + 0.006125577259808779, + 0.005000261124223471, + 0.011003381572663784, + 0.004903236869722605, + 0.010286571457982063, + 0.014327393844723701, + 0.015511671081185341, + 0.02668859250843525, + 0.01330872718244791, + 0.011077160015702248, + 0.047133807092905045, + 0.00977421086281538, + 0.01553521491587162, + 0.004273889120668173, + 0.005343729630112648, + 0.006821034010499716, + 0.0045592933893203735, + 0.004161912947893143, + 0.012314990162849426, + 0.003584355115890503, + 0.0034900850150734186, + 0.006506211590021849, + 0.012067252770066261, + 0.009984740056097507, + 0.006747357081621885, + 0.005441519897431135, + 0.01041040662676096, + 0.006713794078677893, + 0.004679629113525152, + 0.017917094752192497, + 0.015443860553205013, + 0.022164512425661087, + 0.010883110575377941, + 0.011709660291671753, + 0.03670882806181908, + 0.012199140153825283, + 0.018547669053077698, + 0.0019670024048537016, + 0.004586246330291033, + 0.004015309736132622, + 0.003328981576487422, + 0.002425906714051962, + 0.00349890417419374, + 0.004397075157612562, + 0.0022607003338634968, + 0.010171668604016304, + 0.011980346404016018, + 0.014319631271064281, + 0.01014991756528616, + 0.008251657709479332, + 0.02284945547580719, + 0.01345040276646614, + 0.019769126549363136, + 0.007694715168327093, + 0.007268691435456276, + 0.015303222462534904, + 0.008127152919769287, + 0.005170631222426891, + 0.013689008541405201, + 0.015601681545376778, + 0.007709357887506485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.023294596001505852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328333020210266 + ], + "opposite_logprob_tokens": [ + -0.7578333020210266 + ], + "logprob_reward": -0.6328333020210266, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328333020210266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29788196086883545, + 0.10747069120407104, + 0.03188701346516609, + 0.06803479045629501, + 0.07195103913545609, + 0.08826738595962524, + 0.09542452543973923, + 0.08374727517366409, + 0.07300805300474167, + 0.08232733607292175 + ], + "credit_map": [ + 0.03134718909859657, + 0.045631177723407745, + 0.048200905323028564, + 0.04415925219655037, + 0.03353840485215187, + 0.046445902436971664, + 0.024763770401477814, + 0.0237953532487154, + 0.010294117033481598, + 0.023434922099113464, + 0.014641659334301949, + 0.020092567428946495, + 0.013223598711192608, + 0.00935713667422533, + 0.008075149729847908, + 0.008351525291800499, + 0.0035563891287893057, + 0.0053882659412920475, + 0.00443264190107584, + 0.005056878551840782, + 0.0031379566062241793, + 0.005302958190441132, + 0.0023720816243439913, + 0.0026398368645459414, + 0.006246187724173069, + 0.010296090506017208, + 0.010489385575056076, + 0.008338126353919506, + 0.0067024328745901585, + 0.015289687551558018, + 0.0035865779500454664, + 0.007086292374879122, + 0.008078859187662601, + 0.009657514281570911, + 0.01205646526068449, + 0.006744770333170891, + 0.006975048221647739, + 0.014890329912304878, + 0.007830976508557796, + 0.005717070773243904, + 0.006889887619763613, + 0.012207897379994392, + 0.011935122311115265, + 0.013244280591607094, + 0.00719195744022727, + 0.01355094462633133, + 0.009971234947443008, + 0.013276050798594952, + 0.009141704998910427, + 0.01417368184775114, + 0.014412714168429375, + 0.016640111804008484, + 0.010540329851210117, + 0.010853702202439308, + 0.01319105364382267, + 0.006471221335232258, + 0.008961808867752552, + 0.010807526297867298, + 0.012675734236836433, + 0.008436646312475204, + 0.007599107455462217, + 0.016468556597828865, + 0.010208276100456715, + 0.008589602075517178, + 0.005149433854967356, + 0.014819294214248657, + 0.009580815210938454, + 0.007076822221279144, + 0.004803783725947142, + 0.013144226744771004, + 0.00812444742769003, + 0.010309220291674137, + 0.007695973850786686, + 0.015299119986593723, + 0.010685116052627563, + 0.014433348551392555, + 0.009728529490530491, + 0.0073248050175607204, + 0.009197433479130268, + 0.007963001728057861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.6328333020210266 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, slowly pans forward, focusing on an elderly man with a face deeply wrinkled and furrowed. He wears a thick, dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar. His wispy silver hair gleams softly in the warm light. He sits in an old wooden armchair, a rough stone fireplace behind him. The crackling, orange-red flames in the hearth provide the scene's sole source of light. The flickering flames cast soft, dynamic shadows across his deep wrinkles, accentuating the depths of his dark brown eyes, which shimmer with wisdom and opacity. He leans forward slightly, one hand, covered in age spots, moving slowly and expressively in the air. His lips move slightly as he speaks, leisurely unwinding a forgotten story to an audience whose outlines are blurred and out of focus. The entire picture is shrouded in warm amber light, with the main colors being dark brown and warm orange, creating an intimate atmosphere of extreme tranquility, nostalgia, and wisdom, as if time has frozen at this moment for an ancient story.", + "seed": 2582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_14/index_7439_seed_2582_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10372998416423798, + "vlm_logprob": -0.10372998416423798, + "vlm_yes_no_margin": 3.6499999880790712, + "vlm_token_credit": -0.10372998416423798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014497773721814156 + ], + "opposite_logprob_tokens": [ + -6.264497756958008 + ], + "logprob_reward": -0.014497773721814156, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.014497773721814156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25983309745788574, + 0.08903414756059647, + 0.03773140534758568, + 0.13367612659931183, + 0.0521562322974205, + 0.038866233080625534, + 0.04469408839941025, + 0.19816364347934723, + 0.04257607460021973, + 0.10326901078224182 + ], + "credit_map": [ + 0.019236037507653236, + 0.012805474922060966, + 0.010837954469025135, + 0.01541957352310419, + 0.02169385924935341, + 0.10114908963441849, + 0.016034668311476707, + 0.0626564472913742, + 0.009994671680033207, + 0.007961825467646122, + 0.004778287373483181, + 0.005907377228140831, + 0.013817297294735909, + 0.022081540897488594, + 0.005918816663324833, + 0.018574323505163193, + 0.008297459222376347, + 0.003588350722566247, + 0.0033109155483543873, + 0.004345480352640152, + 0.004795045591890812, + 0.004019679501652718, + 0.005535718519240618, + 0.0038387582171708345, + 0.010419909842312336, + 0.007786965928971767, + 0.0066597796976566315, + 0.008630416356027126, + 0.012268921360373497, + 0.05030956119298935, + 0.006816658657044172, + 0.030783919617533684, + 0.005638527683913708, + 0.00643526716157794, + 0.003717908402904868, + 0.005246473476290703, + 0.010166749358177185, + 0.006551277358084917, + 0.0058583226054906845, + 0.008541705086827278, + 0.009595056064426899, + 0.004275943618267775, + 0.003251690184697509, + 0.003704676404595375, + 0.004976035561412573, + 0.004762187134474516, + 0.0036511728540062904, + 0.004649471025913954, + 0.00948036927729845, + 0.0041590104810893536, + 0.004367252811789513, + 0.0033218194730579853, + 0.00459211366251111, + 0.00863419845700264, + 0.0033936710096895695, + 0.006745653692632914, + 0.03108864463865757, + 0.012304011732339859, + 0.011118331924080849, + 0.012380652129650116, + 0.013676070608198643, + 0.06531628966331482, + 0.0068926080130040646, + 0.04538702592253685, + 0.008562405593693256, + 0.004929844755679369, + 0.004640134982764721, + 0.004648865666240454, + 0.005454943515360355, + 0.00667656259611249, + 0.0028359179850667715, + 0.004827396012842655, + 0.010314098559319973, + 0.01464640349149704, + 0.009240953251719475, + 0.014440934173762798, + 0.015449046157300472, + 0.0161708015948534, + 0.011415990069508553, + 0.011590785346925259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.014497773721814156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016094880178570747 + ], + "opposite_logprob_tokens": [ + -4.14109468460083 + ], + "logprob_reward": -0.016094880178570747, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016094880178570747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10626555979251862, + 0.06839362531900406, + 0.11625396460294724, + 0.13573193550109863, + 0.07464439421892166, + 0.033282194286584854, + 0.04502786323428154, + 0.270620197057724, + 0.03668682649731636, + 0.11309350281953812 + ], + "credit_map": [ + 0.008338776417076588, + 0.01305500790476799, + 0.008254796266555786, + 0.010668111965060234, + 0.012449539266526699, + 0.021138852462172508, + 0.016253221780061722, + 0.01610725186765194, + 0.006255527026951313, + 0.008093993179500103, + 0.003924672491848469, + 0.006440053228288889, + 0.016204770654439926, + 0.011189202778041363, + 0.0079877320677042, + 0.008297673426568508, + 0.03345160558819771, + 0.009537349455058575, + 0.009878912940621376, + 0.013484871946275234, + 0.011778563261032104, + 0.012305566109716892, + 0.010757429525256157, + 0.015059665776789188, + 0.010195937938988209, + 0.005651907995343208, + 0.00591984111815691, + 0.008998159319162369, + 0.01099013164639473, + 0.05386743322014809, + 0.006064566317945719, + 0.03404396399855614, + 0.014805028215050697, + 0.006879310589283705, + 0.005531760863959789, + 0.006839648354798555, + 0.012125516310334206, + 0.009710662066936493, + 0.00911948923021555, + 0.009632977657020092, + 0.005159955006092787, + 0.0028362188022583723, + 0.002251050667837262, + 0.0034505254589021206, + 0.005298776086419821, + 0.00506715988740325, + 0.004222100600600243, + 0.0049964068457484245, + 0.0039080302231013775, + 0.005467391572892666, + 0.002950366586446762, + 0.004449005238711834, + 0.0047997781075537205, + 0.009249331429600716, + 0.004151847213506699, + 0.010052111931145191, + 0.022267933934926987, + 0.017598018050193787, + 0.011831965297460556, + 0.016761846840381622, + 0.02325093373656273, + 0.09874102473258972, + 0.009175361134111881, + 0.07099311053752899, + 0.006563532631844282, + 0.005198969971388578, + 0.00345408683642745, + 0.0040185339748859406, + 0.003932508639991283, + 0.004973380360752344, + 0.004700167570263147, + 0.003845644649118185, + 0.010374514386057854, + 0.015018151141703129, + 0.009811167605221272, + 0.012190822511911392, + 0.019514529034495354, + 0.021304242312908173, + 0.012652954086661339, + 0.012227118946611881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.016094880178570747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308224096894264 + ], + "opposite_logprob_tokens": [ + -3.1680822372436523 + ], + "logprob_reward": -0.04308224096894264, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308224096894264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19493477046489716, + 0.11081968247890472, + 0.027673492208123207, + 0.2980569899082184, + 0.0179537795484066, + 0.04794662818312645, + 0.07061818242073059, + 0.13100208342075348, + 0.060361482203006744, + 0.04063286632299423 + ], + "credit_map": [ + 0.022893313318490982, + 0.013913862407207489, + 0.014005962759256363, + 0.01146908663213253, + 0.01745879091322422, + 0.052582550793886185, + 0.018695713952183723, + 0.0439155139029026, + 0.01237411703914404, + 0.008062302134931087, + 0.005720878019928932, + 0.0070984200574457645, + 0.01586812548339367, + 0.031099587678909302, + 0.008153701201081276, + 0.022442560642957687, + 0.0043702507391572, + 0.002283938927575946, + 0.003279258729889989, + 0.0044543021358549595, + 0.0037011480890214443, + 0.003077652771025896, + 0.0030709593556821346, + 0.003435984021052718, + 0.019885262474417686, + 0.012337948195636272, + 0.011102314107120037, + 0.018508702516555786, + 0.024807950481772423, + 0.12066146731376648, + 0.012894844636321068, + 0.07785852998495102, + 0.0018798061646521091, + 0.0019560083746910095, + 0.0020249856170266867, + 0.0019351189257577062, + 0.002983573591336608, + 0.0019985598046332598, + 0.003272061934694648, + 0.0019036676967516541, + 0.00577209610491991, + 0.00820018071681261, + 0.005041380412876606, + 0.006894029211252928, + 0.00648553716018796, + 0.005748325493186712, + 0.004891942720860243, + 0.00491314148530364, + 0.016457386314868927, + 0.00837999302893877, + 0.008890248835086823, + 0.00607784790918231, + 0.007439187727868557, + 0.008213947527110577, + 0.0048251403495669365, + 0.010334436781704426, + 0.012770713306963444, + 0.009424393996596336, + 0.006275110878050327, + 0.008470118045806885, + 0.013061732053756714, + 0.04394051805138588, + 0.005253258161246777, + 0.03180624544620514, + 0.01012518908828497, + 0.010390663519501686, + 0.005878243129700422, + 0.005349694285541773, + 0.0072458721697330475, + 0.009825658984482288, + 0.004488544072955847, + 0.007057618349790573, + 0.0040200576186180115, + 0.005230029579252005, + 0.003866152372211218, + 0.006396517623215914, + 0.0046197413466870785, + 0.005196044221520424, + 0.006339901592582464, + 0.0049644229002296925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.04308224096894264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32118427753448486, + 0.14653076231479645, + 0.09834437817335129, + 0.07882355898618698, + 0.027213267982006073, + 0.029353657737374306, + 0.049560900777578354, + 0.04412638768553734, + 0.1455601304769516, + 0.05930261313915253 + ], + "credit_map": [ + 0.027205077931284904, + 0.0391877107322216, + 0.012875222600996494, + 0.018396444618701935, + 0.03627634420990944, + 0.10358426719903946, + 0.021686265245079994, + 0.06197298318147659, + 0.01703357696533203, + 0.01095294114202261, + 0.007079722359776497, + 0.011142528615891933, + 0.02269085869193077, + 0.037361159920692444, + 0.011390251107513905, + 0.02887973003089428, + 0.028954945504665375, + 0.00928469467908144, + 0.008566384203732014, + 0.010459888726472855, + 0.010183477774262428, + 0.011735803447663784, + 0.008154544979333878, + 0.011004643514752388, + 0.014291735365986824, + 0.007487276569008827, + 0.00977255031466484, + 0.008814224973320961, + 0.0071247052401304245, + 0.015062768943607807, + 0.006107022520154715, + 0.010163277387619019, + 0.003946290351450443, + 0.004321021493524313, + 0.0022876798175275326, + 0.0026737358421087265, + 0.004462272860109806, + 0.0033850267063826323, + 0.0031250701285898685, + 0.0030121724121272564, + 0.007353026885539293, + 0.0037063828203827143, + 0.003090390469878912, + 0.0025461402256041765, + 0.0033152010291814804, + 0.003213834250345826, + 0.0021215432789176702, + 0.004007141105830669, + 0.005966947879642248, + 0.004324664361774921, + 0.0034872880205512047, + 0.003621638985350728, + 0.0036097723059356213, + 0.01261561457067728, + 0.0028519663028419018, + 0.013083012774586678, + 0.010074908845126629, + 0.004877213854342699, + 0.0036108423955738544, + 0.003438273910433054, + 0.006746732629835606, + 0.006916232407093048, + 0.003187318332493305, + 0.005274870898574591, + 0.02000359073281288, + 0.025488052517175674, + 0.014424304477870464, + 0.012450647540390491, + 0.019367417320609093, + 0.026318365707993507, + 0.010128657333552837, + 0.017379099503159523, + 0.005437525920569897, + 0.009026381187140942, + 0.006117821205407381, + 0.008721567690372467, + 0.005558881908655167, + 0.010212431661784649, + 0.008066887967288494, + 0.00616112444549799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01607329398393631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an elderly man with deeply wrinkled and furrowed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a thick dark brown wool cardigan over an off-white cotton shirt with a slightly yellowed collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting in an old wooden armchair with a rough stone fireplace behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature crackling orange-red flames in a hearth that serve as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward while the man leans forward and moves one hand expressively in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.428901731967926 + ], + "opposite_logprob_tokens": [ + -1.0539016723632812 + ], + "logprob_reward": -0.428901731967926, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.428901731967926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3844110369682312, + 0.04164700582623482, + 0.026491679251194, + 0.03659460321068764, + 0.031777847558259964, + 0.04219948500394821, + 0.024878909811377525, + 0.12036833167076111, + 0.11713793128728867, + 0.17449314892292023 + ], + "credit_map": [ + 0.03192761912941933, + 0.021115873008966446, + 0.014483265578746796, + 0.01910526305437088, + 0.03349536284804344, + 0.1541968286037445, + 0.02391206845641136, + 0.08617479354143143, + 0.00711949123069644, + 0.006776417605578899, + 0.0036748198326677084, + 0.003461363259702921, + 0.007323749829083681, + 0.004625162109732628, + 0.004382132086902857, + 0.004283868242055178, + 0.0039075640961527824, + 0.003128401469439268, + 0.0030674701556563377, + 0.0028068318497389555, + 0.0036811488680541515, + 0.0038758914452046156, + 0.0030755240004509687, + 0.0029488473664969206, + 0.0038918997161090374, + 0.003767185378819704, + 0.00242081331089139, + 0.003686735639348626, + 0.008238089270889759, + 0.0048209792003035545, + 0.005606674589216709, + 0.00416222307831049, + 0.006309144664555788, + 0.0031534929294139147, + 0.0027695733588188887, + 0.003199319588020444, + 0.004370833747088909, + 0.003839920973405242, + 0.004717032425105572, + 0.003418529871851206, + 0.005636001005768776, + 0.005559185985475779, + 0.004528711084276438, + 0.007129506208002567, + 0.0056939497590065, + 0.003727648640051484, + 0.00525783933699131, + 0.004666644148528576, + 0.0027131675742566586, + 0.0020892475731670856, + 0.0026673879474401474, + 0.0026505221612751484, + 0.002365961205214262, + 0.0051628886722028255, + 0.0019861944019794464, + 0.005243539810180664, + 0.016680194064974785, + 0.008314047008752823, + 0.005556932650506496, + 0.007963071577250957, + 0.0098496675491333, + 0.038892634212970734, + 0.00451903510838747, + 0.02859274484217167, + 0.015830030664801598, + 0.01820300705730915, + 0.01233747135847807, + 0.009282446466386318, + 0.014198729768395424, + 0.02426963299512863, + 0.008074871264398098, + 0.014941737987101078, + 0.015965765342116356, + 0.031210990622639656, + 0.013087858445942402, + 0.02567967027425766, + 0.02317737229168415, + 0.03035571426153183, + 0.01794370450079441, + 0.01707208715379238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.428901731967926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_tensor([1478, 1478])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_tensor([1478, 1478])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48d81edec5bb1a327a6ba4a3d0298e530a2b5d46 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_tensor([1478, 1478])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle shot moves swiftly, following a sleek, multi-car light rail train as it teeters on elevated tracks, rendered in a vibrant, animated style. At night, the train negotiates a perilous, sharp turn above a busy, futuristic city street. The entire scene is bathed in a predominantly cool blue, creating a moody, nocturnal atmosphere. The train's polished silver metal surface acts like a distorting mirror, capturing and distorting the chaotic light emanating from the metropolis below. On the street, the light from car headlights and taillights blurs into streaks of white and red, while massive skyscrapers are adorned with vibrant neon signs in vibrant pink, electric yellow, and deep purple. These brilliant, varied colors cast long, shimmering reflections that slide and twist across the train's swaying body, emphasizing its precarious motion. The train itself tilts at an extreme angle, its individual cars jostling against each other as it strains against the curve, creating a palpable sense of high-speed tension and exhilarating danger.", + "seed": 28853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_1478_seed_28853_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05485573160112835, + "vlm_logprob": -0.05485573160112835, + "vlm_yes_no_margin": 4.799999952316284, + "vlm_token_credit": -0.05485573160112835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034219395369291306 + ], + "opposite_logprob_tokens": [ + -4.034219264984131 + ], + "logprob_reward": -0.034219395369291306, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.034219395369291306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05079753324389458, + 0.08185694366693497, + 0.09162759780883789, + 0.14118030667304993, + 0.21686722338199615, + 0.07988860458135605, + 0.031906191259622574, + 0.18912193179130554, + 0.10270846635103226, + 0.014045223593711853 + ], + "credit_map": [ + 0.005715341307222843, + 0.007028139661997557, + 0.006645005661994219, + 0.005295325070619583, + 0.0069233449175953865, + 0.006379440426826477, + 0.005803483072668314, + 0.007007450796663761, + 0.010910463519394398, + 0.009245578199625015, + 0.006376191042363644, + 0.010407193563878536, + 0.012084641493856907, + 0.004655597265809774, + 0.0208221934735775, + 0.007355084177106619, + 0.011363099329173565, + 0.011152110062539577, + 0.01427046861499548, + 0.011560591869056225, + 0.010973288677632809, + 0.006815713364630938, + 0.013400567695498466, + 0.012091757729649544, + 0.01043932605534792, + 0.01951736956834793, + 0.009996479377150536, + 0.020315635949373245, + 0.024507755413651466, + 0.020840585231781006, + 0.018537040799856186, + 0.017026111483573914, + 0.02729959785938263, + 0.029865754768252373, + 0.0180036723613739, + 0.030824368819594383, + 0.0395854115486145, + 0.02549225650727749, + 0.02708391100168228, + 0.018712271004915237, + 0.017673246562480927, + 0.00970639567822218, + 0.0031501122284680605, + 0.0035876380279660225, + 0.008055330254137516, + 0.027609393000602722, + 0.004782705567777157, + 0.0053237807005643845, + 0.00390150910243392, + 0.004396222997456789, + 0.003350566839799285, + 0.0034383859019726515, + 0.005243374966084957, + 0.004346413072198629, + 0.004125514067709446, + 0.0031042038463056087, + 0.031188664957880974, + 0.05341934785246849, + 0.014530492946505547, + 0.014174641110002995, + 0.02209354005753994, + 0.028272805735468864, + 0.014277808368206024, + 0.011164628900587559, + 0.014059705659747124, + 0.013699072413146496, + 0.010644959285855293, + 0.011028564535081387, + 0.014168476685881615, + 0.014375349506735802, + 0.013955787755548954, + 0.010776548646390438, + 0.00188098952639848, + 0.003412393154576421, + 0.0018732436001300812, + 0.0015494361286982894, + 0.0017710612155497074, + 0.001354083651676774, + 0.0009173594298772514, + 0.0012866563629359007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.034219395369291306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256160531193018 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.01256160531193018, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256160531193018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08047457784414291, + 0.08253375440835953, + 0.04895533621311188, + 0.042405661195516586, + 0.11944080889225006, + 0.0613679476082325, + 0.07931093126535416, + 0.38549545407295227, + 0.05714059993624687, + 0.04287497326731682 + ], + "credit_map": [ + 0.008794988505542278, + 0.010923604480922222, + 0.008484694175422192, + 0.008373569697141647, + 0.012507947161793709, + 0.011225774884223938, + 0.00938666332513094, + 0.010777324438095093, + 0.008420158177614212, + 0.008280763402581215, + 0.008983694016933441, + 0.01182505302131176, + 0.010611788369715214, + 0.005368213169276714, + 0.018015416339039803, + 0.011028662323951721, + 0.005714460276067257, + 0.008192566223442554, + 0.005652822554111481, + 0.005674990359693766, + 0.006572785787284374, + 0.004306843038648367, + 0.0067779081873595715, + 0.006062955129891634, + 0.002884219167754054, + 0.005283936392515898, + 0.002696373499929905, + 0.007231193128973246, + 0.007291864138096571, + 0.0060633509419858456, + 0.006022236775606871, + 0.00493248226121068, + 0.022923851385712624, + 0.016171863302588463, + 0.013952512294054031, + 0.01414747815579176, + 0.01864052750170231, + 0.014410954900085926, + 0.01085647102445364, + 0.008337145671248436, + 0.00916950311511755, + 0.0055357227101922035, + 0.007689410354942083, + 0.009445189498364925, + 0.006360440980643034, + 0.013458128087222576, + 0.0052388315089046955, + 0.004470714833587408, + 0.01048801839351654, + 0.010166856460273266, + 0.007864399813115597, + 0.008260685950517654, + 0.011413782835006714, + 0.011543232016265392, + 0.0101384362205863, + 0.00943551491945982, + 0.07059651613235474, + 0.13341742753982544, + 0.02711966447532177, + 0.03466382250189781, + 0.021425118669867516, + 0.07324936985969543, + 0.011533677577972412, + 0.013489828445017338, + 0.008004876784980297, + 0.009807740338146687, + 0.008534422144293785, + 0.0071038356982171535, + 0.004617015365511179, + 0.006435318849980831, + 0.0056641292758286, + 0.006973256357014179, + 0.006492688320577145, + 0.004796020220965147, + 0.004661940038204193, + 0.003680621972307563, + 0.005324182566255331, + 0.005883459001779556, + 0.004466527607291937, + 0.007569529116153717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.01256160531193018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00044109628652222455 + ], + "opposite_logprob_tokens": [ + -7.750441074371338 + ], + "logprob_reward": -0.00044109628652222455, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00044109628652222455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05610130727291107, + 0.1926604062318802, + 0.029659459367394447, + 0.016375044360756874, + 0.03575119748711586, + 0.4717438817024231, + 0.027674242854118347, + 0.10864928364753723, + 0.03946318477392197, + 0.021922074258327484 + ], + "credit_map": [ + 0.004751673899590969, + 0.00740463612601161, + 0.00549423648044467, + 0.007950996048748493, + 0.00841831136494875, + 0.007052389904856682, + 0.0070385099388659, + 0.007990548387169838, + 0.022459441795945168, + 0.018864499405026436, + 0.03143957629799843, + 0.040670715272426605, + 0.023984014987945557, + 0.01314634457230568, + 0.027045490220189095, + 0.015050322748720646, + 0.003371000289916992, + 0.0025052607525140047, + 0.0037398971617221832, + 0.0030596647411584854, + 0.003540255129337311, + 0.00264886231161654, + 0.004474427551031113, + 0.006320089101791382, + 0.001574822934344411, + 0.001838446594774723, + 0.002929012756794691, + 0.0031779909040778875, + 0.0021349750459194183, + 0.0010502120712772012, + 0.0017677549039945006, + 0.0019018268212676048, + 0.004521324764937162, + 0.004719689022749662, + 0.0028075079899281263, + 0.004886130336672068, + 0.0066192555241286755, + 0.004674885887652636, + 0.004762586671859026, + 0.0027598144952207804, + 0.10238656401634216, + 0.05310696363449097, + 0.0283200666308403, + 0.03540126979351044, + 0.03912835940718651, + 0.16340136528015137, + 0.026861069723963737, + 0.02313818596303463, + 0.0035338338930159807, + 0.003990926779806614, + 0.0024861996062099934, + 0.003139079548418522, + 0.0036082216538488865, + 0.0037952233105897903, + 0.004061190877109766, + 0.003059564856812358, + 0.021157285198569298, + 0.03946928679943085, + 0.007684886921197176, + 0.008777126669883728, + 0.005569344852119684, + 0.019863750785589218, + 0.0028257747180759907, + 0.0033018195535987616, + 0.005149576347321272, + 0.006275128107517958, + 0.004332852549850941, + 0.004540936090052128, + 0.004544050432741642, + 0.005928119644522667, + 0.004738458897918463, + 0.00395406037569046, + 0.0035951787140220404, + 0.0028475564904510975, + 0.00264284061267972, + 0.0017143312143161893, + 0.0024714788887649775, + 0.002649674890562892, + 0.0024252929724752903, + 0.0035757189616560936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.00044109628652222455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225540891289711 + ], + "opposite_logprob_tokens": [ + -1.6005408763885498 + ], + "logprob_reward": -0.225540891289711, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225540891289711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08971775323152542, + 0.09979235380887985, + 0.1087770015001297, + 0.048209067434072495, + 0.09583670645952225, + 0.2243696004152298, + 0.050891511142253876, + 0.1565447598695755, + 0.08375728875398636, + 0.04210397228598595 + ], + "credit_map": [ + 0.011468800716102123, + 0.009673059917986393, + 0.011071795597672462, + 0.00875189807265997, + 0.0119458744302392, + 0.012301982380449772, + 0.01469371747225523, + 0.00981062650680542, + 0.010387334041297436, + 0.013159562833607197, + 0.010556245222687721, + 0.018610991537570953, + 0.016136344522237778, + 0.0073480186983942986, + 0.01493208296597004, + 0.008661769330501556, + 0.00926392711699009, + 0.007272099610418081, + 0.015758804976940155, + 0.010226809419691563, + 0.012766793370246887, + 0.011458217166364193, + 0.01929328776896, + 0.022737065330147743, + 0.004378029145300388, + 0.005868174601346254, + 0.00509990518912673, + 0.008924229070544243, + 0.008602879010140896, + 0.00420925859361887, + 0.006562386639416218, + 0.004564205650240183, + 0.020085301250219345, + 0.013741903007030487, + 0.01110114436596632, + 0.012440900318324566, + 0.009113160893321037, + 0.01598489284515381, + 0.007648613769561052, + 0.00572078675031662, + 0.030033554881811142, + 0.026106921955943108, + 0.02190505340695381, + 0.02976064756512642, + 0.04147379472851753, + 0.028363827615976334, + 0.025558017194271088, + 0.021167762577533722, + 0.006563866976648569, + 0.008285594172775745, + 0.007971490733325481, + 0.00638301158323884, + 0.005036490969359875, + 0.0054082851856946945, + 0.006484083831310272, + 0.004758686758577824, + 0.016380248591303825, + 0.027779020369052887, + 0.02260003238916397, + 0.0166909359395504, + 0.021178212016820908, + 0.023853246122598648, + 0.01696835644543171, + 0.011094709858298302, + 0.012010514736175537, + 0.010340305045247078, + 0.00960606336593628, + 0.0087459497153759, + 0.009752144105732441, + 0.01188863068819046, + 0.010665363632142544, + 0.010748316533863544, + 0.00527133559808135, + 0.007604357320815325, + 0.00563004519790411, + 0.004172127228230238, + 0.007172984536737204, + 0.004900386091321707, + 0.003656689776107669, + 0.003696043509989977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.225540891289711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015156697481870651 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015156697481870651, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015156697481870651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06367632746696472, + 0.06617677211761475, + 0.21655817329883575, + 0.030713390558958054, + 0.09474080801010132, + 0.31305649876594543, + 0.07586965709924698, + 0.0499589741230011, + 0.06258261948823929, + 0.026666780933737755 + ], + "credit_map": [ + 0.005948265083134174, + 0.008260571397840977, + 0.008415303193032742, + 0.006749236490577459, + 0.009860584512352943, + 0.010593240149319172, + 0.006723070051521063, + 0.007126053795218468, + 0.009474002756178379, + 0.008908367715775967, + 0.007060437463223934, + 0.008347855880856514, + 0.009461427107453346, + 0.004778134170919657, + 0.01133914664387703, + 0.006807398516684771, + 0.01581750251352787, + 0.015008479356765747, + 0.03335072845220566, + 0.028413617983460426, + 0.025051020085811615, + 0.020793139934539795, + 0.037841565907001495, + 0.04028210788965225, + 0.0028528482653200626, + 0.004629948176443577, + 0.00425499863922596, + 0.00369844026863575, + 0.004855628591030836, + 0.0034502362832427025, + 0.0036328695714473724, + 0.003338419832289219, + 0.014110137708485126, + 0.012856889516115189, + 0.010867122560739517, + 0.011643975973129272, + 0.015611015260219574, + 0.012066017836332321, + 0.009597095660865307, + 0.00798855535686016, + 0.06604541093111038, + 0.03871263191103935, + 0.014422534964978695, + 0.01651633158326149, + 0.03275366500020027, + 0.10809527337551117, + 0.02020835503935814, + 0.016302278265357018, + 0.00817068200558424, + 0.011800138279795647, + 0.005297005642205477, + 0.006837162654846907, + 0.011122572235763073, + 0.013700319454073906, + 0.008396686054766178, + 0.010545083321630955, + 0.006744973827153444, + 0.011038037948310375, + 0.004128139931708574, + 0.005731658544391394, + 0.008234385401010513, + 0.004345500376075506, + 0.0056394850835204124, + 0.004096790216863155, + 0.008895817212760448, + 0.008258642628788948, + 0.006342199165374041, + 0.005722675938159227, + 0.007523380685597658, + 0.00827140361070633, + 0.007996981032192707, + 0.009571508504450321, + 0.003724141512066126, + 0.0043567270040512085, + 0.0034576691687107086, + 0.0027143312618136406, + 0.004167845938354731, + 0.0026823768857866526, + 0.002213821280747652, + 0.003349867882207036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.0015156697481870651 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle shot moves swiftly, following a sleek, multi-car light rail train as it teeters on elevated tracks, rendered in a vibrant, animated style. At night, the train negotiates a perilous, sharp turn above a busy, futuristic city street. The entire scene is bathed in a predominantly cool blue, creating a moody, nocturnal atmosphere. The train's polished silver metal surface acts like a distorting mirror, capturing and distorting the chaotic light emanating from the metropolis below. On the street, the light from car headlights and taillights blurs into streaks of white and red, while massive skyscrapers are adorned with vibrant neon signs in vibrant pink, electric yellow, and deep purple. These brilliant, varied colors cast long, shimmering reflections that slide and twist across the train's swaying body, emphasizing its precarious motion. The train itself tilts at an extreme angle, its individual cars jostling against each other as it strains against the curve, creating a palpable sense of high-speed tension and exhilarating danger.", + "seed": 28854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_1478_seed_28854_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08885756688541732, + "vlm_logprob": -0.08885756688541732, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.08885756688541732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37778428196907043 + ], + "opposite_logprob_tokens": [ + -1.252784252166748 + ], + "logprob_reward": -0.37778428196907043, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.37778428196907043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20436899363994598, + 0.07635190337896347, + 0.06616068631410599, + 0.1024610623717308, + 0.13889217376708984, + 0.057352516800165176, + 0.16469809412956238, + 0.029548637568950653, + 0.04969668388366699, + 0.11046919226646423 + ], + "credit_map": [ + 0.01861501857638359, + 0.014351561665534973, + 0.03113381750881672, + 0.01947050727903843, + 0.027901101857423782, + 0.026229456067085266, + 0.03145688772201538, + 0.035210661590099335, + 0.01119488850235939, + 0.007576290052384138, + 0.012837975285947323, + 0.008588861674070358, + 0.008386457338929176, + 0.009024248458445072, + 0.00933163147419691, + 0.00941154919564724, + 0.008660813793540001, + 0.005985502153635025, + 0.00968835037201643, + 0.007726970594376326, + 0.006637034006416798, + 0.011252416297793388, + 0.007881545461714268, + 0.00832805410027504, + 0.013928193598985672, + 0.013756032101809978, + 0.010970295406877995, + 0.0175116416066885, + 0.010269148275256157, + 0.012195209041237831, + 0.012773266062140465, + 0.011057271622121334, + 0.012791769579052925, + 0.01626676321029663, + 0.0171796977519989, + 0.02757679484784603, + 0.015246869064867496, + 0.013225706294178963, + 0.021322572603821754, + 0.015281992964446545, + 0.006077168975025415, + 0.00880074966698885, + 0.006539857015013695, + 0.010320181027054787, + 0.008012515492737293, + 0.005977666936814785, + 0.006504720076918602, + 0.00511966273188591, + 0.012487338855862617, + 0.014311925508081913, + 0.018155205994844437, + 0.02862739749252796, + 0.019948774948716164, + 0.01190191600471735, + 0.02886786498129368, + 0.030397659167647362, + 0.0026388659607619047, + 0.0026940919924527407, + 0.00323634990490973, + 0.004849803168326616, + 0.004088971763849258, + 0.0032568180467933416, + 0.004052485805004835, + 0.004731250926852226, + 0.00397455133497715, + 0.005600424483418465, + 0.006719659082591534, + 0.005583771038800478, + 0.009067176841199398, + 0.004723294638097286, + 0.006978703197091818, + 0.007049103267490864, + 0.005366954952478409, + 0.009420515038073063, + 0.011352053843438625, + 0.018430426716804504, + 0.007016428280621767, + 0.008259925059974194, + 0.023878002539277077, + 0.02674488164484501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.37778428196907043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026396818459033966 + ], + "opposite_logprob_tokens": [ + -3.6513967514038086 + ], + "logprob_reward": -0.026396818459033966, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026396818459033966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18137480318546295, + 0.0855550691485405, + 0.03925888240337372, + 0.08728101849555969, + 0.11387564241886139, + 0.04171315208077431, + 0.3460444211959839, + 0.01292384508997202, + 0.01455978024750948, + 0.07741335779428482 + ], + "credit_map": [ + 0.012957380153238773, + 0.01317409984767437, + 0.026429474353790283, + 0.017421269789338112, + 0.028816167265176773, + 0.028201434761285782, + 0.026914583519101143, + 0.02746039442718029, + 0.011442308314144611, + 0.010239734314382076, + 0.0074852886609733105, + 0.007952370680868626, + 0.019003242254257202, + 0.012274809181690216, + 0.007649851962924004, + 0.009507463313639164, + 0.004327286500483751, + 0.005193501710891724, + 0.0051653734408319, + 0.004516794811934233, + 0.0037708603776991367, + 0.004730877932161093, + 0.00648490572348237, + 0.005069280043244362, + 0.010140208527445793, + 0.008692261762917042, + 0.0055432869121432304, + 0.012375548481941223, + 0.01310100220143795, + 0.01459879707545042, + 0.01201205886900425, + 0.010817854665219784, + 0.010374086908996105, + 0.012149583548307419, + 0.01388358511030674, + 0.019434170797467232, + 0.014825037680566311, + 0.011602606624364853, + 0.019219914451241493, + 0.012386659160256386, + 0.0031167257111519575, + 0.006038069725036621, + 0.0056322054006159306, + 0.004507087171077728, + 0.008598350919783115, + 0.005572077352553606, + 0.004417288117110729, + 0.003831349778920412, + 0.029348120093345642, + 0.03272540494799614, + 0.03804681822657585, + 0.062268227338790894, + 0.052482232451438904, + 0.03221791610121727, + 0.04890557378530502, + 0.05005013570189476, + 0.0008612137171439826, + 0.0013236430240795016, + 0.001217180979438126, + 0.002887468785047531, + 0.0013334404211491346, + 0.0016385340131819248, + 0.0019125038525089622, + 0.0017498609377071261, + 0.001185137778520584, + 0.0026644247118383646, + 0.0017717473674565554, + 0.001457897131331265, + 0.0028356865514069796, + 0.0014285638462752104, + 0.0013103170786052942, + 0.0019060058984905481, + 0.0015535682905465364, + 0.0019516912288963795, + 0.002295701066032052, + 0.0018811416812241077, + 0.0017678237054497004, + 0.0035226228646934032, + 0.03184249624609947, + 0.03259831294417381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.026396818459033966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003071551618631929 + ], + "opposite_logprob_tokens": [ + -8.125307083129883 + ], + "logprob_reward": -0.0003071551618631929, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003071551618631929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13641467690467834, + 0.07367632538080215, + 0.10382208228111267, + 0.08201555907726288, + 0.04009043052792549, + 0.060495197772979736, + 0.04915561154484749, + 0.14615938067436218, + 0.2463749498128891, + 0.06179577112197876 + ], + "credit_map": [ + 0.01291797123849392, + 0.014383603818714619, + 0.017963409423828125, + 0.014302635565400124, + 0.01859133131802082, + 0.01759309321641922, + 0.019929733127355576, + 0.020732911303639412, + 0.012374427169561386, + 0.00866142101585865, + 0.007131252903491259, + 0.009899416007101536, + 0.00770858908072114, + 0.006946495734155178, + 0.012407657690346241, + 0.008547063916921616, + 0.012951230630278587, + 0.01268365141004324, + 0.012780650518834591, + 0.01100506167858839, + 0.011409306898713112, + 0.013740475289523602, + 0.016632841899991035, + 0.012618857435882092, + 0.01139046996831894, + 0.00856312271207571, + 0.007667554542422295, + 0.011366480961441994, + 0.010103456676006317, + 0.013426795601844788, + 0.010299975983798504, + 0.009197697043418884, + 0.004539166111499071, + 0.004196862690150738, + 0.0049479566514492035, + 0.007091824430972338, + 0.004764475394040346, + 0.004245657008141279, + 0.005146782845258713, + 0.005157706793397665, + 0.006752551067620516, + 0.00663616880774498, + 0.008456769399344921, + 0.011405245400965214, + 0.007412362843751907, + 0.00727462861686945, + 0.007079162169247866, + 0.00547830993309617, + 0.004071731120347977, + 0.004292943514883518, + 0.005695555359125137, + 0.005804714281111956, + 0.004893501754850149, + 0.00528253149241209, + 0.011050441302359104, + 0.008064189925789833, + 0.014833210036158562, + 0.014567937701940536, + 0.016448233276605606, + 0.022110549733042717, + 0.012436059303581715, + 0.016245760023593903, + 0.028808418661355972, + 0.02070920541882515, + 0.022112904116511345, + 0.043996669352054596, + 0.02370624989271164, + 0.022713731974363327, + 0.033523183315992355, + 0.03166959062218666, + 0.027771838009357452, + 0.040880776941776276, + 0.0026746292132884264, + 0.0029318719170987606, + 0.003707300405949354, + 0.0028579197824001312, + 0.004784617107361555, + 0.00488682184368372, + 0.020063627511262894, + 0.019888985902071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0003071551618631929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808479383587837 + ], + "opposite_logprob_tokens": [ + -3.2880847454071045 + ], + "logprob_reward": -0.03808479383587837, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808479383587837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09184925258159637, + 0.08762217313051224, + 0.051735058426856995, + 0.07441714406013489, + 0.12724773585796356, + 0.08198245614767075, + 0.2976112365722656, + 0.04494635388255119, + 0.059520941227674484, + 0.0830676406621933 + ], + "credit_map": [ + 0.010300766676664352, + 0.0067048389464616776, + 0.012624005787074566, + 0.007980971597135067, + 0.01373544242233038, + 0.013057530857622623, + 0.013288304209709167, + 0.014157394878566265, + 0.014121894724667072, + 0.010125684551894665, + 0.011661669239401817, + 0.011249787174165249, + 0.009817942976951599, + 0.010290767066180706, + 0.009931884706020355, + 0.010422547347843647, + 0.006278980989009142, + 0.004601727239787579, + 0.007809164002537727, + 0.006161913275718689, + 0.007178163155913353, + 0.0069266739301383495, + 0.006978447083383799, + 0.005799984559416771, + 0.012319887056946754, + 0.00807169359177351, + 0.005908302031457424, + 0.010004837065935135, + 0.008132764138281345, + 0.010351255536079407, + 0.0112770302221179, + 0.008351371623575687, + 0.013086973689496517, + 0.012939964421093464, + 0.018384726718068123, + 0.017169233411550522, + 0.018576210364699364, + 0.013702952302992344, + 0.0196821428835392, + 0.01370551809668541, + 0.007671277038753033, + 0.00898381881415844, + 0.011383566074073315, + 0.014840244315564632, + 0.013305749744176865, + 0.0073347860015928745, + 0.00953070167452097, + 0.00893231201916933, + 0.021572504192590714, + 0.023877298459410667, + 0.02662050724029541, + 0.04482513666152954, + 0.025812573730945587, + 0.025517454370856285, + 0.07063750922679901, + 0.0587482787668705, + 0.00373454624786973, + 0.006629697512835264, + 0.005586537998169661, + 0.0067729223519563675, + 0.004602843429893255, + 0.004946809262037277, + 0.00774803664535284, + 0.004924958571791649, + 0.005125150084495544, + 0.010048381052911282, + 0.008324586786329746, + 0.009251373820006847, + 0.00718369334936142, + 0.005754990037530661, + 0.006399782840162516, + 0.007432985119521618, + 0.006717415526509285, + 0.007480154279619455, + 0.009662539698183537, + 0.011498812586069107, + 0.006757029797881842, + 0.006669509690254927, + 0.018880121409893036, + 0.015402060002088547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.03808479383587837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017147850012406707 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017147850012406707, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017147850012406707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10093692690134048, + 0.056791115552186966, + 0.0542508140206337, + 0.09374172985553741, + 0.11903710663318634, + 0.050723712891340256, + 0.08010200411081314, + 0.08485192060470581, + 0.1047670841217041, + 0.2547975480556488 + ], + "credit_map": [ + 0.009863805957138538, + 0.008108611218631268, + 0.015049299225211143, + 0.009013687260448933, + 0.011608812026679516, + 0.01216178759932518, + 0.01680184341967106, + 0.018329087644815445, + 0.008953144773840904, + 0.006298573687672615, + 0.007921050302684307, + 0.006871517281979322, + 0.00607633963227272, + 0.007251420058310032, + 0.005941886454820633, + 0.007477183826267719, + 0.006391056813299656, + 0.008136029355227947, + 0.0071614524349570274, + 0.005845753010362387, + 0.005995484068989754, + 0.0063604265451431274, + 0.00780657259747386, + 0.0065540443174541, + 0.010318036191165447, + 0.011508294381201267, + 0.008755365386605263, + 0.010732441209256649, + 0.01339061837643385, + 0.013693859800696373, + 0.013762320391833782, + 0.01158080529421568, + 0.013612830080091953, + 0.012989447452127934, + 0.018686428666114807, + 0.016841717064380646, + 0.01835973560810089, + 0.011769372969865799, + 0.013991234824061394, + 0.012786353006958961, + 0.0039265695959329605, + 0.005298377480357885, + 0.007867076434195042, + 0.012845735996961594, + 0.005126379895955324, + 0.00548378424718976, + 0.004971133079379797, + 0.005204659886658192, + 0.004785249009728432, + 0.005771778058260679, + 0.008335175924003124, + 0.011176490224897861, + 0.007529060356318951, + 0.007600916083902121, + 0.019665995612740517, + 0.015237346291542053, + 0.008272266946732998, + 0.006617526523768902, + 0.008386734873056412, + 0.01820831373333931, + 0.007398154586553574, + 0.009407452307641506, + 0.013658993877470493, + 0.01290248055011034, + 0.008347894996404648, + 0.017886066809296608, + 0.009590189903974533, + 0.009379378519952297, + 0.01607736200094223, + 0.013828355818986893, + 0.012350337579846382, + 0.01730751246213913, + 0.005448941607028246, + 0.004970627836883068, + 0.007748802658170462, + 0.01063399389386177, + 0.007782394532114267, + 0.01223982684314251, + 0.10294194519519806, + 0.10303103923797607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.0017147850012406707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle shot moves swiftly, following a sleek, multi-car light rail train as it teeters on elevated tracks, rendered in a vibrant, animated style. At night, the train negotiates a perilous, sharp turn above a busy, futuristic city street. The entire scene is bathed in a predominantly cool blue, creating a moody, nocturnal atmosphere. The train's polished silver metal surface acts like a distorting mirror, capturing and distorting the chaotic light emanating from the metropolis below. On the street, the light from car headlights and taillights blurs into streaks of white and red, while massive skyscrapers are adorned with vibrant neon signs in vibrant pink, electric yellow, and deep purple. These brilliant, varied colors cast long, shimmering reflections that slide and twist across the train's swaying body, emphasizing its precarious motion. The train itself tilts at an extreme angle, its individual cars jostling against each other as it strains against the curve, creating a palpable sense of high-speed tension and exhilarating danger.", + "seed": 28855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_1478_seed_28855_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05406411627773196, + "vlm_logprob": -0.05406411627773196, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.05406411627773196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17569169402122498 + ], + "opposite_logprob_tokens": [ + -1.9256917238235474 + ], + "logprob_reward": -0.17569169402122498, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17569169402122498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4329604506492615, + 0.053527817130088806, + 0.06802180409431458, + 0.0250083114951849, + 0.045717354863882065, + 0.05115058645606041, + 0.07344699651002884, + 0.06972707808017731, + 0.1219099909067154, + 0.058529578149318695 + ], + "credit_map": [ + 0.042268168181180954, + 0.028935631737113, + 0.03667338192462921, + 0.1333661526441574, + 0.0239734947681427, + 0.021283533424139023, + 0.057666052132844925, + 0.08879402279853821, + 0.006435132585465908, + 0.006149292923510075, + 0.00694687757641077, + 0.01074347086250782, + 0.005041196476668119, + 0.004283295478671789, + 0.007042859215289354, + 0.006885692942887545, + 0.0047975340858101845, + 0.012367363087832928, + 0.007028651889413595, + 0.010115329176187515, + 0.0070036062970757484, + 0.009549337439239025, + 0.006050620693713427, + 0.011109366081655025, + 0.0023051092866808176, + 0.0038215494714677334, + 0.002825285540893674, + 0.0029375702142715454, + 0.002869688207283616, + 0.002496458822861314, + 0.0033916833344846964, + 0.00436096778139472, + 0.006814137101173401, + 0.004886454902589321, + 0.005711370147764683, + 0.0045806635171175, + 0.005135921761393547, + 0.007737438660115004, + 0.00479148281738162, + 0.006059888284653425, + 0.004934404511004686, + 0.0055883536115288734, + 0.007866528816521168, + 0.008160395547747612, + 0.005910767242312431, + 0.007123955059796572, + 0.005582145415246487, + 0.005984034854918718, + 0.00952564924955368, + 0.010856729932129383, + 0.0070992703549563885, + 0.008739696815609932, + 0.011303294450044632, + 0.010987079702317715, + 0.008197817951440811, + 0.006737454328685999, + 0.009629610925912857, + 0.01344543881714344, + 0.007060763891786337, + 0.013397308997809887, + 0.005179220344871283, + 0.007723522372543812, + 0.007157390005886555, + 0.006133822724223137, + 0.00993312057107687, + 0.015262672677636147, + 0.01918177492916584, + 0.013724190182983875, + 0.014783822931349277, + 0.01833903044462204, + 0.014007951132953167, + 0.016677431762218475, + 0.007236789911985397, + 0.004417646676301956, + 0.007389674428850412, + 0.011295357719063759, + 0.007933547720313072, + 0.004481486510485411, + 0.008778807707130909, + 0.006996269803494215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.17569169402122498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870905354619026 + ], + "opposite_logprob_tokens": [ + -3.0487091541290283 + ], + "logprob_reward": -0.04870905354619026, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870905354619026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5071579217910767, + 0.08330317586660385, + 0.059445787221193314, + 0.023339783772826195, + 0.0849132388830185, + 0.027800526469945908, + 0.028012044727802277, + 0.04753347113728523, + 0.07366630434989929, + 0.06482771039009094 + ], + "credit_map": [ + 0.05563369765877724, + 0.050345003604888916, + 0.05497334152460098, + 0.07770964503288269, + 0.04682862013578415, + 0.04093245044350624, + 0.09250082075595856, + 0.08823435008525848, + 0.00827347207814455, + 0.00898266863077879, + 0.012930911034345627, + 0.017207374796271324, + 0.009791565127670765, + 0.008631461299955845, + 0.00798084493726492, + 0.009504875168204308, + 0.005776088684797287, + 0.008630827069282532, + 0.010547864250838757, + 0.008677304722368717, + 0.006405087653547525, + 0.006525627337396145, + 0.0045907325111329556, + 0.008292254991829395, + 0.002033284166827798, + 0.0032512512989342213, + 0.0023094285279512405, + 0.0036271419376134872, + 0.0031916263978928328, + 0.0023819892667233944, + 0.0030701386276632547, + 0.0034749244805425406, + 0.005932730622589588, + 0.004252552520483732, + 0.011287580244243145, + 0.00603095255792141, + 0.010709255933761597, + 0.021411577239632607, + 0.014020757749676704, + 0.011267829686403275, + 0.0020907840225845575, + 0.00226609967648983, + 0.0025893186684697866, + 0.0030088722705841064, + 0.005210015457123518, + 0.004595979582518339, + 0.0034700583200901747, + 0.004569398704916239, + 0.00268981559202075, + 0.0034801075235009193, + 0.003041025483980775, + 0.0034982762299478054, + 0.004634707700461149, + 0.004287930205464363, + 0.003263946156948805, + 0.0031162360683083534, + 0.003881636541336775, + 0.004977581091225147, + 0.005469373427331448, + 0.008445190265774727, + 0.005327789578586817, + 0.007844804786145687, + 0.004922420252114534, + 0.006664675194770098, + 0.005885770544409752, + 0.008293463848531246, + 0.013938309624791145, + 0.012164963409304619, + 0.00820515863597393, + 0.007913653738796711, + 0.00843733549118042, + 0.008827646262943745, + 0.007911737076938152, + 0.007430777885019779, + 0.005344690289348364, + 0.008194726891815662, + 0.01052361074835062, + 0.007553874980658293, + 0.0098249027505517, + 0.008043386042118073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04870905354619026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009247555863112211 + ], + "opposite_logprob_tokens": [ + -7.000924587249756 + ], + "logprob_reward": -0.0009247555863112211, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009247555863112211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15952420234680176, + 0.08925370126962662, + 0.061946820467710495, + 0.07590343058109283, + 0.08135537058115005, + 0.07547411322593689, + 0.07011966407299042, + 0.10356801003217697, + 0.2104811817407608, + 0.07237344980239868 + ], + "credit_map": [ + 0.018913865089416504, + 0.019839134067296982, + 0.01770593784749508, + 0.02172659896314144, + 0.014059220440685749, + 0.011882982216775417, + 0.034866005182266235, + 0.020530471578240395, + 0.008259963244199753, + 0.011911377310752869, + 0.011009372770786285, + 0.022604716941714287, + 0.010189846158027649, + 0.006888598203659058, + 0.008748799562454224, + 0.009641032665967941, + 0.0048400177620351315, + 0.00830359198153019, + 0.006335647776722908, + 0.010507036931812763, + 0.005150746088474989, + 0.008482206612825394, + 0.007277528755366802, + 0.011050047352910042, + 0.005502633284777403, + 0.009025908075273037, + 0.008072299882769585, + 0.01338264811784029, + 0.01009655836969614, + 0.00430726632475853, + 0.011885059997439384, + 0.013631054200232029, + 0.00977837760001421, + 0.009330288507044315, + 0.009837016463279724, + 0.008962067775428295, + 0.007923069410026073, + 0.011337295174598694, + 0.013231505639851093, + 0.010955752804875374, + 0.006165305618196726, + 0.007778410334140062, + 0.006641368847340345, + 0.009035235270857811, + 0.014003483578562737, + 0.009954124689102173, + 0.011786755174398422, + 0.010109431110322475, + 0.006841226946562529, + 0.008684084750711918, + 0.006248136516660452, + 0.009132031351327896, + 0.008644813671708107, + 0.011295231990516186, + 0.010607949458062649, + 0.008666195906698704, + 0.010351622477173805, + 0.011640377342700958, + 0.01038922369480133, + 0.014869214035570621, + 0.0153195196762681, + 0.011639350093901157, + 0.012907334603369236, + 0.01645137183368206, + 0.01869170181453228, + 0.022225338965654373, + 0.023317871615290642, + 0.025937989354133606, + 0.0280083566904068, + 0.028024327009916306, + 0.03418242186307907, + 0.030093176290392876, + 0.008935702033340931, + 0.007638371549546719, + 0.0048607587814331055, + 0.007114931009709835, + 0.014634395018219948, + 0.008749756962060928, + 0.011018267832696438, + 0.009421264752745628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0009247555863112211 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043052781373262405 + ], + "opposite_logprob_tokens": [ + -3.1680526733398438 + ], + "logprob_reward": -0.043052781373262405, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043052781373262405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21037882566452026, + 0.08150561898946762, + 0.11807700246572495, + 0.10364973545074463, + 0.1496576964855194, + 0.10833139717578888, + 0.07533861696720123, + 0.028360402211546898, + 0.06241703778505325, + 0.06228364631533623 + ], + "credit_map": [ + 0.016886333003640175, + 0.02062167041003704, + 0.016423948109149933, + 0.054763954132795334, + 0.01951458305120468, + 0.014707493595778942, + 0.024482326582074165, + 0.042978521436452866, + 0.009662078693509102, + 0.010769160464406013, + 0.009942804463207722, + 0.017661308869719505, + 0.007611141540110111, + 0.008855554275214672, + 0.008460458368062973, + 0.00854311604052782, + 0.00980035774409771, + 0.020616373047232628, + 0.016242660582065582, + 0.017927145585417747, + 0.013417094945907593, + 0.014758369885385036, + 0.008172914385795593, + 0.01714208535850048, + 0.007602072786539793, + 0.011822700500488281, + 0.016428563743829727, + 0.014734553173184395, + 0.013597389683127403, + 0.010046972893178463, + 0.012221507728099823, + 0.017195969820022583, + 0.012441535480320454, + 0.011552037671208382, + 0.017345314845442772, + 0.013991111889481544, + 0.016975505277514458, + 0.03511347621679306, + 0.022258777171373367, + 0.019979940727353096, + 0.010402960702776909, + 0.010271809063851833, + 0.011501148343086243, + 0.012020202353596687, + 0.02289404720067978, + 0.014874021522700787, + 0.013388892635703087, + 0.01297831442207098, + 0.007203333079814911, + 0.008329257369041443, + 0.005367138888686895, + 0.009926803410053253, + 0.012177852913737297, + 0.014053023420274258, + 0.009034735150635242, + 0.00924647692590952, + 0.0024385880678892136, + 0.00343241891823709, + 0.0030166786164045334, + 0.004727368708699942, + 0.0027006715536117554, + 0.00483968248590827, + 0.0034241576213389635, + 0.0037808355409651995, + 0.00501839118078351, + 0.006902993191033602, + 0.015285423025488853, + 0.011175117455422878, + 0.005345145706087351, + 0.006318220868706703, + 0.004906685557216406, + 0.007465063128620386, + 0.0061072614043951035, + 0.005780818872153759, + 0.008092690259218216, + 0.014172098599374294, + 0.005960703361779451, + 0.004449260886758566, + 0.009466346353292465, + 0.008254464715719223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.043052781373262405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019422968616709113 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019422968616709113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019422968616709113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25116807222366333, + 0.05911697447299957, + 0.0710771307349205, + 0.11869117617607117, + 0.07420097291469574, + 0.07012026011943817, + 0.15077830851078033, + 0.04733048006892204, + 0.0914473682641983, + 0.06606928259134293 + ], + "credit_map": [ + 0.02690526470541954, + 0.01782459393143654, + 0.0211943369358778, + 0.07244408875703812, + 0.017736054956912994, + 0.012770452536642551, + 0.030429383739829063, + 0.051863908767700195, + 0.006448572035878897, + 0.007570141926407814, + 0.009576309472322464, + 0.012561432085931301, + 0.005305962637066841, + 0.00410487549379468, + 0.006432592403143644, + 0.007117083761841059, + 0.007234639022499323, + 0.011795738711953163, + 0.012273645959794521, + 0.008418833836913109, + 0.008369313552975655, + 0.009170428849756718, + 0.006934135686606169, + 0.006880386266857386, + 0.009171594865620136, + 0.01658504083752632, + 0.015051792375743389, + 0.01914113201200962, + 0.014089935459196568, + 0.009209812618792057, + 0.015157650224864483, + 0.020284218713641167, + 0.007569139823317528, + 0.008377416990697384, + 0.009376589208841324, + 0.007951205596327782, + 0.009794667363166809, + 0.012218574061989784, + 0.01033442746847868, + 0.0085789505392313, + 0.005767088383436203, + 0.006515127141028643, + 0.006451081484556198, + 0.008942890912294388, + 0.01512335054576397, + 0.010034923441708088, + 0.008480051532387733, + 0.008805747143924236, + 0.013600142672657967, + 0.016976825892925262, + 0.009322878904640675, + 0.020900556817650795, + 0.02461414784193039, + 0.025651859119534492, + 0.020515384152531624, + 0.019196508452296257, + 0.004547401797026396, + 0.005288383923470974, + 0.004208825528621674, + 0.009576118551194668, + 0.00655789952725172, + 0.00664402823895216, + 0.00492096645757556, + 0.005586852785199881, + 0.007135659456253052, + 0.009470091201364994, + 0.019744371995329857, + 0.015495610423386097, + 0.01039023045450449, + 0.009736116975545883, + 0.008943389169871807, + 0.010531903244554996, + 0.008064250461757183, + 0.007348688784986734, + 0.006145423743873835, + 0.00941462256014347, + 0.01083027571439743, + 0.007436174899339676, + 0.008629914373159409, + 0.008199933916330338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0019422968616709113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle shot moves swiftly, following a sleek, multi-car light rail train as it teeters on elevated tracks, rendered in a vibrant, animated style. At night, the train negotiates a perilous, sharp turn above a busy, futuristic city street. The entire scene is bathed in a predominantly cool blue, creating a moody, nocturnal atmosphere. The train's polished silver metal surface acts like a distorting mirror, capturing and distorting the chaotic light emanating from the metropolis below. On the street, the light from car headlights and taillights blurs into streaks of white and red, while massive skyscrapers are adorned with vibrant neon signs in vibrant pink, electric yellow, and deep purple. These brilliant, varied colors cast long, shimmering reflections that slide and twist across the train's swaying body, emphasizing its precarious motion. The train itself tilts at an extreme angle, its individual cars jostling against each other as it strains against the curve, creating a palpable sense of high-speed tension and exhilarating danger.", + "seed": 28856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_15/index_1478_seed_28856_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0325365913216956, + "vlm_logprob": -0.0325365913216956, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.0325365913216956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07168631255626678 + ], + "opposite_logprob_tokens": [ + -2.821686267852783 + ], + "logprob_reward": -0.07168631255626678, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07168631255626678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05944928154349327, + 0.05921245366334915, + 0.10288828611373901, + 0.024908659979701042, + 0.07786010205745697, + 0.41997402906417847, + 0.10185795277357101, + 0.05081206560134888, + 0.057747967541217804, + 0.04528922215104103 + ], + "credit_map": [ + 0.006994855124503374, + 0.009203873574733734, + 0.008320060558617115, + 0.009648101404309273, + 0.005189092364162207, + 0.008003193885087967, + 0.007343033794313669, + 0.004747075494378805, + 0.005277220159769058, + 0.0031376711558550596, + 0.0093739228323102, + 0.00988037046045065, + 0.0033600025344640017, + 0.0037315571680665016, + 0.015325869433581829, + 0.009125839918851852, + 0.008557654917240143, + 0.003292037872597575, + 0.004099296871572733, + 0.006667983252555132, + 0.011928380466997623, + 0.03268299996852875, + 0.020359385758638382, + 0.015300549566745758, + 0.0020667132921516895, + 0.0029782482888549566, + 0.002937420504167676, + 0.00202224962413311, + 0.0038130527827888727, + 0.0037193475291132927, + 0.004023587331175804, + 0.003348040394484997, + 0.005785138811916113, + 0.007456734776496887, + 0.006981560494750738, + 0.009647808969020844, + 0.015303872525691986, + 0.008521050214767456, + 0.011408179998397827, + 0.012755759991705418, + 0.014832966960966587, + 0.016806725412607193, + 0.01278007309883833, + 0.021623216569423676, + 0.018209848552942276, + 0.01658608950674534, + 0.16865113377571106, + 0.15048396587371826, + 0.012441270984709263, + 0.013433550484478474, + 0.01116963941603899, + 0.014335084706544876, + 0.014880241826176643, + 0.010624254122376442, + 0.014233777299523354, + 0.010740133933722973, + 0.0049884445033967495, + 0.005583667196333408, + 0.003476397367194295, + 0.01060596201568842, + 0.004837115295231342, + 0.005198607221245766, + 0.009247533977031708, + 0.006874335929751396, + 0.006021937821060419, + 0.007360872346907854, + 0.004125477280467749, + 0.007135974708944559, + 0.008352124132215977, + 0.0048486050218343735, + 0.010332401841878891, + 0.009570574387907982, + 0.0037022032774984837, + 0.00355786201544106, + 0.004612690303474665, + 0.004848724231123924, + 0.00612053694203496, + 0.003402737667784095, + 0.009578028693795204, + 0.009466438554227352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.07168631255626678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018208375200629234 + ], + "opposite_logprob_tokens": [ + -4.0182085037231445 + ], + "logprob_reward": -0.018208375200629234, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018208375200629234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08429676294326782, + 0.11041324585676193, + 0.1032978966832161, + 0.03007291629910469, + 0.06630674004554749, + 0.1409517079591751, + 0.12511906027793884, + 0.1668318808078766, + 0.12943942844867706, + 0.04327034577727318 + ], + "credit_map": [ + 0.0075166174210608006, + 0.006849858444184065, + 0.008937870152294636, + 0.014396604150533676, + 0.007073874585330486, + 0.019813159480690956, + 0.011652150191366673, + 0.008056632243096828, + 0.010604094713926315, + 0.005122837144881487, + 0.020409079268574715, + 0.022309046238660812, + 0.0059602833352983, + 0.006744490470737219, + 0.02266434207558632, + 0.016599074006080627, + 0.010612661018967628, + 0.003397637279704213, + 0.004262709058821201, + 0.00693637877702713, + 0.013243261724710464, + 0.031520795077085495, + 0.01858634129166603, + 0.01473811361938715, + 0.0030764932744205, + 0.002403436228632927, + 0.004467654973268509, + 0.002042079344391823, + 0.0037961704656481743, + 0.005140780005604029, + 0.003254583803936839, + 0.00589172076433897, + 0.007694796659052372, + 0.009133223444223404, + 0.0068718912079930305, + 0.0071006761863827705, + 0.007692000828683376, + 0.007121125236153603, + 0.01125182956457138, + 0.009441200643777847, + 0.0058904485777020454, + 0.005484168883413076, + 0.004150665830820799, + 0.009985754266381264, + 0.010236437432467937, + 0.009900679811835289, + 0.04941270872950554, + 0.04589085280895233, + 0.021157409995794296, + 0.012199717573821545, + 0.011887815780937672, + 0.021755415946245193, + 0.017417769879102707, + 0.010605587624013424, + 0.018665624782443047, + 0.011429729871451855, + 0.018369944766163826, + 0.013955775648355484, + 0.010312446393072605, + 0.02216394431889057, + 0.017895594239234924, + 0.016610654070973396, + 0.03216775506734848, + 0.03535577654838562, + 0.015168079175055027, + 0.020846514031291008, + 0.015073769725859165, + 0.013612470589578152, + 0.01553090289235115, + 0.013256183825433254, + 0.01771629974246025, + 0.018235208466649055, + 0.004413231275975704, + 0.0040377299301326275, + 0.004594198893755674, + 0.005750678945332766, + 0.0059441677294671535, + 0.005867303349077702, + 0.006521215662360191, + 0.006141824182122946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.018208375200629234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00030501006403937936 + ], + "opposite_logprob_tokens": [ + -8.12530517578125 + ], + "logprob_reward": -0.00030501006403937936, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.00030501006403937936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07856986671686172, + 0.06879465281963348, + 0.06116728112101555, + 0.046456228941679, + 0.05805429071187973, + 0.41960325837135315, + 0.04554871842265129, + 0.07613592594861984, + 0.12038014084100723, + 0.025289686396718025 + ], + "credit_map": [ + 0.009892643429338932, + 0.00900341384112835, + 0.012472333386540413, + 0.016534609720110893, + 0.008002745918929577, + 0.006963240914046764, + 0.008849387988448143, + 0.006851480342447758, + 0.00620174640789628, + 0.0055038719438016415, + 0.01126949768513441, + 0.01135961338877678, + 0.003751093754544854, + 0.006713650655001402, + 0.011783841997385025, + 0.012211333960294724, + 0.006110704503953457, + 0.006129477173089981, + 0.0030617506708949804, + 0.007087948266416788, + 0.013610769994556904, + 0.008028799667954445, + 0.0072792405262589455, + 0.009858581237494946, + 0.0033310255967080593, + 0.004781665280461311, + 0.0055144173093140125, + 0.00366468308493495, + 0.004498600959777832, + 0.0064778984524309635, + 0.00796979758888483, + 0.010218137875199318, + 0.0061803776770830154, + 0.008883689530193806, + 0.005603462923318148, + 0.004993277136236429, + 0.007455271203070879, + 0.006088579073548317, + 0.010767430067062378, + 0.008082197979092598, + 0.011275343596935272, + 0.0066936383955180645, + 0.008998197503387928, + 0.01930909790098667, + 0.018007759004831314, + 0.013966496102511883, + 0.18295808136463165, + 0.15839460492134094, + 0.005218063481152058, + 0.005554103758186102, + 0.0045387279242277145, + 0.006808807607740164, + 0.005966487806290388, + 0.004772166255861521, + 0.007791617419570684, + 0.0048987423069775105, + 0.0062277307733893394, + 0.0061939372681081295, + 0.0044580972753465176, + 0.007605958729982376, + 0.009824573993682861, + 0.007478306069970131, + 0.018145347014069557, + 0.016201969236135483, + 0.01282865833491087, + 0.016198787838220596, + 0.01402492169290781, + 0.018647609278559685, + 0.014538682997226715, + 0.008027002215385437, + 0.017130780965089798, + 0.01898367702960968, + 0.0032120265532284975, + 0.0031323193106800318, + 0.0021150587126612663, + 0.0030738727655261755, + 0.0025413380935788155, + 0.0018787315348163247, + 0.004960971884429455, + 0.004375365097075701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.00030501006403937936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999547779560089 + ], + "opposite_logprob_tokens": [ + -2.694995403289795 + ], + "logprob_reward": -0.06999547779560089, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999547779560089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14701557159423828, + 0.08797610551118851, + 0.11997103691101074, + 0.1461247056722641, + 0.12646819651126862, + 0.18221406638622284, + 0.03518717736005783, + 0.04423531889915466, + 0.03713452070951462, + 0.07367333769798279 + ], + "credit_map": [ + 0.016822898760437965, + 0.017241453751921654, + 0.019571488723158836, + 0.01555575616657734, + 0.013307549990713596, + 0.03514423966407776, + 0.015239515341818333, + 0.014132666401565075, + 0.006378789432346821, + 0.005246332380920649, + 0.016245516017079353, + 0.01660083793103695, + 0.0043517849408090115, + 0.00854012742638588, + 0.013365516439080238, + 0.01724720560014248, + 0.010534431785345078, + 0.007145113777369261, + 0.005325185135006905, + 0.008784478530287743, + 0.01926553249359131, + 0.03226009011268616, + 0.018644437193870544, + 0.01801176927983761, + 0.00790477730333805, + 0.020413324236869812, + 0.011629455722868443, + 0.01923760212957859, + 0.017328571528196335, + 0.02080760896205902, + 0.017864366993308067, + 0.03093899041414261, + 0.013773850165307522, + 0.012811833992600441, + 0.018738042563199997, + 0.012653794139623642, + 0.02130115032196045, + 0.011650928296148777, + 0.018452079966664314, + 0.017086511477828026, + 0.01884259097278118, + 0.014261770993471146, + 0.013997907750308514, + 0.020619848743081093, + 0.01401500217616558, + 0.01471434161067009, + 0.05127175897359848, + 0.034490834921598434, + 0.004272014368325472, + 0.003155993763357401, + 0.0050358944572508335, + 0.0052229478023946285, + 0.004009345080703497, + 0.003422618145123124, + 0.006223616190254688, + 0.003844747319817543, + 0.005321926902979612, + 0.004273900296539068, + 0.0025381871964782476, + 0.0060986620374023914, + 0.0041019800119102, + 0.005607216618955135, + 0.00803303625434637, + 0.008260408416390419, + 0.0034789862111210823, + 0.003777230391278863, + 0.0019915092270821333, + 0.00439500855281949, + 0.005946729332208633, + 0.0030726122204214334, + 0.006778141483664513, + 0.0076943039894104, + 0.006555734667927027, + 0.004959404934197664, + 0.00932042021304369, + 0.008832677267491817, + 0.007680114824324846, + 0.003908668179064989, + 0.016175832599401474, + 0.01624048314988613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.06999547779560089 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-car light rail train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train traveling on elevated tracks above a busy, futuristic city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at night with a predominantly cool blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train tilt at an extreme angle while navigating a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot dynamic and moving swiftly to follow the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1558781862258911, + 0.07722026109695435, + 0.03637176379561424, + 0.06753290444612503, + 0.1602500081062317, + 0.21908588707447052, + 0.11538422852754593, + 0.05258876457810402, + 0.042341578751802444, + 0.07334648072719574 + ], + "credit_map": [ + 0.011990864761173725, + 0.019522923976182938, + 0.016235629096627235, + 0.021849287673830986, + 0.014497793279588223, + 0.03834584727883339, + 0.01865026354789734, + 0.014785561710596085, + 0.006540116388350725, + 0.0043565318919718266, + 0.011109300889074802, + 0.015025480650365353, + 0.0049288636073470116, + 0.00954626314342022, + 0.014000123366713524, + 0.011713584885001183, + 0.004070500377565622, + 0.003271025838330388, + 0.0020972287748008966, + 0.0049562714993953705, + 0.0069785453379154205, + 0.005151002202183008, + 0.004755731672048569, + 0.005091459956020117, + 0.004619178362190723, + 0.0076897405087947845, + 0.005424021743237972, + 0.006850359495729208, + 0.009780353866517544, + 0.00868779793381691, + 0.011082889512181282, + 0.013398565351963043, + 0.016675226390361786, + 0.016477534547448158, + 0.02333257533609867, + 0.021325238049030304, + 0.02384108304977417, + 0.015370748937129974, + 0.021383341401815414, + 0.021844269707798958, + 0.007829184643924236, + 0.006635324098169804, + 0.00555978761985898, + 0.012559237889945507, + 0.01144158560782671, + 0.009970542043447495, + 0.08926733583211899, + 0.07582288235425949, + 0.009898655116558075, + 0.00845001358538866, + 0.010882052592933178, + 0.013716747984290123, + 0.010601050220429897, + 0.01047795731574297, + 0.030118543654680252, + 0.021239202469587326, + 0.004815710708498955, + 0.005472386255860329, + 0.003512513590976596, + 0.008249201811850071, + 0.00530741922557354, + 0.0044424403458833694, + 0.011961868032813072, + 0.008827226236462593, + 0.003779915627092123, + 0.005623059347271919, + 0.003351198509335518, + 0.007725628092885017, + 0.005846290849149227, + 0.0023781463969498873, + 0.006810351274907589, + 0.006826986558735371, + 0.007317528128623962, + 0.007211955264210701, + 0.009080708958208561, + 0.007697599940001965, + 0.009200378321111202, + 0.005347951781004667, + 0.01333116926252842, + 0.014159188605844975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0024877809919416904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_tensor([3336, 3336])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_tensor([3336, 3336])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f675a61e065a304889c4c6657109dda87f78a37c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_tensor([3336, 3336])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, held at eye level, presents the scene in a striking pop art style, defined by highly saturated colors and bold, thick black outlines. At its center lies a high-speed blender, its base a gleaming chrome surface, and a pitcher, a thick, transparent container, resting on a bright canary-yellow countertop. Inside the pitcher, a vibrant, unnatural pink liquid boils violently, stirring constantly into a chaotic vortex. Large, graphic bubbles, depicted as perfect circles with sharp white highlights, expand rapidly across the bubbling surface before bursting with exaggerated energy. From the blender's lid, stylized steam rises not as soft steam but in a distinctive pattern of bold black lines and Ben Day dots against a solid cobalt blue background. The entire composition is illuminated by bright, flat lighting, casting hard-edged, minimal shadows, creating an atmosphere of high contrast, energy, and graphic dynamism that feels both chaotic and meticulously designed.", + "seed": 17787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_3336_seed_17787_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008222060813568532, + "vlm_logprob": -0.008222060813568532, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.008222060813568532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009840651182457805 + ], + "opposite_logprob_tokens": [ + -8.250984191894531 + ], + "logprob_reward": -0.0009840651182457805, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0009840651182457805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18668733537197113, + 0.09068603813648224, + 0.11677315831184387, + 0.0888284221291542, + 0.08283691853284836, + 0.0688590332865715, + 0.13570033013820648, + 0.034083493053913116, + 0.08446450531482697, + 0.11108075082302094 + ], + "credit_map": [ + 0.01913689449429512, + 0.0339667946100235, + 0.024186575785279274, + 0.030125951394438744, + 0.022214096039533615, + 0.01355737168341875, + 0.025699006393551826, + 0.01780063286423683, + 0.00810899119824171, + 0.02673012763261795, + 0.007749147247523069, + 0.018571168184280396, + 0.00888579897582531, + 0.0047345091588795185, + 0.008995337411761284, + 0.00691096018999815, + 0.01576271653175354, + 0.025166643783450127, + 0.014368358999490738, + 0.01870020665228367, + 0.012312319129705429, + 0.007945232093334198, + 0.014257113449275494, + 0.008260565809905529, + 0.008986604399979115, + 0.025027137249708176, + 0.008090886287391186, + 0.021449072286486626, + 0.008957838639616966, + 0.004625075962394476, + 0.00788294617086649, + 0.0038088674191385508, + 0.008769877254962921, + 0.020708303898572922, + 0.006988616194576025, + 0.016060376539826393, + 0.009218476712703705, + 0.005588831380009651, + 0.010051275603473186, + 0.005451162811368704, + 0.006184323225170374, + 0.01704704947769642, + 0.0066300807520747185, + 0.01412722934037447, + 0.007257549092173576, + 0.005271231755614281, + 0.007455878425389528, + 0.004885692149400711, + 0.012732602655887604, + 0.04227510094642639, + 0.011199511587619781, + 0.027843181043863297, + 0.016791796311736107, + 0.007096403278410435, + 0.010654975660145283, + 0.007106756325811148, + 0.0037317529786378145, + 0.005914514418691397, + 0.0032339398749172688, + 0.006601949688047171, + 0.00362966931425035, + 0.003206429770216346, + 0.004963168408721685, + 0.002802069066092372, + 0.012138085439801216, + 0.015391917899250984, + 0.0087360804900527, + 0.014012102037668228, + 0.010087305679917336, + 0.00549595570191741, + 0.01174651738256216, + 0.006856548599898815, + 0.011914718896150589, + 0.029799604788422585, + 0.009974948130548, + 0.021722868084907532, + 0.010336929000914097, + 0.006972033530473709, + 0.01292326208204031, + 0.007436383981257677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0009840651182457805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204584993422031 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204584993422031, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204584993422031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34189900755882263, + 0.07413600385189056, + 0.05474760755896568, + 0.17060136795043945, + 0.04850185289978981, + 0.04886873438954353, + 0.1005985364317894, + 0.04765491187572479, + 0.028132742270827293, + 0.08485924452543259 + ], + "credit_map": [ + 0.032282039523124695, + 0.06356006860733032, + 0.040916673839092255, + 0.07196491956710815, + 0.028715597465634346, + 0.021929744631052017, + 0.05030543729662895, + 0.03222452104091644, + 0.007065141573548317, + 0.01961739920079708, + 0.007759984582662582, + 0.014186439104378223, + 0.006479834672063589, + 0.004887736402451992, + 0.009858744218945503, + 0.004280729219317436, + 0.005567413289099932, + 0.011440095491707325, + 0.0070825680159032345, + 0.010697728022933006, + 0.006035351660102606, + 0.004095661919564009, + 0.005427852738648653, + 0.004400942474603653, + 0.025187036022543907, + 0.035424891859292984, + 0.016619211062788963, + 0.03133365884423256, + 0.016688231378793716, + 0.013742910698056221, + 0.019383816048502922, + 0.012221617624163628, + 0.006314891390502453, + 0.011347096413373947, + 0.004494182765483856, + 0.00867706909775734, + 0.005322034005075693, + 0.0034786290489137173, + 0.005671477410942316, + 0.0031964743975549936, + 0.005459708161652088, + 0.008653965778648853, + 0.005449533928185701, + 0.011290411464869976, + 0.004546857438981533, + 0.004121487028896809, + 0.005834260955452919, + 0.003512514056637883, + 0.010485104285180569, + 0.028896929696202278, + 0.00901695154607296, + 0.01982688158750534, + 0.01029386930167675, + 0.0063298786990344524, + 0.009859035722911358, + 0.00588989956304431, + 0.00565483421087265, + 0.009717407636344433, + 0.004370532464236021, + 0.008041747845709324, + 0.006329844705760479, + 0.00362916337326169, + 0.00569906085729599, + 0.004212323576211929, + 0.0027313116006553173, + 0.006842513103038073, + 0.003634167369455099, + 0.005772265139967203, + 0.00264540733769536, + 0.0017967591993510723, + 0.002726740902289748, + 0.0019835804123431444, + 0.012019130401313305, + 0.018533090129494667, + 0.00922216847538948, + 0.01361755933612585, + 0.00908234529197216, + 0.006390937604010105, + 0.009149959310889244, + 0.006844061892479658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.014204584993422031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21138238906860352, + 0.16496877372264862, + 0.07314903289079666, + 0.0698370561003685, + 0.07727576047182083, + 0.05891215428709984, + 0.19838668406009674, + 0.027974586933851242, + 0.060881175100803375, + 0.0572323240339756 + ], + "credit_map": [ + 0.019047383219003677, + 0.041400957852602005, + 0.02602108009159565, + 0.03481379896402359, + 0.026564721018075943, + 0.01754624955356121, + 0.022191137075424194, + 0.02379707433283329, + 0.01514242123812437, + 0.023614905774593353, + 0.017732087522745132, + 0.033921342343091965, + 0.014492620714008808, + 0.014710230752825737, + 0.02892446704208851, + 0.016430696472525597, + 0.00754719041287899, + 0.015582840889692307, + 0.008114480413496494, + 0.011626028455793858, + 0.010794239118695259, + 0.005075851920992136, + 0.008654430508613586, + 0.005753975361585617, + 0.005434202495962381, + 0.017985370010137558, + 0.006630028132349253, + 0.014051397331058979, + 0.007254649884998798, + 0.00422899192199111, + 0.009967709891498089, + 0.004284708760678768, + 0.007145869079977274, + 0.023260818794369698, + 0.006398445460945368, + 0.013897585682570934, + 0.00725075276568532, + 0.004560795146971941, + 0.008083980530500412, + 0.006677510682493448, + 0.0049423715099692345, + 0.014578551985323429, + 0.006439160089939833, + 0.01036777812987566, + 0.006490054540336132, + 0.004857093561440706, + 0.007417396176606417, + 0.003819752251729369, + 0.018149567767977715, + 0.07060156017541885, + 0.012531514279544353, + 0.04256213828921318, + 0.0234389528632164, + 0.009124173782765865, + 0.013064646162092686, + 0.008914124220609665, + 0.0030308053828775883, + 0.006167367100715637, + 0.0023470281157642603, + 0.006841642316430807, + 0.0025579759385436773, + 0.001932490966282785, + 0.003268653992563486, + 0.0018286245176568627, + 0.006422059144824743, + 0.017110096290707588, + 0.006701961625367403, + 0.011957607232034206, + 0.004511853214353323, + 0.004200691822916269, + 0.00651782751083374, + 0.0034590819850564003, + 0.004070573952049017, + 0.0190390944480896, + 0.005053304601460695, + 0.011720609851181507, + 0.004543370567262173, + 0.003095644060522318, + 0.00583991315215826, + 0.0038698106072843075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0036151548847556114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774206206202507 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774206206202507, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774206206202507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2794862687587738, + 0.14710931479930878, + 0.103612020611763, + 0.08567880094051361, + 0.0637560784816742, + 0.09505951404571533, + 0.08150219172239304, + 0.049932681024074554, + 0.02673792652785778, + 0.06712522357702255 + ], + "credit_map": [ + 0.02063177153468132, + 0.06408506631851196, + 0.028724567964673042, + 0.051966387778520584, + 0.034393057227134705, + 0.02442220039665699, + 0.028100809082388878, + 0.02716241404414177, + 0.01241202000528574, + 0.034070197492837906, + 0.010226027108728886, + 0.034186650067567825, + 0.015077929012477398, + 0.01131019089370966, + 0.017759114503860474, + 0.012067187577486038, + 0.009652015753090382, + 0.026900174096226692, + 0.008656442165374756, + 0.021900061517953873, + 0.012455377727746964, + 0.006952124647796154, + 0.010331206023693085, + 0.00676462659612298, + 0.009735836647450924, + 0.020829375833272934, + 0.009549509733915329, + 0.016656339168548584, + 0.009406615979969501, + 0.006494205445051193, + 0.007923192344605923, + 0.005083727650344372, + 0.006096117198467255, + 0.014431631192564964, + 0.0079379016533494, + 0.009891986846923828, + 0.006927832495421171, + 0.00549929728731513, + 0.0069822329096496105, + 0.005989080294966698, + 0.007798336446285248, + 0.02968665212392807, + 0.007099158130586147, + 0.01971207931637764, + 0.010097022168338299, + 0.0066549391485750675, + 0.008496749214828014, + 0.005514573305845261, + 0.006969373673200607, + 0.02303195185959339, + 0.006818504072725773, + 0.016862936317920685, + 0.00797903724014759, + 0.0050683761946856976, + 0.009617404080927372, + 0.005154611077159643, + 0.006991970352828503, + 0.009347961284220219, + 0.0049967532977461815, + 0.010166977532207966, + 0.0055532678961753845, + 0.003959936089813709, + 0.004684042185544968, + 0.004231772385537624, + 0.0028864622581750154, + 0.005209650844335556, + 0.0031850708182901144, + 0.004644730128347874, + 0.0022343015298247337, + 0.0027488283812999725, + 0.003823835402727127, + 0.0020050446037203074, + 0.005631070584058762, + 0.02149013802409172, + 0.005639601033180952, + 0.012462606653571129, + 0.007118484936654568, + 0.004338440019637346, + 0.006235688924789429, + 0.004209189210087061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009774206206202507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532292865216732 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532292865216732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532292865216732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3292029798030853, + 0.13168317079544067, + 0.09356047958135605, + 0.09765076637268066, + 0.06319308280944824, + 0.04322924464941025, + 0.07072705030441284, + 0.048478011041879654, + 0.03506145998835564, + 0.08721373975276947 + ], + "credit_map": [ + 0.024179136380553246, + 0.07176295667886734, + 0.033141784369945526, + 0.06123876944184303, + 0.04028136283159256, + 0.03179996460676193, + 0.037364356219768524, + 0.029434649273753166, + 0.013118403032422066, + 0.02685471437871456, + 0.017168233171105385, + 0.024144308641552925, + 0.011204896494746208, + 0.011992749758064747, + 0.01603563316166401, + 0.011164236813783646, + 0.00956992618739605, + 0.013531778939068317, + 0.010354752652347088, + 0.021476149559020996, + 0.010440492071211338, + 0.006368269678205252, + 0.015346930362284184, + 0.006472187582403421, + 0.013554055243730545, + 0.024867165833711624, + 0.01033454854041338, + 0.016867216676473618, + 0.008761768229305744, + 0.006309079006314278, + 0.010383046232163906, + 0.00657389871776104, + 0.0066563403233885765, + 0.016859188675880432, + 0.006969859823584557, + 0.011801854707300663, + 0.005386607721447945, + 0.00404462730512023, + 0.006489202845841646, + 0.004985402338206768, + 0.004229903686791658, + 0.0072345612570643425, + 0.005858483724296093, + 0.007073186803609133, + 0.005090104416012764, + 0.004579921718686819, + 0.005199545528739691, + 0.003963538911193609, + 0.006711553316563368, + 0.01943904533982277, + 0.007211192976683378, + 0.013704727403819561, + 0.006891684141010046, + 0.004940994083881378, + 0.006723023951053619, + 0.005104838404804468, + 0.005675339139997959, + 0.011167312040925026, + 0.004944042302668095, + 0.00934822391718626, + 0.004681271500885487, + 0.003578687785193324, + 0.005616307258605957, + 0.003466830588877201, + 0.0038669565692543983, + 0.006358546670526266, + 0.00471592228859663, + 0.006001345347613096, + 0.0034864796325564384, + 0.0032317289151251316, + 0.004552173428237438, + 0.002848308300599456, + 0.010404342785477638, + 0.024991221725940704, + 0.009182888083159924, + 0.015664825215935707, + 0.007840093225240707, + 0.005543556530028582, + 0.008630717173218727, + 0.0049561006017029285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.012532292865216732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, presents the scene in a striking pop art style, defined by highly saturated colors and bold, thick black outlines. At its center lies a high-speed blender, its base a gleaming chrome surface, and a pitcher, a thick, transparent container, resting on a bright canary-yellow countertop. Inside the pitcher, a vibrant, unnatural pink liquid boils violently, stirring constantly into a chaotic vortex. Large, graphic bubbles, depicted as perfect circles with sharp white highlights, expand rapidly across the bubbling surface before bursting with exaggerated energy. From the blender's lid, stylized steam rises not as soft steam but in a distinctive pattern of bold black lines and Ben Day dots against a solid cobalt blue background. The entire composition is illuminated by bright, flat lighting, casting hard-edged, minimal shadows, creating an atmosphere of high contrast, energy, and graphic dynamism that feels both chaotic and meticulously designed.", + "seed": 17788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_3336_seed_17788_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022420704108662905, + "vlm_logprob": -0.022420704108662905, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.022420704108662905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013513966696336865 + ], + "opposite_logprob_tokens": [ + -7.376351356506348 + ], + "logprob_reward": -0.0013513966696336865, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0013513966696336865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19481757283210754, + 0.2604638636112213, + 0.14011922478675842, + 0.09432929754257202, + 0.08204547315835953, + 0.035146314650774, + 0.029283836483955383, + 0.03874598443508148, + 0.07896573096513748, + 0.04608263447880745 + ], + "credit_map": [ + 0.022852234542369843, + 0.05058997496962547, + 0.012061932124197483, + 0.02189832739531994, + 0.026936938986182213, + 0.014798292890191078, + 0.015970507636666298, + 0.029709357768297195, + 0.024515794590115547, + 0.011272033676505089, + 0.08428860455751419, + 0.016825711354613304, + 0.011146292090415955, + 0.025652475655078888, + 0.013668129220604897, + 0.07309482991695404, + 0.01966497302055359, + 0.0216277576982975, + 0.021863849833607674, + 0.011168415658175945, + 0.010007939301431179, + 0.012198965065181255, + 0.009622149169445038, + 0.03396518528461456, + 0.00883857998996973, + 0.005177006125450134, + 0.02273380197584629, + 0.006041239481419325, + 0.0038758055306971073, + 0.010902725160121918, + 0.004468696657568216, + 0.03229144215583801, + 0.006337461527436972, + 0.007236351724714041, + 0.008154388517141342, + 0.0065209041349589825, + 0.009495480917394161, + 0.012798023410141468, + 0.008277758955955505, + 0.023225106298923492, + 0.006081764120608568, + 0.005669885780662298, + 0.002556993393227458, + 0.0035568978637456894, + 0.004206689540296793, + 0.0040529347024858, + 0.0027507676277309656, + 0.006270383484661579, + 0.003017675131559372, + 0.004258147440850735, + 0.004892359487712383, + 0.003503975924104452, + 0.003449083538725972, + 0.002928393194451928, + 0.0020662208553403616, + 0.005167982075363398, + 0.0037296765949577093, + 0.007043296936899424, + 0.004423587117344141, + 0.005600709468126297, + 0.004476042464375496, + 0.004245852120220661, + 0.003333175089210272, + 0.005893642548471689, + 0.007113599684089422, + 0.012889984995126724, + 0.01569167897105217, + 0.00602746894583106, + 0.004715527407824993, + 0.009073046036064625, + 0.0038219690322875977, + 0.019632456824183464, + 0.004453711211681366, + 0.00822420697659254, + 0.0064412252977490425, + 0.003651598934084177, + 0.004271912854164839, + 0.0037938859313726425, + 0.002845058450475335, + 0.012401029467582703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0013513966696336865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898768037557602 + ], + "opposite_logprob_tokens": [ + -2.5789875984191895 + ], + "logprob_reward": -0.07898768037557602, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898768037557602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20226521790027618, + 0.15518519282341003, + 0.12389765679836273, + 0.10586046427488327, + 0.1581282913684845, + 0.13995030522346497, + 0.017059946432709694, + 0.04827786609530449, + 0.019987162202596664, + 0.029387831687927246 + ], + "credit_map": [ + 0.019229471683502197, + 0.03968289494514465, + 0.02131151780486107, + 0.028928160667419434, + 0.024739807471632957, + 0.013850015588104725, + 0.021413253620266914, + 0.03311010077595711, + 0.014489968307316303, + 0.010986101813614368, + 0.048785604536533356, + 0.009953049942851067, + 0.007309434935450554, + 0.011339236982166767, + 0.0075769610702991486, + 0.04474484175443649, + 0.016114819794893265, + 0.021480580791831017, + 0.0246575977653265, + 0.010739359073340893, + 0.011115302331745625, + 0.011489464901387691, + 0.009077928960323334, + 0.01922261342406273, + 0.009892120957374573, + 0.0057115317322313786, + 0.024586381390690804, + 0.006977125070989132, + 0.00548856146633625, + 0.01146444957703352, + 0.006284103728830814, + 0.03545619919896126, + 0.011816338635981083, + 0.01617184467613697, + 0.013298500329256058, + 0.013235006481409073, + 0.01610267348587513, + 0.028136037290096283, + 0.016830971464514732, + 0.042536936700344086, + 0.023512860760092735, + 0.01646915264427662, + 0.016846787184476852, + 0.012134397402405739, + 0.014830934815108776, + 0.028460584580898285, + 0.012507284060120583, + 0.015188314951956272, + 0.0018943777540698647, + 0.0017282661283388734, + 0.003094636369496584, + 0.0017016016645357013, + 0.0021281871013343334, + 0.002306578913703561, + 0.0014032804174348712, + 0.002803018083795905, + 0.0033846714068204165, + 0.011241147294640541, + 0.004580766428261995, + 0.007689138874411583, + 0.004107131157070398, + 0.003589913249015808, + 0.0038725058548152447, + 0.009812595322728157, + 0.0015133429551497102, + 0.006067028269171715, + 0.002461950061842799, + 0.0023192893713712692, + 0.0023192265070974827, + 0.0021401653066277504, + 0.001168392482213676, + 0.0019977684132754803, + 0.0029276481363922358, + 0.0032359843607991934, + 0.006444993894547224, + 0.0018437467515468597, + 0.0037657066714018583, + 0.0024759129155427217, + 0.0024120232556015253, + 0.006281817331910133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07898768037557602 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637557547539473 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637557547539473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637557547539473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34868937730789185, + 0.10065577179193497, + 0.12877599895000458, + 0.06641478091478348, + 0.10820959508419037, + 0.05195197835564613, + 0.05589156970381737, + 0.026788227260112762, + 0.07158119231462479, + 0.041041597723960876 + ], + "credit_map": [ + 0.02987802028656006, + 0.09750837087631226, + 0.02546381764113903, + 0.039970699697732925, + 0.03731496259570122, + 0.025827499106526375, + 0.02616444043815136, + 0.0665615126490593, + 0.008654965087771416, + 0.019469883292913437, + 0.01961408741772175, + 0.013899700716137886, + 0.007807862479239702, + 0.00933397002518177, + 0.006170189939439297, + 0.01570509746670723, + 0.015143885277211666, + 0.014310100115835667, + 0.029145821928977966, + 0.012533082626760006, + 0.009924240410327911, + 0.012281454168260098, + 0.009163403883576393, + 0.026273997500538826, + 0.006633258890360594, + 0.003977774642407894, + 0.01922648772597313, + 0.004673261661082506, + 0.0042307754047214985, + 0.009260591119527817, + 0.0037452843971550465, + 0.014667343348264694, + 0.009012287482619286, + 0.01302473247051239, + 0.009780695661902428, + 0.009591851383447647, + 0.010396579280495644, + 0.016583392396569252, + 0.010378357954323292, + 0.029441699385643005, + 0.0064737750217318535, + 0.007993921637535095, + 0.005067591089755297, + 0.0062531060539186, + 0.005331214051693678, + 0.008409650065004826, + 0.004889047704637051, + 0.007533671334385872, + 0.006115357391536236, + 0.005422113463282585, + 0.010389548726379871, + 0.006378191988915205, + 0.00615763058885932, + 0.008917942643165588, + 0.004000245593488216, + 0.008510536514222622, + 0.0028950267005711794, + 0.006002010311931372, + 0.002195583190768957, + 0.004607930779457092, + 0.004057526588439941, + 0.0021487066987901926, + 0.0022375076077878475, + 0.002643930958583951, + 0.008723846636712551, + 0.003579857060685754, + 0.02093324065208435, + 0.004193529020994902, + 0.0024031451903283596, + 0.007218020036816597, + 0.0032015375327318907, + 0.02132800780236721, + 0.004581152461469173, + 0.006092862691730261, + 0.00427777087315917, + 0.004535491578280926, + 0.006419852841645479, + 0.004897358827292919, + 0.0037025618366897106, + 0.006534543354064226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.004637557547539473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605921797454357 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.01605921797454357, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605921797454357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17580147087574005, + 0.11186212301254272, + 0.0876353532075882, + 0.12039418518543243, + 0.13712111115455627, + 0.1283566802740097, + 0.04207722097635269, + 0.028829345479607582, + 0.10443317145109177, + 0.0634894073009491 + ], + "credit_map": [ + 0.02201714925467968, + 0.0398385189473629, + 0.015625257045030594, + 0.026235129684209824, + 0.018677160143852234, + 0.01494420226663351, + 0.01476994063705206, + 0.023694107308983803, + 0.01009692158550024, + 0.02036222256720066, + 0.021136505529284477, + 0.011045098304748535, + 0.010709420777857304, + 0.013450214639306068, + 0.0064055719412863255, + 0.018656153231859207, + 0.008147268556058407, + 0.017867807298898697, + 0.00763359759002924, + 0.009952363558113575, + 0.00904655922204256, + 0.007544358726590872, + 0.007359520066529512, + 0.020083880051970482, + 0.012701690196990967, + 0.005628349259495735, + 0.03759909048676491, + 0.008371280506253242, + 0.005229264963418245, + 0.010726924985647202, + 0.007270684000104666, + 0.03286689147353172, + 0.011186719872057438, + 0.013330214656889439, + 0.013898292556405067, + 0.01173868216574192, + 0.013738472945988178, + 0.02308826893568039, + 0.014660376124083996, + 0.0354800820350647, + 0.017941588535904884, + 0.014960342086851597, + 0.016698751598596573, + 0.011543146334588528, + 0.014306453056633472, + 0.020691871643066406, + 0.012588154524564743, + 0.01962636038661003, + 0.004205181263387203, + 0.005844898987561464, + 0.003576389979571104, + 0.005366224329918623, + 0.006703540682792664, + 0.0069914585910737514, + 0.0031833057291805744, + 0.006206217687577009, + 0.0022972857113927603, + 0.004154092166572809, + 0.004132944159209728, + 0.0031017388682812452, + 0.0024978527799248695, + 0.003106747753918171, + 0.0020930967293679714, + 0.007445584982633591, + 0.010769934393465519, + 0.012948994524776936, + 0.028655538335442543, + 0.009389715269207954, + 0.006333240773528814, + 0.008262762799859047, + 0.005682088900357485, + 0.02239089086651802, + 0.00475835707038641, + 0.02046806924045086, + 0.006586397532373667, + 0.008826901204884052, + 0.009027532301843166, + 0.0047573549672961235, + 0.004197323694825172, + 0.0048674666322767735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.01605921797454357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067667976021767 + ], + "opposite_logprob_tokens": [ + -4.511067867279053 + ], + "logprob_reward": -0.011067667976021767, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067667976021767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6850301027297974, + 0.03306759148836136, + 0.07537587732076645, + 0.04209136962890625, + 0.04684608802199364, + 0.03678499162197113, + 0.014308704063296318, + 0.04546218737959862, + 0.010360801592469215, + 0.010672231204807758 + ], + "credit_map": [ + 0.06277210265398026, + 0.20258744060993195, + 0.042749978601932526, + 0.07629366219043732, + 0.07071344554424286, + 0.0478878878057003, + 0.04641450196504593, + 0.13561108708381653, + 0.0029081476386636496, + 0.003345990087836981, + 0.00648076506331563, + 0.0028380276635289192, + 0.0024117801804095507, + 0.005170536693185568, + 0.001834832364693284, + 0.008077512495219707, + 0.00818585604429245, + 0.006765440572053194, + 0.010972932912409306, + 0.006158933043479919, + 0.006083302665501833, + 0.007204749155789614, + 0.006165557540953159, + 0.023839106783270836, + 0.004334959667176008, + 0.002018506405875087, + 0.01350769679993391, + 0.0029140966944396496, + 0.0018755634082481265, + 0.0035487196873873472, + 0.0022133421152830124, + 0.011678483337163925, + 0.0036148536019027233, + 0.003771333023905754, + 0.0069935517385602, + 0.003912041895091534, + 0.005317385774105787, + 0.007279582787305117, + 0.004630687180906534, + 0.011326649226248264, + 0.003448331030085683, + 0.003412711899727583, + 0.005477599333971739, + 0.003184017725288868, + 0.0036895587109029293, + 0.005859348922967911, + 0.003323266515508294, + 0.008390158414840698, + 0.002502884017303586, + 0.0021562136244028807, + 0.0011669083032757044, + 0.002273495076224208, + 0.0015073573449626565, + 0.0018514697439968586, + 0.0013547338312491775, + 0.0014956422382965684, + 0.003113130107522011, + 0.00943334586918354, + 0.004229262005537748, + 0.008691391907632351, + 0.0044347671791911125, + 0.004122980870306492, + 0.00418528076261282, + 0.00725203100591898, + 0.001498492667451501, + 0.0013144771801307797, + 0.0013423152267932892, + 0.0013779009459540248, + 0.001117301289923489, + 0.0016610423335805535, + 0.0008798264316283166, + 0.0011694451095536351, + 0.0009427726618014276, + 0.001709897886030376, + 0.0013301870785653591, + 0.001819959026761353, + 0.001926789409480989, + 0.000941515201702714, + 0.0010048963595181704, + 0.0009962136391550303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.011067667976021767 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, presents the scene in a striking pop art style, defined by highly saturated colors and bold, thick black outlines. At its center lies a high-speed blender, its base a gleaming chrome surface, and a pitcher, a thick, transparent container, resting on a bright canary-yellow countertop. Inside the pitcher, a vibrant, unnatural pink liquid boils violently, stirring constantly into a chaotic vortex. Large, graphic bubbles, depicted as perfect circles with sharp white highlights, expand rapidly across the bubbling surface before bursting with exaggerated energy. From the blender's lid, stylized steam rises not as soft steam but in a distinctive pattern of bold black lines and Ben Day dots against a solid cobalt blue background. The entire composition is illuminated by bright, flat lighting, casting hard-edged, minimal shadows, creating an atmosphere of high contrast, energy, and graphic dynamism that feels both chaotic and meticulously designed.", + "seed": 17789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_3336_seed_17789_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04678641618229449, + "vlm_logprob": -0.04678641618229449, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.04678641618229449, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002434744266793132 + ], + "opposite_logprob_tokens": [ + -6.377434730529785 + ], + "logprob_reward": -0.002434744266793132, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.002434744266793132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3461899757385254, + 0.07181601971387863, + 0.051321640610694885, + 0.08313644677400589, + 0.11869920045137405, + 0.07668796181678772, + 0.06942853331565857, + 0.08194036781787872, + 0.06127937138080597, + 0.03950043395161629 + ], + "credit_map": [ + 0.04472527280449867, + 0.0788472592830658, + 0.03523295000195503, + 0.03248446807265282, + 0.06575482338666916, + 0.02406136691570282, + 0.03867766633629799, + 0.026406196877360344, + 0.010283167473971844, + 0.013303681276738644, + 0.008027459494769573, + 0.006004930939525366, + 0.009279719553887844, + 0.007529976777732372, + 0.007719962392002344, + 0.00966713298112154, + 0.007355882320553064, + 0.007867580279707909, + 0.00628062104806304, + 0.006026442628353834, + 0.008680913597345352, + 0.0036847221199423075, + 0.006477666553109884, + 0.004947818350046873, + 0.009590172208845615, + 0.017536194995045662, + 0.009232611395418644, + 0.01003238745033741, + 0.012062030844390392, + 0.007805509027093649, + 0.010379469953477383, + 0.006498077418655157, + 0.01556355506181717, + 0.02548103779554367, + 0.012102115899324417, + 0.01069473847746849, + 0.022576071321964264, + 0.008766316808760166, + 0.012739605270326138, + 0.010775763541460037, + 0.006672850344330072, + 0.011037502437829971, + 0.009536745958030224, + 0.011320327408611774, + 0.007332656532526016, + 0.006096140947192907, + 0.013873039744794369, + 0.010818704031407833, + 0.00801999494433403, + 0.013455510139465332, + 0.006829483434557915, + 0.007398112677037716, + 0.014952979981899261, + 0.004959031939506531, + 0.0071097188629209995, + 0.006703705061227083, + 0.008332116529345512, + 0.017434999346733093, + 0.008033723570406437, + 0.009806079789996147, + 0.0178119707852602, + 0.00707078306004405, + 0.006867952179163694, + 0.006582751404494047, + 0.0061754449270665646, + 0.015085444785654545, + 0.006051450967788696, + 0.008710287511348724, + 0.007362529635429382, + 0.0036946686450392008, + 0.00743988947942853, + 0.006759658921509981, + 0.004089827183634043, + 0.006871283985674381, + 0.004127682186663151, + 0.003348683938384056, + 0.008788944222033024, + 0.0032037340570241213, + 0.00483829528093338, + 0.0042319889180362225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.002434744266793132 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015840709209442 + ], + "opposite_logprob_tokens": [ + -1.7015841007232666 + ], + "logprob_reward": -0.2015840709209442, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015840709209442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1578332483768463, + 0.09688086062669754, + 0.04139574617147446, + 0.09775224328041077, + 0.26783886551856995, + 0.048230331391096115, + 0.18710030615329742, + 0.02771248295903206, + 0.02187695913016796, + 0.053378958255052567 + ], + "credit_map": [ + 0.01398998498916626, + 0.0297402236610651, + 0.0163788590580225, + 0.017491519451141357, + 0.02606930397450924, + 0.018559519201517105, + 0.018607163801789284, + 0.01699667051434517, + 0.017411822453141212, + 0.017832385376095772, + 0.012164182960987091, + 0.008263758383691311, + 0.011081904172897339, + 0.008192691951990128, + 0.00916697271168232, + 0.012767150066792965, + 0.005224095657467842, + 0.00606393301859498, + 0.005370361264795065, + 0.005016516428440809, + 0.0048523242585361, + 0.003440992673859, + 0.005496078636497259, + 0.005931444466114044, + 0.01589321903884411, + 0.014553221873939037, + 0.013420525006949902, + 0.010202757082879543, + 0.012787614949047565, + 0.010837532579898834, + 0.01204010657966137, + 0.008017264306545258, + 0.035253241658210754, + 0.07692699134349823, + 0.02292870543897152, + 0.023169342428445816, + 0.040265828371047974, + 0.02448876015841961, + 0.02679154835641384, + 0.01801445335149765, + 0.005387572571635246, + 0.007433236576616764, + 0.006334931124001741, + 0.007343587465584278, + 0.005254087503999472, + 0.004056798759847879, + 0.006004126742482185, + 0.006415991112589836, + 0.022621750831604004, + 0.022760536521673203, + 0.031726449728012085, + 0.025836048647761345, + 0.024096466600894928, + 0.011800414882600307, + 0.0300480704754591, + 0.018210554495453835, + 0.0024469660129398108, + 0.0058326516300439835, + 0.002514083869755268, + 0.003212105017155409, + 0.004089074674993753, + 0.0032630576752126217, + 0.0030116771813482046, + 0.003342864103615284, + 0.0023254144471138716, + 0.003633167129009962, + 0.00237957458011806, + 0.0031437817960977554, + 0.003244949970394373, + 0.0015599074540659785, + 0.0034579455386847258, + 0.0021322162356227636, + 0.0055242665112018585, + 0.0069001661613583565, + 0.005379479378461838, + 0.0063229636289179325, + 0.005997074767947197, + 0.005591711960732937, + 0.00820943247526884, + 0.009453864768147469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.2015840709209442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22167129814624786, + 0.14511218667030334, + 0.06570245325565338, + 0.09546166658401489, + 0.14220456779003143, + 0.025903182104229927, + 0.20057952404022217, + 0.03413660079240799, + 0.0532442070543766, + 0.015984369441866875 + ], + "credit_map": [ + 0.02331460453569889, + 0.055620186030864716, + 0.023368068039417267, + 0.023321643471717834, + 0.04377567768096924, + 0.014078162610530853, + 0.025502106174826622, + 0.01269082073122263, + 0.013155817054212093, + 0.030743857845664024, + 0.016629021614789963, + 0.018949663266539574, + 0.02551259659230709, + 0.012489750981330872, + 0.015726134181022644, + 0.011905327439308167, + 0.0054623340256512165, + 0.012900627218186855, + 0.009310105815529823, + 0.007833242416381836, + 0.011714117601513863, + 0.00541033735498786, + 0.008323167450726032, + 0.004748515319079161, + 0.007492292672395706, + 0.019464027136564255, + 0.01060977391898632, + 0.009279749356210232, + 0.01874304562807083, + 0.010128213092684746, + 0.012157240882515907, + 0.007587317377328873, + 0.020213468000292778, + 0.035852789878845215, + 0.015011092647910118, + 0.01247396506369114, + 0.02059747837483883, + 0.01458681933581829, + 0.013812676072120667, + 0.009656257927417755, + 0.003607609309256077, + 0.0044738249853253365, + 0.0028051426634192467, + 0.002710773143917322, + 0.0033328086137771606, + 0.0019951602444052696, + 0.003938641399145126, + 0.0030392210464924574, + 0.0409696139395237, + 0.03066086769104004, + 0.023788323625922203, + 0.021423891186714172, + 0.02733781933784485, + 0.015515081584453583, + 0.021773317828774452, + 0.01911059021949768, + 0.0027297958731651306, + 0.007995796389877796, + 0.003823718288913369, + 0.003607033286243677, + 0.005941880866885185, + 0.002628587419167161, + 0.0034435081761330366, + 0.003966276068240404, + 0.004382026381790638, + 0.010819539427757263, + 0.005282858852297068, + 0.005752746481448412, + 0.0097225746139884, + 0.00525859696790576, + 0.007309759967029095, + 0.004716099239885807, + 0.00132581009529531, + 0.003439859254285693, + 0.0016058051260188222, + 0.0014231277164071798, + 0.002627460053190589, + 0.0023401801008731127, + 0.001755429315380752, + 0.0014666975475847721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187425374984741 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.014187425374984741, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187425374984741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1867404282093048, + 0.09888901561498642, + 0.06146826222538948, + 0.12922382354736328, + 0.08871818333864212, + 0.09463929384946823, + 0.13544084131717682, + 0.11904367804527283, + 0.050758082419633865, + 0.03507833927869797 + ], + "credit_map": [ + 0.02124089188873768, + 0.03323594853281975, + 0.015574757009744644, + 0.016599152237176895, + 0.025003910064697266, + 0.03388136252760887, + 0.01746482029557228, + 0.023739593103528023, + 0.007491901516914368, + 0.021709302440285683, + 0.010197286494076252, + 0.013734379783272743, + 0.017069349065423012, + 0.007785086985677481, + 0.011060969904065132, + 0.009840735234320164, + 0.005767466500401497, + 0.009629754349589348, + 0.009267541579902172, + 0.009096317924559116, + 0.008595958352088928, + 0.0050028045661747456, + 0.007241419516503811, + 0.006866996642202139, + 0.012002565898001194, + 0.032381948083639145, + 0.011796594597399235, + 0.01792784035205841, + 0.023722708225250244, + 0.01049877516925335, + 0.012819305993616581, + 0.008074095472693443, + 0.012735451571643353, + 0.024594752117991447, + 0.00844776164740324, + 0.007628659252077341, + 0.013419241644442081, + 0.007460110355168581, + 0.008504634723067284, + 0.005927576217800379, + 0.008956825360655785, + 0.015429537743330002, + 0.01662062667310238, + 0.018955063074827194, + 0.01086003053933382, + 0.0053790356032550335, + 0.01195661909878254, + 0.006481560412794352, + 0.031705573201179504, + 0.02111954614520073, + 0.01622605137526989, + 0.01267277728766203, + 0.015702249482274055, + 0.011190167628228664, + 0.013525937683880329, + 0.013298538513481617, + 0.009425465948879719, + 0.033376701176166534, + 0.012630179524421692, + 0.014141900464892387, + 0.02441578358411789, + 0.007766283117234707, + 0.009850352071225643, + 0.007437011692672968, + 0.006475238595157862, + 0.01067802868783474, + 0.006153430789709091, + 0.00802844762802124, + 0.007209023926407099, + 0.0036553810350596905, + 0.004620712716132402, + 0.003937818575650454, + 0.0046292622573673725, + 0.006252619903534651, + 0.004155480302870274, + 0.004156875889748335, + 0.004669265355914831, + 0.003368041245266795, + 0.00421120785176754, + 0.00363558460958302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.014187425374984741 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48590677976608276, + 0.06670068204402924, + 0.04275253415107727, + 0.09272207319736481, + 0.05066557228565216, + 0.09177737683057785, + 0.03717976063489914, + 0.0450558215379715, + 0.05117018520832062, + 0.03606925532221794 + ], + "credit_map": [ + 0.048387255519628525, + 0.11660224199295044, + 0.04468608275055885, + 0.04399384185671806, + 0.10196410119533539, + 0.043729543685913086, + 0.048794057220220566, + 0.03774963319301605, + 0.009332574903964996, + 0.012495906092226505, + 0.006046164780855179, + 0.007801261264830828, + 0.007513908203691244, + 0.00656518479809165, + 0.00808245874941349, + 0.008863220922648907, + 0.005851722788065672, + 0.008874955587089062, + 0.004569671582430601, + 0.0034752420615404844, + 0.0063210707157850266, + 0.003209940157830715, + 0.003888885024935007, + 0.006561045069247484, + 0.015067609958350658, + 0.014257127419114113, + 0.014850410632789135, + 0.010961078107357025, + 0.013088200241327286, + 0.006058529019355774, + 0.010255386121571064, + 0.00818373542279005, + 0.0037801198195666075, + 0.014859906397759914, + 0.00454165693372488, + 0.006066890899091959, + 0.008423001505434513, + 0.0028695398941636086, + 0.006120461970567703, + 0.00400399649515748, + 0.012356681749224663, + 0.013750400394201279, + 0.010356090031564236, + 0.011030523106455803, + 0.010002829134464264, + 0.007573096081614494, + 0.014110401272773743, + 0.012597352266311646, + 0.004023951943963766, + 0.007446504198014736, + 0.00530066667124629, + 0.0037893301341682673, + 0.006730709690600634, + 0.0019698813557624817, + 0.005073062609881163, + 0.002845655893906951, + 0.0048186699859797955, + 0.005462037865072489, + 0.005751671269536018, + 0.004594765137881041, + 0.004890460520982742, + 0.00808753166347742, + 0.005305285565555096, + 0.006145399063825607, + 0.004026568029075861, + 0.013747465796768665, + 0.005592650733888149, + 0.00694994954392314, + 0.00817989930510521, + 0.0032389077823609114, + 0.005195789970457554, + 0.004238952416926622, + 0.0032234538812190294, + 0.006171592976897955, + 0.0042567471973598, + 0.003908050246536732, + 0.006626033689826727, + 0.0027548980433493853, + 0.004423258826136589, + 0.004705221392214298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0059450226835906506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, presents the scene in a striking pop art style, defined by highly saturated colors and bold, thick black outlines. At its center lies a high-speed blender, its base a gleaming chrome surface, and a pitcher, a thick, transparent container, resting on a bright canary-yellow countertop. Inside the pitcher, a vibrant, unnatural pink liquid boils violently, stirring constantly into a chaotic vortex. Large, graphic bubbles, depicted as perfect circles with sharp white highlights, expand rapidly across the bubbling surface before bursting with exaggerated energy. From the blender's lid, stylized steam rises not as soft steam but in a distinctive pattern of bold black lines and Ben Day dots against a solid cobalt blue background. The entire composition is illuminated by bright, flat lighting, casting hard-edged, minimal shadows, creating an atmosphere of high contrast, energy, and graphic dynamism that feels both chaotic and meticulously designed.", + "seed": 17790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_2/index_3336_seed_17790_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.18473543375730514, + "vlm_logprob": -0.18473543375730514, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.18473543375730514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03474663943052292 + ], + "opposite_logprob_tokens": [ + -3.4097466468811035 + ], + "logprob_reward": -0.03474663943052292, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03474663943052292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6795167922973633, + 0.09447316825389862, + 0.03680916130542755, + 0.011019731871783733, + 0.024019088596105576, + 0.007789235562086105, + 0.08508549630641937, + 0.017705274745821953, + 0.01650606095790863, + 0.027075961232185364 + ], + "credit_map": [ + 0.17804838716983795, + 0.03838495537638664, + 0.035129256546497345, + 0.021956155076622963, + 0.06296877562999725, + 0.26839539408683777, + 0.03807977959513664, + 0.03655405342578888, + 0.00874265469610691, + 0.010916013270616531, + 0.007591987028717995, + 0.010973735712468624, + 0.022422991693019867, + 0.011402259580790997, + 0.017797091975808144, + 0.004626427311450243, + 0.004284996073693037, + 0.004118653479963541, + 0.0022577536292374134, + 0.0040619042702019215, + 0.009412643499672413, + 0.004658849909901619, + 0.005929119419306517, + 0.002085241489112377, + 0.0012357410741969943, + 0.0021933442912995815, + 0.0011237465078011155, + 0.0010519056813791394, + 0.001656915876083076, + 0.0015465603210031986, + 0.0009972923435270786, + 0.0012142261257395148, + 0.003053572028875351, + 0.003041609888896346, + 0.0019709698390215635, + 0.0021503327880054712, + 0.003881955286487937, + 0.0045801736414432526, + 0.002862584311515093, + 0.0024778912775218487, + 0.0010818974114954472, + 0.001025655190460384, + 0.000810942321550101, + 0.0007066044490784407, + 0.0009754772763699293, + 0.0009896581759676337, + 0.000882491935044527, + 0.0013165087439119816, + 0.01585274748504162, + 0.012204764410853386, + 0.007508950773626566, + 0.0076172566041350365, + 0.007685051299631596, + 0.008320609107613564, + 0.004606105852872133, + 0.021290017291903496, + 0.0015470398357138038, + 0.00254535092972219, + 0.0010126722045242786, + 0.002082186285406351, + 0.0036797872744500637, + 0.0018954603001475334, + 0.003157243365421891, + 0.0017855354817584157, + 0.0017985508311539888, + 0.0035549027379602194, + 0.0009242954547517002, + 0.002445246558636427, + 0.0024957938585430384, + 0.0020328026730567217, + 0.0016254126094281673, + 0.001629055943340063, + 0.005261728074401617, + 0.0024394840002059937, + 0.001773990225046873, + 0.0020686655770987272, + 0.0036079217679798603, + 0.0053748623467981815, + 0.0019089130219072104, + 0.004640395753085613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03474663943052292 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328403949737549 + ], + "opposite_logprob_tokens": [ + -0.7578403949737549 + ], + "logprob_reward": -0.6328403949737549, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328403949737549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5838698148727417, + 0.12867742776870728, + 0.048199158161878586, + 0.02913893200457096, + 0.028185756877064705, + 0.009342403151094913, + 0.13084612786769867, + 0.013678101822733879, + 0.01547475066035986, + 0.012587571516633034 + ], + "credit_map": [ + 0.1469968557357788, + 0.03660454601049423, + 0.030130062252283096, + 0.019207246601581573, + 0.061249908059835434, + 0.2226373702287674, + 0.03510107472538948, + 0.03194271773099899, + 0.014396131038665771, + 0.014706388115882874, + 0.011228696443140507, + 0.013564526103436947, + 0.030588584020733833, + 0.01546978484839201, + 0.022512180730700493, + 0.006211136467754841, + 0.007947424426674843, + 0.004544053692370653, + 0.0027638820465654135, + 0.0035460353828966618, + 0.015427552163600922, + 0.006624043453484774, + 0.004554077982902527, + 0.0027920883148908615, + 0.0037647257559001446, + 0.005448806565254927, + 0.0020030413288623095, + 0.0042200698517262936, + 0.003361644223332405, + 0.003987238742411137, + 0.004278330598026514, + 0.0020750761032104492, + 0.004067940171808004, + 0.0036410961765795946, + 0.0020650913938879967, + 0.002982958685606718, + 0.003883717115968466, + 0.005262203514575958, + 0.0036307312548160553, + 0.0026520194951444864, + 0.0011600482976064086, + 0.0012073733378201723, + 0.0007833637646399438, + 0.0010105775436386466, + 0.0020189383067190647, + 0.0011208723299205303, + 0.0011181584559381008, + 0.0009230704745277762, + 0.01926947571337223, + 0.03134353458881378, + 0.0071566905826330185, + 0.014761185273528099, + 0.021974720060825348, + 0.01705341413617134, + 0.010228642262518406, + 0.009058461524546146, + 0.0013190534664317966, + 0.0021008809562772512, + 0.0011512574274092913, + 0.0015856185927987099, + 0.002043114509433508, + 0.0020910780876874924, + 0.0016928835539147258, + 0.0016942141810432076, + 0.0017530302284285426, + 0.0028777215629816055, + 0.0013399653835222125, + 0.002258269116282463, + 0.002327115973457694, + 0.001788633526302874, + 0.0017016301862895489, + 0.001428384566679597, + 0.0014459188096225262, + 0.0012743796687573195, + 0.0011258983286097646, + 0.0015659767668694258, + 0.002037121681496501, + 0.0019377253483980894, + 0.0014040535315871239, + 0.0017964969156309962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.6328403949737549 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204604357481003 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.06204604357481003, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204604357481003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.694412887096405, + 0.06935793906450272, + 0.05061691999435425, + 0.015233681537210941, + 0.02157127670943737, + 0.019039930775761604, + 0.04474582150578499, + 0.028573034331202507, + 0.03150293231010437, + 0.024945519864559174 + ], + "credit_map": [ + 0.18260976672172546, + 0.040550827980041504, + 0.03660714998841286, + 0.021773552522063255, + 0.06447374820709229, + 0.2733387351036072, + 0.038146473467350006, + 0.03691265732049942, + 0.009765053167939186, + 0.008590810932219028, + 0.0053921835497021675, + 0.007094196975231171, + 0.015929831191897392, + 0.0087357796728611, + 0.010768004693090916, + 0.0030820760875940323, + 0.005071608349680901, + 0.005025890655815601, + 0.0030882402788847685, + 0.007737221196293831, + 0.010756674222648144, + 0.005462647881358862, + 0.010500884614884853, + 0.0029737474396824837, + 0.0014940230175852776, + 0.0026644200552254915, + 0.0012359662214294076, + 0.0014000883093103766, + 0.0024257823824882507, + 0.0026507910806685686, + 0.001403758767992258, + 0.0019588498398661613, + 0.0027645560912787914, + 0.002885870635509491, + 0.0017654551193118095, + 0.0020543753635138273, + 0.0035012196749448776, + 0.0037594474852085114, + 0.0027591667603701353, + 0.0020811832509934902, + 0.0025190319865942, + 0.002711109584197402, + 0.00177720095962286, + 0.0017583128064870834, + 0.0032550084870308638, + 0.002098530763760209, + 0.002161687472835183, + 0.002759047318249941, + 0.009310985915362835, + 0.004010715056210756, + 0.00380166107788682, + 0.0031995591707527637, + 0.004497036337852478, + 0.0052639152854681015, + 0.0030224367510527372, + 0.011639509350061417, + 0.003400739748030901, + 0.005151108838617802, + 0.0017327972454950213, + 0.0030262297950685024, + 0.005273325834423304, + 0.003720986656844616, + 0.003774139564484358, + 0.0024937032721936703, + 0.0038491059094667435, + 0.005818822421133518, + 0.0019519089255481958, + 0.0041647194884717464, + 0.004024442285299301, + 0.004929082002490759, + 0.003105861833319068, + 0.003658985486254096, + 0.003641621908172965, + 0.002904606284573674, + 0.001619763788767159, + 0.0029192997608333826, + 0.0037656500935554504, + 0.00372088560834527, + 0.0024599977768957615, + 0.003913691267371178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06204604357481003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336940661072731 + ], + "opposite_logprob_tokens": [ + -3.40869402885437 + ], + "logprob_reward": -0.0336940661072731, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336940661072731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3621025085449219, + 0.1491561084985733, + 0.06578212976455688, + 0.036245621740818024, + 0.03958766162395477, + 0.03133746236562729, + 0.13152144849300385, + 0.07673142850399017, + 0.04140311852097511, + 0.06613247841596603 + ], + "credit_map": [ + 0.0979650691151619, + 0.02368411421775818, + 0.020990217104554176, + 0.012722046114504337, + 0.029463859274983406, + 0.13915875554084778, + 0.018864134326577187, + 0.019254334270954132, + 0.021816667169332504, + 0.016948731616139412, + 0.011927524581551552, + 0.01263404916971922, + 0.04093976691365242, + 0.01672987826168537, + 0.021283112466335297, + 0.006876376457512379, + 0.008457478135824203, + 0.011362134478986263, + 0.004893602803349495, + 0.0045564500615000725, + 0.012965850532054901, + 0.009992549195885658, + 0.0079778041690588, + 0.0055762589909136295, + 0.004547206684947014, + 0.00549484184011817, + 0.003061556490138173, + 0.002780255861580372, + 0.005566077772527933, + 0.006758555769920349, + 0.004701336845755577, + 0.003335789078846574, + 0.0049157594330608845, + 0.004456138703972101, + 0.003679482964798808, + 0.0048193661496043205, + 0.00572782289236784, + 0.004853167571127415, + 0.006429572589695454, + 0.004706351086497307, + 0.004441787488758564, + 0.0035213083028793335, + 0.0032353955321013927, + 0.0030578570440411568, + 0.005721786990761757, + 0.0031670203898102045, + 0.0038889236748218536, + 0.00430338317528367, + 0.0259642843157053, + 0.01679554022848606, + 0.008767127990722656, + 0.009652379900217056, + 0.01668611541390419, + 0.017431851476430893, + 0.009262902662158012, + 0.026961255818605423, + 0.009602897800505161, + 0.012836024165153503, + 0.0054029421880841255, + 0.008187760598957539, + 0.012065283954143524, + 0.013122406788170338, + 0.008406905457377434, + 0.007107201498001814, + 0.005077987443655729, + 0.007719141896814108, + 0.0023375549353659153, + 0.005444403737783432, + 0.005774473771452904, + 0.00580081669613719, + 0.004952011629939079, + 0.004296727944165468, + 0.011707259342074394, + 0.006848515011370182, + 0.004310728050768375, + 0.006459267344325781, + 0.009628391824662685, + 0.009681141935288906, + 0.006219248287379742, + 0.011277930811047554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0336940661072731 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a pop art style with bold, thick black outlines and highly saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a high-speed blender with a gleaming chrome base visible on a bright canary-yellow countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, transparent pitcher containing a vibrant, unnatural pink liquid that boils violently into a chaotic vortex visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, graphic bubbles depicted as perfect circles with sharp white highlights expanding and bursting on the liquid surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid cobalt blue color with stylized steam rising in bold black lines and Ben Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603500247001648 + ], + "opposite_logprob_tokens": [ + -1.9103500843048096 + ], + "logprob_reward": -0.1603500247001648, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603500247001648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18940091133117676, + 0.06427401304244995, + 0.14803743362426758, + 0.05204122141003609, + 0.02882487140595913, + 0.02437148429453373, + 0.4085206985473633, + 0.044206149876117706, + 0.01909712888300419, + 0.02122611366212368 + ], + "credit_map": [ + 0.04349149763584137, + 0.01883450709283352, + 0.015318256802856922, + 0.013341973535716534, + 0.024075541645288467, + 0.042789798229932785, + 0.013892327435314655, + 0.01765700802206993, + 0.00998140498995781, + 0.005437877960503101, + 0.0070211929269135, + 0.008055330254137516, + 0.01030629500746727, + 0.007329870015382767, + 0.006773661356419325, + 0.009368370287120342, + 0.019322309643030167, + 0.012409460730850697, + 0.008434252813458443, + 0.016962338238954544, + 0.043567027896642685, + 0.015243210829794407, + 0.025062646716833115, + 0.007036176975816488, + 0.006381528917700052, + 0.01038374099880457, + 0.00553131615743041, + 0.004969946574419737, + 0.007864479906857014, + 0.005977096036076546, + 0.004334023222327232, + 0.006599082611501217, + 0.0036945934407413006, + 0.0033466157037764788, + 0.0022498376201838255, + 0.002846110612154007, + 0.003300596261397004, + 0.00509580597281456, + 0.0028490759432315826, + 0.005442233290523291, + 0.003915685694664717, + 0.0026342272758483887, + 0.0022090596612542868, + 0.0028200463857501745, + 0.004220421891659498, + 0.0023968389723449945, + 0.0037015106063336134, + 0.002473692875355482, + 0.0990523174405098, + 0.0568021759390831, + 0.02426588535308838, + 0.033460069447755814, + 0.03793494030833244, + 0.06056682765483856, + 0.025782203301787376, + 0.07065626978874207, + 0.005612317472696304, + 0.005436452571302652, + 0.002743291202932596, + 0.005010556895285845, + 0.007632352411746979, + 0.005187977571040392, + 0.007135398685932159, + 0.005447798874229193, + 0.003638381138443947, + 0.002692731563001871, + 0.002178248018026352, + 0.0016690026968717575, + 0.002803549636155367, + 0.0026703665498644114, + 0.0015176230808719993, + 0.0019272243371233344, + 0.004629961680620909, + 0.0016986908158287406, + 0.0017857609782367945, + 0.0018357480876147747, + 0.0023539613466709852, + 0.0029223510064184666, + 0.0013986012199893594, + 0.0046010371297597885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.1603500247001648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_tensor([4598, 4598])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_tensor([4598, 4598])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..565497365eb7013a0f1c827c144ae2761af412dd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_tensor([4598, 4598])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up meticulously depicts a pile of clay-composite poker chips, precariously balanced, resting on the smooth, dark, highly polished surface of a wooden table. The chips, with their matte texture and slightly worn edges, are arranged in a vibrant pattern of red, white, and blue. A warm, overhead light casts a striking, focused glow, creating sharp, mirror-like highlights on the chips' rounded edges and bright reflections on the mirror-like tabletop, while the background dissolves into a soft, blurred darkness. The scene is tense and quiet. Suddenly, a subtle vibration travels through the surface, and the pile collapses instantly; the chips begin sliding outward in a chaotic cascade. They move rapidly at first, their colors blurring as they glide across the polished wood. Then, a steady rubbing and deliberate slowing of their movement causes them to sway and twirl, finally coming to rest in a scattered, overlapping pile—a moment frozen in chaos on the dark, gleaming surface.", + "seed": 26270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_4598_seed_26270_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03658465584740043, + "vlm_logprob": -0.03658465584740043, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03658465584740043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027412209659814835 + ], + "opposite_logprob_tokens": [ + -3.777412176132202 + ], + "logprob_reward": -0.027412209659814835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027412209659814835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.352845162153244, + 0.08332483470439911, + 0.21829165518283844, + 0.07490519434213638, + 0.07236044108867645, + 0.07436198741197586, + 0.032522305846214294, + 0.027464257553219795, + 0.035739172250032425, + 0.028185011819005013 + ], + "credit_map": [ + 0.04797928407788277, + 0.018802911043167114, + 0.049308255314826965, + 0.029136745259165764, + 0.017341231927275658, + 0.051151763647794724, + 0.04153638705611229, + 0.09758859127759933, + 0.007087063975632191, + 0.008367843925952911, + 0.010588424280285835, + 0.009936259128153324, + 0.007110794074833393, + 0.021616876125335693, + 0.00774026894941926, + 0.010877308435738087, + 0.03987305983901024, + 0.020219217985868454, + 0.05198991671204567, + 0.025386475026607513, + 0.021007375791668892, + 0.02444525808095932, + 0.015535294078290462, + 0.01983506605029106, + 0.009474335238337517, + 0.0063186343759298325, + 0.010417136363685131, + 0.00858165044337511, + 0.007745812647044659, + 0.011178265325725079, + 0.009040344506502151, + 0.012149017304182053, + 0.009857064113020897, + 0.005690326914191246, + 0.016089947894215584, + 0.006250179372727871, + 0.0053514232859015465, + 0.012234109453856945, + 0.0066525936126708984, + 0.010234802961349487, + 0.009136814624071121, + 0.010799295268952847, + 0.011385368183255196, + 0.007801754400134087, + 0.006860166788101196, + 0.00810158159583807, + 0.007916863076388836, + 0.012360140681266785, + 0.004381961654871702, + 0.0029635103419423103, + 0.004943192005157471, + 0.003243926912546158, + 0.004498198162764311, + 0.006764205638319254, + 0.0022710394114255905, + 0.003456271253526211, + 0.003958307672291994, + 0.0024926364421844482, + 0.007990281097590923, + 0.002462215255945921, + 0.002166859805583954, + 0.0027064175810664892, + 0.002652263268828392, + 0.003035277593880892, + 0.004029321018606424, + 0.003908918704837561, + 0.004314347170293331, + 0.0030204891227185726, + 0.007221107836812735, + 0.005863790865987539, + 0.004077808000147343, + 0.0033033897634595633, + 0.00384945934638381, + 0.0024953752290457487, + 0.003214736934751272, + 0.001869676518253982, + 0.00553792342543602, + 0.0056156134232878685, + 0.002499575028195977, + 0.0031026508659124374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.027412209659814835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296301335096359 + ], + "opposite_logprob_tokens": [ + -2.2379629611968994 + ], + "logprob_reward": -0.11296301335096359, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296301335096359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29454928636550903, + 0.0911865159869194, + 0.09590180963277817, + 0.27684804797172546, + 0.04709922894835472, + 0.04523741081357002, + 0.033118922263383865, + 0.04514208436012268, + 0.04935211315751076, + 0.021564587950706482 + ], + "credit_map": [ + 0.045551598072052, + 0.021072207018733025, + 0.04039870575070381, + 0.03139157593250275, + 0.016724305227398872, + 0.03322172164916992, + 0.04098840057849884, + 0.06520076096057892, + 0.010602355003356934, + 0.010807206854224205, + 0.014582586474716663, + 0.01427070889621973, + 0.009298304095864296, + 0.00985697377473116, + 0.010590564459562302, + 0.011177817359566689, + 0.014481659047305584, + 0.009167201817035675, + 0.01931151933968067, + 0.012798153795301914, + 0.00894052442163229, + 0.00991840474307537, + 0.011269507929682732, + 0.010014842264354229, + 0.03273891285061836, + 0.026769505813717842, + 0.04104692116379738, + 0.027512459084391594, + 0.025677315890789032, + 0.04121280089020729, + 0.030460277572274208, + 0.05142985284328461, + 0.008290122263133526, + 0.004607583861798048, + 0.009338654577732086, + 0.005508081987500191, + 0.004783229902386665, + 0.005285013932734728, + 0.004684473387897015, + 0.004602071363478899, + 0.005345455836504698, + 0.00490952655673027, + 0.007660882547497749, + 0.004688003566116095, + 0.003639501053839922, + 0.006828363984823227, + 0.004481473006308079, + 0.007684205658733845, + 0.004635936580598354, + 0.004563888534903526, + 0.004329466726630926, + 0.003958179149776697, + 0.002718224423006177, + 0.0022063367068767548, + 0.006915968377143145, + 0.0037909226957708597, + 0.007190583273768425, + 0.004563757684081793, + 0.006809245329350233, + 0.005229496862739325, + 0.0038904151879251003, + 0.00684267608448863, + 0.004629421513527632, + 0.005986487492918968, + 0.007722107227891684, + 0.004714712034910917, + 0.005840247496962547, + 0.0049388594925403595, + 0.005355296190828085, + 0.007910914719104767, + 0.005660629365593195, + 0.00720934197306633, + 0.0028482447378337383, + 0.001683984068222344, + 0.003654665779322386, + 0.0019937572069466114, + 0.003701196750625968, + 0.002050265669822693, + 0.0031696895603090525, + 0.002462784992530942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.11296301335096359 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041066608391702175 + ], + "opposite_logprob_tokens": [ + -5.504106521606445 + ], + "logprob_reward": -0.0041066608391702175, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041066608391702175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.466790109872818, + 0.07583869248628616, + 0.06233274191617966, + 0.09553419053554535, + 0.14120829105377197, + 0.026305468752980232, + 0.019545597955584526, + 0.02240939997136593, + 0.02651573158800602, + 0.06351973116397858 + ], + "credit_map": [ + 0.05118943750858307, + 0.02405649796128273, + 0.07151944935321808, + 0.0449475459754467, + 0.025165952742099762, + 0.07554709911346436, + 0.04645029082894325, + 0.12791383266448975, + 0.00767298461869359, + 0.007990265265107155, + 0.013256694190204144, + 0.011954176239669323, + 0.007192594930529594, + 0.0084579112008214, + 0.007296946831047535, + 0.012017118744552135, + 0.011713185347616673, + 0.004329670686274767, + 0.015365672297775745, + 0.004902413114905357, + 0.005953398533165455, + 0.008848126046359539, + 0.005034203641116619, + 0.0061860778369009495, + 0.012394663877785206, + 0.008917505852878094, + 0.007939688861370087, + 0.008827480487525463, + 0.01665913313627243, + 0.0180733073502779, + 0.010356089100241661, + 0.012366322800517082, + 0.02781582996249199, + 0.014322644099593163, + 0.03271738812327385, + 0.014031674712896347, + 0.01162085123360157, + 0.01163223572075367, + 0.015300522558391094, + 0.013767153024673462, + 0.00403759116306901, + 0.0032388472463935614, + 0.003899527946487069, + 0.002229805802926421, + 0.003799873637035489, + 0.003608534811064601, + 0.0024450658820569515, + 0.003046222496777773, + 0.0021740873344242573, + 0.0023899993393570185, + 0.002405621111392975, + 0.002216694876551628, + 0.003025206970050931, + 0.0034486702643334866, + 0.0020321852061897516, + 0.0018531326204538345, + 0.004244551062583923, + 0.0023749361280351877, + 0.0061179036274552345, + 0.0017255846178159118, + 0.002381377387791872, + 0.0018019316485151649, + 0.001875229412689805, + 0.001887885038740933, + 0.004029149655252695, + 0.002944831969216466, + 0.003398546250537038, + 0.0030677183531224728, + 0.004371209070086479, + 0.0020546892192214727, + 0.003831271082162857, + 0.002818315289914608, + 0.00935384351760149, + 0.005073817912489176, + 0.014218209311366081, + 0.004966937471181154, + 0.011298921890556812, + 0.0062059517949819565, + 0.007263080682605505, + 0.005138960666954517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0041066608391702175 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24484194815158844, + 0.1830720454454422, + 0.09563374519348145, + 0.2757348120212555, + 0.06911975890398026, + 0.03174204006791115, + 0.020513148978352547, + 0.028453869745135307, + 0.025367371737957, + 0.025521263480186462 + ], + "credit_map": [ + 0.027906283736228943, + 0.015551350079476833, + 0.0411037839949131, + 0.02459968440234661, + 0.01483780238777399, + 0.03010488674044609, + 0.029998905956745148, + 0.06073923781514168, + 0.01507814321666956, + 0.027166293933987617, + 0.02341150864958763, + 0.03561154752969742, + 0.020154954865574837, + 0.02979022078216076, + 0.014522360637784004, + 0.017337024211883545, + 0.011894387193024158, + 0.00997634045779705, + 0.014762945473194122, + 0.013223251327872276, + 0.015039488673210144, + 0.013910848647356033, + 0.009821729734539986, + 0.007004754617810249, + 0.0351552739739418, + 0.023518675938248634, + 0.031095009297132492, + 0.033372145146131516, + 0.030507748946547508, + 0.0563473105430603, + 0.0359991118311882, + 0.02973952516913414, + 0.014252745546400547, + 0.006962795276194811, + 0.014783303253352642, + 0.0066900127567350864, + 0.005942556541413069, + 0.0044188653118908405, + 0.008330287411808968, + 0.007739189080893993, + 0.00462839612737298, + 0.0044935159385204315, + 0.005296341609209776, + 0.00402990635484457, + 0.002311632502824068, + 0.003984812181442976, + 0.002936063101515174, + 0.004061370622366667, + 0.0031437373254448175, + 0.002367421519011259, + 0.003497079014778137, + 0.0021192103158682585, + 0.0018291505984961987, + 0.0030223424546420574, + 0.0019699577242136, + 0.0025642511900514364, + 0.005184061359614134, + 0.0035187972243875265, + 0.0036179088056087494, + 0.0022739465348422527, + 0.003682023612782359, + 0.002876982092857361, + 0.003431124845519662, + 0.003869025968015194, + 0.0035109901800751686, + 0.0018362082773819566, + 0.003742135828360915, + 0.001957154832780361, + 0.003042887430638075, + 0.005582062527537346, + 0.002807073760777712, + 0.0028888594824820757, + 0.003790965536609292, + 0.0027201662305742502, + 0.004085945431143045, + 0.0026250716764479876, + 0.004026329144835472, + 0.0026108219753950834, + 0.0029000863432884216, + 0.0027618769090622663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02980063669383526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640758693218231 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640758693218231, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640758693218231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1598825603723526, + 0.11327870190143585, + 0.13721127808094025, + 0.18108515441417694, + 0.06921110302209854, + 0.04665733501315117, + 0.042905546724796295, + 0.04212493821978569, + 0.07014059275388718, + 0.1375027745962143 + ], + "credit_map": [ + 0.014943907037377357, + 0.015386544167995453, + 0.02394367754459381, + 0.020925991237163544, + 0.013546600937843323, + 0.02191098779439926, + 0.01817134954035282, + 0.03105350397527218, + 0.010043272748589516, + 0.008329452946782112, + 0.015371290035545826, + 0.011452010832726955, + 0.011567140929400921, + 0.028135105967521667, + 0.011328532360494137, + 0.01705189049243927, + 0.02202215977013111, + 0.012241877615451813, + 0.030697626993060112, + 0.016105307266116142, + 0.015043429099023342, + 0.01500923465937376, + 0.011968878097832203, + 0.01412276178598404, + 0.02571430802345276, + 0.017597725614905357, + 0.033914417028427124, + 0.017483292147517204, + 0.012199541553854942, + 0.0302970539778471, + 0.01699153147637844, + 0.026887282729148865, + 0.013310771435499191, + 0.007198765408247709, + 0.013324808329343796, + 0.0070625473745167255, + 0.005625986959785223, + 0.00623280368745327, + 0.008398107253015041, + 0.008057317696511745, + 0.006957707926630974, + 0.004969144240021706, + 0.007062245160341263, + 0.004792769905179739, + 0.005207535810768604, + 0.005289427004754543, + 0.005132478196173906, + 0.007246024440973997, + 0.006733099464327097, + 0.005088247358798981, + 0.008117224089801311, + 0.004123417194932699, + 0.003267754567787051, + 0.0047304960899055, + 0.00558563694357872, + 0.0052596693858504295, + 0.0066076889634132385, + 0.004639310296624899, + 0.006033306010067463, + 0.004677006974816322, + 0.005117623135447502, + 0.00579238822683692, + 0.004370011854916811, + 0.00488760182633996, + 0.011453652754426003, + 0.006622633431106806, + 0.012405906803905964, + 0.006897187791764736, + 0.010045972652733326, + 0.00632678484544158, + 0.009285162203013897, + 0.0071032983250916, + 0.01818942092359066, + 0.012313284911215305, + 0.029814770445227623, + 0.014622504822909832, + 0.018785735592246056, + 0.015408975072205067, + 0.013205270282924175, + 0.015162807889282703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.008640758693218231 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up meticulously depicts a pile of clay-composite poker chips, precariously balanced, resting on the smooth, dark, highly polished surface of a wooden table. The chips, with their matte texture and slightly worn edges, are arranged in a vibrant pattern of red, white, and blue. A warm, overhead light casts a striking, focused glow, creating sharp, mirror-like highlights on the chips' rounded edges and bright reflections on the mirror-like tabletop, while the background dissolves into a soft, blurred darkness. The scene is tense and quiet. Suddenly, a subtle vibration travels through the surface, and the pile collapses instantly; the chips begin sliding outward in a chaotic cascade. They move rapidly at first, their colors blurring as they glide across the polished wood. Then, a steady rubbing and deliberate slowing of their movement causes them to sway and twirl, finally coming to rest in a scattered, overlapping pile—a moment frozen in chaos on the dark, gleaming surface.", + "seed": 26271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_4598_seed_26271_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11219695415347815, + "vlm_logprob": -0.11219695415347815, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.11219695415347815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1837121993303299 + ], + "opposite_logprob_tokens": [ + -1.8087122440338135 + ], + "logprob_reward": -0.1837121993303299, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1837121993303299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1702021062374115, + 0.07388844341039658, + 0.05214589834213257, + 0.038486745208501816, + 0.06586138904094696, + 0.09519679099321365, + 0.050853319466114044, + 0.05566208064556122, + 0.29995349049568176, + 0.0977497547864914 + ], + "credit_map": [ + 0.015177858993411064, + 0.013943974860012531, + 0.02611387148499489, + 0.01891917735338211, + 0.015624990686774254, + 0.024796834215521812, + 0.019174039363861084, + 0.03645135089755058, + 0.008887292817234993, + 0.007196025922894478, + 0.010111210867762566, + 0.008802894502878189, + 0.0055921939201653, + 0.012379701249301434, + 0.008832993917167187, + 0.012086127884685993, + 0.005389083176851273, + 0.00777607224881649, + 0.006972822360694408, + 0.007337472401559353, + 0.006375498604029417, + 0.007052311208099127, + 0.005400499328970909, + 0.005842138081789017, + 0.005694486200809479, + 0.0032614856027066708, + 0.006002747919410467, + 0.006763605400919914, + 0.002726631937548518, + 0.003840845078229904, + 0.00555260106921196, + 0.004644342698156834, + 0.005259418394416571, + 0.007631869520992041, + 0.008471296168863773, + 0.010407046414911747, + 0.008620242588222027, + 0.009374421089887619, + 0.009445572271943092, + 0.006651520263403654, + 0.01625320501625538, + 0.011393590830266476, + 0.01353680994361639, + 0.01313850749284029, + 0.01089927926659584, + 0.007102516945451498, + 0.010259442962706089, + 0.012613438069820404, + 0.007600807584822178, + 0.003115247469395399, + 0.019230710342526436, + 0.010131093673408031, + 0.0016716101672500372, + 0.0017174737295135856, + 0.0036882904823869467, + 0.0036980891600251198, + 0.010085251182317734, + 0.004190127830952406, + 0.005783688277006149, + 0.013953630812466145, + 0.002878813538700342, + 0.0033118866849690676, + 0.005454180762171745, + 0.010004500858485699, + 0.036548685282468796, + 0.1101873442530632, + 0.010996895842254162, + 0.06466566771268845, + 0.02066764421761036, + 0.019512547180056572, + 0.025705084204673767, + 0.01166960597038269, + 0.005775713361799717, + 0.00793424341827631, + 0.0035792463459074497, + 0.0277313981205225, + 0.004265185445547104, + 0.0021165236830711365, + 0.042595066130161285, + 0.0037523715291172266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.1837121993303299 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28140735626220703 + ], + "opposite_logprob_tokens": [ + -1.406407356262207 + ], + "logprob_reward": -0.28140735626220703, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28140735626220703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12496883422136307, + 0.0747176930308342, + 0.06383994221687317, + 0.050858039408922195, + 0.10243014246225357, + 0.039812203496694565, + 0.04726565256714821, + 0.040305618196725845, + 0.2530084550380707, + 0.20279347896575928 + ], + "credit_map": [ + 0.011653992347419262, + 0.012684437446296215, + 0.016285333782434464, + 0.01768919825553894, + 0.015962744131684303, + 0.01624302566051483, + 0.01493529137223959, + 0.019514799118041992, + 0.006569002289324999, + 0.00507608987390995, + 0.011343829333782196, + 0.009266029112040997, + 0.005951165221631527, + 0.00897293258458376, + 0.016304079443216324, + 0.011234558187425137, + 0.005231477785855532, + 0.007334760390222073, + 0.008916137740015984, + 0.011652495712041855, + 0.006598459556698799, + 0.008625216782093048, + 0.007659675553441048, + 0.007821710780262947, + 0.004979813937097788, + 0.004099622368812561, + 0.008261901326477528, + 0.009402592666447163, + 0.004114085342735052, + 0.004596711602061987, + 0.009347692131996155, + 0.0060556186363101006, + 0.00834625493735075, + 0.010916299186646938, + 0.01276964321732521, + 0.01636613719165325, + 0.011383715085685253, + 0.01183324959129095, + 0.016124190762639046, + 0.014690647833049297, + 0.0028809711802750826, + 0.003439750289544463, + 0.008795160800218582, + 0.009041176177561283, + 0.003487675916403532, + 0.0030533496756106615, + 0.004485229030251503, + 0.0046288869343698025, + 0.008161943405866623, + 0.003898580791428685, + 0.01496097818017006, + 0.0075248111970722675, + 0.0030076808761805296, + 0.002845929702743888, + 0.003385507734492421, + 0.0034802183508872986, + 0.005354945082217455, + 0.0033026649616658688, + 0.007889383472502232, + 0.009298882447183132, + 0.002371510025113821, + 0.0028866904322057962, + 0.00352809252217412, + 0.0056734466925263405, + 0.017083020880818367, + 0.05181382596492767, + 0.01205603126436472, + 0.0941547304391861, + 0.011687593534588814, + 0.012735473923385143, + 0.04142298176884651, + 0.012054777704179287, + 0.008562001399695873, + 0.012585150077939034, + 0.008669442497193813, + 0.06441446393728256, + 0.00889535341411829, + 0.004939664155244827, + 0.0848134234547615, + 0.009913969784975052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.28140735626220703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647049613296986 + ], + "opposite_logprob_tokens": [ + -4.8826470375061035 + ], + "logprob_reward": -0.007647049613296986, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647049613296986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08101561665534973, + 0.08599274605512619, + 0.15427882969379425, + 0.04616968333721161, + 0.11742347478866577, + 0.12736725807189941, + 0.09880340844392776, + 0.050255365669727325, + 0.14515258371829987, + 0.09354103356599808 + ], + "credit_map": [ + 0.008553590625524521, + 0.008093937300145626, + 0.011382813565433025, + 0.008745796047151089, + 0.009279429912567139, + 0.010718422941863537, + 0.0088988421484828, + 0.015342782251536846, + 0.008335556834936142, + 0.00573127344250679, + 0.011795946396887302, + 0.00935707800090313, + 0.006958364509046078, + 0.007980232127010822, + 0.023339636623859406, + 0.01249465811997652, + 0.01478924322873354, + 0.020661024376749992, + 0.01952480338513851, + 0.019799938425421715, + 0.01864617131650448, + 0.022751623764634132, + 0.01778183877468109, + 0.020324185490608215, + 0.005606179591268301, + 0.004881716798990965, + 0.006736233830451965, + 0.006497799418866634, + 0.006173148285597563, + 0.006109439302235842, + 0.00560630951076746, + 0.004558857996016741, + 0.009598390199244022, + 0.012818368151783943, + 0.01314598973840475, + 0.018361572176218033, + 0.01124893594533205, + 0.01527140848338604, + 0.01932682655751705, + 0.017651982605457306, + 0.019776972010731697, + 0.013498636893928051, + 0.019418083131313324, + 0.021157683804631233, + 0.01287646684795618, + 0.010448222979903221, + 0.011769959703087807, + 0.018421238288283348, + 0.017001304775476456, + 0.007942859083414078, + 0.022968575358390808, + 0.016970515251159668, + 0.00701451301574707, + 0.007028198800981045, + 0.008396536111831665, + 0.011480908840894699, + 0.006214008200913668, + 0.004478040151298046, + 0.008104225620627403, + 0.011225120164453983, + 0.0030677812173962593, + 0.0035770356189459562, + 0.0058584813959896564, + 0.0077306716702878475, + 0.02058306522667408, + 0.05419926345348358, + 0.005829290021210909, + 0.0280414167791605, + 0.009572378359735012, + 0.00939386710524559, + 0.011748949065804482, + 0.00578435929492116, + 0.005867605563253164, + 0.007410435471683741, + 0.008437899872660637, + 0.021959224715828896, + 0.004734770394861698, + 0.004445584490895271, + 0.03218386322259903, + 0.00850165355950594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.007647049613296986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002559304237366 + ], + "opposite_logprob_tokens": [ + -2.695025682449341 + ], + "logprob_reward": -0.07002559304237366, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002559304237366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05583013594150543, + 0.10325915366411209, + 0.06736619770526886, + 0.018917908892035484, + 0.034628719091415405, + 0.03885708004236221, + 0.06489566713571548, + 0.031051110476255417, + 0.5077106952667236, + 0.07748333364725113 + ], + "credit_map": [ + 0.005796062760055065, + 0.004632981028407812, + 0.00783692765980959, + 0.007751774042844772, + 0.0077653974294662476, + 0.005549015011638403, + 0.006559721194207668, + 0.009938254952430725, + 0.008864178322255611, + 0.007208641618490219, + 0.015875261276960373, + 0.012609738856554031, + 0.006639508996158838, + 0.01211157813668251, + 0.023213203996419907, + 0.01673703081905842, + 0.007242032792419195, + 0.008003821596503258, + 0.00986800529062748, + 0.013346864841878414, + 0.006823042407631874, + 0.008452038280665874, + 0.007244403474032879, + 0.0063859871588647366, + 0.001497384742833674, + 0.002034416189417243, + 0.0027858607936650515, + 0.0026044894475489855, + 0.002354654250666499, + 0.0017698426963761449, + 0.002974854316562414, + 0.0028964041266590357, + 0.004274948500096798, + 0.004994278308004141, + 0.0051885643042624, + 0.0061944881454110146, + 0.0030354373157024384, + 0.002770014340057969, + 0.004357305821031332, + 0.0038136786315590143, + 0.004997066222131252, + 0.00345284934155643, + 0.008432388305664062, + 0.0053610894829034805, + 0.0034494551364332438, + 0.003085837699472904, + 0.004759359173476696, + 0.005319029558449984, + 0.013677088543772697, + 0.00618177792057395, + 0.01174242701381445, + 0.008746184408664703, + 0.005279705859720707, + 0.005078548099845648, + 0.005902172531932592, + 0.008287751115858555, + 0.003071764251217246, + 0.0030973264947533607, + 0.005443561356514692, + 0.006283627822995186, + 0.002254761289805174, + 0.003227957058697939, + 0.0038042888045310974, + 0.003867817111313343, + 0.12054398655891418, + 0.20765723288059235, + 0.02233380451798439, + 0.032306112349033356, + 0.03646870702505112, + 0.03245287761092186, + 0.03462184965610504, + 0.021326081827282906, + 0.006441744044423103, + 0.007635707035660744, + 0.005663403309881687, + 0.0176392775028944, + 0.0034956769086420536, + 0.002963610226288438, + 0.02963922917842865, + 0.004004674963653088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.07002559304237366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05412016063928604, + 0.07215011119842529, + 0.057768844068050385, + 0.0384162999689579, + 0.034050263464450836, + 0.09451513737440109, + 0.0376310721039772, + 0.06684176623821259, + 0.41226470470428467, + 0.1322416365146637 + ], + "credit_map": [ + 0.005173926241695881, + 0.004671553149819374, + 0.008778306655585766, + 0.005788113921880722, + 0.00863152276724577, + 0.005089765414595604, + 0.004723886027932167, + 0.01126308087259531, + 0.008076472207903862, + 0.005251422058790922, + 0.010327134281396866, + 0.008791175670921803, + 0.005319817457348108, + 0.007863696664571762, + 0.014353408478200436, + 0.012166979722678661, + 0.00716604245826602, + 0.007306226063519716, + 0.007766249589622021, + 0.007563829887658358, + 0.006572731304913759, + 0.008228790946304798, + 0.006179340649396181, + 0.006985630374401808, + 0.005675573367625475, + 0.004272700287401676, + 0.005376216024160385, + 0.0070195975713431835, + 0.003163763089105487, + 0.0034668997395783663, + 0.005189010873436928, + 0.004252532962709665, + 0.003498294623568654, + 0.002853632904589176, + 0.004202455282211304, + 0.004789207596331835, + 0.0026829440612345934, + 0.005300021264702082, + 0.004821717273443937, + 0.00590198440477252, + 0.018544059246778488, + 0.008813099004328251, + 0.01609954982995987, + 0.013549414463341236, + 0.007895289920270443, + 0.007296356372535229, + 0.008737960830330849, + 0.013579398393630981, + 0.004228543024510145, + 0.0035819753538817167, + 0.005892014130949974, + 0.00576603040099144, + 0.003883713623508811, + 0.003940245136618614, + 0.004681531339883804, + 0.005657016299664974, + 0.010123353451490402, + 0.006792791187763214, + 0.00993096362799406, + 0.010580810718238354, + 0.004807415883988142, + 0.005334523040801287, + 0.008043453097343445, + 0.011228449642658234, + 0.06628664582967758, + 0.15497274696826935, + 0.017457060515880585, + 0.07289116084575653, + 0.027636002749204636, + 0.02619992196559906, + 0.030337274074554443, + 0.016483841463923454, + 0.014216850511729717, + 0.00992509163916111, + 0.01004253514111042, + 0.02718573436141014, + 0.005827749148011208, + 0.00559601653367281, + 0.05088423192501068, + 0.008563424460589886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.01819257251918316 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up meticulously depicts a pile of clay-composite poker chips, precariously balanced, resting on the smooth, dark, highly polished surface of a wooden table. The chips, with their matte texture and slightly worn edges, are arranged in a vibrant pattern of red, white, and blue. A warm, overhead light casts a striking, focused glow, creating sharp, mirror-like highlights on the chips' rounded edges and bright reflections on the mirror-like tabletop, while the background dissolves into a soft, blurred darkness. The scene is tense and quiet. Suddenly, a subtle vibration travels through the surface, and the pile collapses instantly; the chips begin sliding outward in a chaotic cascade. They move rapidly at first, their colors blurring as they glide across the polished wood. Then, a steady rubbing and deliberate slowing of their movement causes them to sway and twirl, finally coming to rest in a scattered, overlapping pile—a moment frozen in chaos on the dark, gleaming surface.", + "seed": 26272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_4598_seed_26272_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05448886808007956, + "vlm_logprob": -0.05448886808007956, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05448886808007956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020170921459794044 + ], + "opposite_logprob_tokens": [ + -4.27017068862915 + ], + "logprob_reward": -0.020170921459794044, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020170921459794044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14699402451515198, + 0.07103439420461655, + 0.052073970437049866, + 0.06543612480163574, + 0.0659472718834877, + 0.044168874621391296, + 0.1466994285583496, + 0.12775854766368866, + 0.1295071840286255, + 0.15038016438484192 + ], + "credit_map": [ + 0.021916652098298073, + 0.010272457264363766, + 0.019534626975655556, + 0.011768054217100143, + 0.01006778609007597, + 0.03051859512925148, + 0.024315889924764633, + 0.018599973991513252, + 0.01163521595299244, + 0.007634728215634823, + 0.009804033674299717, + 0.011365729384124279, + 0.006831987760961056, + 0.005406437907367945, + 0.00952927302569151, + 0.008826990611851215, + 0.006330593954771757, + 0.0060056536458432674, + 0.007232040632516146, + 0.004446345381438732, + 0.0075902557000517845, + 0.007159488741308451, + 0.0077784243039786816, + 0.005531167611479759, + 0.010762248188257217, + 0.0062059382908046246, + 0.0088868523016572, + 0.007710941135883331, + 0.007464970462024212, + 0.008113795891404152, + 0.005902226082980633, + 0.010389151982963085, + 0.00899573415517807, + 0.008235990069806576, + 0.005394594743847847, + 0.0074630300514400005, + 0.009870599955320358, + 0.009157122112810612, + 0.0065422882325947285, + 0.010287908837199211, + 0.0050064269453287125, + 0.0036288469564169645, + 0.002134959679096937, + 0.0037734925281256437, + 0.010038608685135841, + 0.013587124645709991, + 0.0030989896040409803, + 0.0029004255775362253, + 0.02343369647860527, + 0.009616740047931671, + 0.00564265251159668, + 0.004491365514695644, + 0.034398894757032394, + 0.05523747578263283, + 0.006971707567572594, + 0.006906887982040644, + 0.014800955541431904, + 0.017439981922507286, + 0.014535505324602127, + 0.011203627102077007, + 0.025702888146042824, + 0.019094988703727722, + 0.013918490149080753, + 0.011062117293477058, + 0.021815959364175797, + 0.02136846072971821, + 0.011493903584778309, + 0.007448446471244097, + 0.021292582154273987, + 0.032570257782936096, + 0.00638366537168622, + 0.007133902981877327, + 0.0338360071182251, + 0.04535777121782303, + 0.007213171571493149, + 0.008010141551494598, + 0.010872233659029007, + 0.026952315121889114, + 0.007543691899627447, + 0.010594837367534637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.020170921459794044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296471208333969 + ], + "opposite_logprob_tokens": [ + -2.237964630126953 + ], + "logprob_reward": -0.11296471208333969, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296471208333969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10849249362945557, + 0.07574771344661713, + 0.052244797348976135, + 0.16966605186462402, + 0.15077324211597443, + 0.11000388115644455, + 0.04235995560884476, + 0.08270632475614548, + 0.1316099762916565, + 0.07639559358358383 + ], + "credit_map": [ + 0.01441933959722519, + 0.0074724480509757996, + 0.009827392175793648, + 0.008185471408069134, + 0.0075855618342757225, + 0.03051799349486828, + 0.00953891035169363, + 0.020945372059941292, + 0.010562971234321594, + 0.006295415107160807, + 0.007997510954737663, + 0.008943737484514713, + 0.009498928673565388, + 0.013797461986541748, + 0.010330633260309696, + 0.008321053348481655, + 0.006361035164445639, + 0.00502136442810297, + 0.004786973353475332, + 0.004752257373183966, + 0.00971173495054245, + 0.009175240062177181, + 0.005382022820413113, + 0.007054167799651623, + 0.026519833132624626, + 0.01634356565773487, + 0.027730248868465424, + 0.020273029804229736, + 0.02131565846502781, + 0.019244709983468056, + 0.01751530170440674, + 0.020723698660731316, + 0.007135594263672829, + 0.007415503729134798, + 0.006456084083765745, + 0.006363990716636181, + 0.05214376747608185, + 0.05050904303789139, + 0.00997293833643198, + 0.010776320472359657, + 0.008658085018396378, + 0.0071936361491680145, + 0.004746494349092245, + 0.005049834493547678, + 0.034480541944503784, + 0.040191031992435455, + 0.005491693038493395, + 0.0041925543919205666, + 0.007298648823052645, + 0.004724671132862568, + 0.002767100464552641, + 0.004491102881729603, + 0.006343969143927097, + 0.008895862847566605, + 0.0034553708974272013, + 0.004383228253573179, + 0.007091766223311424, + 0.007553871721029282, + 0.003863288089632988, + 0.004407922737300396, + 0.022177964448928833, + 0.0256649199873209, + 0.007470138370990753, + 0.00447644479572773, + 0.019280212000012398, + 0.009465046226978302, + 0.005827414337545633, + 0.003696658182889223, + 0.032252661883831024, + 0.05200599506497383, + 0.0040235067717731, + 0.005058473441749811, + 0.021342609077692032, + 0.022800715640187263, + 0.004187706392258406, + 0.004461060743778944, + 0.005615290720015764, + 0.008804154582321644, + 0.004415824078023434, + 0.004768225830048323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.11296471208333969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647168796509504 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647168796509504, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647168796509504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11284632235765457, + 0.0997418463230133, + 0.037293124943971634, + 0.07763542234897614, + 0.11608375608921051, + 0.01630466803908348, + 0.08181452751159668, + 0.18636713922023773, + 0.04769112169742584, + 0.2242220789194107 + ], + "credit_map": [ + 0.015150229446589947, + 0.010955986566841602, + 0.010115593671798706, + 0.012325947172939777, + 0.009195868857204914, + 0.018783267587423325, + 0.019918637350201607, + 0.016400787979364395, + 0.015980787575244904, + 0.010995853692293167, + 0.010851276107132435, + 0.012245865538716316, + 0.012645404785871506, + 0.012537330389022827, + 0.013121383264660835, + 0.011363948695361614, + 0.005105026066303253, + 0.00503366906195879, + 0.00448942743241787, + 0.0030794234480708838, + 0.005171906668692827, + 0.004768917802721262, + 0.003536943346261978, + 0.006107810419052839, + 0.012182452715933323, + 0.006956302095204592, + 0.011154571548104286, + 0.00979948416352272, + 0.00944566261023283, + 0.010115230455994606, + 0.008010243065655231, + 0.00997147150337696, + 0.01217572670429945, + 0.012527319602668285, + 0.01146437507122755, + 0.009547933004796505, + 0.02678576484322548, + 0.020488815382122993, + 0.009866763837635517, + 0.013227062299847603, + 0.002106285886839032, + 0.0022718911059200764, + 0.002187988953664899, + 0.0015268160495907068, + 0.0023848735727369785, + 0.0034230260644108057, + 0.0013138994108885527, + 0.001089886762201786, + 0.01329964678734541, + 0.004692143760621548, + 0.002481436589732766, + 0.0021723280660808086, + 0.020416153594851494, + 0.032810039818286896, + 0.0024474794045090675, + 0.0034952927380800247, + 0.02183559164404869, + 0.024975726380944252, + 0.00846779067069292, + 0.007599011063575745, + 0.04391768202185631, + 0.06562317162752151, + 0.006547835655510426, + 0.007400325033813715, + 0.009041215293109417, + 0.008211708627641201, + 0.0029912421014159918, + 0.0026398946065455675, + 0.007675627712160349, + 0.010905072093009949, + 0.0033884576987475157, + 0.0028379077557474375, + 0.05506904050707817, + 0.07285874336957932, + 0.010936444625258446, + 0.010341489687561989, + 0.014642219990491867, + 0.03883001580834389, + 0.009649479761719704, + 0.011894638650119305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.004647168796509504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703010439872742 + ], + "opposite_logprob_tokens": [ + -2.12703013420105 + ], + "logprob_reward": -0.12703010439872742, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703010439872742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1835634559392929, + 0.060901109129190445, + 0.12455331534147263, + 0.07248258590698242, + 0.33359917998313904, + 0.0168673824518919, + 0.022350860759615898, + 0.02564306929707527, + 0.013800009153783321, + 0.1462389975786209 + ], + "credit_map": [ + 0.0344802588224411, + 0.014556796289980412, + 0.01931910216808319, + 0.009735639207065105, + 0.013925166800618172, + 0.0487903468310833, + 0.016813576221466064, + 0.025942588225007057, + 0.008912255056202412, + 0.009178625419735909, + 0.006081932689994574, + 0.009962971322238445, + 0.006069737020879984, + 0.004750118590891361, + 0.007959776557981968, + 0.007985695265233517, + 0.014970844611525536, + 0.012064453214406967, + 0.01565106026828289, + 0.011990825645625591, + 0.019181273877620697, + 0.02637011744081974, + 0.011071731336414814, + 0.013253017328679562, + 0.010245207697153091, + 0.00675814226269722, + 0.009798556566238403, + 0.011604405008256435, + 0.007545717991888523, + 0.010028536431491375, + 0.007411705330014229, + 0.009090312756597996, + 0.013027309440076351, + 0.01754070445895195, + 0.008624257519841194, + 0.008880591951310635, + 0.13882064819335938, + 0.12470735609531403, + 0.011162366718053818, + 0.010835940018296242, + 0.002273771446198225, + 0.0021371422335505486, + 0.001305522513575852, + 0.0014356836909428239, + 0.003053306369110942, + 0.004270852077752352, + 0.0012484126491472125, + 0.0011426913551986217, + 0.0035317004658281803, + 0.0021213446743786335, + 0.0016028331592679024, + 0.0013847948284819722, + 0.0041213082149624825, + 0.006234798580408096, + 0.0014488365268334746, + 0.0019052451243624091, + 0.0034210761077702045, + 0.004600964952260256, + 0.0016835746355354786, + 0.0013997489586472511, + 0.004356759134680033, + 0.00663891714066267, + 0.0013616863870993257, + 0.002180343959480524, + 0.00203335820697248, + 0.0022743805311620235, + 0.0016950367717072368, + 0.0014043493429198861, + 0.0012398312101140618, + 0.0020300487522035837, + 0.0014468523440882564, + 0.0016761525766924024, + 0.017923809587955475, + 0.029337868094444275, + 0.008808174170553684, + 0.006898925639688969, + 0.022122632712125778, + 0.04646351933479309, + 0.006847096607089043, + 0.007836981676518917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.12703010439872742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15894971787929535, + 0.09266354888677597, + 0.04683174192905426, + 0.07383600622415543, + 0.2790282666683197, + 0.05968261510133743, + 0.01707693375647068, + 0.09208681434392929, + 0.12360639125108719, + 0.05623798072338104 + ], + "credit_map": [ + 0.026578469201922417, + 0.014251845888793468, + 0.014271087013185024, + 0.011663813143968582, + 0.010731060057878494, + 0.03825254738330841, + 0.01613554172217846, + 0.027065351605415344, + 0.012660262174904346, + 0.009379777126014233, + 0.01249307207763195, + 0.012541892006993294, + 0.007522230967879295, + 0.016628528013825417, + 0.009674972854554653, + 0.011762818321585655, + 0.00587299233302474, + 0.0050467755645513535, + 0.005733822472393513, + 0.003495176089927554, + 0.00804970320314169, + 0.009308147244155407, + 0.0033935995306819677, + 0.005931525491178036, + 0.010532103478908539, + 0.006391932722181082, + 0.010510683991014957, + 0.01284467801451683, + 0.007672639563679695, + 0.008489490486681461, + 0.008317702449858189, + 0.009076776914298534, + 0.009182834066450596, + 0.011301044374704361, + 0.007809841074049473, + 0.006755192764103413, + 0.11457620561122894, + 0.10753978788852692, + 0.011618084274232388, + 0.010245277546346188, + 0.0034552335273474455, + 0.00423164339736104, + 0.002136422321200371, + 0.002307882299646735, + 0.019944824278354645, + 0.0213735643774271, + 0.003973302431404591, + 0.0022597433999180794, + 0.002574034733697772, + 0.0020197308622300625, + 0.001492472249083221, + 0.0013605221174657345, + 0.0026842961087822914, + 0.004087076056748629, + 0.0013728768099099398, + 0.0014859243528917432, + 0.007095640990883112, + 0.01055427361279726, + 0.004281181376427412, + 0.004610228352248669, + 0.024223364889621735, + 0.03152720257639885, + 0.005474206060171127, + 0.004320713225752115, + 0.01714012585580349, + 0.009545857086777687, + 0.005042010452598333, + 0.004163101781159639, + 0.029956338927149773, + 0.0477394200861454, + 0.004150630906224251, + 0.005868913140147924, + 0.014347364194691181, + 0.016575606539845467, + 0.0034543322399258614, + 0.003885583719238639, + 0.003399054054170847, + 0.007744918577373028, + 0.003237779252231121, + 0.0035933437757194042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.007631433662027121 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up meticulously depicts a pile of clay-composite poker chips, precariously balanced, resting on the smooth, dark, highly polished surface of a wooden table. The chips, with their matte texture and slightly worn edges, are arranged in a vibrant pattern of red, white, and blue. A warm, overhead light casts a striking, focused glow, creating sharp, mirror-like highlights on the chips' rounded edges and bright reflections on the mirror-like tabletop, while the background dissolves into a soft, blurred darkness. The scene is tense and quiet. Suddenly, a subtle vibration travels through the surface, and the pile collapses instantly; the chips begin sliding outward in a chaotic cascade. They move rapidly at first, their colors blurring as they glide across the polished wood. Then, a steady rubbing and deliberate slowing of their movement causes them to sway and twirl, finally coming to rest in a scattered, overlapping pile—a moment frozen in chaos on the dark, gleaming surface.", + "seed": 26273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_3/index_4598_seed_26273_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.051346811652183535, + "vlm_logprob": -0.051346811652183535, + "vlm_yes_no_margin": 3.674999952316284, + "vlm_token_credit": -0.051346811652183535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019733767956495285 + ], + "opposite_logprob_tokens": [ + -4.769733905792236 + ], + "logprob_reward": -0.019733767956495285, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.019733767956495285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1349768489599228, + 0.1269264817237854, + 0.059468552470207214, + 0.14323323965072632, + 0.019588135182857513, + 0.21248099207878113, + 0.09230541437864304, + 0.026591479778289795, + 0.15622161328792572, + 0.02820725180208683 + ], + "credit_map": [ + 0.022687962278723717, + 0.008800171315670013, + 0.025441018864512444, + 0.008014283142983913, + 0.01376058254390955, + 0.022916125133633614, + 0.008434437215328217, + 0.024922272190451622, + 0.03226350620388985, + 0.008849828504025936, + 0.012477041222155094, + 0.0067232814617455006, + 0.011422467418015003, + 0.024941086769104004, + 0.008616461418569088, + 0.021632809191942215, + 0.011262929067015648, + 0.0045365262776613235, + 0.009931381791830063, + 0.0035888426937162876, + 0.00623302161693573, + 0.0097896046936512, + 0.003748121438547969, + 0.01037812139838934, + 0.022840453311800957, + 0.014119356870651245, + 0.025826964527368546, + 0.013290098868310452, + 0.01837056316435337, + 0.01587842032313347, + 0.015168598853051662, + 0.01773878000676632, + 0.002409439766779542, + 0.002631597686558962, + 0.0024772202596068382, + 0.0020681358873844147, + 0.003223762148991227, + 0.0021902937442064285, + 0.002409192966297269, + 0.002178492024540901, + 0.010561028495430946, + 0.07003186643123627, + 0.010503777302801609, + 0.012309269048273563, + 0.07350439578294754, + 0.013750148937106133, + 0.01247670128941536, + 0.00934381503611803, + 0.016210271045565605, + 0.012907166965305805, + 0.00640058983117342, + 0.0068300808779895306, + 0.013762826099991798, + 0.012586530297994614, + 0.016493016853928566, + 0.0071149300783872604, + 0.003956464119255543, + 0.004730757791548967, + 0.0021931836381554604, + 0.0030503589659929276, + 0.0045949374325573444, + 0.00294471881352365, + 0.0027617490850389004, + 0.002359311794862151, + 0.013972318731248379, + 0.017035089433193207, + 0.022188453003764153, + 0.011143489740788937, + 0.04045000299811363, + 0.035792142152786255, + 0.007066753227263689, + 0.008573370054364204, + 0.00409328006207943, + 0.0044339741580188274, + 0.0023148516193032265, + 0.0024303486570715904, + 0.004056398291140795, + 0.003689755452796817, + 0.0033586828503757715, + 0.003829962108284235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.019733767956495285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283962547779083 + ], + "opposite_logprob_tokens": [ + -2.0178396701812744 + ], + "logprob_reward": -0.14283962547779083, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283962547779083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0807323306798935, + 0.13434138894081116, + 0.046100400388240814, + 0.07677580416202545, + 0.050138428807258606, + 0.22185873985290527, + 0.05288108438253403, + 0.10238874703645706, + 0.16280139982700348, + 0.07198165357112885 + ], + "credit_map": [ + 0.013707131147384644, + 0.004703061189502478, + 0.010889793746173382, + 0.0056737810373306274, + 0.01057397574186325, + 0.01315704733133316, + 0.008098413236439228, + 0.013929128646850586, + 0.03414308279752731, + 0.009652120061218739, + 0.015007277950644493, + 0.006812747567892075, + 0.012565604411065578, + 0.02513226307928562, + 0.007432826794683933, + 0.023595446720719337, + 0.00675042113289237, + 0.003643327858299017, + 0.00690797483548522, + 0.0030622852500528097, + 0.0061043789610266685, + 0.007037276402115822, + 0.003922218922525644, + 0.00867251493036747, + 0.011146986857056618, + 0.006983495783060789, + 0.012416333891451359, + 0.0071502490900456905, + 0.010566049255430698, + 0.012267702259123325, + 0.007701068185269833, + 0.008543917909264565, + 0.005485325586050749, + 0.005661186762154102, + 0.006211759988218546, + 0.0057161469012498856, + 0.010219060815870762, + 0.004663196858018637, + 0.006531668361276388, + 0.005650081671774387, + 0.0184757262468338, + 0.04180781543254852, + 0.020492201671004295, + 0.020236721262335777, + 0.044133126735687256, + 0.02718924731016159, + 0.0306260846555233, + 0.01889782026410103, + 0.0064185974188148975, + 0.007375161163508892, + 0.0039627430960536, + 0.005911098327487707, + 0.007914853282272816, + 0.007131209131330252, + 0.007382615003734827, + 0.006784807424992323, + 0.012293001636862755, + 0.013915604911744595, + 0.009915006347000599, + 0.013601578772068024, + 0.015282677486538887, + 0.01693938858807087, + 0.010522449389100075, + 0.00991904642432928, + 0.013890044763684273, + 0.026050839573144913, + 0.012967864982783794, + 0.011533195152878761, + 0.030302098020911217, + 0.03649985045194626, + 0.016868937760591507, + 0.014688569121062756, + 0.0114445136860013, + 0.016052644699811935, + 0.004732124507427216, + 0.005492834839969873, + 0.011003297753632069, + 0.010501449927687645, + 0.006728718988597393, + 0.006026070099323988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.14283962547779083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962086841464043 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005962086841464043, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962086841464043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12557587027549744, + 0.15785950422286987, + 0.032169997692108154, + 0.12533901631832123, + 0.05827987939119339, + 0.13329961895942688, + 0.06633476912975311, + 0.03630943223834038, + 0.23313777148723602, + 0.03169412538409233 + ], + "credit_map": [ + 0.027710089460015297, + 0.008766611106693745, + 0.015550417825579643, + 0.0076454561203718185, + 0.013979489915072918, + 0.021766290068626404, + 0.011728839948773384, + 0.018428685143589973, + 0.037699926644563675, + 0.011095136404037476, + 0.016555938869714737, + 0.008040719665586948, + 0.015213429927825928, + 0.03019472025334835, + 0.011326128616929054, + 0.027733493596315384, + 0.007304775528609753, + 0.0020609369967132807, + 0.0046556429006159306, + 0.0014304508222267032, + 0.003417661879211664, + 0.0059971935115754604, + 0.0016994625329971313, + 0.005603873170912266, + 0.014753302559256554, + 0.013333361595869064, + 0.020324504002928734, + 0.008853420615196228, + 0.022006338462233543, + 0.016250237822532654, + 0.015796970576047897, + 0.014020866714417934, + 0.009420948103070259, + 0.00532968295738101, + 0.006978434044867754, + 0.00623758789151907, + 0.00837752502411604, + 0.007056611590087414, + 0.00827004574239254, + 0.006609044969081879, + 0.011407327838242054, + 0.04274607077240944, + 0.006239889655262232, + 0.008567151613533497, + 0.04119873046875, + 0.008800539188086987, + 0.006442660000175238, + 0.007897257804870605, + 0.009773056022822857, + 0.010592855513095856, + 0.005036543123424053, + 0.005245461128652096, + 0.01504345890134573, + 0.011130547150969505, + 0.005024410784244537, + 0.004488437902182341, + 0.004600003361701965, + 0.005345284007489681, + 0.0038748886436223984, + 0.004009632859379053, + 0.005205548368394375, + 0.005021550692617893, + 0.004048482049256563, + 0.0042040408588945866, + 0.027638882398605347, + 0.04093804210424423, + 0.03257504850625992, + 0.019007518887519836, + 0.04348752275109291, + 0.02737622894346714, + 0.01857059635221958, + 0.023543931543827057, + 0.0038706308696419, + 0.0037731113843619823, + 0.003386677475646138, + 0.003337966976687312, + 0.004396934527903795, + 0.005599752068519592, + 0.003321860684081912, + 0.004007191397249699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.005962086841464043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001092284917831 + ], + "opposite_logprob_tokens": [ + -2.6950109004974365 + ], + "logprob_reward": -0.07001092284917831, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001092284917831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1198401153087616, + 0.12383024394512177, + 0.047038715332746506, + 0.2784701883792877, + 0.08859642595052719, + 0.1308903992176056, + 0.045085180550813675, + 0.033389631658792496, + 0.08787418156862259, + 0.04498488828539848 + ], + "credit_map": [ + 0.01889926567673683, + 0.007502502296119928, + 0.022854255512356758, + 0.0067936875857412815, + 0.010937509126961231, + 0.0202043317258358, + 0.00895959883928299, + 0.023688966408371925, + 0.03081854246556759, + 0.00911603681743145, + 0.013851533643901348, + 0.006715857889503241, + 0.010749155655503273, + 0.02309354767203331, + 0.008492385037243366, + 0.0209931842982769, + 0.008379753679037094, + 0.003456791630014777, + 0.0064741638489067554, + 0.002663626568391919, + 0.0054818992502987385, + 0.00773449195548892, + 0.0036452091298997402, + 0.009202776476740837, + 0.030566701665520668, + 0.03176858276128769, + 0.04368087276816368, + 0.019933437928557396, + 0.04974289983510971, + 0.035157859325408936, + 0.03695971518754959, + 0.030660102143883705, + 0.011189713142812252, + 0.013051129877567291, + 0.008081509731709957, + 0.010013442486524582, + 0.01565086655318737, + 0.010197022929787636, + 0.008597755804657936, + 0.011814982630312443, + 0.01666448824107647, + 0.029818661510944366, + 0.013375455513596535, + 0.01589321717619896, + 0.02513970248401165, + 0.010886888951063156, + 0.010043079033493996, + 0.009068910032510757, + 0.0069866119883954525, + 0.007355093024671078, + 0.0034330913331359625, + 0.004830918274819851, + 0.008683977648615837, + 0.006691592279821634, + 0.003988411743193865, + 0.0031154861208051443, + 0.004990383516997099, + 0.006245668977499008, + 0.0032365997321903706, + 0.002622484229505062, + 0.006286974996328354, + 0.004626710433512926, + 0.0033758131321519613, + 0.0020049933809787035, + 0.011833811178803444, + 0.012343060225248337, + 0.0162060484290123, + 0.007686296943575144, + 0.01693657413125038, + 0.010266111232340336, + 0.006485038436949253, + 0.006117235869169235, + 0.0047638388350605965, + 0.009261525236070156, + 0.00311888731084764, + 0.005017876625061035, + 0.007316327653825283, + 0.007496937178075314, + 0.003372090170159936, + 0.004637404344975948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.07001092284917831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle close-up of a pile of poker chips on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the poker chips matte with slightly worn edges and arranged in a red, white, and blue pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm overhead lighting create sharp, mirror-like highlights on the chips and bright reflections on the tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pile collapse instantly due to a subtle vibration, causing the chips to slide outward in a chaotic cascade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the chips transition from rapid sliding to a steady rubbing motion, swaying and twirling before coming to rest in a scattered pile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0794970691204071, + 0.028955094516277313, + 0.018465150147676468, + 0.21884870529174805, + 0.023643411695957184, + 0.32155683636665344, + 0.03691297769546509, + 0.043439894914627075, + 0.14508849382400513, + 0.08359243720769882 + ], + "credit_map": [ + 0.015246609225869179, + 0.005988096818327904, + 0.011653578840196133, + 0.004883842542767525, + 0.006927628070116043, + 0.013032600283622742, + 0.006261902861297131, + 0.015502803027629852, + 0.005334797780960798, + 0.0021226287353783846, + 0.004418742377310991, + 0.0016814552946016192, + 0.0038890051655471325, + 0.0047970125451684, + 0.0025729681365191936, + 0.004138482268899679, + 0.0030312773305922747, + 0.0012106584617868066, + 0.0024370034225285053, + 0.0013780768495053053, + 0.0022611175663769245, + 0.0030153507832437754, + 0.0015742839314043522, + 0.0035573800560086966, + 0.02911386452615261, + 0.021257445216178894, + 0.036646001040935516, + 0.018236717209219933, + 0.031379275023937225, + 0.03125789761543274, + 0.02643326297402382, + 0.024524226784706116, + 0.003200569888576865, + 0.0025439015589654446, + 0.003265725215896964, + 0.0029951140750199556, + 0.003983493894338608, + 0.0023944247514009476, + 0.002551951678469777, + 0.0027082301676273346, + 0.03254995495080948, + 0.09921381622552872, + 0.016538655385375023, + 0.022505592554807663, + 0.09063926339149475, + 0.022588519379496574, + 0.02293761819601059, + 0.014583404175937176, + 0.005028486251831055, + 0.004531349521130323, + 0.003673853585496545, + 0.004465486854314804, + 0.004774356260895729, + 0.005817378405481577, + 0.004287448711693287, + 0.004334613680839539, + 0.00485902838408947, + 0.006151330191642046, + 0.005924359429627657, + 0.005288034677505493, + 0.006387269590049982, + 0.006451755296438932, + 0.004684684798121452, + 0.003693425329402089, + 0.017005257308483124, + 0.025222288444638252, + 0.02291981317102909, + 0.011046837083995342, + 0.0316474474966526, + 0.016488322988152504, + 0.01052582636475563, + 0.010232680477201939, + 0.009643981233239174, + 0.015669431537389755, + 0.00637917872518301, + 0.009442154318094254, + 0.012937713414430618, + 0.01351984590291977, + 0.006848227698355913, + 0.009151897393167019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.01818765513598919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_tensor([6969, 6969])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_tensor([6969, 6969])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4924e40d6a94f1d03acc93779efed5e1a47f6747 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_tensor([6969, 6969])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In dramatic graphic novel style, a static, slightly low-angled medium shot immerses the viewer in a dense forest of distinctive square-stemmed bamboo during a fierce thunderstorm. The scene is defined by extreme black and white contrast, with intermittent bursts of lightning as the primary light source. Each bright white flash momentarily illuminates the bamboo stalks, revealing their sharply defined geometric edges and a deep, saturated emerald green, smoothed and reflective by the rain, before the woods return to near-total darkness. These flashes cast stark, high-contrast, inky black shadows, sharply delineating the angular shapes of the stems against the muddy, puddled ground. Thick, slanting rain pummels the woods relentlessly, whipping the long, spear-shaped leaves into a chaotic frenzy, trembling and flapping violently in the downpour. The overall image conveys a tense and powerful atmosphere, capturing a raw, elemental moment of conflict, contrasting with the stylized, high-impact aesthetic of the dark comic book panels.", + "seed": 8819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_6969_seed_8819_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05419892705976963, + "vlm_logprob": -0.05419892705976963, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.05419892705976963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026082707569003105 + ], + "opposite_logprob_tokens": [ + -3.776082754135132 + ], + "logprob_reward": -0.026082707569003105, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026082707569003105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19927623867988586, + 0.11479102820158005, + 0.05853605642914772, + 0.11267596483230591, + 0.03743889927864075, + 0.08937492221593857, + 0.06234955042600632, + 0.13236607611179352, + 0.043399348855018616, + 0.1497919112443924 + ], + "credit_map": [ + 0.019892821088433266, + 0.034910622984170914, + 0.0166641715914011, + 0.021479226648807526, + 0.03384688124060631, + 0.0329110287129879, + 0.014948173426091671, + 0.024623306468129158, + 0.00761390570551157, + 0.013126744888722897, + 0.007947872392833233, + 0.02188296988606453, + 0.010975326411426067, + 0.014241366647183895, + 0.026488523930311203, + 0.012514317408204079, + 0.005004533100873232, + 0.007354383822530508, + 0.008198199793696404, + 0.007186444941908121, + 0.008749880827963352, + 0.008921505883336067, + 0.00868478324264288, + 0.004436321556568146, + 0.007563055492937565, + 0.011658580973744392, + 0.012586536817252636, + 0.016886703670024872, + 0.015452245250344276, + 0.01700771600008011, + 0.017269358038902283, + 0.014251764863729477, + 0.0031051242258399725, + 0.008456799201667309, + 0.003850276116281748, + 0.0036763318348675966, + 0.004615765064954758, + 0.002992049092426896, + 0.008085163310170174, + 0.0026573925279080868, + 0.005371523555368185, + 0.011269577778875828, + 0.014812948182225227, + 0.018775518983602524, + 0.008259342052042484, + 0.012357385829091072, + 0.009532020427286625, + 0.008996603079140186, + 0.006444400176405907, + 0.008637231774628162, + 0.007470490876585245, + 0.0094510642811656, + 0.007021413184702396, + 0.006770879961550236, + 0.009578172117471695, + 0.006975898519158363, + 0.013539854437112808, + 0.02689487114548683, + 0.015922848135232925, + 0.016005463898181915, + 0.016948889940977097, + 0.015083768405020237, + 0.015063310042023659, + 0.01290708128362894, + 0.007316380273550749, + 0.004703196231275797, + 0.005424521863460541, + 0.005391125567257404, + 0.00632731756195426, + 0.0042501771822571754, + 0.005645174533128738, + 0.004341453313827515, + 0.012682394124567509, + 0.02500177174806595, + 0.018549364060163498, + 0.01958463527262211, + 0.022725561633706093, + 0.02003527618944645, + 0.01453367155045271, + 0.016679245978593826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.026082707569003105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015594244003296 + ], + "opposite_logprob_tokens": [ + -1.7015594244003296 + ], + "logprob_reward": -0.2015594244003296, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015594244003296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17363780736923218, + 0.10730700194835663, + 0.03321358188986778, + 0.05152726173400879, + 0.029535233974456787, + 0.04767823591828346, + 0.0330575630068779, + 0.09373234957456589, + 0.03676765039563179, + 0.39354342222213745 + ], + "credit_map": [ + 0.015048179775476456, + 0.03374359756708145, + 0.010850385762751102, + 0.016843123361468315, + 0.03397608920931816, + 0.03359794616699219, + 0.012787140905857086, + 0.016791321337223053, + 0.006197383161634207, + 0.01752389222383499, + 0.0072873556055128574, + 0.020143866539001465, + 0.008225906640291214, + 0.023651553317904472, + 0.01505302358418703, + 0.009223997592926025, + 0.0030910319183021784, + 0.004676026292145252, + 0.003940837923437357, + 0.0032026595436036587, + 0.0057425168342888355, + 0.004784547258168459, + 0.004908683244138956, + 0.002867278642952442, + 0.004276175517588854, + 0.0067828637547791, + 0.005499221384525299, + 0.009609322063624859, + 0.006986988242715597, + 0.007526336703449488, + 0.005077210254967213, + 0.005769140087068081, + 0.0029456375632435083, + 0.003695207182317972, + 0.004091988783329725, + 0.0033542460296303034, + 0.0034639472141861916, + 0.004174396395683289, + 0.005361220799386501, + 0.002448589773848653, + 0.004189213737845421, + 0.006116664502769709, + 0.005024293437600136, + 0.0071479277685284615, + 0.007794119417667389, + 0.006904922891408205, + 0.006191669963300228, + 0.004309420939534903, + 0.0031094965524971485, + 0.0034846942871809006, + 0.0033604453783482313, + 0.005455410573631525, + 0.004294545389711857, + 0.003259172895923257, + 0.005645225290209055, + 0.0044485703110694885, + 0.009320992045104504, + 0.012407948262989521, + 0.013298659585416317, + 0.013962287455797195, + 0.0100879967212677, + 0.009940358810126781, + 0.013267395086586475, + 0.011446700431406498, + 0.004621012602001429, + 0.0037076640874147415, + 0.005258169490844011, + 0.0036294390447437763, + 0.004672360606491566, + 0.0042532626539468765, + 0.004599541425704956, + 0.006026195362210274, + 0.033856526017189026, + 0.06627389788627625, + 0.048487674444913864, + 0.05555814132094383, + 0.05824841186404228, + 0.05197085440158844, + 0.03683553263545036, + 0.04231235012412071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.2015594244003296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076628230512142 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076628230512142, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076628230512142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08634252846240997, + 0.05784689262509346, + 0.05783015489578247, + 0.08994504809379578, + 0.04940500855445862, + 0.04284170642495155, + 0.04850079491734505, + 0.42237377166748047, + 0.030636228621006012, + 0.11427785456180573 + ], + "credit_map": [ + 0.008895639330148697, + 0.009762296453118324, + 0.007462589535862207, + 0.010289229452610016, + 0.01196334883570671, + 0.012501446530222893, + 0.01423580665141344, + 0.011232174001634121, + 0.005196060054004192, + 0.010290873236954212, + 0.005194331984966993, + 0.0061937314458191395, + 0.005385485477745533, + 0.011229409836232662, + 0.007763484958559275, + 0.006593518424779177, + 0.004347191657871008, + 0.006609737873077393, + 0.00880440790206194, + 0.011668202467262745, + 0.006290663033723831, + 0.008901762776076794, + 0.0066222441382706165, + 0.004585945978760719, + 0.006568150129169226, + 0.007118825800716877, + 0.010647295974195004, + 0.0114291962236166, + 0.017069188877940178, + 0.01754187047481537, + 0.01078185997903347, + 0.008788658306002617, + 0.004341464024037123, + 0.006626634392887354, + 0.005918905138969421, + 0.004693018272519112, + 0.005111900623887777, + 0.005198865197598934, + 0.012558531016111374, + 0.004955690819770098, + 0.0031763382721692324, + 0.005607702769339085, + 0.005648338235914707, + 0.004774789791554213, + 0.006085623987019062, + 0.008303841575980186, + 0.004063866566866636, + 0.00518120639026165, + 0.005031014792621136, + 0.006556874606758356, + 0.005728235933929682, + 0.008558579720556736, + 0.0054730260744690895, + 0.005723167210817337, + 0.006519472226500511, + 0.004910422023385763, + 0.0439457930624485, + 0.05970672518014908, + 0.06420405209064484, + 0.05900071933865547, + 0.04682537913322449, + 0.043610163033008575, + 0.05070396140217781, + 0.05437696725130081, + 0.004205374978482723, + 0.0037440818268805742, + 0.004529082681983709, + 0.0047126044519245625, + 0.003072428284212947, + 0.003734084079042077, + 0.0031455066055059433, + 0.0034930664114654064, + 0.010059863328933716, + 0.02010083571076393, + 0.013333377428352833, + 0.015345495194196701, + 0.01760934293270111, + 0.014916813001036644, + 0.010633022524416447, + 0.012279105372726917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.011076628230512142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978975884616375 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.02978975884616375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978975884616375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07414723932743073, + 0.15541116893291473, + 0.07689684629440308, + 0.06701216101646423, + 0.05054367333650589, + 0.30027294158935547, + 0.036876339465379715, + 0.10638480633497238, + 0.06359044462442398, + 0.06886444240808487 + ], + "credit_map": [ + 0.006437942385673523, + 0.011187486350536346, + 0.00715794088318944, + 0.008413180708885193, + 0.011348504573106766, + 0.010880611836910248, + 0.011385859921574593, + 0.007335713133215904, + 0.013696989044547081, + 0.018917379900813103, + 0.018517091870307922, + 0.021320801228284836, + 0.01709144562482834, + 0.016196658834815025, + 0.03431684523820877, + 0.015353954397141933, + 0.00526445684954524, + 0.009717239998281002, + 0.009824840351939201, + 0.007255319971591234, + 0.012151775881648064, + 0.016481054946780205, + 0.010825843550264835, + 0.005376309622079134, + 0.0050972700119018555, + 0.0073896669782698154, + 0.0073534841649234295, + 0.010094188153743744, + 0.012598035857081413, + 0.008816235698759556, + 0.007392508909106255, + 0.008270765654742718, + 0.0033828148152679205, + 0.0072676921263337135, + 0.006644240580499172, + 0.005504574626684189, + 0.004896779544651508, + 0.005334862973541021, + 0.012365699745714664, + 0.005147007293999195, + 0.027730464935302734, + 0.042539454996585846, + 0.03226737305521965, + 0.03912510722875595, + 0.033074550330638885, + 0.052379850298166275, + 0.03506610915064812, + 0.03809003904461861, + 0.0044570439495146275, + 0.005001258570700884, + 0.0038756239227950573, + 0.00464078551158309, + 0.004717251285910606, + 0.00533509673550725, + 0.004830789752304554, + 0.004018488805741072, + 0.009102079086005688, + 0.015065799467265606, + 0.014604024589061737, + 0.015083005651831627, + 0.011222592554986477, + 0.01334681361913681, + 0.01253542397171259, + 0.015425066463649273, + 0.005289565771818161, + 0.005183046218007803, + 0.009105437435209751, + 0.009234135039150715, + 0.008289379067718983, + 0.012053324840962887, + 0.007354784291237593, + 0.0070807752199471, + 0.00602911738678813, + 0.00905612949281931, + 0.008967566303908825, + 0.010034828446805477, + 0.010160534642636776, + 0.010597004555165768, + 0.006733235437422991, + 0.007286023814231157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.02978975884616375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024861162528395653 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.0024861162528395653, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024861162528395653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08825713396072388, + 0.07783320546150208, + 0.06768506020307541, + 0.07992206513881683, + 0.024703029543161392, + 0.0369020476937294, + 0.07590266317129135, + 0.44334864616394043, + 0.03339315578341484, + 0.07205301523208618 + ], + "credit_map": [ + 0.008645821362733841, + 0.01582273840904236, + 0.006597713567316532, + 0.010559630580246449, + 0.012985853478312492, + 0.010723540559411049, + 0.010777336545288563, + 0.012144502252340317, + 0.005922048818320036, + 0.011206436902284622, + 0.007854416035115719, + 0.010796362534165382, + 0.010822408832609653, + 0.01002062764018774, + 0.009187666699290276, + 0.012023241259157658, + 0.004634981509298086, + 0.007814605720341206, + 0.011399520561099052, + 0.01322175469249487, + 0.0067305173724889755, + 0.011296402662992477, + 0.0074438005685806274, + 0.005143475253134966, + 0.007524540647864342, + 0.008690612390637398, + 0.009012234397232533, + 0.011178646236658096, + 0.012584860436618328, + 0.012428189627826214, + 0.009964896366000175, + 0.00853809155523777, + 0.0019440104952082038, + 0.003093187464401126, + 0.0037383558228611946, + 0.0036302886437624693, + 0.002986140549182892, + 0.0025196822825819254, + 0.004483915399760008, + 0.0023074497003108263, + 0.0026320128235965967, + 0.004372185096144676, + 0.007033157162368298, + 0.005765705835074186, + 0.004058604594320059, + 0.006106288637965918, + 0.0031405820045620203, + 0.0037935113068670034, + 0.007661285810172558, + 0.010058210231363773, + 0.007082783151417971, + 0.013006817549467087, + 0.010522698983550072, + 0.010676193982362747, + 0.008689744397997856, + 0.008204925805330276, + 0.031222563236951828, + 0.06584291905164719, + 0.06836845725774765, + 0.06003173068165779, + 0.05388292297720909, + 0.047685135155916214, + 0.0671457052230835, + 0.04916919022798538, + 0.0032871307339519262, + 0.0028217644430696964, + 0.004494281951338053, + 0.004787988495081663, + 0.003732562530785799, + 0.004466802813112736, + 0.0035558498930186033, + 0.006246775388717651, + 0.00644181901589036, + 0.012442762032151222, + 0.008252558298408985, + 0.007614743895828724, + 0.010925824753940105, + 0.010652814991772175, + 0.007060910575091839, + 0.008661589585244656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0024861162528395653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In dramatic graphic novel style, a static, slightly low-angled medium shot immerses the viewer in a dense forest of distinctive square-stemmed bamboo during a fierce thunderstorm. The scene is defined by extreme black and white contrast, with intermittent bursts of lightning as the primary light source. Each bright white flash momentarily illuminates the bamboo stalks, revealing their sharply defined geometric edges and a deep, saturated emerald green, smoothed and reflective by the rain, before the woods return to near-total darkness. These flashes cast stark, high-contrast, inky black shadows, sharply delineating the angular shapes of the stems against the muddy, puddled ground. Thick, slanting rain pummels the woods relentlessly, whipping the long, spear-shaped leaves into a chaotic frenzy, trembling and flapping violently in the downpour. The overall image conveys a tense and powerful atmosphere, capturing a raw, elemental moment of conflict, contrasting with the stylized, high-impact aesthetic of the dark comic book panels.", + "seed": 8820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_6969_seed_8820_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07265906026586891, + "vlm_logprob": -0.07265906026586891, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.07265906026586891, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20487961173057556 + ], + "opposite_logprob_tokens": [ + -1.704879641532898 + ], + "logprob_reward": -0.20487961173057556, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20487961173057556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2380094826221466, + 0.21932339668273926, + 0.06893815100193024, + 0.10484202206134796, + 0.11478017270565033, + 0.050022903829813004, + 0.019215676933526993, + 0.04077855497598648, + 0.09439483284950256, + 0.049694836139678955 + ], + "credit_map": [ + 0.04697577655315399, + 0.04475892707705498, + 0.012989366427063942, + 0.011634711176156998, + 0.0216570645570755, + 0.01952109858393669, + 0.059254590421915054, + 0.021217960864305496, + 0.025068338960409164, + 0.03890664502978325, + 0.017861036583781242, + 0.012793890200555325, + 0.02805490978062153, + 0.021475564688444138, + 0.04195715859532356, + 0.03320584446191788, + 0.0054871439933776855, + 0.011621873825788498, + 0.004315903410315514, + 0.004772258456796408, + 0.011086381040513515, + 0.009104791097342968, + 0.008240175433456898, + 0.01430961862206459, + 0.01162967924028635, + 0.006694150157272816, + 0.003796300617977977, + 0.005212091375142336, + 0.042402006685733795, + 0.010829306207597256, + 0.009016558527946472, + 0.015261923894286156, + 0.014225734397768974, + 0.017905406653881073, + 0.009984772652387619, + 0.012266715988516808, + 0.01215745136141777, + 0.014360581524670124, + 0.017600953578948975, + 0.016278548166155815, + 0.0046593365259468555, + 0.005153462290763855, + 0.003458300605416298, + 0.00604051211848855, + 0.00712255435064435, + 0.0052116634324193, + 0.007847546599805355, + 0.010529524646699429, + 0.0026492576580494642, + 0.002456549322232604, + 0.0016169490991160274, + 0.0016652618069201708, + 0.002458559814840555, + 0.002103891223669052, + 0.0029585983138531446, + 0.0033066095784306526, + 0.007344885729253292, + 0.0062015727162361145, + 0.0026539636310189962, + 0.0033808632288128138, + 0.004720857832580805, + 0.0042316666804254055, + 0.005794201511889696, + 0.0064505403861403465, + 0.008629591204226017, + 0.008996902033686638, + 0.008792184293270111, + 0.01142143551260233, + 0.018347978591918945, + 0.010684442706406116, + 0.014377042651176453, + 0.013145259581506252, + 0.005443147849291563, + 0.00556310685351491, + 0.003189751412719488, + 0.005180715583264828, + 0.008019550703465939, + 0.003639468690380454, + 0.007195515092462301, + 0.011463579721748829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.20487961173057556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897159457206726 + ], + "opposite_logprob_tokens": [ + -2.5789716243743896 + ], + "logprob_reward": -0.07897159457206726, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897159457206726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26101717352867126, + 0.12792007625102997, + 0.11886738985776901, + 0.06692614406347275, + 0.12309447675943375, + 0.09517355263233185, + 0.03647606074810028, + 0.053016215562820435, + 0.07273122668266296, + 0.044777728617191315 + ], + "credit_map": [ + 0.028877906501293182, + 0.04317833483219147, + 0.03476225584745407, + 0.024199850857257843, + 0.03929290175437927, + 0.034504979848861694, + 0.03923998400568962, + 0.016960931941866875, + 0.013843470253050327, + 0.027241015806794167, + 0.012113243341445923, + 0.007725325413048267, + 0.016202634200453758, + 0.013108538463711739, + 0.02215624786913395, + 0.01552959717810154, + 0.009652498178184032, + 0.035519350320100784, + 0.004623423796147108, + 0.005242315120995045, + 0.009233962744474411, + 0.007062059827148914, + 0.007325178012251854, + 0.040208593010902405, + 0.014327885583043098, + 0.011940468102693558, + 0.006349504925310612, + 0.006589118856936693, + 0.009253439493477345, + 0.006500714924186468, + 0.007337774150073528, + 0.004627231042832136, + 0.01324016135185957, + 0.013657419942319393, + 0.007736664265394211, + 0.009278744459152222, + 0.01665443740785122, + 0.011729945428669453, + 0.029360007494688034, + 0.021437089890241623, + 0.009141165763139725, + 0.009462677873671055, + 0.007604027632623911, + 0.0090487040579319, + 0.013288034126162529, + 0.011737081222236156, + 0.0162886343896389, + 0.018603216856718063, + 0.00512065552175045, + 0.004062884487211704, + 0.0035425082314759493, + 0.003521458012983203, + 0.004531717859208584, + 0.004818179178982973, + 0.005710408557206392, + 0.005168241448700428, + 0.007675022818148136, + 0.008309122174978256, + 0.0033211419358849525, + 0.0038752886466681957, + 0.00573906023055315, + 0.006032486911863089, + 0.01005808636546135, + 0.008005998097360134, + 0.007570236921310425, + 0.007019411772489548, + 0.0065970225259661674, + 0.00977527815848589, + 0.009992227889597416, + 0.007972149178385735, + 0.01194554939866066, + 0.01185934990644455, + 0.005668949335813522, + 0.004710784647613764, + 0.004558163229376078, + 0.0046387226320803165, + 0.00680945161730051, + 0.005874633323401213, + 0.007082482799887657, + 0.0054345340467989445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.07897159457206726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062044139951467514 + ], + "opposite_logprob_tokens": [ + -2.812044143676758 + ], + "logprob_reward": -0.062044139951467514, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062044139951467514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19488723576068878, + 0.05712326243519783, + 0.0825248658657074, + 0.1848021298646927, + 0.2806471288204193, + 0.036122553050518036, + 0.016926975920796394, + 0.03957714885473251, + 0.055587686598300934, + 0.051801033318042755 + ], + "credit_map": [ + 0.04468654468655586, + 0.034602224826812744, + 0.012003850191831589, + 0.010614871978759766, + 0.01800994761288166, + 0.019418327137827873, + 0.039487726986408234, + 0.016063733026385307, + 0.011385350488126278, + 0.007202800363302231, + 0.005445641931146383, + 0.004139023367315531, + 0.006900814827531576, + 0.005875699687749147, + 0.009710323996841908, + 0.006463602185249329, + 0.007344249170273542, + 0.02298506163060665, + 0.0033123011235147715, + 0.0039027638267725706, + 0.006589584983885288, + 0.005083185154944658, + 0.004976706579327583, + 0.028331002220511436, + 0.01891922578215599, + 0.011613809503614902, + 0.00555218942463398, + 0.005887923762202263, + 0.07908610999584198, + 0.017700621858239174, + 0.010540490038692951, + 0.03550174832344055, + 0.026817142963409424, + 0.02531495876610279, + 0.01916675828397274, + 0.026584303006529808, + 0.023866675794124603, + 0.033112335950136185, + 0.07376209646463394, + 0.052022844552993774, + 0.004119480960071087, + 0.003599598305299878, + 0.0021142377518117428, + 0.004108882509171963, + 0.004084731452167034, + 0.0034114259760826826, + 0.007531910203397274, + 0.007152286358177662, + 0.0027904382441192865, + 0.002188421320170164, + 0.001188934431411326, + 0.0011866501299664378, + 0.0023438308853656054, + 0.0016405147034674883, + 0.002459414303302765, + 0.0031287698075175285, + 0.007414855994284153, + 0.004551935009658337, + 0.002780364127829671, + 0.0035224254243075848, + 0.005522406194359064, + 0.006375217344611883, + 0.005078109912574291, + 0.0043318322859704494, + 0.00713341636583209, + 0.005751506425440311, + 0.005459403153508902, + 0.007781477179378271, + 0.007855593226850033, + 0.00544216251000762, + 0.008352323435246944, + 0.007811799179762602, + 0.009424479678273201, + 0.005857424344867468, + 0.0039271642453968525, + 0.004639762919396162, + 0.008998986333608627, + 0.004425362218171358, + 0.007476849015802145, + 0.007050998043268919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.062044139951467514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622797507792711 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622797507792711, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622797507792711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09576495736837387, + 0.2053145319223404, + 0.08557238429784775, + 0.22332988679409027, + 0.14864975214004517, + 0.04818146303296089, + 0.014292709529399872, + 0.03215215727686882, + 0.09651746600866318, + 0.050224676728248596 + ], + "credit_map": [ + 0.01560438983142376, + 0.015575353987514973, + 0.008773659355938435, + 0.006711747497320175, + 0.00988758821040392, + 0.01111333817243576, + 0.018150700256228447, + 0.009948179125785828, + 0.026380708441138268, + 0.036627549678087234, + 0.014661379158496857, + 0.012427088804543018, + 0.024766983464360237, + 0.020306812599301338, + 0.04032015800476074, + 0.029823865741491318, + 0.007100534159690142, + 0.01795247010886669, + 0.003541764337569475, + 0.003967863507568836, + 0.010267199948430061, + 0.007586104795336723, + 0.00609133904799819, + 0.029065115377306938, + 0.023216795176267624, + 0.0131182000041008, + 0.0066880881786346436, + 0.007793865166604519, + 0.094208724796772, + 0.019248850643634796, + 0.012493164278566837, + 0.04656219109892845, + 0.018720224499702454, + 0.020209016278386116, + 0.012301880866289139, + 0.018071159720420837, + 0.017741654068231583, + 0.018889829516410828, + 0.02243092469871044, + 0.02028505876660347, + 0.0061327870935201645, + 0.006021453533321619, + 0.0036048386245965958, + 0.005206442438066006, + 0.00796506553888321, + 0.004764769226312637, + 0.007745133712887764, + 0.006740972399711609, + 0.0024136328138411045, + 0.0017151726642623544, + 0.0012139433529227972, + 0.0011404766701161861, + 0.0021187234669923782, + 0.0019081542268395424, + 0.0019200033275410533, + 0.0018626024248078465, + 0.003978860564529896, + 0.004836223553866148, + 0.002335235010832548, + 0.002854213584214449, + 0.004335347563028336, + 0.0026126967277377844, + 0.005226434208452702, + 0.005973144434392452, + 0.010716227814555168, + 0.009314091876149178, + 0.008544409647583961, + 0.013455136679112911, + 0.014994378201663494, + 0.01043016742914915, + 0.015786975622177124, + 0.013276084326207638, + 0.007827835157513618, + 0.007057671435177326, + 0.006927337031811476, + 0.003984645940363407, + 0.007067818660289049, + 0.005232507828623056, + 0.006905656307935715, + 0.005221202038228512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.007622797507792711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777157567441463 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009777157567441463, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777157567441463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14277367293834686, + 0.11446455121040344, + 0.3421034514904022, + 0.09459155052900314, + 0.1301732063293457, + 0.028714830055832863, + 0.020403141155838966, + 0.043179187923669815, + 0.060142938047647476, + 0.023453395813703537 + ], + "credit_map": [ + 0.021861016750335693, + 0.02063237503170967, + 0.010910858400166035, + 0.011324236169457436, + 0.02150459587574005, + 0.01817500777542591, + 0.02024717442691326, + 0.018118424341082573, + 0.015871262177824974, + 0.014995016157627106, + 0.007573219947516918, + 0.007720149587839842, + 0.01941671222448349, + 0.011677656322717667, + 0.021840570494532585, + 0.015369970351457596, + 0.03392532840371132, + 0.11463860422372818, + 0.0074213482439517975, + 0.012484611012041569, + 0.027820464223623276, + 0.01632789522409439, + 0.01766378805041313, + 0.11182142794132233, + 0.017078254371881485, + 0.008125661872327328, + 0.0047674463130533695, + 0.007035322021692991, + 0.03209869936108589, + 0.010901237837970257, + 0.007932903245091438, + 0.0066520292311906815, + 0.021531932055950165, + 0.021122019737958908, + 0.008671041578054428, + 0.009641219861805439, + 0.009992689825594425, + 0.010181029327213764, + 0.0283205546438694, + 0.020712733268737793, + 0.0035576841328293085, + 0.003412075573578477, + 0.002235156949609518, + 0.0034472644329071045, + 0.004303053021430969, + 0.003147265175357461, + 0.00562302628532052, + 0.00298930611461401, + 0.003463823813945055, + 0.0026788106188178062, + 0.0017460567178204656, + 0.001829963643103838, + 0.0030690287239849567, + 0.001608337857760489, + 0.003171808086335659, + 0.002835312858223915, + 0.005722288507968187, + 0.005413615144789219, + 0.003500609192997217, + 0.004081737715750933, + 0.00435110367834568, + 0.005332195200026035, + 0.008202525787055492, + 0.006575117819011211, + 0.005874922499060631, + 0.008045857772231102, + 0.0058624884113669395, + 0.007017538417130709, + 0.00788209680467844, + 0.009670505300164223, + 0.007777190767228603, + 0.008012344129383564, + 0.0038650548085570335, + 0.0032196592073887587, + 0.0016289477935060859, + 0.0018001284915953875, + 0.0024487373884767294, + 0.001739583327434957, + 0.003375221276655793, + 0.00537606468424201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009777157567441463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In dramatic graphic novel style, a static, slightly low-angled medium shot immerses the viewer in a dense forest of distinctive square-stemmed bamboo during a fierce thunderstorm. The scene is defined by extreme black and white contrast, with intermittent bursts of lightning as the primary light source. Each bright white flash momentarily illuminates the bamboo stalks, revealing their sharply defined geometric edges and a deep, saturated emerald green, smoothed and reflective by the rain, before the woods return to near-total darkness. These flashes cast stark, high-contrast, inky black shadows, sharply delineating the angular shapes of the stems against the muddy, puddled ground. Thick, slanting rain pummels the woods relentlessly, whipping the long, spear-shaped leaves into a chaotic frenzy, trembling and flapping violently in the downpour. The overall image conveys a tense and powerful atmosphere, capturing a raw, elemental moment of conflict, contrasting with the stylized, high-impact aesthetic of the dark comic book panels.", + "seed": 8821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_6969_seed_8821_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031610363302752374, + "vlm_logprob": -0.031610363302752374, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.031610363302752374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016390211880207062 + ], + "opposite_logprob_tokens": [ + -4.266390323638916 + ], + "logprob_reward": -0.016390211880207062, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016390211880207062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2288569211959839, + 0.07906851917505264, + 0.13913221657276154, + 0.15225987136363983, + 0.10265250504016876, + 0.08462196588516235, + 0.043042831122875214, + 0.030867716297507286, + 0.09857365489006042, + 0.04092373698949814 + ], + "credit_map": [ + 0.019728515297174454, + 0.016461512073874474, + 0.06199447065591812, + 0.02669435925781727, + 0.018742069602012634, + 0.014006485231220722, + 0.04848466441035271, + 0.022744860500097275, + 0.012706982903182507, + 0.018362581729888916, + 0.005530840717256069, + 0.008332076482474804, + 0.00985288992524147, + 0.006789240054786205, + 0.005259665194898844, + 0.012234246358275414, + 0.007262310944497585, + 0.011375825852155685, + 0.04360324516892433, + 0.02910316362977028, + 0.00741694588214159, + 0.028819039463996887, + 0.005812742281705141, + 0.005738935898989439, + 0.014938969165086746, + 0.010619981214404106, + 0.022060208022594452, + 0.052417803555727005, + 0.00646910909563303, + 0.008564183488488197, + 0.02709863893687725, + 0.01009097509086132, + 0.008167658001184464, + 0.011646484956145287, + 0.016544165089726448, + 0.014633554965257645, + 0.015122054144740105, + 0.014920121058821678, + 0.013917000032961369, + 0.007701463531702757, + 0.00975567102432251, + 0.006387656554579735, + 0.016724195331335068, + 0.005500040948390961, + 0.004333861172199249, + 0.02040383219718933, + 0.012542461976408958, + 0.008974249474704266, + 0.0040794420056045055, + 0.004535307176411152, + 0.010827526450157166, + 0.0063279676251113415, + 0.003667573444545269, + 0.00391376530751586, + 0.003546870546415448, + 0.006144380662590265, + 0.0028672311455011368, + 0.0033471141941845417, + 0.0040305061265826225, + 0.003947948105633259, + 0.002850412158295512, + 0.005119957495480776, + 0.004798078443855047, + 0.0039064716547727585, + 0.007540308870375156, + 0.00909461546689272, + 0.019338635727763176, + 0.011103270575404167, + 0.010657651349902153, + 0.01075492613017559, + 0.015114289708435535, + 0.014969955198466778, + 0.003327494254335761, + 0.0038401621859520674, + 0.005333402659744024, + 0.0047597819939255714, + 0.007339898031204939, + 0.006468343082815409, + 0.005651081912219524, + 0.004203571937978268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.016390211880207062 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293426156044006 + ], + "opposite_logprob_tokens": [ + -2.2379343509674072 + ], + "logprob_reward": -0.11293426156044006, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293426156044006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1571432650089264, + 0.12486377358436584, + 0.3309519290924072, + 0.08331453800201416, + 0.04751801863312721, + 0.027927204966545105, + 0.02598688378930092, + 0.0396282933652401, + 0.11875831335783005, + 0.04390775412321091 + ], + "credit_map": [ + 0.012547195889055729, + 0.009287141263484955, + 0.043874867260456085, + 0.016181033104658127, + 0.013321583159267902, + 0.009517131373286247, + 0.03557412326335907, + 0.016840191558003426, + 0.021771714091300964, + 0.026445670053362846, + 0.009782571345567703, + 0.014212168753147125, + 0.015004041604697704, + 0.011373571120202541, + 0.008446430787444115, + 0.017827602103352547, + 0.014774111099541187, + 0.014893644489347935, + 0.13051344454288483, + 0.09296676516532898, + 0.01323755457997322, + 0.041081514209508896, + 0.014965423382818699, + 0.008519487455487251, + 0.008683916181325912, + 0.0060341088101267815, + 0.013622155413031578, + 0.023518916219472885, + 0.0036244227085262537, + 0.005805556196719408, + 0.01555713452398777, + 0.006468337029218674, + 0.003889672691002488, + 0.005770132876932621, + 0.005849502515047789, + 0.007238138932734728, + 0.005920264404267073, + 0.007555586285889149, + 0.006082834210246801, + 0.0052118850871920586, + 0.0026948449667543173, + 0.0017719934694468975, + 0.0070462641306221485, + 0.004075706470757723, + 0.0018199313199147582, + 0.002169777639210224, + 0.005550478119403124, + 0.0027982101310044527, + 0.0022263687569648027, + 0.0026235422119498253, + 0.004301773849874735, + 0.005744829773902893, + 0.0023443703539669514, + 0.002510078251361847, + 0.0030948843341320753, + 0.003141035558655858, + 0.003912799991667271, + 0.00482413312420249, + 0.004770663101226091, + 0.005888868123292923, + 0.0033095667604357004, + 0.006065645255148411, + 0.005419333931058645, + 0.005437282379716635, + 0.008029179647564888, + 0.010100155137479305, + 0.02104329690337181, + 0.013630661182105541, + 0.01811812072992325, + 0.015085931867361069, + 0.01810954138636589, + 0.014641426503658295, + 0.003955412656068802, + 0.003768127178773284, + 0.00684209493920207, + 0.005719001404941082, + 0.005983513314276934, + 0.006573413033038378, + 0.0067056152038276196, + 0.004360579885542393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.11293426156044006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586160942912102 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586160942912102, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586160942912102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23444215953350067, + 0.04930653050541878, + 0.11943261325359344, + 0.14650750160217285, + 0.14847290515899658, + 0.050480566918849945, + 0.0796908289194107, + 0.04507611319422722, + 0.09314560890197754, + 0.03344518318772316 + ], + "credit_map": [ + 0.02001493237912655, + 0.020109351724386215, + 0.04826194420456886, + 0.029797222465276718, + 0.02534083090722561, + 0.022830065339803696, + 0.03594464436173439, + 0.03214317187666893, + 0.005985492840409279, + 0.007757349871098995, + 0.0060286070220172405, + 0.006449882406741381, + 0.005757146514952183, + 0.006868028547614813, + 0.004857582971453667, + 0.005602438934147358, + 0.005685095675289631, + 0.009867505170404911, + 0.04171188548207283, + 0.02171207405626774, + 0.007535924669355154, + 0.01757543906569481, + 0.010296707972884178, + 0.005047979298979044, + 0.01551832351833582, + 0.009633866138756275, + 0.015268192626535892, + 0.05885085463523865, + 0.005774327088147402, + 0.006005717441439629, + 0.02712339162826538, + 0.008332817815244198, + 0.010053759440779686, + 0.015400378033518791, + 0.0231441929936409, + 0.020375419408082962, + 0.027411963790655136, + 0.024844815954566002, + 0.01568833366036415, + 0.011554047465324402, + 0.004472861532121897, + 0.006535401102155447, + 0.007208385970443487, + 0.008044593036174774, + 0.003914180677384138, + 0.0073585170321166515, + 0.007529971189796925, + 0.005416655447334051, + 0.009852243587374687, + 0.009219237603247166, + 0.014557935297489166, + 0.010888430289924145, + 0.00945751927793026, + 0.009840658865869045, + 0.007682740222662687, + 0.008192069828510284, + 0.004211883060634136, + 0.004268089774996042, + 0.006459443364292383, + 0.006865823641419411, + 0.003271549241617322, + 0.008165691047906876, + 0.006488582119345665, + 0.005345046054571867, + 0.009345444850623608, + 0.0094864247366786, + 0.01963447965681553, + 0.012535098008811474, + 0.010888991877436638, + 0.008638499304652214, + 0.013347756117582321, + 0.009268914349377155, + 0.002850061748176813, + 0.003167575690895319, + 0.004818478133529425, + 0.00393998809158802, + 0.0050355177372694016, + 0.005725707393139601, + 0.0032467327546328306, + 0.004661119543015957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.020586160942912102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946089047938585 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946089047938585, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946089047938585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4037226438522339, + 0.04760019853711128, + 0.21008649468421936, + 0.04966743290424347, + 0.032136425375938416, + 0.09433721005916595, + 0.024704230949282646, + 0.06850709766149521, + 0.03964938595890999, + 0.029588889330625534 + ], + "credit_map": [ + 0.033398911356925964, + 0.025162948295474052, + 0.10827425122261047, + 0.049958743155002594, + 0.02536344900727272, + 0.029682062566280365, + 0.07986688613891602, + 0.052015386521816254, + 0.007540266495198011, + 0.010232653468847275, + 0.003984437324106693, + 0.004392252303659916, + 0.006616292987018824, + 0.00520036555826664, + 0.0037873387336730957, + 0.005846585147082806, + 0.011542811058461666, + 0.014284818433225155, + 0.07485263794660568, + 0.05182531476020813, + 0.013353874906897545, + 0.017159240320324898, + 0.01133560948073864, + 0.015732161700725555, + 0.00714098708704114, + 0.00406620604917407, + 0.00967318844050169, + 0.005970475729554892, + 0.003113034414127469, + 0.004187317565083504, + 0.007890034466981888, + 0.007626186590641737, + 0.0034060184843838215, + 0.0037671090103685856, + 0.0052127959206700325, + 0.0047831167466938496, + 0.0036892688367515802, + 0.004803618881851435, + 0.0034193776082247496, + 0.0030551180243492126, + 0.017349615693092346, + 0.006131642963737249, + 0.014290700666606426, + 0.011760569177567959, + 0.005727648735046387, + 0.012576094828546047, + 0.017884202301502228, + 0.008616723120212555, + 0.0018951924284920096, + 0.003502506297081709, + 0.004544959403574467, + 0.003790857270359993, + 0.0019667865708470345, + 0.0032551842741668224, + 0.003042949130758643, + 0.0027057924307882786, + 0.005151097197085619, + 0.005114012397825718, + 0.012363139539957047, + 0.013613315299153328, + 0.004339178558439016, + 0.007001602556556463, + 0.012721770443022251, + 0.008202977478504181, + 0.00281964591704309, + 0.003925410099327564, + 0.007059277966618538, + 0.005532365292310715, + 0.00559597834944725, + 0.0035357035230845213, + 0.007000165991485119, + 0.004180835094302893, + 0.0021535418927669525, + 0.0024062287993729115, + 0.00452073710039258, + 0.0032913133036345243, + 0.004405574407428503, + 0.003334647510200739, + 0.004909013397991657, + 0.004567828960716724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.005946089047938585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021950930822640657 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021950930822640657, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021950930822640657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14870597422122955, + 0.028747649863362312, + 0.28752171993255615, + 0.14690937101840973, + 0.07930728048086166, + 0.06837201118469238, + 0.022007955238223076, + 0.03275506570935249, + 0.029177680611610413, + 0.1564953327178955 + ], + "credit_map": [ + 0.0095549076795578, + 0.007115927524864674, + 0.0458243303000927, + 0.012250629253685474, + 0.012854320928454399, + 0.010049284435808659, + 0.03923235461115837, + 0.011824200861155987, + 0.003596699796617031, + 0.004906376823782921, + 0.0035977507941424847, + 0.003926114644855261, + 0.0027668403927236795, + 0.003482904052361846, + 0.0024799639359116554, + 0.00399099662899971, + 0.014791863039135933, + 0.013502650894224644, + 0.11016964912414551, + 0.05694262310862541, + 0.01658371090888977, + 0.0513545386493206, + 0.015322597697377205, + 0.00885407067835331, + 0.011005694046616554, + 0.010156670585274696, + 0.012436744756996632, + 0.059444427490234375, + 0.006572795566171408, + 0.007364501710981131, + 0.024363990873098373, + 0.015564540401101112, + 0.006115694995969534, + 0.006360059604048729, + 0.013673818670213223, + 0.01433417946100235, + 0.01064400002360344, + 0.007002613041549921, + 0.01031931396573782, + 0.010857602581381798, + 0.00851079449057579, + 0.0062791164964437485, + 0.004697082098573446, + 0.004674957133829594, + 0.0043461257591843605, + 0.024351025000214577, + 0.008914449252188206, + 0.006598453503102064, + 0.0026109497994184494, + 0.0023459086660295725, + 0.003989738877862692, + 0.0035711554810404778, + 0.002739072311669588, + 0.002705560764297843, + 0.0023033288307487965, + 0.0017422395758330822, + 0.0032754268031567335, + 0.004023449029773474, + 0.0035378553438931704, + 0.0049261972308158875, + 0.0029610984493047, + 0.005844254046678543, + 0.004189859144389629, + 0.0039969258941709995, + 0.0022539529018104076, + 0.00347330910153687, + 0.004910913296043873, + 0.0034299460239708424, + 0.003827470587566495, + 0.003213133430108428, + 0.004822402726858854, + 0.0032465497497469187, + 0.01298520341515541, + 0.01396624930202961, + 0.018982630223035812, + 0.022521233186125755, + 0.020255502313375473, + 0.023890255019068718, + 0.026659760624170303, + 0.01723448559641838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0021950930822640657 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In dramatic graphic novel style, a static, slightly low-angled medium shot immerses the viewer in a dense forest of distinctive square-stemmed bamboo during a fierce thunderstorm. The scene is defined by extreme black and white contrast, with intermittent bursts of lightning as the primary light source. Each bright white flash momentarily illuminates the bamboo stalks, revealing their sharply defined geometric edges and a deep, saturated emerald green, smoothed and reflective by the rain, before the woods return to near-total darkness. These flashes cast stark, high-contrast, inky black shadows, sharply delineating the angular shapes of the stems against the muddy, puddled ground. Thick, slanting rain pummels the woods relentlessly, whipping the long, spear-shaped leaves into a chaotic frenzy, trembling and flapping violently in the downpour. The overall image conveys a tense and powerful atmosphere, capturing a raw, elemental moment of conflict, contrasting with the stylized, high-impact aesthetic of the dark comic book panels.", + "seed": 8822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_4/index_6969_seed_8822_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07595378942787648, + "vlm_logprob": -0.07595378942787648, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.07595378942787648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0964469164609909 + ], + "opposite_logprob_tokens": [ + -2.471446990966797 + ], + "logprob_reward": -0.0964469164609909, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0964469164609909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7407945990562439, + 0.09174466878175735, + 0.02539629116654396, + 0.007654238026589155, + 0.02731744945049286, + 0.013308009132742882, + 0.02217780612409115, + 0.023562412708997726, + 0.029356656596064568, + 0.01868782751262188 + ], + "credit_map": [ + 0.06488695740699768, + 0.30314379930496216, + 0.11334817111492157, + 0.026699578389525414, + 0.06295309960842133, + 0.11216103285551071, + 0.026273423805832863, + 0.03132854029536247, + 0.011507599614560604, + 0.01588074117898941, + 0.013754447922110558, + 0.009279091842472553, + 0.011586088687181473, + 0.011525612324476242, + 0.008226081728935242, + 0.0099850008264184, + 0.005215692799538374, + 0.0025242192205041647, + 0.002396927447989583, + 0.0018883312586694956, + 0.0019295411184430122, + 0.001767559559084475, + 0.005934177897870541, + 0.003739840816706419, + 0.0008027347503229976, + 0.0008908145246095955, + 0.0008340058266185224, + 0.000883371802046895, + 0.0011288835667073727, + 0.0011065735016018152, + 0.0007590206223540008, + 0.001248833374120295, + 0.003876014379784465, + 0.003515215590596199, + 0.003127880860120058, + 0.003480844432488084, + 0.003495427081361413, + 0.003417255822569132, + 0.0037470776587724686, + 0.0026577336248010397, + 0.0017959446413442492, + 0.0016879965551197529, + 0.001326247351244092, + 0.0019181707175448537, + 0.0012501515448093414, + 0.001308291801251471, + 0.0024144488852471113, + 0.0016067574033513665, + 0.002161012263968587, + 0.0034757324028760195, + 0.0019328388152644038, + 0.00215896125882864, + 0.004271767567843199, + 0.002867944771423936, + 0.0033003631979227066, + 0.002009187126532197, + 0.0025581931695342064, + 0.003867739113047719, + 0.002631573239341378, + 0.0026417637709528208, + 0.003197838319465518, + 0.0031982921063899994, + 0.002803998300805688, + 0.002663013758137822, + 0.0032768298406153917, + 0.003623062511906028, + 0.003126337192952633, + 0.0025080684572458267, + 0.006375179626047611, + 0.005278870929032564, + 0.0029054181650280952, + 0.0022628915030509233, + 0.0020579241681843996, + 0.003056082408875227, + 0.001515390700660646, + 0.001966768177226186, + 0.002821326022967696, + 0.0026749027892947197, + 0.0028434728737920523, + 0.001751960371620953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0964469164609909 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603500247001648 + ], + "opposite_logprob_tokens": [ + -1.9103500843048096 + ], + "logprob_reward": -0.1603500247001648, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603500247001648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6989858150482178, + 0.06186122074723244, + 0.03288101404905319, + 0.01922081969678402, + 0.03984110429883003, + 0.03184787184000015, + 0.03898463398218155, + 0.013789888471364975, + 0.039670176804065704, + 0.022917570546269417 + ], + "credit_map": [ + 0.060262300074100494, + 0.2803928256034851, + 0.10440731048583984, + 0.026256993412971497, + 0.05815136805176735, + 0.10829214006662369, + 0.03177228197455406, + 0.02945047989487648, + 0.009032278321683407, + 0.007379063870757818, + 0.006982069928199053, + 0.007395527791231871, + 0.011590978130698204, + 0.0073653473518788815, + 0.005721784196794033, + 0.006394160445779562, + 0.0038491173181682825, + 0.004723615478724241, + 0.0033849207684397697, + 0.0030831163749098778, + 0.003394259372726083, + 0.0037346333265304565, + 0.004025895614176989, + 0.0066854520700871944, + 0.002111303387209773, + 0.0036347736604511738, + 0.0017939573153853416, + 0.001769596477970481, + 0.0029894341714680195, + 0.0025079508777707815, + 0.001686576404608786, + 0.002727225422859192, + 0.005206461064517498, + 0.005952558014541864, + 0.004270677454769611, + 0.003685446921736002, + 0.006817301735281944, + 0.005655097309499979, + 0.004761197604238987, + 0.003492360934615135, + 0.003934504929929972, + 0.0043288785964250565, + 0.003173269098624587, + 0.004243264906108379, + 0.0035763727501034737, + 0.0025830084923654795, + 0.004360353108495474, + 0.005648217163980007, + 0.003635078202933073, + 0.008129251189529896, + 0.004650818649679422, + 0.0032981361728161573, + 0.006079486571252346, + 0.005123391281813383, + 0.004930026363581419, + 0.0031384387984871864, + 0.0019990154542028904, + 0.001916184788569808, + 0.0017403375823050737, + 0.0015388518804684281, + 0.0020717817824333906, + 0.0020668895449489355, + 0.0012943799374625087, + 0.0011624458711594343, + 0.005892864428460598, + 0.004679879639297724, + 0.004569196607917547, + 0.0040242550894618034, + 0.006508298683911562, + 0.00540161831304431, + 0.003986211493611336, + 0.004607845563441515, + 0.0027257297188043594, + 0.0038110115565359592, + 0.002021342981606722, + 0.002286418341100216, + 0.002829601988196373, + 0.00415803911164403, + 0.0025674793869256973, + 0.0025179435033351183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.1603500247001648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293362081050873 + ], + "opposite_logprob_tokens": [ + -2.23793363571167 + ], + "logprob_reward": -0.11293362081050873, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293362081050873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6905079483985901, + 0.10174515098333359, + 0.02919737435877323, + 0.013648062013089657, + 0.026891818270087242, + 0.03135428950190544, + 0.037168409675359726, + 0.019989825785160065, + 0.022737883031368256, + 0.02675916627049446 + ], + "credit_map": [ + 0.06351212412118912, + 0.27803507447242737, + 0.08731278032064438, + 0.02761015109717846, + 0.06269098073244095, + 0.10337278246879578, + 0.036024704575538635, + 0.031949371099472046, + 0.012046054936945438, + 0.013644254766404629, + 0.01783233880996704, + 0.010075600817799568, + 0.01394120417535305, + 0.012057549320161343, + 0.012041359208524227, + 0.010106786154210567, + 0.0038881690707057714, + 0.004106997046619654, + 0.004533863160759211, + 0.0025251025799661875, + 0.0028874026611447334, + 0.002692600479349494, + 0.004103959072381258, + 0.004459278658032417, + 0.001527742831967771, + 0.001964591210708022, + 0.0012863331940025091, + 0.0012640286004170775, + 0.0023671239614486694, + 0.0020162847358733416, + 0.0012954574776813388, + 0.0019265010487288237, + 0.003378363326191902, + 0.003711109282448888, + 0.0030838686507195234, + 0.0030986496713012457, + 0.0044003133662045, + 0.0030871527269482613, + 0.00341478968039155, + 0.0027175708673894405, + 0.004823022987693548, + 0.004221995361149311, + 0.0027719330973923206, + 0.0024134593550115824, + 0.003522473154589534, + 0.005288082640618086, + 0.004893187433481216, + 0.0034201364032924175, + 0.004422900266945362, + 0.0048246802762150764, + 0.005739808548241854, + 0.0036197099834680557, + 0.0042371503077447414, + 0.005227348767220974, + 0.005715699400752783, + 0.003381109330803156, + 0.001844056067056954, + 0.0034247722942382097, + 0.002203033771365881, + 0.0020165927708148956, + 0.002678929129615426, + 0.0036487735342234373, + 0.0019134897738695145, + 0.002260177629068494, + 0.0029099397361278534, + 0.002933299634605646, + 0.0015791482292115688, + 0.001661534421145916, + 0.005581504199653864, + 0.0038105167914181948, + 0.0018592411652207375, + 0.0024026986211538315, + 0.002441487507894635, + 0.00419633137062192, + 0.0023620964493602514, + 0.002688065404072404, + 0.0047905948013067245, + 0.003997649531811476, + 0.003958523739129305, + 0.002324416534975171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.11293362081050873 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945970769971609 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005945970769971609, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945970769971609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16942401230335236, + 0.1422278732061386, + 0.10932166129350662, + 0.022723719477653503, + 0.06592492014169693, + 0.054364435374736786, + 0.13751250505447388, + 0.1030905544757843, + 0.05636155232787132, + 0.1390487551689148 + ], + "credit_map": [ + 0.015618349425494671, + 0.026863930746912956, + 0.01868510991334915, + 0.015490586869418621, + 0.0174855999648571, + 0.03245669603347778, + 0.02274017035961151, + 0.02008357271552086, + 0.015721507370471954, + 0.01723231002688408, + 0.022093405947089195, + 0.012471972964704037, + 0.027916481718420982, + 0.01628321222960949, + 0.017781581729650497, + 0.01272741798311472, + 0.019947456195950508, + 0.010527168400585651, + 0.015670767053961754, + 0.009213399142026901, + 0.00726495124399662, + 0.006759287789463997, + 0.018588460981845856, + 0.021350175142288208, + 0.003034744178876281, + 0.0024924888275563717, + 0.0019950978457927704, + 0.0022280090488493443, + 0.003458211896941066, + 0.0031995372846722603, + 0.0025307482574135065, + 0.0037848849315196276, + 0.008157100528478622, + 0.00943934265524149, + 0.008465261198580265, + 0.0075027127750217915, + 0.0077085648663342, + 0.010211391374468803, + 0.006720768287777901, + 0.007719785440713167, + 0.005336141213774681, + 0.008214752189815044, + 0.0044359564781188965, + 0.0070795766077935696, + 0.006759530398994684, + 0.006547205150127411, + 0.008328873664140701, + 0.007662403397262096, + 0.011520436964929104, + 0.0137871028855443, + 0.01763026975095272, + 0.0163585152477026, + 0.015923310071229935, + 0.018298808485269547, + 0.03361279517412186, + 0.010381271131336689, + 0.011397061869502068, + 0.014761807397007942, + 0.0121374586597085, + 0.0088910898193717, + 0.012462190352380276, + 0.014026444405317307, + 0.013835664838552475, + 0.015578842721879482, + 0.006044252775609493, + 0.009108560159802437, + 0.00471501937136054, + 0.0053785997442901134, + 0.01347381342202425, + 0.008178789168596268, + 0.005871118046343327, + 0.0035914010368287563, + 0.017000406980514526, + 0.016913805156946182, + 0.01356714777648449, + 0.015996966511011124, + 0.019640492275357246, + 0.019153548404574394, + 0.025073055177927017, + 0.011703336611390114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005945970769971609 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a dramatic graphic novel or dark comic book aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stalks depicted with distinctive square stems and sharp geometric edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by intermittent bursts of lightning acting as the primary light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bamboo stalks appear deep emerald green, smoothed and reflective due to rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the long, spear-shaped leaves shown whipping chaotically in thick, slanting rain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3067958652973175, + 0.18017207086086273, + 0.06579770892858505, + 0.048685599118471146, + 0.06229959800839424, + 0.12733829021453857, + 0.05812507122755051, + 0.02644883655011654, + 0.06857747584581375, + 0.05575945973396301 + ], + "credit_map": [ + 0.025726554915308952, + 0.10672418773174286, + 0.013752609491348267, + 0.020973099395632744, + 0.023347649723291397, + 0.07950377464294434, + 0.021083613857626915, + 0.015684394165873528, + 0.01839505322277546, + 0.021226901561021805, + 0.03672327473759651, + 0.01904115080833435, + 0.03391262888908386, + 0.015884196385741234, + 0.019618013873696327, + 0.015370851382613182, + 0.012731451541185379, + 0.007569131441414356, + 0.0074598821811378, + 0.005624460522085428, + 0.006489002611488104, + 0.006460290402173996, + 0.010575054213404655, + 0.008888442069292068, + 0.005648092832416296, + 0.005578422453254461, + 0.004825552925467491, + 0.004720926750451326, + 0.009107336401939392, + 0.006141865625977516, + 0.005196722224354744, + 0.007466684095561504, + 0.008380386047065258, + 0.008486987091600895, + 0.006669134832918644, + 0.007384252734482288, + 0.009213930927217007, + 0.00902455486357212, + 0.0077488841488957405, + 0.005391473416239023, + 0.015045988373458385, + 0.017009001225233078, + 0.010368561372160912, + 0.01371092814952135, + 0.01973590813577175, + 0.01877317577600479, + 0.021610427647829056, + 0.01108430977910757, + 0.0043609910644590855, + 0.005243915598839521, + 0.004964192397892475, + 0.009009744971990585, + 0.00679683918133378, + 0.006145720835775137, + 0.016462992876768112, + 0.005140678025782108, + 0.0033632104750722647, + 0.004191670101135969, + 0.002947527216747403, + 0.0028928883839398623, + 0.003791278460994363, + 0.003441536333411932, + 0.003215141361579299, + 0.0026055865455418825, + 0.010157519020140171, + 0.01183178648352623, + 0.0059644849970936775, + 0.007623129524290562, + 0.010842626914381981, + 0.009512243792414665, + 0.005985526368021965, + 0.006660162005573511, + 0.0046641831286251545, + 0.007344712968915701, + 0.004375532269477844, + 0.005733997095376253, + 0.008225984871387482, + 0.008053481578826904, + 0.010452581569552422, + 0.006908993236720562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0040924143977463245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_tensor([5673, 5673])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_tensor([5673, 5673])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3ac5fc6f99dcc088a9712866cb89362c8c9a5d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_tensor([5673, 5673])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic wide-angle aerial shot, the camera slowly and gracefully orbits an ancient stone tower suspended in the twilight sky. A mythical dragon, its vast, sinuous form, circling the tower in majestic presence. Each scale, seemingly carved from sapphires and emeralds, emanates a steady, luminous bioluminescence from within, pulsing through the deep, scaly hues like breathing with every flex of its muscles. Its head, majestic and regal, bears translucent crystal horns, and gazes forth from golden eyes ablaze with ancient wisdom. The tower itself is constructed of weathered, dark gray stone, its surface covered in intricate carvings and mottled with moss. Its Gothic spire reaches up to a sky tinged with deep indigo and violet. The dragon wasn't flying at a swift speed, but rather gliding through the air with a near-weightless, leisurely grace. Its massive, membranous wings flapped only occasionally and slowly, each time silently stirring the surrounding mist. Its luminous form cast ever-shifting, shimmering reflections on the rough stone face of the ancient pagoda. The entire scene, serene and profound, exuded an awe-inspiring atmosphere of magical power and grandeur.", + "seed": 30568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_5673_seed_30568_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14967977879568933, + "vlm_logprob": -0.14967977879568933, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.14967977879568933, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11642259359359741 + ], + "opposite_logprob_tokens": [ + -2.241422653198242 + ], + "logprob_reward": -0.11642259359359741, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11642259359359741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10591408610343933, + 0.0842457264661789, + 0.06759051233530045, + 0.40693625807762146, + 0.03704548999667168, + 0.11148479580879211, + 0.11503402888774872, + 0.02589324675500393, + 0.03816084563732147, + 0.0076949698850512505 + ], + "credit_map": [ + 0.0085238516330719, + 0.00833346601575613, + 0.014941433444619179, + 0.02130972407758236, + 0.005888942163437605, + 0.015508178621530533, + 0.012915407307446003, + 0.018493087962269783, + 0.007142293732613325, + 0.007243797183036804, + 0.014849314466118813, + 0.008276144042611122, + 0.005338552873581648, + 0.013461688533425331, + 0.009089391678571701, + 0.018844537436962128, + 0.006790281273424625, + 0.007808944210410118, + 0.010018103756010532, + 0.009389610961079597, + 0.006376041565090418, + 0.008528253994882107, + 0.008643371984362602, + 0.010035907849669456, + 0.039888277649879456, + 0.056583985686302185, + 0.03820079192519188, + 0.05242609232664108, + 0.06265102326869965, + 0.0447770357131958, + 0.07391034811735153, + 0.03849870339035988, + 0.005886937491595745, + 0.00465355534106493, + 0.003053983673453331, + 0.0034615558106452227, + 0.002897002035751939, + 0.006846692878752947, + 0.007364583667367697, + 0.0028811802621930838, + 0.017040081322193146, + 0.00711416732519865, + 0.01636481285095215, + 0.011850268580019474, + 0.007000562734901905, + 0.010690531693398952, + 0.024034416303038597, + 0.01738995499908924, + 0.03440522775053978, + 0.008522026240825653, + 0.03435797244310379, + 0.011393410153687, + 0.005169837269932032, + 0.008138935081660748, + 0.0069331335835158825, + 0.006113480776548386, + 0.003735767910256982, + 0.0035236349795013666, + 0.0023305786307901144, + 0.0030019294936209917, + 0.0028870292007923126, + 0.005072901491075754, + 0.003080989932641387, + 0.0022604160476475954, + 0.00743004959076643, + 0.006107669789344072, + 0.0066226269118487835, + 0.005094911437481642, + 0.002850615419447422, + 0.003529648296535015, + 0.0039044590666890144, + 0.002620863728225231, + 0.0008569415658712387, + 0.0008665634668432176, + 0.0007534598698839545, + 0.0010264237644150853, + 0.000841636850964278, + 0.001114252139814198, + 0.0012423540465533733, + 0.0009933380642905831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.11642259359359741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968174546957016 + ], + "opposite_logprob_tokens": [ + -2.9299681186676025 + ], + "logprob_reward": -0.054968174546957016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968174546957016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17435021698474884, + 0.12905022501945496, + 0.11599976569414139, + 0.14774090051651, + 0.12123595923185349, + 0.06826574355363846, + 0.12354866415262222, + 0.02892988733947277, + 0.0688924565911293, + 0.021986236795783043 + ], + "credit_map": [ + 0.023229006677865982, + 0.01774528995156288, + 0.022117743268609047, + 0.030818304046988487, + 0.014636131003499031, + 0.020858926698565483, + 0.019944041967391968, + 0.025000758469104767, + 0.015734771266579628, + 0.011829636059701443, + 0.020070958882570267, + 0.02721915766596794, + 0.010277730412781239, + 0.015193985775113106, + 0.015981463715434074, + 0.012742499820888042, + 0.009961234405636787, + 0.014496776275336742, + 0.012986531481146812, + 0.018793458119034767, + 0.013599098660051823, + 0.015685033053159714, + 0.01563849300146103, + 0.014839128591120243, + 0.013842430897057056, + 0.012438852339982986, + 0.012772377580404282, + 0.031313277781009674, + 0.013185454532504082, + 0.017629291862249374, + 0.038392677903175354, + 0.008166520856320858, + 0.028914617374539375, + 0.008876413106918335, + 0.017101580277085304, + 0.006470857188105583, + 0.00657119182869792, + 0.029597170650959015, + 0.013323377817869186, + 0.010380737483501434, + 0.009712640196084976, + 0.006863846909254789, + 0.01308534387499094, + 0.006829801015555859, + 0.005045719910413027, + 0.007221677806228399, + 0.012447059154510498, + 0.007059648633003235, + 0.04313257709145546, + 0.00677749328315258, + 0.042718999087810516, + 0.007924774661660194, + 0.003973867744207382, + 0.00613842997699976, + 0.006586688105016947, + 0.006295811850577593, + 0.005475476384162903, + 0.003683295100927353, + 0.0033424729481339455, + 0.0036946251057088375, + 0.0030205738730728626, + 0.003051277482882142, + 0.003843587590381503, + 0.0028185774572193623, + 0.012588145211338997, + 0.008720006793737411, + 0.011477562598884106, + 0.007028147578239441, + 0.005780929699540138, + 0.010118316859006882, + 0.006906640715897083, + 0.006272698286920786, + 0.003520283382385969, + 0.0032732391264289618, + 0.002444666111841798, + 0.0032031165901571512, + 0.001416706945747137, + 0.003443839494138956, + 0.002764329547062516, + 0.0019200514070689678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.054968174546957016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486032485961914 + ], + "opposite_logprob_tokens": [ + -1.2236032485961914 + ], + "logprob_reward": -0.3486032485961914, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486032485961914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11039666086435318, + 0.17031045258045197, + 0.37179556488990784, + 0.0531715452671051, + 0.16651812195777893, + 0.03100348636507988, + 0.034861475229263306, + 0.01624077372252941, + 0.02201605960726738, + 0.023685839027166367 + ], + "credit_map": [ + 0.01401728205382824, + 0.01077287644147873, + 0.015178981237113476, + 0.020193425938487053, + 0.007235495839267969, + 0.013941019773483276, + 0.011479617096483707, + 0.017577970400452614, + 0.02707546204328537, + 0.013163677416741848, + 0.020068049430847168, + 0.029849175363779068, + 0.016483217477798462, + 0.01945551671087742, + 0.019923288375139236, + 0.02429206296801567, + 0.03654281049966812, + 0.03144507110118866, + 0.051136650145053864, + 0.04458731412887573, + 0.03536598011851311, + 0.04565908759832382, + 0.04491365700960159, + 0.08214502036571503, + 0.005683688446879387, + 0.007001716177910566, + 0.004083575680851936, + 0.008718723431229591, + 0.00643010763451457, + 0.0067658680491149426, + 0.009033792652189732, + 0.005454068537801504, + 0.03129313513636589, + 0.016355052590370178, + 0.02220272831618786, + 0.011040683835744858, + 0.011515581980347633, + 0.03309805318713188, + 0.031885966658592224, + 0.009126923978328705, + 0.003906056983396411, + 0.0037710892502218485, + 0.005407823249697685, + 0.004025018308311701, + 0.002136167837306857, + 0.0024836768861860037, + 0.005981822032481432, + 0.003291831584647298, + 0.005102560389786959, + 0.003897265065461397, + 0.003651794046163559, + 0.005776144098490477, + 0.0033079758286476135, + 0.004496216308325529, + 0.005025945603847504, + 0.003603575052693486, + 0.002286303788423538, + 0.0021509926300495863, + 0.001376780797727406, + 0.002915787510573864, + 0.001726327114738524, + 0.0022588653955608606, + 0.0020478330552577972, + 0.0014778829645365477, + 0.0027737836353480816, + 0.0032434465829283, + 0.0025770231150090694, + 0.002772534964606166, + 0.0022526264656335115, + 0.0036102565936744213, + 0.002545315772294998, + 0.002241073874756694, + 0.0026411644648760557, + 0.003377302782610059, + 0.0019988534040749073, + 0.0028157129418104887, + 0.0032337335869669914, + 0.003537376644089818, + 0.003348112804815173, + 0.0027335812337696552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.3486032485961914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225589781999588 + ], + "opposite_logprob_tokens": [ + -1.6005897521972656 + ], + "logprob_reward": -0.225589781999588, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225589781999588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10857941210269928, + 0.09472765028476715, + 0.0761723518371582, + 0.20222057402133942, + 0.2777380645275116, + 0.12962444126605988, + 0.057261873036623, + 0.015276843681931496, + 0.026588765904307365, + 0.011810000985860825 + ], + "credit_map": [ + 0.023698022589087486, + 0.011691175401210785, + 0.014101504348218441, + 0.011019454337656498, + 0.008559197187423706, + 0.013101162388920784, + 0.011165845207870007, + 0.015243051573634148, + 0.016547955572605133, + 0.010609542950987816, + 0.015541356056928635, + 0.008641937747597694, + 0.008957775309681892, + 0.009151248261332512, + 0.01023879460990429, + 0.015039040707051754, + 0.008913049474358559, + 0.008007326163351536, + 0.00976286269724369, + 0.010636964812874794, + 0.011760242283344269, + 0.008771865628659725, + 0.011583478190004826, + 0.006736563052982092, + 0.02189973182976246, + 0.03190557658672333, + 0.01564035192131996, + 0.03759287670254707, + 0.02076365053653717, + 0.02093607746064663, + 0.03946778550744057, + 0.01401453372091055, + 0.04667288064956665, + 0.022383838891983032, + 0.039131831377744675, + 0.01988520473241806, + 0.014866913668811321, + 0.05104592815041542, + 0.06883428990840912, + 0.014917178079485893, + 0.019522063434123993, + 0.01472645252943039, + 0.018876977264881134, + 0.013159850612282753, + 0.007195585872977972, + 0.010343750938773155, + 0.02746322564780712, + 0.018336545675992966, + 0.008859818801283836, + 0.006410346832126379, + 0.0067614042200148106, + 0.0068806130439043045, + 0.005171601660549641, + 0.009113580919802189, + 0.007822067476809025, + 0.006242442410439253, + 0.002819347893819213, + 0.002574480837211013, + 0.0015424324665218592, + 0.0015986873768270016, + 0.0013589771697297692, + 0.0027639081235975027, + 0.0012763107661157846, + 0.0013426982332020998, + 0.0038806567899882793, + 0.003133082063868642, + 0.003319638315588236, + 0.0027192046400159597, + 0.0022339290007948875, + 0.004699812736362219, + 0.0035761548206210136, + 0.0030262877698987722, + 0.0015517794527113438, + 0.0016560290241613984, + 0.0012864219024777412, + 0.0016188869485631585, + 0.00127767666708678, + 0.0018480761209502816, + 0.001571978209540248, + 0.000999151961877942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.225589781999588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815095242112875 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815095242112875, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815095242112875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16944298148155212, + 0.04640180990099907, + 0.06659811735153198, + 0.0803331583738327, + 0.11169275641441345, + 0.3206478953361511, + 0.11893793195486069, + 0.021664217114448547, + 0.034842342138290405, + 0.029438799247145653 + ], + "credit_map": [ + 0.032190654426813126, + 0.01617308333516121, + 0.019060708582401276, + 0.02082163840532303, + 0.01410624012351036, + 0.021237408742308617, + 0.017045853659510612, + 0.028807399794459343, + 0.008339224383234978, + 0.0057999505661427975, + 0.0072578261606395245, + 0.004201414529234171, + 0.005096845794469118, + 0.005510777700692415, + 0.004445772618055344, + 0.005749991629272699, + 0.0064231110736727715, + 0.008446484804153442, + 0.008714916184544563, + 0.009329930879175663, + 0.0068894606083631516, + 0.00997287780046463, + 0.008350701071321964, + 0.00847062561661005, + 0.006959879770874977, + 0.010617422871291637, + 0.0077299210242927074, + 0.012076593935489655, + 0.01208314299583435, + 0.008570132777094841, + 0.013506582006812096, + 0.0087894843891263, + 0.016172263771295547, + 0.008929576724767685, + 0.012732462957501411, + 0.008827411569654942, + 0.006620274391025305, + 0.019976694136857986, + 0.029517587274312973, + 0.008916482329368591, + 0.056270092725753784, + 0.02339049242436886, + 0.05978943035006523, + 0.021611619740724564, + 0.022978736087679863, + 0.03035667911171913, + 0.06804845482110977, + 0.03820238634943962, + 0.04401983693242073, + 0.006965023931115866, + 0.04307888075709343, + 0.006962027866393328, + 0.0037670519668608904, + 0.00564228231087327, + 0.0035238536074757576, + 0.004978966899216175, + 0.003399122040718794, + 0.0037281110417097807, + 0.0022166846320033073, + 0.002281530760228634, + 0.002402368001639843, + 0.003378922352567315, + 0.0026891182642430067, + 0.0015683576930314302, + 0.006635046564042568, + 0.004953055176883936, + 0.0063159530982375145, + 0.003096567466855049, + 0.0031552643049508333, + 0.004863964393734932, + 0.003022903809323907, + 0.0027995831333100796, + 0.00438646087422967, + 0.004090179223567247, + 0.003547731786966324, + 0.003917590249329805, + 0.0023576735984534025, + 0.003748005721718073, + 0.003969890996813774, + 0.0034212637692689896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.002815095242112875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle aerial shot, the camera slowly and gracefully orbits an ancient stone tower suspended in the twilight sky. A mythical dragon, its vast, sinuous form, circling the tower in majestic presence. Each scale, seemingly carved from sapphires and emeralds, emanates a steady, luminous bioluminescence from within, pulsing through the deep, scaly hues like breathing with every flex of its muscles. Its head, majestic and regal, bears translucent crystal horns, and gazes forth from golden eyes ablaze with ancient wisdom. The tower itself is constructed of weathered, dark gray stone, its surface covered in intricate carvings and mottled with moss. Its Gothic spire reaches up to a sky tinged with deep indigo and violet. The dragon wasn't flying at a swift speed, but rather gliding through the air with a near-weightless, leisurely grace. Its massive, membranous wings flapped only occasionally and slowly, each time silently stirring the surrounding mist. Its luminous form cast ever-shifting, shimmering reflections on the rough stone face of the ancient pagoda. The entire scene, serene and profound, exuded an awe-inspiring atmosphere of magical power and grandeur.", + "seed": 30569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_5673_seed_30569_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17183703016489743, + "vlm_logprob": -0.17183703016489743, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.17183703016489743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.581439733505249 + ], + "opposite_logprob_tokens": [ + -0.831439733505249 + ], + "logprob_reward": -0.581439733505249, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.581439733505249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13912621140480042, + 0.14277109503746033, + 0.12309527397155762, + 0.02546434849500656, + 0.034255560487508774, + 0.05736007168889046, + 0.05297565832734108, + 0.10335250943899155, + 0.0933740958571434, + 0.22822509706020355 + ], + "credit_map": [ + 0.01468704454600811, + 0.010773911140859127, + 0.023507019504904747, + 0.01683415099978447, + 0.012591429054737091, + 0.02309240587055683, + 0.014034257270395756, + 0.02360600233078003, + 0.01000705175101757, + 0.011583509854972363, + 0.034108128398656845, + 0.017016708850860596, + 0.015847040340304375, + 0.023510506376624107, + 0.012927831150591373, + 0.017770322039723396, + 0.014188673347234726, + 0.012927201576530933, + 0.019000718370079994, + 0.012093138881027699, + 0.013774363324046135, + 0.010002806782722473, + 0.022922584787011147, + 0.01818578690290451, + 0.002848308067768812, + 0.0022491575218737125, + 0.006627218797802925, + 0.003680169116705656, + 0.003250324400141835, + 0.0035695377737283707, + 0.0017587757902219892, + 0.0014808576088398695, + 0.004863242618739605, + 0.005178755149245262, + 0.003495696932077408, + 0.003992334008216858, + 0.005211475770920515, + 0.004884805530309677, + 0.0027469638735055923, + 0.003882286138832569, + 0.0052312989719212055, + 0.006258814129978418, + 0.007101021707057953, + 0.008333874866366386, + 0.0053903949446976185, + 0.006790202111005783, + 0.009484297595918179, + 0.008770166896283627, + 0.004170594736933708, + 0.004973676986992359, + 0.008690510876476765, + 0.007809056434780359, + 0.00751854432746768, + 0.00501939607784152, + 0.007378047797828913, + 0.007415832951664925, + 0.006892422679811716, + 0.010552975349128246, + 0.010693751275539398, + 0.018540600314736366, + 0.006651845760643482, + 0.010549137368798256, + 0.026053212583065033, + 0.01341856736689806, + 0.007837695069611073, + 0.011945723555982113, + 0.010769704356789589, + 0.013770056888461113, + 0.00861946027725935, + 0.011136536486446857, + 0.015451001934707165, + 0.013843919150531292, + 0.01773366518318653, + 0.016255836933851242, + 0.06248632073402405, + 0.021417614072561264, + 0.01999458484351635, + 0.01507918443530798, + 0.023914435878396034, + 0.05134345591068268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.581439733505249 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059713751077652 + ], + "opposite_logprob_tokens": [ + -3.895597219467163 + ], + "logprob_reward": -0.02059713751077652, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059713751077652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12785939872264862, + 0.08370897173881531, + 0.14285168051719666, + 0.08413279056549072, + 0.06996636092662811, + 0.029152877628803253, + 0.03794000670313835, + 0.1309497058391571, + 0.06759122759103775, + 0.22584697604179382 + ], + "credit_map": [ + 0.016311561688780785, + 0.01053562667220831, + 0.012202338315546513, + 0.011313219554722309, + 0.016095299273729324, + 0.020786063745617867, + 0.015211209654808044, + 0.025404075160622597, + 0.006826144643127918, + 0.009198903106153011, + 0.012939057312905788, + 0.007697084452956915, + 0.013680096715688705, + 0.0140414172783494, + 0.00801768433302641, + 0.011308578774333, + 0.018383968621492386, + 0.01671554706990719, + 0.013622856698930264, + 0.012844845652580261, + 0.018784049898386, + 0.017265560105443, + 0.026455989107489586, + 0.018778864294290543, + 0.008076396770775318, + 0.008232620544731617, + 0.02180253341794014, + 0.01016849186271429, + 0.006134252529591322, + 0.011727545410394669, + 0.010702835395932198, + 0.007288113236427307, + 0.007248465437442064, + 0.010371644049882889, + 0.010447140783071518, + 0.0060640620067715645, + 0.00748019153252244, + 0.009516808204352856, + 0.00669318251311779, + 0.012144867330789566, + 0.002530045574530959, + 0.0031721352133899927, + 0.004112918861210346, + 0.0033437171950936317, + 0.0026265615597367287, + 0.0037409684155136347, + 0.004976601339876652, + 0.004649929236620665, + 0.0032831851858645678, + 0.003223798470571637, + 0.007869943976402283, + 0.0035355237778276205, + 0.0041673011146485806, + 0.003507287474349141, + 0.0033964707981795073, + 0.00895649567246437, + 0.010265507735311985, + 0.018213286995887756, + 0.011185985989868641, + 0.01635139249265194, + 0.012987358495593071, + 0.009786991402506828, + 0.028864659368991852, + 0.023294519633054733, + 0.007652599364519119, + 0.00766031676903367, + 0.009896463714540005, + 0.008930694311857224, + 0.008913284167647362, + 0.006307907402515411, + 0.009262938052415848, + 0.008967030793428421, + 0.02082263119518757, + 0.016522875055670738, + 0.057622019201517105, + 0.019851895049214363, + 0.023286525160074234, + 0.015565354377031326, + 0.02211497537791729, + 0.0500607006251812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.02059713751077652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155824720859528 + ], + "opposite_logprob_tokens": [ + -1.701558232307434 + ], + "logprob_reward": -0.20155824720859528, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155824720859528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20878329873085022, + 0.13355344533920288, + 0.10511932522058487, + 0.11085823178291321, + 0.06636989861726761, + 0.06333835422992706, + 0.023166323080658913, + 0.0917423889040947, + 0.08275431394577026, + 0.11431446671485901 + ], + "credit_map": [ + 0.028735270723700523, + 0.016543572768568993, + 0.033334095031023026, + 0.023840580135583878, + 0.019455846399068832, + 0.030244173482060432, + 0.01933859847486019, + 0.037291157990694046, + 0.014301076531410217, + 0.011246914975345135, + 0.020680410787463188, + 0.012191420421004295, + 0.019663890823721886, + 0.024930568411946297, + 0.012015928514301777, + 0.018523234874010086, + 0.012783266603946686, + 0.013835860416293144, + 0.012446925044059753, + 0.008700728416442871, + 0.010640781372785568, + 0.011356418021023273, + 0.01915759965777397, + 0.016197746619582176, + 0.011400742456316948, + 0.009827851317822933, + 0.03467313200235367, + 0.013668718747794628, + 0.009630645625293255, + 0.016413262113928795, + 0.007639341522008181, + 0.007604536600410938, + 0.008851096034049988, + 0.009873290546238422, + 0.010361634194850922, + 0.007458765525370836, + 0.007037211209535599, + 0.00910896621644497, + 0.006101681385189295, + 0.007577258162200451, + 0.008631229400634766, + 0.008991539478302002, + 0.007108659949153662, + 0.006932433694601059, + 0.005573710892349482, + 0.008755096234381199, + 0.006648845039308071, + 0.010696837678551674, + 0.002039820421487093, + 0.002170526422560215, + 0.004801598843187094, + 0.0024882385041564703, + 0.0029315908905118704, + 0.0018079446163028479, + 0.0023958394303917885, + 0.004530763253569603, + 0.007214365992695093, + 0.013385522179305553, + 0.007989090867340565, + 0.012796849012374878, + 0.010211561806499958, + 0.006842983886599541, + 0.015790507197380066, + 0.017511509358882904, + 0.005751722026616335, + 0.012744470499455929, + 0.008809163235127926, + 0.013124805875122547, + 0.01356565486639738, + 0.01023270282894373, + 0.009157577529549599, + 0.009368212893605232, + 0.009004306979477406, + 0.012601830065250397, + 0.019131511449813843, + 0.014522603712975979, + 0.018399324268102646, + 0.011897778138518333, + 0.014471659436821938, + 0.014285454526543617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.20155824720859528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305985942482948 + ], + "opposite_logprob_tokens": [ + -3.168059825897217 + ], + "logprob_reward": -0.04305985942482948, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305985942482948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12399175763130188, + 0.07368553429841995, + 0.05828595161437988, + 0.07546021789312363, + 0.10039571672677994, + 0.07399934530258179, + 0.14935415983200073, + 0.09264099597930908, + 0.18548212945461273, + 0.06670419126749039 + ], + "credit_map": [ + 0.015718622133135796, + 0.014505656436085701, + 0.022558677941560745, + 0.014233277179300785, + 0.011475291103124619, + 0.016949722543358803, + 0.013093330897390842, + 0.01545717753469944, + 0.008389698341488838, + 0.009123998694121838, + 0.014697466045618057, + 0.006815912667661905, + 0.007814310491085052, + 0.010590311139822006, + 0.006657099351286888, + 0.009596736170351505, + 0.008466197177767754, + 0.005638533271849155, + 0.008510402403771877, + 0.0062151686288416386, + 0.007160883396863937, + 0.005797583144158125, + 0.008462140336632729, + 0.008035045117139816, + 0.008743040263652802, + 0.007845435291528702, + 0.018801338970661163, + 0.009318496100604534, + 0.007282486651092768, + 0.010679083876311779, + 0.006875641644001007, + 0.005914703011512756, + 0.009237624704837799, + 0.011876133270561695, + 0.015599813312292099, + 0.012515462934970856, + 0.015465187840163708, + 0.013414600864052773, + 0.010547359474003315, + 0.011739538982510567, + 0.008169489912688732, + 0.005555749870836735, + 0.009216704405844212, + 0.010449408553540707, + 0.0084133455529809, + 0.00935132335871458, + 0.008683627471327782, + 0.014159704558551311, + 0.011238412000238895, + 0.018540266901254654, + 0.023976873606443405, + 0.016977224498987198, + 0.02095569111406803, + 0.015297232195734978, + 0.015851425006985664, + 0.026517052203416824, + 0.006466188933700323, + 0.008117105811834335, + 0.00921362079679966, + 0.015493139624595642, + 0.0059646419249475, + 0.012263844721019268, + 0.02252456359565258, + 0.012597889639437199, + 0.027028430253267288, + 0.024189632385969162, + 0.024120893329381943, + 0.01980769634246826, + 0.026623420417308807, + 0.020677782595157623, + 0.022561607882380486, + 0.020472658798098564, + 0.005166432820260525, + 0.004949580412358046, + 0.020614327862858772, + 0.0054977089166641235, + 0.004387691617012024, + 0.004119729623198509, + 0.006817860063165426, + 0.01515086181461811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.04305985942482948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012530173175036907 + ], + "opposite_logprob_tokens": [ + -4.387530326843262 + ], + "logprob_reward": -0.012530173175036907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012530173175036907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16162435710430145, + 0.0213344544172287, + 0.045590128749608994, + 0.025302063673734665, + 0.12037473917007446, + 0.04853380098938942, + 0.053345687687397, + 0.039211783558130264, + 0.14278285205364227, + 0.341900110244751 + ], + "credit_map": [ + 0.01441054418683052, + 0.01186488289386034, + 0.020703081041574478, + 0.013475559651851654, + 0.01268617995083332, + 0.034573644399642944, + 0.015618125908076763, + 0.03829234093427658, + 0.0022742950823158026, + 0.0025236334186047316, + 0.0028087703976780176, + 0.0026517570950090885, + 0.0032248804345726967, + 0.0029192545916885138, + 0.0022158604115247726, + 0.0027160036843270063, + 0.006329691968858242, + 0.004121824167668819, + 0.0051039825193583965, + 0.0038883856032043695, + 0.006653352174907923, + 0.006295503117144108, + 0.0066157374531030655, + 0.006581650581210852, + 0.0033002851996570826, + 0.0024846557062119246, + 0.0060646324418485165, + 0.004016191232949495, + 0.0017344775842502713, + 0.003636361565440893, + 0.0019315020181238651, + 0.0021339573431760073, + 0.011382527649402618, + 0.015795700252056122, + 0.01600547693669796, + 0.010620239190757275, + 0.014410730451345444, + 0.016511080786585808, + 0.014721223153173923, + 0.02092777192592621, + 0.005393619649112225, + 0.004806336481124163, + 0.006599374581128359, + 0.006390327122062445, + 0.006192288361489773, + 0.00635767774656415, + 0.005464720539748669, + 0.007329457905143499, + 0.004108543507754803, + 0.0043306234292685986, + 0.010379606857895851, + 0.0066702645272016525, + 0.008335838094353676, + 0.006663826759904623, + 0.005703756585717201, + 0.007153227925300598, + 0.0032178102992475033, + 0.0045767962001264095, + 0.003047073958441615, + 0.006356851197779179, + 0.0043698265217244625, + 0.0049349116161465645, + 0.007224353030323982, + 0.005484164692461491, + 0.012797161005437374, + 0.021685829386115074, + 0.016256097704172134, + 0.026756474748253822, + 0.01699579879641533, + 0.009541070088744164, + 0.019014915451407433, + 0.01973550207912922, + 0.030454901978373528, + 0.011453538201749325, + 0.1250423938035965, + 0.021958155557513237, + 0.014813625253736973, + 0.014507701620459557, + 0.027515439316630363, + 0.0961543545126915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012530173175036907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle aerial shot, the camera slowly and gracefully orbits an ancient stone tower suspended in the twilight sky. A mythical dragon, its vast, sinuous form, circling the tower in majestic presence. Each scale, seemingly carved from sapphires and emeralds, emanates a steady, luminous bioluminescence from within, pulsing through the deep, scaly hues like breathing with every flex of its muscles. Its head, majestic and regal, bears translucent crystal horns, and gazes forth from golden eyes ablaze with ancient wisdom. The tower itself is constructed of weathered, dark gray stone, its surface covered in intricate carvings and mottled with moss. Its Gothic spire reaches up to a sky tinged with deep indigo and violet. The dragon wasn't flying at a swift speed, but rather gliding through the air with a near-weightless, leisurely grace. Its massive, membranous wings flapped only occasionally and slowly, each time silently stirring the surrounding mist. Its luminous form cast ever-shifting, shimmering reflections on the rough stone face of the ancient pagoda. The entire scene, serene and profound, exuded an awe-inspiring atmosphere of magical power and grandeur.", + "seed": 30570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_5673_seed_30570_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04690790278837085, + "vlm_logprob": -0.04690790278837085, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04690790278837085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0482504665851593 + ], + "opposite_logprob_tokens": [ + -3.173250436782837 + ], + "logprob_reward": -0.0482504665851593, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0482504665851593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3726791739463806, + 0.10001525282859802, + 0.07629551738500595, + 0.0625341460108757, + 0.05194832384586334, + 0.031830936670303345, + 0.057088252156972885, + 0.13066013157367706, + 0.08320885896682739, + 0.03373946249485016 + ], + "credit_map": [ + 0.06488848477602005, + 0.034161526709795, + 0.11928590387105942, + 0.05808635428547859, + 0.02716759219765663, + 0.021451439708471298, + 0.027850808575749397, + 0.019787076860666275, + 0.007950159721076488, + 0.009630069136619568, + 0.024344507604837418, + 0.02219172939658165, + 0.006912384647876024, + 0.009596162475645542, + 0.00940927304327488, + 0.009980972856283188, + 0.007777045946568251, + 0.010628917254507542, + 0.013022231869399548, + 0.011866793967783451, + 0.008769742213189602, + 0.0067193713039159775, + 0.009916570037603378, + 0.0075948419980704784, + 0.005666568409651518, + 0.005778875667601824, + 0.007420444395393133, + 0.007753355428576469, + 0.006474556401371956, + 0.009044121019542217, + 0.010988096706569195, + 0.009408121928572655, + 0.0055407630279660225, + 0.0057511599734425545, + 0.008172597736120224, + 0.005585118662565947, + 0.005650558974593878, + 0.00999457947909832, + 0.0037761060521006584, + 0.007477441802620888, + 0.004094510804861784, + 0.0034969449043273926, + 0.0048850481398403645, + 0.006724975537508726, + 0.0023902684915810823, + 0.0035239863209426403, + 0.0032651752699166536, + 0.0034500265028327703, + 0.004009583033621311, + 0.006052196491509676, + 0.015794124454259872, + 0.013636641204357147, + 0.005008822772651911, + 0.005821019411087036, + 0.0034854616969823837, + 0.003280405420809984, + 0.011837102472782135, + 0.017648771405220032, + 0.008372630923986435, + 0.014600030146539211, + 0.02701408788561821, + 0.02012798935174942, + 0.012270931154489517, + 0.018788589164614677, + 0.015562303364276886, + 0.009570784866809845, + 0.02027548849582672, + 0.007192525081336498, + 0.012188872322440147, + 0.006448103580623865, + 0.0070248250849545, + 0.004945950582623482, + 0.005143618676811457, + 0.004853763151913881, + 0.0037341779097914696, + 0.0040984260849654675, + 0.0056229024194180965, + 0.0031248207669705153, + 0.003916549030691385, + 0.003245204919949174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0482504665851593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371066227555275 + ], + "opposite_logprob_tokens": [ + -3.4087107181549072 + ], + "logprob_reward": -0.03371066227555275, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371066227555275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14533662796020508, + 0.08915705233812332, + 0.0623735748231411, + 0.02723446488380432, + 0.03314826637506485, + 0.046047184616327286, + 0.02329258620738983, + 0.2565143406391144, + 0.19773173332214355, + 0.11916410177946091 + ], + "credit_map": [ + 0.015097575262188911, + 0.021291201934218407, + 0.026890603825449944, + 0.019737394526600838, + 0.021898755803704262, + 0.009770050644874573, + 0.01624847762286663, + 0.014402585104107857, + 0.00826278980821371, + 0.010415250435471535, + 0.02268430031836033, + 0.02319241315126419, + 0.00461986567825079, + 0.007712421007454395, + 0.005034218542277813, + 0.007235807832330465, + 0.005725158378481865, + 0.007628840859979391, + 0.012850023806095123, + 0.010740278288722038, + 0.006642739288508892, + 0.007729321252554655, + 0.005273995455354452, + 0.005783226806670427, + 0.0028266794979572296, + 0.00207724841311574, + 0.002867253730073571, + 0.003310384228825569, + 0.002483430551365018, + 0.004546639043837786, + 0.004803179297596216, + 0.0043196529150009155, + 0.0028965778183192015, + 0.003632664680480957, + 0.0051516322419047356, + 0.00364707806147635, + 0.0029986239969730377, + 0.007883713580667973, + 0.0025890441611409187, + 0.004348937887698412, + 0.005993477534502745, + 0.004854815546423197, + 0.006087837275117636, + 0.00669722119346261, + 0.004401938524097204, + 0.006260276772081852, + 0.004619094543159008, + 0.007132525555789471, + 0.0018651237478479743, + 0.0019091450376436114, + 0.0062640816904604435, + 0.004427265375852585, + 0.003268220229074359, + 0.00212485552765429, + 0.002013380406424403, + 0.0014205160550773144, + 0.021406998857855797, + 0.025947192683815956, + 0.021765796467661858, + 0.03160698711872101, + 0.04308971390128136, + 0.041064050048589706, + 0.023896649479866028, + 0.047736965119838715, + 0.05443919450044632, + 0.013101283460855484, + 0.06906244158744812, + 0.011797476559877396, + 0.015813196077942848, + 0.011020279489457607, + 0.009033670648932457, + 0.013464213348925114, + 0.015818998217582703, + 0.016856960952281952, + 0.01701539196074009, + 0.012219938449561596, + 0.012986152432858944, + 0.016419995576143265, + 0.011628080159425735, + 0.01621859148144722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.03371066227555275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048643991351127625 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.048643991351127625, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048643991351127625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5987443327903748, + 0.074068583548069, + 0.04365924745798111, + 0.01970168575644493, + 0.02837735414505005, + 0.03157706931233406, + 0.025630950927734375, + 0.046885278075933456, + 0.103782519698143, + 0.027572963386774063 + ], + "credit_map": [ + 0.1452092081308365, + 0.03571819141507149, + 0.257029265165329, + 0.11749479174613953, + 0.009908132255077362, + 0.009592707268893719, + 0.0123678520321846, + 0.011424144729971886, + 0.010920306667685509, + 0.014796524308621883, + 0.01575871929526329, + 0.013520343229174614, + 0.005344809964299202, + 0.005151070188730955, + 0.004378957208245993, + 0.004197850823402405, + 0.005517963785678148, + 0.008918672800064087, + 0.006747996434569359, + 0.006384616252034903, + 0.0047695753164589405, + 0.004361448809504509, + 0.0035857108887284994, + 0.003373261308297515, + 0.002059357473626733, + 0.002014746190980077, + 0.00266525661572814, + 0.0033011066261678934, + 0.0019078279146924615, + 0.0020958827808499336, + 0.00309635978192091, + 0.0025611468590795994, + 0.0032725848723202944, + 0.0038306377828121185, + 0.0034013583790510893, + 0.002615487203001976, + 0.003203916596248746, + 0.006109876092523336, + 0.0025235444772988558, + 0.0034199461806565523, + 0.004413883667439222, + 0.004886928014457226, + 0.004512096755206585, + 0.00361950253136456, + 0.0024352711625397205, + 0.003943065647035837, + 0.0032326215878129005, + 0.004533701576292515, + 0.003964518662542105, + 0.0033242106437683105, + 0.0030141689348965883, + 0.005228590220212936, + 0.0028389173094183207, + 0.0014914883067831397, + 0.0035522703547030687, + 0.002216783119365573, + 0.00501338392496109, + 0.0036462307907640934, + 0.005604836158454418, + 0.005478326231241226, + 0.006766976322978735, + 0.008786069229245186, + 0.0046208444982767105, + 0.006968606263399124, + 0.021249715238809586, + 0.011864698491990566, + 0.025787657126784325, + 0.008036059327423573, + 0.012478135526180267, + 0.008667102083563805, + 0.007410568650811911, + 0.008288570679724216, + 0.0037551759742200375, + 0.004133488982915878, + 0.0038506595883518457, + 0.0027736136689782143, + 0.0044312975369393826, + 0.0029916749335825443, + 0.0029660428408533335, + 0.002671007765457034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.048643991351127625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032197088003159 + ], + "opposite_logprob_tokens": [ + -2.3503220081329346 + ], + "logprob_reward": -0.10032197088003159, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032197088003159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5631439089775085, + 0.09790810942649841, + 0.049685344099998474, + 0.027377767488360405, + 0.038068391382694244, + 0.0389341339468956, + 0.012539003044366837, + 0.022959375753998756, + 0.12459007650613785, + 0.024793844670057297 + ], + "credit_map": [ + 0.11967825144529343, + 0.030415797606110573, + 0.22081778943538666, + 0.1042974665760994, + 0.023715810850262642, + 0.015671951696276665, + 0.027439482510089874, + 0.021107343956828117, + 0.007725797593593597, + 0.00849338248372078, + 0.023148098960518837, + 0.023226363584399223, + 0.005674238316714764, + 0.010620137676596642, + 0.008013520389795303, + 0.011006576009094715, + 0.0037156036123633385, + 0.0046559846960008144, + 0.008494578301906586, + 0.008355215191841125, + 0.006472094915807247, + 0.0073456596583127975, + 0.005437556188553572, + 0.005208650138229132, + 0.0022067250683903694, + 0.002204463118687272, + 0.00346362148411572, + 0.0035606659948825836, + 0.003011261811479926, + 0.004436127375811338, + 0.00477452902123332, + 0.0037203719839453697, + 0.0025195188354700804, + 0.0027691670693457127, + 0.004180761519819498, + 0.0033802625257521868, + 0.0035248692147433758, + 0.011874997057020664, + 0.002566432347521186, + 0.007252384442836046, + 0.0062420242466032505, + 0.0042971037328243256, + 0.004846669267863035, + 0.004232832230627537, + 0.0032622837461531162, + 0.005523758474737406, + 0.0037512192502617836, + 0.006778241600841284, + 0.001002974109724164, + 0.0010531633161008358, + 0.0034585073590278625, + 0.0026550753973424435, + 0.0013913653092458844, + 0.0014460962265729904, + 0.000856775208376348, + 0.000675046001560986, + 0.002845859620720148, + 0.0023042820394039154, + 0.0031659617088735104, + 0.003376150270923972, + 0.0026658731512725353, + 0.0029165646992623806, + 0.0028436484280973673, + 0.002841036766767502, + 0.03789810836315155, + 0.007962940260767937, + 0.04897148534655571, + 0.006184187717735767, + 0.006664188113063574, + 0.007378880865871906, + 0.0048518371768295765, + 0.004678451456129551, + 0.003231724724173546, + 0.004080444574356079, + 0.0028928646352142096, + 0.002283873502165079, + 0.0031275912187993526, + 0.004178754985332489, + 0.002272237092256546, + 0.0027263567317277193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.10032197088003159 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003612422849982977 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.003612422849982977, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003612422849982977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29515016078948975, + 0.119092658162117, + 0.031847041100263596, + 0.08130194246768951, + 0.03688216581940651, + 0.028629356995224953, + 0.1917010396718979, + 0.030587391927838326, + 0.09084311872720718, + 0.09396520256996155 + ], + "credit_map": [ + 0.06629444658756256, + 0.021803932264447212, + 0.11713486909866333, + 0.05776306241750717, + 0.01010184921324253, + 0.0073512159287929535, + 0.008072705008089542, + 0.006628050003200769, + 0.011799965985119343, + 0.017821617424488068, + 0.02961176261305809, + 0.03037678264081478, + 0.0072518568485975266, + 0.005522702354937792, + 0.007653533015400171, + 0.009054427966475487, + 0.0029969799797981977, + 0.004492498002946377, + 0.005544425919651985, + 0.004495292901992798, + 0.0024077941197901964, + 0.005376676097512245, + 0.0037380054127424955, + 0.002795363776385784, + 0.0074845547787845135, + 0.006386871915310621, + 0.010690081864595413, + 0.013371697627007961, + 0.007676613051444292, + 0.010733346454799175, + 0.013671842403709888, + 0.011286924593150616, + 0.004448098596185446, + 0.004224078264087439, + 0.0054021417163312435, + 0.004256404936313629, + 0.0028745110612362623, + 0.007884480059146881, + 0.002834012033417821, + 0.0049584354273974895, + 0.0033536499831825495, + 0.0033031038474291563, + 0.005492642521858215, + 0.0034587252885103226, + 0.0024411617778241634, + 0.002966636326164007, + 0.003933410160243511, + 0.0036800249945372343, + 0.009358920156955719, + 0.010876252315938473, + 0.06853941828012466, + 0.058907248079776764, + 0.015338778495788574, + 0.01266917958855629, + 0.009042499586939812, + 0.006968738045543432, + 0.004911419935524464, + 0.002857624553143978, + 0.004575290251523256, + 0.004532014485448599, + 0.003527865745127201, + 0.004216575995087624, + 0.002917526289820671, + 0.003049070481210947, + 0.026868462562561035, + 0.005278395488858223, + 0.035562414675951004, + 0.0050328937359154224, + 0.005378051195293665, + 0.004367187153548002, + 0.004412247333675623, + 0.0039434535428881645, + 0.01327518466860056, + 0.016413727775216103, + 0.008239776827394962, + 0.008477690629661083, + 0.012353073805570602, + 0.01734675094485283, + 0.008320561610162258, + 0.009538430720567703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.003612422849982977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle aerial shot, the camera slowly and gracefully orbits an ancient stone tower suspended in the twilight sky. A mythical dragon, its vast, sinuous form, circling the tower in majestic presence. Each scale, seemingly carved from sapphires and emeralds, emanates a steady, luminous bioluminescence from within, pulsing through the deep, scaly hues like breathing with every flex of its muscles. Its head, majestic and regal, bears translucent crystal horns, and gazes forth from golden eyes ablaze with ancient wisdom. The tower itself is constructed of weathered, dark gray stone, its surface covered in intricate carvings and mottled with moss. Its Gothic spire reaches up to a sky tinged with deep indigo and violet. The dragon wasn't flying at a swift speed, but rather gliding through the air with a near-weightless, leisurely grace. Its massive, membranous wings flapped only occasionally and slowly, each time silently stirring the surrounding mist. Its luminous form cast ever-shifting, shimmering reflections on the rough stone face of the ancient pagoda. The entire scene, serene and profound, exuded an awe-inspiring atmosphere of magical power and grandeur.", + "seed": 30571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_5/index_5673_seed_30571_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11432339213788509, + "vlm_logprob": -0.11432339213788509, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.11432339213788509, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16371554136276245 + ], + "opposite_logprob_tokens": [ + -1.9137156009674072 + ], + "logprob_reward": -0.16371554136276245, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16371554136276245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25142115354537964, + 0.04873748868703842, + 0.09108010679483414, + 0.06915125995874405, + 0.1575804203748703, + 0.03771404176950455, + 0.1738126277923584, + 0.04816513881087303, + 0.08345872908830643, + 0.03887900710105896 + ], + "credit_map": [ + 0.029444009065628052, + 0.027678614482283592, + 0.058244772255420685, + 0.03760450333356857, + 0.0218726247549057, + 0.024910058826208115, + 0.02256506122648716, + 0.02910149097442627, + 0.006058470346033573, + 0.004898902028799057, + 0.006415721029043198, + 0.004686776082962751, + 0.004801676142960787, + 0.00712410407140851, + 0.006489933468401432, + 0.008261905051767826, + 0.010166834108531475, + 0.017293699085712433, + 0.01288190484046936, + 0.013480376452207565, + 0.008928608149290085, + 0.00927214976400137, + 0.009127008728682995, + 0.009929527528584003, + 0.008789642713963985, + 0.009788228198885918, + 0.013426966033875942, + 0.009825496934354305, + 0.004771062172949314, + 0.006980448495596647, + 0.007528489921241999, + 0.008040927350521088, + 0.01770445518195629, + 0.028807172551751137, + 0.04269612953066826, + 0.02752450294792652, + 0.009941082447767258, + 0.011040788143873215, + 0.010218409821391106, + 0.009647897444665432, + 0.005289224907755852, + 0.004483566153794527, + 0.006676898337900639, + 0.00478703435510397, + 0.0032084716949611902, + 0.004918948281556368, + 0.0033978803548961878, + 0.0049520209431648254, + 0.020615261048078537, + 0.013927114196121693, + 0.022634366527199745, + 0.018284078687429428, + 0.020387114956974983, + 0.022349361330270767, + 0.029989944770932198, + 0.02562538906931877, + 0.005087404511868954, + 0.005461036693304777, + 0.006958255544304848, + 0.006487551145255566, + 0.006320860702544451, + 0.005707846023142338, + 0.006032105069607496, + 0.006110080052167177, + 0.010466604493558407, + 0.012226821854710579, + 0.008829538710415363, + 0.013212621212005615, + 0.01362809631973505, + 0.010718042962253094, + 0.006487250328063965, + 0.007889754138886929, + 0.004953821189701557, + 0.003956852946430445, + 0.005189741030335426, + 0.004028009250760078, + 0.004926231224089861, + 0.005685047246515751, + 0.004554391838610172, + 0.005584913771599531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.16371554136276245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306941613554955 + ], + "opposite_logprob_tokens": [ + -3.7733068466186523 + ], + "logprob_reward": -0.023306941613554955, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306941613554955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16332828998565674, + 0.04688871651887894, + 0.06527519226074219, + 0.050728973001241684, + 0.041512928903102875, + 0.10572651028633118, + 0.06554529815912247, + 0.12949128448963165, + 0.18908822536468506, + 0.14241458475589752 + ], + "credit_map": [ + 0.01236230880022049, + 0.017900682985782623, + 0.04316854104399681, + 0.020288005471229553, + 0.016923945397138596, + 0.015523356385529041, + 0.019347013905644417, + 0.017814436927437782, + 0.007604727055877447, + 0.006607487797737122, + 0.006050016265362501, + 0.0065001556649804115, + 0.003947047051042318, + 0.005162927322089672, + 0.005017753690481186, + 0.0059985993430018425, + 0.011429904028773308, + 0.016940578818321228, + 0.012655498459935188, + 0.008382235653698444, + 0.004110488574951887, + 0.00415674177929759, + 0.0030315008480101824, + 0.0045682466588914394, + 0.005806799978017807, + 0.008560573682188988, + 0.008485701866447926, + 0.007392409257590771, + 0.003673186991363764, + 0.005430690478533506, + 0.005958266090601683, + 0.00542134465649724, + 0.005342458840459585, + 0.0077545833773911, + 0.009983506985008717, + 0.007426874712109566, + 0.002723169978708029, + 0.0025946444366127253, + 0.003014235058799386, + 0.002673454349860549, + 0.012355403043329716, + 0.009159625507891178, + 0.01271307747811079, + 0.020162859931588173, + 0.005010032560676336, + 0.018684061244130135, + 0.007354893255978823, + 0.020286552608013153, + 0.008471659384667873, + 0.005783500149846077, + 0.007455381564795971, + 0.009280559606850147, + 0.005830467212945223, + 0.008004325442016125, + 0.010969070717692375, + 0.009750336408615112, + 0.017030630260705948, + 0.009747182950377464, + 0.02021609991788864, + 0.011884109117090702, + 0.011349901556968689, + 0.01588117517530918, + 0.02718711644411087, + 0.016195068135857582, + 0.020184975117444992, + 0.027621688321232796, + 0.021139299497008324, + 0.036365196108818054, + 0.031205635517835617, + 0.01486122701317072, + 0.023957455530762672, + 0.013752746395766735, + 0.017315305769443512, + 0.015477852895855904, + 0.022502174600958824, + 0.014264349825680256, + 0.015369368717074394, + 0.02368924207985401, + 0.013878395780920982, + 0.019917888566851616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.023306941613554955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904164284467697 + ], + "opposite_logprob_tokens": [ + -2.4640417098999023 + ], + "logprob_reward": -0.08904164284467697, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904164284467697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10316949337720871, + 0.020235316827893257, + 0.19153134524822235, + 0.03080051764845848, + 0.09615329653024673, + 0.17303241789340973, + 0.1672280728816986, + 0.04146214574575424, + 0.11995415389537811, + 0.05643320456147194 + ], + "credit_map": [ + 0.01508314348757267, + 0.01655704155564308, + 0.014858638867735863, + 0.014139006845653057, + 0.01310338731855154, + 0.007394193205982447, + 0.012270954437553883, + 0.009763132780790329, + 0.001899364055134356, + 0.002885730005800724, + 0.0025486978702247143, + 0.0025281840935349464, + 0.002085056621581316, + 0.0021142393816262484, + 0.0028515050653368235, + 0.0033225419465452433, + 0.02679325081408024, + 0.053307726979255676, + 0.028758086264133453, + 0.03214306756854057, + 0.011797013692557812, + 0.013103439472615719, + 0.01201754342764616, + 0.013611222617328167, + 0.0044651273638010025, + 0.0036823097616434097, + 0.004545470234006643, + 0.004198170732706785, + 0.0026717078872025013, + 0.0031948122195899487, + 0.00418914295732975, + 0.0038537776563316584, + 0.012933874502778053, + 0.01598062366247177, + 0.024350713938474655, + 0.02069084346294403, + 0.005033103749155998, + 0.0056460704654455185, + 0.006580809131264687, + 0.004937262274324894, + 0.03881204128265381, + 0.01271170936524868, + 0.01892857439815998, + 0.042205099016427994, + 0.008695798926055431, + 0.0202656127512455, + 0.009814583696424961, + 0.02159900777041912, + 0.01745588332414627, + 0.016739385202527046, + 0.019431356340646744, + 0.02169269695878029, + 0.016408788040280342, + 0.014612116850912571, + 0.04246235638856888, + 0.01842549443244934, + 0.004827003926038742, + 0.005470343865454197, + 0.0055334134958684444, + 0.004690270870923996, + 0.005903887562453747, + 0.004744510166347027, + 0.006992146838456392, + 0.003300571581348777, + 0.0159447118639946, + 0.022649474442005157, + 0.011917047202587128, + 0.01741122268140316, + 0.02140573039650917, + 0.010056506842374802, + 0.012594935484230518, + 0.007974525913596153, + 0.00957560259848833, + 0.005500490311533213, + 0.009075174108147621, + 0.0054934266954660416, + 0.007490376476198435, + 0.006916174665093422, + 0.007089328020811081, + 0.005292628891766071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.08904164284467697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813643515110016 + ], + "opposite_logprob_tokens": [ + -1.4063643217086792 + ], + "logprob_reward": -0.2813643515110016, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813643515110016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34280186891555786, + 0.04646237939596176, + 0.09131709486246109, + 0.044981129467487335, + 0.05313144996762276, + 0.13547919690608978, + 0.05631813779473305, + 0.07287896424531937, + 0.05779704451560974, + 0.09883270412683487 + ], + "credit_map": [ + 0.06850562244653702, + 0.05929965898394585, + 0.055686865001916885, + 0.04769274219870567, + 0.029056064784526825, + 0.027232885360717773, + 0.02697582356631756, + 0.02835220843553543, + 0.0046505737118422985, + 0.003914352040737867, + 0.008715156465768814, + 0.005297080613672733, + 0.005334206856787205, + 0.004821909125894308, + 0.006456704344600439, + 0.007272394374012947, + 0.01125708594918251, + 0.024801429361104965, + 0.01524833682924509, + 0.011653527617454529, + 0.0056685409508645535, + 0.007389766164124012, + 0.0051134186796844006, + 0.01018499955534935, + 0.003135313745588064, + 0.00486272107809782, + 0.007277281954884529, + 0.004842663649469614, + 0.006275705061852932, + 0.006347562186419964, + 0.006794310174882412, + 0.005445574410259724, + 0.004484367556869984, + 0.0032090279273688793, + 0.011263782158493996, + 0.017824115231633186, + 0.0026655506808310747, + 0.006832734681665897, + 0.002876478713005781, + 0.003975392784923315, + 0.011355690658092499, + 0.0070817731320858, + 0.00836268812417984, + 0.01964462362229824, + 0.008047646842896938, + 0.033952467143535614, + 0.009063699282705784, + 0.037970609962940216, + 0.009011154994368553, + 0.007108309771865606, + 0.007579362019896507, + 0.009490018710494041, + 0.005655297543853521, + 0.0037319567054510117, + 0.009073793888092041, + 0.0046682474203407764, + 0.008334869518876076, + 0.006920752581208944, + 0.011991615407168865, + 0.007516935002058744, + 0.006429475266486406, + 0.007623516023159027, + 0.016144175082445145, + 0.007917629554867744, + 0.006995203904807568, + 0.007385077886283398, + 0.007258099503815174, + 0.008155114017426968, + 0.011525355279445648, + 0.00523249339312315, + 0.005909226834774017, + 0.005336475558578968, + 0.013058683834969997, + 0.009756271727383137, + 0.01733635552227497, + 0.008645321242511272, + 0.008859214372932911, + 0.01803841069340706, + 0.00976602267473936, + 0.01337242592126131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.2813643515110016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle aerial view that orbits an ancient stone tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mythical dragon visible with scales appearing to be made of sapphires and emeralds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon glide slowly and leisurely with membranous wings that flap only occasionally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the stone tower constructed of weathered dark gray stone with intricate carvings and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinged with deep indigo and violet colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188483357429504 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188483357429504, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188483357429504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42643362283706665, + 0.03853820264339447, + 0.09569872915744781, + 0.27163583040237427, + 0.02727404423058033, + 0.012745480053126812, + 0.02779957465827465, + 0.031235642731189728, + 0.02771761082112789, + 0.04092131182551384 + ], + "credit_map": [ + 0.09117014706134796, + 0.07386817783117294, + 0.08597555011510849, + 0.07742215692996979, + 0.02355814166367054, + 0.02665986493229866, + 0.02170604281127453, + 0.026073504239320755, + 0.005764361936599016, + 0.005862243473529816, + 0.006555601954460144, + 0.007160588633269072, + 0.0027578638400882483, + 0.0030941397417336702, + 0.003439755644649267, + 0.003903642762452364, + 0.011413692496716976, + 0.025521868839859962, + 0.01459208782762289, + 0.017844462767243385, + 0.00675198296085, + 0.006065463647246361, + 0.005068641621619463, + 0.008440525270998478, + 0.03925418481230736, + 0.05477095767855644, + 0.04245532304048538, + 0.049514688551425934, + 0.01705888658761978, + 0.020916800945997238, + 0.024280168116092682, + 0.023384781554341316, + 0.002703312784433365, + 0.0026332722045481205, + 0.006763688754290342, + 0.008412155322730541, + 0.0011587004410102963, + 0.0022153686732053757, + 0.0015583985950797796, + 0.001829142915084958, + 0.0025365245528519154, + 0.0013952185399830341, + 0.0014290801482275128, + 0.002962912665680051, + 0.0010098506463691592, + 0.0011009523877874017, + 0.0011005047708749771, + 0.0012104338966310024, + 0.0029426889959722757, + 0.0020703282207250595, + 0.003967910539358854, + 0.0028580245561897755, + 0.0028109378181397915, + 0.0035361950285732746, + 0.005562730599194765, + 0.004050757735967636, + 0.0037043297197669744, + 0.0027958264108747244, + 0.004267295356839895, + 0.0028789693024009466, + 0.004088305402547121, + 0.004133857786655426, + 0.005765743087977171, + 0.003601313102990389, + 0.00363263301551342, + 0.004423854406923056, + 0.0022964882664382458, + 0.00438937870785594, + 0.004506061319261789, + 0.002995917107909918, + 0.002770473947748542, + 0.0027027996256947517, + 0.005347694735974073, + 0.004501940682530403, + 0.006550963502377272, + 0.003742465516552329, + 0.00465805409476161, + 0.006793614476919174, + 0.0037145819514989853, + 0.005611993372440338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.014188483357429504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_tensor([804, 804])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_tensor([804, 804])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..290476d15b00a8e0e1339b828f313819a598677e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_tensor([804, 804])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An unwavering, static wide-angle shot captures a dramatic scene rendered entirely in the iconic Ukiyo-e woodcut style. The mountainous landscape, its rolling forms defined by simple, elegant lines and a flattened perspective, is completely covered in thick, stylized kudzu foliage. Each leaf is a unique heart-shaped form, with a bold black outline and flat planes of soft green and earthy brown, creating a richly textured graphic tapestry across the mountain. From the edge of the frame, a stylized flame advances relentlessly; the flames are not real, but depicted as vibrant, flat orange and yellow shapes, each lick and curl defined by a thick, inky black outline, devouring the foliage with artistic ferocity. As the fire spreads, billowing black smoke rises into the pale ochre sky, revealing solid, swirling gray and black forms with identical outlines, resembling the dynamic, stylized waves of a classic print. The camera's unwavering gaze transforms the destructive events into a living, moving photograph, the fire's relentless advance and the rising plumes of smoke creating a mesmerizing and powerful atmosphere of controlled, beautiful disaster.", + "seed": 22852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_804_seed_22852_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030428836960345505, + "vlm_logprob": -0.030428836960345505, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.030428836960345505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01501414179801941 + ], + "opposite_logprob_tokens": [ + -4.265014171600342 + ], + "logprob_reward": -0.01501414179801941, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01501414179801941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5904838442802429, + 0.01940220408141613, + 0.01480718795210123, + 0.027201922610402107, + 0.02111741714179516, + 0.015263844281435013, + 0.05608784407377243, + 0.014094275422394276, + 0.040525633841753006, + 0.20101585984230042 + ], + "credit_map": [ + 0.04977850615978241, + 0.042763493955135345, + 0.10931885242462158, + 0.15581148862838745, + 0.037818606942892075, + 0.048292264342308044, + 0.08258030563592911, + 0.06412029266357422, + 0.0016597151989117265, + 0.0017637875862419605, + 0.0020220023579895496, + 0.0023730620741844177, + 0.002895749406889081, + 0.002687501022592187, + 0.0020997109822928905, + 0.0039006753358989954, + 0.001990846125409007, + 0.002519288333132863, + 0.0012703437823802233, + 0.0025859251618385315, + 0.0018484315369278193, + 0.0010169451124966145, + 0.0016733204247429967, + 0.0019020885229110718, + 0.0029375371523201466, + 0.0029949783347547054, + 0.004315220285207033, + 0.003398523898795247, + 0.0023340131156146526, + 0.002682601334527135, + 0.0032626076135784388, + 0.005276442039757967, + 0.0018726844573393464, + 0.0019647434819489717, + 0.0026996745727956295, + 0.0025866844225674868, + 0.003476604586467147, + 0.0021620499901473522, + 0.004023159854114056, + 0.0023318149615079165, + 0.0013813595287501812, + 0.0015418734401464462, + 0.0015371779445558786, + 0.0019131331937387586, + 0.0016997933853417635, + 0.001255848677828908, + 0.002260589739307761, + 0.003674069419503212, + 0.004150785971432924, + 0.005069664213806391, + 0.004774753469973803, + 0.003424336202442646, + 0.01216750405728817, + 0.004852615762501955, + 0.005351593717932701, + 0.016296591609716415, + 0.0014097719686105847, + 0.0017959561664611101, + 0.0020370695274323225, + 0.0022728480398654938, + 0.001715976744890213, + 0.001662389375269413, + 0.0015470260987058282, + 0.001653237035498023, + 0.0023572545032948256, + 0.0022072114516049623, + 0.0028385641053318977, + 0.014327275566756725, + 0.004969243425875902, + 0.0030567969661206007, + 0.0033480811398476362, + 0.007421205751597881, + 0.01301503460854292, + 0.012426398694515228, + 0.016563784331083298, + 0.08327724039554596, + 0.01498789619654417, + 0.006986773107200861, + 0.015291460789740086, + 0.038467273116111755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.01501414179801941 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864796806126833 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.00864796806126833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864796806126833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6286638379096985, + 0.016099564731121063, + 0.11415236443281174, + 0.016738861799240112, + 0.024780970066785812, + 0.013245676644146442, + 0.04398077726364136, + 0.012811752036213875, + 0.08391877263784409, + 0.04560747370123863 + ], + "credit_map": [ + 0.052259914577007294, + 0.04071562737226486, + 0.11758800595998764, + 0.1661411076784134, + 0.038948800414800644, + 0.050208866596221924, + 0.09087477624416351, + 0.07192670553922653, + 0.0017539294203743339, + 0.0015440807910636067, + 0.0023675919510424137, + 0.0026535114739090204, + 0.0020963356364518404, + 0.0015969944652169943, + 0.0022478660102933645, + 0.0018392563797533512, + 0.016326826065778732, + 0.015947120264172554, + 0.008618753403425217, + 0.009615158662199974, + 0.02035594917833805, + 0.022751882672309875, + 0.008356140926480293, + 0.01218054536730051, + 0.0016539104981347919, + 0.0016488423570990562, + 0.004156097769737244, + 0.001968222204595804, + 0.001507675158791244, + 0.001191148767247796, + 0.0026181135326623917, + 0.001994852442294359, + 0.0019124501850456, + 0.002584339352324605, + 0.0031047160737216473, + 0.0032658290583640337, + 0.004127176944166422, + 0.001691871671937406, + 0.005250067450106144, + 0.0028445201460272074, + 0.0019099147757515311, + 0.0014193095266819, + 0.0012367775198072195, + 0.0014451112365350127, + 0.0013157854555174708, + 0.0011806944385170937, + 0.0017229599179700017, + 0.0030151247046887875, + 0.0034734720829874277, + 0.00404810905456543, + 0.003620615229010582, + 0.0022405770141631365, + 0.010935958474874496, + 0.0035100537352263927, + 0.0038828235119581223, + 0.01226917002350092, + 0.0013531957520172, + 0.0016397318104282022, + 0.0023472164757549763, + 0.0016776564298197627, + 0.0013091492000967264, + 0.001335485023446381, + 0.0019318010890856385, + 0.0012175175361335278, + 0.005887117702513933, + 0.006245445925742388, + 0.004957421682775021, + 0.024683581665158272, + 0.009805379435420036, + 0.009983045049011707, + 0.006172160618007183, + 0.016184618696570396, + 0.003165879985317588, + 0.0038579164538532495, + 0.005086622200906277, + 0.009519054554402828, + 0.008957183919847012, + 0.0025187027640640736, + 0.004628149792551994, + 0.00787396915256977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.00864796806126833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43727150559425354, + 0.08555497229099274, + 0.0543438196182251, + 0.08130321651697159, + 0.026848234236240387, + 0.08190225064754486, + 0.12670035660266876, + 0.01687857322394848, + 0.026856696233153343, + 0.0623403824865818 + ], + "credit_map": [ + 0.03637126833200455, + 0.028406651690602303, + 0.07260863482952118, + 0.10407373309135437, + 0.034569595009088516, + 0.04025863856077194, + 0.06440002471208572, + 0.05658295750617981, + 0.009521338157355785, + 0.009107251651585102, + 0.012467208318412304, + 0.01197343785315752, + 0.00903515424579382, + 0.013084670528769493, + 0.0101819122210145, + 0.010183999314904213, + 0.006176943890750408, + 0.008297384716570377, + 0.00439932057633996, + 0.006419547833502293, + 0.00985932257026434, + 0.009451944380998611, + 0.0040856581181287766, + 0.005653697531670332, + 0.0068811578676104546, + 0.006346502341330051, + 0.01706952415406704, + 0.011122087016701698, + 0.006394712720066309, + 0.00957737397402525, + 0.010641505941748619, + 0.013270355761051178, + 0.0025037128943949938, + 0.003005614969879389, + 0.004794215317815542, + 0.003064111340790987, + 0.003123397007584572, + 0.002862275345250964, + 0.005027989856898785, + 0.00246691657230258, + 0.005234959535300732, + 0.004970767069607973, + 0.009400710463523865, + 0.02896135114133358, + 0.019645987078547478, + 0.0035494535695761442, + 0.006670387927442789, + 0.003468634793534875, + 0.007881652563810349, + 0.010486130602657795, + 0.010935572907328606, + 0.006698476616293192, + 0.029499737545847893, + 0.008295564912259579, + 0.012555783614516258, + 0.04034743830561638, + 0.0017630872316658497, + 0.0028589689172804356, + 0.003274129703640938, + 0.00195085140876472, + 0.002163379220291972, + 0.001236100448295474, + 0.00224490137770772, + 0.0013871551491320133, + 0.00148703099694103, + 0.0022341234143823385, + 0.002347675384953618, + 0.008765311911702156, + 0.0015770000172778964, + 0.0018657709006220102, + 0.0021570834796875715, + 0.006422699894756079, + 0.006054735742509365, + 0.005843809340149164, + 0.00836087390780449, + 0.01271688099950552, + 0.008829808793962002, + 0.004946338478475809, + 0.008291810750961304, + 0.007296125870198011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057156339287758 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.02057156339287758, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057156339287758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6077335476875305, + 0.03055238537490368, + 0.042814891785383224, + 0.058070939034223557, + 0.024484187364578247, + 0.13146890699863434, + 0.027911582961678505, + 0.00508250342682004, + 0.05360518395900726, + 0.018275851383805275 + ], + "credit_map": [ + 0.050062522292137146, + 0.03952114284038544, + 0.1065480038523674, + 0.1530386507511139, + 0.041012492030858994, + 0.05251810699701309, + 0.09123757481575012, + 0.07379504293203354, + 0.002842961112037301, + 0.003269340144470334, + 0.003916097804903984, + 0.004517111927270889, + 0.004356446210294962, + 0.0031500658951699734, + 0.003026020945981145, + 0.005474339239299297, + 0.0053043002262711525, + 0.006260289344936609, + 0.0031112206634134054, + 0.004899112973362207, + 0.007456551771610975, + 0.007717011496424675, + 0.00306705292314291, + 0.004999352153390646, + 0.0056155030615627766, + 0.0048865508288145065, + 0.00890179630368948, + 0.007399354130029678, + 0.0054229008965194225, + 0.006958328653126955, + 0.006804476026445627, + 0.012082023546099663, + 0.0029436431359499693, + 0.0026582046411931515, + 0.002353219548240304, + 0.002564318012446165, + 0.0050942059606313705, + 0.002283062320202589, + 0.004082069266587496, + 0.0025054635480046272, + 0.008665950037539005, + 0.007523112930357456, + 0.015376347117125988, + 0.04650171473622322, + 0.033818867057561874, + 0.0055092410184443, + 0.008578257635235786, + 0.005495405290275812, + 0.0025637305807322264, + 0.0026545580476522446, + 0.0014824233949184418, + 0.0019031729316338897, + 0.007134525571018457, + 0.002718671690672636, + 0.0021168403327465057, + 0.007337659131735563, + 0.00044373926357366145, + 0.0007682246505282819, + 0.0007143566035665572, + 0.0006917130667716265, + 0.0006339853280223906, + 0.0005350902210921049, + 0.0005510904011316597, + 0.0007443039212375879, + 0.003316895104944706, + 0.0031479173339903355, + 0.0037407316267490387, + 0.017781082540750504, + 0.004798735026270151, + 0.005047556944191456, + 0.004450752399861813, + 0.01132151111960411, + 0.0023232230450958014, + 0.001955627929419279, + 0.0016102419467642903, + 0.0013379966840147972, + 0.004274955950677395, + 0.0016433941200375557, + 0.0017060007667168975, + 0.0034244097769260406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.02057156339287758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002800241112709 + ], + "opposite_logprob_tokens": [ + -2.3502800464630127 + ], + "logprob_reward": -0.1002800241112709, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002800241112709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32043391466140747, + 0.09099871665239334, + 0.09349290281534195, + 0.060464564710855484, + 0.05347737669944763, + 0.03364742547273636, + 0.09059524536132812, + 0.13859151303768158, + 0.09698370844125748, + 0.021314578130841255 + ], + "credit_map": [ + 0.028964731842279434, + 0.021782681345939636, + 0.0727660059928894, + 0.09883613884449005, + 0.016934113577008247, + 0.015347760170698166, + 0.03340235725045204, + 0.03240012750029564, + 0.008290973491966724, + 0.009269706904888153, + 0.010021070018410683, + 0.013411885127425194, + 0.014594140462577343, + 0.009570598602294922, + 0.007428792770951986, + 0.018411554396152496, + 0.012388118542730808, + 0.012811688706278801, + 0.00721884798258543, + 0.009458895772695541, + 0.015327196568250656, + 0.018698925152420998, + 0.008027341216802597, + 0.009561889804899693, + 0.006155718583613634, + 0.005826258100569248, + 0.006712336093187332, + 0.008824119344353676, + 0.005740141961723566, + 0.006150851957499981, + 0.006744010373950005, + 0.014311124570667744, + 0.004797080997377634, + 0.009624306112527847, + 0.006614910904318094, + 0.007814588025212288, + 0.007091077975928783, + 0.005880117882043123, + 0.006550000049173832, + 0.0051052928902208805, + 0.002802799455821514, + 0.0026144145522266626, + 0.0034611420705914497, + 0.006788875907659531, + 0.006552493665367365, + 0.0029387848917394876, + 0.0030880281701684, + 0.005400889553129673, + 0.005645463243126869, + 0.007058179005980492, + 0.0066887675784528255, + 0.006674795411527157, + 0.021091081202030182, + 0.00607534172013402, + 0.007914142683148384, + 0.02944747731089592, + 0.014643888920545578, + 0.023683037608861923, + 0.026650575920939445, + 0.016526170074939728, + 0.021004807204008102, + 0.01070715207606554, + 0.012261825613677502, + 0.013114044442772865, + 0.0052786776795983315, + 0.004941892810165882, + 0.007039097603410482, + 0.03919472545385361, + 0.004794130101799965, + 0.003101460635662079, + 0.007602272555232048, + 0.025031454861164093, + 0.0021841132547706366, + 0.002497752197086811, + 0.002158666728064418, + 0.002475476823747158, + 0.0038681766018271446, + 0.0029871074948459864, + 0.00219868216663599, + 0.0029446042608469725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.1002800241112709 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unwavering, static wide-angle shot captures a dramatic scene rendered entirely in the iconic Ukiyo-e woodcut style. The mountainous landscape, its rolling forms defined by simple, elegant lines and a flattened perspective, is completely covered in thick, stylized kudzu foliage. Each leaf is a unique heart-shaped form, with a bold black outline and flat planes of soft green and earthy brown, creating a richly textured graphic tapestry across the mountain. From the edge of the frame, a stylized flame advances relentlessly; the flames are not real, but depicted as vibrant, flat orange and yellow shapes, each lick and curl defined by a thick, inky black outline, devouring the foliage with artistic ferocity. As the fire spreads, billowing black smoke rises into the pale ochre sky, revealing solid, swirling gray and black forms with identical outlines, resembling the dynamic, stylized waves of a classic print. The camera's unwavering gaze transforms the destructive events into a living, moving photograph, the fire's relentless advance and the rising plumes of smoke creating a mesmerizing and powerful atmosphere of controlled, beautiful disaster.", + "seed": 22853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_804_seed_22853_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024873434333130717, + "vlm_logprob": -0.024873434333130717, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.024873434333130717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002273952355608344 + ], + "opposite_logprob_tokens": [ + -6.377274036407471 + ], + "logprob_reward": -0.002273952355608344, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.002273952355608344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13040241599082947, + 0.17905068397521973, + 0.0461537167429924, + 0.18850596249103546, + 0.06481029093265533, + 0.08288197219371796, + 0.03882884979248047, + 0.0815897211432457, + 0.1310846358537674, + 0.056691717356443405 + ], + "credit_map": [ + 0.018150705844163895, + 0.016350746154785156, + 0.01648201048374176, + 0.01579805091023445, + 0.010832675732672215, + 0.019275812432169914, + 0.01641657203435898, + 0.01709585264325142, + 0.03075150027871132, + 0.033325377851724625, + 0.019027451053261757, + 0.019254034385085106, + 0.018334150314331055, + 0.02175326645374298, + 0.018254099413752556, + 0.018350817263126373, + 0.003424488240852952, + 0.00472654402256012, + 0.005452282726764679, + 0.01469944417476654, + 0.0041373111307621, + 0.0028544871602207422, + 0.004165998660027981, + 0.006693162024021149, + 0.0122977951541543, + 0.009299548342823982, + 0.07426963001489639, + 0.054469503462314606, + 0.006159254349768162, + 0.008196900598704815, + 0.01126175932586193, + 0.012551581487059593, + 0.008315998129546642, + 0.00678634736686945, + 0.007638096809387207, + 0.0074269711039960384, + 0.010387982241809368, + 0.0074668810702860355, + 0.00802350603044033, + 0.008764516562223434, + 0.008135844022035599, + 0.0104364650323987, + 0.008381572552025318, + 0.01131274364888668, + 0.00964711420238018, + 0.009170661680400372, + 0.014034166000783443, + 0.011763402260839939, + 0.0040904683992266655, + 0.004084807354956865, + 0.004457661882042885, + 0.004097544122487307, + 0.004664110951125622, + 0.003661922411993146, + 0.004061941523104906, + 0.009710394777357578, + 0.006372183095663786, + 0.007059409748762846, + 0.004095384385436773, + 0.00816292129456997, + 0.014143838547170162, + 0.00694232527166605, + 0.022975536063313484, + 0.011838124133646488, + 0.01219220831990242, + 0.01213893387466669, + 0.017719393596053123, + 0.019090881571173668, + 0.010247805155813694, + 0.010598457418382168, + 0.018159955739974976, + 0.03093700483441353, + 0.005641149822622538, + 0.004637003410607576, + 0.016772156581282616, + 0.00478089926764369, + 0.009854596108198166, + 0.004103500861674547, + 0.004259401001036167, + 0.006643018685281277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.002273952355608344 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955095402896404 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955095402896404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955095402896404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18406064808368683, + 0.11475606262683868, + 0.024148840457201004, + 0.125517800450325, + 0.18232129514217377, + 0.12943294644355774, + 0.11923731118440628, + 0.05711920186877251, + 0.044497858732938766, + 0.018908068537712097 + ], + "credit_map": [ + 0.021296441555023193, + 0.03071802295744419, + 0.028904490172863007, + 0.020665181800723076, + 0.0200266782194376, + 0.025094065815210342, + 0.01717541366815567, + 0.02018035016953945, + 0.020499370992183685, + 0.011104839853942394, + 0.009246673434972763, + 0.010566670447587967, + 0.008766237646341324, + 0.013396009802818298, + 0.026612721383571625, + 0.014563530683517456, + 0.0024724500253796577, + 0.0031293670181185007, + 0.004368029534816742, + 0.0035828188993036747, + 0.002717254450544715, + 0.002204430988058448, + 0.002461513038724661, + 0.0032129758037626743, + 0.009642304852604866, + 0.008714447729289532, + 0.04507400095462799, + 0.03357888013124466, + 0.006101685576140881, + 0.006378354504704475, + 0.006833542138338089, + 0.00919459480792284, + 0.00917681585997343, + 0.00856108870357275, + 0.020147105678915977, + 0.053052306175231934, + 0.01121357548981905, + 0.008657035417854786, + 0.02816213294863701, + 0.04335123300552368, + 0.01104333158582449, + 0.008455718867480755, + 0.013012923300266266, + 0.018604636192321777, + 0.009462797082960606, + 0.01853920891880989, + 0.03116818144917488, + 0.0191461481153965, + 0.01344083622097969, + 0.013364508748054504, + 0.014590099453926086, + 0.012125568464398384, + 0.01793851889669895, + 0.01204909197986126, + 0.01165823545306921, + 0.02407045289874077, + 0.0028973722364753485, + 0.0026046009734272957, + 0.0029477509669959545, + 0.012322384864091873, + 0.0026745758950710297, + 0.0022926554083824158, + 0.021934788674116135, + 0.009445073083043098, + 0.0046021584421396255, + 0.004838734399527311, + 0.007908645085990429, + 0.007033693138509989, + 0.003672589547932148, + 0.003956545609980822, + 0.004595401231199503, + 0.00789009127765894, + 0.0021657794713974, + 0.0024311088491231203, + 0.002118217758834362, + 0.0015807666350156069, + 0.002533920109272003, + 0.0028177837375551462, + 0.0019286666065454483, + 0.003331824205815792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.005955095402896404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606731116771698 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606731116771698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606731116771698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08264026790857315, + 0.10802111774682999, + 0.03905565291643143, + 0.1743142306804657, + 0.1410830020904541, + 0.2767515778541565, + 0.028996456414461136, + 0.053909141570329666, + 0.023820405825972557, + 0.07140808552503586 + ], + "credit_map": [ + 0.011124607175588608, + 0.015423323027789593, + 0.00929633155465126, + 0.008857840672135353, + 0.008921528235077858, + 0.011391527950763702, + 0.00970933586359024, + 0.007915783673524857, + 0.02094966173171997, + 0.02007218264043331, + 0.012147270143032074, + 0.008360904641449451, + 0.009818896651268005, + 0.011724095791578293, + 0.014662813395261765, + 0.010285308584570885, + 0.0054358565248548985, + 0.0035414095036685467, + 0.005218155216425657, + 0.0059000845067203045, + 0.005341173615306616, + 0.0037135600578039885, + 0.005370726343244314, + 0.004534694366157055, + 0.012798678129911423, + 0.010180598124861717, + 0.061018429696559906, + 0.03696718439459801, + 0.013858381658792496, + 0.010631256736814976, + 0.016643958166241646, + 0.012215769849717617, + 0.008917618542909622, + 0.01073423307389021, + 0.015649281442165375, + 0.03528645634651184, + 0.013551919721066952, + 0.009071065112948418, + 0.022608540952205658, + 0.02526390179991722, + 0.01622427999973297, + 0.014820078387856483, + 0.04535004869103432, + 0.03634108975529671, + 0.018899602815508842, + 0.020939603447914124, + 0.0639512911438942, + 0.060225628316402435, + 0.0033392615150660276, + 0.003539284458383918, + 0.00544073386117816, + 0.0034949269611388445, + 0.003080124268308282, + 0.002513237064704299, + 0.0033607420045882463, + 0.00422814954072237, + 0.003099885070696473, + 0.005075407214462757, + 0.006874681916087866, + 0.011217205785214901, + 0.0051255254074931145, + 0.0033644340001046658, + 0.014230195432901382, + 0.0049218107014894485, + 0.002650987124070525, + 0.0017310390248894691, + 0.002807105891406536, + 0.002526909811422229, + 0.002565797185525298, + 0.002540386514738202, + 0.004313607234507799, + 0.004684577230364084, + 0.006075337994843721, + 0.005209163296967745, + 0.012838669121265411, + 0.006278443615883589, + 0.01309544499963522, + 0.00966597255319357, + 0.006946195848286152, + 0.01129886507987976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.01606731116771698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069082655012608 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069082655012608, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069082655012608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12493918836116791, + 0.13341890275478363, + 0.12521743774414062, + 0.21153125166893005, + 0.05780981853604317, + 0.09955696016550064, + 0.0765799880027771, + 0.06303726881742477, + 0.019474884495139122, + 0.08843432366847992 + ], + "credit_map": [ + 0.01041521318256855, + 0.01584235206246376, + 0.019415808841586113, + 0.020410867407917976, + 0.010803247801959515, + 0.009836130775511265, + 0.025220856070518494, + 0.012994717806577682, + 0.02232440561056137, + 0.008297965861856937, + 0.013194913975894451, + 0.013482616282999516, + 0.010163280181586742, + 0.01011184323579073, + 0.04086422175168991, + 0.014979643747210503, + 0.015529357828199863, + 0.007284887135028839, + 0.021148039028048515, + 0.04778164625167847, + 0.0048166015185415745, + 0.006341993343085051, + 0.010439160279929638, + 0.01187574677169323, + 0.01056924369186163, + 0.006993185728788376, + 0.08991856873035431, + 0.06263048946857452, + 0.008394870907068253, + 0.006972386967390776, + 0.015120926313102245, + 0.010931563563644886, + 0.006822613999247551, + 0.005084628704935312, + 0.005882779601961374, + 0.008715491741895676, + 0.005643312819302082, + 0.008078454062342644, + 0.007703593000769615, + 0.009878944605588913, + 0.004268142860382795, + 0.007757398299872875, + 0.008518580347299576, + 0.014546981081366539, + 0.0048731123097240925, + 0.004847617354243994, + 0.033792488276958466, + 0.02095264010131359, + 0.007592586800456047, + 0.006612845696508884, + 0.010897502303123474, + 0.009357575327157974, + 0.010484439320862293, + 0.006712432950735092, + 0.006571609526872635, + 0.01835099793970585, + 0.005204637534916401, + 0.003974586259573698, + 0.004624584224075079, + 0.007055035792291164, + 0.011321456171572208, + 0.010037493892014027, + 0.015628375113010406, + 0.0051911016926169395, + 0.002070765011012554, + 0.002269004937261343, + 0.0024944248143583536, + 0.002466903766617179, + 0.0019047147361561656, + 0.002377103315666318, + 0.002355910139158368, + 0.0035360571928322315, + 0.007368636317551136, + 0.0044176229275763035, + 0.028705265372991562, + 0.007431988604366779, + 0.017336474731564522, + 0.006641769781708717, + 0.007533449679613113, + 0.008999118581414223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.011069082655012608 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900173008441925 + ], + "opposite_logprob_tokens": [ + -2.4640016555786133 + ], + "logprob_reward": -0.08900173008441925, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900173008441925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14766117930412292, + 0.34708189964294434, + 0.04853740707039833, + 0.07099118083715439, + 0.08309760689735413, + 0.06939700990915298, + 0.02852487564086914, + 0.032341741025447845, + 0.017936332151293755, + 0.1544308215379715 + ], + "credit_map": [ + 0.014044888317584991, + 0.01921016164124012, + 0.023770786821842194, + 0.01923002488911152, + 0.011795527301728725, + 0.0199778750538826, + 0.019075307995080948, + 0.020556610077619553, + 0.040995966643095016, + 0.04546748101711273, + 0.024693768471479416, + 0.03763769939541817, + 0.028563998639583588, + 0.024947144091129303, + 0.10748288780450821, + 0.037292953580617905, + 0.005052622873336077, + 0.012470354326069355, + 0.004063161555677652, + 0.0038632552605122328, + 0.011147040873765945, + 0.004793948028236628, + 0.003418586915358901, + 0.003728439100086689, + 0.0051332260482013226, + 0.0052815829403698444, + 0.01827923208475113, + 0.019349591806530952, + 0.005158029496669769, + 0.005065776873379946, + 0.005581781733781099, + 0.007141954265534878, + 0.004391962196677923, + 0.005160799715667963, + 0.00886306632310152, + 0.021308770403265953, + 0.00889144092798233, + 0.0048376345075666904, + 0.012732420116662979, + 0.01691151224076748, + 0.004471419844776392, + 0.010120859369635582, + 0.004498107358813286, + 0.006441275589168072, + 0.008955797180533409, + 0.004163856618106365, + 0.016102802008390427, + 0.014642895199358463, + 0.003360038623213768, + 0.00377101032063365, + 0.0034018666483461857, + 0.003152201185002923, + 0.003705766284838319, + 0.00438581733033061, + 0.0023407635744661093, + 0.004407412372529507, + 0.002023420762270689, + 0.0029824385419487953, + 0.004206860903650522, + 0.0067381481640040874, + 0.003198903752490878, + 0.0024086404591798782, + 0.005556115880608559, + 0.00522721279412508, + 0.0018195647280663252, + 0.0018507263157516718, + 0.002751298248767853, + 0.0020275237038731575, + 0.001835329458117485, + 0.002089784014970064, + 0.002645473927259445, + 0.002916631055995822, + 0.01046007964760065, + 0.008087157271802425, + 0.05391260236501694, + 0.01408461481332779, + 0.02464456483721733, + 0.012644167989492416, + 0.012460259720683098, + 0.018137387931346893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.08900173008441925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unwavering, static wide-angle shot captures a dramatic scene rendered entirely in the iconic Ukiyo-e woodcut style. The mountainous landscape, its rolling forms defined by simple, elegant lines and a flattened perspective, is completely covered in thick, stylized kudzu foliage. Each leaf is a unique heart-shaped form, with a bold black outline and flat planes of soft green and earthy brown, creating a richly textured graphic tapestry across the mountain. From the edge of the frame, a stylized flame advances relentlessly; the flames are not real, but depicted as vibrant, flat orange and yellow shapes, each lick and curl defined by a thick, inky black outline, devouring the foliage with artistic ferocity. As the fire spreads, billowing black smoke rises into the pale ochre sky, revealing solid, swirling gray and black forms with identical outlines, resembling the dynamic, stylized waves of a classic print. The camera's unwavering gaze transforms the destructive events into a living, moving photograph, the fire's relentless advance and the rising plumes of smoke creating a mesmerizing and powerful atmosphere of controlled, beautiful disaster.", + "seed": 22854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_804_seed_22854_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018616931326687337, + "vlm_logprob": -0.018616931326687337, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.018616931326687337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004054422490298748 + ], + "opposite_logprob_tokens": [ + -5.629054546356201 + ], + "logprob_reward": -0.004054422490298748, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.004054422490298748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1269581913948059, + 0.046283021569252014, + 0.04390972480177879, + 0.06157638877630234, + 0.4879107177257538, + 0.07681354880332947, + 0.05154424160718918, + 0.017858531326055527, + 0.024904591962695122, + 0.06224101409316063 + ], + "credit_map": [ + 0.035305097699165344, + 0.010627292096614838, + 0.010496314615011215, + 0.009765387512743473, + 0.009997946210205555, + 0.013404669240117073, + 0.008946729823946953, + 0.02841474860906601, + 0.006464606616646051, + 0.005970129277557135, + 0.00407859031111002, + 0.004333231598138809, + 0.005235517863184214, + 0.010698338970541954, + 0.003386203432455659, + 0.006116404198110104, + 0.00769072026014328, + 0.004375574644654989, + 0.00695808557793498, + 0.003964538220316172, + 0.00467320391908288, + 0.006324748508632183, + 0.005401413422077894, + 0.004521443508565426, + 0.008604271337389946, + 0.007203142624348402, + 0.007370712235569954, + 0.008144855499267578, + 0.006552916951477528, + 0.005247000139206648, + 0.0066038160584867, + 0.011849671602249146, + 0.11860128492116928, + 0.032380327582359314, + 0.04190868139266968, + 0.044148094952106476, + 0.017345061525702477, + 0.0323924794793129, + 0.04180227220058441, + 0.1593325138092041, + 0.01662362553179264, + 0.026397565379738808, + 0.006276692729443312, + 0.004274709615856409, + 0.008978856727480888, + 0.005458017811179161, + 0.004885713569819927, + 0.003918368369340897, + 0.008661671541631222, + 0.005618060007691383, + 0.0033933536615222692, + 0.002541485009714961, + 0.004495435394346714, + 0.017911789938807487, + 0.005738125182688236, + 0.003184322267770767, + 0.0016612507170066237, + 0.0018349129240959883, + 0.005164541769772768, + 0.0012645575916394591, + 0.0015419229166582227, + 0.0016674600774422288, + 0.001520114135928452, + 0.003203771309927106, + 0.0031570754945278168, + 0.0016372454119846225, + 0.002627458656206727, + 0.0060771736316382885, + 0.0022714370861649513, + 0.002223614137619734, + 0.0029322667978703976, + 0.003978322260081768, + 0.014259614050388336, + 0.0036477893590927124, + 0.004942368250340223, + 0.004454006440937519, + 0.004192044958472252, + 0.003195270663127303, + 0.019290773198008537, + 0.008259148336946964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.004054422490298748 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646813031286001 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646813031286001, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646813031286001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1775805503129959, + 0.0821615681052208, + 0.057393960654735565, + 0.152817964553833, + 0.10429081320762634, + 0.06119494140148163, + 0.2019401490688324, + 0.041896600276231766, + 0.08470036834478378, + 0.03602319210767746 + ], + "credit_map": [ + 0.048548586666584015, + 0.01309717632830143, + 0.015390767715871334, + 0.01556658186018467, + 0.010754148475825787, + 0.012415644712746143, + 0.013922407291829586, + 0.04788520559668541, + 0.01020700391381979, + 0.013544965535402298, + 0.006914905738085508, + 0.006313086021691561, + 0.010130963288247585, + 0.01864314079284668, + 0.006007891148328781, + 0.010399598628282547, + 0.006081348285079002, + 0.005741438828408718, + 0.00964826438575983, + 0.0056260572746396065, + 0.006016404367983341, + 0.008866981603205204, + 0.008492917753756046, + 0.006920539308339357, + 0.020483210682868958, + 0.013910477980971336, + 0.011804192326962948, + 0.012270571663975716, + 0.01563531532883644, + 0.011724209412932396, + 0.02485146000981331, + 0.0421384759247303, + 0.0268919188529253, + 0.006629191339015961, + 0.008210723288357258, + 0.009867039509117603, + 0.004143470898270607, + 0.008921617642045021, + 0.008642748929560184, + 0.03098408505320549, + 0.014148684218525887, + 0.019035885110497475, + 0.003874767804518342, + 0.0037260870449244976, + 0.0071200174279510975, + 0.004872563760727644, + 0.004514558240771294, + 0.0039023649878799915, + 0.027739521116018295, + 0.023606469854712486, + 0.013530476950109005, + 0.009676262736320496, + 0.018666395917534828, + 0.07427558302879333, + 0.02325722947716713, + 0.011188162490725517, + 0.004134754184633493, + 0.005244490224868059, + 0.008225700818002224, + 0.005503135733306408, + 0.003616634290665388, + 0.0028570948634296656, + 0.004829403944313526, + 0.007485376205295324, + 0.010773862712085247, + 0.00913245789706707, + 0.006719326600432396, + 0.023012474179267883, + 0.008197004906833172, + 0.011171039193868637, + 0.00826174020767212, + 0.007432451006025076, + 0.00794197991490364, + 0.0026658501010388136, + 0.0035650180652737617, + 0.002901925938203931, + 0.002137022325769067, + 0.0022458930034190416, + 0.009740824811160564, + 0.004824669566005468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.004646813031286001 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953317973762751 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953317973762751, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953317973762751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1205177828669548, + 0.06768056005239487, + 0.2473081350326538, + 0.031708694994449615, + 0.09396810084581375, + 0.03311034291982651, + 0.2877386510372162, + 0.014962996356189251, + 0.03186717629432678, + 0.07113759219646454 + ], + "credit_map": [ + 0.027986297383904457, + 0.012372458353638649, + 0.010773006826639175, + 0.012697640806436539, + 0.008361306972801685, + 0.008427975699305534, + 0.011190852150321007, + 0.028708243742585182, + 0.008326071314513683, + 0.005508346948772669, + 0.00747198611497879, + 0.010610956698656082, + 0.005057002417743206, + 0.006167304702103138, + 0.009270435199141502, + 0.015268459916114807, + 0.03510383144021034, + 0.02152758464217186, + 0.026618583127856255, + 0.026712724938988686, + 0.021774129942059517, + 0.02080100029706955, + 0.050499167293310165, + 0.04427110031247139, + 0.0032810652628540993, + 0.002550588920712471, + 0.004087039269506931, + 0.006436132360249758, + 0.003306017257273197, + 0.00343054230324924, + 0.004996701143682003, + 0.0036206080112606287, + 0.019533580169081688, + 0.005589307751506567, + 0.010025584138929844, + 0.007971720770001411, + 0.0035645938478410244, + 0.0058919950388371944, + 0.008612863719463348, + 0.03277845308184624, + 0.008152641355991364, + 0.006779232993721962, + 0.0030630144756287336, + 0.0027124369516968727, + 0.0025870304089039564, + 0.002392844296991825, + 0.004548271186649799, + 0.0028748714830726385, + 0.03141266480088234, + 0.03213705122470856, + 0.02095760405063629, + 0.015077807009220123, + 0.026471424847841263, + 0.09050719439983368, + 0.051141880452632904, + 0.020033027976751328, + 0.0017443232936784625, + 0.0012915813131257892, + 0.0027336589992046356, + 0.0016982581000775099, + 0.0015156315639615059, + 0.0014073796337470412, + 0.0016270705964416265, + 0.002945092972368002, + 0.00427980674430728, + 0.002726014005020261, + 0.002899269573390484, + 0.007183436304330826, + 0.0027930994983762503, + 0.0026107728481292725, + 0.003984028473496437, + 0.005390746984630823, + 0.013158267363905907, + 0.0037129882257431746, + 0.006790475454181433, + 0.0058598592877388, + 0.004686964210122824, + 0.004651903174817562, + 0.022261328995227814, + 0.010015804320573807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.005953317973762751 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795082286000252 + ], + "opposite_logprob_tokens": [ + -3.5297951698303223 + ], + "logprob_reward": -0.029795082286000252, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795082286000252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2650192081928253, + 0.03574288263916969, + 0.10379244387149811, + 0.052104659378528595, + 0.13666164875030518, + 0.03162194415926933, + 0.2782205641269684, + 0.016275158151984215, + 0.04596130549907684, + 0.034600187093019485 + ], + "credit_map": [ + 0.08073969185352325, + 0.020515723153948784, + 0.020118458196520805, + 0.019179651513695717, + 0.014570171944797039, + 0.017569510266184807, + 0.018681420013308525, + 0.07364460080862045, + 0.00517760356888175, + 0.00447875214740634, + 0.004121014848351479, + 0.003559643868356943, + 0.004263552837073803, + 0.0045553939417004585, + 0.0038186288438737392, + 0.0057682921178638935, + 0.011570130474865437, + 0.010159510187804699, + 0.018454456701874733, + 0.010197887197136879, + 0.009794237092137337, + 0.016457363963127136, + 0.0162062831223011, + 0.01095257792621851, + 0.007284923922270536, + 0.004618286620825529, + 0.0039035554509609938, + 0.0041915038600564, + 0.004311554599553347, + 0.004048310685902834, + 0.00865252036601305, + 0.01509400550276041, + 0.03442397341132164, + 0.008380716666579247, + 0.011133057996630669, + 0.012511981651186943, + 0.0048100147396326065, + 0.01017601415514946, + 0.011798379011452198, + 0.043427515774965286, + 0.007073625922203064, + 0.010642312467098236, + 0.0019804402254521847, + 0.0016918082255870104, + 0.003347506280988455, + 0.003373516257852316, + 0.00196273485198617, + 0.0015499977162107825, + 0.03532356396317482, + 0.0327158160507679, + 0.020105261355638504, + 0.013606542721390724, + 0.026188630610704422, + 0.09776496887207031, + 0.037244416773319244, + 0.015271374955773354, + 0.0015483356546610594, + 0.001536383992061019, + 0.0031032157130539417, + 0.002744184108451009, + 0.0016987051349133253, + 0.0012591014383360744, + 0.00200222828425467, + 0.002383004641160369, + 0.0064100767485797405, + 0.004513678606599569, + 0.0033577431458979845, + 0.013379312120378017, + 0.00419212132692337, + 0.005494598764926195, + 0.004709844943135977, + 0.0039039321709424257, + 0.006712048780173063, + 0.0033535303082317114, + 0.0021499679423868656, + 0.002105348976328969, + 0.002682965714484453, + 0.0028344979509711266, + 0.01032661460340023, + 0.004435213282704353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.029795082286000252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863502085208893 + ], + "opposite_logprob_tokens": [ + -3.0486350059509277 + ], + "logprob_reward": -0.04863502085208893, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863502085208893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15103211998939514, + 0.1132519543170929, + 0.17013907432556152, + 0.04276539757847786, + 0.0596950501203537, + 0.08741293102502823, + 0.12923365831375122, + 0.04977411404252052, + 0.06357359886169434, + 0.1331220120191574 + ], + "credit_map": [ + 0.026015475392341614, + 0.022153712809085846, + 0.016804952174425125, + 0.014738918282091618, + 0.012085054069757462, + 0.020739544183015823, + 0.01891668699681759, + 0.019577791914343834, + 0.012457622215151787, + 0.01450714748352766, + 0.01225963607430458, + 0.015509228222072124, + 0.012080463580787182, + 0.014696501195430756, + 0.014027080498635769, + 0.01771428808569908, + 0.02893482893705368, + 0.019581973552703857, + 0.027620265260338783, + 0.013985656201839447, + 0.017741896212100983, + 0.02596999891102314, + 0.01904500648379326, + 0.017259476706385612, + 0.006331531796604395, + 0.004092804621905088, + 0.005602480378001928, + 0.005607572849839926, + 0.003926072735339403, + 0.004040847532451153, + 0.005428981967270374, + 0.007735109888017178, + 0.01360279694199562, + 0.009642884135246277, + 0.004964443854987621, + 0.005759247578680515, + 0.00603505689650774, + 0.0038224330637604, + 0.0038646147586405277, + 0.012003579176962376, + 0.02006734535098076, + 0.028591353446245193, + 0.005511315073817968, + 0.005097729619592428, + 0.010188551619648933, + 0.0066155195236206055, + 0.006513615138828754, + 0.004827520810067654, + 0.01766130141913891, + 0.014483959414064884, + 0.0091858534142375, + 0.010073008015751839, + 0.010701560415327549, + 0.043923746794462204, + 0.011713454499840736, + 0.011490787379443645, + 0.005507329944521189, + 0.003143718233332038, + 0.015054527670145035, + 0.005476304795593023, + 0.004019416403025389, + 0.003720243461430073, + 0.0038441414944827557, + 0.009008443914353848, + 0.006964327301830053, + 0.01719817705452442, + 0.004613522905856371, + 0.005183475092053413, + 0.007803162094205618, + 0.011971727944910526, + 0.005863272584974766, + 0.0039759436622262, + 0.03322046622633934, + 0.007361575961112976, + 0.011771934106945992, + 0.008973391726613045, + 0.008953172713518143, + 0.006699034944176674, + 0.03886336460709572, + 0.017279090359807014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.04863502085208893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unwavering, static wide-angle shot captures a dramatic scene rendered entirely in the iconic Ukiyo-e woodcut style. The mountainous landscape, its rolling forms defined by simple, elegant lines and a flattened perspective, is completely covered in thick, stylized kudzu foliage. Each leaf is a unique heart-shaped form, with a bold black outline and flat planes of soft green and earthy brown, creating a richly textured graphic tapestry across the mountain. From the edge of the frame, a stylized flame advances relentlessly; the flames are not real, but depicted as vibrant, flat orange and yellow shapes, each lick and curl defined by a thick, inky black outline, devouring the foliage with artistic ferocity. As the fire spreads, billowing black smoke rises into the pale ochre sky, revealing solid, swirling gray and black forms with identical outlines, resembling the dynamic, stylized waves of a classic print. The camera's unwavering gaze transforms the destructive events into a living, moving photograph, the fire's relentless advance and the rising plumes of smoke creating a mesmerizing and powerful atmosphere of controlled, beautiful disaster.", + "seed": 22855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_6/index_804_seed_22855_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03682713028974831, + "vlm_logprob": -0.03682713028974831, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03682713028974831, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024006405845284462 + ], + "opposite_logprob_tokens": [ + -3.7740063667297363 + ], + "logprob_reward": -0.024006405845284462, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024006405845284462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21109941601753235, + 0.03074614517390728, + 0.02432878315448761, + 0.14347273111343384, + 0.08886571228504181, + 0.24921900033950806, + 0.03162449225783348, + 0.040701910853385925, + 0.02058100886642933, + 0.1593608260154724 + ], + "credit_map": [ + 0.013529768213629723, + 0.012082614004611969, + 0.040250781923532486, + 0.06753871589899063, + 0.017061755061149597, + 0.02080090530216694, + 0.024141566827893257, + 0.01569329760968685, + 0.004575967323035002, + 0.004019871354103088, + 0.004240250680595636, + 0.003315117908641696, + 0.004045819863677025, + 0.0038419889751821756, + 0.0038373558782041073, + 0.0028697694651782513, + 0.0023826996330171824, + 0.0025453120470046997, + 0.002741229720413685, + 0.0027106348425149918, + 0.002257334068417549, + 0.0038707638159394264, + 0.004763176664710045, + 0.0030576293356716633, + 0.010114097036421299, + 0.04319467768073082, + 0.005723968148231506, + 0.017733579501509666, + 0.01157798245549202, + 0.035560060292482376, + 0.007530653849244118, + 0.012037697248160839, + 0.01022940780967474, + 0.006517344620078802, + 0.006251139100641012, + 0.022902511060237885, + 0.006367215886712074, + 0.023084374144673347, + 0.007018259726464748, + 0.006495442241430283, + 0.02223501354455948, + 0.021525613963603973, + 0.03200136497616768, + 0.03751518204808235, + 0.029005110263824463, + 0.02524089813232422, + 0.055218104273080826, + 0.026477694511413574, + 0.0031895418651401997, + 0.002720032585784793, + 0.001817435142584145, + 0.00959681160748005, + 0.002473553642630577, + 0.006736713461577892, + 0.0027084501925855875, + 0.002381951780989766, + 0.009063858538866043, + 0.004513946827501059, + 0.0045395209454, + 0.004295932129025459, + 0.0037543277721852064, + 0.005587753839790821, + 0.0038549716118723154, + 0.005091593600809574, + 0.0027527709025889635, + 0.0022957101464271545, + 0.0026771710254251957, + 0.003128205193206668, + 0.0019956494215875864, + 0.0021710034925490618, + 0.0017840915825217962, + 0.0037764038424938917, + 0.010696938261389732, + 0.011218308471143246, + 0.0107404300943017, + 0.05644591525197029, + 0.00908771064132452, + 0.03545863926410675, + 0.01390942931175232, + 0.011803451925516129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.024006405845284462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650095202028751 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.008650095202028751, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650095202028751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13098983466625214, + 0.08670621365308762, + 0.0699709877371788, + 0.10368162393569946, + 0.1234419047832489, + 0.30628442764282227, + 0.020133445039391518, + 0.028641164302825928, + 0.02555251680314541, + 0.10459791868925095 + ], + "credit_map": [ + 0.010718555189669132, + 0.008872512727975845, + 0.023622171953320503, + 0.02633153274655342, + 0.018139485269784927, + 0.015078701078891754, + 0.015011339448392391, + 0.013215538114309311, + 0.010016683489084244, + 0.010061288252472878, + 0.013034448027610779, + 0.010958591476082802, + 0.01337135024368763, + 0.008766315877437592, + 0.01149401068687439, + 0.009003529325127602, + 0.006514059845358133, + 0.007409796584397554, + 0.008692769333720207, + 0.008204859681427479, + 0.006123502738773823, + 0.009334535337984562, + 0.014085648581385612, + 0.009605815634131432, + 0.009467476978898048, + 0.03207099437713623, + 0.0036402903497219086, + 0.00746195949614048, + 0.007879111915826797, + 0.03223275765776634, + 0.004157994873821735, + 0.006771044805645943, + 0.016035979613661766, + 0.006132481154054403, + 0.008559131063520908, + 0.031362611800432205, + 0.0074905287474393845, + 0.036111604422330856, + 0.009156325832009315, + 0.008593235164880753, + 0.03003077767789364, + 0.02487625740468502, + 0.040128711611032486, + 0.0467863604426384, + 0.032726358622312546, + 0.030462097376585007, + 0.06833495944738388, + 0.0329388789832592, + 0.003672500140964985, + 0.0023058070801198483, + 0.0016377873253077269, + 0.003108024364337325, + 0.0018828866304829717, + 0.0025884518399834633, + 0.001865219441242516, + 0.003072768682613969, + 0.005077929235994816, + 0.0052667781710624695, + 0.00298807630315423, + 0.0023123999126255512, + 0.0050512938760221004, + 0.0032432221341878176, + 0.002652302850037813, + 0.002049160422757268, + 0.0027910631615668535, + 0.0035264245234429836, + 0.0030295182950794697, + 0.0037186620756983757, + 0.002960932906717062, + 0.0026352598797529936, + 0.002360300859436393, + 0.004530356731265783, + 0.007884977385401726, + 0.008287760429084301, + 0.0069787525571882725, + 0.034738823771476746, + 0.008472170680761337, + 0.022108737379312515, + 0.008807449601590633, + 0.007319247350096703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.008650095202028751 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024905160535126925 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024905160535126925, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024905160535126925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.141707181930542, + 0.08016520738601685, + 0.03172978386282921, + 0.1777437925338745, + 0.2205972820520401, + 0.09619399905204773, + 0.05260780453681946, + 0.06741174310445786, + 0.026751531288027763, + 0.1050916537642479 + ], + "credit_map": [ + 0.008307022042572498, + 0.010041000321507454, + 0.023603113368153572, + 0.04544108361005783, + 0.02053898759186268, + 0.009722169488668442, + 0.014237194322049618, + 0.009816611185669899, + 0.009331272915005684, + 0.011699811555445194, + 0.010873144492506981, + 0.00786428153514862, + 0.01371132954955101, + 0.008720099925994873, + 0.00978587381541729, + 0.00817939918488264, + 0.0035047426354140043, + 0.003574667265638709, + 0.0037484269123524427, + 0.0042409771122038364, + 0.0029644761234521866, + 0.003519987454637885, + 0.0055981287732720375, + 0.004578377120196819, + 0.018022894859313965, + 0.048827290534973145, + 0.008159643970429897, + 0.016932900995016098, + 0.01201026700437069, + 0.05346314609050751, + 0.00824705045670271, + 0.012080604210495949, + 0.025946835055947304, + 0.011857877485454082, + 0.015918485820293427, + 0.06098036840558052, + 0.013066033832728863, + 0.06092802807688713, + 0.01734093390405178, + 0.014558728784322739, + 0.008923876099288464, + 0.009272624738514423, + 0.01083974540233612, + 0.014165800996124744, + 0.012641798704862595, + 0.008747055195271969, + 0.02129916474223137, + 0.010303929448127747, + 0.005594417452812195, + 0.010917133651673794, + 0.0037691080942749977, + 0.004754961002618074, + 0.011048085987567902, + 0.005621288437396288, + 0.005248041823506355, + 0.005654769018292427, + 0.015537467785179615, + 0.008124256506562233, + 0.007399125024676323, + 0.007480401545763016, + 0.007045152597129345, + 0.007997351698577404, + 0.006133577786386013, + 0.0076944115571677685, + 0.005366533529013395, + 0.0022420608438551426, + 0.0031449340749531984, + 0.0034840088337659836, + 0.0031258573289960623, + 0.0023187645711004734, + 0.002194942906498909, + 0.004874430131167173, + 0.009725499898195267, + 0.0068899281322956085, + 0.007132053375244141, + 0.036864880472421646, + 0.005096507258713245, + 0.02291291207075119, + 0.008905775845050812, + 0.007564104162156582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0024905160535126925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789971575140953 + ], + "opposite_logprob_tokens": [ + -2.5789971351623535 + ], + "logprob_reward": -0.0789971575140953, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789971575140953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12633883953094482, + 0.046347398310899734, + 0.044305846095085144, + 0.053742438554763794, + 0.03305617347359657, + 0.4751385450363159, + 0.06769170612096786, + 0.029911350458860397, + 0.010816396214067936, + 0.11265132576227188 + ], + "credit_map": [ + 0.0068197608925402164, + 0.006926620379090309, + 0.022192038595676422, + 0.04511525481939316, + 0.010126566514372826, + 0.009956036694347858, + 0.017118502408266068, + 0.008084064349532127, + 0.0068075163289904594, + 0.006842134986072779, + 0.004419011063873768, + 0.004410001449286938, + 0.006637746002525091, + 0.006681612692773342, + 0.00641366234049201, + 0.004135717172175646, + 0.004406753461807966, + 0.004575794097036123, + 0.004851054400205612, + 0.005031251814216375, + 0.004047013353556395, + 0.006433755625039339, + 0.009620190598070621, + 0.00534003134816885, + 0.006584370508790016, + 0.012218753807246685, + 0.0028310581110417843, + 0.003808698384091258, + 0.0035935158375650644, + 0.01823192834854126, + 0.0024301372468471527, + 0.0040439763106405735, + 0.004057752434164286, + 0.0036901612766087055, + 0.002686607651412487, + 0.006470049265772104, + 0.0034060056786984205, + 0.00477882893756032, + 0.003936012275516987, + 0.0040307557210326195, + 0.04359722509980202, + 0.039689138531684875, + 0.06179600954055786, + 0.07342271506786346, + 0.052611906081438065, + 0.04557489976286888, + 0.10690723359584808, + 0.051539406180381775, + 0.0061837974935770035, + 0.0052812425419688225, + 0.004302136600017548, + 0.02178255096077919, + 0.003906253259629011, + 0.013696995563805103, + 0.005975083447992802, + 0.006563647650182247, + 0.005520240869373083, + 0.003683770541101694, + 0.003584622172638774, + 0.003462836379185319, + 0.003034251043573022, + 0.004090454895049334, + 0.0029270662926137447, + 0.0036081073340028524, + 0.0016322375740855932, + 0.0012262328527867794, + 0.0013622945407405496, + 0.0014995761448517442, + 0.0013151676394045353, + 0.0011475965147837996, + 0.0008708427776582539, + 0.0017624485772103071, + 0.008084647357463837, + 0.007593107409775257, + 0.0075121368281543255, + 0.03977687284350395, + 0.005949110258370638, + 0.02605929784476757, + 0.009653373621404171, + 0.00802277959883213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0789971575140953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in the Ukiyo-e woodcut style with bold black outlines and flat planes of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a mountainous landscape covered in thick, stylized heart-shaped kudzu foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stylized flame advancing from the edge of the frame, depicted as vibrant flat orange and yellow shapes with black outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is billowing black smoke rising into a pale ochre sky, resembling stylized waves with solid gray and black forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999147683382034 + ], + "opposite_logprob_tokens": [ + -2.6949915885925293 + ], + "logprob_reward": -0.06999147683382034, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999147683382034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09993470460176468, + 0.0816645398736, + 0.07741856575012207, + 0.10039090365171432, + 0.373676061630249, + 0.17929229140281677, + 0.014467572793364525, + 0.009262834675610065, + 0.013219931162893772, + 0.05067263916134834 + ], + "credit_map": [ + 0.006168424151837826, + 0.005989154800772667, + 0.018731171265244484, + 0.026142194867134094, + 0.018082652240991592, + 0.010075890459120274, + 0.009107002057135105, + 0.0056382156908512115, + 0.006522864103317261, + 0.010501103475689888, + 0.013041412457823753, + 0.008255328051745892, + 0.009383167140185833, + 0.008808379992842674, + 0.015026367269456387, + 0.010125922970473766, + 0.006723752710968256, + 0.007497663609683514, + 0.00917187798768282, + 0.008905347436666489, + 0.006998015101999044, + 0.013144191354513168, + 0.014182822778820992, + 0.01079489104449749, + 0.007374304812401533, + 0.03052620030939579, + 0.004693301394581795, + 0.015282798558473587, + 0.009135488420724869, + 0.018605496734380722, + 0.005832695867866278, + 0.008940617553889751, + 0.04421048238873482, + 0.018864162266254425, + 0.026937875896692276, + 0.10823577642440796, + 0.018112218007445335, + 0.10653072595596313, + 0.028140898793935776, + 0.022643934935331345, + 0.01663048379123211, + 0.015035583637654781, + 0.022938044741749763, + 0.026928210631012917, + 0.02068607695400715, + 0.018084047362208366, + 0.04007093980908394, + 0.01891888864338398, + 0.0024497259873896837, + 0.002512310165911913, + 0.0011167594930157065, + 0.0012411409988999367, + 0.002708055544644594, + 0.0015889490023255348, + 0.0011779447086155415, + 0.0016726878238841891, + 0.003035040805116296, + 0.0009250340517610312, + 0.0007383738411590457, + 0.0007726708427071571, + 0.0008910584729164839, + 0.001207535620778799, + 0.0007606943836435676, + 0.0009324267157353461, + 0.003992982674390078, + 0.0018343263072893023, + 0.0010055284947156906, + 0.0012668374693021178, + 0.0016809331718832254, + 0.001207550405524671, + 0.0008447191212326288, + 0.0013870534021407366, + 0.005188435781747103, + 0.0037213678006082773, + 0.0033044235315173864, + 0.016851942986249924, + 0.0034494763240218163, + 0.010187944397330284, + 0.004084599204361439, + 0.0038844491355121136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.06999147683382034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_tensor([8397, 8397])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_tensor([8397, 8397])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f2030b1db5e1baeb2b2925d8d588bc03f618de5f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_tensor([8397, 8397])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly low-angle, full-body tracking shot smoothly follows a female ice skater, maintaining the perfect center of the frame as she glides across a vast, pristine ice rink. Her dark brown hair is pulled back into a sleek, tight bun, her cheeks flushed from the cold, an expression of intense, elegant concentration. She wears a stunning deep blue figure skating costume made of stretch velvet, the bodice embroidered with silver thread and shimmering sequins, capturing the light with every movement. Beneath, her muscular legs, clad in flesh-colored tights, culminate in pristine white leather skates, their smooth steel blades silently slicing through the ice. The camera moves parallel to her as she powerfully pushes off with one leg, the other extending behind her in a long, graceful glide, her arms extended for perfect balance. The setting is a vast, dark indoor arena, with powerful, cool-toned stadium spotlights shining down from above, isolating her in a pool of bright light. This high-contrast lighting casts a long, sharp, dramatic shadow extending directly onto the mirrored ice behind her, dynamically lengthening and contracting with her fluid movements, creating a powerful atmosphere of solitary focus, athleticism, and breathtaking artistry.", + "seed": 23661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23661_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021797859109938145, + "vlm_logprob": -0.021797859109938145, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.021797859109938145, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01270756870508194 + ], + "opposite_logprob_tokens": [ + -5.137707710266113 + ], + "logprob_reward": -0.01270756870508194, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01270756870508194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08474970608949661, + 0.18983842432498932, + 0.22639445960521698, + 0.13110150396823883, + 0.06767189502716064, + 0.05638131871819496, + 0.04702490195631981, + 0.058702632784843445, + 0.06947436183691025, + 0.06866075843572617 + ], + "credit_map": [ + 0.010981915518641472, + 0.01089698076248169, + 0.007451943121850491, + 0.010187136009335518, + 0.010574040934443474, + 0.011255046352744102, + 0.01286999974399805, + 0.010532642714679241, + 0.04214918985962868, + 0.06479252874851227, + 0.007262246683239937, + 0.008147011511027813, + 0.0185678880661726, + 0.014995012432336807, + 0.010370154865086079, + 0.023554394021630287, + 0.02877512015402317, + 0.0465724803507328, + 0.011374426074326038, + 0.022574909031391144, + 0.019654905423521996, + 0.04610621929168701, + 0.020998777821660042, + 0.030337607488036156, + 0.011670374311506748, + 0.009124122560024261, + 0.022312255576252937, + 0.012014049105346203, + 0.02213217504322529, + 0.029394224286079407, + 0.012106574140489101, + 0.012347734533250332, + 0.00647716224193573, + 0.009548566304147243, + 0.008195395581424236, + 0.007407739758491516, + 0.014300638809800148, + 0.00624398747459054, + 0.008118885569274426, + 0.007379517890512943, + 0.007906652055680752, + 0.005648329388350248, + 0.006778344511985779, + 0.0061874836683273315, + 0.007392127066850662, + 0.0069093285128474236, + 0.007861673831939697, + 0.007697377819567919, + 0.008134173229336739, + 0.004529814701527357, + 0.007641926873475313, + 0.004913472570478916, + 0.00573014048859477, + 0.006305545102804899, + 0.005024230573326349, + 0.004745597951114178, + 0.005079930182546377, + 0.007423819042742252, + 0.005036157090216875, + 0.0030000912956893444, + 0.012800989672541618, + 0.015006356872618198, + 0.004673713352531195, + 0.00568157946690917, + 0.005897922907024622, + 0.011016180738806725, + 0.007661759853363037, + 0.008464556187391281, + 0.00977430958300829, + 0.008303630165755749, + 0.010388251394033432, + 0.00796775333583355, + 0.0054475124925374985, + 0.007720057386904955, + 0.006411375477910042, + 0.007567253895103931, + 0.01435254979878664, + 0.011369673535227776, + 0.010489375330507755, + 0.005302964244037867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.01270756870508194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210934370756149 + ], + "opposite_logprob_tokens": [ + -2.8121092319488525 + ], + "logprob_reward": -0.06210934370756149, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210934370756149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13088490068912506, + 0.1587425023317337, + 0.10646482557058334, + 0.07969575375318527, + 0.05196765437722206, + 0.06257156282663345, + 0.16172486543655396, + 0.0807739645242691, + 0.08380748331546783, + 0.08336640894412994 + ], + "credit_map": [ + 0.020651105791330338, + 0.018228620290756226, + 0.013534323312342167, + 0.017289800569415092, + 0.01115501020103693, + 0.019052855670452118, + 0.01414654403924942, + 0.016826648265123367, + 0.033453065901994705, + 0.052500784397125244, + 0.005880287848412991, + 0.0066074468195438385, + 0.01260034553706646, + 0.014869587495923042, + 0.009867079555988312, + 0.022963931784033775, + 0.017372049391269684, + 0.01202821172773838, + 0.005419320892542601, + 0.007520412094891071, + 0.011766734533011913, + 0.029178520664572716, + 0.008726322092115879, + 0.01445327140390873, + 0.007455100305378437, + 0.007789646275341511, + 0.016134144738316536, + 0.00738113047555089, + 0.0141633041203022, + 0.013687502592802048, + 0.006269202567636967, + 0.006815733853727579, + 0.005356726236641407, + 0.004818754270672798, + 0.005333742126822472, + 0.0040834518149495125, + 0.012679691426455975, + 0.005836767610162497, + 0.007322733756154776, + 0.006535790860652924, + 0.007481591776013374, + 0.0049940673634409904, + 0.007453169673681259, + 0.007153305225074291, + 0.006899212021380663, + 0.008366965688765049, + 0.007801887113600969, + 0.012421369552612305, + 0.018313635140657425, + 0.01323108933866024, + 0.029706435278058052, + 0.011898559518158436, + 0.022498028352856636, + 0.035396285355091095, + 0.01642148196697235, + 0.014259375631809235, + 0.009180007502436638, + 0.010148869827389717, + 0.007317253854125738, + 0.0045747170224785805, + 0.01979139819741249, + 0.016245942562818527, + 0.006184338126331568, + 0.007331447210162878, + 0.00870361365377903, + 0.014227948151528835, + 0.01006233412772417, + 0.009845350868999958, + 0.01343490555882454, + 0.009832651354372501, + 0.01026875339448452, + 0.007431930862367153, + 0.008500817231833935, + 0.010650057345628738, + 0.007671221625059843, + 0.006655256263911724, + 0.020333586260676384, + 0.01620352268218994, + 0.006037626415491104, + 0.0073143248446285725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.06210934370756149 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590949803590775 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020590949803590775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590949803590775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11719238758087158, + 0.2424633800983429, + 0.06332464516162872, + 0.056328851729631424, + 0.05822959542274475, + 0.0732470452785492, + 0.1374100148677826, + 0.0765211284160614, + 0.08838150650262833, + 0.08690145611763 + ], + "credit_map": [ + 0.018602024763822556, + 0.01403780747205019, + 0.014059362933039665, + 0.016581617295742035, + 0.00978696160018444, + 0.01488647423684597, + 0.013544263318181038, + 0.01569388434290886, + 0.05286521464586258, + 0.08348111808300018, + 0.008288169279694557, + 0.009015041403472424, + 0.020537495613098145, + 0.02113082818686962, + 0.013644256629049778, + 0.0335012823343277, + 0.0080342385917902, + 0.008553712628781796, + 0.004493810702115297, + 0.005798902362585068, + 0.01026215497404337, + 0.011936047114431858, + 0.006855854764580727, + 0.0073899170383811, + 0.006106408778578043, + 0.0074611385352909565, + 0.011176055297255516, + 0.006357448175549507, + 0.009196419268846512, + 0.005409434903413057, + 0.004464918281883001, + 0.00615703035145998, + 0.006095916964113712, + 0.006079287733882666, + 0.006345330271869898, + 0.005109105259180069, + 0.014464177191257477, + 0.006102391052991152, + 0.00835031270980835, + 0.005683074705302715, + 0.009490064345300198, + 0.005518489051610231, + 0.008064555004239082, + 0.008044244721531868, + 0.008476724848151207, + 0.011787217110395432, + 0.0074984608218073845, + 0.014367293566465378, + 0.017113303765654564, + 0.013043594546616077, + 0.024395179003477097, + 0.01179402694106102, + 0.020781198516488075, + 0.020726628601551056, + 0.01665225625038147, + 0.012903827242553234, + 0.008895759470760822, + 0.008217421360313892, + 0.007068159990012646, + 0.003715441096574068, + 0.01906680315732956, + 0.016064252704381943, + 0.0056194718927145, + 0.00787382386624813, + 0.009612545371055603, + 0.015535818412899971, + 0.010933712124824524, + 0.01032287534326315, + 0.013381045311689377, + 0.010559195652604103, + 0.01074264943599701, + 0.007293663918972015, + 0.008542751893401146, + 0.011061392724514008, + 0.007517896592617035, + 0.006909320130944252, + 0.021364327520132065, + 0.017178233712911606, + 0.006919757928699255, + 0.007407773286104202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.020590949803590775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039071377366781235, + 0.06716708093881607, + 0.3957313001155853, + 0.15758925676345825, + 0.06058067828416824, + 0.10480601340532303, + 0.059129808098077774, + 0.020626386627554893, + 0.04530422389507294, + 0.0499938540160656 + ], + "credit_map": [ + 0.004338769242167473, + 0.004895173478871584, + 0.0067023346200585365, + 0.004662891384214163, + 0.004897718783468008, + 0.003619671333581209, + 0.005207581911236048, + 0.0047472380101680756, + 0.010624285787343979, + 0.016859939321875572, + 0.004371154122054577, + 0.006854576524347067, + 0.00947700534015894, + 0.006305679678916931, + 0.00630214624106884, + 0.006372296251356602, + 0.0702490508556366, + 0.02880394645035267, + 0.02029735967516899, + 0.03677460178732872, + 0.043815724551677704, + 0.12443794310092926, + 0.028947122395038605, + 0.04240552708506584, + 0.014398571103811264, + 0.017269456759095192, + 0.024659983813762665, + 0.012027599848806858, + 0.03438033536076546, + 0.029686154797673225, + 0.014158554375171661, + 0.011008607223629951, + 0.006749086081981659, + 0.006197800859808922, + 0.006764201447367668, + 0.007270144298672676, + 0.015524889342486858, + 0.006865760777145624, + 0.006295601837337017, + 0.004913195036351681, + 0.011206896044313908, + 0.010086633265018463, + 0.018086154013872147, + 0.012171179056167603, + 0.015872212126851082, + 0.010980737395584583, + 0.01431412622332573, + 0.01208807248622179, + 0.006721964571624994, + 0.0028750691562891006, + 0.010605799034237862, + 0.0031133329030126333, + 0.00862403679639101, + 0.020450793206691742, + 0.0034280535764992237, + 0.00331075768917799, + 0.002750708954408765, + 0.002214539097622037, + 0.0024610913824290037, + 0.001357169938273728, + 0.0033936691470444202, + 0.005562151782214642, + 0.001470404677093029, + 0.0014166502514854074, + 0.004161820746958256, + 0.008013817481696606, + 0.005610497202724218, + 0.005217045545578003, + 0.008030922152101994, + 0.005626629572361708, + 0.005655498243868351, + 0.0029879938811063766, + 0.005436798557639122, + 0.004834969062358141, + 0.0035442793741822243, + 0.0035653968807309866, + 0.012507537379860878, + 0.010110399685800076, + 0.0053911954164505005, + 0.004603275563567877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949999671429396 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949999671429396, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949999671429396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04252933710813522, + 0.15152516961097717, + 0.16285134851932526, + 0.08820880949497223, + 0.10771243274211884, + 0.06408844888210297, + 0.21274594962596893, + 0.0750400722026825, + 0.06409943848848343, + 0.031198957934975624 + ], + "credit_map": [ + 0.003974102903157473, + 0.005581636447459459, + 0.004976434633135796, + 0.005271558649837971, + 0.006407546810805798, + 0.006399225443601608, + 0.005600676406174898, + 0.004318153019994497, + 0.026364972814917564, + 0.0426417738199234, + 0.009736325591802597, + 0.008012892678380013, + 0.020355287939310074, + 0.013108748011291027, + 0.012256723828613758, + 0.01904843933880329, + 0.0305037684738636, + 0.010271022096276283, + 0.007975428365170956, + 0.010777197778224945, + 0.016733776777982712, + 0.05403899401426315, + 0.011050650849938393, + 0.021500516682863235, + 0.006905613001435995, + 0.008450339548289776, + 0.015751365572214127, + 0.00990314967930317, + 0.0156480111181736, + 0.01769392006099224, + 0.007889587432146072, + 0.0059668212197721004, + 0.01059110276401043, + 0.01270216703414917, + 0.014566845260560513, + 0.01592279225587845, + 0.01990591548383236, + 0.010131938382983208, + 0.012827937491238117, + 0.011063735000789165, + 0.006860201247036457, + 0.006578427739441395, + 0.007002346217632294, + 0.006740286946296692, + 0.010739869438111782, + 0.00854206271469593, + 0.0112296212464571, + 0.0063956319354474545, + 0.022677067667245865, + 0.013780838809907436, + 0.04003370925784111, + 0.013647029176354408, + 0.03038170374929905, + 0.05858684331178665, + 0.017922021448612213, + 0.01571674272418022, + 0.0055063278414309025, + 0.008073895238339901, + 0.004262398928403854, + 0.003853096393868327, + 0.017679408192634583, + 0.019144805148243904, + 0.00871478021144867, + 0.007805362809449434, + 0.006632177624851465, + 0.011685878969728947, + 0.005891745910048485, + 0.007154570426791906, + 0.011265826411545277, + 0.007713099475950003, + 0.009986650198698044, + 0.003769490635022521, + 0.002915217773988843, + 0.005017071031033993, + 0.0037534458097070456, + 0.0036324383690953255, + 0.00600070646032691, + 0.004810229875147343, + 0.0025876888539642096, + 0.0024821588303893805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.005949999671429396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, full-body tracking shot smoothly follows a female ice skater, maintaining the perfect center of the frame as she glides across a vast, pristine ice rink. Her dark brown hair is pulled back into a sleek, tight bun, her cheeks flushed from the cold, an expression of intense, elegant concentration. She wears a stunning deep blue figure skating costume made of stretch velvet, the bodice embroidered with silver thread and shimmering sequins, capturing the light with every movement. Beneath, her muscular legs, clad in flesh-colored tights, culminate in pristine white leather skates, their smooth steel blades silently slicing through the ice. The camera moves parallel to her as she powerfully pushes off with one leg, the other extending behind her in a long, graceful glide, her arms extended for perfect balance. The setting is a vast, dark indoor arena, with powerful, cool-toned stadium spotlights shining down from above, isolating her in a pool of bright light. This high-contrast lighting casts a long, sharp, dramatic shadow extending directly onto the mirrored ice behind her, dynamically lengthening and contracting with her fluid movements, creating a powerful atmosphere of solitary focus, athleticism, and breathtaking artistry.", + "seed": 23662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23662_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.060640210006386044, + "vlm_logprob": -0.060640210006386044, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.060640210006386044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04681622236967087 + ], + "opposite_logprob_tokens": [ + -3.671816110610962 + ], + "logprob_reward": -0.04681622236967087, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04681622236967087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15146097540855408, + 0.07094278186559677, + 0.09891881048679352, + 0.13167081773281097, + 0.09426718950271606, + 0.04600470885634422, + 0.07455183565616608, + 0.24611513316631317, + 0.03620075806975365, + 0.04986700043082237 + ], + "credit_map": [ + 0.012190800160169601, + 0.02452916093170643, + 0.013969876803457737, + 0.02442997135221958, + 0.018990661948919296, + 0.017080655321478844, + 0.021775079891085625, + 0.01849476620554924, + 0.006900351494550705, + 0.007294448558241129, + 0.006947464309632778, + 0.005664442665874958, + 0.005584653001278639, + 0.020635239779949188, + 0.006895809434354305, + 0.011020376347005367, + 0.007455482613295317, + 0.008576331660151482, + 0.02472713403403759, + 0.009727110154926777, + 0.011165300384163857, + 0.006607499904930592, + 0.014255506917834282, + 0.01640445552766323, + 0.014641543850302696, + 0.015194443054497242, + 0.01436677761375904, + 0.015282508917152882, + 0.00998235959559679, + 0.007344177924096584, + 0.036185845732688904, + 0.018673168495297432, + 0.00954714510589838, + 0.007738420274108648, + 0.014522209763526917, + 0.018125565722584724, + 0.007032363209873438, + 0.006968645378947258, + 0.013981791213154793, + 0.016351046040654182, + 0.0030989376828074455, + 0.0023718636948615313, + 0.011365829035639763, + 0.00561025645583868, + 0.003078550100326538, + 0.003130929311737418, + 0.006461859680712223, + 0.010886484757065773, + 0.0052448539063334465, + 0.005478057079017162, + 0.02458711713552475, + 0.010604391805827618, + 0.004404714796692133, + 0.0038419319316744804, + 0.005826663691550493, + 0.014564107172191143, + 0.013008599169552326, + 0.00810711644589901, + 0.0836663618683815, + 0.03213343769311905, + 0.007663121446967125, + 0.03938719257712364, + 0.012762251310050488, + 0.04938705265522003, + 0.002652000868692994, + 0.0017374880844727159, + 0.010140868835151196, + 0.008341490291059017, + 0.0019457365851849318, + 0.0036432957276701927, + 0.0035851586144417524, + 0.004154720343649387, + 0.003393007442355156, + 0.005309378262609243, + 0.006633060984313488, + 0.006724764592945576, + 0.006254562642425299, + 0.0036868916358798742, + 0.006983570288866758, + 0.010881763882935047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.04681622236967087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16052711009979248 + ], + "opposite_logprob_tokens": [ + -1.9105271100997925 + ], + "logprob_reward": -0.16052711009979248, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16052711009979248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16950677335262299, + 0.04969401657581329, + 0.09357629716396332, + 0.1321433186531067, + 0.06690389662981033, + 0.012920294888317585, + 0.033644769340753555, + 0.3792647421360016, + 0.029477450996637344, + 0.0328684002161026 + ], + "credit_map": [ + 0.014954236336052418, + 0.032038405537605286, + 0.01687568426132202, + 0.02596173994243145, + 0.022452564910054207, + 0.01709740236401558, + 0.021533530205488205, + 0.018593231216073036, + 0.005109226331114769, + 0.004496915265917778, + 0.005683469586074352, + 0.004529368598014116, + 0.004776277579367161, + 0.012606695294380188, + 0.00416138768196106, + 0.008330685086548328, + 0.007134178187698126, + 0.008505603298544884, + 0.014702693559229374, + 0.008788387291133404, + 0.011622657999396324, + 0.007772611454129219, + 0.020330920815467834, + 0.014719259925186634, + 0.015361898578703403, + 0.015225454233586788, + 0.015386750921607018, + 0.016363900154829025, + 0.010016001760959625, + 0.007765476126223803, + 0.03307793661952019, + 0.018945910036563873, + 0.00663213012740016, + 0.006184806115925312, + 0.010073192417621613, + 0.011904115788638592, + 0.005514935590326786, + 0.004933943040668964, + 0.010959614999592304, + 0.010701166465878487, + 0.0010616859653964639, + 0.0009123578201979399, + 0.0025170710869133472, + 0.0018436888931319118, + 0.0010893851285800338, + 0.0010112691670656204, + 0.0019133477471768856, + 0.002571490127593279, + 0.003516797674819827, + 0.0047723157331347466, + 0.006425421219319105, + 0.006184217054396868, + 0.0024624867364764214, + 0.002332210773602128, + 0.003680877387523651, + 0.004270447418093681, + 0.017543071880936623, + 0.02372129261493683, + 0.08595511317253113, + 0.03449716791510582, + 0.028980065137147903, + 0.08848554641008377, + 0.035507626831531525, + 0.06457488983869553, + 0.002353443531319499, + 0.001675726380199194, + 0.008202937431633472, + 0.006078030448406935, + 0.0020239504519850016, + 0.0036288772244006395, + 0.001947654876857996, + 0.003566835541278124, + 0.002044428139925003, + 0.0032510655000805855, + 0.004755497444421053, + 0.005286485888063908, + 0.003933911211788654, + 0.002291529905050993, + 0.005748577881604433, + 0.005556906573474407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.16052711009979248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308326914906502 + ], + "opposite_logprob_tokens": [ + -3.1680831909179688 + ], + "logprob_reward": -0.04308326914906502, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308326914906502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09493870288133621, + 0.04461721330881119, + 0.09060358256101608, + 0.10885223746299744, + 0.09210502356290817, + 0.04597840458154678, + 0.02586994878947735, + 0.4505827724933624, + 0.019664183259010315, + 0.026787949725985527 + ], + "credit_map": [ + 0.00719843152910471, + 0.011584168300032616, + 0.009528201073408127, + 0.015130001120269299, + 0.013259612955152988, + 0.015360510908067226, + 0.010673591867089272, + 0.012204189784824848, + 0.005011796019971371, + 0.002510620979592204, + 0.005759354215115309, + 0.0028736351523548365, + 0.0029762890189886093, + 0.01116451621055603, + 0.003849327564239502, + 0.010471670888364315, + 0.0050765275955200195, + 0.004291351418942213, + 0.033854853361845016, + 0.011116357520222664, + 0.003351499792188406, + 0.005016673356294632, + 0.007533190306276083, + 0.020363125950098038, + 0.014964083209633827, + 0.012994655407965183, + 0.011858618818223476, + 0.012289305217564106, + 0.008045484311878681, + 0.007247631903737783, + 0.02529279515147209, + 0.01615966483950615, + 0.008577678352594376, + 0.00797357503324747, + 0.013830099254846573, + 0.01755363680422306, + 0.007910631597042084, + 0.00609414279460907, + 0.014766662381589413, + 0.015398597344756126, + 0.003256754716858268, + 0.005063035525381565, + 0.006401133257895708, + 0.007282691542059183, + 0.0052503552287817, + 0.003832163754850626, + 0.007316055707633495, + 0.00757621368393302, + 0.002192296087741852, + 0.002813126193359494, + 0.006998230237513781, + 0.00351949711330235, + 0.0016261457931250334, + 0.0015305201523005962, + 0.0024367121513932943, + 0.004753421526402235, + 0.02091432921588421, + 0.02795618213713169, + 0.10326512157917023, + 0.04078051075339317, + 0.03391474485397339, + 0.10482225567102432, + 0.0415273979306221, + 0.07740223407745361, + 0.001406600116752088, + 0.0018179371254518628, + 0.003935896325856447, + 0.003407557960599661, + 0.0018779236124828458, + 0.0024873462971299887, + 0.001994515536352992, + 0.0027364057023078203, + 0.001604190212674439, + 0.003059748560190201, + 0.003780379891395569, + 0.004612281918525696, + 0.0028995403554290533, + 0.0018571802647784352, + 0.00513567216694355, + 0.0038389561232179403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.04308326914906502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867669194936752 + ], + "opposite_logprob_tokens": [ + -3.0486767292022705 + ], + "logprob_reward": -0.04867669194936752, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867669194936752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15808561444282532, + 0.04878180846571922, + 0.08823540806770325, + 0.07086168229579926, + 0.18303240835666656, + 0.07029248028993607, + 0.07470154762268066, + 0.23978716135025024, + 0.029134785756468773, + 0.03708711639046669 + ], + "credit_map": [ + 0.015492678619921207, + 0.02567053586244583, + 0.01637142337858677, + 0.024052003398537636, + 0.016933498904109, + 0.022588634863495827, + 0.017015205696225166, + 0.01996162347495556, + 0.004395173862576485, + 0.004475169349461794, + 0.012299160473048687, + 0.005831918679177761, + 0.004051525611430407, + 0.0046649412252008915, + 0.005366536322981119, + 0.007697379216551781, + 0.006422726903110743, + 0.005806987173855305, + 0.028950227424502373, + 0.01084722951054573, + 0.005789276212453842, + 0.005490061827003956, + 0.00824161246418953, + 0.016687285155057907, + 0.0114485127851367, + 0.009807790629565716, + 0.006715838331729174, + 0.0069692619144916534, + 0.0052156997844576836, + 0.005830399692058563, + 0.014944098889827728, + 0.009930072352290154, + 0.019529862329363823, + 0.01600196585059166, + 0.028380125761032104, + 0.03342154994606972, + 0.013376381248235703, + 0.013637019321322441, + 0.027052288874983788, + 0.03163321316242218, + 0.005260044243186712, + 0.005580740515142679, + 0.011800841428339481, + 0.009409750811755657, + 0.007309336680918932, + 0.005134751088917255, + 0.012391429394483566, + 0.013405585661530495, + 0.0077052246779203415, + 0.01152859628200531, + 0.013053870759904385, + 0.012278109788894653, + 0.0051532103680074215, + 0.005337348207831383, + 0.008464498445391655, + 0.011180692352354527, + 0.012329120188951492, + 0.01722702570259571, + 0.040491100400686264, + 0.022737648338079453, + 0.021629249677062035, + 0.060876455157995224, + 0.027736760675907135, + 0.03675977140665054, + 0.0018430209020152688, + 0.0014269824605435133, + 0.008116433396935463, + 0.00740178395062685, + 0.0016876704758033156, + 0.0028172440361231565, + 0.0023423198144882917, + 0.0034993302542716265, + 0.002692539943382144, + 0.004581528715789318, + 0.004320258740335703, + 0.005266030319035053, + 0.005414347629994154, + 0.0031923989299684763, + 0.004935584496706724, + 0.006684425752609968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.04867669194936752 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10214386135339737, + 0.062267448753118515, + 0.09170455485582352, + 0.08724234253168106, + 0.04293161630630493, + 0.024276887997984886, + 0.028403636068105698, + 0.4828408360481262, + 0.05215571075677872, + 0.026033200323581696 + ], + "credit_map": [ + 0.01477909553796053, + 0.027128316462039948, + 0.011871635913848877, + 0.011309267021715641, + 0.011696829460561275, + 0.0070770373567938805, + 0.009280186146497726, + 0.009001482278108597, + 0.005892644636332989, + 0.005338871385902166, + 0.007277046795934439, + 0.00614236481487751, + 0.003962310031056404, + 0.01440346147865057, + 0.007615113165229559, + 0.011635632254183292, + 0.007428443059325218, + 0.011269456706941128, + 0.011106940917670727, + 0.008707760833203793, + 0.01388456393033266, + 0.00931695569306612, + 0.017477823421359062, + 0.012512593530118465, + 0.008762032724916935, + 0.008641141466796398, + 0.008564835414290428, + 0.011562429368495941, + 0.004479314666241407, + 0.009257822297513485, + 0.020787697285413742, + 0.015187056735157967, + 0.002992260968312621, + 0.005451755132526159, + 0.006048231851309538, + 0.0060363165102899075, + 0.004414892755448818, + 0.004723389632999897, + 0.0066076964139938354, + 0.006657067686319351, + 0.0028731857892125845, + 0.002508696634322405, + 0.002707886742427945, + 0.003221415914595127, + 0.0025153576862066984, + 0.0019781538285315037, + 0.0043921866454184055, + 0.004080003127455711, + 0.0015896244440227747, + 0.0020246414933353662, + 0.00903158076107502, + 0.004715250339359045, + 0.001923360861837864, + 0.001786754815839231, + 0.0028454617131501436, + 0.004486958961933851, + 0.021707141771912575, + 0.02575145661830902, + 0.12502849102020264, + 0.048285774886608124, + 0.029133476316928864, + 0.11095855385065079, + 0.0346464067697525, + 0.0873294547200203, + 0.0035043195821344852, + 0.0031520742923021317, + 0.012493116781115532, + 0.009960570372641087, + 0.0037006656639277935, + 0.00702590961009264, + 0.00512700853869319, + 0.007192038930952549, + 0.002338133519515395, + 0.0037324915174394846, + 0.0040024579502642155, + 0.003567260457202792, + 0.0032608946785330772, + 0.0021558699663728476, + 0.0038834381848573685, + 0.003092651953920722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.004097756464034319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, full-body tracking shot smoothly follows a female ice skater, maintaining the perfect center of the frame as she glides across a vast, pristine ice rink. Her dark brown hair is pulled back into a sleek, tight bun, her cheeks flushed from the cold, an expression of intense, elegant concentration. She wears a stunning deep blue figure skating costume made of stretch velvet, the bodice embroidered with silver thread and shimmering sequins, capturing the light with every movement. Beneath, her muscular legs, clad in flesh-colored tights, culminate in pristine white leather skates, their smooth steel blades silently slicing through the ice. The camera moves parallel to her as she powerfully pushes off with one leg, the other extending behind her in a long, graceful glide, her arms extended for perfect balance. The setting is a vast, dark indoor arena, with powerful, cool-toned stadium spotlights shining down from above, isolating her in a pool of bright light. This high-contrast lighting casts a long, sharp, dramatic shadow extending directly onto the mirrored ice behind her, dynamically lengthening and contracting with her fluid movements, creating a powerful atmosphere of solitary focus, athleticism, and breathtaking artistry.", + "seed": 23663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23663_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028047707863152027, + "vlm_logprob": -0.028047707863152027, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.028047707863152027, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021850910037755966 + ], + "opposite_logprob_tokens": [ + -4.896851062774658 + ], + "logprob_reward": -0.021850910037755966, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.021850910037755966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11517822742462158, + 0.0830053985118866, + 0.06288701295852661, + 0.03841443732380867, + 0.17928609251976013, + 0.19738063216209412, + 0.21473278105258942, + 0.03723714128136635, + 0.03895715996623039, + 0.032921090722084045 + ], + "credit_map": [ + 0.018913598731160164, + 0.0158814899623394, + 0.008526177145540714, + 0.009348821826279163, + 0.01921863853931427, + 0.016942618414759636, + 0.013584616594016552, + 0.012762272730469704, + 0.00647442601621151, + 0.012807274237275124, + 0.010224212892353535, + 0.015388760715723038, + 0.009440717287361622, + 0.008605063892900944, + 0.01312743779271841, + 0.006937509868294001, + 0.00583308469504118, + 0.005797328427433968, + 0.008839010260999203, + 0.006485544145107269, + 0.009295087307691574, + 0.01187881175428629, + 0.007375531364232302, + 0.007382620126008987, + 0.0030113549437373877, + 0.004963081330060959, + 0.0032599030528217554, + 0.007624388672411442, + 0.004532062914222479, + 0.0036189905367791653, + 0.006868112366646528, + 0.0045365504920482635, + 0.023594003170728683, + 0.008274550549685955, + 0.010806250385940075, + 0.006811645347625017, + 0.03599041327834129, + 0.06319016963243484, + 0.010449460707604885, + 0.020169619470834732, + 0.02653447911143303, + 0.01868429221212864, + 0.017524998635053635, + 0.021265720948576927, + 0.04001879692077637, + 0.03872619941830635, + 0.018055401742458344, + 0.016570743173360825, + 0.022025160491466522, + 0.0299542173743248, + 0.029131634160876274, + 0.035416968166828156, + 0.026663275435566902, + 0.02318725362420082, + 0.0262410007417202, + 0.022113291546702385, + 0.004510618280619383, + 0.003655076026916504, + 0.008236425928771496, + 0.004661387763917446, + 0.003053858643397689, + 0.0029826625250279903, + 0.004752311389893293, + 0.0053848023526370525, + 0.004969960078597069, + 0.005118648521602154, + 0.0031552293803542852, + 0.004294938407838345, + 0.004822137299925089, + 0.005217176862061024, + 0.007505332585424185, + 0.003873737994581461, + 0.003373384475708008, + 0.004015376325696707, + 0.004085032735019922, + 0.0033360410016030073, + 0.0036667045205831528, + 0.0034077672753483057, + 0.00515303248539567, + 0.005883755628019571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.021850910037755966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905498147010803 + ], + "opposite_logprob_tokens": [ + -2.579055070877075 + ], + "logprob_reward": -0.07905498147010803, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905498147010803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3315609395503998, + 0.23853738605976105, + 0.08549068868160248, + 0.0481632761657238, + 0.045630376785993576, + 0.07967036217451096, + 0.058117784559726715, + 0.04478377848863602, + 0.037341322749853134, + 0.030704017728567123 + ], + "credit_map": [ + 0.047482505440711975, + 0.052427809685468674, + 0.02969053015112877, + 0.036959704011678696, + 0.046215616166591644, + 0.04548364132642746, + 0.04833853244781494, + 0.02496262639760971, + 0.02463860996067524, + 0.03063499927520752, + 0.021092401817440987, + 0.02305561490356922, + 0.044749144464731216, + 0.03799957409501076, + 0.03561137244105339, + 0.02075568027794361, + 0.009586293250322342, + 0.007157876156270504, + 0.012202410958707333, + 0.006566987838596106, + 0.012993276119232178, + 0.020812857896089554, + 0.007558135781437159, + 0.00861285999417305, + 0.007288869936019182, + 0.006281075533479452, + 0.0043191928416490555, + 0.006182172801345587, + 0.009982872754335403, + 0.005991974379867315, + 0.004514662548899651, + 0.003602459095418453, + 0.006260918918997049, + 0.004165396559983492, + 0.0026515722274780273, + 0.004173755645751953, + 0.008477220311760902, + 0.01227500569075346, + 0.004805820994079113, + 0.0028206862043589354, + 0.009052536450326443, + 0.009027915075421333, + 0.006137039512395859, + 0.012013742700219154, + 0.01207918394356966, + 0.010764224454760551, + 0.012110230512917042, + 0.008485489524900913, + 0.005430553574115038, + 0.009269582107663155, + 0.00789390504360199, + 0.00951655488461256, + 0.007178188767284155, + 0.006590987090021372, + 0.007510391995310783, + 0.004727622494101524, + 0.004720231983810663, + 0.004273483995348215, + 0.0055791218765079975, + 0.006832953076809645, + 0.004355203825980425, + 0.007277266588062048, + 0.006442684680223465, + 0.005302839446812868, + 0.004986634943634272, + 0.005425137002021074, + 0.0033545519690960646, + 0.0037042065523564816, + 0.005264048930257559, + 0.005584497936069965, + 0.0064979237504303455, + 0.0025243214331567287, + 0.0031918028835207224, + 0.003932715859264135, + 0.00288976076990366, + 0.003236432559788227, + 0.003686523297801614, + 0.003271134104579687, + 0.0057619474828243256, + 0.004733703564852476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07905498147010803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355599984526634 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355599984526634, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355599984526634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4444037079811096, + 0.1036548987030983, + 0.045991867780685425, + 0.034951601177453995, + 0.09959283471107483, + 0.06791368871927261, + 0.07313040643930435, + 0.045481208711862564, + 0.03264721855521202, + 0.052232496440410614 + ], + "credit_map": [ + 0.0660400316119194, + 0.07223594933748245, + 0.03708602115511894, + 0.050242096185684204, + 0.05589476227760315, + 0.059099771082401276, + 0.07004716247320175, + 0.03375793620944023, + 0.010652974247932434, + 0.013317891396582127, + 0.011495986022055149, + 0.013170444406569004, + 0.018367614597082138, + 0.013207145035266876, + 0.015058116987347603, + 0.008384719491004944, + 0.004198487848043442, + 0.007072369568049908, + 0.004468713887035847, + 0.006133613176643848, + 0.006695328280329704, + 0.006957142613828182, + 0.006717350799590349, + 0.003748863935470581, + 0.0051102349534630775, + 0.004942866042256355, + 0.0038966769352555275, + 0.005497111473232508, + 0.005078935995697975, + 0.003067114856094122, + 0.0033074424136430025, + 0.0040512168779969215, + 0.013440241105854511, + 0.006629277486354113, + 0.0066938563250005245, + 0.005157071631401777, + 0.018413087353110313, + 0.03277399763464928, + 0.0067345155403018, + 0.00975078996270895, + 0.00825540255755186, + 0.007655725814402103, + 0.0051657455042004585, + 0.010113918222486973, + 0.00977740716189146, + 0.008575436659157276, + 0.011135698296129704, + 0.0072343540377914906, + 0.006595039274543524, + 0.010326310992240906, + 0.010211142711341381, + 0.012509055435657501, + 0.007490064948797226, + 0.0077767930924892426, + 0.00938018411397934, + 0.00884181447327137, + 0.004923705942928791, + 0.005697154905647039, + 0.004869435913860798, + 0.0063932607881724834, + 0.004867916461080313, + 0.006042473949491978, + 0.007266780827194452, + 0.005420483183115721, + 0.006507611833512783, + 0.004431817214936018, + 0.0035629495978355408, + 0.0030565382912755013, + 0.0038791366387158632, + 0.003803827101364732, + 0.003873081412166357, + 0.0035322559997439384, + 0.005191932898014784, + 0.0077546387910842896, + 0.005282466299831867, + 0.004735472612082958, + 0.006722922902554274, + 0.005192297976464033, + 0.00892960000783205, + 0.00842316448688507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.026355599984526634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21859031915664673, + 0.3609955906867981, + 0.06299521774053574, + 0.08085792511701584, + 0.019239516928792, + 0.04596927389502525, + 0.03845953941345215, + 0.08532292395830154, + 0.028321247547864914, + 0.05924844369292259 + ], + "credit_map": [ + 0.025321630761027336, + 0.018161596730351448, + 0.01490189041942358, + 0.01192852109670639, + 0.05006982758641243, + 0.0573347806930542, + 0.014349480159580708, + 0.026522601023316383, + 0.033051762729883194, + 0.04278047755360603, + 0.03403306007385254, + 0.043435901403427124, + 0.07595247030258179, + 0.054139282554388046, + 0.04867392033338547, + 0.02892872877418995, + 0.008288424462080002, + 0.004478279035538435, + 0.010116741061210632, + 0.006110247224569321, + 0.007982907816767693, + 0.010616857558488846, + 0.005839105229824781, + 0.0095626600086689, + 0.010385087691247463, + 0.006783672142773867, + 0.007658583112061024, + 0.00725483987480402, + 0.019909489899873734, + 0.01290600560605526, + 0.009335141628980637, + 0.006625100504606962, + 0.0022061150521039963, + 0.0017412329325452447, + 0.0018195273587480187, + 0.0017075232462957501, + 0.0032270182855427265, + 0.0048233079724013805, + 0.0018489020876586437, + 0.0018658890621736646, + 0.005817836616188288, + 0.005018871743232012, + 0.003853544592857361, + 0.0050285980105400085, + 0.007477742154151201, + 0.007592897396534681, + 0.005052919965237379, + 0.006126864813268185, + 0.00413336930796504, + 0.0041337343864142895, + 0.004015109967440367, + 0.0044989269226789474, + 0.004871123470366001, + 0.007278904784470797, + 0.0035025784745812416, + 0.006025792099535465, + 0.012594764120876789, + 0.005539600271731615, + 0.006839284673333168, + 0.007573005743324757, + 0.011531657539308071, + 0.022823607549071312, + 0.00793225597590208, + 0.010488749481737614, + 0.005282552447170019, + 0.0038075794000178576, + 0.0031966071110218763, + 0.0027611153200268745, + 0.0036521228030323982, + 0.003501360770314932, + 0.0033295280300080776, + 0.002790382830426097, + 0.005945253185927868, + 0.0055259643122553825, + 0.014567454345524311, + 0.009805495850741863, + 0.004506703466176987, + 0.006094467826187611, + 0.005904103629291058, + 0.006898998748511076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19843176007270813, + 0.14380989968776703, + 0.05653740093111992, + 0.06054670363664627, + 0.12363076955080032, + 0.1445372998714447, + 0.05302957817912102, + 0.06274080276489258, + 0.07132285088300705, + 0.08541296422481537 + ], + "credit_map": [ + 0.0217159204185009, + 0.012331812642514706, + 0.014682188630104065, + 0.011666967533528805, + 0.038466449826955795, + 0.05292098596692085, + 0.015763241797685623, + 0.030884191393852234, + 0.014997958205640316, + 0.02017834410071373, + 0.014127026312053204, + 0.018035126850008965, + 0.024888666346669197, + 0.022309014573693275, + 0.018189795315265656, + 0.011083971709012985, + 0.0069232662208378315, + 0.0066038272343575954, + 0.0062647671438753605, + 0.006889026612043381, + 0.007282581180334091, + 0.007392860949039459, + 0.008703597821295261, + 0.006477476097643375, + 0.005421042907983065, + 0.006816126871854067, + 0.007296568714082241, + 0.010418246500194073, + 0.007858959957957268, + 0.006062312982976437, + 0.010146909393370152, + 0.00652653444558382, + 0.016265282407402992, + 0.013322360813617706, + 0.007581645157188177, + 0.012678387574851513, + 0.02026357501745224, + 0.03227836638689041, + 0.012564458884298801, + 0.008676703087985516, + 0.013808270916342735, + 0.01634129136800766, + 0.011224856600165367, + 0.0200106929987669, + 0.018090607598423958, + 0.01766461879014969, + 0.024369312450289726, + 0.02302764542400837, + 0.005340057425200939, + 0.006910570431500673, + 0.006364185828715563, + 0.00816530641168356, + 0.008381523191928864, + 0.00670934421941638, + 0.006383985280990601, + 0.004774600267410278, + 0.01058223471045494, + 0.00392926437780261, + 0.014091220684349537, + 0.008956425823271275, + 0.005230434238910675, + 0.007155071943998337, + 0.006455299910157919, + 0.006340850610285997, + 0.00759816262871027, + 0.010675117373466492, + 0.006866428069770336, + 0.006034319754689932, + 0.011192506179213524, + 0.011552036739885807, + 0.012181412428617477, + 0.005222870968282223, + 0.008261914364993572, + 0.00997835025191307, + 0.009616570547223091, + 0.007312373258173466, + 0.009828736074268818, + 0.008769690059125423, + 0.011705460958182812, + 0.019939865916967392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0031929248943924904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, full-body tracking shot smoothly follows a female ice skater, maintaining the perfect center of the frame as she glides across a vast, pristine ice rink. Her dark brown hair is pulled back into a sleek, tight bun, her cheeks flushed from the cold, an expression of intense, elegant concentration. She wears a stunning deep blue figure skating costume made of stretch velvet, the bodice embroidered with silver thread and shimmering sequins, capturing the light with every movement. Beneath, her muscular legs, clad in flesh-colored tights, culminate in pristine white leather skates, their smooth steel blades silently slicing through the ice. The camera moves parallel to her as she powerfully pushes off with one leg, the other extending behind her in a long, graceful glide, her arms extended for perfect balance. The setting is a vast, dark indoor arena, with powerful, cool-toned stadium spotlights shining down from above, isolating her in a pool of bright light. This high-contrast lighting casts a long, sharp, dramatic shadow extending directly onto the mirrored ice behind her, dynamically lengthening and contracting with her fluid movements, creating a powerful atmosphere of solitary focus, athleticism, and breathtaking artistry.", + "seed": 23664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_7/index_8397_seed_23664_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02747051939368248, + "vlm_logprob": -0.02747051939368248, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02747051939368248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030045099556446075 + ], + "opposite_logprob_tokens": [ + -5.030045032501221 + ], + "logprob_reward": -0.030045099556446075, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.030045099556446075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05593482777476311, + 0.03373555839061737, + 0.06179213896393776, + 0.0886281207203865, + 0.03781009465456009, + 0.029649559408426285, + 0.1505308449268341, + 0.08823342621326447, + 0.07344797998666763, + 0.3802374303340912 + ], + "credit_map": [ + 0.007705202791839838, + 0.006352385971695185, + 0.004896956961601973, + 0.006310156546533108, + 0.009280610829591751, + 0.0068145617842674255, + 0.00908360444009304, + 0.005491351708769798, + 0.0039274427108466625, + 0.004560814239084721, + 0.003322568256407976, + 0.004961710888892412, + 0.004566403105854988, + 0.003235498210415244, + 0.003503755433484912, + 0.005657365545630455, + 0.003075334709137678, + 0.00394890271127224, + 0.019858431071043015, + 0.005524830427020788, + 0.0028901498299092054, + 0.0023464413825422525, + 0.004279610235244036, + 0.019868437200784683, + 0.006507732905447483, + 0.00656293798238039, + 0.008876601234078407, + 0.0087820990011096, + 0.01050928607583046, + 0.008310257457196712, + 0.015591847710311413, + 0.023487361147999763, + 0.0047554997727274895, + 0.004390893969684839, + 0.0038281683810055256, + 0.007213369477540255, + 0.004873845260590315, + 0.003119452390819788, + 0.004094065632671118, + 0.005534799303859472, + 0.0044520203955471516, + 0.003953868523240089, + 0.003682090900838375, + 0.0025976214092224836, + 0.006799808703362942, + 0.002476146910339594, + 0.0029623601585626602, + 0.0027256407774984837, + 0.01754462718963623, + 0.011774498969316483, + 0.024492474272847176, + 0.017241893336176872, + 0.02241424471139908, + 0.012949026189744473, + 0.019398333504796028, + 0.024715743958950043, + 0.008152936585247517, + 0.007665084209293127, + 0.014877410605549812, + 0.007849644869565964, + 0.010495365597307682, + 0.014210838824510574, + 0.01078918855637312, + 0.014192963019013405, + 0.005921355914324522, + 0.008137574419379234, + 0.012943827547132969, + 0.013963190838694572, + 0.006489151157438755, + 0.0049112034030258656, + 0.010830041952431202, + 0.010251631028950214, + 0.07758920639753342, + 0.03154156729578972, + 0.027316881343722343, + 0.012853950262069702, + 0.019054746255278587, + 0.1458224356174469, + 0.020267143845558167, + 0.045791491866111755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.030045099556446075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06211438402533531 + ], + "opposite_logprob_tokens": [ + -2.8121144771575928 + ], + "logprob_reward": -0.06211438402533531, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06211438402533531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10666372627019882, + 0.08296336978673935, + 0.18030697107315063, + 0.09884334355592728, + 0.05314148962497711, + 0.01726393960416317, + 0.12822814285755157, + 0.11465343832969666, + 0.040907859802246094, + 0.17702771723270416 + ], + "credit_map": [ + 0.01102663017809391, + 0.012878426350653172, + 0.008998407982289791, + 0.011315648443996906, + 0.016731826588511467, + 0.010218349285423756, + 0.02160247601568699, + 0.013891969807446003, + 0.010387281887233257, + 0.014877748675644398, + 0.010981028899550438, + 0.010443841107189655, + 0.012013755738735199, + 0.007886958308517933, + 0.006530835758894682, + 0.009841921739280224, + 0.006625282112509012, + 0.006004414055496454, + 0.06491810083389282, + 0.012819221243262291, + 0.006315193604677916, + 0.006348868366330862, + 0.012355486862361431, + 0.06492040306329727, + 0.007506250403821468, + 0.006954541429877281, + 0.010589909739792347, + 0.011617036536335945, + 0.0073279873467981815, + 0.007061718963086605, + 0.01920224539935589, + 0.02858366072177887, + 0.005924204830080271, + 0.010142560116946697, + 0.0053994618356227875, + 0.011079997755587101, + 0.005222382489591837, + 0.005736078135669231, + 0.004901213105767965, + 0.0047355880960822105, + 0.00214679935015738, + 0.0017985431477427483, + 0.0024833418428897858, + 0.0018231236608698964, + 0.002310957992449403, + 0.001398965367116034, + 0.0015527595533058047, + 0.003749449271708727, + 0.0111885666847229, + 0.015569230541586876, + 0.013832498341798782, + 0.020843366160988808, + 0.021836677566170692, + 0.011309395544230938, + 0.018764864653348923, + 0.014883545227348804, + 0.0075147636234760284, + 0.015358246862888336, + 0.014535339549183846, + 0.014982280321419239, + 0.014187115244567394, + 0.011079419404268265, + 0.01862410455942154, + 0.01837216690182686, + 0.0033203274942934513, + 0.005523175932466984, + 0.005960694048553705, + 0.006172860506922007, + 0.004376822151243687, + 0.004149158019572496, + 0.005941926036030054, + 0.0054628970101475716, + 0.03556174039840698, + 0.014945310540497303, + 0.022929051890969276, + 0.006079176440834999, + 0.008172796107828617, + 0.06894461810588837, + 0.006595554295927286, + 0.013799475505948067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.06211438402533531 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814867302775383 + ], + "opposite_logprob_tokens": [ + -3.5298149585723877 + ], + "logprob_reward": -0.029814867302775383, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814867302775383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13464413583278656, + 0.08033464103937149, + 0.11571944504976273, + 0.15554241836071014, + 0.12295075505971909, + 0.05541320517659187, + 0.07315617054700851, + 0.11550169438123703, + 0.026102544739842415, + 0.12063499540090561 + ], + "credit_map": [ + 0.01340335700660944, + 0.010984484106302261, + 0.011534054763615131, + 0.014557629823684692, + 0.022325240075588226, + 0.011233935132622719, + 0.032242219895124435, + 0.018363216891884804, + 0.008889280259609222, + 0.013741404749453068, + 0.012352566234767437, + 0.011078228242695332, + 0.00919515360146761, + 0.0066262828186154366, + 0.00546484999358654, + 0.012986873276531696, + 0.005628429353237152, + 0.006427582819014788, + 0.041282374411821365, + 0.0055612665601074696, + 0.005162383895367384, + 0.005618763621896505, + 0.006792782805860043, + 0.039245858788490295, + 0.013803095556795597, + 0.011246243491768837, + 0.018567226827144623, + 0.022178998216986656, + 0.012645409442484379, + 0.012570835649967194, + 0.025075120851397514, + 0.03945547714829445, + 0.013895965181291103, + 0.02002115547657013, + 0.011620208621025085, + 0.02630108781158924, + 0.01476152054965496, + 0.012475013732910156, + 0.011293239891529083, + 0.01258255634456873, + 0.006723340600728989, + 0.008897223509848118, + 0.005739116575568914, + 0.005641943775117397, + 0.011237566359341145, + 0.00442692544311285, + 0.005761218257248402, + 0.006985872518271208, + 0.008096513338387012, + 0.006304949056357145, + 0.011246065609157085, + 0.007321916986256838, + 0.012877163477241993, + 0.00793157797306776, + 0.008820747025310993, + 0.010557238012552261, + 0.009806367568671703, + 0.008913354948163033, + 0.019489021971821785, + 0.010810978710651398, + 0.012851947918534279, + 0.014122232794761658, + 0.0179350133985281, + 0.021572770550847054, + 0.002406903775408864, + 0.0032144798897206783, + 0.003792297560721636, + 0.004851094912737608, + 0.0023404534440487623, + 0.002462556352838874, + 0.0030727845150977373, + 0.00396197522059083, + 0.01992175728082657, + 0.008772354573011398, + 0.027626540511846542, + 0.005263492465019226, + 0.0046502999030053616, + 0.04120916128158569, + 0.004477646667510271, + 0.008713748306035995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.029814867302775383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642767556011677 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642767556011677, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642767556011677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10949374735355377, + 0.1469414383172989, + 0.1675596386194229, + 0.28090062737464905, + 0.06388057768344879, + 0.008618849329650402, + 0.0924878790974617, + 0.049476563930511475, + 0.022074390202760696, + 0.05856632813811302 + ], + "credit_map": [ + 0.009946155361831188, + 0.008976617828011513, + 0.009616672992706299, + 0.01313952449709177, + 0.015593557618558407, + 0.008206562139093876, + 0.027569225057959557, + 0.016445431858301163, + 0.02546139433979988, + 0.014865822158753872, + 0.013571529649198055, + 0.011783851310610771, + 0.03344905376434326, + 0.0200347937643528, + 0.014575363136827946, + 0.013199622742831707, + 0.0066553750075399876, + 0.005696193315088749, + 0.05735873430967331, + 0.007812430150806904, + 0.007406033109873533, + 0.007008247543126345, + 0.01233197282999754, + 0.06329067051410675, + 0.01886369101703167, + 0.014865766279399395, + 0.027258286252617836, + 0.021062660962343216, + 0.04068675637245178, + 0.030476972460746765, + 0.049321554601192474, + 0.07836493104696274, + 0.00840023159980774, + 0.006347253918647766, + 0.007483102381229401, + 0.006911813747137785, + 0.007335667964071035, + 0.007709832396358252, + 0.006662654224783182, + 0.01303002517670393, + 0.0011346905957907438, + 0.0010896915337070823, + 0.001075042993761599, + 0.0009768059244379401, + 0.0011463214177638292, + 0.000774179759901017, + 0.0007731085061095655, + 0.0016490088310092688, + 0.011322199366986752, + 0.00773744098842144, + 0.013954090885818005, + 0.010690794326364994, + 0.013807890936732292, + 0.008608831092715263, + 0.011746615171432495, + 0.014620007947087288, + 0.004729015287011862, + 0.0036740442737936974, + 0.010982004925608635, + 0.004004964604973793, + 0.005224801134318113, + 0.006227730307728052, + 0.00608123978599906, + 0.008552761748433113, + 0.001883627031929791, + 0.0025411839596927166, + 0.003375458298251033, + 0.0028798694256693125, + 0.0030606184154748917, + 0.0020751538686454296, + 0.003691538004204631, + 0.0025669410824775696, + 0.011697345413267612, + 0.004590779077261686, + 0.009048320353031158, + 0.002142447279766202, + 0.0024301467929035425, + 0.02324085496366024, + 0.0018418324179947376, + 0.0035746032372117043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.008642767556011677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle full-body tracking shot of a female ice skater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater wearing a deep blue velvet costume with silver embroidery and sequins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater have dark brown hair in a tight bun and flushed cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast dark indoor arena with cool-toned spotlights creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track parallel to the skater while she glides with one leg pushing off and the other extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735478527843952 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735478527843952, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735478527843952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0755220428109169, + 0.0732962042093277, + 0.020348984748125076, + 0.07343573123216629, + 0.036734919995069504, + 0.1293589323759079, + 0.05773751065135002, + 0.16135992109775543, + 0.0658998042345047, + 0.3063059151172638 + ], + "credit_map": [ + 0.008800446055829525, + 0.009716007858514786, + 0.005341017618775368, + 0.006254383362829685, + 0.014593958854675293, + 0.007597662974148989, + 0.014131618663668633, + 0.00908694788813591, + 0.009569861926138401, + 0.007097602356225252, + 0.008580245077610016, + 0.008376896381378174, + 0.012498652562499046, + 0.0061052413657307625, + 0.005855777766555548, + 0.015211929567158222, + 0.0015661041252315044, + 0.001806519110687077, + 0.004503244999796152, + 0.0027655563317239285, + 0.002232260536402464, + 0.001586434431374073, + 0.0023031935561448336, + 0.0035856717731803656, + 0.004403955768793821, + 0.004828895442187786, + 0.00748349167406559, + 0.0058607193641364574, + 0.007294332142919302, + 0.006323125213384628, + 0.014786039479076862, + 0.02245517447590828, + 0.003782375482842326, + 0.004258869215846062, + 0.0050080628134310246, + 0.006198864430189133, + 0.004905197769403458, + 0.002849886193871498, + 0.0038006058894097805, + 0.005931057967245579, + 0.015191764570772648, + 0.020357243716716766, + 0.018203161656856537, + 0.01702621392905712, + 0.018929433077573776, + 0.009533069096505642, + 0.017745843157172203, + 0.012372193858027458, + 0.0054650385864079, + 0.006391975563019514, + 0.007122652605175972, + 0.008823098614811897, + 0.008393310010433197, + 0.003954221028834581, + 0.007397145964205265, + 0.010190068744122982, + 0.015353305265307426, + 0.011889791116118431, + 0.031510647386312485, + 0.014419284649193287, + 0.01764867454767227, + 0.02246103249490261, + 0.02151292935013771, + 0.026564259082078934, + 0.006040051579475403, + 0.00818867702037096, + 0.00834578461945057, + 0.007689803373068571, + 0.011626922525465488, + 0.006518270820379257, + 0.009667878039181232, + 0.007822414860129356, + 0.06318813562393188, + 0.025352951139211655, + 0.016208454966545105, + 0.00929259043186903, + 0.014237971045076847, + 0.11821631342172623, + 0.012834769673645496, + 0.04697470739483833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.006735478527843952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_tensor([2326, 2326])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_tensor([2326, 2326])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a43e6f6f4c696efa19b2b6049766320958dea18f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_tensor([2326, 2326])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, ultra-wide shot, the camera moving slowly and steadily backward, captures a lone figure walking along a straight, dusty path. The figure's features are obscured by the deep hood of a rough, earth-colored cloak, and he moves with a steady, deliberate pace. Each footstep stirs up a lingering puff of pale, fine dust from the cracked earth. Seen from above, a vast geometric grid of towering steel power lines dominates the frame; their intricate, repetitive pattern of towers and sagging cables stretches from the foreground to a vanishing point on the distant horizon, dwarfing the lone traveler. In the evening, harsh golden sunlight casts long shadows across the barren landscape, highlighting the textures of the parched earth and the weathered, rusted metal towers. The color palette is a gentle symphony of ochre, brown, and a hazy, faded blue sky, creating an overwhelming atmosphere of desolation, quiet determination, and an endless journey.", + "seed": 13459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_2326_seed_13459_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08094926318153739, + "vlm_logprob": -0.08094926318153739, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.08094926318153739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14698071777820587 + ], + "opposite_logprob_tokens": [ + -2.7719807624816895 + ], + "logprob_reward": -0.14698071777820587, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.14698071777820587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12058188766241074, + 0.2517448961734772, + 0.12257511168718338, + 0.06313342601060867, + 0.155110701918602, + 0.03878963738679886, + 0.02714693173766136, + 0.03136410564184189, + 0.15265163779258728, + 0.03690166026353836 + ], + "credit_map": [ + 0.015574922785162926, + 0.012204975821077824, + 0.016778359189629555, + 0.020418347790837288, + 0.009223919361829758, + 0.012250345200300217, + 0.016791343688964844, + 0.017339667305350304, + 0.025389881804585457, + 0.02770637348294258, + 0.026929251849651337, + 0.03937990963459015, + 0.023264653980731964, + 0.025972094386816025, + 0.03155602887272835, + 0.05154670029878616, + 0.009264450520277023, + 0.02005188725888729, + 0.008373023010790348, + 0.016265342012047768, + 0.018659785389900208, + 0.011551146395504475, + 0.026118626818060875, + 0.01229084562510252, + 0.0060313125140964985, + 0.006508346181362867, + 0.006794782355427742, + 0.012066999450325966, + 0.0067679076455533504, + 0.006985235493630171, + 0.008356898091733456, + 0.009621944278478622, + 0.01039675809442997, + 0.020294364541769028, + 0.0124065475538373, + 0.016542330384254456, + 0.031566306948661804, + 0.022549858316779137, + 0.025381915271282196, + 0.015972623601555824, + 0.0026161738205701113, + 0.004255266394466162, + 0.003619440132752061, + 0.004358646459877491, + 0.0058321417309343815, + 0.002708291634917259, + 0.010419425554573536, + 0.004980252590030432, + 0.001589481602422893, + 0.0027610452380031347, + 0.003272342961281538, + 0.004574644844979048, + 0.0029449386056512594, + 0.0016701376298442483, + 0.006844712421298027, + 0.0034896277356892824, + 0.002342470223084092, + 0.0036816326901316643, + 0.003566063242033124, + 0.0034764003939926624, + 0.004727750550955534, + 0.0027044392190873623, + 0.006641930900514126, + 0.004223421216011047, + 0.008519424125552177, + 0.020346766337752342, + 0.019611401483416557, + 0.017948882654309273, + 0.017680402845144272, + 0.008568956516683102, + 0.04466567933559418, + 0.015310131013393402, + 0.004653521813452244, + 0.006415494251996279, + 0.004743334371596575, + 0.0029788282699882984, + 0.006441596429795027, + 0.003585608908906579, + 0.004940677434206009, + 0.0031426004134118557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.14698071777820587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215408205986023 + ], + "opposite_logprob_tokens": [ + -1.5021541118621826 + ], + "logprob_reward": -0.25215408205986023, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215408205986023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13364051282405853, + 0.09293573349714279, + 0.2013334333896637, + 0.08020757138729095, + 0.22560879588127136, + 0.09176936000585556, + 0.038214802742004395, + 0.028367504477500916, + 0.07980425655841827, + 0.028118111193180084 + ], + "credit_map": [ + 0.011476104147732258, + 0.016200898215174675, + 0.014333460479974747, + 0.04226124659180641, + 0.010218414478003979, + 0.007723986636847258, + 0.01835479587316513, + 0.013071595691144466, + 0.008045537397265434, + 0.01909617893397808, + 0.00847693718969822, + 0.013527153991162777, + 0.011828040704131126, + 0.006009778007864952, + 0.015153632499277592, + 0.010798459872603416, + 0.014514335431158543, + 0.03422480821609497, + 0.014166446402668953, + 0.02816585637629032, + 0.03763167932629585, + 0.023495113477110863, + 0.029495341703295708, + 0.019639834761619568, + 0.008951651863753796, + 0.007507788483053446, + 0.011196695268154144, + 0.013771777041256428, + 0.007791469339281321, + 0.006464977283030748, + 0.010457915253937244, + 0.014065289869904518, + 0.018433695659041405, + 0.030697694048285484, + 0.017775004729628563, + 0.018123198300600052, + 0.049510080367326736, + 0.03537431359291077, + 0.03713436424732208, + 0.01856043189764023, + 0.004223409108817577, + 0.007077335380017757, + 0.009327840059995651, + 0.017934013158082962, + 0.007375394459813833, + 0.007137601263821125, + 0.026495467871427536, + 0.012198295444250107, + 0.0032911489252001047, + 0.0038605553563684225, + 0.0030052601359784603, + 0.0060783992521464825, + 0.005084441974759102, + 0.0038319139275699854, + 0.008301912806928158, + 0.004761167801916599, + 0.002453773282468319, + 0.003012752393260598, + 0.0030653499998152256, + 0.0033955578692257404, + 0.003716415725648403, + 0.0027009693440049887, + 0.00600140867754817, + 0.004021272528916597, + 0.006221574265509844, + 0.008460215292870998, + 0.011413619853556156, + 0.010885043069720268, + 0.006997267249971628, + 0.007416192442178726, + 0.01712566614151001, + 0.011284672655165195, + 0.003124529728665948, + 0.0028331733774393797, + 0.002850158140063286, + 0.005370319355279207, + 0.003388825338333845, + 0.0023378883488476276, + 0.004513364750891924, + 0.003699850058183074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.25215408205986023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017219253350049257 + ], + "opposite_logprob_tokens": [ + -6.3767218589782715 + ], + "logprob_reward": -0.0017219253350049257, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017219253350049257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0442187562584877, + 0.10520722717046738, + 0.05922601372003555, + 0.021830979734659195, + 0.2616425156593323, + 0.05756672844290733, + 0.042081933468580246, + 0.027374843135476112, + 0.3562853932380676, + 0.024565642699599266 + ], + "credit_map": [ + 0.0029959785751998425, + 0.003955945372581482, + 0.004076610319316387, + 0.012705087661743164, + 0.003777153789997101, + 0.002964244456961751, + 0.008803239092230797, + 0.0049404967576265335, + 0.01476992852985859, + 0.009389298968017101, + 0.01131246704608202, + 0.01239651907235384, + 0.011211277917027473, + 0.012737731449306011, + 0.013045345433056355, + 0.02034466527402401, + 0.0044935098849236965, + 0.009000372141599655, + 0.005081539042294025, + 0.0072655389085412025, + 0.0080794682726264, + 0.005104060284793377, + 0.012972339987754822, + 0.007229182869195938, + 0.0023493794724345207, + 0.002057685749605298, + 0.002224109834060073, + 0.0041134534403681755, + 0.0020504810381680727, + 0.0022519745398312807, + 0.0028565051034092903, + 0.003927391022443771, + 0.021346934139728546, + 0.04044072702527046, + 0.011789814569056034, + 0.02100561559200287, + 0.06675601750612259, + 0.03970339894294739, + 0.04007859155535698, + 0.02052142843604088, + 0.0022837393917143345, + 0.004832101054489613, + 0.005642620846629143, + 0.010011859238147736, + 0.005153663456439972, + 0.004941069055348635, + 0.01741601899266243, + 0.007285652682185173, + 0.002970058238133788, + 0.005802019499242306, + 0.003315961454063654, + 0.007265625987201929, + 0.006142384838312864, + 0.0033268288243561983, + 0.008900023996829987, + 0.004359029233455658, + 0.0014369447017088532, + 0.0024201832711696625, + 0.002382260514423251, + 0.004324162378907204, + 0.0030525994952768087, + 0.0019915346056222916, + 0.008163787424564362, + 0.00360336946323514, + 0.01681179739534855, + 0.04492245614528656, + 0.040946703404188156, + 0.04756813496351242, + 0.0392475463449955, + 0.0176131259649992, + 0.10746293514966965, + 0.041712693870067596, + 0.0013861906481906772, + 0.0024831993505358696, + 0.002550381002947688, + 0.0053968727588653564, + 0.003122548572719097, + 0.001684323069639504, + 0.004652279894798994, + 0.0032898494973778725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.0017219253350049257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019480077316984534 + ], + "opposite_logprob_tokens": [ + -6.25194787979126 + ], + "logprob_reward": -0.0019480077316984534, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019480077316984534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07952769100666046, + 0.19503311812877655, + 0.03926512226462364, + 0.02585119567811489, + 0.47133761644363403, + 0.020814118906855583, + 0.016361987218260765, + 0.04368801414966583, + 0.08784212172031403, + 0.020279044285416603 + ], + "credit_map": [ + 0.009914245456457138, + 0.008187921717762947, + 0.007162981666624546, + 0.02263094298541546, + 0.004551938269287348, + 0.006162646226584911, + 0.009947464801371098, + 0.010969548486173153, + 0.024538323283195496, + 0.026007208973169327, + 0.022425713017582893, + 0.02754845842719078, + 0.02330319955945015, + 0.022416315972805023, + 0.026540575549006462, + 0.022253340110182762, + 0.003469800343737006, + 0.005899752955883741, + 0.004392331466078758, + 0.004213401116430759, + 0.005663692951202393, + 0.003795191179960966, + 0.0068190088495612144, + 0.005011940374970436, + 0.0017175190150737762, + 0.0020339982584118843, + 0.0033375525381416082, + 0.004337301477789879, + 0.0022056582383811474, + 0.0031415470875799656, + 0.004574204329401255, + 0.0045034149661660194, + 0.03695378825068474, + 0.07054360955953598, + 0.023804927244782448, + 0.043891023844480515, + 0.11358263343572617, + 0.06569908559322357, + 0.07509223371744156, + 0.04177030175924301, + 0.0011071222834289074, + 0.0022450783289968967, + 0.0020783995278179646, + 0.003242962993681431, + 0.002407953841611743, + 0.002206322969868779, + 0.005218690726906061, + 0.0023075866047292948, + 0.000975422328338027, + 0.0014310244005173445, + 0.0017918922239914536, + 0.002678653923794627, + 0.0019508945988491178, + 0.0013456778833642602, + 0.004100010264664888, + 0.0020884121768176556, + 0.0023006638512015343, + 0.004937555640935898, + 0.003776072058826685, + 0.00588236702606082, + 0.0061463648453354836, + 0.003806656925007701, + 0.011320974677801132, + 0.005517360754311085, + 0.0038618443068116903, + 0.010202791541814804, + 0.008148483000695705, + 0.014701656065881252, + 0.010003847070038319, + 0.0042067887261509895, + 0.026393432170152664, + 0.01032327301800251, + 0.001564737525768578, + 0.002412493573501706, + 0.00181075488217175, + 0.0037464466877281666, + 0.002626999979838729, + 0.0013343717437237501, + 0.004058755934238434, + 0.0027244852390140295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.0019480077316984534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019415830029174685 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019415830029174685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019415830029174685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23705759644508362, + 0.19728520512580872, + 0.058132436126470566, + 0.07167711853981018, + 0.2052817940711975, + 0.0617617592215538, + 0.02839614264667034, + 0.036000799387693405, + 0.0764462947845459, + 0.027960823848843575 + ], + "credit_map": [ + 0.03558104857802391, + 0.025754328817129135, + 0.03976893424987793, + 0.03315011411905289, + 0.016629865393042564, + 0.025118032470345497, + 0.02127758041024208, + 0.0397777184844017, + 0.01851099170744419, + 0.03354868292808533, + 0.016887672245502472, + 0.02275891602039337, + 0.0188058540225029, + 0.018503280356526375, + 0.022044245153665543, + 0.04622558876872063, + 0.005741342436522245, + 0.005203270353376865, + 0.005624822806566954, + 0.00857179332524538, + 0.00614912947639823, + 0.006284564267843962, + 0.013763613998889923, + 0.00679390924051404, + 0.007683624513447285, + 0.006700172554701567, + 0.00816397462040186, + 0.013987534679472446, + 0.006864118855446577, + 0.006390335038304329, + 0.009783460758626461, + 0.012103904969990253, + 0.015377743169665337, + 0.030233509838581085, + 0.010536986403167248, + 0.01853836700320244, + 0.04958410933613777, + 0.027815455570816994, + 0.034745924174785614, + 0.01844971440732479, + 0.004907791968435049, + 0.006486466620117426, + 0.005535326898097992, + 0.008476107381284237, + 0.01051434688270092, + 0.006478180177509785, + 0.011815890669822693, + 0.00754765048623085, + 0.002942359074950218, + 0.002914711134508252, + 0.0024515697732567787, + 0.003265928477048874, + 0.004141851328313351, + 0.002425341634079814, + 0.00711979903280735, + 0.0031345870811492205, + 0.005166210234165192, + 0.005050553474575281, + 0.00497432891279459, + 0.00257584429346025, + 0.007653118576854467, + 0.003543229540809989, + 0.0036826978903263807, + 0.0033548185601830482, + 0.005481785628944635, + 0.008669917471706867, + 0.008281523361802101, + 0.00964354444295168, + 0.00854996033012867, + 0.007243457715958357, + 0.01942056603729725, + 0.00915555004030466, + 0.0035769445821642876, + 0.003577044466510415, + 0.002879355102777481, + 0.0028510941192507744, + 0.004543562885373831, + 0.0030004915315657854, + 0.004864410497248173, + 0.0026679213624447584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0019415830029174685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, ultra-wide shot, the camera moving slowly and steadily backward, captures a lone figure walking along a straight, dusty path. The figure's features are obscured by the deep hood of a rough, earth-colored cloak, and he moves with a steady, deliberate pace. Each footstep stirs up a lingering puff of pale, fine dust from the cracked earth. Seen from above, a vast geometric grid of towering steel power lines dominates the frame; their intricate, repetitive pattern of towers and sagging cables stretches from the foreground to a vanishing point on the distant horizon, dwarfing the lone traveler. In the evening, harsh golden sunlight casts long shadows across the barren landscape, highlighting the textures of the parched earth and the weathered, rusted metal towers. The color palette is a gentle symphony of ochre, brown, and a hazy, faded blue sky, creating an overwhelming atmosphere of desolation, quiet determination, and an endless journey.", + "seed": 13460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_2326_seed_13460_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07753551080822944, + "vlm_logprob": -0.07753551080822944, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.07753551080822944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09659531712532043 + ], + "opposite_logprob_tokens": [ + -3.096595287322998 + ], + "logprob_reward": -0.09659531712532043, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.09659531712532043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23720711469650269, + 0.07024282217025757, + 0.06403354555368423, + 0.09417400509119034, + 0.16613906621932983, + 0.033648472279310226, + 0.14979077875614166, + 0.07108636945486069, + 0.0667242705821991, + 0.04695352539420128 + ], + "credit_map": [ + 0.02755514718592167, + 0.025597630068659782, + 0.01823689416050911, + 0.02638581208884716, + 0.03337683156132698, + 0.025650179013609886, + 0.046488914638757706, + 0.03391570970416069, + 0.009446032345294952, + 0.009263859130442142, + 0.006054059136658907, + 0.006301832385361195, + 0.0070575810968875885, + 0.008628902956843376, + 0.01378875132650137, + 0.009701801463961601, + 0.006827355828136206, + 0.006956065073609352, + 0.005931796971708536, + 0.0065636662766337395, + 0.00781233748421073, + 0.007539799902588129, + 0.012384948320686817, + 0.010017572902143002, + 0.008294597268104553, + 0.00977646466344595, + 0.00753584923222661, + 0.010072383098304272, + 0.018199527636170387, + 0.011233764700591564, + 0.018809298053383827, + 0.010252121835947037, + 0.013991927728056908, + 0.014371071010828018, + 0.013964083977043629, + 0.018542615696787834, + 0.03367741033434868, + 0.018185129389166832, + 0.03149228170514107, + 0.02191455289721489, + 0.0036852830089628696, + 0.005976058542728424, + 0.002391753951087594, + 0.0038250924553722143, + 0.004809573758393526, + 0.0041246190667152405, + 0.004430866334587336, + 0.004405225161463022, + 0.01290871948003769, + 0.010105554945766926, + 0.007851283065974712, + 0.015165507793426514, + 0.03078000247478485, + 0.01651395484805107, + 0.04050508514046669, + 0.015960659831762314, + 0.007632782217115164, + 0.007239820435643196, + 0.004671138245612383, + 0.007070666179060936, + 0.015009071677923203, + 0.008240786381065845, + 0.015583562664687634, + 0.005638544913381338, + 0.007246243301779032, + 0.007212815340608358, + 0.004078072030097246, + 0.006356560159474611, + 0.017275016754865646, + 0.00992819294333458, + 0.009015623480081558, + 0.005611747968941927, + 0.004790924955159426, + 0.0039304629899561405, + 0.004014893900603056, + 0.004774230532348156, + 0.01027387473732233, + 0.006298169959336519, + 0.0071687898598611355, + 0.005702175199985504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.09659531712532043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813586890697479 + ], + "opposite_logprob_tokens": [ + -1.4063587188720703 + ], + "logprob_reward": -0.2813586890697479, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813586890697479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14164118468761444, + 0.05718545243144035, + 0.11693938076496124, + 0.07238972932100296, + 0.06896422803401947, + 0.0456819050014019, + 0.17970706522464752, + 0.14586642384529114, + 0.1317211091518402, + 0.039903488010168076 + ], + "credit_map": [ + 0.02994932234287262, + 0.013811567798256874, + 0.010361019521951675, + 0.019332680851221085, + 0.015317878685891628, + 0.018222395330667496, + 0.019577376544475555, + 0.015068946406245232, + 0.005234485026448965, + 0.00661466782912612, + 0.004206891171634197, + 0.0053784954361617565, + 0.012793072499334812, + 0.006580615881830454, + 0.009557457640767097, + 0.006819767877459526, + 0.01350284181535244, + 0.012386322021484375, + 0.008219040930271149, + 0.013516079634428024, + 0.01901836507022381, + 0.010805569589138031, + 0.022653888911008835, + 0.016837267205119133, + 0.006012753117829561, + 0.005534663796424866, + 0.005860485136508942, + 0.00929681770503521, + 0.01098349317908287, + 0.007030929438769817, + 0.014902672730386257, + 0.012767909094691277, + 0.006768439896404743, + 0.007813293486833572, + 0.004540832247585058, + 0.009320381097495556, + 0.010854755528271198, + 0.007418669760227203, + 0.014289083890616894, + 0.007958773523569107, + 0.00623754458501935, + 0.006974998861551285, + 0.0035757585428655148, + 0.005605420097708702, + 0.0069397324696183205, + 0.005494515877217054, + 0.006317486986517906, + 0.004536446649581194, + 0.020240243524312973, + 0.015185787342488766, + 0.009257372468709946, + 0.012895901687443256, + 0.05489557608962059, + 0.027266256511211395, + 0.027753261849284172, + 0.01221266109496355, + 0.017958350479602814, + 0.012621497735381126, + 0.008511830121278763, + 0.014812495559453964, + 0.03478119894862175, + 0.018531469628214836, + 0.023957278579473495, + 0.014692301861941814, + 0.021335560828447342, + 0.01898096688091755, + 0.009589671157300472, + 0.015113906934857368, + 0.020925486460328102, + 0.015186688862740993, + 0.013633234426379204, + 0.016955599188804626, + 0.005082564428448677, + 0.004908778704702854, + 0.00319710117764771, + 0.003580785822123289, + 0.008449268527328968, + 0.004673591814935207, + 0.005587814841419458, + 0.0044235833920538425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.2813586890697479 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010471820132806897 + ], + "opposite_logprob_tokens": [ + -6.876047134399414 + ], + "logprob_reward": -0.0010471820132806897, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010471820132806897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10157819092273712, + 0.0757511705160141, + 0.17148086428642273, + 0.12284781783819199, + 0.07367472350597382, + 0.08676449954509735, + 0.0771622583270073, + 0.14551100134849548, + 0.11332450062036514, + 0.031904980540275574 + ], + "credit_map": [ + 0.012836185283958912, + 0.010600004345178604, + 0.009219981729984283, + 0.01214581448584795, + 0.014271749183535576, + 0.010311413556337357, + 0.018481625244021416, + 0.013711413368582726, + 0.008922604843974113, + 0.010238022543489933, + 0.004981620237231255, + 0.00866700615733862, + 0.014056730084121227, + 0.008081940934062004, + 0.011993774212896824, + 0.00880947895348072, + 0.013820718973875046, + 0.016347358003258705, + 0.008939631283283234, + 0.020571226254105568, + 0.029378972947597504, + 0.018111135810613632, + 0.04211466386914253, + 0.02219715155661106, + 0.012692997232079506, + 0.012851260602474213, + 0.008813080377876759, + 0.015425477176904678, + 0.019817965105175972, + 0.013826940208673477, + 0.025279002264142036, + 0.014141098596155643, + 0.006951893679797649, + 0.007137907203286886, + 0.005265900399535894, + 0.010001921094954014, + 0.012250564992427826, + 0.007947191596031189, + 0.014609766192734241, + 0.009509576484560966, + 0.00882594846189022, + 0.009287587366998196, + 0.006110284477472305, + 0.008274833671748638, + 0.01144073624163866, + 0.009675209410488605, + 0.020212799310684204, + 0.012937100604176521, + 0.01053136121481657, + 0.009749455377459526, + 0.005808583460748196, + 0.008292934857308865, + 0.01207266841083765, + 0.009175139479339123, + 0.011935213580727577, + 0.009596899151802063, + 0.01796986721456051, + 0.012697234749794006, + 0.007858715020120144, + 0.01732730120420456, + 0.022165264934301376, + 0.015341938473284245, + 0.03615100309252739, + 0.015999680384993553, + 0.016368990764021873, + 0.014447740279138088, + 0.00843085814267397, + 0.012911273166537285, + 0.020907647907733917, + 0.013253478333353996, + 0.012953506782650948, + 0.014051010832190514, + 0.004120672587305307, + 0.0029069564770907164, + 0.0036619575694203377, + 0.0033926889300346375, + 0.004265160765498877, + 0.0037525035440921783, + 0.004594884812831879, + 0.005210156086832285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.0010471820132806897 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019446763908490539 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019446763908490539, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019446763908490539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1866142898797989, + 0.1320342719554901, + 0.13344870507717133, + 0.11802058666944504, + 0.11609461903572083, + 0.04681222513318062, + 0.05281517654657364, + 0.09715420752763748, + 0.07034151256084442, + 0.04666437953710556 + ], + "credit_map": [ + 0.01793651469051838, + 0.015156126581132412, + 0.015775101259350777, + 0.024571135640144348, + 0.022158870473504066, + 0.01702408492565155, + 0.04968111962080002, + 0.024311330169439316, + 0.010134580545127392, + 0.012847552075982094, + 0.007562160026282072, + 0.012253416702151299, + 0.03743787854909897, + 0.012767030857503414, + 0.024594631046056747, + 0.01443703006953001, + 0.014522488228976727, + 0.01499137282371521, + 0.01079501025378704, + 0.016381962224841118, + 0.026630299165844917, + 0.013985807076096535, + 0.021901097148656845, + 0.014240667223930359, + 0.01277603954076767, + 0.014478150755167007, + 0.009348316118121147, + 0.014635970816016197, + 0.020489098504185677, + 0.013180267065763474, + 0.020006999373435974, + 0.01310575008392334, + 0.010487372986972332, + 0.009697308763861656, + 0.010194933973252773, + 0.014074304141104221, + 0.021120447665452957, + 0.013438677415251732, + 0.02297152951359749, + 0.014110054820775986, + 0.003816240467131138, + 0.006217926740646362, + 0.003034996334463358, + 0.005475110374391079, + 0.0069340732879936695, + 0.005398935172706842, + 0.010240684263408184, + 0.005694254767149687, + 0.008280429989099503, + 0.005170381627976894, + 0.004006173927336931, + 0.006436720956116915, + 0.006973391864448786, + 0.005925976671278477, + 0.010208780877292156, + 0.00581332016736269, + 0.014162067323923111, + 0.009247721172869205, + 0.006472261622548103, + 0.010903128422796726, + 0.013361305929720402, + 0.01053820364177227, + 0.021940037608146667, + 0.010529481805860996, + 0.009538734331727028, + 0.009233223274350166, + 0.004502380732446909, + 0.006119122263044119, + 0.017841968685388565, + 0.009779958054423332, + 0.007114409003406763, + 0.006211717147380114, + 0.005341117735952139, + 0.006351982709020376, + 0.00438733259215951, + 0.0054018073715269566, + 0.010054466314613819, + 0.005914664827287197, + 0.0048920763656497, + 0.004320928361266851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0019446763908490539 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731689441949129 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731689441949129, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731689441949129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41239774227142334, + 0.06512219458818436, + 0.07625842839479446, + 0.06081739440560341, + 0.08804666996002197, + 0.03326256200671196, + 0.0952552855014801, + 0.0903070718050003, + 0.041988153010606766, + 0.036544475704431534 + ], + "credit_map": [ + 0.12160882353782654, + 0.058105167001485825, + 0.03127209097146988, + 0.030438067391514778, + 0.02644430100917816, + 0.06652610003948212, + 0.04570809379220009, + 0.032295092940330505, + 0.01029790285974741, + 0.008709174580872059, + 0.006345272995531559, + 0.007462944369763136, + 0.00812318455427885, + 0.007011450361460447, + 0.00852786935865879, + 0.008644395507872105, + 0.009935849346220493, + 0.010147924534976482, + 0.005002399440854788, + 0.008462747558951378, + 0.011651498265564442, + 0.010447448119521141, + 0.01179164182394743, + 0.008818923495709896, + 0.008533850312232971, + 0.011310561560094357, + 0.004111276473850012, + 0.00726418849080801, + 0.008426281623542309, + 0.007701749913394451, + 0.006471159867942333, + 0.006998320110142231, + 0.008862587623298168, + 0.011399277485907078, + 0.0062119425274431705, + 0.009572246111929417, + 0.014728809706866741, + 0.010684214532375336, + 0.016518162563443184, + 0.010069435462355614, + 0.004312639590352774, + 0.005376875400543213, + 0.002535704057663679, + 0.003604777855798602, + 0.004817527253180742, + 0.004886192735284567, + 0.004401726648211479, + 0.0033271191641688347, + 0.008676395751535892, + 0.009889883920550346, + 0.004308338742703199, + 0.007355852052569389, + 0.026003185659646988, + 0.013058208860456944, + 0.017579512670636177, + 0.008383912034332752, + 0.010230290703475475, + 0.009006944485008717, + 0.00563547620549798, + 0.010751260444521904, + 0.013279279693961143, + 0.009500603191554546, + 0.020316071808338165, + 0.011587146669626236, + 0.006164427846670151, + 0.004710033070296049, + 0.004531864542514086, + 0.004225252661854029, + 0.007002570666372776, + 0.004601513035595417, + 0.006283355411142111, + 0.004469139501452446, + 0.003446205286309123, + 0.0036532292142510414, + 0.00279602431692183, + 0.002581425243988633, + 0.009311442263424397, + 0.005495141725987196, + 0.005047808866947889, + 0.004213196225464344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.006731689441949129 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, ultra-wide shot, the camera moving slowly and steadily backward, captures a lone figure walking along a straight, dusty path. The figure's features are obscured by the deep hood of a rough, earth-colored cloak, and he moves with a steady, deliberate pace. Each footstep stirs up a lingering puff of pale, fine dust from the cracked earth. Seen from above, a vast geometric grid of towering steel power lines dominates the frame; their intricate, repetitive pattern of towers and sagging cables stretches from the foreground to a vanishing point on the distant horizon, dwarfing the lone traveler. In the evening, harsh golden sunlight casts long shadows across the barren landscape, highlighting the textures of the parched earth and the weathered, rusted metal towers. The color palette is a gentle symphony of ochre, brown, and a hazy, faded blue sky, creating an overwhelming atmosphere of desolation, quiet determination, and an endless journey.", + "seed": 13461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_2326_seed_13461_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11882664213189856, + "vlm_logprob": -0.11882664213189856, + "vlm_yes_no_margin": 4.5499999761581424, + "vlm_token_credit": -0.11882664213189856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20243558287620544 + ], + "opposite_logprob_tokens": [ + -2.2024354934692383 + ], + "logprob_reward": -0.20243558287620544, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.20243558287620544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28068310022354126, + 0.051991283893585205, + 0.09148692339658737, + 0.04932946711778641, + 0.06713485717773438, + 0.20755842328071594, + 0.08837223798036575, + 0.09236334264278412, + 0.04653998836874962, + 0.02454047091305256 + ], + "credit_map": [ + 0.02259182743728161, + 0.02943602204322815, + 0.024264542385935783, + 0.03277301788330078, + 0.04220133647322655, + 0.03738727420568466, + 0.03645089268684387, + 0.05557817593216896, + 0.004671084228903055, + 0.005817857105284929, + 0.0055628870613873005, + 0.007401817478239536, + 0.005666386801749468, + 0.005786197260022163, + 0.01150655560195446, + 0.005578497424721718, + 0.0049598501063883305, + 0.013570626266300678, + 0.005780282896012068, + 0.009359641931951046, + 0.020055411383509636, + 0.012809479609131813, + 0.015883127227425575, + 0.009068501181900501, + 0.0045550307258963585, + 0.006104134023189545, + 0.0036949191708117723, + 0.005970703437924385, + 0.007699578534811735, + 0.00625642342492938, + 0.006868704687803984, + 0.008179972879588604, + 0.007703082170337439, + 0.00937188882380724, + 0.005773015785962343, + 0.0070430622436106205, + 0.009954346343874931, + 0.009241346269845963, + 0.007322858553379774, + 0.010725253261625767, + 0.013163604773581028, + 0.031793590635061264, + 0.009362492710351944, + 0.01597604714334011, + 0.06198872625827789, + 0.035109054297208786, + 0.01872829534113407, + 0.02143661305308342, + 0.004263721406459808, + 0.01305356901139021, + 0.0035063924733549356, + 0.008771270513534546, + 0.02372867241501808, + 0.014212098903954029, + 0.010369677096605301, + 0.010466831736266613, + 0.006155120208859444, + 0.014227500185370445, + 0.004256655462086201, + 0.00823834165930748, + 0.025289595127105713, + 0.015536302700638771, + 0.0082289082929492, + 0.010430912487208843, + 0.0026325727812945843, + 0.006654621101915836, + 0.0033039934933185577, + 0.0037397569976747036, + 0.012489480897784233, + 0.007104760501533747, + 0.004610119387507439, + 0.006004683207720518, + 0.0014699603198096156, + 0.003939444199204445, + 0.0012924306793138385, + 0.002321244217455387, + 0.005545709282159805, + 0.00421613734215498, + 0.0024537041317671537, + 0.0033018419053405523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.20243558287620544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871077299118042 + ], + "opposite_logprob_tokens": [ + -1.1371077299118042 + ], + "logprob_reward": -0.3871077299118042, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871077299118042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052831847220659256, + 0.04205552116036415, + 0.0629788339138031, + 0.1716383695602417, + 0.1759270876646042, + 0.13432684540748596, + 0.2052040547132492, + 0.11483319848775864, + 0.021159959957003593, + 0.019044218584895134 + ], + "credit_map": [ + 0.006045275367796421, + 0.0080990856513381, + 0.004115165676921606, + 0.0061118341982364655, + 0.007513165008276701, + 0.007067135535180569, + 0.006994135212153196, + 0.00688605522736907, + 0.004167153034359217, + 0.0057848370634019375, + 0.005267854314297438, + 0.005689449142664671, + 0.005922555457800627, + 0.005549455527216196, + 0.003758256323635578, + 0.005915962625294924, + 0.004063780419528484, + 0.009515466168522835, + 0.00457265879958868, + 0.006243351846933365, + 0.013725525699555874, + 0.00881488248705864, + 0.006844671443104744, + 0.009198511950671673, + 0.013420234434306622, + 0.030941378325223923, + 0.009231485426425934, + 0.012797589413821697, + 0.04055563732981682, + 0.026741042733192444, + 0.02320968359708786, + 0.014741339720785618, + 0.017579151317477226, + 0.02565252035856247, + 0.012721894308924675, + 0.017470577731728554, + 0.026647450402379036, + 0.02288076840341091, + 0.029929108917713165, + 0.023045629262924194, + 0.007730691693723202, + 0.017791254445910454, + 0.007992277853190899, + 0.01171867921948433, + 0.03614567592740059, + 0.020796746015548706, + 0.013385902158915997, + 0.018765637651085854, + 0.01362086832523346, + 0.017817018553614616, + 0.011816048994660378, + 0.0213175006210804, + 0.031228309497237206, + 0.027549028396606445, + 0.04134182631969452, + 0.04051348194479942, + 0.0080616045743227, + 0.014942127279937267, + 0.006442612502723932, + 0.013075102120637894, + 0.026045219972729683, + 0.016625739634037018, + 0.013008455745875835, + 0.01663235016167164, + 0.0017933835042640567, + 0.002716587856411934, + 0.0015534305712208152, + 0.0022794343531131744, + 0.0042792921885848045, + 0.0028365149628371, + 0.002817164408043027, + 0.0028841570019721985, + 0.0016268473118543625, + 0.0029211826622486115, + 0.0017178638372570276, + 0.0017344868974760175, + 0.0037414291873574257, + 0.0028142216615378857, + 0.00216995133087039, + 0.002318238839507103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.3871077299118042 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001050040009431541 + ], + "opposite_logprob_tokens": [ + -6.876049995422363 + ], + "logprob_reward": -0.001050040009431541, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001050040009431541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09149401634931564, + 0.19145604968070984, + 0.1666269749403, + 0.1168980598449707, + 0.13472656905651093, + 0.09891786426305771, + 0.08683709055185318, + 0.04060249403119087, + 0.05276316776871681, + 0.019677773118019104 + ], + "credit_map": [ + 0.010651622898876667, + 0.0132923424243927, + 0.006045368034392595, + 0.01209022756665945, + 0.01396457850933075, + 0.012141327373683453, + 0.011526055634021759, + 0.011782494373619556, + 0.01043106522411108, + 0.025086740031838417, + 0.021505340933799744, + 0.02892550267279148, + 0.025262676179409027, + 0.02146599441766739, + 0.033608660101890564, + 0.025170080363750458, + 0.009334994480013847, + 0.0169638991355896, + 0.017186708748340607, + 0.022128790616989136, + 0.01976209133863449, + 0.016919197514653206, + 0.04414477199316025, + 0.020186521112918854, + 0.007319246418774128, + 0.012803868390619755, + 0.00863100215792656, + 0.012053982354700565, + 0.01558687724173069, + 0.01357683353126049, + 0.03002690151333809, + 0.016899349167943, + 0.01678656041622162, + 0.015517362393438816, + 0.0109342560172081, + 0.01308359857648611, + 0.01731073297560215, + 0.01409263163805008, + 0.02804148569703102, + 0.018959930166602135, + 0.005317321512848139, + 0.013546808622777462, + 0.005238139070570469, + 0.009989049285650253, + 0.026156850159168243, + 0.015354089438915253, + 0.009273168630897999, + 0.014042433351278305, + 0.0064471117220819, + 0.007330457679927349, + 0.006454877555370331, + 0.00854104571044445, + 0.011066136881709099, + 0.012140888720750809, + 0.01876167580485344, + 0.01609489507973194, + 0.0023316824808716774, + 0.005961290560662746, + 0.0027127733919769526, + 0.004513501189649105, + 0.004347732290625572, + 0.005855032708495855, + 0.00644025718793273, + 0.008440223522484303, + 0.003960023634135723, + 0.007136467378586531, + 0.0035406388342380524, + 0.007844101637601852, + 0.0064329407177865505, + 0.007239038124680519, + 0.006893154699355364, + 0.009716805070638657, + 0.0016941193025559187, + 0.002510243561118841, + 0.001891650608740747, + 0.0026803016662597656, + 0.0021995201241225004, + 0.002416180679574609, + 0.001872599357739091, + 0.004413157235831022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.001050040009431541 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007221474661491811 + ], + "opposite_logprob_tokens": [ + -7.2507219314575195 + ], + "logprob_reward": -0.0007221474661491811, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007221474661491811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25602325797080994, + 0.19838519394397736, + 0.0980340912938118, + 0.08794241398572922, + 0.0926966741681099, + 0.04905664920806885, + 0.10575877875089645, + 0.063289113342762, + 0.024402564391493797, + 0.024411235004663467 + ], + "credit_map": [ + 0.02904168888926506, + 0.03632430359721184, + 0.01943391188979149, + 0.02684384398162365, + 0.027217121794819832, + 0.034221094101667404, + 0.029795562848448753, + 0.05314573273062706, + 0.012074122205376625, + 0.031429070979356766, + 0.015006225556135178, + 0.021166302263736725, + 0.035662535578012466, + 0.02551249787211418, + 0.03408510610461235, + 0.023449325934052467, + 0.006643456872552633, + 0.009648163802921772, + 0.009417110122740269, + 0.013827578164637089, + 0.011265602894127369, + 0.010572012513875961, + 0.024670103564858437, + 0.011990070343017578, + 0.008432203903794289, + 0.010600492358207703, + 0.009392622858285904, + 0.007012990768998861, + 0.010164531879127026, + 0.009168505668640137, + 0.022941024973988533, + 0.01023004949092865, + 0.009232165291905403, + 0.009844932705163956, + 0.007723433896899223, + 0.010981481522321701, + 0.009854343719780445, + 0.009651015512645245, + 0.02166260965168476, + 0.013746690936386585, + 0.004636681172996759, + 0.0058695171028375626, + 0.004470737650990486, + 0.004851857200264931, + 0.008579147048294544, + 0.006582191679626703, + 0.006215051747858524, + 0.007851465605199337, + 0.005703767295926809, + 0.00924302265048027, + 0.006077073514461517, + 0.0102333789691329, + 0.020202133804559708, + 0.01496700756251812, + 0.0207660049200058, + 0.01856638863682747, + 0.0035950627643615007, + 0.00879924651235342, + 0.0032509842421859503, + 0.005201716907322407, + 0.016980374231934547, + 0.01027729269117117, + 0.0069573246873915195, + 0.008227117359638214, + 0.0025582972448319197, + 0.0037811819929629564, + 0.001453568460419774, + 0.002174901543185115, + 0.004957579541951418, + 0.0035189036279916763, + 0.0025132352020591497, + 0.003444898873567581, + 0.002115995390340686, + 0.003602417651563883, + 0.0026738059241324663, + 0.0025199151132255793, + 0.004920728038996458, + 0.0035025468096137047, + 0.0022557838819921017, + 0.0028200424276292324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0007221474661491811 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2532086968421936, + 0.10522294789552689, + 0.24426864087581635, + 0.12384679913520813, + 0.04887859523296356, + 0.06033530831336975, + 0.05306759476661682, + 0.05857714265584946, + 0.01435345970094204, + 0.03824083134531975 + ], + "credit_map": [ + 0.04140974208712578, + 0.04608498886227608, + 0.017717240378260612, + 0.015544255264103413, + 0.0573049820959568, + 0.039995044469833374, + 0.018114803358912468, + 0.01703762449324131, + 0.009713484905660152, + 0.023001713678240776, + 0.008272168226540089, + 0.012609967961907387, + 0.014924535527825356, + 0.017220942303538322, + 0.0085756191983819, + 0.010904517024755478, + 0.019545605406165123, + 0.044806428253650665, + 0.009435804560780525, + 0.023723836988210678, + 0.06906036287546158, + 0.03793790191411972, + 0.01902802288532257, + 0.02073066122829914, + 0.012425577268004417, + 0.020017966628074646, + 0.01528748869895935, + 0.010678447782993317, + 0.021693993359804153, + 0.015317624434828758, + 0.020628459751605988, + 0.007797248661518097, + 0.007267724256962538, + 0.006105280015617609, + 0.004351494833827019, + 0.004271359648555517, + 0.007625728379935026, + 0.006066438276320696, + 0.006867548450827599, + 0.006323018576949835, + 0.003603343153372407, + 0.009611220099031925, + 0.00377651653252542, + 0.004914646968245506, + 0.016778888180851936, + 0.009892932139337063, + 0.005098175257444382, + 0.006659585051238537, + 0.00374218774959445, + 0.0071079060435295105, + 0.003476554760709405, + 0.004258391913026571, + 0.012278883717954159, + 0.008411657996475697, + 0.007217444945126772, + 0.006574571132659912, + 0.00725087383762002, + 0.007623491808772087, + 0.0059410929679870605, + 0.005317742936313152, + 0.010387325659394264, + 0.008597485721111298, + 0.00711583299562335, + 0.0063432990573346615, + 0.0011959295952692628, + 0.0020898431539535522, + 0.0010674266377463937, + 0.001195285003632307, + 0.0029225372709333897, + 0.0019791433587670326, + 0.0016091789584606886, + 0.002294116420671344, + 0.0035325291100889444, + 0.004539442714303732, + 0.0030963041353970766, + 0.004758665803819895, + 0.005416718777269125, + 0.004831238649785519, + 0.0036539591383188963, + 0.008411971852183342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0028177103959023952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, ultra-wide shot, the camera moving slowly and steadily backward, captures a lone figure walking along a straight, dusty path. The figure's features are obscured by the deep hood of a rough, earth-colored cloak, and he moves with a steady, deliberate pace. Each footstep stirs up a lingering puff of pale, fine dust from the cracked earth. Seen from above, a vast geometric grid of towering steel power lines dominates the frame; their intricate, repetitive pattern of towers and sagging cables stretches from the foreground to a vanishing point on the distant horizon, dwarfing the lone traveler. In the evening, harsh golden sunlight casts long shadows across the barren landscape, highlighting the textures of the parched earth and the weathered, rusted metal towers. The color palette is a gentle symphony of ochre, brown, and a hazy, faded blue sky, creating an overwhelming atmosphere of desolation, quiet determination, and an endless journey.", + "seed": 13462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_8/index_2326_seed_13462_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10786571286153049, + "vlm_logprob": -0.10786571286153049, + "vlm_yes_no_margin": 4.199999988079071, + "vlm_token_credit": -0.10786571286153049, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1422051042318344 + ], + "opposite_logprob_tokens": [ + -3.017204999923706 + ], + "logprob_reward": -0.1422051042318344, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.1422051042318344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14891482889652252, + 0.07802118360996246, + 0.2593640089035034, + 0.028196508064866066, + 0.02929837815463543, + 0.10040606558322906, + 0.06809847056865692, + 0.031357575207948685, + 0.0830869972705841, + 0.1732560396194458 + ], + "credit_map": [ + 0.01665770448744297, + 0.019750742241740227, + 0.01192049216479063, + 0.026837723329663277, + 0.019298028200864792, + 0.01770474761724472, + 0.021809669211506844, + 0.01493571326136589, + 0.011582968756556511, + 0.007581288926303387, + 0.016743076965212822, + 0.013205685652792454, + 0.006947943475097418, + 0.008314172737300396, + 0.007255619391798973, + 0.006390427704900503, + 0.05619491636753082, + 0.03602077439427376, + 0.02054251730442047, + 0.0317082516849041, + 0.0330791138112545, + 0.036779891699552536, + 0.023131584748625755, + 0.02190694585442543, + 0.003326419508084655, + 0.003826029133051634, + 0.0025340246502310038, + 0.00360330892726779, + 0.0045660571195185184, + 0.0030385844875127077, + 0.003704593051224947, + 0.0035974886268377304, + 0.00321215926669538, + 0.002955182921141386, + 0.005247332155704498, + 0.0033348584547638893, + 0.002959145698696375, + 0.0031307595781981945, + 0.0048439656384289265, + 0.003614973509684205, + 0.0140544343739748, + 0.009688173420727253, + 0.017515169456601143, + 0.018364878371357918, + 0.009807446040213108, + 0.01197489257901907, + 0.008757496252655983, + 0.010243571363389492, + 0.010111727751791477, + 0.008141990751028061, + 0.007676451466977596, + 0.011313486844301224, + 0.00867383647710085, + 0.007131882011890411, + 0.009675589390099049, + 0.005373505409806967, + 0.0036823060363531113, + 0.003518851473927498, + 0.0037659427616745234, + 0.004048177506774664, + 0.004626949783414602, + 0.00403950922191143, + 0.003947332501411438, + 0.003728504292666912, + 0.01048438623547554, + 0.00905642006546259, + 0.007123053539544344, + 0.011380924843251705, + 0.009658889845013618, + 0.012081204913556576, + 0.014525220729410648, + 0.00877689104527235, + 0.02547541819512844, + 0.019307788461446762, + 0.019537243992090225, + 0.041334714740514755, + 0.01627252623438835, + 0.01831500045955181, + 0.020661145448684692, + 0.01235217321664095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.1422051042318344 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709932565689087 + ], + "opposite_logprob_tokens": [ + -1.137099266052246 + ], + "logprob_reward": -0.38709932565689087, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38709932565689087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07878448069095612, + 0.042366307228803635, + 0.41629675030708313, + 0.019923552870750427, + 0.015444186516106129, + 0.05641555041074753, + 0.029120590537786484, + 0.0760638490319252, + 0.05484712868928909, + 0.21073760092258453 + ], + "credit_map": [ + 0.010875221341848373, + 0.009521923959255219, + 0.008241914212703705, + 0.013979107141494751, + 0.008902394212782383, + 0.007998262532055378, + 0.01133725419640541, + 0.007928402163088322, + 0.006608474534004927, + 0.004122457932680845, + 0.011777298524975777, + 0.006379830650985241, + 0.0036388300359249115, + 0.0024096653796732426, + 0.0033656558953225613, + 0.004064097069203854, + 0.08341250568628311, + 0.05218822881579399, + 0.041107479482889175, + 0.06613919138908386, + 0.04906563088297844, + 0.056833378970623016, + 0.03521009534597397, + 0.03234025835990906, + 0.002187688136473298, + 0.0026831654831767082, + 0.001546358224004507, + 0.001940085319802165, + 0.0032669086940586567, + 0.002113903407007456, + 0.003779173595830798, + 0.002406270010396838, + 0.001609154511243105, + 0.001727294409647584, + 0.0014364148955792189, + 0.0018146911170333624, + 0.001913464511744678, + 0.0018287422135472298, + 0.003712435020133853, + 0.001401989720761776, + 0.006936989724636078, + 0.00657321373000741, + 0.007986676879227161, + 0.006274326238781214, + 0.007602138910442591, + 0.007302288431674242, + 0.007936353795230389, + 0.005803566891700029, + 0.004587478004395962, + 0.003115505212917924, + 0.004859587177634239, + 0.006107561755925417, + 0.002546047791838646, + 0.0032473315950483084, + 0.0027559546288102865, + 0.0019011235563084483, + 0.012209364213049412, + 0.008521423675119877, + 0.006422678008675575, + 0.015035144053399563, + 0.006595739163458347, + 0.009889070875942707, + 0.011184191331267357, + 0.006206240970641375, + 0.006141860969364643, + 0.0063440375961363316, + 0.006424357648938894, + 0.006225175689905882, + 0.008057821542024612, + 0.006203960627317429, + 0.009759057313203812, + 0.005690861027687788, + 0.03919205814599991, + 0.027120552957057953, + 0.022584211081266403, + 0.033533092588186264, + 0.024353601038455963, + 0.02774360217154026, + 0.020711015909910202, + 0.015499458648264408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.38709932565689087 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019471748964861035 + ], + "opposite_logprob_tokens": [ + -6.251947402954102 + ], + "logprob_reward": -0.0019471748964861035, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019471748964861035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08515842258930206, + 0.040105029940605164, + 0.45711401104927063, + 0.012887605465948582, + 0.020197972655296326, + 0.021261025220155716, + 0.01774970255792141, + 0.04181835427880287, + 0.17496496438980103, + 0.12874287366867065 + ], + "credit_map": [ + 0.011845376342535019, + 0.009511299431324005, + 0.015064418315887451, + 0.012638581916689873, + 0.0070413630455732346, + 0.007097828201949596, + 0.01327291876077652, + 0.00868663378059864, + 0.00720421876758337, + 0.005157863721251488, + 0.004461038392037153, + 0.004774259869009256, + 0.00476827472448349, + 0.004520315211266279, + 0.005504928063601255, + 0.0037141332868486643, + 0.08622585982084274, + 0.05362597852945328, + 0.05270598828792572, + 0.07914818823337555, + 0.04810037463903427, + 0.060242220759391785, + 0.04225417971611023, + 0.03481123596429825, + 0.0017547269817441702, + 0.0012998550664633512, + 0.002085310872644186, + 0.0020158400293439627, + 0.0011643976904451847, + 0.0012529561063274741, + 0.0020996234379708767, + 0.0012148951645940542, + 0.002290657488629222, + 0.00218202848918736, + 0.0016368480864912271, + 0.003839245531708002, + 0.0017707027727738023, + 0.002902008593082428, + 0.0039059072732925415, + 0.001670575002208352, + 0.0022537028416991234, + 0.0024760407395660877, + 0.003392247250303626, + 0.0027786954306066036, + 0.003042602678760886, + 0.0017712953267619014, + 0.003281456883996725, + 0.002264986280351877, + 0.003257067408412695, + 0.0021557025611400604, + 0.0019538146443665028, + 0.0032799418549984694, + 0.001690046745352447, + 0.002082176972180605, + 0.002147352322936058, + 0.0011836005141958594, + 0.0031593898311257362, + 0.004175434820353985, + 0.003574541537091136, + 0.006095050368458033, + 0.005316964816302061, + 0.0041611576452851295, + 0.01078071165829897, + 0.00455510662868619, + 0.033275652676820755, + 0.022196263074874878, + 0.019202250987291336, + 0.03264540806412697, + 0.019065169617533684, + 0.022150594741106033, + 0.015293411910533905, + 0.01113621424883604, + 0.020155973732471466, + 0.017900416627526283, + 0.010438257828354836, + 0.018970808014273643, + 0.01617967151105404, + 0.013156390748918056, + 0.020949846133589745, + 0.01099151186645031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.0019471748964861035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259486846625805 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.005259486846625805, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259486846625805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1117645725607872, + 0.01788448542356491, + 0.552293062210083, + 0.014639727771282196, + 0.030637649819254875, + 0.023354222998023033, + 0.017364999279379845, + 0.01317946519702673, + 0.016376961022615433, + 0.20250490307807922 + ], + "credit_map": [ + 0.009953509084880352, + 0.011138980276882648, + 0.00922625046223402, + 0.02789480611681938, + 0.009514695964753628, + 0.008389155380427837, + 0.02461099810898304, + 0.011036159470677376, + 0.0019902659114450216, + 0.0016579311341047287, + 0.0034850467927753925, + 0.0031363230664283037, + 0.0017478616209700704, + 0.0017226791242137551, + 0.002419805619865656, + 0.0017245709896087646, + 0.11114907264709473, + 0.0694330707192421, + 0.05354797840118408, + 0.08647610992193222, + 0.0632447749376297, + 0.07420601695775986, + 0.05086148902773857, + 0.04337450861930847, + 0.0014596945839002728, + 0.00153994036372751, + 0.0015798279782757163, + 0.0018461027648299932, + 0.0018252222798764706, + 0.0018115746788680553, + 0.0026623348239809275, + 0.0019150290172547102, + 0.003754065837711096, + 0.0034321655984967947, + 0.0023243301548063755, + 0.004504399839788675, + 0.0029604171868413687, + 0.004140709061175585, + 0.006644330453127623, + 0.0028772270306944847, + 0.0026197845581918955, + 0.002705714665353298, + 0.0018575344001874328, + 0.0037131458520889282, + 0.0031587709672749043, + 0.003737163497135043, + 0.0031239520758390427, + 0.0024381543044000864, + 0.0023285429924726486, + 0.0018031146610155702, + 0.002551186131313443, + 0.0037899345625191927, + 0.0019204162526875734, + 0.00167734338901937, + 0.0020023086108267307, + 0.0012921516317874193, + 0.0013869339600205421, + 0.001558663439936936, + 0.0014005872653797269, + 0.001411505276337266, + 0.0019436436705291271, + 0.0016673634527251124, + 0.0024231355637311935, + 0.0013876311713829637, + 0.002165167359635234, + 0.001711905119009316, + 0.0013465454103425145, + 0.0025634297635406256, + 0.0018241660436615348, + 0.0016130277654156089, + 0.0033850909676402807, + 0.0017676276620477438, + 0.03670496866106987, + 0.025466313585639, + 0.017247386276721954, + 0.03513297438621521, + 0.022619059309363365, + 0.026149388402700424, + 0.022728750482201576, + 0.016456041485071182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.005259486846625805 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure walking along a straight, dusty path visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a rough, earth-colored cloak with a deep hood obscuring their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vast geometric grid of towering steel power lines visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18960130214691162, + 0.19170917570590973, + 0.08851009607315063, + 0.036517634987831116, + 0.06133577972650528, + 0.13756491243839264, + 0.05399717763066292, + 0.05946623906493187, + 0.049525417387485504, + 0.13177229464054108 + ], + "credit_map": [ + 0.024203768000006676, + 0.021939443424344063, + 0.022515421733260155, + 0.03070097789168358, + 0.019757887348532677, + 0.019686156883835793, + 0.035072553902864456, + 0.01572508178651333, + 0.031322602182626724, + 0.02080213651061058, + 0.030816270038485527, + 0.03376513347029686, + 0.01785963401198387, + 0.022919822484254837, + 0.019720863550901413, + 0.014502698555588722, + 0.009352892637252808, + 0.010173500515520573, + 0.014984886161983013, + 0.017306555062532425, + 0.010964583605527878, + 0.0073790168389678, + 0.010907306335866451, + 0.007441356312483549, + 0.0036991967353969812, + 0.00433345464989543, + 0.005431836470961571, + 0.006887081544846296, + 0.00519402464851737, + 0.0035101748071610928, + 0.004700829274952412, + 0.002761034294962883, + 0.006532036233693361, + 0.006922103464603424, + 0.005032627377659082, + 0.0074662696570158005, + 0.0070381020195782185, + 0.008337793871760368, + 0.014244534075260162, + 0.005762314889580011, + 0.01429675705730915, + 0.014238850213587284, + 0.010367132723331451, + 0.02484513260424137, + 0.016432730481028557, + 0.02125333808362484, + 0.02206559292972088, + 0.014065382070839405, + 0.009838377125561237, + 0.0069711520336568356, + 0.006127048749476671, + 0.009971714578568935, + 0.005468470510095358, + 0.007088729180395603, + 0.005139251705259085, + 0.003392434911802411, + 0.007309278007596731, + 0.006293153390288353, + 0.004355293698608875, + 0.007383345626294613, + 0.007556856609880924, + 0.006881771143525839, + 0.014215543866157532, + 0.005470998119562864, + 0.00951382052153349, + 0.006438045296818018, + 0.0042082504369318485, + 0.005800585262477398, + 0.005442592315375805, + 0.008235572837293148, + 0.005769676994532347, + 0.004116871859878302, + 0.022934023290872574, + 0.015106366947293282, + 0.012040861882269382, + 0.015813598409295082, + 0.016666173934936523, + 0.015943879261612892, + 0.0183198731392622, + 0.014947529882192612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0028174726758152246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_tensor([1979, 1979])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_tensor([1979, 1979])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ca252a03451f9bbac633a9d00091ae3ce6cdd581 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_tensor([1979, 1979])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide, long shot, the camera panning slowly to the right along the broad crest of an ancient city wall from a slightly bird's-eye view, captures a group of children chasing and playing freely. The wall, constructed of massive, mottled gray bricks, is weathered and covered with moss over time. The surface is wide and flat, wide enough for them to run freely. The setting sun casts a warm, golden-orange glow over the world, the sky gradating from fiery orange to soft violet. A boy, wearing a faded blue short-sleeved shirt and beige shorts, leads the group, laughing excitedly as he glances back. Following closely behind is a girl with twin pigtails in a simple white dress, her hands reaching out, intently trying to catch him. Their faces, rosy in the warm glow of the setting sun, are filled with pure joy. Most striking is the dramatic, elongated, deep shadows cast behind them as they run. These shadows leap and dance across the ancient stone pavement, almost as vibrant as the people themselves. In the distance, the dark blue rooftops of the ancient city loom in the twilight, creating a tranquil and poetic scene, imbued with nostalgia and the carefree spirit of childhood.", + "seed": 24068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_1979_seed_24068_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14294294654391707, + "vlm_logprob": -0.14294294654391707, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.14294294654391707, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5884647369384766 + ], + "opposite_logprob_tokens": [ + -0.8384647369384766 + ], + "logprob_reward": -0.5884647369384766, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5884647369384766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1689094603061676, + 0.1113574206829071, + 0.05097944661974907, + 0.0741293728351593, + 0.07950951159000397, + 0.04429005831480026, + 0.08244755119085312, + 0.1785169392824173, + 0.1453261524438858, + 0.06453414261341095 + ], + "credit_map": [ + 0.015077930875122547, + 0.01939762756228447, + 0.017690451815724373, + 0.022717850282788277, + 0.02935068868100643, + 0.02232345938682556, + 0.018137909471988678, + 0.024213522672653198, + 0.01165102794766426, + 0.010529430583119392, + 0.01022846158593893, + 0.016085442155599594, + 0.01738126017153263, + 0.013738664798438549, + 0.020205695182085037, + 0.011537428945302963, + 0.004270407371222973, + 0.004634829703718424, + 0.006190011743456125, + 0.00997767224907875, + 0.007140750996768475, + 0.004840907175093889, + 0.00893646664917469, + 0.004988394677639008, + 0.006746359635144472, + 0.009485033340752125, + 0.009564814157783985, + 0.010043038986623287, + 0.010303087532520294, + 0.008498042821884155, + 0.012994622811675072, + 0.006494363769888878, + 0.004830751568078995, + 0.0053034196607768536, + 0.023602081462740898, + 0.010154466144740582, + 0.006745857652276754, + 0.017319438979029655, + 0.007604948710650206, + 0.003948538098484278, + 0.003242850536480546, + 0.003973404876887798, + 0.005518125835806131, + 0.008946743793785572, + 0.006345320958644152, + 0.004923620726913214, + 0.005784755572676659, + 0.005555231124162674, + 0.004507740028202534, + 0.007277016062289476, + 0.009891883470118046, + 0.013792693614959717, + 0.011204025708138943, + 0.013039267621934414, + 0.013039682991802692, + 0.00969522725790739, + 0.015317153185606003, + 0.048596762120723724, + 0.05696370452642441, + 0.017242902889847755, + 0.013846861198544502, + 0.011092673987150192, + 0.007890858687460423, + 0.007565999869257212, + 0.014250461012125015, + 0.023492248728871346, + 0.013783267699182034, + 0.014079497195780277, + 0.010240412317216396, + 0.04696807637810707, + 0.007888026535511017, + 0.014624149538576603, + 0.006538048852235079, + 0.008148548193275928, + 0.008134161122143269, + 0.004794332664459944, + 0.007646061480045319, + 0.018440252169966698, + 0.0046761250123381615, + 0.00615660659968853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.5884647369384766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207147613167763 + ], + "opposite_logprob_tokens": [ + -2.8120715618133545 + ], + "logprob_reward": -0.06207147613167763, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207147613167763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41167664527893066, + 0.06778700649738312, + 0.042141143232584, + 0.042394913733005524, + 0.06330814212560654, + 0.01698262058198452, + 0.058902591466903687, + 0.12556111812591553, + 0.13074344396591187, + 0.0405023992061615 + ], + "credit_map": [ + 0.06082336977124214, + 0.08776725083589554, + 0.040597669780254364, + 0.07331839948892593, + 0.02696230448782444, + 0.06515323370695114, + 0.03222047910094261, + 0.024833930656313896, + 0.00598955387249589, + 0.006544227246195078, + 0.007988041266798973, + 0.01088715996593237, + 0.009640795178711414, + 0.008117079734802246, + 0.011337594129145145, + 0.007282558362931013, + 0.0032518883235752583, + 0.00474570645019412, + 0.005558612756431103, + 0.008844596333801746, + 0.005226456094533205, + 0.004040094092488289, + 0.006866871379315853, + 0.003606917802244425, + 0.004616798833012581, + 0.003595442743971944, + 0.007965677417814732, + 0.005739410873502493, + 0.004740765783935785, + 0.0074742636643350124, + 0.0054674940183758736, + 0.0027950643561780453, + 0.0047726877965033054, + 0.004230842925608158, + 0.0204880740493536, + 0.007947678677737713, + 0.004325654823333025, + 0.013625390827655792, + 0.004288931377232075, + 0.0036288779228925705, + 0.0027904037851840258, + 0.001395408995449543, + 0.0019382660975679755, + 0.003404738148674369, + 0.001585736172273755, + 0.0022627003490924835, + 0.0019137130584567785, + 0.0016916525783017278, + 0.004135564900934696, + 0.007270104251801968, + 0.007866223342716694, + 0.010304365307092667, + 0.006039035506546497, + 0.00893825851380825, + 0.00768681988120079, + 0.006662219297140837, + 0.01042403094470501, + 0.03535093367099762, + 0.04448087140917778, + 0.013988190330564976, + 0.006239788141101599, + 0.006463083904236555, + 0.004701289813965559, + 0.003912921063601971, + 0.007718001026660204, + 0.0052012065425515175, + 0.052547287195920944, + 0.013298865407705307, + 0.0046122148633003235, + 0.00891733169555664, + 0.00638843746855855, + 0.032060105353593826, + 0.004153325688093901, + 0.004503345582634211, + 0.0055959937162697315, + 0.003696338040754199, + 0.005301176570355892, + 0.011565810069441795, + 0.002650434849783778, + 0.003035977017134428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.06207147613167763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048657048493623734 + ], + "opposite_logprob_tokens": [ + -3.048656940460205 + ], + "logprob_reward": -0.048657048493623734, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048657048493623734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06342065334320068, + 0.031816352158784866, + 0.022043349221348763, + 0.02235054038465023, + 0.07239504903554916, + 0.05079933628439903, + 0.1895216405391693, + 0.1577071100473404, + 0.34835532307624817, + 0.041590578854084015 + ], + "credit_map": [ + 0.00804013293236494, + 0.010981015861034393, + 0.004423428326845169, + 0.011442741379141808, + 0.00578190665692091, + 0.0127524109557271, + 0.005069945473223925, + 0.00492906803265214, + 0.0031721536070108414, + 0.003034882480278611, + 0.003029105719178915, + 0.005252965725958347, + 0.004566441755741835, + 0.0033170480746775866, + 0.005820602644234896, + 0.003623153315857053, + 0.001740410691127181, + 0.0023803310468792915, + 0.002330557443201542, + 0.005353117827326059, + 0.0018465450266376138, + 0.002996521070599556, + 0.0038199606351554394, + 0.0015759048983454704, + 0.00296281767077744, + 0.004236319102346897, + 0.0021896357648074627, + 0.002591331023722887, + 0.0025143292732536793, + 0.003412125399336219, + 0.0024934199173003435, + 0.0019505626987665892, + 0.004533133935183287, + 0.0037122517824172974, + 0.023288095369935036, + 0.00917495135217905, + 0.0036839728709310293, + 0.016838429495692253, + 0.005651200655847788, + 0.005513021256774664, + 0.003133606631308794, + 0.004536787047982216, + 0.005843753460794687, + 0.007520449347794056, + 0.006837442982941866, + 0.004287490155547857, + 0.012151144444942474, + 0.006488665007054806, + 0.012412301264703274, + 0.017739275470376015, + 0.02448461949825287, + 0.033353567123413086, + 0.025592973455786705, + 0.02632685750722885, + 0.025382086634635925, + 0.024229958653450012, + 0.01203837152570486, + 0.04033762589097023, + 0.056376587599515915, + 0.020082399249076843, + 0.007907443679869175, + 0.008333513513207436, + 0.007900232449173927, + 0.0047309217043221, + 0.02044278383255005, + 0.014369325712323189, + 0.13163216412067413, + 0.03654513508081436, + 0.015619735233485699, + 0.019170498475432396, + 0.02127320133149624, + 0.08930248767137527, + 0.0038094944320619106, + 0.0053743100725114346, + 0.0051597473211586475, + 0.004182754550129175, + 0.007324913050979376, + 0.005105545278638601, + 0.006680222228169441, + 0.003953590989112854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.048657048493623734 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013363964390009642 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013363964390009642, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013363964390009642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2571444511413574, + 0.10318978875875473, + 0.04824139177799225, + 0.04951314255595207, + 0.12888504564762115, + 0.06999097019433975, + 0.10668545216321945, + 0.05161565914750099, + 0.1386069357395172, + 0.04612724110484123 + ], + "credit_map": [ + 0.04328211769461632, + 0.07111221551895142, + 0.019705187529325485, + 0.031031867489218712, + 0.0265789981931448, + 0.026791928336024284, + 0.020545141771435738, + 0.01809699460864067, + 0.008262180723249912, + 0.00826974306255579, + 0.014940942637622356, + 0.01770876906812191, + 0.014009619131684303, + 0.011161098256707191, + 0.01792214624583721, + 0.010915280319750309, + 0.0038873841986060143, + 0.004248271696269512, + 0.007082381285727024, + 0.009953703731298447, + 0.006376272067427635, + 0.00389842689037323, + 0.00864680390805006, + 0.004148145671933889, + 0.0048277853056788445, + 0.005814054049551487, + 0.008621465414762497, + 0.0068291667848825455, + 0.006513544823974371, + 0.004777510650455952, + 0.00760189164429903, + 0.004527719225734472, + 0.008200197480618954, + 0.0060309977270662785, + 0.0395091250538826, + 0.021555095911026, + 0.010463660582900047, + 0.024140967056155205, + 0.011013200506567955, + 0.007971802726387978, + 0.008219936862587929, + 0.004698030184954405, + 0.008889374323189259, + 0.012869971804320812, + 0.010786034166812897, + 0.00919501855969429, + 0.009290988557040691, + 0.006041610613465309, + 0.007047334685921669, + 0.011010403744876385, + 0.014270666986703873, + 0.01945614442229271, + 0.011912176385521889, + 0.016548000276088715, + 0.01479018572717905, + 0.011650525964796543, + 0.00415264256298542, + 0.005548940505832434, + 0.006493997760117054, + 0.006061050109565258, + 0.006916670594364405, + 0.005834368988871574, + 0.007959519512951374, + 0.008648463524878025, + 0.013390997424721718, + 0.019318141043186188, + 0.02285456471145153, + 0.01064140722155571, + 0.00973682664334774, + 0.03840658441185951, + 0.007447661366313696, + 0.016810735687613487, + 0.004629749804735184, + 0.006204526871442795, + 0.008017324842512608, + 0.005875277798622847, + 0.005642462521791458, + 0.005795964039862156, + 0.005364157725125551, + 0.004597772844135761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0013363964390009642 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185074716806412 + ], + "opposite_logprob_tokens": [ + -4.264184951782227 + ], + "logprob_reward": -0.014185074716806412, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185074716806412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3330087959766388, + 0.08393387496471405, + 0.054551441222429276, + 0.04541993513703346, + 0.037965934723615646, + 0.05706743523478508, + 0.0944601446390152, + 0.07598403096199036, + 0.18730638921260834, + 0.030302030965685844 + ], + "credit_map": [ + 0.027619587257504463, + 0.050541531294584274, + 0.04504132270812988, + 0.05620170384645462, + 0.025433486327528954, + 0.047721151262521744, + 0.043574269860982895, + 0.03687573969364166, + 0.005957016721367836, + 0.006526767276227474, + 0.013429178856313229, + 0.014436070807278156, + 0.009216401726007462, + 0.010483318008482456, + 0.01399411540478468, + 0.00989101268351078, + 0.004067562520503998, + 0.006771985441446304, + 0.005103047471493483, + 0.008446846157312393, + 0.008970229886472225, + 0.005910775158554316, + 0.009617447853088379, + 0.005663550924509764, + 0.0028063261415809393, + 0.003877163864672184, + 0.009476658888161182, + 0.006450369954109192, + 0.006552783772349358, + 0.008222924545407295, + 0.005150890909135342, + 0.002882818691432476, + 0.0021816131193190813, + 0.002741921227425337, + 0.009669133462011814, + 0.004989316686987877, + 0.0033488008193671703, + 0.007629890460520983, + 0.004923189990222454, + 0.002482071751728654, + 0.01059979572892189, + 0.0038546521682292223, + 0.0056123132817447186, + 0.005234249401837587, + 0.006189071573317051, + 0.010721206665039062, + 0.006910624913871288, + 0.007945522665977478, + 0.006729878950864077, + 0.01034474652260542, + 0.012485266663134098, + 0.015924101695418358, + 0.011975140310823917, + 0.013406114652752876, + 0.012913965620100498, + 0.010680926032364368, + 0.004935724660754204, + 0.015390141867101192, + 0.021650008857250214, + 0.009120680391788483, + 0.006433058530092239, + 0.005735412705689669, + 0.0063774241134524345, + 0.006341576110571623, + 0.011420052498579025, + 0.014794859103858471, + 0.055773619562387466, + 0.015588912181556225, + 0.007827814668416977, + 0.033715926110744476, + 0.008871576748788357, + 0.039313625544309616, + 0.002868765266612172, + 0.003204319393262267, + 0.004327081143856049, + 0.003281558630988002, + 0.004276613239198923, + 0.005930375307798386, + 0.003938403911888599, + 0.0024749154690653086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.014185074716806412 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot, the camera panning slowly to the right along the broad crest of an ancient city wall from a slightly bird's-eye view, captures a group of children chasing and playing freely. The wall, constructed of massive, mottled gray bricks, is weathered and covered with moss over time. The surface is wide and flat, wide enough for them to run freely. The setting sun casts a warm, golden-orange glow over the world, the sky gradating from fiery orange to soft violet. A boy, wearing a faded blue short-sleeved shirt and beige shorts, leads the group, laughing excitedly as he glances back. Following closely behind is a girl with twin pigtails in a simple white dress, her hands reaching out, intently trying to catch him. Their faces, rosy in the warm glow of the setting sun, are filled with pure joy. Most striking is the dramatic, elongated, deep shadows cast behind them as they run. These shadows leap and dance across the ancient stone pavement, almost as vibrant as the people themselves. In the distance, the dark blue rooftops of the ancient city loom in the twilight, creating a tranquil and poetic scene, imbued with nostalgia and the carefree spirit of childhood.", + "seed": 24069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_1979_seed_24069_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15925907636992634, + "vlm_logprob": -0.15925907636992634, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.15925907636992634, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.589671790599823 + ], + "opposite_logprob_tokens": [ + -0.839671790599823 + ], + "logprob_reward": -0.589671790599823, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.589671790599823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.282482385635376, + 0.06795214861631393, + 0.1516903042793274, + 0.08642704039812088, + 0.05603025108575821, + 0.059346914291381836, + 0.05116485431790352, + 0.06781250983476639, + 0.1253606528043747, + 0.05173289030790329 + ], + "credit_map": [ + 0.01505840290337801, + 0.013026675209403038, + 0.022048551589250565, + 0.02055581845343113, + 0.03395697847008705, + 0.04560232535004616, + 0.06424890458583832, + 0.06798472255468369, + 0.006095260847359896, + 0.005305319558829069, + 0.009327051229774952, + 0.007925181649625301, + 0.009443337097764015, + 0.009807126596570015, + 0.008938200771808624, + 0.011110672727227211, + 0.010319261811673641, + 0.015979282557964325, + 0.020431842654943466, + 0.016701187938451767, + 0.017632311210036278, + 0.015637611970305443, + 0.022854845970869064, + 0.03213396295905113, + 0.0051931701600551605, + 0.005923525895923376, + 0.010167742148041725, + 0.007996384054422379, + 0.012712243013083935, + 0.016651947051286697, + 0.018882790580391884, + 0.008899233303964138, + 0.006863823160529137, + 0.009986592456698418, + 0.005809133872389793, + 0.006691764574497938, + 0.009338358417153358, + 0.005943650845438242, + 0.006111063528805971, + 0.005285864695906639, + 0.00523933582007885, + 0.007485765498131514, + 0.008419155143201351, + 0.008541852235794067, + 0.008171807043254375, + 0.006014131009578705, + 0.009311456233263016, + 0.006163412239402533, + 0.0036329308059066534, + 0.0044081611558794975, + 0.005147925112396479, + 0.006895072292536497, + 0.005207375157624483, + 0.007306852377951145, + 0.007320521865040064, + 0.011246017180383205, + 0.005277290008962154, + 0.009051923640072346, + 0.006766295526176691, + 0.008975576609373093, + 0.010441907681524754, + 0.005362080875784159, + 0.011077290400862694, + 0.010860142298042774, + 0.004140659235417843, + 0.006951155140995979, + 0.007156950421631336, + 0.008781126700341702, + 0.008791003376245499, + 0.01120833307504654, + 0.050885070115327835, + 0.027446357533335686, + 0.005180983804166317, + 0.006958661135286093, + 0.0040138219483196735, + 0.006169178057461977, + 0.007343436125665903, + 0.010373864322900772, + 0.005752089433372021, + 0.0059408522211015224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.589671790599823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270606517791748 + ], + "opposite_logprob_tokens": [ + -2.127060651779175 + ], + "logprob_reward": -0.1270606517791748, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270606517791748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19266550242900848, + 0.04712242633104324, + 0.07022791355848312, + 0.22333642840385437, + 0.14424045383930206, + 0.13067024946212769, + 0.0379876010119915, + 0.049649666994810104, + 0.048700135201215744, + 0.05539964511990547 + ], + "credit_map": [ + 0.014062779024243355, + 0.012480317614972591, + 0.01947116106748581, + 0.02708451636135578, + 0.0258327629417181, + 0.03119024634361267, + 0.031626004725694656, + 0.03091771900653839, + 0.00468987924978137, + 0.004521587863564491, + 0.00523914722725749, + 0.004976168274879456, + 0.006564134266227484, + 0.007662428542971611, + 0.006174411159008741, + 0.007294670213013887, + 0.005578345153480768, + 0.008787311613559723, + 0.010123961605131626, + 0.008685596287250519, + 0.008219693787395954, + 0.007901501841843128, + 0.010144318453967571, + 0.010787179693579674, + 0.014668802730739117, + 0.019367065280675888, + 0.024242591112852097, + 0.018571224063634872, + 0.024642961099743843, + 0.05817142128944397, + 0.03478001430630684, + 0.028892330825328827, + 0.010605518706142902, + 0.011700511910021305, + 0.014873003587126732, + 0.014919755049049854, + 0.015607251785695553, + 0.024786991998553276, + 0.03224237263202667, + 0.019505050033330917, + 0.009431756101548672, + 0.011621691286563873, + 0.011461173184216022, + 0.012199263088405132, + 0.01815238781273365, + 0.02367166429758072, + 0.029311414808034897, + 0.014820899814367294, + 0.0020135471131652594, + 0.0027214293368160725, + 0.0034397493582218885, + 0.0038327083457261324, + 0.004444558639079332, + 0.007791721262037754, + 0.005326807498931885, + 0.008417081087827682, + 0.004903472028672695, + 0.0044473144225776196, + 0.005450224503874779, + 0.006121470592916012, + 0.006867602001875639, + 0.006800161208957434, + 0.0065592690370976925, + 0.00850015226751566, + 0.0034663330297917128, + 0.0049148909747600555, + 0.004136030096560717, + 0.004230855032801628, + 0.007013479247689247, + 0.010015252977609634, + 0.008352208882570267, + 0.006571086123585701, + 0.00457935594022274, + 0.006854983512312174, + 0.005999252200126648, + 0.006116877775639296, + 0.009957660920917988, + 0.006293406710028648, + 0.00937263760715723, + 0.006225472316145897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.1270606517791748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370305523276329 + ], + "opposite_logprob_tokens": [ + -3.408703088760376 + ], + "logprob_reward": -0.03370305523276329, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370305523276329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2206183224916458, + 0.07440365105867386, + 0.06194226071238518, + 0.11448495090007782, + 0.1362011432647705, + 0.13258866965770721, + 0.0678374320268631, + 0.055514510720968246, + 0.06956572085618973, + 0.06684327125549316 + ], + "credit_map": [ + 0.011112973093986511, + 0.01502502616494894, + 0.015819784253835678, + 0.020944392308592796, + 0.030701134353876114, + 0.04835362732410431, + 0.04441312327980995, + 0.034248266369104385, + 0.005053625907748938, + 0.006352226249873638, + 0.007075035944581032, + 0.007927941158413887, + 0.011251836083829403, + 0.012796623632311821, + 0.010841377079486847, + 0.013104980811476707, + 0.006559373345226049, + 0.006072771269828081, + 0.009074014611542225, + 0.00593182910233736, + 0.00889199785888195, + 0.007290491834282875, + 0.007777982857078314, + 0.010343796573579311, + 0.01055605337023735, + 0.012144186533987522, + 0.016125811263918877, + 0.011102673597633839, + 0.014435543678700924, + 0.016728553920984268, + 0.020848864689469337, + 0.012543260119855404, + 0.017779657617211342, + 0.019176211208105087, + 0.025112025439739227, + 0.014382743276655674, + 0.022647932171821594, + 0.010988296009600163, + 0.01306528877466917, + 0.013048999011516571, + 0.015651186928153038, + 0.01084856316447258, + 0.01599840633571148, + 0.018158745020627975, + 0.012914253398776054, + 0.011550302617251873, + 0.031949482858181, + 0.015517734922468662, + 0.007575187366455793, + 0.008691755123436451, + 0.008379933424293995, + 0.0073674824088811874, + 0.008417252451181412, + 0.007420036010444164, + 0.009206900373101234, + 0.010778888128697872, + 0.004651040770113468, + 0.005748480558395386, + 0.006948936730623245, + 0.008072562515735626, + 0.00541186798363924, + 0.006401553750038147, + 0.00726904533803463, + 0.011011023074388504, + 0.0033279310446232557, + 0.0049550957046449184, + 0.006511061452329159, + 0.00546868285164237, + 0.007052667438983917, + 0.006575391627848148, + 0.02217624895274639, + 0.013498641550540924, + 0.004827968310564756, + 0.008926229551434517, + 0.004830173682421446, + 0.007672296836972237, + 0.010525419376790524, + 0.009089834988117218, + 0.01168983057141304, + 0.009281513281166553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.03370305523276329 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820919966325164 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.002820919966325164, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820919966325164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15423522889614105, + 0.15554386377334595, + 0.11853958666324615, + 0.10814294219017029, + 0.046505361795425415, + 0.07081698626279831, + 0.15411870181560516, + 0.08469691127538681, + 0.06707420945167542, + 0.04032622650265694 + ], + "credit_map": [ + 0.014931743033230305, + 0.01083894819021225, + 0.017841968685388565, + 0.019373485818505287, + 0.014821675606071949, + 0.024787403643131256, + 0.025628911331295967, + 0.02601110003888607, + 0.01074378751218319, + 0.01524923276156187, + 0.01906169392168522, + 0.01717730052769184, + 0.02426787279546261, + 0.022058749571442604, + 0.02327040210366249, + 0.02371481992304325, + 0.008951612748205662, + 0.015264433808624744, + 0.015569210983812809, + 0.015546762384474277, + 0.01612248457968235, + 0.01482396386563778, + 0.015650484710931778, + 0.016610631719231606, + 0.006603079382330179, + 0.008278405293822289, + 0.012569308280944824, + 0.008154695853590965, + 0.012469940818846226, + 0.0294190663844347, + 0.01574186235666275, + 0.01490657776594162, + 0.00488032354041934, + 0.005970777478069067, + 0.004855609964579344, + 0.004151812754571438, + 0.006276784930378199, + 0.005476521328091621, + 0.009399388916790485, + 0.005494142882525921, + 0.007298662792891264, + 0.0065452042035758495, + 0.008941872045397758, + 0.007857350632548332, + 0.009694002568721771, + 0.009565446525812149, + 0.01280379481613636, + 0.008110652677714825, + 0.011056328192353249, + 0.01815267838537693, + 0.01257045567035675, + 0.02242894284427166, + 0.020282872021198273, + 0.022702835500240326, + 0.02098970301449299, + 0.02593488246202469, + 0.008889001794159412, + 0.013352724723517895, + 0.009184585884213448, + 0.011685699224472046, + 0.014622214250266552, + 0.005990966223180294, + 0.01162338349968195, + 0.009348333813250065, + 0.0027244859375059605, + 0.004564549773931503, + 0.0037225638516247272, + 0.004572260193526745, + 0.008354496210813522, + 0.006619974970817566, + 0.0232979916036129, + 0.013217885047197342, + 0.003937609493732452, + 0.004934919998049736, + 0.0042976378463208675, + 0.003555353032425046, + 0.007810838520526886, + 0.0060522714629769325, + 0.005061381962150335, + 0.0046762144193053246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.002820919966325164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303896427154541 + ], + "opposite_logprob_tokens": [ + -3.168038845062256 + ], + "logprob_reward": -0.04303896427154541, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303896427154541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44673478603363037, + 0.06384007632732391, + 0.10130597651004791, + 0.10817459225654602, + 0.03756808117032051, + 0.041601020842790604, + 0.04482804983854294, + 0.05882295221090317, + 0.0431155264377594, + 0.054008904844522476 + ], + "credit_map": [ + 0.02684488520026207, + 0.024432005360722542, + 0.04057900235056877, + 0.04893555864691734, + 0.06116258352994919, + 0.09997738897800446, + 0.08264485001564026, + 0.062158506363630295, + 0.005415500607341528, + 0.007142444606870413, + 0.0074935611337423325, + 0.008377252146601677, + 0.010999135673046112, + 0.008217744529247284, + 0.008466705679893494, + 0.007727735210210085, + 0.008381078951060772, + 0.011742962524294853, + 0.011875098571181297, + 0.013681773096323013, + 0.008408823050558567, + 0.008934250101447105, + 0.01662347838282585, + 0.021658513695001602, + 0.008829336613416672, + 0.015277443453669548, + 0.017377380281686783, + 0.016347132623195648, + 0.011409364640712738, + 0.010578290559351444, + 0.01749461330473423, + 0.010861021466553211, + 0.0035780721809715033, + 0.004570121876895428, + 0.0034668247681111097, + 0.003767195390537381, + 0.005960746668279171, + 0.004377576056867838, + 0.007392728701233864, + 0.004454813897609711, + 0.004525945521891117, + 0.0040672821924090385, + 0.00552715826779604, + 0.004854027181863785, + 0.005676755215972662, + 0.005006568040698767, + 0.007343565113842487, + 0.004599716980010271, + 0.0027448872569948435, + 0.0034981691278517246, + 0.005345507524907589, + 0.006727767176926136, + 0.0037288726307451725, + 0.00512655358761549, + 0.006639206316322088, + 0.011017085053026676, + 0.005951856728643179, + 0.006681425962597132, + 0.006599347107112408, + 0.008269215002655983, + 0.013195476494729519, + 0.005006819032132626, + 0.0068850042298436165, + 0.006233809981495142, + 0.003013869049027562, + 0.0036988023202866316, + 0.0035606122110038996, + 0.0034893404226750135, + 0.004630954470485449, + 0.006128726992756128, + 0.01156013272702694, + 0.0070330859161913395, + 0.0037562670186161995, + 0.005018582567572594, + 0.0037784406449645758, + 0.00437044957652688, + 0.00791885145008564, + 0.010648069903254509, + 0.011183909140527248, + 0.007334333378821611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04303896427154541 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot, the camera panning slowly to the right along the broad crest of an ancient city wall from a slightly bird's-eye view, captures a group of children chasing and playing freely. The wall, constructed of massive, mottled gray bricks, is weathered and covered with moss over time. The surface is wide and flat, wide enough for them to run freely. The setting sun casts a warm, golden-orange glow over the world, the sky gradating from fiery orange to soft violet. A boy, wearing a faded blue short-sleeved shirt and beige shorts, leads the group, laughing excitedly as he glances back. Following closely behind is a girl with twin pigtails in a simple white dress, her hands reaching out, intently trying to catch him. Their faces, rosy in the warm glow of the setting sun, are filled with pure joy. Most striking is the dramatic, elongated, deep shadows cast behind them as they run. These shadows leap and dance across the ancient stone pavement, almost as vibrant as the people themselves. In the distance, the dark blue rooftops of the ancient city loom in the twilight, creating a tranquil and poetic scene, imbued with nostalgia and the carefree spirit of childhood.", + "seed": 24070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_1979_seed_24070_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16758157163858414, + "vlm_logprob": -0.16758157163858414, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.16758157163858414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.595786452293396 + ], + "opposite_logprob_tokens": [ + -0.845786452293396 + ], + "logprob_reward": -0.595786452293396, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.595786452293396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21044009923934937, + 0.1418413668870926, + 0.12422345578670502, + 0.10089973360300064, + 0.1322791576385498, + 0.11039424687623978, + 0.0454702153801918, + 0.05634884163737297, + 0.034134797751903534, + 0.043968070298433304 + ], + "credit_map": [ + 0.02809068188071251, + 0.03659683093428612, + 0.02364889159798622, + 0.027033230289816856, + 0.014471007511019707, + 0.02088262513279915, + 0.03222149610519409, + 0.02749534510076046, + 0.019676441326737404, + 0.024685891345143318, + 0.00984963122755289, + 0.017613448202610016, + 0.019305787980556488, + 0.01272905059158802, + 0.017047757282853127, + 0.020933356136083603, + 0.01571974717080593, + 0.01912669837474823, + 0.009575320407748222, + 0.011927014216780663, + 0.018586475402116776, + 0.01308701653033495, + 0.01804877631366253, + 0.01815241575241089, + 0.011974955908954144, + 0.010857423767447472, + 0.008737495169043541, + 0.008427095599472523, + 0.011773785576224327, + 0.015168339014053345, + 0.015978936105966568, + 0.017981713637709618, + 0.027815470471978188, + 0.013186384923756123, + 0.01844026893377304, + 0.013922961428761482, + 0.009284420870244503, + 0.013327005319297314, + 0.011028125882148743, + 0.025274522602558136, + 0.012071060948073864, + 0.011247354559600353, + 0.011974054388701916, + 0.02090466395020485, + 0.007614366244524717, + 0.017489377409219742, + 0.012159658595919609, + 0.01693372055888176, + 0.005526408087462187, + 0.003584858262911439, + 0.004562683403491974, + 0.004671063739806414, + 0.00316895660944283, + 0.010402325540781021, + 0.0046080974861979485, + 0.008945820853114128, + 0.006240225862711668, + 0.003971989266574383, + 0.005127891432493925, + 0.006359545513987541, + 0.004870912525802851, + 0.01490013301372528, + 0.007256787735968828, + 0.007621361408382654, + 0.003502012463286519, + 0.0046230703592300415, + 0.004651765804737806, + 0.005674214567989111, + 0.003567886771634221, + 0.004873303696513176, + 0.003195889061316848, + 0.0040466561913490295, + 0.003556372132152319, + 0.004982946906238794, + 0.006372184492647648, + 0.007244466338306665, + 0.004164564423263073, + 0.005190586671233177, + 0.00459370156750083, + 0.007863247767090797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.595786452293396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603977531194687 + ], + "opposite_logprob_tokens": [ + -1.9103977680206299 + ], + "logprob_reward": -0.1603977531194687, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603977531194687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.298661470413208, + 0.10722895711660385, + 0.2982445955276489, + 0.1078910231590271, + 0.01803348958492279, + 0.04081599786877632, + 0.013563853688538074, + 0.023200655356049538, + 0.021848415955901146, + 0.07051155716180801 + ], + "credit_map": [ + 0.04417916014790535, + 0.03524401783943176, + 0.020649490877985954, + 0.02332322858273983, + 0.025280630216002464, + 0.04367808252573013, + 0.04537217319011688, + 0.06093468889594078, + 0.014743022620677948, + 0.011423136107623577, + 0.007734374143183231, + 0.009729556739330292, + 0.011522461660206318, + 0.015913670882582664, + 0.015603413805365562, + 0.020559312775731087, + 0.03502265363931656, + 0.021051418036222458, + 0.016129665076732635, + 0.022644855082035065, + 0.03986819088459015, + 0.047929711639881134, + 0.04981163889169693, + 0.0657864511013031, + 0.012501624412834644, + 0.008551809005439281, + 0.00826938170939684, + 0.011235780082643032, + 0.009887685999274254, + 0.018856436014175415, + 0.0177299901843071, + 0.020858315750956535, + 0.002720580203458667, + 0.002265747170895338, + 0.002412262838333845, + 0.002314049983397126, + 0.0014295371947810054, + 0.0016900500049814582, + 0.0025284362491220236, + 0.002672826172783971, + 0.00631336634978652, + 0.004067524801939726, + 0.0053656878881156445, + 0.005694842431694269, + 0.002053976757451892, + 0.007007529027760029, + 0.0036513449158519506, + 0.00666172755882144, + 0.001936845830641687, + 0.0012063197791576385, + 0.0015361725818365812, + 0.0012704735854640603, + 0.001000239746645093, + 0.002267650794237852, + 0.0012211604043841362, + 0.0031249909661710262, + 0.003774631302803755, + 0.001957009779289365, + 0.002718047471717, + 0.0026046789716929197, + 0.0014303631614893675, + 0.0028269938193261623, + 0.0025113315787166357, + 0.005377599503844976, + 0.0027653526049107313, + 0.004610022064298391, + 0.0021212741266936064, + 0.003750658594071865, + 0.001721277367323637, + 0.003605123609304428, + 0.0015563255874440074, + 0.0017183821182698011, + 0.0055359662510454655, + 0.006815425585955381, + 0.013208971358835697, + 0.02018408291041851, + 0.004314687568694353, + 0.007027080748230219, + 0.006084065418690443, + 0.00734128151088953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1603977531194687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349911466240883 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026349911466240883, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349911466240883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1138165295124054, + 0.05732977017760277, + 0.1167130321264267, + 0.13629208505153656, + 0.13169340789318085, + 0.061256542801856995, + 0.05263202637434006, + 0.20652352273464203, + 0.07963913679122925, + 0.04410398006439209 + ], + "credit_map": [ + 0.015233469195663929, + 0.016836684197187424, + 0.011082769371569157, + 0.011017474345862865, + 0.011286516673862934, + 0.017154183238744736, + 0.01623372733592987, + 0.014971710741519928, + 0.008045087568461895, + 0.00940252561122179, + 0.008351659402251244, + 0.008643319830298424, + 0.004056553356349468, + 0.004772442393004894, + 0.006647017318755388, + 0.007411167491227388, + 0.014052786864340305, + 0.020708171650767326, + 0.010998683981597424, + 0.011660325340926647, + 0.01388365589082241, + 0.015222469344735146, + 0.014664367772638798, + 0.015522565692663193, + 0.015557079575955868, + 0.01745155081152916, + 0.018260270357131958, + 0.021064894273877144, + 0.01037017721682787, + 0.016257185488939285, + 0.019965460523962975, + 0.017365459352731705, + 0.017025921493768692, + 0.027683649212121964, + 0.02216283418238163, + 0.019102564081549644, + 0.00791081041097641, + 0.008413992822170258, + 0.01554287038743496, + 0.01385076530277729, + 0.007717758417129517, + 0.005750598851591349, + 0.009048192761838436, + 0.009056131355464458, + 0.003726152703166008, + 0.009226270020008087, + 0.007635489571839571, + 0.00909594725817442, + 0.00629270076751709, + 0.005211447365581989, + 0.007928014732897282, + 0.00656269071623683, + 0.003580368822440505, + 0.007476864382624626, + 0.00565832294523716, + 0.00992161501199007, + 0.04075274616479874, + 0.014727851375937462, + 0.021810978651046753, + 0.0169926006346941, + 0.007502753287553787, + 0.023733526468276978, + 0.01968993805348873, + 0.06131312996149063, + 0.007496916688978672, + 0.009363747201859951, + 0.009964211843907833, + 0.013545097783207893, + 0.008439980447292328, + 0.010242563672363758, + 0.006319832988083363, + 0.014266785234212875, + 0.004449295345693827, + 0.004671452566981316, + 0.00814406480640173, + 0.007294161710888147, + 0.004092436283826828, + 0.006197064183652401, + 0.004276281222701073, + 0.004979225341230631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.026349911466240883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739741191267967 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739741191267967, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739741191267967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1921754777431488, + 0.13635782897472382, + 0.13137251138687134, + 0.0717681273818016, + 0.035703182220458984, + 0.23945902287960052, + 0.020517630502581596, + 0.06737757474184036, + 0.021847672760486603, + 0.08342096954584122 + ], + "credit_map": [ + 0.028976205736398697, + 0.03394349664449692, + 0.016157491132616997, + 0.022498006001114845, + 0.01811010017991066, + 0.0253007635474205, + 0.019968217238783836, + 0.02722121775150299, + 0.02291213907301426, + 0.030535852536559105, + 0.010031554847955704, + 0.013583789579570293, + 0.021918324753642082, + 0.011639646254479885, + 0.0140366917476058, + 0.01169982273131609, + 0.015545705333352089, + 0.008788932114839554, + 0.0069641172885894775, + 0.009680526331067085, + 0.017012758180499077, + 0.024442033842206, + 0.019505256786942482, + 0.02943318523466587, + 0.007948887534439564, + 0.008445489220321178, + 0.005864836275577545, + 0.008967481553554535, + 0.0056620026007294655, + 0.008787566795945168, + 0.013901010155677795, + 0.01219085045158863, + 0.0038879667408764362, + 0.004112877883017063, + 0.004602012224495411, + 0.0043790885247290134, + 0.0038067596033215523, + 0.005417976528406143, + 0.005101282149553299, + 0.004395219497382641, + 0.029685985296964645, + 0.025835737586021423, + 0.04712017625570297, + 0.04878777638077736, + 0.015042882412672043, + 0.02358247898519039, + 0.0186057947576046, + 0.030798189342021942, + 0.0022385898046195507, + 0.0017393076559528708, + 0.0033496683463454247, + 0.0026269610971212387, + 0.0014035750646144152, + 0.0040618120692670345, + 0.0022267147433012724, + 0.002871001372113824, + 0.012772984802722931, + 0.00432681804522872, + 0.005312138237059116, + 0.005534024443477392, + 0.0035520708188414574, + 0.008065597154200077, + 0.0072207399643957615, + 0.020593196153640747, + 0.0021976931020617485, + 0.0032455166801810265, + 0.0023506791330873966, + 0.0031733294017612934, + 0.0013177403016015887, + 0.0025187560822814703, + 0.0031614673789590597, + 0.003882490796968341, + 0.006814386695623398, + 0.009348146617412567, + 0.014868293888866901, + 0.023547863587737083, + 0.005259133409708738, + 0.008544554933905602, + 0.007664731238037348, + 0.007373860105872154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.006739741191267967 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863400012254715 + ], + "opposite_logprob_tokens": [ + -3.0486340522766113 + ], + "logprob_reward": -0.04863400012254715, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863400012254715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3483780026435852, + 0.08634981513023376, + 0.21647995710372925, + 0.05677508935332298, + 0.03770144283771515, + 0.043855853378772736, + 0.022960670292377472, + 0.05558858439326286, + 0.03662211447954178, + 0.09528849273920059 + ], + "credit_map": [ + 0.03947468847036362, + 0.04648645222187042, + 0.0504104346036911, + 0.04071871191263199, + 0.02882436476647854, + 0.03136444464325905, + 0.05305905640125275, + 0.058039866387844086, + 0.014119025319814682, + 0.01555333100259304, + 0.006270482204854488, + 0.010325322858989239, + 0.0052434965036809444, + 0.008692296221852303, + 0.011444972828030586, + 0.01470088865607977, + 0.027235517278313637, + 0.013165684416890144, + 0.010861453600227833, + 0.01542896032333374, + 0.028292521834373474, + 0.04078405350446701, + 0.0321853831410408, + 0.04852638766169548, + 0.0059053050354123116, + 0.005462986882776022, + 0.004052170552313328, + 0.006770160980522633, + 0.0046297647058963776, + 0.008203072473406792, + 0.010677128098905087, + 0.01107450295239687, + 0.006885241251438856, + 0.004421854391694069, + 0.005812920164316893, + 0.003841299330815673, + 0.002953009447082877, + 0.003858640557155013, + 0.0037226774729788303, + 0.006205798592418432, + 0.0035195809323340654, + 0.0032937314826995134, + 0.00764800887554884, + 0.004896074999123812, + 0.003108459757640958, + 0.011039767414331436, + 0.0039488342590630054, + 0.006401394959539175, + 0.00244327075779438, + 0.0019502420909702778, + 0.0022754017263650894, + 0.0022614162880927324, + 0.0019170630257576704, + 0.006029856391251087, + 0.002504373202100396, + 0.003579044947400689, + 0.01184095535427332, + 0.0038433545269072056, + 0.007325105369091034, + 0.0045823403634130955, + 0.0022334796376526356, + 0.006763070821762085, + 0.005427790805697441, + 0.01357248518615961, + 0.0038936834316700697, + 0.006399267353117466, + 0.00295140384696424, + 0.004522000905126333, + 0.002551102777943015, + 0.003911353647708893, + 0.004143564961850643, + 0.00824973825365305, + 0.00686107762157917, + 0.008655830286443233, + 0.010826465673744678, + 0.025163991376757622, + 0.0066489242017269135, + 0.01195430476218462, + 0.012929899618029594, + 0.012248001992702484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.04863400012254715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot, the camera panning slowly to the right along the broad crest of an ancient city wall from a slightly bird's-eye view, captures a group of children chasing and playing freely. The wall, constructed of massive, mottled gray bricks, is weathered and covered with moss over time. The surface is wide and flat, wide enough for them to run freely. The setting sun casts a warm, golden-orange glow over the world, the sky gradating from fiery orange to soft violet. A boy, wearing a faded blue short-sleeved shirt and beige shorts, leads the group, laughing excitedly as he glances back. Following closely behind is a girl with twin pigtails in a simple white dress, her hands reaching out, intently trying to catch him. Their faces, rosy in the warm glow of the setting sun, are filled with pure joy. Most striking is the dramatic, elongated, deep shadows cast behind them as they run. These shadows leap and dance across the ancient stone pavement, almost as vibrant as the people themselves. In the distance, the dark blue rooftops of the ancient city loom in the twilight, creating a tranquil and poetic scene, imbued with nostalgia and the carefree spirit of childhood.", + "seed": 24071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000012/group_9/index_1979_seed_24071_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12165121454745531, + "vlm_logprob": -0.12165121454745531, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.12165121454745531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3568446934223175 + ], + "opposite_logprob_tokens": [ + -1.2318446636199951 + ], + "logprob_reward": -0.3568446934223175, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3568446934223175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11565391719341278, + 0.08303945511579514, + 0.039745911955833435, + 0.0852547213435173, + 0.04390362277626991, + 0.05943746119737625, + 0.0398760624229908, + 0.43888941407203674, + 0.06196700781583786, + 0.03223245218396187 + ], + "credit_map": [ + 0.01149530429393053, + 0.010766305029392242, + 0.01060700323432684, + 0.017352689057588577, + 0.01030878908932209, + 0.015554468147456646, + 0.021294740960001945, + 0.018274608999490738, + 0.008361605927348137, + 0.00849178433418274, + 0.009974472224712372, + 0.013793906196951866, + 0.007370139472186565, + 0.010527570731937885, + 0.014750681817531586, + 0.00976929534226656, + 0.003579656360670924, + 0.003005803329870105, + 0.003956172615289688, + 0.004970855545252562, + 0.004528635647147894, + 0.009098812006413937, + 0.0056877597235143185, + 0.00491821626201272, + 0.007010978180915117, + 0.009967006742954254, + 0.008079202845692635, + 0.020609483122825623, + 0.009798369370400906, + 0.007927444763481617, + 0.008396239019930363, + 0.013465994037687778, + 0.004596170037984848, + 0.004211653023958206, + 0.003886889200657606, + 0.0059885079972445965, + 0.005064959637820721, + 0.005875058006495237, + 0.006415970623493195, + 0.00786441657692194, + 0.003568233922123909, + 0.0042052678763866425, + 0.0028377757407724857, + 0.006746100261807442, + 0.02068587765097618, + 0.004108084831386805, + 0.0035107664298266172, + 0.013775357976555824, + 0.004553121514618397, + 0.004511811304837465, + 0.0030734091997146606, + 0.003869244595989585, + 0.01009119488298893, + 0.004946371540427208, + 0.003217424266040325, + 0.0056134858168661594, + 0.020121116191148758, + 0.046661797910928726, + 0.11490607261657715, + 0.01545960083603859, + 0.1651000827550888, + 0.02829475700855255, + 0.025914009660482407, + 0.022431975230574608, + 0.005829215981066227, + 0.007558678742498159, + 0.009034475311636925, + 0.007712379097938538, + 0.007765735499560833, + 0.009518437087535858, + 0.007677445653825998, + 0.006870637182146311, + 0.002732794266194105, + 0.0052136583253741264, + 0.0034274940844625235, + 0.0036013892386108637, + 0.0031775333918631077, + 0.004938906989991665, + 0.006265491247177124, + 0.0028751855716109276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.3568446934223175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991216480731964 + ], + "opposite_logprob_tokens": [ + -1.8049122095108032 + ], + "logprob_reward": -0.17991216480731964, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991216480731964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21941477060317993, + 0.08581280708312988, + 0.13613656163215637, + 0.05042368918657303, + 0.09891696274280548, + 0.04680003225803375, + 0.14460746943950653, + 0.05744985118508339, + 0.08202892541885376, + 0.0784088596701622 + ], + "credit_map": [ + 0.015399935655295849, + 0.01635114662349224, + 0.012336667627096176, + 0.02448473498225212, + 0.015170632861554623, + 0.05072510614991188, + 0.028938595205545425, + 0.05600797012448311, + 0.008319171145558357, + 0.008119923062622547, + 0.011668833903968334, + 0.011307865381240845, + 0.006927108857780695, + 0.01780860312283039, + 0.012338140048086643, + 0.009323161095380783, + 0.01065182127058506, + 0.018522894009947777, + 0.008885822258889675, + 0.009024415165185928, + 0.028028497472405434, + 0.020438194274902344, + 0.01420969981700182, + 0.02637522853910923, + 0.004864880815148354, + 0.006985238753259182, + 0.003529130946844816, + 0.009732699021697044, + 0.00508864177390933, + 0.006219921167939901, + 0.006826541852205992, + 0.00717663811519742, + 0.008610482327640057, + 0.011243262328207493, + 0.008086671121418476, + 0.007767690345644951, + 0.013684532605111599, + 0.015594180673360825, + 0.020931297913193703, + 0.012998849153518677, + 0.003566877916455269, + 0.002611675066873431, + 0.0017942136619240046, + 0.003030296415090561, + 0.019440943375229836, + 0.003548914100974798, + 0.0018722891109064221, + 0.010934822261333466, + 0.014658346772193909, + 0.008219419047236443, + 0.00488963583484292, + 0.0036807304713875055, + 0.06075660139322281, + 0.010542694479227066, + 0.006162267178297043, + 0.035697780549526215, + 0.004293606150895357, + 0.007395730819553137, + 0.006470866501331329, + 0.004321436397731304, + 0.008517134934663773, + 0.008713140152394772, + 0.00562223419547081, + 0.012115707620978355, + 0.006405622232705355, + 0.009200993925333023, + 0.0090387724339962, + 0.009011741727590561, + 0.013948867097496986, + 0.009367157705128193, + 0.016613328829407692, + 0.00844244472682476, + 0.007823779247701168, + 0.009574822150170803, + 0.010211233049631119, + 0.0071858554147183895, + 0.00843753106892109, + 0.015946365892887115, + 0.009439162909984589, + 0.009790120646357536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.17991216480731964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809615224599838 + ], + "opposite_logprob_tokens": [ + -3.2880961894989014 + ], + "logprob_reward": -0.03809615224599838, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809615224599838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0845799520611763, + 0.038535699248313904, + 0.27277812361717224, + 0.07669974863529205, + 0.03075985424220562, + 0.08347340673208237, + 0.22507992386817932, + 0.030728986486792564, + 0.031327586621046066, + 0.12603673338890076 + ], + "credit_map": [ + 0.008030351251363754, + 0.005134629085659981, + 0.0046949805691838264, + 0.008919532410800457, + 0.005373666062951088, + 0.018817206844687462, + 0.012194906361401081, + 0.021414674818515778, + 0.0032463176175951958, + 0.004733073525130749, + 0.0053942943923175335, + 0.005279351957142353, + 0.00506471935659647, + 0.005793028045445681, + 0.003718146588653326, + 0.005306764971464872, + 0.021345408633351326, + 0.030957143753767014, + 0.017923355102539062, + 0.024478666484355927, + 0.050631020218133926, + 0.03577372804284096, + 0.02991381101310253, + 0.0617549829185009, + 0.007699941750615835, + 0.006573137827217579, + 0.005873523652553558, + 0.018564283847808838, + 0.006581867579370737, + 0.011684976518154144, + 0.010242141783237457, + 0.00947987474501133, + 0.0028627803549170494, + 0.004740903154015541, + 0.0023402327205985785, + 0.004739385563880205, + 0.0036551502998918295, + 0.004762282129377127, + 0.0033022454008460045, + 0.004356874153017998, + 0.006072322838008404, + 0.008985145948827267, + 0.005001513287425041, + 0.01549991499632597, + 0.021776651963591576, + 0.0052451821975409985, + 0.0072844610549509525, + 0.013608221895992756, + 0.020759379491209984, + 0.012409212067723274, + 0.008637168444693089, + 0.007540825288742781, + 0.09211678802967072, + 0.015186483971774578, + 0.010631791315972805, + 0.057798270136117935, + 0.0023940536193549633, + 0.0025113862939178944, + 0.002803223906084895, + 0.007102080620825291, + 0.003792828181758523, + 0.0034497317392379045, + 0.004303690046072006, + 0.004371993709355593, + 0.0028155434411019087, + 0.005133331287652254, + 0.0035652518272399902, + 0.003613368608057499, + 0.00333631900139153, + 0.0027228991966694593, + 0.0061411624774336815, + 0.0039997100830078125, + 0.012789351865649223, + 0.022040212526917458, + 0.020283248275518417, + 0.0144200399518013, + 0.009737005457282066, + 0.0193696990609169, + 0.01868191547691822, + 0.008715258911252022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.03809615224599838 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11910833418369293, + 0.1294015645980835, + 0.060482390224933624, + 0.06434491276741028, + 0.32122766971588135, + 0.02507142163813114, + 0.07144162058830261, + 0.1293793022632599, + 0.02781037800014019, + 0.05173245072364807 + ], + "credit_map": [ + 0.010863223113119602, + 0.007244332227855921, + 0.009667180478572845, + 0.019168827682733536, + 0.007608909159898758, + 0.02626858465373516, + 0.019055165350437164, + 0.019232098013162613, + 0.014054464176297188, + 0.01252969540655613, + 0.01404963806271553, + 0.019260268658399582, + 0.010815047658979893, + 0.01649642549455166, + 0.019358010962605476, + 0.022838007658720016, + 0.005076741334050894, + 0.008256404660642147, + 0.0039015382062643766, + 0.005901396740227938, + 0.011351258493959904, + 0.005634717643260956, + 0.005943766795098782, + 0.014416559599339962, + 0.006678086705505848, + 0.009093443863093853, + 0.004384355619549751, + 0.013417491689324379, + 0.007529670372605324, + 0.006254934705793858, + 0.007842471823096275, + 0.009144448675215244, + 0.038882531225681305, + 0.049778565764427185, + 0.024405257776379585, + 0.031463682651519775, + 0.0356719009578228, + 0.05794782564043999, + 0.04940316081047058, + 0.03367473557591438, + 0.0031920187175273895, + 0.002638455480337143, + 0.0022334614768624306, + 0.003250685054808855, + 0.005146978888660669, + 0.002766242017969489, + 0.0020035654306411743, + 0.003840015269815922, + 0.007123552728444338, + 0.004133252892643213, + 0.00370734091848135, + 0.003051123348996043, + 0.027985315769910812, + 0.005922263953834772, + 0.0035393729340285063, + 0.015979386866092682, + 0.008213622495532036, + 0.012807605788111687, + 0.030899468809366226, + 0.004472013562917709, + 0.04656517505645752, + 0.00914874766021967, + 0.00758445356041193, + 0.009688202291727066, + 0.0021832804195582867, + 0.002087764209136367, + 0.0033878476824611425, + 0.0034172460436820984, + 0.003058097558096051, + 0.004581154324114323, + 0.005915958900004625, + 0.0031790274661034346, + 0.006616448517888784, + 0.00483997818082571, + 0.013279610313475132, + 0.004890641663223505, + 0.0034999700728803873, + 0.010047903284430504, + 0.004758880473673344, + 0.003799016121774912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.0036151548847556114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide, long shot with a bird's-eye view panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on an ancient city wall made of massive, mottled gray bricks covered with moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a boy in a faded blue short-sleeved shirt and beige shorts and a girl with twin pigtails in a white dress visible chasing each other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a setting sun casting a warm, golden-orange glow with dramatic, elongated deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dark blue rooftops of an ancient city visible in the distance against a sky gradating from fiery orange to soft violet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787907376885414 + ], + "opposite_logprob_tokens": [ + -3.529788017272949 + ], + "logprob_reward": -0.029787907376885414, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787907376885414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20971915125846863, + 0.23919939994812012, + 0.09191824495792389, + 0.05014064535498619, + 0.07679273188114166, + 0.01919335126876831, + 0.04008260741829872, + 0.17874671518802643, + 0.04073856398463249, + 0.05346860736608505 + ], + "credit_map": [ + 0.021680962294340134, + 0.015867363661527634, + 0.018515948206186295, + 0.030781086534261703, + 0.017170803621411324, + 0.029242433607578278, + 0.047201551496982574, + 0.02925899811089039, + 0.023129645735025406, + 0.021887268871068954, + 0.02955826371908188, + 0.04372095689177513, + 0.018576908856630325, + 0.02676583081483841, + 0.04349781200289726, + 0.03206270933151245, + 0.007494558580219746, + 0.010930608958005905, + 0.006087697111070156, + 0.010246257297694683, + 0.014460158534348011, + 0.013119949027895927, + 0.0107958959415555, + 0.018783124163746834, + 0.004338060040026903, + 0.005960867740213871, + 0.005764510948210955, + 0.007096314802765846, + 0.00519242137670517, + 0.008633318357169628, + 0.007179635111242533, + 0.005975520703941584, + 0.010067488066852093, + 0.005657333415001631, + 0.00741664320230484, + 0.016072478145360947, + 0.007855336181819439, + 0.007157586980611086, + 0.012638822197914124, + 0.00992704275995493, + 0.0022571447771042585, + 0.002189723076298833, + 0.001490552444010973, + 0.0034246330615133047, + 0.003944873809814453, + 0.0015315755736082792, + 0.0017808295087888837, + 0.00257401866838336, + 0.003687390126287937, + 0.002860713517293334, + 0.002375636948272586, + 0.003138233907520771, + 0.013372553512454033, + 0.002894006669521332, + 0.002276394050568342, + 0.00947768148034811, + 0.010008363984525204, + 0.01987353153526783, + 0.04238462448120117, + 0.004725612699985504, + 0.07008354365825653, + 0.011207628063857555, + 0.009834781289100647, + 0.010628646239638329, + 0.003509540343657136, + 0.004339908715337515, + 0.005702685099095106, + 0.004011020064353943, + 0.006361953914165497, + 0.005819316022098064, + 0.005572063382714987, + 0.005422077141702175, + 0.0057016476057469845, + 0.009595154784619808, + 0.006512199528515339, + 0.006074865348637104, + 0.004363846965134144, + 0.007727078162133694, + 0.009483159519731998, + 0.004010656848549843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.029787907376885414 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_tensor([9828, 9828])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_tensor([9828, 9828])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b574f12f4e135b95dad9de12d11da4cf60d5fd9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_tensor([9828, 9828])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, majestic drone shot, sweeping in an elegant arc, slowly circles a massive stone Buddha statue carved directly into a towering crimson mountainside. Seated in a lotus position, its massive form almost merges with the mountain. In the golden afterglow of sunset, the weathered stone surface takes on a unique matte texture, reminiscent of the ancient Dunhuang murals. As the camera steadily pans, the Buddha's solemn features gradually reveal themselves: his eyes slightly closed, revealing a deep compassion and introspection. A high nose bridge and full lips frame a serene smile, and a full white hair rests between his brows. The statue's monotonous stone color is not the only hue. Mottled traces of mineral pigments—faded ochre, russet, and a faint turquoise—still linger in the deep folds of the monk's robe and the crevices of his body. These colors appear remarkably soft under the low-angle, warm light, casting long, sharply defined shadows within the carved recesses, dramatically enhancing the statue's three-dimensionality and the sense of its age. The camera hovers steadily, taking in the grandeur of the Buddha statue and the surrounding desolate and magnificent mountain landscape, creating a sacred and peaceful atmosphere that transcends time and space.", + "seed": 19026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_9828_seed_19026_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06701642833650112, + "vlm_logprob": -0.06701642833650112, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.06701642833650112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.129731684923172 + ], + "opposite_logprob_tokens": [ + -2.1297316551208496 + ], + "logprob_reward": -0.129731684923172, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.129731684923172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12231024354696274, + 0.26926881074905396, + 0.05777271091938019, + 0.18251429498195648, + 0.048226166516542435, + 0.028521955013275146, + 0.03421144559979439, + 0.07257293909788132, + 0.13452395796775818, + 0.05007750913500786 + ], + "credit_map": [ + 0.016530191525816917, + 0.010851086117327213, + 0.01196149829775095, + 0.010982693172991276, + 0.01996075175702572, + 0.01751885376870632, + 0.010143113322556019, + 0.02436205744743347, + 0.03730690851807594, + 0.033448319882154465, + 0.027043499052524567, + 0.027077315375208855, + 0.044680919498205185, + 0.02581707574427128, + 0.032907579094171524, + 0.04098718240857124, + 0.010760841891169548, + 0.008392942138016224, + 0.004968722816556692, + 0.007871582172811031, + 0.005758668761700392, + 0.00672647962346673, + 0.005178290884941816, + 0.008115183562040329, + 0.012289679609239101, + 0.007781653665006161, + 0.0144750215113163, + 0.005691405851393938, + 0.004253181628882885, + 0.06407374143600464, + 0.005454404279589653, + 0.06849521398544312, + 0.004065073095262051, + 0.00688764825463295, + 0.003616195172071457, + 0.010913887992501259, + 0.003422769485041499, + 0.005808862391859293, + 0.006785011850297451, + 0.0067267147824168205, + 0.002591855823993683, + 0.004820420406758785, + 0.00547961238771677, + 0.003898242022842169, + 0.002341979881748557, + 0.0031669349409639835, + 0.0028395347762852907, + 0.003383373375982046, + 0.0028017046861350536, + 0.0034947474487125874, + 0.007683744188398123, + 0.0042861695401370525, + 0.004237262532114983, + 0.0033309536520391703, + 0.004901157692074776, + 0.0034757046960294247, + 0.0051580448634922504, + 0.009586332365870476, + 0.007752911653369665, + 0.008860070258378983, + 0.008095535449683666, + 0.008923856541514397, + 0.015301221050322056, + 0.008894966915249825, + 0.008800994604825974, + 0.017336802557110786, + 0.02154095470905304, + 0.016015201807022095, + 0.01976594142615795, + 0.014384041540324688, + 0.01727704517543316, + 0.019402984529733658, + 0.00647687679156661, + 0.0055338735692203045, + 0.007892766036093235, + 0.004968269728124142, + 0.006502022035419941, + 0.004896998405456543, + 0.008271985687315464, + 0.005534714553505182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.129731684923172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496930330991745 + ], + "opposite_logprob_tokens": [ + -2.929969310760498 + ], + "logprob_reward": -0.05496930330991745, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496930330991745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13057884573936462, + 0.1238822266459465, + 0.16989603638648987, + 0.0995093584060669, + 0.0750906765460968, + 0.06295855343341827, + 0.1003245934844017, + 0.1297595202922821, + 0.05054578185081482, + 0.05745438486337662 + ], + "credit_map": [ + 0.015459476038813591, + 0.021337129175662994, + 0.012444130145013332, + 0.013214473612606525, + 0.023879870772361755, + 0.014269592240452766, + 0.012689853087067604, + 0.017284316942095757, + 0.015765240415930748, + 0.015592255629599094, + 0.012247846461832523, + 0.012259408831596375, + 0.015823058784008026, + 0.011690154671669006, + 0.014836137183010578, + 0.025668127462267876, + 0.02183183655142784, + 0.02629675902426243, + 0.013306129723787308, + 0.022658783942461014, + 0.023455359041690826, + 0.021524066105484962, + 0.015333015471696854, + 0.025490082800388336, + 0.007021610625088215, + 0.008521764539182186, + 0.007592641748487949, + 0.006751230452209711, + 0.005387363024055958, + 0.0273623988032341, + 0.005573698785156012, + 0.03129865229129791, + 0.0071840048767626286, + 0.010547267273068428, + 0.005118261557072401, + 0.010941536165773869, + 0.004573178011924028, + 0.014105522073805332, + 0.008795556612312794, + 0.013825349509716034, + 0.005451912060379982, + 0.014549147337675095, + 0.007450944744050503, + 0.00868336670100689, + 0.005700943525880575, + 0.00708187697455287, + 0.007904709316790104, + 0.006135653238743544, + 0.006854419130831957, + 0.013052612543106079, + 0.02109530195593834, + 0.011271909810602665, + 0.009562546387314796, + 0.014945181086659431, + 0.011746028438210487, + 0.011796585284173489, + 0.009273267351090908, + 0.01811240054666996, + 0.01541562844067812, + 0.013222116976976395, + 0.014545106329023838, + 0.022357996553182602, + 0.02339744195342064, + 0.01343555934727192, + 0.004099038429558277, + 0.006052534561604261, + 0.00796967651695013, + 0.004308673553168774, + 0.007430543191730976, + 0.005840195808559656, + 0.007897764444351196, + 0.0069473558105528355, + 0.009413238614797592, + 0.00798958819359541, + 0.008437431417405605, + 0.005535082425922155, + 0.005984690971672535, + 0.006564074661582708, + 0.006235287990421057, + 0.007294992450624704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.05496930330991745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289239034056664 + ], + "opposite_logprob_tokens": [ + -3.773289203643799 + ], + "logprob_reward": -0.023289239034056664, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289239034056664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14300528168678284, + 0.2741469740867615, + 0.08205297589302063, + 0.12095535546541214, + 0.12530303001403809, + 0.028561625629663467, + 0.037911515682935715, + 0.07216203957796097, + 0.05733644962310791, + 0.05856478586792946 + ], + "credit_map": [ + 0.016271866858005524, + 0.01708310842514038, + 0.013141129165887833, + 0.012573392130434513, + 0.026329467073082924, + 0.01697331666946411, + 0.014479093253612518, + 0.026153896003961563, + 0.03550393506884575, + 0.037732694298028946, + 0.029963379725813866, + 0.029264627024531364, + 0.034179091453552246, + 0.02435903809964657, + 0.03394918516278267, + 0.04919502139091492, + 0.013878666795790195, + 0.00903995055705309, + 0.007558455690741539, + 0.011682139709591866, + 0.00681623350828886, + 0.009507606737315655, + 0.007496124133467674, + 0.016073791310191154, + 0.007797863334417343, + 0.007515824865549803, + 0.009039256721735, + 0.004395226016640663, + 0.0036521125584840775, + 0.040362607687711716, + 0.0037139386404305696, + 0.04447852447628975, + 0.008820144459605217, + 0.0100922966375947, + 0.018283329904079437, + 0.012564662843942642, + 0.009668052196502686, + 0.009212078526616096, + 0.03553323820233345, + 0.02112923376262188, + 0.0027493410743772984, + 0.005100993439555168, + 0.00357730221003294, + 0.004108280874788761, + 0.0019166748970746994, + 0.003941109403967857, + 0.0036133022513240576, + 0.0035546226426959038, + 0.002668512286618352, + 0.004878719802945852, + 0.008177574723958969, + 0.0044103325344622135, + 0.002769696293398738, + 0.006657717749476433, + 0.00378577783703804, + 0.00456318398937583, + 0.0053316098637878895, + 0.01041830237954855, + 0.008546371944248676, + 0.010718992911279202, + 0.010671248659491539, + 0.0067789265885949135, + 0.010284916497766972, + 0.009411665610969067, + 0.004036220721900463, + 0.007903983816504478, + 0.007555672898888588, + 0.00689959991723299, + 0.008352233096957207, + 0.00786835141479969, + 0.007823477499186993, + 0.0068969083949923515, + 0.010532003827393055, + 0.008279705420136452, + 0.009206186980009079, + 0.005491033662110567, + 0.005575984250754118, + 0.006040033884346485, + 0.006211817730218172, + 0.007228019647300243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.023289239034056664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038081347942352295 + ], + "opposite_logprob_tokens": [ + -3.288081407546997 + ], + "logprob_reward": -0.038081347942352295, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038081347942352295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10358140617609024, + 0.10773836821317673, + 0.06528345495462418, + 0.22510908544063568, + 0.04544023051857948, + 0.08046799153089523, + 0.26090431213378906, + 0.0568137988448143, + 0.028773892670869827, + 0.025887491181492805 + ], + "credit_map": [ + 0.014481297694146633, + 0.00978586170822382, + 0.01424644235521555, + 0.010759120807051659, + 0.010059445165097713, + 0.017285840585827827, + 0.00965291541069746, + 0.017310483381152153, + 0.013334812596440315, + 0.01433110237121582, + 0.011673929169774055, + 0.012294416315853596, + 0.015219476073980331, + 0.00987642165273428, + 0.0126388780772686, + 0.018369341269135475, + 0.009977644309401512, + 0.007598500233143568, + 0.0054010432213544846, + 0.012723132967948914, + 0.005624200217425823, + 0.00661578169092536, + 0.006162369158118963, + 0.011180787347257137, + 0.015323559753596783, + 0.010086041875183582, + 0.017838608473539352, + 0.008828168734908104, + 0.006881913635879755, + 0.07631201297044754, + 0.007080592680722475, + 0.08275818824768066, + 0.0037599685601890087, + 0.007116217166185379, + 0.0038261935114860535, + 0.007888426072895527, + 0.0053998446092009544, + 0.004916023928672075, + 0.006628616247326136, + 0.005904940888285637, + 0.008235082030296326, + 0.014448686502873898, + 0.015487984754145145, + 0.008734486065804958, + 0.006741657387465239, + 0.007690131198614836, + 0.007515099365264177, + 0.011614865623414516, + 0.019666584208607674, + 0.0290598813444376, + 0.06314130872488022, + 0.03081425465643406, + 0.026379188522696495, + 0.03449403494596481, + 0.02648206613957882, + 0.03086698427796364, + 0.0038330708630383015, + 0.008440990000963211, + 0.006298294756561518, + 0.006443284451961517, + 0.008349773474037647, + 0.00816267542541027, + 0.00870092585682869, + 0.006584787275642157, + 0.002104690996930003, + 0.0030087335035204887, + 0.004465925041586161, + 0.0034824719186872244, + 0.004218816291540861, + 0.00301924510858953, + 0.004085257183760405, + 0.004388753790408373, + 0.0030693490989506245, + 0.00260790903121233, + 0.0046870349906384945, + 0.002805348951369524, + 0.0027910899370908737, + 0.001696016639471054, + 0.004427087493240833, + 0.0038036550395190716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.038081347942352295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890105664730072 + ], + "opposite_logprob_tokens": [ + -2.46401047706604 + ], + "logprob_reward": -0.0890105664730072, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890105664730072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09781262278556824, + 0.41104447841644287, + 0.06612128019332886, + 0.0684211254119873, + 0.04456597939133644, + 0.032226596027612686, + 0.031941428780555725, + 0.0380839966237545, + 0.02234668843448162, + 0.18743574619293213 + ], + "credit_map": [ + 0.012743569910526276, + 0.008786668069660664, + 0.010400270111858845, + 0.007329727988690138, + 0.011516149155795574, + 0.017791274935007095, + 0.008247172459959984, + 0.020997799932956696, + 0.05482533574104309, + 0.051720090210437775, + 0.04426652565598488, + 0.04290451854467392, + 0.06028221175074577, + 0.03768046200275421, + 0.04905889555811882, + 0.0703064575791359, + 0.012384475208818913, + 0.010901900008320808, + 0.00591403478756547, + 0.0071326978504657745, + 0.0065840911120176315, + 0.00782075710594654, + 0.006070875097066164, + 0.009312457405030727, + 0.0048632510006427765, + 0.0033364640548825264, + 0.005174425430595875, + 0.0024365531280636787, + 0.0016469410620629787, + 0.02383209764957428, + 0.002033401047810912, + 0.025097999721765518, + 0.004519541282206774, + 0.006573301274329424, + 0.004314136691391468, + 0.005408768076449633, + 0.00375796714797616, + 0.007322021294385195, + 0.005627595353871584, + 0.007042648270726204, + 0.0026775612495839596, + 0.004559310153126717, + 0.005311603657901287, + 0.004639966879040003, + 0.00295107695274055, + 0.003755909390747547, + 0.003356830682605505, + 0.004974336363375187, + 0.001955987885594368, + 0.005165616981685162, + 0.005795164965093136, + 0.002737303962931037, + 0.003494573524221778, + 0.004975668620318174, + 0.0038549446035176516, + 0.003962168004363775, + 0.0028262094128876925, + 0.0051621925085783005, + 0.004715269897133112, + 0.004720763769000769, + 0.003933402244001627, + 0.006051524542272091, + 0.004492024891078472, + 0.006182608660310507, + 0.00217362679541111, + 0.002518554450944066, + 0.0031486235093325377, + 0.0026489845477044582, + 0.0022653646301478148, + 0.002772840904071927, + 0.0028918671887367964, + 0.003926826640963554, + 0.032932814210653305, + 0.025675732642412186, + 0.028580598533153534, + 0.020271893590688705, + 0.015442186035215855, + 0.020005080848932266, + 0.020067187026143074, + 0.02446025051176548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0890105664730072 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, majestic drone shot, sweeping in an elegant arc, slowly circles a massive stone Buddha statue carved directly into a towering crimson mountainside. Seated in a lotus position, its massive form almost merges with the mountain. In the golden afterglow of sunset, the weathered stone surface takes on a unique matte texture, reminiscent of the ancient Dunhuang murals. As the camera steadily pans, the Buddha's solemn features gradually reveal themselves: his eyes slightly closed, revealing a deep compassion and introspection. A high nose bridge and full lips frame a serene smile, and a full white hair rests between his brows. The statue's monotonous stone color is not the only hue. Mottled traces of mineral pigments—faded ochre, russet, and a faint turquoise—still linger in the deep folds of the monk's robe and the crevices of his body. These colors appear remarkably soft under the low-angle, warm light, casting long, sharply defined shadows within the carved recesses, dramatically enhancing the statue's three-dimensionality and the sense of its age. The camera hovers steadily, taking in the grandeur of the Buddha statue and the surrounding desolate and magnificent mountain landscape, creating a sacred and peaceful atmosphere that transcends time and space.", + "seed": 19027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_9828_seed_19027_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03734454587101936, + "vlm_logprob": -0.03734454587101936, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.03734454587101936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06449320167303085 + ], + "opposite_logprob_tokens": [ + -2.814493179321289 + ], + "logprob_reward": -0.06449320167303085, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06449320167303085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16681748628616333, + 0.1817161589860916, + 0.0775289461016655, + 0.08698373287916183, + 0.049269404262304306, + 0.06228622794151306, + 0.023353343829512596, + 0.018281113356351852, + 0.060889679938554764, + 0.2728739082813263 + ], + "credit_map": [ + 0.020054616034030914, + 0.02084924839437008, + 0.023433469235897064, + 0.015504816547036171, + 0.019356390461325645, + 0.027858193963766098, + 0.019907504320144653, + 0.01985323801636696, + 0.022240150719881058, + 0.021664662286639214, + 0.0265539288520813, + 0.015962105244398117, + 0.02788836881518364, + 0.019278259947896004, + 0.024640953168272972, + 0.02348773553967476, + 0.008333920501172543, + 0.014853325672447681, + 0.014323742128908634, + 0.009103567339479923, + 0.006452908739447594, + 0.004944109823554754, + 0.014681866392493248, + 0.0048355041071772575, + 0.011117140762507915, + 0.012642924673855305, + 0.009017618373036385, + 0.010141384787857533, + 0.012125302106142044, + 0.00931991171091795, + 0.01465990673750639, + 0.007959544658660889, + 0.006922371219843626, + 0.00707238307222724, + 0.00625667255371809, + 0.005142650566995144, + 0.007706927601248026, + 0.005096832290291786, + 0.007404163479804993, + 0.0036674044094979763, + 0.007484887260943651, + 0.006131602916866541, + 0.007794533856213093, + 0.009887686930596828, + 0.00561790494248271, + 0.009137091226875782, + 0.007492126896977425, + 0.00874039251357317, + 0.002305499278008938, + 0.0025420032907277346, + 0.0029817044269293547, + 0.003750270465388894, + 0.0033471330534666777, + 0.0021095057018101215, + 0.00426624296233058, + 0.002050987211987376, + 0.001930407015606761, + 0.002368964022025466, + 0.0022201247047632933, + 0.0021515977568924427, + 0.0025344635359942913, + 0.0015654397429898381, + 0.0030392564367502928, + 0.0024708593264222145, + 0.005126115866005421, + 0.0033084950409829617, + 0.013610566034913063, + 0.004960118792951107, + 0.0035285004414618015, + 0.005746808368712664, + 0.006064360495656729, + 0.018544718623161316, + 0.0637892410159111, + 0.02313908562064171, + 0.04430854320526123, + 0.027815496549010277, + 0.030656715855002403, + 0.02879880741238594, + 0.030240820720791817, + 0.024125216528773308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.06449320167303085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496637150645256 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.05496637150645256, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496637150645256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0729036033153534, + 0.04204742610454559, + 0.03595883771777153, + 0.09742407500743866, + 0.10102798044681549, + 0.0685853436589241, + 0.1887674331665039, + 0.10761591792106628, + 0.04388657212257385, + 0.2417827844619751 + ], + "credit_map": [ + 0.011019707657396793, + 0.009716348722577095, + 0.0097822779789567, + 0.009907986037433147, + 0.007830619812011719, + 0.00717937434092164, + 0.011683878488838673, + 0.00578341307118535, + 0.0038243394810706377, + 0.003580070799216628, + 0.008258428424596786, + 0.005719837732613087, + 0.003651062725111842, + 0.004404901526868343, + 0.009341995231807232, + 0.0032667936757206917, + 0.005817863158881664, + 0.006637521553784609, + 0.0061045014299452305, + 0.004579971078783274, + 0.0031324364244937897, + 0.0034292780328541994, + 0.003047409700229764, + 0.003209860296919942, + 0.007133833598345518, + 0.028351977467536926, + 0.018164945766329765, + 0.010105921886861324, + 0.008839249610900879, + 0.006360189989209175, + 0.012006239965558052, + 0.006461727432906628, + 0.012074366211891174, + 0.014550567604601383, + 0.010426507331430912, + 0.020158851519227028, + 0.012256416492164135, + 0.011499309912323952, + 0.010085627436637878, + 0.009976347908377647, + 0.007609165273606777, + 0.005434911698102951, + 0.008001952432096004, + 0.012711345218122005, + 0.007921325974166393, + 0.007809328380972147, + 0.011146621778607368, + 0.007950699888169765, + 0.016467174515128136, + 0.027858760207891464, + 0.018872741609811783, + 0.027940046042203903, + 0.03389044106006622, + 0.012313720770180225, + 0.03653545305132866, + 0.0148891257122159, + 0.016167450696229935, + 0.00939092319458723, + 0.01548534445464611, + 0.009608627296984196, + 0.018201464787125587, + 0.01574084348976612, + 0.015601662918925285, + 0.007419615518301725, + 0.004897405859082937, + 0.004649599082767963, + 0.009100407361984253, + 0.0030869299080222845, + 0.004783227574080229, + 0.003953895997256041, + 0.0028820496518164873, + 0.010533064603805542, + 0.05656158924102783, + 0.01940183714032173, + 0.038252342492341995, + 0.025254633277654648, + 0.0269510205835104, + 0.026638654991984367, + 0.026561612263321877, + 0.02216111496090889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.05496637150645256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290986195206642 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.023290986195206642, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290986195206642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17871426045894623, + 0.14392651617527008, + 0.07366853207349777, + 0.0999397560954094, + 0.10193757712841034, + 0.02993236854672432, + 0.08764161169528961, + 0.1603011041879654, + 0.04856250807642937, + 0.07537572085857391 + ], + "credit_map": [ + 0.02211141400039196, + 0.023286448791623116, + 0.02531055174767971, + 0.022140318527817726, + 0.019737450405955315, + 0.015005922876298428, + 0.03539026156067848, + 0.015731889754533768, + 0.018800843507051468, + 0.017264897003769875, + 0.018607884645462036, + 0.018765626475214958, + 0.015295004472136497, + 0.011692222207784653, + 0.03046969324350357, + 0.013030343689024448, + 0.008105306886136532, + 0.016852309927344322, + 0.01439755316823721, + 0.00885733775794506, + 0.006272883620113134, + 0.006264922209084034, + 0.005903130397200584, + 0.00701508903875947, + 0.011239312589168549, + 0.024249417707324028, + 0.015450399369001389, + 0.008506172336637974, + 0.01138316560536623, + 0.011550875380635262, + 0.009069934487342834, + 0.008490475825965405, + 0.009714723564684391, + 0.011427586898207664, + 0.01308581791818142, + 0.010915619321167469, + 0.013711821287870407, + 0.013551212847232819, + 0.019332613795995712, + 0.010198179632425308, + 0.0037078980822116137, + 0.0027733035385608673, + 0.0039034292567521334, + 0.004301239736378193, + 0.0035260606091469526, + 0.003617112757638097, + 0.005198092199862003, + 0.002905234694480896, + 0.010726514272391796, + 0.010737131349742413, + 0.009144645184278488, + 0.011843412183225155, + 0.011323190294206142, + 0.007030286826193333, + 0.013993474654853344, + 0.012842955067753792, + 0.021010279655456543, + 0.013317940756678581, + 0.0253361277282238, + 0.01380406878888607, + 0.022358965128660202, + 0.03303812816739082, + 0.021517282351851463, + 0.00991831999272108, + 0.004880399908870459, + 0.004291339311748743, + 0.007079428527504206, + 0.005170386750251055, + 0.00495722284540534, + 0.007040349300950766, + 0.0064692022278904915, + 0.008674174547195435, + 0.01642751879990101, + 0.008193974383175373, + 0.010447192937135696, + 0.007590731140226126, + 0.006553140468895435, + 0.009173691272735596, + 0.007397310808300972, + 0.009592157788574696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023290986195206642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978721261024475 + ], + "opposite_logprob_tokens": [ + -3.529787302017212 + ], + "logprob_reward": -0.02978721261024475, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978721261024475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09959524869918823, + 0.09035886079072952, + 0.16978183388710022, + 0.05628709867596626, + 0.09132423251867294, + 0.054020848125219345, + 0.05908089131116867, + 0.0918048620223999, + 0.19127701222896576, + 0.09646908193826675 + ], + "credit_map": [ + 0.01253348495811224, + 0.012430526316165924, + 0.013075371272861958, + 0.01330506894737482, + 0.009673860855400562, + 0.009891211986541748, + 0.017644492909312248, + 0.011041241697967052, + 0.008284792304039001, + 0.009749260731041431, + 0.016312189400196075, + 0.013041161932051182, + 0.009239706210792065, + 0.007561365142464638, + 0.018731489777565002, + 0.007438899017870426, + 0.017253214493393898, + 0.04202670231461525, + 0.0399542897939682, + 0.021654725074768066, + 0.014013233594596386, + 0.010466882027685642, + 0.015494514256715775, + 0.00891829002648592, + 0.005400771740823984, + 0.006771270185709, + 0.006428417284041643, + 0.009628999046981335, + 0.006623372435569763, + 0.006033807527273893, + 0.010321259498596191, + 0.005079201888293028, + 0.013751358725130558, + 0.00894543994218111, + 0.011776087805628777, + 0.01237483974546194, + 0.01161335501819849, + 0.008173619396984577, + 0.014956307597458363, + 0.009733228012919426, + 0.006721421144902706, + 0.005629474297165871, + 0.006686629261821508, + 0.009053600020706654, + 0.007305706851184368, + 0.006976539269089699, + 0.005778043530881405, + 0.005869435146450996, + 0.006757026072591543, + 0.0068373242393136024, + 0.005636878311634064, + 0.007140908390283585, + 0.00834089145064354, + 0.004874737933278084, + 0.01254546269774437, + 0.006947664078325033, + 0.008775869384407997, + 0.0066345082595944405, + 0.017785584554076195, + 0.006554159801453352, + 0.007796559017151594, + 0.02583339810371399, + 0.012209661304950714, + 0.0062151215970516205, + 0.016496462747454643, + 0.02169431746006012, + 0.030802778899669647, + 0.018261026591062546, + 0.024448702111840248, + 0.020746802911162376, + 0.01980801299214363, + 0.0390189029276371, + 0.021346187219023705, + 0.012478276155889034, + 0.014254064299166203, + 0.00890672579407692, + 0.010025971569120884, + 0.009640123695135117, + 0.011881153099238873, + 0.007936582900583744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.02978721261024475 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418495737016201 + ], + "opposite_logprob_tokens": [ + -4.264184951782227 + ], + "logprob_reward": -0.01418495737016201, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418495737016201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3088849186897278, + 0.10115109384059906, + 0.03487091138958931, + 0.044658441096544266, + 0.08299319446086884, + 0.04409213364124298, + 0.04790804535150528, + 0.06977079808712006, + 0.03721955046057701, + 0.2284509390592575 + ], + "credit_map": [ + 0.06072612479329109, + 0.03864079341292381, + 0.04023045301437378, + 0.03965412452816963, + 0.022806325927376747, + 0.03604268282651901, + 0.04354339838027954, + 0.027241021394729614, + 0.01670682802796364, + 0.010073542594909668, + 0.013821770437061787, + 0.011890879832208157, + 0.01264076679944992, + 0.007874478586018085, + 0.020184393972158432, + 0.007958441972732544, + 0.003378787776455283, + 0.0076797129586339, + 0.008487686514854431, + 0.0042456998489797115, + 0.0030917061958462, + 0.0024446516763418913, + 0.0034623127430677414, + 0.002080355305224657, + 0.004325537476688623, + 0.00545906089246273, + 0.005929138977080584, + 0.0064685530960559845, + 0.00615786574780941, + 0.004153441172093153, + 0.008622005581855774, + 0.0035428402479737997, + 0.010327911004424095, + 0.01117737963795662, + 0.008456671610474586, + 0.015856528654694557, + 0.008531816303730011, + 0.009640962816774845, + 0.009864065796136856, + 0.00913785770535469, + 0.005273665767163038, + 0.0038822663482278585, + 0.005383396986871958, + 0.007565164007246494, + 0.004470464773476124, + 0.004948773887008429, + 0.0070658582262694836, + 0.005502542480826378, + 0.007156155537813902, + 0.004960744176059961, + 0.005202673841267824, + 0.005103856325149536, + 0.007513534277677536, + 0.00447390042245388, + 0.006440413184463978, + 0.007056770380586386, + 0.0091429203748703, + 0.005968431942164898, + 0.011426247656345367, + 0.0059885564260184765, + 0.009988336823880672, + 0.014391153119504452, + 0.008588786236941814, + 0.004276372026652098, + 0.0033356742933392525, + 0.002092022681608796, + 0.007805865723639727, + 0.0030902281869202852, + 0.003145017195492983, + 0.004153393674641848, + 0.004234131891280413, + 0.009363219141960144, + 0.04731181636452675, + 0.031093847006559372, + 0.03166057541966438, + 0.024689385667443275, + 0.020682552829384804, + 0.020984390750527382, + 0.03142184019088745, + 0.020606499165296555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01418495737016201 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, majestic drone shot, sweeping in an elegant arc, slowly circles a massive stone Buddha statue carved directly into a towering crimson mountainside. Seated in a lotus position, its massive form almost merges with the mountain. In the golden afterglow of sunset, the weathered stone surface takes on a unique matte texture, reminiscent of the ancient Dunhuang murals. As the camera steadily pans, the Buddha's solemn features gradually reveal themselves: his eyes slightly closed, revealing a deep compassion and introspection. A high nose bridge and full lips frame a serene smile, and a full white hair rests between his brows. The statue's monotonous stone color is not the only hue. Mottled traces of mineral pigments—faded ochre, russet, and a faint turquoise—still linger in the deep folds of the monk's robe and the crevices of his body. These colors appear remarkably soft under the low-angle, warm light, casting long, sharply defined shadows within the carved recesses, dramatically enhancing the statue's three-dimensionality and the sense of its age. The camera hovers steadily, taking in the grandeur of the Buddha statue and the surrounding desolate and magnificent mountain landscape, creating a sacred and peaceful atmosphere that transcends time and space.", + "seed": 19028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_9828_seed_19028_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03194810994900763, + "vlm_logprob": -0.03194810994900763, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03194810994900763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033789940644055605 + ], + "opposite_logprob_tokens": [ + -6.128378868103027 + ], + "logprob_reward": -0.0033789940644055605, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0033789940644055605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13275714218616486, + 0.05971986800432205, + 0.052470818161964417, + 0.03298071026802063, + 0.01668170653283596, + 0.03930846229195595, + 0.1974629908800125, + 0.10994669049978256, + 0.12688934803009033, + 0.23178230226039886 + ], + "credit_map": [ + 0.016745980829000473, + 0.014757661148905754, + 0.022826246917247772, + 0.031050659716129303, + 0.011021394282579422, + 0.008554182015359402, + 0.019441785290837288, + 0.008359214290976524, + 0.005970851052552462, + 0.006308319512754679, + 0.01470522303134203, + 0.011392437852919102, + 0.005068821832537651, + 0.004127020016312599, + 0.007693973369896412, + 0.004453218076378107, + 0.002798435976728797, + 0.005551447160542011, + 0.0059961313381791115, + 0.009333021938800812, + 0.0073068877682089806, + 0.005571863614022732, + 0.012380574829876423, + 0.0035324476193636656, + 0.002371330978348851, + 0.004350904840976, + 0.00417869770899415, + 0.004591621924191713, + 0.005086191929876804, + 0.0038878193590790033, + 0.005403163377195597, + 0.0031109782867133617, + 0.002244079951196909, + 0.002288983203470707, + 0.0021218606270849705, + 0.0020894480403512716, + 0.0023405123502016068, + 0.0014436024939641356, + 0.0029062312096357346, + 0.0012469885405153036, + 0.003779742633923888, + 0.006066149100661278, + 0.00492411432787776, + 0.005069354083389044, + 0.005958911031484604, + 0.0025599109940230846, + 0.008763397112488747, + 0.002186877653002739, + 0.016775989904999733, + 0.010186576284468174, + 0.06551704555749893, + 0.011934082955121994, + 0.007586716674268246, + 0.013812833465635777, + 0.011374903842806816, + 0.06027483195066452, + 0.009058688767254353, + 0.013488451018929482, + 0.015493998304009438, + 0.02135959453880787, + 0.008986814878880978, + 0.00694591598585248, + 0.02434740774333477, + 0.01026580948382616, + 0.015447077341377735, + 0.02176908403635025, + 0.025085581466555595, + 0.01308977510780096, + 0.008738051168620586, + 0.011532769538462162, + 0.012743406929075718, + 0.0184836033731699, + 0.025249993428587914, + 0.011033719405531883, + 0.05675250664353371, + 0.01168337557464838, + 0.006253348663449287, + 0.011284217238426208, + 0.01687443256378174, + 0.09265068173408508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0033789940644055605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904840797185898 + ], + "opposite_logprob_tokens": [ + -2.464048385620117 + ], + "logprob_reward": -0.08904840797185898, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904840797185898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11567674577236176, + 0.05966847389936447, + 0.057250361889600754, + 0.02302323840558529, + 0.03134339302778244, + 0.029320882633328438, + 0.17253412306308746, + 0.10116341710090637, + 0.18849849700927734, + 0.22152088582515717 + ], + "credit_map": [ + 0.014103585854172707, + 0.010226423852145672, + 0.025910664349794388, + 0.02708417922258377, + 0.008836708031594753, + 0.007763992995023727, + 0.015194370411336422, + 0.0065568238496780396, + 0.004532691556960344, + 0.006047292612493038, + 0.009879914112389088, + 0.011068250052630901, + 0.006642181891947985, + 0.005851218942552805, + 0.010507176630198956, + 0.005139749497175217, + 0.0034033367410302162, + 0.0057629309594631195, + 0.0076291244477033615, + 0.009778546169400215, + 0.008096749894320965, + 0.006049878895282745, + 0.012038425542414188, + 0.004491370636969805, + 0.0023066133726388216, + 0.0025024444330483675, + 0.0027158865705132484, + 0.0033147830981761217, + 0.0028636064380407333, + 0.0030416122172027826, + 0.0035016778856515884, + 0.0027766136918216944, + 0.003141976660117507, + 0.004606036935001612, + 0.004281173460185528, + 0.004163918551057577, + 0.003701025154441595, + 0.002950705587863922, + 0.00629399111494422, + 0.0022045685909688473, + 0.0026772681158035994, + 0.003315956564620137, + 0.004547590389847755, + 0.004296872764825821, + 0.0036648239474743605, + 0.0018755297642201185, + 0.006807307247072458, + 0.002135532908141613, + 0.02001333422958851, + 0.01258818432688713, + 0.022045563906431198, + 0.018421584740281105, + 0.009734652936458588, + 0.01517071109265089, + 0.018888326361775398, + 0.05567177012562752, + 0.011234022676944733, + 0.013383070006966591, + 0.025101471692323685, + 0.012733533047139645, + 0.007610724773257971, + 0.006309082265943289, + 0.01591416262090206, + 0.008877351880073547, + 0.01935737021267414, + 0.035612668842077255, + 0.03464193642139435, + 0.016312457621097565, + 0.019031405448913574, + 0.0174015574157238, + 0.020303547382354736, + 0.025837553665041924, + 0.024647828191518784, + 0.02351815067231655, + 0.021336061879992485, + 0.018107395619153976, + 0.01238945871591568, + 0.026196258142590523, + 0.035639286041259766, + 0.05968645215034485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.08904840797185898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809546306729317 + ], + "opposite_logprob_tokens": [ + -3.288095474243164 + ], + "logprob_reward": -0.03809546306729317, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809546306729317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12664319574832916, + 0.09612662345170975, + 0.08923231065273285, + 0.04915750399231911, + 0.038155317306518555, + 0.02251334860920906, + 0.2093495875597, + 0.028232155367732048, + 0.12260495871305466, + 0.21798501908779144 + ], + "credit_map": [ + 0.011016514152288437, + 0.013239776715636253, + 0.017219072207808495, + 0.020872244611382484, + 0.013223288580775261, + 0.00967063382267952, + 0.03253389522433281, + 0.008867775090038776, + 0.0057364571839571, + 0.009250165894627571, + 0.015479560941457748, + 0.01859252341091633, + 0.011102026328444481, + 0.0061501264572143555, + 0.02314060367643833, + 0.006675161886960268, + 0.004602927714586258, + 0.00865876954048872, + 0.01138221099972725, + 0.014984812587499619, + 0.013655354268848896, + 0.009660356678068638, + 0.01981949806213379, + 0.006468380801379681, + 0.004015209153294563, + 0.005465660709887743, + 0.006442535202950239, + 0.008957347832620144, + 0.008250974118709564, + 0.0038130113389343023, + 0.009005866944789886, + 0.0032068954315036535, + 0.003934594802558422, + 0.0036437332164496183, + 0.006132013164460659, + 0.005045739468187094, + 0.00539137190207839, + 0.0037235585041344166, + 0.007303453516215086, + 0.0029808508697897196, + 0.002182879252359271, + 0.0026336379814893007, + 0.0030550851952284575, + 0.0024747755378484726, + 0.003387817647308111, + 0.0017771224956959486, + 0.00540586793795228, + 0.0015961633762344718, + 0.02179671637713909, + 0.01495889201760292, + 0.0319875068962574, + 0.014621026813983917, + 0.008937687613070011, + 0.020841429010033607, + 0.018634503707289696, + 0.0775718241930008, + 0.0031606359407305717, + 0.0038157617673277855, + 0.004479756113141775, + 0.00291747460141778, + 0.0029100971296429634, + 0.0025229284074157476, + 0.005552970804274082, + 0.0028725331649184227, + 0.010468454100191593, + 0.016695231199264526, + 0.021071994677186012, + 0.01619665138423443, + 0.014526762999594212, + 0.009099820628762245, + 0.019501106813549995, + 0.015044941566884518, + 0.02769242785871029, + 0.02450620010495186, + 0.020647667348384857, + 0.017529327422380447, + 0.011509372852742672, + 0.01824241876602173, + 0.027406105771660805, + 0.07045150548219681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.03809546306729317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274913430213928 + ], + "opposite_logprob_tokens": [ + -3.7732748985290527 + ], + "logprob_reward": -0.023274913430213928, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274913430213928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04888938367366791, + 0.024815021082758904, + 0.03460942953824997, + 0.033610571175813675, + 0.05687401443719864, + 0.04279549792408943, + 0.11723420768976212, + 0.18524520099163055, + 0.17594486474990845, + 0.279981791973114 + ], + "credit_map": [ + 0.004481889307498932, + 0.004680203273892403, + 0.01010109018534422, + 0.008647864684462547, + 0.004351561889052391, + 0.0036107292398810387, + 0.00979218166321516, + 0.0032238629646599293, + 0.001787968329153955, + 0.0026594826485961676, + 0.00358714465983212, + 0.005274457391351461, + 0.0026020335499197245, + 0.0019401366589590907, + 0.005231074057519436, + 0.0017327237874269485, + 0.0025887659285217524, + 0.0034735368099063635, + 0.005504774395376444, + 0.0052629150450229645, + 0.004471207968890667, + 0.0028541286010295153, + 0.008217832073569298, + 0.0022362666204571724, + 0.002531675389036536, + 0.0041665127500891685, + 0.004438410047441721, + 0.005055954214185476, + 0.00534096360206604, + 0.003045093733817339, + 0.006340974010527134, + 0.0026909871958196163, + 0.0045594340190291405, + 0.010478794574737549, + 0.006852747872471809, + 0.007221605628728867, + 0.008528857491910458, + 0.0042090993374586105, + 0.011310315690934658, + 0.003713158657774329, + 0.0033494196832180023, + 0.0065138861536979675, + 0.005842226091772318, + 0.006797879934310913, + 0.00722092017531395, + 0.002952168695628643, + 0.007222775369882584, + 0.0028962206561118364, + 0.014566672034561634, + 0.00617310730740428, + 0.04191043600440025, + 0.00918427761644125, + 0.003643123898655176, + 0.007065974175930023, + 0.009004058316349983, + 0.025686560198664665, + 0.019177978858351707, + 0.015125860460102558, + 0.062365949153900146, + 0.01758236438035965, + 0.01260422170162201, + 0.012217722833156586, + 0.020974505692720413, + 0.025196613743901253, + 0.01946934498846531, + 0.03604770451784134, + 0.021997570991516113, + 0.024632995948195457, + 0.015237974934279919, + 0.01902303658425808, + 0.025100698694586754, + 0.014435549266636372, + 0.02828516997396946, + 0.01958354003727436, + 0.0592915341258049, + 0.017343124374747276, + 0.00920049287378788, + 0.01639619655907154, + 0.024501511827111244, + 0.10538021475076675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.023274913430213928 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942771211266518 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942771211266518, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942771211266518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1925630122423172, + 0.07063959538936615, + 0.06520885974168777, + 0.04583413898944855, + 0.02633158676326275, + 0.03618347644805908, + 0.1981400102376938, + 0.12631620466709137, + 0.128043532371521, + 0.11073961853981018 + ], + "credit_map": [ + 0.018162058666348457, + 0.030400870367884636, + 0.035055335611104965, + 0.030778607353568077, + 0.01707233488559723, + 0.01579681597650051, + 0.030519381165504456, + 0.014777597039937973, + 0.00928743276745081, + 0.00614952202886343, + 0.014107788912951946, + 0.013494308106601238, + 0.007934481836855412, + 0.0041847238317132, + 0.011231105774641037, + 0.004250233061611652, + 0.005674256943166256, + 0.006354480050504208, + 0.011959775350987911, + 0.010445328429341316, + 0.007163769565522671, + 0.007132700644433498, + 0.01181404571980238, + 0.004664503503590822, + 0.004369724541902542, + 0.00510042579844594, + 0.007196271326392889, + 0.006358269136399031, + 0.006258595734834671, + 0.0045762937515974045, + 0.007758779916912317, + 0.00421577924862504, + 0.0028825297486037016, + 0.004149141255766153, + 0.0036777020432054996, + 0.003637848421931267, + 0.003249183762818575, + 0.0024074672255665064, + 0.004336187615990639, + 0.0019915285520255566, + 0.003054764587432146, + 0.00505074393004179, + 0.004858736880123615, + 0.00639467453584075, + 0.003715292550623417, + 0.004072984214872122, + 0.005950485821813345, + 0.003085796721279621, + 0.03154953569173813, + 0.02190045826137066, + 0.02666916325688362, + 0.018277674913406372, + 0.012473933398723602, + 0.015713606029748917, + 0.02081938274204731, + 0.05073625221848488, + 0.013042637147009373, + 0.013530722819268703, + 0.03461851179599762, + 0.01804523915052414, + 0.00869777612388134, + 0.008289031684398651, + 0.017066869884729385, + 0.013025418855249882, + 0.013592377305030823, + 0.027036253362894058, + 0.017554469406604767, + 0.015505759976804256, + 0.01551831979304552, + 0.012310169637203217, + 0.015596232376992702, + 0.010929937474429607, + 0.012975933961570263, + 0.01941663585603237, + 0.012142142280936241, + 0.011904289945960045, + 0.0076125687919557095, + 0.010427197441458702, + 0.018097275868058205, + 0.01816357672214508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.005942771211266518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, majestic drone shot, sweeping in an elegant arc, slowly circles a massive stone Buddha statue carved directly into a towering crimson mountainside. Seated in a lotus position, its massive form almost merges with the mountain. In the golden afterglow of sunset, the weathered stone surface takes on a unique matte texture, reminiscent of the ancient Dunhuang murals. As the camera steadily pans, the Buddha's solemn features gradually reveal themselves: his eyes slightly closed, revealing a deep compassion and introspection. A high nose bridge and full lips frame a serene smile, and a full white hair rests between his brows. The statue's monotonous stone color is not the only hue. Mottled traces of mineral pigments—faded ochre, russet, and a faint turquoise—still linger in the deep folds of the monk's robe and the crevices of his body. These colors appear remarkably soft under the low-angle, warm light, casting long, sharply defined shadows within the carved recesses, dramatically enhancing the statue's three-dimensionality and the sense of its age. The camera hovers steadily, taking in the grandeur of the Buddha statue and the surrounding desolate and magnificent mountain landscape, creating a sacred and peaceful atmosphere that transcends time and space.", + "seed": 19029, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_0/index_9828_seed_19029_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027858832292258738, + "vlm_logprob": -0.027858832292258738, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.027858832292258738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01171636302024126 + ], + "opposite_logprob_tokens": [ + -4.636716365814209 + ], + "logprob_reward": -0.01171636302024126, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01171636302024126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03987022116780281, + 0.050420913845300674, + 0.018796106800436974, + 0.0523940846323967, + 0.026346832513809204, + 0.04143574461340904, + 0.10213267803192139, + 0.144302099943161, + 0.39362436532974243, + 0.1306769996881485 + ], + "credit_map": [ + 0.003481036750599742, + 0.0056868791580200195, + 0.00430689100176096, + 0.005557216703891754, + 0.004840614274144173, + 0.005467155482620001, + 0.004206877667456865, + 0.006323545705527067, + 0.003999646753072739, + 0.006686416454613209, + 0.0065744430758059025, + 0.007722084876149893, + 0.005788976326584816, + 0.007867724634706974, + 0.0066535016521811485, + 0.0051281144842505455, + 0.0020670685917139053, + 0.0028608515858650208, + 0.00302836112678051, + 0.0018569545354694128, + 0.0023292878177016973, + 0.0027391347102820873, + 0.00202469015493989, + 0.0018897559493780136, + 0.004791584797203541, + 0.004795898217707872, + 0.014804045669734478, + 0.004431460052728653, + 0.003882817691192031, + 0.009273633360862732, + 0.003415857907384634, + 0.0069987839087843895, + 0.003710863646119833, + 0.004035759251564741, + 0.0031546312384307384, + 0.0035162053536623716, + 0.0023210830986499786, + 0.004061053041368723, + 0.0025112710427492857, + 0.003035964211449027, + 0.004178749863058329, + 0.008529574610292912, + 0.0044189151376485825, + 0.006372198928147554, + 0.004975372459739447, + 0.0041095660999417305, + 0.00478555029258132, + 0.004065811168402433, + 0.011815832927823067, + 0.009559405036270618, + 0.012525283731520176, + 0.017458710819482803, + 0.014098758809268475, + 0.013784367591142654, + 0.01131437812000513, + 0.011575940065085888, + 0.013814995065331459, + 0.020209727808833122, + 0.016005713492631912, + 0.018410418182611465, + 0.009699208661913872, + 0.02403428964316845, + 0.015117447823286057, + 0.027010295540094376, + 0.06967304646968842, + 0.04976746067404747, + 0.03392840549349785, + 0.02575034834444523, + 0.03774571046233177, + 0.07294151186943054, + 0.03733706846833229, + 0.06648077815771103, + 0.011890538036823273, + 0.021925289183855057, + 0.01269588340073824, + 0.014208325184881687, + 0.014002400450408459, + 0.01952706091105938, + 0.017292873933911324, + 0.01913462206721306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.01171636302024126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307573288679123 + ], + "opposite_logprob_tokens": [ + -3.1680757999420166 + ], + "logprob_reward": -0.04307573288679123, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307573288679123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06768923252820969, + 0.09067533165216446, + 0.03852159157395363, + 0.020270276814699173, + 0.046155061572790146, + 0.046317312866449356, + 0.13017700612545013, + 0.06186565011739731, + 0.18453513085842133, + 0.313793420791626 + ], + "credit_map": [ + 0.00797558855265379, + 0.009503304958343506, + 0.006426210980862379, + 0.009255442768335342, + 0.007943721488118172, + 0.008370144292712212, + 0.008768608793616295, + 0.009446211159229279, + 0.00879141315817833, + 0.015846246853470802, + 0.012107590213418007, + 0.011174267157912254, + 0.01158122718334198, + 0.007728829514235258, + 0.009684551507234573, + 0.013761206530034542, + 0.005833934992551804, + 0.004905750975012779, + 0.006454369984567165, + 0.003758711740374565, + 0.00443161791190505, + 0.005515643395483494, + 0.0036822394467890263, + 0.003939323592931032, + 0.0014289615210145712, + 0.0031916943844407797, + 0.00233424361795187, + 0.004693299997597933, + 0.002062633167952299, + 0.0017326041124761105, + 0.003148235846310854, + 0.0016786037012934685, + 0.005578355863690376, + 0.006361905485391617, + 0.008241566829383373, + 0.005436733830720186, + 0.004663500469177961, + 0.005326645448803902, + 0.0064980266615748405, + 0.0040483251214027405, + 0.004388716071844101, + 0.006840051617473364, + 0.0050313640385866165, + 0.0056864675134420395, + 0.004150725435465574, + 0.005199160426855087, + 0.0049102529883384705, + 0.010110574774444103, + 0.014410172589123249, + 0.013067019172012806, + 0.014208339154720306, + 0.015722736716270447, + 0.022778263315558434, + 0.01948801428079605, + 0.01213039830327034, + 0.0183720700442791, + 0.007637665141373873, + 0.008236566558480263, + 0.015856074169278145, + 0.004496609326452017, + 0.004271047189831734, + 0.005829983856528997, + 0.004902999382466078, + 0.010634707286953926, + 0.017407242208719254, + 0.029181035235524178, + 0.015697475522756577, + 0.029357999563217163, + 0.012646626681089401, + 0.03166389837861061, + 0.01822831481695175, + 0.03035254031419754, + 0.02261829935014248, + 0.03925508260726929, + 0.036114927381277084, + 0.04683341085910797, + 0.038879670202732086, + 0.04103008657693863, + 0.05208254233002663, + 0.03697938844561577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.04307573288679123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282833397388458 + ], + "opposite_logprob_tokens": [ + -3.773282766342163 + ], + "logprob_reward": -0.023282833397388458, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282833397388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0849909633398056, + 0.050189342349767685, + 0.08267144113779068, + 0.023336611688137054, + 0.022343460470438004, + 0.12407950311899185, + 0.19948387145996094, + 0.07831946760416031, + 0.15454000234603882, + 0.18004533648490906 + ], + "credit_map": [ + 0.008998038247227669, + 0.0097612040117383, + 0.012700870633125305, + 0.01095648668706417, + 0.007388895843178034, + 0.00927812047302723, + 0.009338232688605785, + 0.01656910963356495, + 0.004248961806297302, + 0.008144861087203026, + 0.008675235323607922, + 0.006301851011812687, + 0.006303441245108843, + 0.005139620043337345, + 0.00567430816590786, + 0.005701063200831413, + 0.01160190999507904, + 0.009464573115110397, + 0.0150729576125741, + 0.010897211730480194, + 0.007603852078318596, + 0.013594427146017551, + 0.006895648781210184, + 0.007540850434452295, + 0.003543213475495577, + 0.003685516072437167, + 0.003146121744066477, + 0.0027942752931267023, + 0.002361099235713482, + 0.0023754502180963755, + 0.0034899876918643713, + 0.0019409487722441554, + 0.002884497633203864, + 0.0028485930524766445, + 0.00436426792293787, + 0.0022077939938753843, + 0.002253588056191802, + 0.003021564334630966, + 0.002680795965716243, + 0.0020823588129132986, + 0.010770185850560665, + 0.014196561649441719, + 0.008424931205809116, + 0.013110656291246414, + 0.008596714586019516, + 0.03189767152070999, + 0.0074746375903487206, + 0.029608137905597687, + 0.016332637518644333, + 0.01706096902489662, + 0.014893422834575176, + 0.015352636575698853, + 0.014960060827434063, + 0.0537944957613945, + 0.011842414736747742, + 0.0552472285926342, + 0.009810950607061386, + 0.010832215659320354, + 0.020902583375573158, + 0.0067163752391934395, + 0.005209363531321287, + 0.00865556113421917, + 0.007711703889071941, + 0.008480709977447987, + 0.03128165751695633, + 0.017079893499612808, + 0.012092478573322296, + 0.009791271761059761, + 0.015300818718969822, + 0.024375781416893005, + 0.0159405916929245, + 0.028677504509687424, + 0.014624721370637417, + 0.026824500411748886, + 0.016773534938693047, + 0.014790532179176807, + 0.01911635510623455, + 0.03557189553976059, + 0.025425419211387634, + 0.026918374001979828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.023282833397388458 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304410144686699 + ], + "opposite_logprob_tokens": [ + -3.168044090270996 + ], + "logprob_reward": -0.04304410144686699, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304410144686699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13704407215118408, + 0.06802798062562943, + 0.08125043660402298, + 0.08196794986724854, + 0.04023809731006622, + 0.09378143399953842, + 0.26299849152565, + 0.07704083621501923, + 0.06466784328222275, + 0.09298285096883774 + ], + "credit_map": [ + 0.009915788657963276, + 0.02498096600174904, + 0.014106771908700466, + 0.01858346164226532, + 0.024088846519589424, + 0.01592988148331642, + 0.014913218095898628, + 0.01452514436095953, + 0.004828617442399263, + 0.008554072119295597, + 0.006679284851998091, + 0.011952104978263378, + 0.008686515502631664, + 0.007104390766471624, + 0.010074862278997898, + 0.010148131288588047, + 0.010513211600482464, + 0.00763117428869009, + 0.015017771162092686, + 0.012836930342018604, + 0.007642628159373999, + 0.009310985915362835, + 0.008578653447329998, + 0.009719082154333591, + 0.004998331423848867, + 0.0141077795997262, + 0.01075380016118288, + 0.015333497896790504, + 0.009023122489452362, + 0.008867265656590462, + 0.011844856664538383, + 0.007039299700409174, + 0.0037121532950550318, + 0.008246025070548058, + 0.004462987184524536, + 0.004534480627626181, + 0.006650264840573072, + 0.004968874156475067, + 0.004939425736665726, + 0.0027238866314291954, + 0.008265777491033077, + 0.013472836464643478, + 0.01258618663996458, + 0.014258877374231815, + 0.00814457144588232, + 0.018711309880018234, + 0.011464212089776993, + 0.006877659820020199, + 0.02099759876728058, + 0.02141600288450718, + 0.019100848585367203, + 0.017528505995869637, + 0.019630564376711845, + 0.07393157482147217, + 0.01871979981660843, + 0.07167359441518784, + 0.009281537495553493, + 0.008056140504777431, + 0.017550578340888023, + 0.008677402511239052, + 0.004656681790947914, + 0.014803431928157806, + 0.00604380713775754, + 0.007971263490617275, + 0.009526979178190231, + 0.008458700962364674, + 0.005268136505037546, + 0.005571743007749319, + 0.006384632084518671, + 0.01060271356254816, + 0.00832724291831255, + 0.010527699254453182, + 0.0075460015796124935, + 0.0120674604550004, + 0.00872743409126997, + 0.011654431000351906, + 0.011217404156923294, + 0.010621839202940464, + 0.019216516986489296, + 0.011931760236620903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04304410144686699 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive stone Buddha statue carved directly into a towering crimson mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Buddha statue seated in a lotus position with a full white hair resting between his brows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone surface show mottled traces of faded ochre, russet, and faint turquoise mineral pigments in the folds and crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, sweeping arc that circles the statue while steadily panning to reveal the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the golden afterglow of sunset, casting long, sharply defined shadows within the carved recesses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817513071000576 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.01817513071000576, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817513071000576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15130187571048737, + 0.18024766445159912, + 0.08992505818605423, + 0.04075745865702629, + 0.04075292497873306, + 0.11640715599060059, + 0.09961678832769394, + 0.06475158035755157, + 0.1046581044793129, + 0.11158141493797302 + ], + "credit_map": [ + 0.011091864667832851, + 0.02041132189333439, + 0.02149038575589657, + 0.01933245360851288, + 0.020109379664063454, + 0.02438913844525814, + 0.015313690528273582, + 0.01916363835334778, + 0.01445452868938446, + 0.033663325011730194, + 0.027712706476449966, + 0.02084369771182537, + 0.02249579131603241, + 0.01994880475103855, + 0.020334163680672646, + 0.02079465426504612, + 0.010098135098814964, + 0.010312856175005436, + 0.021509744226932526, + 0.007949626073241234, + 0.008340584114193916, + 0.014426157809793949, + 0.007383022923022509, + 0.009904924780130386, + 0.0031193699687719345, + 0.0054145995527505875, + 0.008953901939094067, + 0.005660983733832836, + 0.003702563000842929, + 0.005100857466459274, + 0.00395928556099534, + 0.004845894407480955, + 0.0039033826906234026, + 0.007783809211105108, + 0.004181441850960255, + 0.006831902079284191, + 0.004023930057883263, + 0.0043618204072117805, + 0.006666268687695265, + 0.0030003658030182123, + 0.010380439460277557, + 0.011607903987169266, + 0.010043519549071789, + 0.012383094057440758, + 0.0064679901115596294, + 0.029744060710072517, + 0.006719003431499004, + 0.029061144217848778, + 0.008372636511921883, + 0.009329953230917454, + 0.006993721704930067, + 0.008282418362796307, + 0.0053955381736159325, + 0.02466721460223198, + 0.00703370152041316, + 0.029541602358222008, + 0.006037516985088587, + 0.005278584081679583, + 0.010592513717710972, + 0.008513699285686016, + 0.005563108250498772, + 0.008768756873905659, + 0.007653214503079653, + 0.012344188056886196, + 0.020011570304632187, + 0.0130538335070014, + 0.009260058403015137, + 0.009033826179802418, + 0.00832945667207241, + 0.01387500949203968, + 0.010928156785666943, + 0.020166194066405296, + 0.008521683514118195, + 0.016438404098153114, + 0.01550053246319294, + 0.015409566462039948, + 0.012874037027359009, + 0.015939339995384216, + 0.012437012977898121, + 0.014460830949246883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.01817513071000576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_tensor([7647, 7647])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_tensor([7647, 7647])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..28945b80ac8278c8b4cc6d4cd2a1dedb2bfdb82c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_tensor([7647, 7647])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, full-body shot establishes a scene of profound tranquility, styled with calm and concentration, within a minimalist dojo. In the center of the frame, a practitioner, her long black hair neatly tied back, wears loose, flowing off-white silk garments that ripple with their every movement. The dojo's floor, crafted from highly polished, light-colored wood, perfectly reflects the soft, golden glow of early morning sunlight streaming through a large shoji paper screen window closed to the camera. This diffuse light casts long, soft shadows that stretch across the floor and brightly illuminates countless motes dancing in the air, creating an atmosphere of intimacy and introspection. With expressions of focused calm, the practitioners move deliberately in a continuous, circular pattern, their feet gliding across the polished surface. Their bodies turn fluidly from the waist, guiding their arms in a seamless, serpentine motion as their open palms constantly change direction, carving into the air as if in a silent and graceful dialogue with an unseen force.", + "seed": 28568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_7647_seed_28568_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038863139878958466, + "vlm_logprob": -0.038863139878958466, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.038863139878958466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012726990506052971 + ], + "opposite_logprob_tokens": [ + -5.512726783752441 + ], + "logprob_reward": -0.012726990506052971, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.012726990506052971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09768907725811005, + 0.3175964951515198, + 0.05203719437122345, + 0.023031018674373627, + 0.040206439793109894, + 0.1525028944015503, + 0.12191109359264374, + 0.06278746575117111, + 0.05980713665485382, + 0.07243113964796066 + ], + "credit_map": [ + 0.009900853037834167, + 0.011475622653961182, + 0.014645545743405819, + 0.025702906772494316, + 0.005045457277446985, + 0.006787564139813185, + 0.009956392459571362, + 0.01417473703622818, + 0.023590343073010445, + 0.008727647364139557, + 0.13528461754322052, + 0.03245961293578148, + 0.0041773817501962185, + 0.00807582214474678, + 0.01266182865947485, + 0.09261925518512726, + 0.005674042273312807, + 0.010088982991874218, + 0.009871174581348896, + 0.00932799931615591, + 0.0016172241885215044, + 0.002464615274220705, + 0.005108713638037443, + 0.007884446531534195, + 0.0015415131347253919, + 0.004087553359568119, + 0.0037620430812239647, + 0.0040499442256987095, + 0.0011062085395678878, + 0.000962133752182126, + 0.004962317645549774, + 0.0025593070313334465, + 0.0033617261797189713, + 0.0033145311754196882, + 0.004898492246866226, + 0.007633761502802372, + 0.0031098697800189257, + 0.0023159433621913195, + 0.008770894259214401, + 0.006801225244998932, + 0.03750627487897873, + 0.010807516053318977, + 0.015146929770708084, + 0.018844323232769966, + 0.019341079518198967, + 0.010653110221028328, + 0.020487075671553612, + 0.019716598093509674, + 0.011565323919057846, + 0.0035091766621917486, + 0.021107178181409836, + 0.011353092268109322, + 0.008168556727468967, + 0.0030347921419888735, + 0.04856983199715614, + 0.014603153802454472, + 0.010387178510427475, + 0.0030914549715816975, + 0.01623835414648056, + 0.004269910044968128, + 0.004862039349973202, + 0.002515421248972416, + 0.015835700556635857, + 0.005587410647422075, + 0.006712000817060471, + 0.004878920502960682, + 0.006563122849911451, + 0.008100118488073349, + 0.010947522707283497, + 0.003317378694191575, + 0.012531396932899952, + 0.00675667729228735, + 0.00369866075925529, + 0.0025280790869146585, + 0.010693343356251717, + 0.007271459326148033, + 0.006560649257153273, + 0.003689131001010537, + 0.013321178033947945, + 0.02466864325106144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.012726990506052971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007651190273463726 + ], + "opposite_logprob_tokens": [ + -4.882651329040527 + ], + "logprob_reward": -0.007651190273463726, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007651190273463726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09321711212396622, + 0.07738123834133148, + 0.039379507303237915, + 0.0794510468840599, + 0.05528145283460617, + 0.1857011318206787, + 0.26886484026908875, + 0.061369914561510086, + 0.04531291872262955, + 0.09404086321592331 + ], + "credit_map": [ + 0.009966926649212837, + 0.013269346207380295, + 0.013135246932506561, + 0.01998107321560383, + 0.004417668096721172, + 0.005937481299042702, + 0.014235513284802437, + 0.012273857370018959, + 0.00568626681342721, + 0.006826869677752256, + 0.02403620444238186, + 0.013923323713243008, + 0.0015398591058328748, + 0.002178192138671875, + 0.0043976204469799995, + 0.018792903050780296, + 0.004188920371234417, + 0.0045627118088305, + 0.008921338245272636, + 0.008230105973780155, + 0.0016024982323870063, + 0.0021628173999488354, + 0.0044729807414114475, + 0.005238134879618883, + 0.007987416349351406, + 0.014405699446797371, + 0.013955804519355297, + 0.01610434800386429, + 0.0035075487103313208, + 0.002665393752977252, + 0.013299989514052868, + 0.007524844259023666, + 0.003373859217390418, + 0.005305007565766573, + 0.015817711129784584, + 0.008131984621286392, + 0.004613782279193401, + 0.0018822378478944302, + 0.010769732296466827, + 0.005387140437960625, + 0.04296920448541641, + 0.017417699098587036, + 0.027607182040810585, + 0.01964646950364113, + 0.02513423003256321, + 0.01098482683300972, + 0.023951169103384018, + 0.017990341410040855, + 0.0224477406591177, + 0.009282772429287434, + 0.060137923806905746, + 0.02064005471765995, + 0.015438977628946304, + 0.005266804713755846, + 0.11072573810815811, + 0.02492482401430607, + 0.005839563440531492, + 0.0021135525312274694, + 0.015370273031294346, + 0.005278541706502438, + 0.008400512859225273, + 0.002324989764019847, + 0.011975957080721855, + 0.01006652507930994, + 0.00466569559648633, + 0.0042298221960663795, + 0.004103507846593857, + 0.004206609446555376, + 0.008568357676267624, + 0.004012975376099348, + 0.007889301516115665, + 0.007636645343154669, + 0.003582906676456332, + 0.002873324090614915, + 0.016938485205173492, + 0.009553887881338596, + 0.006941572763025761, + 0.0048482995480299, + 0.01714288629591465, + 0.03215949609875679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.007651190273463726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059498813934624195 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.0059498813934624195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059498813934624195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10810457915067673, + 0.3645578920841217, + 0.0346454456448555, + 0.0811079666018486, + 0.07499291747808456, + 0.10163212567567825, + 0.06812668591737747, + 0.11819065362215042, + 0.024012045934796333, + 0.024629749357700348 + ], + "credit_map": [ + 0.00893446896225214, + 0.011497377417981625, + 0.026224108412861824, + 0.016980787739157677, + 0.0076009794138371944, + 0.007265981286764145, + 0.01052042841911316, + 0.019080450758337975, + 0.026703190058469772, + 0.007276467047631741, + 0.15639202296733856, + 0.03490493819117546, + 0.004990450572222471, + 0.009257528930902481, + 0.014112219214439392, + 0.11092108488082886, + 0.0036392437759786844, + 0.004067748785018921, + 0.005011865869164467, + 0.005134607199579477, + 0.0036953757517039776, + 0.0036381930112838745, + 0.00457005575299263, + 0.004888358525931835, + 0.008017522282898426, + 0.01578528620302677, + 0.015438438393175602, + 0.016146482899785042, + 0.0036707057151943445, + 0.002804502844810486, + 0.010822112672030926, + 0.008422913029789925, + 0.009153234772384167, + 0.006651836447417736, + 0.020210497081279755, + 0.010071901604533195, + 0.00539572024717927, + 0.0038519231602549553, + 0.011515290476381779, + 0.008142508566379547, + 0.026038331910967827, + 0.007700216956436634, + 0.010382313281297684, + 0.012964732013642788, + 0.0130832614377141, + 0.007371449377387762, + 0.01351325586438179, + 0.01057856809347868, + 0.006135911680757999, + 0.0023487901780754328, + 0.01563539169728756, + 0.0056613958440721035, + 0.003221123246476054, + 0.0019182543037459254, + 0.02771027758717537, + 0.005495535675436258, + 0.018924230709671974, + 0.0052077798172831535, + 0.03349630534648895, + 0.007264701183885336, + 0.00816237274557352, + 0.005160835571587086, + 0.027644610032439232, + 0.012329818680882454, + 0.0029511190950870514, + 0.001880760071799159, + 0.003061730647459626, + 0.0029107790905982256, + 0.0041765980422496796, + 0.001996283885091543, + 0.0036802049726247787, + 0.0033545726910233498, + 0.0016072941944003105, + 0.002119947923347354, + 0.004097449127584696, + 0.0034714259672909975, + 0.003151638898998499, + 0.0014351473655551672, + 0.003595122369006276, + 0.0051517244428396225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0059498813934624195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890246331691742 + ], + "opposite_logprob_tokens": [ + -2.464024543762207 + ], + "logprob_reward": -0.0890246331691742, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890246331691742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12700098752975464, + 0.06874939799308777, + 0.0426424965262413, + 0.05332738161087036, + 0.03207578510046005, + 0.12277040630578995, + 0.44917118549346924, + 0.025668570771813393, + 0.025807850062847137, + 0.05278598517179489 + ], + "credit_map": [ + 0.011731178499758244, + 0.015021336264908314, + 0.02976897731423378, + 0.024897603318095207, + 0.004489624407142401, + 0.006171532440930605, + 0.01607591286301613, + 0.01884479448199272, + 0.0067197587341070175, + 0.007899805903434753, + 0.01752147451043129, + 0.013374408707022667, + 0.0017707358347252011, + 0.0027098439168184996, + 0.005423998460173607, + 0.013329367153346539, + 0.005580069962888956, + 0.0061681512743234634, + 0.007486490998417139, + 0.007641701493412256, + 0.0019227353623136878, + 0.0028037845622748137, + 0.003812340786680579, + 0.007227219175547361, + 0.0050673699006438255, + 0.010666722431778908, + 0.009458382613956928, + 0.010582035407423973, + 0.002392243128269911, + 0.0018261000514030457, + 0.008200891315937042, + 0.005133635364472866, + 0.004601850174367428, + 0.0024956960696727037, + 0.0056586600840091705, + 0.004578847438097, + 0.0025379934813827276, + 0.0019029907416552305, + 0.006385094951838255, + 0.0039146519266068935, + 0.02827831543982029, + 0.012339982204139233, + 0.014376843348145485, + 0.013002418912947178, + 0.02191149815917015, + 0.00696706073358655, + 0.013912657275795937, + 0.01198161393404007, + 0.033922504633665085, + 0.016255993396043777, + 0.11560582369565964, + 0.032665498554706573, + 0.027529140934348106, + 0.008675063028931618, + 0.1797734946012497, + 0.034743595868349075, + 0.004742610268294811, + 0.0027541134040802717, + 0.003222550731152296, + 0.0025855121202766895, + 0.0028096865862607956, + 0.001562541234306991, + 0.0038480847142636776, + 0.00414347043260932, + 0.00246060430072248, + 0.002271532779559493, + 0.002718721516430378, + 0.002746103797107935, + 0.004020379856228828, + 0.0025699043180793524, + 0.004167408216744661, + 0.0048531945794820786, + 0.002597005106508732, + 0.0032973194029182196, + 0.006591225974261761, + 0.0070692552253603935, + 0.00790372397750616, + 0.0025395555421710014, + 0.010058392770588398, + 0.012729507870972157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0890246331691742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896300405263901 + ], + "opposite_logprob_tokens": [ + -2.578963041305542 + ], + "logprob_reward": -0.07896300405263901, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896300405263901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036336686462163925, + 0.13190008699893951, + 0.02851971797645092, + 0.012947958894073963, + 0.03935220092535019, + 0.034073811024427414, + 0.6109840273857117, + 0.01954362541437149, + 0.017177753150463104, + 0.06916411966085434 + ], + "credit_map": [ + 0.003780112834647298, + 0.004159973934292793, + 0.007937519811093807, + 0.007490038406103849, + 0.001318749855272472, + 0.002587774069979787, + 0.003321208758279681, + 0.005741308443248272, + 0.00887155719101429, + 0.005163018591701984, + 0.05535702779889107, + 0.014684994705021381, + 0.001603544456884265, + 0.003322902135550976, + 0.004987697582691908, + 0.03790934011340141, + 0.0029904996044933796, + 0.004446221515536308, + 0.007263603620231152, + 0.005916889291256666, + 0.0008977213874459267, + 0.0014045920688658953, + 0.00226629339158535, + 0.003333898028358817, + 0.0009751914767548442, + 0.002725880825892091, + 0.0018463152227923274, + 0.002461114199832082, + 0.0009658236522227526, + 0.0005913791828788817, + 0.0020388104021549225, + 0.0013434445718303323, + 0.002710506319999695, + 0.0022966477554291487, + 0.013226895593106747, + 0.004851045086979866, + 0.0027654627338051796, + 0.0015760880196467042, + 0.006079599726945162, + 0.005845954176038504, + 0.005928021389991045, + 0.004112237598747015, + 0.005461289081722498, + 0.0039142221212387085, + 0.004016280639916658, + 0.0019771254155784845, + 0.004727919120341539, + 0.0039367168210446835, + 0.04794226959347725, + 0.017564158886671066, + 0.14320257306098938, + 0.0471174418926239, + 0.03599558770656586, + 0.013415543362498283, + 0.2506582736968994, + 0.05508819594979286, + 0.0027700704522430897, + 0.0012639984488487244, + 0.002076201606541872, + 0.0015184045769274235, + 0.002246385207399726, + 0.001335998298600316, + 0.0031584235839545727, + 0.005174141842871904, + 0.0017245388589799404, + 0.0013686135644093156, + 0.0021635449957102537, + 0.0018464653985574841, + 0.003401497146114707, + 0.0012729141162708402, + 0.0025980195496231318, + 0.0028021596372127533, + 0.0028367231134325266, + 0.0027777983341366053, + 0.009104063734412193, + 0.0076833548955619335, + 0.008912038058042526, + 0.0032136293593794107, + 0.014538513496518135, + 0.020097997039556503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.07896300405263901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, full-body shot establishes a scene of profound tranquility, styled with calm and concentration, within a minimalist dojo. In the center of the frame, a practitioner, her long black hair neatly tied back, wears loose, flowing off-white silk garments that ripple with their every movement. The dojo's floor, crafted from highly polished, light-colored wood, perfectly reflects the soft, golden glow of early morning sunlight streaming through a large shoji paper screen window closed to the camera. This diffuse light casts long, soft shadows that stretch across the floor and brightly illuminates countless motes dancing in the air, creating an atmosphere of intimacy and introspection. With expressions of focused calm, the practitioners move deliberately in a continuous, circular pattern, their feet gliding across the polished surface. Their bodies turn fluidly from the waist, guiding their arms in a seamless, serpentine motion as their open palms constantly change direction, carving into the air as if in a silent and graceful dialogue with an unseen force.", + "seed": 28569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_7647_seed_28569_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06692494433373213, + "vlm_logprob": -0.06692494433373213, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06692494433373213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02737347036600113 + ], + "opposite_logprob_tokens": [ + -3.7773735523223877 + ], + "logprob_reward": -0.02737347036600113, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02737347036600113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7234428524971008, + 0.06393270939588547, + 0.01779792830348015, + 0.0281648151576519, + 0.02511323243379593, + 0.03292388841509819, + 0.024415049701929092, + 0.02641332894563675, + 0.019344842061400414, + 0.038451340049505234 + ], + "credit_map": [ + 0.03333885967731476, + 0.018310312181711197, + 0.19842343032360077, + 0.03872557729482651, + 0.09510786086320877, + 0.014849627390503883, + 0.28263360261917114, + 0.04205358028411865, + 0.004733377601951361, + 0.004934546072036028, + 0.011705602519214153, + 0.007235103286802769, + 0.013803227804601192, + 0.003639772068709135, + 0.012273951433598995, + 0.0056071230210363865, + 0.0020121328998357058, + 0.0013594329357147217, + 0.002599576488137245, + 0.0022413996048271656, + 0.0034501345362514257, + 0.0014341323403641582, + 0.002586678136140108, + 0.002114440780133009, + 0.002231383929029107, + 0.002620014827698469, + 0.003972140606492758, + 0.003231419250369072, + 0.005914618726819754, + 0.0015590069815516472, + 0.004871812183409929, + 0.003764418186619878, + 0.004199265502393246, + 0.002914871321991086, + 0.00366010214202106, + 0.0023634156677871943, + 0.0031929784454405308, + 0.0013953510206192732, + 0.003341569099575281, + 0.004045677371323109, + 0.0049228728748857975, + 0.0033124068286269903, + 0.005150672513991594, + 0.0046206507831811905, + 0.005475921556353569, + 0.0024273034650832415, + 0.003718459280207753, + 0.0032955999486148357, + 0.002776304492726922, + 0.0040086405351758, + 0.0049200961366295815, + 0.002997988136485219, + 0.001907639903947711, + 0.0015242209192365408, + 0.003387786913663149, + 0.002892371267080307, + 0.0020109873730689287, + 0.003033663146197796, + 0.004427654203027487, + 0.0052350242622196674, + 0.0015860314015299082, + 0.001086325733922422, + 0.004371303599327803, + 0.0046623386442661285, + 0.0017304258653894067, + 0.0027701223734766245, + 0.0033053040970116854, + 0.003921372815966606, + 0.0018156708683818579, + 0.0011829278664663434, + 0.0018641510978341103, + 0.00275486521422863, + 0.002524696057662368, + 0.002841535024344921, + 0.009491763077676296, + 0.004125363193452358, + 0.0024360516108572483, + 0.0021372565533965826, + 0.010449483059346676, + 0.004445191007107496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02737347036600113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809148497879505 + ], + "opposite_logprob_tokens": [ + -4.6348090171813965 + ], + "logprob_reward": -0.009809148497879505, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809148497879505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28918567299842834, + 0.060801707208156586, + 0.06147351488471031, + 0.05489629507064819, + 0.106626495718956, + 0.16693755984306335, + 0.07050871849060059, + 0.03428152948617935, + 0.04421946778893471, + 0.11106901615858078 + ], + "credit_map": [ + 0.015433772467076778, + 0.0069501399993896484, + 0.06824855506420135, + 0.021248437464237213, + 0.044378750026226044, + 0.0071733263321220875, + 0.10696593672037125, + 0.01878676936030388, + 0.004240062087774277, + 0.006852509919553995, + 0.008899467997252941, + 0.00868056807667017, + 0.010877307504415512, + 0.004579213447868824, + 0.009822473861277103, + 0.006850102916359901, + 0.004752525128424168, + 0.0031498598400503397, + 0.008761598728597164, + 0.010939010418951511, + 0.011279059574007988, + 0.004554131533950567, + 0.01095916423946619, + 0.007078160531818867, + 0.003822341561317444, + 0.0039057768881320953, + 0.006333163473755121, + 0.008240792900323868, + 0.012220709584653378, + 0.003264729166403413, + 0.009676679968833923, + 0.007432098034769297, + 0.014011839404702187, + 0.007623689249157906, + 0.018896933645009995, + 0.012923103757202625, + 0.017594974488019943, + 0.004978904500603676, + 0.01741926185786724, + 0.01317779254168272, + 0.014610129408538342, + 0.019925203174352646, + 0.028544927015900612, + 0.022151445969939232, + 0.03146689757704735, + 0.00925489328801632, + 0.02495449036359787, + 0.016029568389058113, + 0.0054952530190348625, + 0.009708651341497898, + 0.008461791090667248, + 0.010855541564524174, + 0.011685592122375965, + 0.0033978295978158712, + 0.011841465719044209, + 0.009062591008841991, + 0.003424320835620165, + 0.005276460666209459, + 0.004963798448443413, + 0.006682629231363535, + 0.0037643350660800934, + 0.001677545951679349, + 0.004332524258643389, + 0.004159914795309305, + 0.0036203693598508835, + 0.007246685214340687, + 0.006446002051234245, + 0.010401916690170765, + 0.004383325111120939, + 0.0026186660397797823, + 0.004294372163712978, + 0.005208129994571209, + 0.012700259685516357, + 0.01939600519835949, + 0.020849043503403664, + 0.017112359404563904, + 0.006554367486387491, + 0.006147617474198341, + 0.013999812304973602, + 0.014309557154774666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.009809148497879505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42351585626602173, + 0.13241064548492432, + 0.07076705992221832, + 0.12676817178726196, + 0.020481565967202187, + 0.04466503486037254, + 0.05093635246157646, + 0.03784176707267761, + 0.02352531999349594, + 0.06908820569515228 + ], + "credit_map": [ + 0.021404488012194633, + 0.012236308306455612, + 0.11079643666744232, + 0.023657871410250664, + 0.05803241953253746, + 0.013300457037985325, + 0.15997625887393951, + 0.024111643433570862, + 0.014619357883930206, + 0.022569620981812477, + 0.014185293577611446, + 0.021281911060214043, + 0.017208604142069817, + 0.012683268636465073, + 0.017678260803222656, + 0.012184329330921173, + 0.0050860014744102955, + 0.006832750514149666, + 0.01141789648681879, + 0.009529715403914452, + 0.012475790455937386, + 0.0038832922000437975, + 0.01141662523150444, + 0.010125001892447472, + 0.012120211496949196, + 0.015107500366866589, + 0.018860556185245514, + 0.017964767292141914, + 0.02099650911986828, + 0.007846279069781303, + 0.02145642228424549, + 0.01241594459861517, + 0.0014473459450528026, + 0.0016541526420041919, + 0.0027728101704269648, + 0.0034443961922079325, + 0.0033354125916957855, + 0.0011008166475221515, + 0.0031663693953305483, + 0.0035602636635303497, + 0.004416507203131914, + 0.0058493767865002155, + 0.006034345366060734, + 0.005898843519389629, + 0.009083389304578304, + 0.003216608427464962, + 0.00658663734793663, + 0.003579327603802085, + 0.003037526970729232, + 0.008324725553393364, + 0.0070983050391077995, + 0.006913428194820881, + 0.0061794305220246315, + 0.003586234524846077, + 0.006604173686355352, + 0.009192532859742641, + 0.0035765774082392454, + 0.0043229940347373486, + 0.006933003198355436, + 0.005636092275381088, + 0.0023148523177951574, + 0.0023116294760257006, + 0.004843928385525942, + 0.007902695797383785, + 0.0022639818489551544, + 0.0041455356404185295, + 0.003409771714359522, + 0.0038276626728475094, + 0.0021171364933252335, + 0.0013857827289029956, + 0.002889399416744709, + 0.003486050060018897, + 0.005845871288329363, + 0.008656305260956287, + 0.013161670416593552, + 0.010190372355282307, + 0.004483983386307955, + 0.0048163942992687225, + 0.008780505508184433, + 0.013153105974197388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989015579223633 + ], + "opposite_logprob_tokens": [ + -1.8048901557922363 + ], + "logprob_reward": -0.17989015579223633, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989015579223633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3081318736076355, + 0.07999730855226517, + 0.07469750940799713, + 0.09878838807344437, + 0.06690224260091782, + 0.06606884300708771, + 0.12556211650371552, + 0.04665742814540863, + 0.051896385848522186, + 0.08129790425300598 + ], + "credit_map": [ + 0.015597367659211159, + 0.011212889105081558, + 0.07347571849822998, + 0.021951790899038315, + 0.0480571873486042, + 0.010549239814281464, + 0.10550907999277115, + 0.021778596565127373, + 0.00552345160394907, + 0.004867364652454853, + 0.013870511204004288, + 0.008014996536076069, + 0.015527380630373955, + 0.0078828614205122, + 0.01303364522755146, + 0.011277101002633572, + 0.005188690032809973, + 0.0038562354166060686, + 0.01146598532795906, + 0.013352691195905209, + 0.013651514425873756, + 0.004971859976649284, + 0.012343631125986576, + 0.00986690167337656, + 0.006748361513018608, + 0.005874623544514179, + 0.013157729059457779, + 0.011981168761849403, + 0.022648949176073074, + 0.004606413654983044, + 0.022182516753673553, + 0.011588627472519875, + 0.009348747320473194, + 0.008559667505323887, + 0.0110686756670475, + 0.00641900347545743, + 0.009778290055692196, + 0.0038958529476076365, + 0.008822162635624409, + 0.009009846486151218, + 0.0055462452583014965, + 0.009108960628509521, + 0.012313498184084892, + 0.009508250281214714, + 0.01197850052267313, + 0.0028481767512857914, + 0.008786031045019627, + 0.0059791794046759605, + 0.011675630696117878, + 0.018595941364765167, + 0.015742644667625427, + 0.019708435982465744, + 0.020317384973168373, + 0.005224414169788361, + 0.0211003627628088, + 0.013197301886975765, + 0.0027479531709104776, + 0.008968613110482693, + 0.00655096722766757, + 0.00992828980088234, + 0.005594780668616295, + 0.002227169694378972, + 0.005753885023295879, + 0.004885768052190542, + 0.004027973860502243, + 0.01080456841737032, + 0.006921445019543171, + 0.011850152164697647, + 0.004040381871163845, + 0.00213046558201313, + 0.004930196795612574, + 0.007191204931586981, + 0.012569618411362171, + 0.016505427658557892, + 0.01166754774749279, + 0.012198224663734436, + 0.006025942973792553, + 0.00621122308075428, + 0.007606138475239277, + 0.008513783104717731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.17989015579223633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291403323411942 + ], + "opposite_logprob_tokens": [ + -2.2379140853881836 + ], + "logprob_reward": -0.11291403323411942, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291403323411942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3642173409461975, + 0.21067404747009277, + 0.035166457295417786, + 0.03297025337815285, + 0.05582879111170769, + 0.03647935017943382, + 0.04363676905632019, + 0.10896029323339462, + 0.07314758747816086, + 0.03891916945576668 + ], + "credit_map": [ + 0.018418796360492706, + 0.006028322968631983, + 0.0961325466632843, + 0.019764406606554985, + 0.050047874450683594, + 0.00881171878427267, + 0.14513929188251495, + 0.019874367862939835, + 0.014742170460522175, + 0.019039766862988472, + 0.03081849031150341, + 0.01964111067354679, + 0.04895481467247009, + 0.02238401398062706, + 0.03773222491145134, + 0.017361456528306007, + 0.0026608447078615427, + 0.0017203245079144835, + 0.005484272725880146, + 0.005221566651016474, + 0.006569088436663151, + 0.0017804458038881421, + 0.00829937495291233, + 0.0034305404406040907, + 0.002705697901546955, + 0.002148303436115384, + 0.005246270447969437, + 0.00308400159701705, + 0.006414440926164389, + 0.0017821532674133778, + 0.008297417312860489, + 0.0032919682562351227, + 0.006018707528710365, + 0.00621946481987834, + 0.00824035331606865, + 0.007465492468327284, + 0.009176102466881275, + 0.0026373949367552996, + 0.009389682672917843, + 0.006681591272354126, + 0.0030672301072627306, + 0.0038205645978450775, + 0.005625068675726652, + 0.005540271755307913, + 0.00708645349368453, + 0.001747212721966207, + 0.00574462441727519, + 0.00384792173281312, + 0.00455363467335701, + 0.006391339935362339, + 0.008748052641749382, + 0.007526923902332783, + 0.005007853265851736, + 0.0018748216098174453, + 0.005569049157202244, + 0.003965090960264206, + 0.007975530810654163, + 0.024225976318120956, + 0.022911369800567627, + 0.013855012133717537, + 0.01107704546302557, + 0.004264787770807743, + 0.013674259185791016, + 0.010976315476000309, + 0.007355930749326944, + 0.015698684379458427, + 0.012889083474874496, + 0.013593409210443497, + 0.006924558896571398, + 0.003244271967560053, + 0.0066556320525705814, + 0.006786020006984472, + 0.0032429227139800787, + 0.006840587127953768, + 0.007386954501271248, + 0.006773302797228098, + 0.0025309426710009575, + 0.002606105525046587, + 0.004646447487175465, + 0.004891910124570131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.11291403323411942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, full-body shot establishes a scene of profound tranquility, styled with calm and concentration, within a minimalist dojo. In the center of the frame, a practitioner, her long black hair neatly tied back, wears loose, flowing off-white silk garments that ripple with their every movement. The dojo's floor, crafted from highly polished, light-colored wood, perfectly reflects the soft, golden glow of early morning sunlight streaming through a large shoji paper screen window closed to the camera. This diffuse light casts long, soft shadows that stretch across the floor and brightly illuminates countless motes dancing in the air, creating an atmosphere of intimacy and introspection. With expressions of focused calm, the practitioners move deliberately in a continuous, circular pattern, their feet gliding across the polished surface. Their bodies turn fluidly from the waist, guiding their arms in a seamless, serpentine motion as their open palms constantly change direction, carving into the air as if in a silent and graceful dialogue with an unseen force.", + "seed": 28570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_7647_seed_28570_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04983407170511782, + "vlm_logprob": -0.04983407170511782, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.04983407170511782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05323219299316406 + ], + "opposite_logprob_tokens": [ + -3.053232192993164 + ], + "logprob_reward": -0.05323219299316406, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05323219299316406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19719840586185455, + 0.13482122123241425, + 0.09008295089006424, + 0.030029773712158203, + 0.043628040701150894, + 0.02317158132791519, + 0.112046979367733, + 0.10497055947780609, + 0.030948709696531296, + 0.23310181498527527 + ], + "credit_map": [ + 0.01896573230624199, + 0.019723035395145416, + 0.02628682553768158, + 0.013065718114376068, + 0.03567555546760559, + 0.03310822322964668, + 0.014889671467244625, + 0.03548363223671913, + 0.02226690575480461, + 0.011554424650967121, + 0.025797106325626373, + 0.008776577189564705, + 0.012179017998278141, + 0.01039653830230236, + 0.03163144737482071, + 0.012219195254147053, + 0.005917280912399292, + 0.010796483606100082, + 0.013749784789979458, + 0.013353761285543442, + 0.01080238725990057, + 0.008197623305022717, + 0.010347563773393631, + 0.01691805012524128, + 0.0020681708119809628, + 0.0027912696823477745, + 0.00462309829890728, + 0.00477217510342598, + 0.004856447223573923, + 0.0025974775198847055, + 0.0052519612945616245, + 0.003069171216338873, + 0.009752810932695866, + 0.0037219005171209574, + 0.006000652443617582, + 0.005586465820670128, + 0.005167336203157902, + 0.004494772292673588, + 0.004148031584918499, + 0.0047560702078044415, + 0.0022461351472884417, + 0.0029054235201328993, + 0.0027652799617499113, + 0.0040818797424435616, + 0.004224988166242838, + 0.0015638326294720173, + 0.0031956457532942295, + 0.002188393147662282, + 0.008009823970496655, + 0.009067993611097336, + 0.01686372235417366, + 0.009532283991575241, + 0.01826544664800167, + 0.020315181463956833, + 0.01168887224048376, + 0.01830364763736725, + 0.006703103892505169, + 0.011027897708117962, + 0.01962108165025711, + 0.012153643183410168, + 0.010635452345013618, + 0.009095368906855583, + 0.026749679818749428, + 0.008984319865703583, + 0.0024690746795386076, + 0.004186925012618303, + 0.006672487128525972, + 0.0027992238756269217, + 0.0048119863495230675, + 0.0017403928795829415, + 0.0046574124135077, + 0.00361120468005538, + 0.013690952211618423, + 0.016773266717791557, + 0.046599071472883224, + 0.03879731893539429, + 0.0256833303719759, + 0.011346879415214062, + 0.045226022601127625, + 0.034984953701496124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.05323219299316406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802891872823238 + ], + "opposite_logprob_tokens": [ + -4.63480281829834 + ], + "logprob_reward": -0.009802891872823238, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802891872823238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08843784034252167, + 0.33674609661102295, + 0.0830560028553009, + 0.053594622761011124, + 0.03603271022439003, + 0.03281098231673241, + 0.1353885382413864, + 0.09301813691854477, + 0.09055683016777039, + 0.050358302891254425 + ], + "credit_map": [ + 0.007114051841199398, + 0.011876240372657776, + 0.011950216256082058, + 0.007810736540704966, + 0.019051002338528633, + 0.011964312754571438, + 0.009454210288822651, + 0.009217068552970886, + 0.05735349282622337, + 0.02576909027993679, + 0.06040911003947258, + 0.01949651539325714, + 0.021051345393061638, + 0.03127644956111908, + 0.07087825238704681, + 0.050511814653873444, + 0.0036317806225270033, + 0.012700063176453114, + 0.0109690697863698, + 0.014545464888215065, + 0.01644645817577839, + 0.0049955272115767, + 0.008977925404906273, + 0.010789714753627777, + 0.005070058163255453, + 0.004215124528855085, + 0.0065712654031813145, + 0.008166727609932423, + 0.011032754555344582, + 0.00503117498010397, + 0.008537795394659042, + 0.004969723988324404, + 0.005785948131233454, + 0.0032615680247545242, + 0.004443891812115908, + 0.004348671063780785, + 0.007375576999038458, + 0.002973138354718685, + 0.004246042110025883, + 0.0035978714004158974, + 0.0022781325969845057, + 0.005046559032052755, + 0.004454835783690214, + 0.004356443416327238, + 0.007117947563529015, + 0.0022538676857948303, + 0.004322682041674852, + 0.002980512101203203, + 0.01115671917796135, + 0.01668483391404152, + 0.024207474663853645, + 0.013573245145380497, + 0.022908542305231094, + 0.014325365424156189, + 0.014869431965053082, + 0.01766292005777359, + 0.008017687126994133, + 0.014552549459040165, + 0.013598182238638401, + 0.009946495294570923, + 0.012198974378407001, + 0.006867764052003622, + 0.02018236555159092, + 0.007654119282960892, + 0.0038306969217956066, + 0.005251337308436632, + 0.029178684577345848, + 0.007725374773144722, + 0.00696998555213213, + 0.0016959402710199356, + 0.026677897199988365, + 0.009226915426552296, + 0.003707238705828786, + 0.00417099054902792, + 0.009407106786966324, + 0.008308043703436852, + 0.004966602195054293, + 0.0030063996091485023, + 0.009640621021389961, + 0.007151299621909857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.009802891872823238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031919742468744516 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031919742468744516, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031919742468744516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07848197966814041, + 0.10736266523599625, + 0.11532820761203766, + 0.2101052850484848, + 0.06581518054008484, + 0.05917125195264816, + 0.046595435589551926, + 0.09322257339954376, + 0.1051582470536232, + 0.11875919252634048 + ], + "credit_map": [ + 0.008170089684426785, + 0.008907804265618324, + 0.012397839687764645, + 0.005262303631752729, + 0.014814185909926891, + 0.012375575490295887, + 0.007107279263436794, + 0.009446904063224792, + 0.014403116889297962, + 0.011286098510026932, + 0.01631094142794609, + 0.0114623187109828, + 0.01206355169415474, + 0.009806091897189617, + 0.015058144927024841, + 0.016972407698631287, + 0.016170579940080643, + 0.011416074819862843, + 0.013240993954241276, + 0.01198973786085844, + 0.017802754417061806, + 0.017283394932746887, + 0.017083216458559036, + 0.01034145150333643, + 0.03533535450696945, + 0.013449335470795631, + 0.04390818253159523, + 0.013367895968258381, + 0.018550077453255653, + 0.034377142786979675, + 0.028297841548919678, + 0.022819453850388527, + 0.012933667749166489, + 0.005912186577916145, + 0.008340002037584782, + 0.006292390637099743, + 0.009274369105696678, + 0.008927159011363983, + 0.006010356359183788, + 0.008125047199428082, + 0.004114128649234772, + 0.009934958070516586, + 0.009622078388929367, + 0.007677007932215929, + 0.008991390466690063, + 0.004784595686942339, + 0.008450351655483246, + 0.005596744362264872, + 0.0033702547661960125, + 0.005930319428443909, + 0.008094560354948044, + 0.0058171371929347515, + 0.0060157678090035915, + 0.006451662164181471, + 0.005785069894045591, + 0.005130662117153406, + 0.007056671194732189, + 0.013813754543662071, + 0.014385775662958622, + 0.008946837857365608, + 0.009350040927529335, + 0.009604258462786674, + 0.017022274434566498, + 0.01304295752197504, + 0.005395146552473307, + 0.00803410168737173, + 0.03311071917414665, + 0.0076163881458342075, + 0.008379492908716202, + 0.004094134084880352, + 0.025044962763786316, + 0.013483311980962753, + 0.007666615769267082, + 0.008539587259292603, + 0.024876588955521584, + 0.02043026126921177, + 0.011467590928077698, + 0.00548589276149869, + 0.021067511290311813, + 0.019225141033530235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.0031919742468744516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129392683506012 + ], + "opposite_logprob_tokens": [ + -2.2379393577575684 + ], + "logprob_reward": -0.1129392683506012, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129392683506012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06667805463075638, + 0.13875919580459595, + 0.04637463018298149, + 0.07166291028261185, + 0.039843056350946426, + 0.05353977158665657, + 0.14010140299797058, + 0.1761370599269867, + 0.14819291234016418, + 0.1187109649181366 + ], + "credit_map": [ + 0.006264262367039919, + 0.010387711226940155, + 0.009831245057284832, + 0.004755303729325533, + 0.009014402516186237, + 0.00950966589152813, + 0.005827280692756176, + 0.011088182218372822, + 0.02558717504143715, + 0.011946482583880424, + 0.02437519282102585, + 0.009406228549778461, + 0.00987635925412178, + 0.013258523307740688, + 0.03019978292286396, + 0.014109454117715359, + 0.004830838646739721, + 0.00552796944975853, + 0.005871997680515051, + 0.006440158933401108, + 0.009454761631786823, + 0.003314817789942026, + 0.005666088778525591, + 0.005267999600619078, + 0.00947191659361124, + 0.004557364620268345, + 0.015732448548078537, + 0.006666137836873531, + 0.009459287859499454, + 0.008099355734884739, + 0.010328837670385838, + 0.007347560953348875, + 0.003057576483115554, + 0.0035270019434392452, + 0.005196855403482914, + 0.004131899680942297, + 0.011033706367015839, + 0.0036545644979923964, + 0.005887900944799185, + 0.0033535496331751347, + 0.0027370143216103315, + 0.007882642559707165, + 0.009493181481957436, + 0.007561523467302322, + 0.008214595727622509, + 0.002990087028592825, + 0.00898806657642126, + 0.005672656465321779, + 0.011739921756088734, + 0.019641999155282974, + 0.020263200625777245, + 0.02168385311961174, + 0.01764487475156784, + 0.013102858327329159, + 0.01926238462328911, + 0.016762306913733482, + 0.013030122965574265, + 0.04156247153878212, + 0.022632433101534843, + 0.01880064606666565, + 0.0250833909958601, + 0.010496622882783413, + 0.029924938455224037, + 0.014606432057917118, + 0.00618496211245656, + 0.00777290528640151, + 0.04872533679008484, + 0.010390316136181355, + 0.013077877461910248, + 0.003352140309289098, + 0.041174694895744324, + 0.017514685168862343, + 0.006787777878344059, + 0.008208658546209335, + 0.023909103125333786, + 0.02254418469965458, + 0.011531723663210869, + 0.005200218874961138, + 0.022486168891191483, + 0.01804312691092491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.1129392683506012 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000403106212616 + ], + "opposite_logprob_tokens": [ + -2.6950039863586426 + ], + "logprob_reward": -0.07000403106212616, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000403106212616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059438735246658325, + 0.10604135692119598, + 0.08601149171590805, + 0.05020353943109512, + 0.04586430639028549, + 0.1088622584939003, + 0.04649228975176811, + 0.16873881220817566, + 0.187978133559227, + 0.1403689980506897 + ], + "credit_map": [ + 0.0072493902407586575, + 0.008160940371453762, + 0.00686222966760397, + 0.005670418497174978, + 0.011514861136674881, + 0.006519796326756477, + 0.0055938586592674255, + 0.007867247797548771, + 0.016651151701807976, + 0.009042132645845413, + 0.01883104257285595, + 0.006336376536637545, + 0.007039979565888643, + 0.009814544580876827, + 0.021603666245937347, + 0.01672247424721718, + 0.003873811336234212, + 0.010953554883599281, + 0.008555819280445576, + 0.01124202273786068, + 0.02512020245194435, + 0.00531611917540431, + 0.012432047165930271, + 0.008517923764884472, + 0.004021155182272196, + 0.0035326911602169275, + 0.00652965297922492, + 0.005673959851264954, + 0.012047266587615013, + 0.004707816056907177, + 0.007725023198872805, + 0.005965979304164648, + 0.0058275023475289345, + 0.004387055058032274, + 0.00613009138032794, + 0.005533447489142418, + 0.009852456860244274, + 0.00403463514521718, + 0.005665190983563662, + 0.004433932714164257, + 0.005214625503867865, + 0.014208728447556496, + 0.018374817445874214, + 0.013697372749447823, + 0.020512862130999565, + 0.005059703253209591, + 0.02034599706530571, + 0.011448162607848644, + 0.0038699752185493708, + 0.006993084214627743, + 0.005630686413496733, + 0.006787079386413097, + 0.008170432411134243, + 0.003743695095181465, + 0.0064786807633936405, + 0.0048186639323830605, + 0.012035044841468334, + 0.04054468125104904, + 0.023035187274217606, + 0.020348137244582176, + 0.020093655213713646, + 0.007488251198083162, + 0.02778397873044014, + 0.01740989461541176, + 0.01210611779242754, + 0.017493104562163353, + 0.024640968069434166, + 0.029684191569685936, + 0.04542664438486099, + 0.008982310071587563, + 0.029732752591371536, + 0.019912073388695717, + 0.010079788975417614, + 0.016514787450432777, + 0.025487912818789482, + 0.024294352158904076, + 0.014158661477267742, + 0.007526781875640154, + 0.02172733098268509, + 0.020579401403665543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.07000403106212616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, full-body shot establishes a scene of profound tranquility, styled with calm and concentration, within a minimalist dojo. In the center of the frame, a practitioner, her long black hair neatly tied back, wears loose, flowing off-white silk garments that ripple with their every movement. The dojo's floor, crafted from highly polished, light-colored wood, perfectly reflects the soft, golden glow of early morning sunlight streaming through a large shoji paper screen window closed to the camera. This diffuse light casts long, soft shadows that stretch across the floor and brightly illuminates countless motes dancing in the air, creating an atmosphere of intimacy and introspection. With expressions of focused calm, the practitioners move deliberately in a continuous, circular pattern, their feet gliding across the polished surface. Their bodies turn fluidly from the waist, guiding their arms in a seamless, serpentine motion as their open palms constantly change direction, carving into the air as if in a silent and graceful dialogue with an unseen force.", + "seed": 28571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_1/index_7647_seed_28571_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0665309812873602, + "vlm_logprob": -0.0665309812873602, + "vlm_yes_no_margin": 3.374999928474426, + "vlm_token_credit": -0.0665309812873602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02342386357486248 + ], + "opposite_logprob_tokens": [ + -4.023423671722412 + ], + "logprob_reward": -0.02342386357486248, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02342386357486248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18991617858409882, + 0.12415725737810135, + 0.03561163693666458, + 0.05181247740983963, + 0.11773762106895447, + 0.08439966291189194, + 0.06598123908042908, + 0.047443583607673645, + 0.26293516159057617, + 0.020005136728286743 + ], + "credit_map": [ + 0.02390192821621895, + 0.03323403000831604, + 0.021182261407375336, + 0.015977315604686737, + 0.015127794817090034, + 0.018657900393009186, + 0.027645599097013474, + 0.03418934717774391, + 0.0051855845376849174, + 0.004935916978865862, + 0.02494930475950241, + 0.010629815049469471, + 0.014154959470033646, + 0.01007644273340702, + 0.02846604399383068, + 0.02575918845832348, + 0.002091915113851428, + 0.0032754549756646156, + 0.004904535133391619, + 0.005688443314284086, + 0.004745746962726116, + 0.00302553060464561, + 0.006461618933826685, + 0.005418390966951847, + 0.0029938367661088705, + 0.0028980602510273457, + 0.007815754972398281, + 0.008825543336570263, + 0.005779728293418884, + 0.0057106479071080685, + 0.010616496205329895, + 0.007172406651079655, + 0.0047060782089829445, + 0.0040839482098817825, + 0.038473695516586304, + 0.019841725006699562, + 0.005433389917016029, + 0.003404183778911829, + 0.029234297573566437, + 0.012560303322970867, + 0.00666105467826128, + 0.007882768288254738, + 0.009128959849476814, + 0.02168843150138855, + 0.010881982743740082, + 0.006443270482122898, + 0.012871804647147655, + 0.008841384202241898, + 0.0032132435590028763, + 0.0038355449214577675, + 0.009149855934083462, + 0.010893621481955051, + 0.011310338042676449, + 0.006340303458273411, + 0.01388105284422636, + 0.0073572807013988495, + 0.006754524540156126, + 0.00409989757463336, + 0.006127124652266502, + 0.005490573588758707, + 0.005183096509426832, + 0.004507651086896658, + 0.009063588455319405, + 0.0062171281315386295, + 0.03313904628157616, + 0.0197068452835083, + 0.03742886334657669, + 0.04651395231485367, + 0.01552261970937252, + 0.02779386378824711, + 0.035410646349191666, + 0.04741932079195976, + 0.0013157513458281755, + 0.000914236472453922, + 0.0017999261617660522, + 0.0019447949016466737, + 0.0025009112432599068, + 0.003908093087375164, + 0.004238728899508715, + 0.0033826951403170824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.02342386357486248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217865653336048 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.014217865653336048, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217865653336048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18948736786842346, + 0.07292381674051285, + 0.04083118587732315, + 0.061305105686187744, + 0.06829050183296204, + 0.06853554397821426, + 0.113123819231987, + 0.2369787096977234, + 0.09533979743719101, + 0.053184106945991516 + ], + "credit_map": [ + 0.014954308047890663, + 0.021916763857007027, + 0.01863931119441986, + 0.027596747502684593, + 0.026080980896949768, + 0.019857225939631462, + 0.0341506227850914, + 0.026291409507393837, + 0.004344915971159935, + 0.005368038080632687, + 0.011659665033221245, + 0.015140029601752758, + 0.009046523831784725, + 0.008347771130502224, + 0.010174683295190334, + 0.008842190727591515, + 0.002770165214315057, + 0.0020842705853283405, + 0.005141630303114653, + 0.006436874624341726, + 0.005715672858059406, + 0.0036763125099241734, + 0.008420715108513832, + 0.006585540249943733, + 0.003383771050721407, + 0.003404873888939619, + 0.0071243601851165295, + 0.007698795758187771, + 0.008331261575222015, + 0.010789765045046806, + 0.011033300310373306, + 0.009538977406919003, + 0.0034551071003079414, + 0.003753902856260538, + 0.01563500240445137, + 0.009511630982160568, + 0.00985482893884182, + 0.004800110589712858, + 0.012831134721636772, + 0.008448784239590168, + 0.005611051339656115, + 0.008179593831300735, + 0.009410536848008633, + 0.017363768070936203, + 0.006243268959224224, + 0.005080163013190031, + 0.010194762609899044, + 0.006452405359596014, + 0.0073175071738660336, + 0.007701218593865633, + 0.015869036316871643, + 0.018426239490509033, + 0.01859200932085514, + 0.011725525371730328, + 0.018026215955615044, + 0.015466063283383846, + 0.01545258890837431, + 0.011098618619143963, + 0.033700693398714066, + 0.03702525794506073, + 0.02583443373441696, + 0.02258540317416191, + 0.05450531095266342, + 0.036776382476091385, + 0.009325206279754639, + 0.006763736717402935, + 0.011092185974121094, + 0.02038300782442093, + 0.006030621938407421, + 0.010471592657268047, + 0.014673394151031971, + 0.016600048169493675, + 0.003034429159015417, + 0.0014292155392467976, + 0.002981260186061263, + 0.0038218910340219736, + 0.005481818225234747, + 0.008271349593997002, + 0.017984161153435707, + 0.010179980657994747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.014217865653336048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628831081092358 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628831081092358, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628831081092358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14563590288162231, + 0.15545745193958282, + 0.13149574398994446, + 0.08720619231462479, + 0.21028128266334534, + 0.021521812304854393, + 0.052836962044239044, + 0.09786298125982285, + 0.03485630825161934, + 0.06284534931182861 + ], + "credit_map": [ + 0.00917467288672924, + 0.013391920365393162, + 0.02549579180777073, + 0.023434920236468315, + 0.015533892437815666, + 0.015521806664764881, + 0.017855234444141388, + 0.02522767148911953, + 0.00806987751275301, + 0.012549234554171562, + 0.028364000841975212, + 0.02606392093002796, + 0.008198141120374203, + 0.012996692210435867, + 0.030928997322916985, + 0.028286591172218323, + 0.01957901567220688, + 0.013236015103757381, + 0.01810980774462223, + 0.028085490688681602, + 0.009544729255139828, + 0.008084826171398163, + 0.019284438341856003, + 0.015571427531540394, + 0.00769613916054368, + 0.00864352099597454, + 0.0123085742816329, + 0.016932638362050056, + 0.007643571123480797, + 0.005698794964700937, + 0.015111726708710194, + 0.013171227648854256, + 0.011687559075653553, + 0.018980760127305984, + 0.07063204795122147, + 0.024822521954774857, + 0.009602044709026814, + 0.0046652620658278465, + 0.0491085983812809, + 0.020782483741641045, + 0.0015329372836276889, + 0.002131060231477022, + 0.0026184904854744673, + 0.004826672840863466, + 0.002899422310292721, + 0.002125805476680398, + 0.003188099479302764, + 0.0021993243135511875, + 0.003164771245792508, + 0.003906702622771263, + 0.009252957068383694, + 0.008610310032963753, + 0.006441842764616013, + 0.003686996642500162, + 0.011987809091806412, + 0.005785576067864895, + 0.01148731354624033, + 0.008712793700397015, + 0.01689351722598076, + 0.010214478708803654, + 0.008424109779298306, + 0.012910687364637852, + 0.015524371527135372, + 0.013695714063942432, + 0.00375154335051775, + 0.003306684782728553, + 0.003946526441723108, + 0.006952066905796528, + 0.003056683111935854, + 0.0037649148143827915, + 0.004394474904984236, + 0.005683414172381163, + 0.007576917763799429, + 0.003921876661479473, + 0.006322775036096573, + 0.007220790721476078, + 0.004884851630777121, + 0.006130799185484648, + 0.0122142992913723, + 0.014573036693036556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.007628831081092358 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034983098506927 + ], + "opposite_logprob_tokens": [ + -1.9103498458862305 + ], + "logprob_reward": -0.16034983098506927, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034983098506927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1678871065378189, + 0.14417967200279236, + 0.05784844234585762, + 0.049218952655792236, + 0.19538746774196625, + 0.11508757621049881, + 0.12645693123340607, + 0.06423932313919067, + 0.035507671535015106, + 0.044186897575855255 + ], + "credit_map": [ + 0.01554830837994814, + 0.02498391456902027, + 0.024618908762931824, + 0.02080988697707653, + 0.014858951792120934, + 0.016142282634973526, + 0.02493934892117977, + 0.025985492393374443, + 0.005707074422389269, + 0.006113077513873577, + 0.02313213236629963, + 0.012458735145628452, + 0.020699696615338326, + 0.012011946178972721, + 0.03796646371483803, + 0.026090538129210472, + 0.0035363060887902975, + 0.0031425675842911005, + 0.007502286694943905, + 0.00825588870793581, + 0.008215462788939476, + 0.005805037450045347, + 0.011200154200196266, + 0.010190737433731556, + 0.0026644417084753513, + 0.003467542352154851, + 0.007937444373965263, + 0.0074435328133404255, + 0.006244621705263853, + 0.007074456661939621, + 0.006966767832636833, + 0.007420143578201532, + 0.007086720783263445, + 0.00905168429017067, + 0.053029853850603104, + 0.02993657998740673, + 0.019326724112033844, + 0.007992027327418327, + 0.047529205679893494, + 0.021434659138321877, + 0.0092342309653759, + 0.013349534012377262, + 0.01664073020219803, + 0.028472308069467545, + 0.011155658401548862, + 0.008095931261777878, + 0.016427798196673393, + 0.01171138696372509, + 0.008756961673498154, + 0.009121036157011986, + 0.016413623467087746, + 0.020066875964403152, + 0.015617948956787586, + 0.016065767034888268, + 0.022243404760956764, + 0.018171312287449837, + 0.004180478397756815, + 0.0025864976923912764, + 0.00798027589917183, + 0.008453418500721455, + 0.008576862514019012, + 0.008327147923409939, + 0.012140586972236633, + 0.011994055472314358, + 0.004219448659569025, + 0.004393537528812885, + 0.003956445958465338, + 0.00488263089209795, + 0.0037704622372984886, + 0.004191570915281773, + 0.004640102852135897, + 0.005453471560031176, + 0.0018305722624063492, + 0.0010255874367430806, + 0.004215093329548836, + 0.004643939435482025, + 0.004485294222831726, + 0.005125440191477537, + 0.014818494208157063, + 0.008042478002607822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.16034983098506927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a minimalist dojo with a highly polished, light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing loose, flowing off-white silk garments with long black hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show soft, golden sunlight streaming through a large shoji paper screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the practitioners moving in a continuous, circular pattern with fluid, serpentine arm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and full-body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270345151424408 + ], + "opposite_logprob_tokens": [ + -2.1270344257354736 + ], + "logprob_reward": -0.1270345151424408, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270345151424408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1709083765745163, + 0.0950670912861824, + 0.03467147424817085, + 0.04178014025092125, + 0.05454229190945625, + 0.16555164754390717, + 0.06979221105575562, + 0.035146575421094894, + 0.23169012367725372, + 0.10085005313158035 + ], + "credit_map": [ + 0.00920156016945839, + 0.01633032038807869, + 0.03108757920563221, + 0.03617798909544945, + 0.015716491267085075, + 0.012455635704100132, + 0.027399323880672455, + 0.022539475932717323, + 0.004960671998560429, + 0.004511076956987381, + 0.03257084637880325, + 0.011823463253676891, + 0.004855461418628693, + 0.005129888653755188, + 0.009279065765440464, + 0.02193661406636238, + 0.0015256976475939155, + 0.0019216490909457207, + 0.004931998904794455, + 0.004501670598983765, + 0.005601038224995136, + 0.0027836826629936695, + 0.007041589822620153, + 0.006364145781844854, + 0.0019162684911862016, + 0.0021635598968714476, + 0.005545056425035, + 0.006781540345400572, + 0.006136121228337288, + 0.004219232127070427, + 0.007987762801349163, + 0.007030600216239691, + 0.0019910617265850306, + 0.0025750703644007444, + 0.013443690724670887, + 0.009456059895455837, + 0.004407680127769709, + 0.0028386865742504597, + 0.012533195316791534, + 0.007296842057257891, + 0.007897053845226765, + 0.013293781317770481, + 0.0285038985311985, + 0.05122336372733116, + 0.016399679705500603, + 0.009055580012500286, + 0.025362486019730568, + 0.013815788552165031, + 0.0036710472777485847, + 0.005408464930951595, + 0.013647962361574173, + 0.008711952716112137, + 0.005882906727492809, + 0.0046296450309455395, + 0.019420143216848373, + 0.008420094847679138, + 0.001990882446989417, + 0.0020200349390506744, + 0.006092010997235775, + 0.005936697591096163, + 0.0029891589656472206, + 0.004198620095849037, + 0.007146441377699375, + 0.004772727843374014, + 0.017754405736923218, + 0.02228722721338272, + 0.02482590824365616, + 0.032273124903440475, + 0.01717851497232914, + 0.02115749567747116, + 0.05676289647817612, + 0.03945053368806839, + 0.00528583163395524, + 0.002495129592716694, + 0.006349760573357344, + 0.008643511682748795, + 0.009553148411214352, + 0.014075199142098427, + 0.03507855162024498, + 0.019368918612599373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.1270345151424408 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_tensor([6743, 6743])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_tensor([6743, 6743])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb847349f5ad6535541084e03440c2137bb76b4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_tensor([6743, 6743])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle drone shot captures a tightly knit group of professional cyclists accelerating down a winding country road, the camera gliding smoothly to keep pace with the lead rider and emphasize their speed. The cyclists, wearing skin-tight jerseys in a range of vibrant, saturated colors like electric blue, neon yellow, and fiery red, hunch low in an aggressive, aerodynamic stance, perched atop the drop bars of their sleek carbon fiber racing bikes. Their legs beat in a powerful, synchronized rhythm, and blurs of motion propel them forward as they lean into a sweeping bend in perfect unison. The road itself is a smooth, dark ribbon of asphalt, carving sinuous curves through a breathtaking landscape of winding, lush green hills beneath a clear blue sky. The entire scene is bathed in the warm golden rays of the afternoon sun, which gleams off the polished bicycle frames and casts long, dramatic shadows that race alongside the bikes, creating an exhilarating atmosphere of speed, focus, and competition as the pack moves like a single, fluid organism against a serene, verdant backdrop.", + "seed": 3159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_6743_seed_3159_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06127568245865404, + "vlm_logprob": -0.06127568245865404, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06127568245865404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20122307538986206 + ], + "opposite_logprob_tokens": [ + -2.201223134994507 + ], + "logprob_reward": -0.20122307538986206, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.20122307538986206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03717774152755737, + 0.009822987951338291, + 0.024320228025317192, + 0.0629769116640091, + 0.022602498531341553, + 0.03749489039182663, + 0.48460009694099426, + 0.2818758189678192, + 0.02882775105535984, + 0.010301036760210991 + ], + "credit_map": [ + 0.0038657074328511953, + 0.004927474539726973, + 0.005688824690878391, + 0.006075562909245491, + 0.0032360570039600134, + 0.003171936608850956, + 0.004955490119755268, + 0.00525668915361166, + 0.0012454611714929342, + 0.001416154089383781, + 0.0009845959721133113, + 0.0013893682043999434, + 0.001212545670568943, + 0.0006978529854677618, + 0.001618731301277876, + 0.0012582790805026889, + 0.0023072361946105957, + 0.0039843362756073475, + 0.0032335235737264156, + 0.003811214817687869, + 0.005040629301220179, + 0.0019085489911958575, + 0.002642574720084667, + 0.0013921639183536172, + 0.010510897263884544, + 0.005139498505741358, + 0.007968035526573658, + 0.0025921857450157404, + 0.026153113692998886, + 0.004812283907085657, + 0.0032212315127253532, + 0.0025796634145081043, + 0.0036284448578953743, + 0.0024013561196625233, + 0.00480392063036561, + 0.0033972649835050106, + 0.0014894771156832576, + 0.0028059189207851887, + 0.0020527979359030724, + 0.0020233187824487686, + 0.004818095825612545, + 0.003988593351095915, + 0.0033010891638696194, + 0.0044740960001945496, + 0.004868427291512489, + 0.005380632355809212, + 0.006910174619406462, + 0.003753781085833907, + 0.025553561747074127, + 0.01890263892710209, + 0.09949205070734024, + 0.03134686127305031, + 0.030356109142303467, + 0.20789842307567596, + 0.0161750391125679, + 0.05487539991736412, + 0.011816921643912792, + 0.010511310771107674, + 0.06911826878786087, + 0.016644343733787537, + 0.008716212585568428, + 0.029738740995526314, + 0.019478019326925278, + 0.11585201323032379, + 0.0019328147172927856, + 0.0018851285567507148, + 0.004604008048772812, + 0.005748652387410402, + 0.0020066951401531696, + 0.0029011068399995565, + 0.005584906321018934, + 0.0041644386947155, + 0.000714461668394506, + 0.000675760384183377, + 0.0014154670061543584, + 0.002328203758224845, + 0.0008681081817485392, + 0.0010355120757594705, + 0.002008432988077402, + 0.0012550908140838146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.20122307538986206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055022332817316055 + ], + "opposite_logprob_tokens": [ + -2.9300222396850586 + ], + "logprob_reward": -0.055022332817316055, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055022332817316055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022641289979219437, + 0.007836559787392616, + 0.01537352241575718, + 0.09407603740692139, + 0.03300842270255089, + 0.022031210362911224, + 0.4191400408744812, + 0.35934072732925415, + 0.009009083732962608, + 0.017543062567710876 + ], + "credit_map": [ + 0.0019015352008864284, + 0.002829022239893675, + 0.003944799769669771, + 0.005844219122081995, + 0.0015836048405617476, + 0.0014963540015742183, + 0.0024636220186948776, + 0.0025781348813325167, + 0.0007995377527549863, + 0.0010460709454491735, + 0.0009234045865014195, + 0.0019832069519907236, + 0.000701151613611728, + 0.0005181656451895833, + 0.0011006435379385948, + 0.0007643787539564073, + 0.0014688216615468264, + 0.0029166850727051497, + 0.0016399736050516367, + 0.0015229977434501052, + 0.003293161978945136, + 0.0010745746549218893, + 0.002046376233920455, + 0.001410932745784521, + 0.018007399514317513, + 0.008045162074267864, + 0.010782874189317226, + 0.003611798631027341, + 0.0391794890165329, + 0.006107483059167862, + 0.004087846726179123, + 0.004254000261425972, + 0.004933900665491819, + 0.002536989049986005, + 0.007838726975023746, + 0.0025898630265146494, + 0.002509767422452569, + 0.003879845142364502, + 0.0061324783600866795, + 0.0025868548545986414, + 0.0023649861104786396, + 0.00287540047429502, + 0.0014576228568330407, + 0.002016780897974968, + 0.003633067011833191, + 0.003086533397436142, + 0.005025410559028387, + 0.0015714119654148817, + 0.026515085250139236, + 0.016602205112576485, + 0.10554464906454086, + 0.031840018928050995, + 0.027148103341460228, + 0.1651943325996399, + 0.019311411306262016, + 0.02698426879942417, + 0.015622747130692005, + 0.01573142036795616, + 0.08641685545444489, + 0.022228099405765533, + 0.012782655656337738, + 0.03677792847156525, + 0.025211824104189873, + 0.1445692479610443, + 0.0007165665156207979, + 0.0008574582170695066, + 0.0012605059891939163, + 0.0015797720989212394, + 0.0009320934768766165, + 0.0011691015679389238, + 0.001134994556196034, + 0.0013585923006758094, + 0.0011681091273203492, + 0.0011435834458097816, + 0.0020816070027649403, + 0.005433444865047932, + 0.0009632001165300608, + 0.0014140225248411298, + 0.0031818051356822252, + 0.002157290931791067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.055022332817316055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636326290667057 + ], + "opposite_logprob_tokens": [ + -3.6513633728027344 + ], + "logprob_reward": -0.02636326290667057, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636326290667057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0487876795232296, + 0.016758741810917854, + 0.02446991577744484, + 0.21220816671848297, + 0.03266998752951622, + 0.03897674381732941, + 0.1546654850244522, + 0.372882604598999, + 0.03751448541879654, + 0.061066191643476486 + ], + "credit_map": [ + 0.003852920373901725, + 0.007275828160345554, + 0.007566599640995264, + 0.008670952171087265, + 0.005467646289616823, + 0.004107504617422819, + 0.006744830869138241, + 0.005101403687149286, + 0.0016743845771998167, + 0.002929354552179575, + 0.0017774361185729504, + 0.002849570708349347, + 0.00207747844979167, + 0.0011530729243531823, + 0.002467008540406823, + 0.0018304369878023863, + 0.002418386982753873, + 0.004864824935793877, + 0.002821501577273011, + 0.002431632950901985, + 0.005536943208426237, + 0.0018377200467512012, + 0.0026953364722430706, + 0.001863570068962872, + 0.031349390745162964, + 0.01821540854871273, + 0.030380740761756897, + 0.01036809477955103, + 0.0858110710978508, + 0.016231967136263847, + 0.011735721491277218, + 0.00811578705906868, + 0.003940316382795572, + 0.0027361921966075897, + 0.0034122122451663017, + 0.0037376713007688522, + 0.003718515858054161, + 0.004836590029299259, + 0.005598035175353289, + 0.004690458066761494, + 0.0055894386023283005, + 0.0038525189738720655, + 0.004621645901352167, + 0.003812511684373021, + 0.0043074642308056355, + 0.005002503748983145, + 0.006439961027354002, + 0.00535069964826107, + 0.004415048751980066, + 0.005496810656040907, + 0.0065833223052322865, + 0.009080241434276104, + 0.0072724297642707825, + 0.06582222133874893, + 0.0055894553661346436, + 0.050405971705913544, + 0.015493189916014671, + 0.013961143791675568, + 0.09081802517175674, + 0.02173999696969986, + 0.011718091554939747, + 0.03861132636666298, + 0.026966743171215057, + 0.15357407927513123, + 0.0023471859749406576, + 0.004080954939126968, + 0.004336177837103605, + 0.006864563096314669, + 0.00427696667611599, + 0.0050278701819479465, + 0.004874443169683218, + 0.005706322379410267, + 0.003489741124212742, + 0.0032412640284746885, + 0.007931319065392017, + 0.016308292746543884, + 0.00373843009583652, + 0.006482897326350212, + 0.012224829755723476, + 0.007649424485862255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.02636326290667057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04852331057190895, + 0.015576464124023914, + 0.009919103235006332, + 0.15042977035045624, + 0.015390430577099323, + 0.04376361891627312, + 0.5759437084197998, + 0.12584109604358673, + 0.007073889020830393, + 0.007538581732660532 + ], + "credit_map": [ + 0.004115671385079622, + 0.010003402829170227, + 0.007429353427141905, + 0.0091714384034276, + 0.004797388333827257, + 0.004064312670379877, + 0.0050419652834534645, + 0.0038997780065983534, + 0.0021426689345389605, + 0.003432982601225376, + 0.0017333169234916568, + 0.0014493479393422604, + 0.0026030386798083782, + 0.0013643598649650812, + 0.001699843560345471, + 0.0011509059695526958, + 0.0012510152300819755, + 0.0017830902943387628, + 0.0012205701787024736, + 0.0011340617202222347, + 0.0015360707184299827, + 0.0010753237875178456, + 0.0009471497614867985, + 0.0009718209621496499, + 0.026543602347373962, + 0.012527277693152428, + 0.017220860347151756, + 0.006572859361767769, + 0.06259214133024216, + 0.010843991301953793, + 0.008149277418851852, + 0.005979756359010935, + 0.0021693764720112085, + 0.001482775085605681, + 0.0032839702907949686, + 0.00220623891800642, + 0.0012557755690068007, + 0.002003449946641922, + 0.0017223863396793604, + 0.0012664591195061803, + 0.007536731660366058, + 0.0034964371006935835, + 0.01153706293553114, + 0.005229900125414133, + 0.0030614500865340233, + 0.006398823577910662, + 0.0032491106539964676, + 0.0032541011460125446, + 0.031140334904193878, + 0.021586786955595016, + 0.1223117932677269, + 0.03800135478377342, + 0.03602147102355957, + 0.24655506014823914, + 0.01926480047404766, + 0.061062101274728775, + 0.005276544950902462, + 0.003552609821781516, + 0.031958840787410736, + 0.00789901427924633, + 0.003992584999650717, + 0.012890209443867207, + 0.008970764465630054, + 0.051300518214702606, + 0.0004919075290672481, + 0.000570282805711031, + 0.0008989706402644515, + 0.0011548653710633516, + 0.0008170740911737084, + 0.0013048640685155988, + 0.0008979729027487338, + 0.0009379521361552179, + 0.0005675730644725263, + 0.0007151183090172708, + 0.0008163996390067041, + 0.001313771354034543, + 0.0007350232335738838, + 0.0014869211008772254, + 0.001019200892187655, + 0.0008845737320370972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057693526148796 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.02057693526148796, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057693526148796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027202973142266273, + 0.013670569285750389, + 0.03862380236387253, + 0.12201434373855591, + 0.04112955182790756, + 0.07205242663621902, + 0.32895705103874207, + 0.3195331394672394, + 0.0175552349537611, + 0.01926085539162159 + ], + "credit_map": [ + 0.0028953265864402056, + 0.0035148579627275467, + 0.004084452986717224, + 0.004987530875951052, + 0.002386388136073947, + 0.0023649504873901606, + 0.003617041278630495, + 0.0033524257596582174, + 0.0016457331366837025, + 0.0017321809427812696, + 0.0016387924551963806, + 0.002277777995914221, + 0.0016872999258339405, + 0.0007930340943858027, + 0.002131490036845207, + 0.0017642602324485779, + 0.003826046595349908, + 0.008599485270678997, + 0.004627279005944729, + 0.0031584480311721563, + 0.010096156969666481, + 0.0031412371899932623, + 0.0028990430291742086, + 0.0022761065047234297, + 0.021655907854437828, + 0.010203812271356583, + 0.01474974025040865, + 0.004701600410044193, + 0.05173903703689575, + 0.007836776785552502, + 0.006236380897462368, + 0.0048910826444625854, + 0.006944393273442984, + 0.002636394463479519, + 0.011698886752128601, + 0.004152611829340458, + 0.0023783775977790356, + 0.005873632151633501, + 0.003383593400940299, + 0.004061662591993809, + 0.008165490813553333, + 0.0065361796878278255, + 0.008349830284714699, + 0.008102948777377605, + 0.007986252196133137, + 0.009483087807893753, + 0.011711901053786278, + 0.011716739274561405, + 0.010068838484585285, + 0.011216232553124428, + 0.021682754158973694, + 0.02020219713449478, + 0.016050146892666817, + 0.14156697690486908, + 0.012116448022425175, + 0.09605348110198975, + 0.013665391132235527, + 0.009967510588467121, + 0.0803992822766304, + 0.019584910944104195, + 0.010276934131979942, + 0.03331654518842697, + 0.022058486938476562, + 0.1302640587091446, + 0.0010034129954874516, + 0.0014495414216071367, + 0.0022079080808907747, + 0.003970901016145945, + 0.0016037869499996305, + 0.0015793745405972004, + 0.0037874476984143257, + 0.0019528622506186366, + 0.0011894848430529237, + 0.0009153321152552962, + 0.003125387942418456, + 0.00446635577827692, + 0.001039306283928454, + 0.0017144952435046434, + 0.004192138556391001, + 0.0026183549780398607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.02057693526148796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle drone shot captures a tightly knit group of professional cyclists accelerating down a winding country road, the camera gliding smoothly to keep pace with the lead rider and emphasize their speed. The cyclists, wearing skin-tight jerseys in a range of vibrant, saturated colors like electric blue, neon yellow, and fiery red, hunch low in an aggressive, aerodynamic stance, perched atop the drop bars of their sleek carbon fiber racing bikes. Their legs beat in a powerful, synchronized rhythm, and blurs of motion propel them forward as they lean into a sweeping bend in perfect unison. The road itself is a smooth, dark ribbon of asphalt, carving sinuous curves through a breathtaking landscape of winding, lush green hills beneath a clear blue sky. The entire scene is bathed in the warm golden rays of the afternoon sun, which gleams off the polished bicycle frames and casts long, dramatic shadows that race alongside the bikes, creating an exhilarating atmosphere of speed, focus, and competition as the pack moves like a single, fluid organism against a serene, verdant backdrop.", + "seed": 3160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_6743_seed_3160_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11372907352633774, + "vlm_logprob": -0.11372907352633774, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.11372907352633774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20571503043174744 + ], + "opposite_logprob_tokens": [ + -1.9557150602340698 + ], + "logprob_reward": -0.20571503043174744, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.20571503043174744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10633112490177155, + 0.0928083062171936, + 0.1369743049144745, + 0.07729753851890564, + 0.20558889210224152, + 0.1339869350194931, + 0.11050959676504135, + 0.018746720626950264, + 0.08580423891544342, + 0.0319523923099041 + ], + "credit_map": [ + 0.011293362826108932, + 0.00836749467998743, + 0.015526633709669113, + 0.026857823133468628, + 0.007064449600875378, + 0.00781321246176958, + 0.018506793305277824, + 0.010901354253292084, + 0.00909099355340004, + 0.0061316643841564655, + 0.014146162196993828, + 0.020344793796539307, + 0.0053375535644590855, + 0.008899974636733532, + 0.015219002030789852, + 0.013638154603540897, + 0.021543635055422783, + 0.00866594910621643, + 0.019687343388795853, + 0.05064964294433594, + 0.003769591683521867, + 0.0061208033002913, + 0.01830437034368515, + 0.008232948370277882, + 0.006208359729498625, + 0.010192914865911007, + 0.018526485189795494, + 0.009049446322023869, + 0.0067742750979959965, + 0.009058893658220768, + 0.008934449404478073, + 0.008552703075110912, + 0.01598404534161091, + 0.007248929236084223, + 0.08460832387208939, + 0.044228557497262955, + 0.008192126639187336, + 0.011724274605512619, + 0.00924009457230568, + 0.02436251752078533, + 0.013846233487129211, + 0.008509143255650997, + 0.014356984756886959, + 0.017665216699242592, + 0.01364388968795538, + 0.021858878433704376, + 0.025312257930636406, + 0.018794329836964607, + 0.012542960233986378, + 0.00616643950343132, + 0.03608739748597145, + 0.024023281410336494, + 0.006159306038171053, + 0.010166146792471409, + 0.0059555908665061, + 0.009408477693796158, + 0.0015542894834652543, + 0.0016065506497398019, + 0.0036260492634028196, + 0.0039374190382659435, + 0.0017394544556736946, + 0.0020896131172776222, + 0.002731632674112916, + 0.001461709849536419, + 0.015896210446953773, + 0.029377488419413567, + 0.00621470483019948, + 0.017955129966139793, + 0.005304140970110893, + 0.002919397782534361, + 0.006171190179884434, + 0.00196596491150558, + 0.0027497222181409597, + 0.004266147967427969, + 0.006243550218641758, + 0.007748438511043787, + 0.001678397529758513, + 0.002799778711050749, + 0.003802570980042219, + 0.002663783263415098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.20571503043174744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351199746131897 + ], + "opposite_logprob_tokens": [ + -1.3135119676589966 + ], + "logprob_reward": -0.31351199746131897, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31351199746131897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10183399170637131, + 0.05256282538175583, + 0.0679430440068245, + 0.047483690083026886, + 0.3776074945926666, + 0.11875404417514801, + 0.1060032770037651, + 0.022094370797276497, + 0.08095168322324753, + 0.024765467271208763 + ], + "credit_map": [ + 0.010582339018583298, + 0.010280939750373363, + 0.010601561516523361, + 0.016473354771733284, + 0.010596107691526413, + 0.014233139343559742, + 0.015237229876220226, + 0.013829324394464493, + 0.003415697021409869, + 0.0034258391242474318, + 0.00632554991170764, + 0.009442860260605812, + 0.005084628239274025, + 0.006155737210065126, + 0.011319386772811413, + 0.007393132895231247, + 0.008547580800950527, + 0.0042887465097010136, + 0.010090352967381477, + 0.023995814844965935, + 0.0035307116340845823, + 0.0037964340299367905, + 0.009171603247523308, + 0.004521805327385664, + 0.003614462912082672, + 0.0036428971216082573, + 0.013258964754641056, + 0.008360488340258598, + 0.002409227192401886, + 0.00409523444250226, + 0.006825451273471117, + 0.005276970099657774, + 0.026949232444167137, + 0.013198194094002247, + 0.1424885392189026, + 0.08150642365217209, + 0.01609978824853897, + 0.03091605007648468, + 0.027796659618616104, + 0.03865264728665352, + 0.006978274788707495, + 0.008224064484238625, + 0.032582249492406845, + 0.01930324360728264, + 0.0074078054167330265, + 0.010607896372675896, + 0.01740187779068947, + 0.016248639672994614, + 0.00869633350521326, + 0.005454889964312315, + 0.030199943110346794, + 0.021579312160611153, + 0.008291342295706272, + 0.013545956462621689, + 0.008180919103324413, + 0.010054590180516243, + 0.002079885685816407, + 0.002627486130222678, + 0.0027343740221112967, + 0.0033422044944018126, + 0.0024902622681111097, + 0.004717547912150621, + 0.0026671120431274176, + 0.0014354975428432226, + 0.01695152558386326, + 0.02845284901559353, + 0.009492692537605762, + 0.008929354138672352, + 0.005378658417612314, + 0.003941877745091915, + 0.005598981864750385, + 0.0022057490423321724, + 0.0025348414201289415, + 0.004088541958481073, + 0.0031812945380806923, + 0.005725334398448467, + 0.0017064412822946906, + 0.0027092834934592247, + 0.003242500824853778, + 0.0015772318001836538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.31351199746131897 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354322209954262 + ], + "opposite_logprob_tokens": [ + -3.6513543128967285 + ], + "logprob_reward": -0.026354322209954262, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354322209954262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1282506138086319, + 0.10870525985956192, + 0.4417819678783417, + 0.03107822872698307, + 0.08216176927089691, + 0.06204056367278099, + 0.052937351167201996, + 0.014980090782046318, + 0.04031844437122345, + 0.037745777517557144 + ], + "credit_map": [ + 0.016148369759321213, + 0.011953487992286682, + 0.014028641395270824, + 0.035172704607248306, + 0.007771061733365059, + 0.009219910018146038, + 0.022296881303191185, + 0.011659553274512291, + 0.011002461425960064, + 0.008770124055445194, + 0.016461782157421112, + 0.02589188702404499, + 0.006297055631875992, + 0.011016258038580418, + 0.01708182878792286, + 0.012183857150375843, + 0.042856503278017044, + 0.018193503841757774, + 0.08966878801584244, + 0.1815275400876999, + 0.009480464272201061, + 0.014786913990974426, + 0.06443938612937927, + 0.020828861743211746, + 0.0022173719480633736, + 0.0023550132755190134, + 0.005948841571807861, + 0.004058447666466236, + 0.0018542914185672998, + 0.002729603787884116, + 0.005224316846579313, + 0.006690343376249075, + 0.007926798425614834, + 0.004057861398905516, + 0.027844691649079323, + 0.016119828447699547, + 0.004281677771359682, + 0.006384008098393679, + 0.009549926966428757, + 0.005996978376060724, + 0.004067020490765572, + 0.005226060748100281, + 0.007433878257870674, + 0.00961680244654417, + 0.0052907587960362434, + 0.008249377831816673, + 0.014562807977199554, + 0.007593862246721983, + 0.004080035723745823, + 0.0027165059000253677, + 0.017120791599154472, + 0.011583371087908745, + 0.0037225757259875536, + 0.005361172836273909, + 0.004356277175247669, + 0.003996624611318111, + 0.0018354562344029546, + 0.0016882759518921375, + 0.0026650824584066868, + 0.0020093810744583607, + 0.0015244893729686737, + 0.0024539220612496138, + 0.0016666555311530828, + 0.0011368291452527046, + 0.008000834845006466, + 0.014294699765741825, + 0.0032558452803641558, + 0.007049663458019495, + 0.0025315522216260433, + 0.001661772606894374, + 0.00256920768879354, + 0.0009548663510940969, + 0.007645239122211933, + 0.012367861345410347, + 0.002413015114143491, + 0.003944106865674257, + 0.0030858723912388086, + 0.0032735634595155716, + 0.0030972016975283623, + 0.001918917871080339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.026354322209954262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024878999684005976 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024878999684005976, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024878999684005976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10064369440078735, + 0.049412500113248825, + 0.36786308884620667, + 0.047170452773571014, + 0.18972782790660858, + 0.03872592747211456, + 0.026919057592749596, + 0.010623586364090443, + 0.13193368911743164, + 0.03698013722896576 + ], + "credit_map": [ + 0.008024502545595169, + 0.010135043412446976, + 0.01385386660695076, + 0.016775421798229218, + 0.008866720832884312, + 0.016024945303797722, + 0.010545575991272926, + 0.016417622566223145, + 0.004489494021981955, + 0.00407770462334156, + 0.010860196314752102, + 0.010086908005177975, + 0.0027400916442275047, + 0.005556903779506683, + 0.005672026425600052, + 0.005929175764322281, + 0.04362265765666962, + 0.017531950026750565, + 0.06669516861438751, + 0.14884644746780396, + 0.007699177134782076, + 0.01320108026266098, + 0.0525052584707737, + 0.017761370167136192, + 0.0036394486669451, + 0.004464765079319477, + 0.008632021956145763, + 0.008075500838458538, + 0.004056388512253761, + 0.0063560050912201405, + 0.006808793172240257, + 0.005137531086802483, + 0.01111114677041769, + 0.005343107972294092, + 0.07915425300598145, + 0.040368303656578064, + 0.006158455274999142, + 0.01315977331250906, + 0.007624661549925804, + 0.02680814452469349, + 0.004791525192558765, + 0.003508821129798889, + 0.011700425297021866, + 0.005809933878481388, + 0.0024167795199900866, + 0.004123251885175705, + 0.0037804727908223867, + 0.0025947177782654762, + 0.0019407259533181787, + 0.0018018768168985844, + 0.005606605671346188, + 0.00435973796993494, + 0.0027784176636487246, + 0.004097312223166227, + 0.0035883516538888216, + 0.00274602766148746, + 0.0008840622031129897, + 0.0007685020682401955, + 0.0025649480521678925, + 0.0021559614688158035, + 0.0009041985031217337, + 0.0013639908283948898, + 0.0012158947065472603, + 0.0007660288247279823, + 0.014543619938194752, + 0.029159478843212128, + 0.015713579952716827, + 0.04275764897465706, + 0.005492021329700947, + 0.0069111306220293045, + 0.013586796820163727, + 0.0037694135680794716, + 0.006631871219724417, + 0.011564739048480988, + 0.004302676767110825, + 0.004023849964141846, + 0.003363720839843154, + 0.0026505819987505674, + 0.002675512805581093, + 0.0017671851674094796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.0024878999684005976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057611756026745 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.02057611756026745, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057611756026745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10698238760232925, + 0.07283001393079758, + 0.12252325564622879, + 0.06754644215106964, + 0.2682722508907318, + 0.08092174679040909, + 0.05144580081105232, + 0.022947940975427628, + 0.15049803256988525, + 0.05603206902742386 + ], + "credit_map": [ + 0.008587721735239029, + 0.007388860918581486, + 0.017749130725860596, + 0.020784631371498108, + 0.009599803946912289, + 0.010583780705928802, + 0.017272749915719032, + 0.015015700832009315, + 0.006757291965186596, + 0.004906232003122568, + 0.011284452863037586, + 0.017346341162919998, + 0.0035913391038775444, + 0.010003509931266308, + 0.00950117502361536, + 0.00943966954946518, + 0.008839224465191364, + 0.005218438804149628, + 0.027234820649027824, + 0.04948705807328224, + 0.003237362252548337, + 0.004637337289750576, + 0.017148364335298538, + 0.006720650475472212, + 0.0041564032435417175, + 0.005117618013173342, + 0.00922033004462719, + 0.011782679706811905, + 0.006773198023438454, + 0.0069969589821994305, + 0.01348942145705223, + 0.010009830817580223, + 0.018144890666007996, + 0.00858406350016594, + 0.11115935444831848, + 0.058990154415369034, + 0.008774235844612122, + 0.018277866765856743, + 0.013449368067085743, + 0.03089231252670288, + 0.007659513037651777, + 0.004720637109130621, + 0.019855095073580742, + 0.0120126623660326, + 0.006875285413116217, + 0.009470759890973568, + 0.012104982510209084, + 0.008222812786698341, + 0.005047131329774857, + 0.002953257877379656, + 0.015967119485139847, + 0.01066647656261921, + 0.003674762789160013, + 0.00406696368008852, + 0.005586487241089344, + 0.003483603708446026, + 0.0022526574321091175, + 0.0021231197752058506, + 0.004465768579393625, + 0.003772432217374444, + 0.0016901090275496244, + 0.0025415236596018076, + 0.0037318095564842224, + 0.0023705216590315104, + 0.019200315698981285, + 0.03953064605593681, + 0.012555061839520931, + 0.04560505971312523, + 0.007150851655751467, + 0.006272517144680023, + 0.015830926597118378, + 0.00435265526175499, + 0.003914660308510065, + 0.0033456722740083933, + 0.00870521366596222, + 0.012778008356690407, + 0.0056447782553732395, + 0.010062127374112606, + 0.0063696508295834064, + 0.005211961921304464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.02057611756026745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle drone shot captures a tightly knit group of professional cyclists accelerating down a winding country road, the camera gliding smoothly to keep pace with the lead rider and emphasize their speed. The cyclists, wearing skin-tight jerseys in a range of vibrant, saturated colors like electric blue, neon yellow, and fiery red, hunch low in an aggressive, aerodynamic stance, perched atop the drop bars of their sleek carbon fiber racing bikes. Their legs beat in a powerful, synchronized rhythm, and blurs of motion propel them forward as they lean into a sweeping bend in perfect unison. The road itself is a smooth, dark ribbon of asphalt, carving sinuous curves through a breathtaking landscape of winding, lush green hills beneath a clear blue sky. The entire scene is bathed in the warm golden rays of the afternoon sun, which gleams off the polished bicycle frames and casts long, dramatic shadows that race alongside the bikes, creating an exhilarating atmosphere of speed, focus, and competition as the pack moves like a single, fluid organism against a serene, verdant backdrop.", + "seed": 3161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_6743_seed_3161_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08808605459053069, + "vlm_logprob": -0.08808605459053069, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.08808605459053069, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14835944771766663 + ], + "opposite_logprob_tokens": [ + -2.523359537124634 + ], + "logprob_reward": -0.14835944771766663, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.14835944771766663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39636293053627014, + 0.04773138836026192, + 0.10918315500020981, + 0.054814089089632034, + 0.08121667057275772, + 0.07954218983650208, + 0.06881610304117203, + 0.10789080709218979, + 0.037614814937114716, + 0.016827864572405815 + ], + "credit_map": [ + 0.035727497190237045, + 0.14659708738327026, + 0.015807999297976494, + 0.02862589992582798, + 0.020817622542381287, + 0.03273264318704605, + 0.01798241399228573, + 0.09807182103395462, + 0.00947695504873991, + 0.014955485239624977, + 0.004429818596690893, + 0.00465407595038414, + 0.0037121926434338093, + 0.002739475341513753, + 0.0032771567348390818, + 0.004486228805035353, + 0.011096622794866562, + 0.007835637778043747, + 0.00811473373323679, + 0.012166584841907024, + 0.015412413515150547, + 0.010594164952635765, + 0.028231758624315262, + 0.015731243416666985, + 0.003676108317449689, + 0.0033083984162658453, + 0.004585645627230406, + 0.006110731512308121, + 0.007406329736113548, + 0.01173305045813322, + 0.009509303607046604, + 0.008484521880745888, + 0.007284970488399267, + 0.008810687810182571, + 0.013926712796092033, + 0.01950886659324169, + 0.005837297532707453, + 0.008476468734443188, + 0.010247529484331608, + 0.0071241362020373344, + 0.00974230095744133, + 0.01613532193005085, + 0.005859958473592997, + 0.005179367493838072, + 0.007649832405149937, + 0.016480613499879837, + 0.009517060592770576, + 0.008977734483778477, + 0.007464895490556955, + 0.007264978252351284, + 0.00656815804541111, + 0.011053824797272682, + 0.004793393891304731, + 0.00854650977998972, + 0.013596854172647, + 0.009527484886348248, + 0.008111156523227692, + 0.00702551519498229, + 0.009141971357166767, + 0.01189894787967205, + 0.0062196990475058556, + 0.014842243865132332, + 0.018984537571668625, + 0.03166673704981804, + 0.002474048174917698, + 0.0020637025590986013, + 0.002644150983542204, + 0.009688753634691238, + 0.0022341192234307528, + 0.004131772089749575, + 0.010070274583995342, + 0.0043079969473183155, + 0.001609606551937759, + 0.001082929433323443, + 0.0019124385435134172, + 0.002468579448759556, + 0.0011525113368406892, + 0.002155561000108719, + 0.0030117565765976906, + 0.003434482030570507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.14835944771766663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225631445646286 + ], + "opposite_logprob_tokens": [ + -1.6006314754486084 + ], + "logprob_reward": -0.225631445646286, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225631445646286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39597904682159424, + 0.2693953216075897, + 0.07444120198488235, + 0.14671757817268372, + 0.031717196106910706, + 0.013191848993301392, + 0.010345104150474072, + 0.01146567240357399, + 0.03221859782934189, + 0.014528496190905571 + ], + "credit_map": [ + 0.028286848217248917, + 0.13104431331157684, + 0.015902206301689148, + 0.02872750349342823, + 0.023156663402915, + 0.02482123300433159, + 0.02803739532828331, + 0.11600283533334732, + 0.028522586449980736, + 0.08344583213329315, + 0.013362539932131767, + 0.013213720172643661, + 0.012688012793660164, + 0.01523577980697155, + 0.01973610371351242, + 0.08319071680307388, + 0.005566259380429983, + 0.006031033117324114, + 0.006250959355384111, + 0.01131207961589098, + 0.008596899919211864, + 0.0066400570794939995, + 0.01915816217660904, + 0.010885736905038357, + 0.010117826983332634, + 0.008446553722023964, + 0.015490049496293068, + 0.012368910014629364, + 0.01824442483484745, + 0.02954479120671749, + 0.028194647282361984, + 0.024310342967510223, + 0.0035490645095705986, + 0.005670634098351002, + 0.0028412139508873224, + 0.003966193646192551, + 0.003442195011302829, + 0.004382209852337837, + 0.0037265510763972998, + 0.0041391304694116116, + 0.0012406888417899609, + 0.0012815198861062527, + 0.0017801538342610002, + 0.0016016217414289713, + 0.0012102058390155435, + 0.0026133849751204252, + 0.0020246002823114395, + 0.0014396724291145802, + 0.0010603901464492083, + 0.0013617071090266109, + 0.0008736674208194017, + 0.002189860213547945, + 0.0005360167706385255, + 0.0008409988367930055, + 0.0016958911437541246, + 0.0017865714617073536, + 0.0012328901793807745, + 0.0014931331388652325, + 0.0009881110163405538, + 0.0015629827976226807, + 0.0006825457676313818, + 0.0017818097257986665, + 0.002657121978700161, + 0.0010670763440430164, + 0.0026082878466695547, + 0.0017546489834785461, + 0.002919364022091031, + 0.00644553080201149, + 0.002055583754554391, + 0.006158854346722364, + 0.005857192445546389, + 0.004419130738824606, + 0.0013008060632273555, + 0.0012755864299833775, + 0.0024494242388755083, + 0.0024444148875772953, + 0.0010053387377411127, + 0.00200887699611485, + 0.002426007529720664, + 0.001618039677850902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.225631445646286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663180321455 + ], + "opposite_logprob_tokens": [ + -3.0486631393432617 + ], + "logprob_reward": -0.048663180321455, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663180321455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33895227313041687, + 0.33447831869125366, + 0.022156085819005966, + 0.1265699863433838, + 0.039540376514196396, + 0.04020819813013077, + 0.009861613623797894, + 0.03965692222118378, + 0.02473619021475315, + 0.02384001389145851 + ], + "credit_map": [ + 0.03834066167473793, + 0.1200244277715683, + 0.016511794179677963, + 0.033157095313072205, + 0.02012554742395878, + 0.03420024365186691, + 0.02593296580016613, + 0.05065954849123955, + 0.0268043652176857, + 0.09250462055206299, + 0.0185881108045578, + 0.02029036357998848, + 0.013451404869556427, + 0.020650770515203476, + 0.026865603402256966, + 0.11532308906316757, + 0.0023916407953947783, + 0.0017171697691082954, + 0.0022697020322084427, + 0.0027181566692888737, + 0.0024062390439212322, + 0.0022535028401762247, + 0.005684744566679001, + 0.0027149314992129803, + 0.010761718265712261, + 0.007288392633199692, + 0.01048679742962122, + 0.01040284801274538, + 0.013615142554044724, + 0.02390417642891407, + 0.025228416547179222, + 0.024882500991225243, + 0.003697702893987298, + 0.005079456139355898, + 0.00479310704395175, + 0.005637727677822113, + 0.003839620389044285, + 0.008016383275389671, + 0.003961918409913778, + 0.004514461848884821, + 0.0038526717107743025, + 0.0034853266552090645, + 0.0054670642130076885, + 0.005152320489287376, + 0.0032652290537953377, + 0.005536078475415707, + 0.007385013625025749, + 0.006064495071768761, + 0.0012666501570492983, + 0.0011656396090984344, + 0.0007458883919753134, + 0.0013841543113812804, + 0.0005875529022887349, + 0.001903114141896367, + 0.0018553987611085176, + 0.0009532151161693037, + 0.0031189045403152704, + 0.004318153951317072, + 0.003871222259476781, + 0.003900503274053335, + 0.0019887201488018036, + 0.004791975021362305, + 0.006958890240639448, + 0.010708550922572613, + 0.002370772883296013, + 0.0022292216308414936, + 0.0032911316957324743, + 0.003725934773683548, + 0.0016865074867382646, + 0.00397418811917305, + 0.0043556103482842445, + 0.0031028238590806723, + 0.002441856311634183, + 0.002561731031164527, + 0.003672257997095585, + 0.003781002713367343, + 0.001598530332557857, + 0.004217703361064196, + 0.0024636615999042988, + 0.0031032694969326258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.048663180321455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001712523982860148 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.001712523982860148, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001712523982860148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11326967179775238, + 0.05994848161935806, + 0.02230921760201454, + 0.052087899297475815, + 0.025364387780427933, + 0.07033402472734451, + 0.054739486426115036, + 0.525113582611084, + 0.048779137432575226, + 0.028054043650627136 + ], + "credit_map": [ + 0.00894202757626772, + 0.012220408767461777, + 0.011404571123421192, + 0.021049806848168373, + 0.008466801606118679, + 0.01454204972833395, + 0.018057135865092278, + 0.01858687400817871, + 0.010194823145866394, + 0.011560590006411076, + 0.007559455931186676, + 0.0049523417837917805, + 0.004048468079417944, + 0.004387121647596359, + 0.004350344650447369, + 0.012895341962575912, + 0.0020776712335646152, + 0.0017728002276271582, + 0.0020248719956725836, + 0.0031235201749950647, + 0.003184403060004115, + 0.0021662835497409105, + 0.004264432936906815, + 0.0036952372174710035, + 0.004504899028688669, + 0.004920888692140579, + 0.005563356913626194, + 0.005429881624877453, + 0.006793903652578592, + 0.007482276763767004, + 0.00971797201782465, + 0.007674727588891983, + 0.003202903550118208, + 0.004107253160327673, + 0.0019968959968537092, + 0.0029735788702964783, + 0.0027359153609722853, + 0.005348807666450739, + 0.0025450338143855333, + 0.0024540028534829617, + 0.0060574933886528015, + 0.006817865185439587, + 0.006944489199668169, + 0.007597474846988916, + 0.005826760083436966, + 0.011040283367037773, + 0.015026789158582687, + 0.011022874154150486, + 0.004858987405896187, + 0.005355600267648697, + 0.0038861981593072414, + 0.008029458113014698, + 0.003743973560631275, + 0.006643922999501228, + 0.013827814720571041, + 0.008393538184463978, + 0.039223331958055496, + 0.02936309017241001, + 0.04519285634160042, + 0.05388883873820305, + 0.022996099665760994, + 0.05473839491605759, + 0.10226332396268845, + 0.17744769155979156, + 0.003682505339384079, + 0.0028069105464965105, + 0.0035339449532330036, + 0.01117931678891182, + 0.003363501513376832, + 0.005511379800736904, + 0.011925876140594482, + 0.006775705609470606, + 0.0020042762625962496, + 0.0014927523443475366, + 0.0031314976513385773, + 0.0035440069623291492, + 0.0021888986229896545, + 0.005306426901370287, + 0.006852274294942617, + 0.003533911192789674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.001712523982860148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606367528438568 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606367528438568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606367528438568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5037698149681091, + 0.12271320819854736, + 0.04685966297984123, + 0.09312618523836136, + 0.041415780782699585, + 0.04060105234384537, + 0.047786857932806015, + 0.062243830412626266, + 0.023063546046614647, + 0.01842004805803299 + ], + "credit_map": [ + 0.0473991259932518, + 0.19309622049331665, + 0.01265795063227415, + 0.03157642111182213, + 0.02616945467889309, + 0.04179561510682106, + 0.022292517125606537, + 0.1287825107574463, + 0.017500393092632294, + 0.04386744648218155, + 0.005184322129935026, + 0.007303858175873756, + 0.007050368469208479, + 0.005362477153539658, + 0.007779693230986595, + 0.02866465225815773, + 0.004472494125366211, + 0.0035825916565954685, + 0.006814519874751568, + 0.00531922047957778, + 0.005280559882521629, + 0.00467094499617815, + 0.010004454292356968, + 0.006714878603816032, + 0.007009871304035187, + 0.007161056622862816, + 0.01548045128583908, + 0.010175138711929321, + 0.008858340792357922, + 0.012480809353291988, + 0.019310027360916138, + 0.012650488875806332, + 0.003537416225299239, + 0.0048323436640203, + 0.007936838082969189, + 0.009873751550912857, + 0.0027768563013523817, + 0.005390816833823919, + 0.0036546802148222923, + 0.0034130795393139124, + 0.004717655014246702, + 0.00815912801772356, + 0.0032471278682351112, + 0.0038990164175629616, + 0.003553610760718584, + 0.004872264806181192, + 0.008017626591026783, + 0.004134623799473047, + 0.005216857418417931, + 0.005991319194436073, + 0.004180858843028545, + 0.007632079534232616, + 0.003458623541519046, + 0.005880169570446014, + 0.00923350639641285, + 0.006193441338837147, + 0.005246271844953299, + 0.0054183718748390675, + 0.005214938428252935, + 0.005520134698599577, + 0.004692187532782555, + 0.008888299576938152, + 0.010676831007003784, + 0.016586795449256897, + 0.0017295610159635544, + 0.0011315642623230815, + 0.003229374997317791, + 0.004810358863323927, + 0.0014655801933258772, + 0.004368419758975506, + 0.0041975947096943855, + 0.002131090499460697, + 0.0015211021527647972, + 0.0015909724170342088, + 0.0023295630235224962, + 0.0028768558986485004, + 0.0015086771454662085, + 0.0020716534927487373, + 0.0030793254263699055, + 0.0034418990835547447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01606367528438568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle drone shot captures a tightly knit group of professional cyclists accelerating down a winding country road, the camera gliding smoothly to keep pace with the lead rider and emphasize their speed. The cyclists, wearing skin-tight jerseys in a range of vibrant, saturated colors like electric blue, neon yellow, and fiery red, hunch low in an aggressive, aerodynamic stance, perched atop the drop bars of their sleek carbon fiber racing bikes. Their legs beat in a powerful, synchronized rhythm, and blurs of motion propel them forward as they lean into a sweeping bend in perfect unison. The road itself is a smooth, dark ribbon of asphalt, carving sinuous curves through a breathtaking landscape of winding, lush green hills beneath a clear blue sky. The entire scene is bathed in the warm golden rays of the afternoon sun, which gleams off the polished bicycle frames and casts long, dramatic shadows that race alongside the bikes, creating an exhilarating atmosphere of speed, focus, and competition as the pack moves like a single, fluid organism against a serene, verdant backdrop.", + "seed": 3162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_10/index_6743_seed_3162_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05461083697155118, + "vlm_logprob": -0.05461083697155118, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.05461083697155118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15677087008953094 + ], + "opposite_logprob_tokens": [ + -2.531770944595337 + ], + "logprob_reward": -0.15677087008953094, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.15677087008953094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0861072987318039, + 0.09780270606279373, + 0.07227236777544022, + 0.1061062291264534, + 0.02469758689403534, + 0.13866150379180908, + 0.06801015138626099, + 0.06817024201154709, + 0.2812274992465973, + 0.05694443732500076 + ], + "credit_map": [ + 0.008744039572775364, + 0.00875809695571661, + 0.01198757067322731, + 0.01424313336610794, + 0.006106168497353792, + 0.010987726971507072, + 0.015435827895998955, + 0.009844731539487839, + 0.01046931091696024, + 0.011724072508513927, + 0.009577578864991665, + 0.013360368087887764, + 0.007187228184193373, + 0.019864322617650032, + 0.01285106036812067, + 0.012768754735589027, + 0.006922094617038965, + 0.010458570905029774, + 0.00893157348036766, + 0.009896930307149887, + 0.005289705004543066, + 0.012678679078817368, + 0.01131568942219019, + 0.006779124960303307, + 0.017613165080547333, + 0.016276299953460693, + 0.009376207366585732, + 0.007866920903325081, + 0.013416892848908901, + 0.017906326800584793, + 0.009551683440804482, + 0.014098728075623512, + 0.0019892051350325346, + 0.001888664555735886, + 0.0027334385085850954, + 0.003989716060459614, + 0.0023448048159480095, + 0.005585001315921545, + 0.0030064107850193977, + 0.003160342574119568, + 0.01321945246309042, + 0.010101497173309326, + 0.03494171053171158, + 0.021412521600723267, + 0.015693241730332375, + 0.01383122056722641, + 0.012991237454116344, + 0.016470609232783318, + 0.007936742156744003, + 0.010439252480864525, + 0.00682667875662446, + 0.009754017926752567, + 0.010877026244997978, + 0.007376998662948608, + 0.007877537049353123, + 0.00692189484834671, + 0.007174661383032799, + 0.004489197861403227, + 0.011494928039610386, + 0.012562795542180538, + 0.010213949717581272, + 0.005919288843870163, + 0.010066094808280468, + 0.006249323487281799, + 0.024887433275580406, + 0.024767857044935226, + 0.04288369417190552, + 0.058753564953804016, + 0.04394477605819702, + 0.03400952368974686, + 0.029777107760310173, + 0.022203505039215088, + 0.004489867482334375, + 0.005373187363147736, + 0.006445645820349455, + 0.009627323597669601, + 0.008978629484772682, + 0.0110849654302001, + 0.005991951562464237, + 0.004952862858772278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.15677087008953094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008039206266403 + ], + "opposite_logprob_tokens": [ + -2.695080280303955 + ], + "logprob_reward": -0.07008039206266403, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008039206266403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17398355901241302, + 0.04359598830342293, + 0.13692013919353485, + 0.24308592081069946, + 0.07242938876152039, + 0.053384676575660706, + 0.04366236552596092, + 0.0072561344131827354, + 0.1198878139257431, + 0.10579396784305573 + ], + "credit_map": [ + 0.016757886856794357, + 0.03455847129225731, + 0.017696965485811234, + 0.04259251430630684, + 0.008800584822893143, + 0.01560684759169817, + 0.020545827224850655, + 0.017424464225769043, + 0.004962000530213118, + 0.003424953669309616, + 0.006510822102427483, + 0.008500558324158192, + 0.0034689793828874826, + 0.005320401396602392, + 0.006050824653357267, + 0.0053574456833302975, + 0.01533190906047821, + 0.014375688508152962, + 0.013783229514956474, + 0.044105544686317444, + 0.006852344144135714, + 0.009147914126515388, + 0.020564289763569832, + 0.012759228236973286, + 0.03058469481766224, + 0.027322610840201378, + 0.029245328158140182, + 0.026390984654426575, + 0.03376620262861252, + 0.040248747915029526, + 0.027080589905381203, + 0.028446752578020096, + 0.006847768556326628, + 0.007233760319650173, + 0.008865469135344028, + 0.017023323103785515, + 0.006990618538111448, + 0.00928389560431242, + 0.009178015403449535, + 0.007006536237895489, + 0.004904676228761673, + 0.004877172876149416, + 0.004446511156857014, + 0.009401507675647736, + 0.012560725212097168, + 0.006603276822715998, + 0.006349773611873388, + 0.004241032991558313, + 0.004351708106696606, + 0.0051903678104281425, + 0.003517480567097664, + 0.0070460494607687, + 0.010940270498394966, + 0.003954618703573942, + 0.004522929433733225, + 0.004138940945267677, + 0.0007472151191905141, + 0.0007166466675698757, + 0.0012465277686715126, + 0.0012601157650351524, + 0.0007051177672110498, + 0.0008887833682820201, + 0.0009547925437800586, + 0.0007369355880655348, + 0.009741400368511677, + 0.009911812841892242, + 0.016268838196992874, + 0.025594575330615044, + 0.0175192728638649, + 0.013743148185312748, + 0.01625441573560238, + 0.010854354128241539, + 0.008833152242004871, + 0.007383314426988363, + 0.006131655536592007, + 0.024591758847236633, + 0.015463631600141525, + 0.01764688827097416, + 0.014933228492736816, + 0.010810342617332935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.07008039206266403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809776693582535 + ], + "opposite_logprob_tokens": [ + -3.5298097133636475 + ], + "logprob_reward": -0.029809776693582535, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809776693582535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11747665703296661, + 0.06940735876560211, + 0.0819680392742157, + 0.09939844161272049, + 0.07199181616306305, + 0.14268822968006134, + 0.06612060964107513, + 0.037579458206892014, + 0.05351070314645767, + 0.2598586976528168 + ], + "credit_map": [ + 0.009872949682176113, + 0.010256966575980186, + 0.015525897964835167, + 0.025443363934755325, + 0.010056997649371624, + 0.015394720248878002, + 0.02041773870587349, + 0.010508028790354729, + 0.005136960186064243, + 0.005166634917259216, + 0.009640407748520374, + 0.01599012315273285, + 0.004837230779230595, + 0.009273996576666832, + 0.007663590833544731, + 0.011698417365550995, + 0.010172228328883648, + 0.010904788039624691, + 0.008372134529054165, + 0.01759328879415989, + 0.00523269223049283, + 0.006361063104122877, + 0.014617962762713432, + 0.008713878691196442, + 0.01374551560729742, + 0.012520416639745235, + 0.011497524566948414, + 0.01140853576362133, + 0.01338966190814972, + 0.014682775363326073, + 0.011028959415853024, + 0.011125054210424423, + 0.00671174842864275, + 0.007020160555839539, + 0.009142148308455944, + 0.016456108540296555, + 0.00799199752509594, + 0.008169732056558132, + 0.009743992239236832, + 0.00675592478364706, + 0.012816102243959904, + 0.011367487721145153, + 0.028378993272781372, + 0.024643218144774437, + 0.02214713580906391, + 0.013389967381954193, + 0.015229742042720318, + 0.014715587720274925, + 0.006872145924717188, + 0.0072899325750768185, + 0.010146592743694782, + 0.012184816412627697, + 0.00816002581268549, + 0.006392759270966053, + 0.007538188248872757, + 0.007536148186773062, + 0.0035372513812035322, + 0.0022361567243933678, + 0.0070187002420425415, + 0.00646985974162817, + 0.004829413257539272, + 0.004326028749346733, + 0.005378251429647207, + 0.003783799707889557, + 0.005183586850762367, + 0.005570264998823404, + 0.00698033208027482, + 0.010236331261694431, + 0.007116386201232672, + 0.007784807123243809, + 0.005616008769720793, + 0.00502298679202795, + 0.026159698143601418, + 0.015898212790489197, + 0.013969075866043568, + 0.07390419393777847, + 0.03129764646291733, + 0.0367279052734375, + 0.03902025148272514, + 0.022881699725985527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.029809776693582535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021992563270032406 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.0021992563270032406, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021992563270032406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1789563000202179, + 0.07815644145011902, + 0.0629638209939003, + 0.0376078225672245, + 0.07906707376241684, + 0.11486662924289703, + 0.07929100841283798, + 0.0961805135011673, + 0.15969492495059967, + 0.11321540921926498 + ], + "credit_map": [ + 0.013223236426711082, + 0.012451838701963425, + 0.031841449439525604, + 0.03562195226550102, + 0.014374423772096634, + 0.03004489466547966, + 0.025389092043042183, + 0.016009412705898285, + 0.007285261992365122, + 0.0064115277491509914, + 0.007504078093916178, + 0.011267882771790028, + 0.00612396327778697, + 0.0154510997235775, + 0.010778907686471939, + 0.013333716429769993, + 0.007032877765595913, + 0.008528007194399834, + 0.006111087277531624, + 0.012961131520569324, + 0.005696258973330259, + 0.008228310383856297, + 0.007500011008232832, + 0.006906131282448769, + 0.006826489232480526, + 0.003613237990066409, + 0.004189610481262207, + 0.004438075702637434, + 0.003081035101786256, + 0.00639398954808712, + 0.0034368429332971573, + 0.005628541111946106, + 0.007111818995326757, + 0.007754988502711058, + 0.00905829668045044, + 0.016494430601596832, + 0.0066272905096411705, + 0.01242840476334095, + 0.010506625287234783, + 0.009085219353437424, + 0.010239394381642342, + 0.008990180678665638, + 0.013307194225490093, + 0.01694645546376705, + 0.020165957510471344, + 0.014813996851444244, + 0.018175574019551277, + 0.012227874249219894, + 0.007128856144845486, + 0.006296910345554352, + 0.013851978816092014, + 0.019108105450868607, + 0.013050106354057789, + 0.0056556034833192825, + 0.00811437051743269, + 0.0060850754380226135, + 0.009143855422735214, + 0.005720074754208326, + 0.024842798709869385, + 0.01692170836031437, + 0.008335663005709648, + 0.010352956131100655, + 0.011019370518624783, + 0.00984408799558878, + 0.01445656269788742, + 0.010427565313875675, + 0.015355854295194149, + 0.0329829677939415, + 0.01948699727654457, + 0.016799533739686012, + 0.032464951276779175, + 0.017720477655529976, + 0.014798911288380623, + 0.006362903863191605, + 0.008331884630024433, + 0.03861768916249275, + 0.00946438405662775, + 0.011350155808031559, + 0.017200740054249763, + 0.007088729180395603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0021992563270032406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are professional cyclists wearing skin-tight jerseys in vibrant colors like electric blue, neon yellow, or fiery red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cyclists hunched in an aggressive, aerodynamic stance on drop bars of carbon fiber bikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a winding country road with lush green hills under a clear blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide smoothly to keep pace with the lead rider while the cyclists accelerate into a sweeping bend?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11572802066802979, + 0.03933913633227348, + 0.0944470763206482, + 0.13673971593379974, + 0.1772070825099945, + 0.08750009536743164, + 0.07006402313709259, + 0.08372429758310318, + 0.14654183387756348, + 0.04870866984128952 + ], + "credit_map": [ + 0.007985301315784454, + 0.009371002204716206, + 0.013857289217412472, + 0.014937663450837135, + 0.01531000342220068, + 0.02761939913034439, + 0.01397746242582798, + 0.012669899500906467, + 0.0032511535100638866, + 0.0032042397651821375, + 0.0039026669692248106, + 0.008281141519546509, + 0.002719278447329998, + 0.006116154603660107, + 0.0065366849303245544, + 0.005327816121280193, + 0.009446714073419571, + 0.0167996808886528, + 0.006818119902163744, + 0.016892680898308754, + 0.004865705966949463, + 0.02093680575489998, + 0.00965732429176569, + 0.009030046872794628, + 0.01975121535360813, + 0.01868288218975067, + 0.015397783368825912, + 0.01003205869346857, + 0.017441650852560997, + 0.02568042278289795, + 0.012222767807543278, + 0.01753092184662819, + 0.01556151919066906, + 0.018130168318748474, + 0.019585898146033287, + 0.04032910242676735, + 0.014557401649653912, + 0.024828843772411346, + 0.023918984457850456, + 0.020295171067118645, + 0.007335139438509941, + 0.006074621807783842, + 0.014030746184289455, + 0.012397302314639091, + 0.016873907297849655, + 0.010353827849030495, + 0.011857777833938599, + 0.008576770313084126, + 0.005687089171260595, + 0.007072697393596172, + 0.012235765345394611, + 0.008900459855794907, + 0.015217523090541363, + 0.007089660502970219, + 0.00592000549659133, + 0.007940824143588543, + 0.007549473084509373, + 0.006071148440241814, + 0.018379030749201775, + 0.01437511295080185, + 0.009230993688106537, + 0.008677123114466667, + 0.0109696164727211, + 0.008471797220408916, + 0.015302954241633415, + 0.011249719187617302, + 0.014288369566202164, + 0.029769642278552055, + 0.018828634172677994, + 0.015182011760771275, + 0.027854397892951965, + 0.014066101051867008, + 0.0036541165318340063, + 0.005363818723708391, + 0.006085160654038191, + 0.006397920195013285, + 0.008853305131196976, + 0.008625552989542484, + 0.00484859524294734, + 0.004880200605839491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.014193889684975147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_tensor([3035, 3035])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_tensor([3035, 3035])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..603e57193dba89db4d971fae050263f9f8023ea8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_tensor([3035, 3035])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "locked, wide-angle lens, held steady at eye level, captures a dynamic, unfolding Chinese ink painting, imbued with the artistic conception of traditional aesthetics. The background is a minimalist, blank composition. Within the vast expanse of rice paper, a few strokes of varying shades of ink outline the distant mountain range, its peaks looming in the misty void. At the center of this ethereal landscape, an anthropomorphic figure with a plump pig face, large ears, and a rotund belly, clad in a tight, blue short coat, grins with a gleaming gaze, vigorously wielding a heavy, nine-toothed rake. He is not fighting, but rather, swaying clumsily in intense excitement and greed, his body tumbling and swaying, the heavy rake slicing irregular arcs in his hands, further unbalancing his already massive frame. His whirlwind is directed with unmistakable purpose: a dreamlike cloud swirling to one side of the painting. This is no ordinary cloud, but a vortex formed by countless large, juicy, pink peaches, steaming white flour buns, and glossy roasted poultry. The vibrant colors of the food contrast sharply with the elegant ink painting of the background. A soft, diffuse light evenly casts the scene, devoid of clear shadows, highlighting its scroll-like quality. The scene, strikingly contrasting the solemnity of traditional aesthetics with the comical greed of the characters, creates a sense of mythical wonder and humor.", + "seed": 6461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_3035_seed_6461_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020715649519115687, + "vlm_logprob": -0.020715649519115687, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.020715649519115687, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04386727884411812 + ], + "opposite_logprob_tokens": [ + -3.168867349624634 + ], + "logprob_reward": -0.04386727884411812, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04386727884411812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2779308557510376, + 0.24597179889678955, + 0.09425053745508194, + 0.12383458018302917, + 0.0633474737405777, + 0.046721749007701874, + 0.04686811938881874, + 0.02961341291666031, + 0.03813633322715759, + 0.03332516551017761 + ], + "credit_map": [ + 0.03897817060351372, + 0.048441022634506226, + 0.0636572390794754, + 0.04557886719703674, + 0.018703125417232513, + 0.02101401798427105, + 0.02213108353316784, + 0.01942729949951172, + 0.03223375231027603, + 0.04236974939703941, + 0.06324437260627747, + 0.0386132188141346, + 0.016938719898462296, + 0.017008667811751366, + 0.021742280572652817, + 0.013821011409163475, + 0.01320852991193533, + 0.013702291063964367, + 0.01559417974203825, + 0.017421798780560493, + 0.006728398613631725, + 0.008872440084815025, + 0.010920224711298943, + 0.007802668958902359, + 0.015773508697748184, + 0.02328178472816944, + 0.029452016577124596, + 0.02197100594639778, + 0.009071631357073784, + 0.00897812657058239, + 0.009914752095937729, + 0.005391744431108236, + 0.009181327186524868, + 0.01017768494784832, + 0.012462276965379715, + 0.009742660447955132, + 0.00597580149769783, + 0.00482455687597394, + 0.006349571049213409, + 0.0046335868537425995, + 0.006473369896411896, + 0.008043588139116764, + 0.01018782239407301, + 0.007833835668861866, + 0.0040086437948048115, + 0.0030345122795552015, + 0.004606789909303188, + 0.0025331832002848387, + 0.007002870552241802, + 0.006900473032146692, + 0.007896642200648785, + 0.007688561454415321, + 0.004504287615418434, + 0.0040252236649394035, + 0.005394825711846352, + 0.00345523189753294, + 0.0044846911914646626, + 0.004831538535654545, + 0.005603599362075329, + 0.004682305734604597, + 0.0024163546040654182, + 0.002222567331045866, + 0.0034061153419315815, + 0.0019662389531731606, + 0.00540567934513092, + 0.0062202634289860725, + 0.009100156836211681, + 0.006888082716614008, + 0.0028492752462625504, + 0.0028556021861732006, + 0.0030979986768215895, + 0.001719271531328559, + 0.00548295583575964, + 0.00487516401335597, + 0.004655723925679922, + 0.005267211701720953, + 0.0028267023153603077, + 0.0031212526373565197, + 0.004393028095364571, + 0.0027031225617974997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04386727884411812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419964898377657 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.01419964898377657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419964898377657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2921465337276459, + 0.144758403301239, + 0.058919720351696014, + 0.06541340053081512, + 0.054219137877225876, + 0.07484563440084457, + 0.08192427456378937, + 0.06813549995422363, + 0.08438501507043839, + 0.07525236904621124 + ], + "credit_map": [ + 0.04035191237926483, + 0.04260865971446037, + 0.049716103821992874, + 0.033930473029613495, + 0.03355301544070244, + 0.03575075790286064, + 0.03111150860786438, + 0.025124121457338333, + 0.01855405606329441, + 0.022256851196289062, + 0.017861828207969666, + 0.02487085573375225, + 0.01850825548171997, + 0.01692923717200756, + 0.013465619646012783, + 0.0123117221519351, + 0.010493947193026543, + 0.009866725653409958, + 0.004758704453706741, + 0.007124844938516617, + 0.005307354498654604, + 0.009056531824171543, + 0.006726227235049009, + 0.005585387349128723, + 0.010083167813718319, + 0.010302671231329441, + 0.0035819020122289658, + 0.005974293686449528, + 0.009159930981695652, + 0.011024538427591324, + 0.008667550049722195, + 0.006619355641305447, + 0.007526908069849014, + 0.007124989293515682, + 0.007911508902907372, + 0.007282205391675234, + 0.0051987734623253345, + 0.006995043717324734, + 0.007393335457891226, + 0.004786379169672728, + 0.010030410252511501, + 0.010245268233120441, + 0.0080637251958251, + 0.009040028788149357, + 0.011402670294046402, + 0.009122949093580246, + 0.011193117126822472, + 0.005747470539063215, + 0.009235077537596226, + 0.012450522743165493, + 0.013101683929562569, + 0.010612034238874912, + 0.008306484669446945, + 0.009500309824943542, + 0.010665529407560825, + 0.008052638731896877, + 0.010537414811551571, + 0.00994943082332611, + 0.008882299065589905, + 0.009088629856705666, + 0.007545003201812506, + 0.007731813937425613, + 0.009118094108998775, + 0.0052828192710876465, + 0.010971682146191597, + 0.012747131288051605, + 0.007433000020682812, + 0.010209387168288231, + 0.007706736214458942, + 0.006645082496106625, + 0.022076411172747612, + 0.006595595274120569, + 0.011939666233956814, + 0.011187935248017311, + 0.01263828482478857, + 0.011645924299955368, + 0.007423695642501116, + 0.007839683443307877, + 0.00706523796543479, + 0.0055119432508945465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01419964898377657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979126386344433 + ], + "opposite_logprob_tokens": [ + -3.5297913551330566 + ], + "logprob_reward": -0.02979126386344433, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979126386344433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37698057293891907, + 0.1879958063364029, + 0.07957715541124344, + 0.07512541860342026, + 0.07793550938367844, + 0.04822753742337227, + 0.03140641748905182, + 0.03415936976671219, + 0.04573148116469383, + 0.04286080226302147 + ], + "credit_map": [ + 0.04752803593873978, + 0.05562789738178253, + 0.05800209566950798, + 0.04645063355565071, + 0.044127535074949265, + 0.05891623720526695, + 0.035254448652267456, + 0.03107365407049656, + 0.0216617863625288, + 0.025152156129479408, + 0.02577853947877884, + 0.021181875839829445, + 0.018139149993658066, + 0.022185390815138817, + 0.039210330694913864, + 0.014686575159430504, + 0.008709694258868694, + 0.012170340865850449, + 0.012871474027633667, + 0.009961545467376709, + 0.009006625972688198, + 0.011380655691027641, + 0.0091816745698452, + 0.006295135710388422, + 0.010046838782727718, + 0.013027850538492203, + 0.013502613641321659, + 0.011418977752327919, + 0.0071479398757219315, + 0.008070006035268307, + 0.006994284223765135, + 0.0049168989062309265, + 0.011044930666685104, + 0.013181503862142563, + 0.016688961535692215, + 0.011798319406807423, + 0.007564167026430368, + 0.007611884735524654, + 0.005786757916212082, + 0.004258979111909866, + 0.005763625260442495, + 0.007480536121875048, + 0.009622961282730103, + 0.006891289725899696, + 0.004114835057407618, + 0.004834755323827267, + 0.005921582225710154, + 0.003597947536036372, + 0.00338693056255579, + 0.004326601047068834, + 0.0032183427829295397, + 0.0038243760354816914, + 0.004041679203510284, + 0.004493867978453636, + 0.004960913676768541, + 0.003153701312839985, + 0.004152059089392424, + 0.005701898131519556, + 0.007975470274686813, + 0.005345250014215708, + 0.0029878835193812847, + 0.00226313597522676, + 0.0038735170383006334, + 0.0018601528136059642, + 0.0066094473004341125, + 0.008817773312330246, + 0.00986145157366991, + 0.008005965501070023, + 0.0029546304140239954, + 0.002916698344051838, + 0.004004241898655891, + 0.0025612683966755867, + 0.006452174391597509, + 0.006731616333127022, + 0.00620232755318284, + 0.0054513243958354, + 0.005639337934553623, + 0.005171632394194603, + 0.004007122479379177, + 0.003205261193215847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02979126386344433 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775033220648766 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775033220648766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775033220648766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1406414955854416, + 0.3377467393875122, + 0.048778727650642395, + 0.09498979151248932, + 0.09088228642940521, + 0.06056835129857063, + 0.08671114593744278, + 0.028980035334825516, + 0.06692861765623093, + 0.04377281665802002 + ], + "credit_map": [ + 0.02770162746310234, + 0.024196244776248932, + 0.02014990709722042, + 0.017420807853341103, + 0.012463103048503399, + 0.013326085172593594, + 0.014653787016868591, + 0.01072993129491806, + 0.04664568230509758, + 0.05218122527003288, + 0.06857478618621826, + 0.05771069973707199, + 0.026066185906529427, + 0.031066618859767914, + 0.03601455315947533, + 0.019486982375383377, + 0.005723831243813038, + 0.006675626616925001, + 0.008067961782217026, + 0.006502730771899223, + 0.005220711696892977, + 0.007139715366065502, + 0.0053123608231544495, + 0.0041357893496751785, + 0.010001333430409431, + 0.015243152156472206, + 0.01835481822490692, + 0.017207114025950432, + 0.010556788183748722, + 0.007967513985931873, + 0.008792081847786903, + 0.006866987328976393, + 0.011109252460300922, + 0.012952197343111038, + 0.011659268289804459, + 0.01237572357058525, + 0.011689330451190472, + 0.012651835568249226, + 0.010449853725731373, + 0.007994827814400196, + 0.007676765322685242, + 0.008380039595067501, + 0.006374485790729523, + 0.007193261757493019, + 0.009650932624936104, + 0.0081408666446805, + 0.008245101198554039, + 0.004906900692731142, + 0.0113297700881958, + 0.012632336467504501, + 0.008336525410413742, + 0.012038249522447586, + 0.008056635968387127, + 0.00994510855525732, + 0.016064496710896492, + 0.008308026939630508, + 0.0044931815937161446, + 0.004305446054786444, + 0.0042719049379229546, + 0.0037859028670936823, + 0.0027134160045534372, + 0.0025916260201483965, + 0.004296478349715471, + 0.0025220774114131927, + 0.009130328893661499, + 0.01060583908110857, + 0.007145436946302652, + 0.007952260784804821, + 0.005801611114293337, + 0.004885069560259581, + 0.01650768145918846, + 0.004900392610579729, + 0.005981370806694031, + 0.006346306297928095, + 0.00936820451170206, + 0.0075019169598817825, + 0.003969219978898764, + 0.003980155102908611, + 0.0033319282811135054, + 0.003293718909844756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.009775033220648766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22543343901634216, + 0.14464440941810608, + 0.09216033667325974, + 0.09336718916893005, + 0.07768552005290985, + 0.05410955101251602, + 0.0684075579047203, + 0.03718426823616028, + 0.09202738106250763, + 0.1149803176522255 + ], + "credit_map": [ + 0.03924852982163429, + 0.0378204882144928, + 0.02780005894601345, + 0.03503406047821045, + 0.02126624435186386, + 0.025529755279421806, + 0.019691677764058113, + 0.01904262602329254, + 0.015992222353816032, + 0.023988045752048492, + 0.03051765076816082, + 0.02432629093527794, + 0.010752796195447445, + 0.008500549010932446, + 0.021945666521787643, + 0.008621185086667538, + 0.012071313336491585, + 0.013029935769736767, + 0.013363318517804146, + 0.016879666596651077, + 0.007732020225375891, + 0.009098672308027744, + 0.012242740020155907, + 0.007742674555629492, + 0.013634607195854187, + 0.016647333279252052, + 0.01869243197143078, + 0.0160337183624506, + 0.007049732841551304, + 0.006380402483046055, + 0.010782814584672451, + 0.004146147519350052, + 0.012394601479172707, + 0.01145495381206274, + 0.01284752693027258, + 0.011210243217647076, + 0.008189447224140167, + 0.008244781754910946, + 0.006976771168410778, + 0.0063671995885670185, + 0.007675639819353819, + 0.007623599376529455, + 0.006781346630305052, + 0.007700465619564056, + 0.006351257208734751, + 0.004998903721570969, + 0.008108507841825485, + 0.004869828000664711, + 0.01025544572621584, + 0.011571419425308704, + 0.009176620282232761, + 0.009549872018396854, + 0.00668915081769228, + 0.007770761847496033, + 0.0073720975778996944, + 0.0060221953317523, + 0.006591422017663717, + 0.005710843484848738, + 0.006247463170439005, + 0.005840206053107977, + 0.0026070435997098684, + 0.004000605549663305, + 0.003381063463166356, + 0.0028056236915290356, + 0.0110849067568779, + 0.013129777275025845, + 0.01320677436888218, + 0.010220705531537533, + 0.008926116861402988, + 0.007845286279916763, + 0.021047541871666908, + 0.006566278170794249, + 0.016871698200702667, + 0.01951833814382553, + 0.027012860402464867, + 0.019761353731155396, + 0.008605959825217724, + 0.00927484780550003, + 0.007123957388103008, + 0.00681129889562726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0059450226835906506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, wide-angle lens, held steady at eye level, captures a dynamic, unfolding Chinese ink painting, imbued with the artistic conception of traditional aesthetics. The background is a minimalist, blank composition. Within the vast expanse of rice paper, a few strokes of varying shades of ink outline the distant mountain range, its peaks looming in the misty void. At the center of this ethereal landscape, an anthropomorphic figure with a plump pig face, large ears, and a rotund belly, clad in a tight, blue short coat, grins with a gleaming gaze, vigorously wielding a heavy, nine-toothed rake. He is not fighting, but rather, swaying clumsily in intense excitement and greed, his body tumbling and swaying, the heavy rake slicing irregular arcs in his hands, further unbalancing his already massive frame. His whirlwind is directed with unmistakable purpose: a dreamlike cloud swirling to one side of the painting. This is no ordinary cloud, but a vortex formed by countless large, juicy, pink peaches, steaming white flour buns, and glossy roasted poultry. The vibrant colors of the food contrast sharply with the elegant ink painting of the background. A soft, diffuse light evenly casts the scene, devoid of clear shadows, highlighting its scroll-like quality. The scene, strikingly contrasting the solemnity of traditional aesthetics with the comical greed of the characters, creates a sense of mythical wonder and humor.", + "seed": 6462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_3035_seed_6462_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008122398424893618, + "vlm_logprob": -0.008122398424893618, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.008122398424893618, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010579572059214115 + ], + "opposite_logprob_tokens": [ + -4.760579586029053 + ], + "logprob_reward": -0.010579572059214115, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010579572059214115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1886104792356491, + 0.06651881337165833, + 0.055349547415971756, + 0.03552509844303131, + 0.058199692517519, + 0.0396135039627552, + 0.06859250366687775, + 0.36923927068710327, + 0.06944748014211655, + 0.04890362173318863 + ], + "credit_map": [ + 0.022104056552052498, + 0.04292232170701027, + 0.01834353618323803, + 0.014765255153179169, + 0.02827918343245983, + 0.025988556444644928, + 0.019389525055885315, + 0.01681804470717907, + 0.0069817532785236835, + 0.013336917385458946, + 0.0076956055127084255, + 0.0040321010164916515, + 0.009874556213617325, + 0.010237685404717922, + 0.007103923708200455, + 0.007256268057972193, + 0.005067111924290657, + 0.009683230891823769, + 0.004392083268612623, + 0.004348237998783588, + 0.009692043997347355, + 0.008982249535620213, + 0.007457489613443613, + 0.005727095063775778, + 0.0067388447932899, + 0.004468934144824743, + 0.005765816196799278, + 0.004408243577927351, + 0.00479567376896739, + 0.003992571961134672, + 0.0031113617587834597, + 0.0022436485160142183, + 0.00560041144490242, + 0.007366118486970663, + 0.006806070450693369, + 0.0110536590218544, + 0.008447673171758652, + 0.007619809824973345, + 0.006110969930887222, + 0.005194975528866053, + 0.004303095396608114, + 0.004627831280231476, + 0.005290779285132885, + 0.006943129003047943, + 0.004620428197085857, + 0.0063892025500535965, + 0.003927143290638924, + 0.003511890536174178, + 0.006556802894920111, + 0.010144525207579136, + 0.0058211516588926315, + 0.010280479677021503, + 0.011696066707372665, + 0.011316196992993355, + 0.0071909017860889435, + 0.005586377345025539, + 0.025867179036140442, + 0.03866729140281677, + 0.09757877886295319, + 0.04491230472922325, + 0.10690880566835403, + 0.032436490058898926, + 0.01313230860978365, + 0.009736103937029839, + 0.005689580459147692, + 0.01082396786659956, + 0.007539777085185051, + 0.009456411935389042, + 0.010373476892709732, + 0.010826694779098034, + 0.009048805572092533, + 0.005688761360943317, + 0.006179703399538994, + 0.00609095161780715, + 0.005847195629030466, + 0.0071767535991966724, + 0.006647794507443905, + 0.007764582056552172, + 0.004510757979005575, + 0.004685880616307259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.010579572059214115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00526150269433856 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.00526150269433856, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00526150269433856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4659811556339264, + 0.0801035612821579, + 0.04368879646062851, + 0.041535794734954834, + 0.025319136679172516, + 0.03229355812072754, + 0.040636368095874786, + 0.10587902367115021, + 0.055196791887283325, + 0.10936572402715683 + ], + "credit_map": [ + 0.0555814653635025, + 0.12808309495449066, + 0.03034348413348198, + 0.03394240513443947, + 0.06552126258611679, + 0.05877165496349335, + 0.054746951907873154, + 0.03899085894227028, + 0.007574979681521654, + 0.01220995094627142, + 0.007248135283589363, + 0.00570311676710844, + 0.01393165998160839, + 0.012235740199685097, + 0.011151040904223919, + 0.010048940777778625, + 0.0033527177292853594, + 0.004996069706976414, + 0.004191154148429632, + 0.004904179368168116, + 0.006625048816204071, + 0.006070503033697605, + 0.00749217439442873, + 0.006056949496269226, + 0.003400742541998625, + 0.004563441965728998, + 0.005446814000606537, + 0.005868372041732073, + 0.006996734067797661, + 0.0054333144798874855, + 0.004853955470025539, + 0.004972420632839203, + 0.0024183376226574183, + 0.0025882485788315535, + 0.002619028091430664, + 0.003428706433624029, + 0.004498060792684555, + 0.0025233046617358923, + 0.005280132871121168, + 0.0019633187912404537, + 0.0029337983578443527, + 0.0038169817999005318, + 0.0024505306500941515, + 0.003448599949479103, + 0.00495436554774642, + 0.005932902917265892, + 0.004455965477973223, + 0.004300410859286785, + 0.004765548277646303, + 0.007890577428042889, + 0.003341330448165536, + 0.0072771962732076645, + 0.0056174942292273045, + 0.004346184898167849, + 0.004107306245714426, + 0.0032907328568398952, + 0.006390705704689026, + 0.011571691371500492, + 0.021365944296121597, + 0.014214807190001011, + 0.028477152809500694, + 0.008961591869592667, + 0.008210865780711174, + 0.006686271168291569, + 0.006023965310305357, + 0.005657161120325327, + 0.009168211370706558, + 0.006789454724639654, + 0.006446406245231628, + 0.006419123150408268, + 0.008153505623340607, + 0.0065389592200517654, + 0.013388806954026222, + 0.013842878863215446, + 0.012262428179383278, + 0.012660479173064232, + 0.019170571118593216, + 0.017185313627123833, + 0.011472702026367188, + 0.009382546879351139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.00526150269433856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191539026796818 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191539026796818, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191539026796818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.177425354719162, + 0.12440977990627289, + 0.05236034840345383, + 0.057287584990262985, + 0.039710305631160736, + 0.03472241759300232, + 0.053453877568244934, + 0.24497105181217194, + 0.12931504845619202, + 0.08634426444768906 + ], + "credit_map": [ + 0.018466191366314888, + 0.043982915580272675, + 0.009318584576249123, + 0.015408921055495739, + 0.024004824459552765, + 0.01892121322453022, + 0.023860422894358635, + 0.023462293669581413, + 0.009615778923034668, + 0.0164127666503191, + 0.008907048031687737, + 0.010334480553865433, + 0.023195886984467506, + 0.02202579751610756, + 0.021044541150331497, + 0.012873481959104538, + 0.004239094443619251, + 0.00785595178604126, + 0.0042025018483400345, + 0.0052811638452112675, + 0.0083321463316679, + 0.008714674971997738, + 0.007359916344285011, + 0.0063749016262590885, + 0.005980118177831173, + 0.0056732636876404285, + 0.007264993153512478, + 0.007151840254664421, + 0.01043419074267149, + 0.007581866811960936, + 0.006823310628533363, + 0.006378002464771271, + 0.002255750121548772, + 0.004858216270804405, + 0.0029054998885840178, + 0.004851219244301319, + 0.007842952385544777, + 0.004742053337395191, + 0.0074357991106808186, + 0.004818817600607872, + 0.0029601941350847483, + 0.003865583101287484, + 0.002127270679920912, + 0.0036174901761114597, + 0.005636947695165873, + 0.007377078756690025, + 0.0052766576409339905, + 0.003861198900267482, + 0.0038498970679938793, + 0.007110094651579857, + 0.003627186641097069, + 0.005366635508835316, + 0.010812380351126194, + 0.007796602323651314, + 0.009624403901398182, + 0.005266678053885698, + 0.017385216429829597, + 0.025412144139409065, + 0.06398485600948334, + 0.028378432616591454, + 0.06861544400453568, + 0.022875843569636345, + 0.011181117966771126, + 0.007137991022318602, + 0.012998907826840878, + 0.021527227014303207, + 0.013719139620661736, + 0.01139198150485754, + 0.017568040639162064, + 0.017196761444211006, + 0.021715931594371796, + 0.013197056949138641, + 0.008236382156610489, + 0.012374112382531166, + 0.008047224022448063, + 0.015199907124042511, + 0.014196332544088364, + 0.011466571129858494, + 0.010521418415009975, + 0.006302319932729006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.014191539026796818 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943956319242716 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943956319242716, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943956319242716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16057491302490234, + 0.04759884253144264, + 0.05938757583498955, + 0.04088277742266655, + 0.03932690992951393, + 0.03189826011657715, + 0.08575068414211273, + 0.2963865399360657, + 0.13627922534942627, + 0.10191424936056137 + ], + "credit_map": [ + 0.021348517388105392, + 0.049679938703775406, + 0.010228698141872883, + 0.013094472698867321, + 0.019296618178486824, + 0.019276555627584457, + 0.014548199251294136, + 0.013101921416819096, + 0.003894491819664836, + 0.007641973905265331, + 0.003587250830605626, + 0.005786543246358633, + 0.008171127177774906, + 0.006733136251568794, + 0.006099244579672813, + 0.005685076583176851, + 0.006476488895714283, + 0.012259600684046745, + 0.004280891735106707, + 0.004285030998289585, + 0.011508896015584469, + 0.0072334627620875835, + 0.008340807631611824, + 0.005002401303499937, + 0.0036965543404221535, + 0.0034559916239231825, + 0.0038032031152397394, + 0.005699926055967808, + 0.007419108878821135, + 0.0067858570255339146, + 0.0050754244439303875, + 0.004946711473166943, + 0.002633090363815427, + 0.004241004586219788, + 0.003366895020008087, + 0.004812884610146284, + 0.006886985152959824, + 0.005883478559553623, + 0.006002354435622692, + 0.005500217899680138, + 0.002770288148894906, + 0.00334380567073822, + 0.0019449165556579828, + 0.0034485585056245327, + 0.004929012153297663, + 0.006445687729865313, + 0.006052807439118624, + 0.002963184379041195, + 0.008336463011801243, + 0.014605329371988773, + 0.004665179643779993, + 0.009582838974893093, + 0.014596909284591675, + 0.012127408757805824, + 0.015919867902994156, + 0.005916686728596687, + 0.020416099578142166, + 0.038116175681352615, + 0.02698335237801075, + 0.01953469216823578, + 0.10797170549631119, + 0.031072400510311127, + 0.014122509397566319, + 0.038169607520103455, + 0.015218858607113361, + 0.02521602436900139, + 0.012517346069216728, + 0.016245631501078606, + 0.020322559401392937, + 0.018125997856259346, + 0.018871476873755455, + 0.009761334396898746, + 0.00995610374957323, + 0.015831127762794495, + 0.0078662084415555, + 0.01675453782081604, + 0.01822850853204727, + 0.015013436786830425, + 0.012419709004461765, + 0.005844618193805218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.005943956319242716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635422024875879 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635422024875879, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635422024875879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21941876411437988, + 0.05176151171326637, + 0.041412945836782455, + 0.03669305518269539, + 0.03119928389787674, + 0.035688892006874084, + 0.039285991340875626, + 0.3282151520252228, + 0.13553299009799957, + 0.0807914063334465 + ], + "credit_map": [ + 0.03143337741494179, + 0.061476968228816986, + 0.01577102206647396, + 0.020876066759228706, + 0.027324073016643524, + 0.025843730196356773, + 0.022553568705916405, + 0.014139961451292038, + 0.005173143930733204, + 0.012235252186655998, + 0.004085286054760218, + 0.004500740207731724, + 0.007100821938365698, + 0.007855616509914398, + 0.006313257850706577, + 0.004497394897043705, + 0.003720010630786419, + 0.00769167672842741, + 0.0035020376089960337, + 0.0033866879530251026, + 0.006924486719071865, + 0.006221673917025328, + 0.004996234085410833, + 0.004970137029886246, + 0.007014271803200245, + 0.004945896100252867, + 0.006199698895215988, + 0.0038819024339318275, + 0.00427365442737937, + 0.004473169334232807, + 0.0036033496726304293, + 0.0023011101875454187, + 0.0035495171323418617, + 0.0037325562443584204, + 0.0034025157801806927, + 0.004575722385197878, + 0.004987457301467657, + 0.004000921733677387, + 0.003803079016506672, + 0.003147514071315527, + 0.003337375819683075, + 0.0038806020747870207, + 0.003393882419914007, + 0.006954054813832045, + 0.004591446369886398, + 0.007124730385839939, + 0.003776615485548973, + 0.0026301839388906956, + 0.003395587671548128, + 0.0043253302574157715, + 0.002847538562491536, + 0.004841173067688942, + 0.0069429911673069, + 0.00851492676883936, + 0.004844530485570431, + 0.003573913825675845, + 0.024362973868846893, + 0.026032311841845512, + 0.105167917907238, + 0.04005233943462372, + 0.06496606022119522, + 0.0248948372900486, + 0.010141298174858093, + 0.03259742259979248, + 0.012055640108883381, + 0.020980894565582275, + 0.015778789296746254, + 0.013542464934289455, + 0.018428344279527664, + 0.019311608746647835, + 0.022764423862099648, + 0.012670831754803658, + 0.0071999384090304375, + 0.011361373588442802, + 0.006053835619240999, + 0.009510451927781105, + 0.013667702674865723, + 0.015759946778416634, + 0.009408771991729736, + 0.007829386740922928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.004635422024875879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, wide-angle lens, held steady at eye level, captures a dynamic, unfolding Chinese ink painting, imbued with the artistic conception of traditional aesthetics. The background is a minimalist, blank composition. Within the vast expanse of rice paper, a few strokes of varying shades of ink outline the distant mountain range, its peaks looming in the misty void. At the center of this ethereal landscape, an anthropomorphic figure with a plump pig face, large ears, and a rotund belly, clad in a tight, blue short coat, grins with a gleaming gaze, vigorously wielding a heavy, nine-toothed rake. He is not fighting, but rather, swaying clumsily in intense excitement and greed, his body tumbling and swaying, the heavy rake slicing irregular arcs in his hands, further unbalancing his already massive frame. His whirlwind is directed with unmistakable purpose: a dreamlike cloud swirling to one side of the painting. This is no ordinary cloud, but a vortex formed by countless large, juicy, pink peaches, steaming white flour buns, and glossy roasted poultry. The vibrant colors of the food contrast sharply with the elegant ink painting of the background. A soft, diffuse light evenly casts the scene, devoid of clear shadows, highlighting its scroll-like quality. The scene, strikingly contrasting the solemnity of traditional aesthetics with the comical greed of the characters, creates a sense of mythical wonder and humor.", + "seed": 6463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_3035_seed_6463_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010307559836655855, + "vlm_logprob": -0.010307559836655855, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.010307559836655855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01297483965754509 + ], + "opposite_logprob_tokens": [ + -4.637974739074707 + ], + "logprob_reward": -0.01297483965754509, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01297483965754509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26571905612945557, + 0.07730352133512497, + 0.08549921959638596, + 0.09452778846025467, + 0.055228009819984436, + 0.04802316427230835, + 0.10555961728096008, + 0.11472266912460327, + 0.08846157044172287, + 0.06495530903339386 + ], + "credit_map": [ + 0.032225314527750015, + 0.03655262291431427, + 0.024380352348089218, + 0.04350428655743599, + 0.026385564357042313, + 0.03876596316695213, + 0.03741693124175072, + 0.026488039642572403, + 0.008396969176828861, + 0.007870873436331749, + 0.01141771674156189, + 0.01114363968372345, + 0.010096407495439053, + 0.009298828430473804, + 0.009119726717472076, + 0.009959361515939236, + 0.009663727134466171, + 0.010305861011147499, + 0.007199629209935665, + 0.015832511708140373, + 0.01144789345562458, + 0.011938782408833504, + 0.010523034259676933, + 0.008587785065174103, + 0.015619689598679543, + 0.01129247061908245, + 0.008096056059002876, + 0.00626840116456151, + 0.011744527146220207, + 0.020806917920708656, + 0.006585126742720604, + 0.014114614576101303, + 0.006571452133357525, + 0.007978579960763454, + 0.006329268682748079, + 0.006408040877431631, + 0.011087900958955288, + 0.00850620400160551, + 0.004275158978998661, + 0.004071414936333895, + 0.006615219172090292, + 0.008099079132080078, + 0.00621764175593853, + 0.00641405675560236, + 0.005496167577803135, + 0.006106378976255655, + 0.004161904565989971, + 0.004912720527499914, + 0.030423881486058235, + 0.041939627379179, + 0.0044035217724740505, + 0.0077603706158697605, + 0.007011627312749624, + 0.00562819791957736, + 0.003643547184765339, + 0.004748859442770481, + 0.03151673823595047, + 0.04152381420135498, + 0.007743362337350845, + 0.010195759125053883, + 0.00741440337151289, + 0.00616637896746397, + 0.004324025474488735, + 0.0058381990529596806, + 0.01045497041195631, + 0.008642660453915596, + 0.008445881307125092, + 0.012331412173807621, + 0.01783132180571556, + 0.011518231593072414, + 0.009155131876468658, + 0.010081974789500237, + 0.009410298429429531, + 0.008209445513784885, + 0.007144611794501543, + 0.008350014686584473, + 0.009543742053210735, + 0.009083215147256851, + 0.0049448879435658455, + 0.00826909951865673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01297483965754509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979085173457861 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.00979085173457861, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979085173457861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27535441517829895, + 0.05129975453019142, + 0.03764016553759575, + 0.03340884670615196, + 0.02974085509777069, + 0.062447354197502136, + 0.04966173321008682, + 0.2777200937271118, + 0.09040272235870361, + 0.09232410788536072 + ], + "credit_map": [ + 0.027234964072704315, + 0.028190424665808678, + 0.026183975860476494, + 0.054826769977808, + 0.035209935158491135, + 0.034814994782209396, + 0.04563846439123154, + 0.02325487695634365, + 0.005081517621874809, + 0.005853353533893824, + 0.007319319993257523, + 0.007978693582117558, + 0.006294538266956806, + 0.007633857429027557, + 0.004906635265797377, + 0.006231837905943394, + 0.0038760530296713114, + 0.0038279229775071144, + 0.0037644614931195974, + 0.005749387200921774, + 0.004722468554973602, + 0.00762646459043026, + 0.0037818225100636482, + 0.004291587509214878, + 0.004918862134218216, + 0.006786544341593981, + 0.00407022051513195, + 0.004663629923015833, + 0.00314875110052526, + 0.003018157323822379, + 0.002796240383759141, + 0.004006441216915846, + 0.004228374920785427, + 0.004451263230293989, + 0.003098957473412156, + 0.0038043332751840353, + 0.005226711742579937, + 0.0034222155809402466, + 0.0023721489124000072, + 0.003136850195005536, + 0.006220145616680384, + 0.0072732409462332726, + 0.006043435074388981, + 0.007632959634065628, + 0.013348531909286976, + 0.00982416607439518, + 0.006312691606581211, + 0.005792188923805952, + 0.008087942376732826, + 0.005145931150764227, + 0.005489507224410772, + 0.007377778645604849, + 0.005373685620725155, + 0.004136902745813131, + 0.004990729503333569, + 0.009059254080057144, + 0.09457501769065857, + 0.10862257331609726, + 0.012706990353763103, + 0.019774677231907845, + 0.014705237932503223, + 0.010819328017532825, + 0.009075099602341652, + 0.0074411737732589245, + 0.01127639226615429, + 0.011285936459898949, + 0.009720450267195702, + 0.017832892015576363, + 0.015433549880981445, + 0.010110431350767612, + 0.007302110083401203, + 0.007440961431711912, + 0.011450479738414288, + 0.011243445798754692, + 0.011541837826371193, + 0.018473302945494652, + 0.012798418290913105, + 0.010939779691398144, + 0.008014174178242683, + 0.007862665690481663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.00979085173457861 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1550283581018448, + 0.1142779067158699, + 0.04217568784952164, + 0.058284297585487366, + 0.07963809370994568, + 0.11662647873163223, + 0.17014816403388977, + 0.09356914460659027, + 0.06422057747840881, + 0.10603132098913193 + ], + "credit_map": [ + 0.02305487170815468, + 0.018685396760702133, + 0.019284287467598915, + 0.017522338777780533, + 0.015711544081568718, + 0.022714825347065926, + 0.013753845356404781, + 0.024301230907440186, + 0.011248085647821426, + 0.011768623255193233, + 0.014949358999729156, + 0.01911322958767414, + 0.01445030141621828, + 0.016133544966578484, + 0.015408605337142944, + 0.011206149123609066, + 0.004154830239713192, + 0.004845719784498215, + 0.004654110409319401, + 0.006637935526669025, + 0.004995743744075298, + 0.008688610047101974, + 0.003926154691725969, + 0.004272581543773413, + 0.010614975355565548, + 0.013292110525071621, + 0.005434860475361347, + 0.006360636558383703, + 0.006511790212243795, + 0.005254148971289396, + 0.004590650554746389, + 0.006225123535841703, + 0.021641556173563004, + 0.027068249881267548, + 0.004575648810714483, + 0.00814797729253769, + 0.007029880303889513, + 0.004119720309972763, + 0.003523019375279546, + 0.003532035043463111, + 0.013589659705758095, + 0.016142480075359344, + 0.01782948710024357, + 0.017845718190073967, + 0.017501212656497955, + 0.01089827436953783, + 0.009621038101613522, + 0.013198602013289928, + 0.05021103844046593, + 0.06696250289678574, + 0.007430692669004202, + 0.011298848316073418, + 0.011787712574005127, + 0.010284925810992718, + 0.006135402712970972, + 0.006037029903382063, + 0.013744254596531391, + 0.008307015523314476, + 0.012923534028232098, + 0.019915960729122162, + 0.009029669687151909, + 0.010140209458768368, + 0.008946027606725693, + 0.010562471114099026, + 0.007725825998932123, + 0.008622619323432446, + 0.009163419716060162, + 0.01476355642080307, + 0.008740326389670372, + 0.005105059128254652, + 0.005061735399067402, + 0.005038032308220863, + 0.012948498129844666, + 0.016070889309048653, + 0.015174257569015026, + 0.01982181705534458, + 0.014897089451551437, + 0.010618060827255249, + 0.00836567860096693, + 0.008135012350976467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.014193184673786163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059434822760522366 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059434822760522366, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059434822760522366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18022805452346802, + 0.05126534029841423, + 0.07250694185495377, + 0.07778416574001312, + 0.04585712030529976, + 0.0404772087931633, + 0.09486892074346542, + 0.07484877854585648, + 0.2755211293697357, + 0.08664233237504959 + ], + "credit_map": [ + 0.018017055466771126, + 0.022045284509658813, + 0.015383363701403141, + 0.03322499990463257, + 0.023220399394631386, + 0.02485724724829197, + 0.02614414691925049, + 0.01733555644750595, + 0.004792451858520508, + 0.005017923656851053, + 0.004222208634018898, + 0.007890738546848297, + 0.007540242746472359, + 0.010694711469113827, + 0.005912451073527336, + 0.005194614175707102, + 0.0075877076014876366, + 0.010801414027810097, + 0.006911345757544041, + 0.011445971205830574, + 0.00982990674674511, + 0.009342702105641365, + 0.009730466641485691, + 0.006857427302747965, + 0.011499387212097645, + 0.014494507573544979, + 0.006129610817879438, + 0.006967813707888126, + 0.011372878216207027, + 0.015222618356347084, + 0.0057236128486692905, + 0.0063737379387021065, + 0.0048917667008936405, + 0.0054900613613426685, + 0.002348052803426981, + 0.004272720776498318, + 0.010070479474961758, + 0.009750979952514172, + 0.003380908165127039, + 0.005652152933180332, + 0.003031180938705802, + 0.004375275690108538, + 0.004819569177925587, + 0.005244989413768053, + 0.00884968787431717, + 0.005722417030483484, + 0.003826057305559516, + 0.004608030430972576, + 0.027895236387848854, + 0.035039205104112625, + 0.004182735923677683, + 0.0057540228590369225, + 0.008242170326411724, + 0.006945531815290451, + 0.0036994183901697397, + 0.003110608085989952, + 0.018424972891807556, + 0.007810048293322325, + 0.006620497442781925, + 0.00837427657097578, + 0.0062822154723107815, + 0.00589662604033947, + 0.006626292131841183, + 0.01481384877115488, + 0.03662452474236488, + 0.03589951619505882, + 0.0164358988404274, + 0.02352752536535263, + 0.05967618152499199, + 0.05461190640926361, + 0.022793427109718323, + 0.025952128693461418, + 0.015439420938491821, + 0.009539982303977013, + 0.006530038546770811, + 0.010845961980521679, + 0.013611150905489922, + 0.012398876249790192, + 0.007244592532515526, + 0.011032313108444214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0059434822760522366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635440841317177 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635440841317177, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635440841317177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21076373755931854, + 0.07377082854509354, + 0.038256313651800156, + 0.048081427812576294, + 0.03940150886774063, + 0.0363503023982048, + 0.26619651913642883, + 0.12739883363246918, + 0.0719657838344574, + 0.08781468123197556 + ], + "credit_map": [ + 0.026625771075487137, + 0.019352231174707413, + 0.022262264043092728, + 0.03578048571944237, + 0.022206244990229607, + 0.024910567328333855, + 0.033933162689208984, + 0.02569301798939705, + 0.009400598704814911, + 0.008036427199840546, + 0.012495885603129864, + 0.009737219661474228, + 0.008392073214054108, + 0.009319758042693138, + 0.008966710418462753, + 0.007422162219882011, + 0.0050566233694553375, + 0.005719928536564112, + 0.005002208054065704, + 0.006944897118955851, + 0.0035731219686567783, + 0.0046067493967711926, + 0.004542847163975239, + 0.0028099380433559418, + 0.009026630781590939, + 0.00972125492990017, + 0.007269402500241995, + 0.005685082171112299, + 0.0054398467764258385, + 0.0034335011150687933, + 0.0038831064011901617, + 0.003622607793658972, + 0.00558058125898242, + 0.007293268572539091, + 0.004188071005046368, + 0.004501704126596451, + 0.005906274076551199, + 0.005504338536411524, + 0.0027918333653360605, + 0.0036354418843984604, + 0.0038845238741487265, + 0.003886453341692686, + 0.005379104521125555, + 0.0057294247671961784, + 0.00603104243054986, + 0.0038434152957051992, + 0.0034672473557293415, + 0.0041290936060249805, + 0.08223796635866165, + 0.11052945256233215, + 0.010493291541934013, + 0.01772954687476158, + 0.01595282554626465, + 0.012028747238218784, + 0.007840764708817005, + 0.009383945725858212, + 0.03106369450688362, + 0.043077051639556885, + 0.010239596478641033, + 0.011462216265499592, + 0.009439411573112011, + 0.007539616897702217, + 0.005853553768247366, + 0.00872371532022953, + 0.009121528826653957, + 0.00942200981080532, + 0.010916098020970821, + 0.009833289310336113, + 0.008749065920710564, + 0.00744262058287859, + 0.006978415418416262, + 0.009502759203314781, + 0.010135248303413391, + 0.009178921580314636, + 0.011395879089832306, + 0.014745975844562054, + 0.015126480720937252, + 0.011788048781454563, + 0.006868263240903616, + 0.008575876243412495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.008635440841317177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, wide-angle lens, held steady at eye level, captures a dynamic, unfolding Chinese ink painting, imbued with the artistic conception of traditional aesthetics. The background is a minimalist, blank composition. Within the vast expanse of rice paper, a few strokes of varying shades of ink outline the distant mountain range, its peaks looming in the misty void. At the center of this ethereal landscape, an anthropomorphic figure with a plump pig face, large ears, and a rotund belly, clad in a tight, blue short coat, grins with a gleaming gaze, vigorously wielding a heavy, nine-toothed rake. He is not fighting, but rather, swaying clumsily in intense excitement and greed, his body tumbling and swaying, the heavy rake slicing irregular arcs in his hands, further unbalancing his already massive frame. His whirlwind is directed with unmistakable purpose: a dreamlike cloud swirling to one side of the painting. This is no ordinary cloud, but a vortex formed by countless large, juicy, pink peaches, steaming white flour buns, and glossy roasted poultry. The vibrant colors of the food contrast sharply with the elegant ink painting of the background. A soft, diffuse light evenly casts the scene, devoid of clear shadows, highlighting its scroll-like quality. The scene, strikingly contrasting the solemnity of traditional aesthetics with the comical greed of the characters, creates a sense of mythical wonder and humor.", + "seed": 6464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_11/index_3035_seed_6464_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010288865026086568, + "vlm_logprob": -0.010288865026086568, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.010288865026086568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006299045868217945 + ], + "opposite_logprob_tokens": [ + -5.506299018859863 + ], + "logprob_reward": -0.006299045868217945, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006299045868217945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0744335949420929, + 0.43680405616760254, + 0.08920868486166, + 0.02007686346769333, + 0.02980239875614643, + 0.024098416790366173, + 0.23763440549373627, + 0.025536004453897476, + 0.02996043860912323, + 0.032445114105939865 + ], + "credit_map": [ + 0.011462463065981865, + 0.010321833193302155, + 0.009982043877243996, + 0.009555855765938759, + 0.009694208391010761, + 0.00650318618863821, + 0.010222849436104298, + 0.006691155023872852, + 0.12253954261541367, + 0.1325768232345581, + 0.05189235880970955, + 0.03833650425076485, + 0.033114735037088394, + 0.02231733314692974, + 0.02051907777786255, + 0.015507696196436882, + 0.011349254287779331, + 0.018459053710103035, + 0.009309543296694756, + 0.021313536912202835, + 0.010434258729219437, + 0.004410381428897381, + 0.008333876729011536, + 0.005598777439445257, + 0.00210527959279716, + 0.0036272050347179174, + 0.002837524050846696, + 0.0030102068558335304, + 0.0025186645798385143, + 0.0014259981689974666, + 0.0026176737155765295, + 0.0019343120511621237, + 0.0035543572157621384, + 0.005759776569902897, + 0.0036342740058898926, + 0.004457891918718815, + 0.006198355462402105, + 0.0016954659949988127, + 0.0028535155579447746, + 0.0016487607499584556, + 0.00275210733525455, + 0.004203521180897951, + 0.003542050952091813, + 0.0041293613612651825, + 0.0032558885868638754, + 0.0018379236571490765, + 0.002920686500146985, + 0.0014568795450031757, + 0.0241736751049757, + 0.04782714322209358, + 0.033247027546167374, + 0.0426812469959259, + 0.031178101897239685, + 0.019406596198678017, + 0.021735602989792824, + 0.01738501712679863, + 0.0031163350213319063, + 0.005588727071881294, + 0.002982618985697627, + 0.004382530227303505, + 0.0028938192408531904, + 0.0015171857085078955, + 0.0024411792401224375, + 0.002613607095554471, + 0.0026720180176198483, + 0.00424520019441843, + 0.0026277066208422184, + 0.0064156549051404, + 0.004701024387031794, + 0.0022572672460228205, + 0.00404345290735364, + 0.002998114563524723, + 0.003803374245762825, + 0.006277402397245169, + 0.0035747590009123087, + 0.004140027798712254, + 0.005145002156496048, + 0.0025796308182179928, + 0.004276474006474018, + 0.002648445311933756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.006299045868217945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547479011118412 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.012547479011118412, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547479011118412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12194705754518509, + 0.23275643587112427, + 0.09320992231369019, + 0.046551499515771866, + 0.057373564690351486, + 0.1977154165506363, + 0.1089491993188858, + 0.04255698248744011, + 0.04395022615790367, + 0.05498969554901123 + ], + "credit_map": [ + 0.022456558421254158, + 0.02489412948489189, + 0.009536540135741234, + 0.014193272218108177, + 0.015409289859235287, + 0.011603633873164654, + 0.013128578662872314, + 0.010725057683885098, + 0.07068389654159546, + 0.07232822477817535, + 0.013704158365726471, + 0.01500164344906807, + 0.02692180871963501, + 0.009257429279386997, + 0.016606405377388, + 0.008252864703536034, + 0.010740616358816624, + 0.01571955718100071, + 0.00965624488890171, + 0.019159015268087387, + 0.01208878867328167, + 0.006560820620507002, + 0.01148894801735878, + 0.007795929443091154, + 0.004980582278221846, + 0.007044687867164612, + 0.00673056161031127, + 0.0066933399066329, + 0.006609505508095026, + 0.003957892302423716, + 0.007333922199904919, + 0.0032010043505579233, + 0.007290162146091461, + 0.010348372161388397, + 0.005470508709549904, + 0.008355457335710526, + 0.013024111278355122, + 0.0031164686661213636, + 0.005604758858680725, + 0.0041637239046394825, + 0.020920727401971817, + 0.039962466806173325, + 0.02873842976987362, + 0.02669154852628708, + 0.03215049207210541, + 0.01704193651676178, + 0.019141649827361107, + 0.013068165630102158, + 0.011187147349119186, + 0.013980710878968239, + 0.015265202149748802, + 0.015828313305974007, + 0.015727022662758827, + 0.012481125071644783, + 0.013326586224138737, + 0.011153084225952625, + 0.004332158714532852, + 0.009297509677708149, + 0.004933783318847418, + 0.009599408134818077, + 0.0052204676903784275, + 0.002458886243402958, + 0.004524274729192257, + 0.0021904949098825455, + 0.005298617295920849, + 0.006473409943282604, + 0.007006020750850439, + 0.00800346303731203, + 0.006679796148091555, + 0.003163917688652873, + 0.0034065062645822763, + 0.003918496426194906, + 0.007313148584216833, + 0.007031702436506748, + 0.010086429305374622, + 0.009120616130530834, + 0.007380851078778505, + 0.003930944949388504, + 0.006524553522467613, + 0.003601450938731432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.012547479011118412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183792009949684 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183792009949684, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183792009949684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15225274860858917, + 0.2486155778169632, + 0.05361989885568619, + 0.08971875160932541, + 0.05590051785111427, + 0.16865487396717072, + 0.05629126355051994, + 0.025883199647068977, + 0.10412494838237762, + 0.04493824392557144 + ], + "credit_map": [ + 0.016831466928124428, + 0.01858285628259182, + 0.013726540841162205, + 0.017555389553308487, + 0.03599198907613754, + 0.011201070621609688, + 0.028888506814837456, + 0.009474937804043293, + 0.06364745646715164, + 0.07196108251810074, + 0.02468014881014824, + 0.017034241929650307, + 0.02901061810553074, + 0.014883928932249546, + 0.017819175496697426, + 0.009578930214047432, + 0.00692983390763402, + 0.009316105395555496, + 0.0077745672315359116, + 0.01111887488514185, + 0.007107789162546396, + 0.0030810763128101826, + 0.004808223340660334, + 0.0034834297839552164, + 0.011543742381036282, + 0.017895421013236046, + 0.011539319530129433, + 0.012393148615956306, + 0.014809384010732174, + 0.006422130856662989, + 0.009815562516450882, + 0.0053000496700406075, + 0.007154165301471949, + 0.011309650726616383, + 0.010349667631089687, + 0.008604612201452255, + 0.006705867126584053, + 0.0030508555937558413, + 0.00542073231190443, + 0.0033049634657800198, + 0.017826048657298088, + 0.03883381187915802, + 0.013700451701879501, + 0.022305304184556007, + 0.027560045942664146, + 0.01739383302628994, + 0.01859019510447979, + 0.012445165775716305, + 0.005763306748121977, + 0.008194074966013432, + 0.0051029641181230545, + 0.011575311422348022, + 0.00978020392358303, + 0.0033951662480831146, + 0.008931759744882584, + 0.0035484752152115107, + 0.0028791753575205803, + 0.006487070582807064, + 0.002964045386761427, + 0.003506240202113986, + 0.003530415240675211, + 0.0016606111312285066, + 0.002950873225927353, + 0.0019047680543735623, + 0.012285560369491577, + 0.015389458276331425, + 0.013249759562313557, + 0.014665941707789898, + 0.02264481410384178, + 0.007889150641858578, + 0.010972084477543831, + 0.007028169464319944, + 0.005749307572841644, + 0.005462200380861759, + 0.008163779973983765, + 0.00604246137663722, + 0.006130609195679426, + 0.0034599320497363806, + 0.006820989307016134, + 0.003108962904661894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.018183792009949684 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16138413548469543, + 0.09500809013843536, + 0.051933273673057556, + 0.04864370450377464, + 0.14234420657157898, + 0.14809376001358032, + 0.13335087895393372, + 0.0610676184296608, + 0.048224758356809616, + 0.10994955152273178 + ], + "credit_map": [ + 0.031082633882761, + 0.043512795120477676, + 0.014264662750065327, + 0.014002247713506222, + 0.018317097797989845, + 0.00906643457710743, + 0.017662784084677696, + 0.013475470244884491, + 0.01611308567225933, + 0.017575610429048538, + 0.017282627522945404, + 0.010158879682421684, + 0.012652977369725704, + 0.005987945478409529, + 0.008833188563585281, + 0.006403778679668903, + 0.006228742189705372, + 0.00854265596717596, + 0.00810740515589714, + 0.010024478659033775, + 0.007284096442162991, + 0.0031111345160752535, + 0.005228917580097914, + 0.003405841300264001, + 0.0063376277685165405, + 0.007683869451284409, + 0.005918204318732023, + 0.007644819561392069, + 0.006819815840572119, + 0.0032649580389261246, + 0.006943443324416876, + 0.004030970856547356, + 0.01635981909930706, + 0.032792553305625916, + 0.01801713928580284, + 0.02744060382246971, + 0.017229678109288216, + 0.009892711415886879, + 0.013212643563747406, + 0.007399052381515503, + 0.016816239804029465, + 0.0331546887755394, + 0.016283676028251648, + 0.01805105060338974, + 0.021360721439123154, + 0.014539610594511032, + 0.014595629647374153, + 0.013292144984006882, + 0.013404606841504574, + 0.022673960775136948, + 0.015591375529766083, + 0.02240978740155697, + 0.020630091428756714, + 0.010461333207786083, + 0.015473154373466969, + 0.012706578709185123, + 0.006700665690004826, + 0.012797720730304718, + 0.005389057565480471, + 0.009225239045917988, + 0.010423882864415646, + 0.004296071827411652, + 0.007900224067270756, + 0.0043347557075321674, + 0.006516417488455772, + 0.005832521710544825, + 0.0081479512155056, + 0.00579875661060214, + 0.009370192885398865, + 0.0033392570912837982, + 0.0048581575974822044, + 0.004361502360552549, + 0.01453381311148405, + 0.014857981353998184, + 0.01466385554522276, + 0.017116177827119827, + 0.016874011605978012, + 0.009717252105474472, + 0.013954018242657185, + 0.00823244359344244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.009776213206350803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of a Chinese ink painting on rice paper with a minimalist, blank background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central character an anthropomorphic figure with a plump pig face, large ears, and a rotund belly wearing a tight blue short coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character vigorously wielding a heavy, nine-toothed rake while swaying and tumbling in excitement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a swirling vortex of large pink peaches, steaming white flour buns, and glossy roasted poultry to the side of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and diffuse with no clear shadows, highlighting the scroll-like quality of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046377950347959995 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.0046377950347959995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046377950347959995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17987601459026337, + 0.18018123507499695, + 0.19809290766716003, + 0.03218669071793556, + 0.06166612729430199, + 0.0691380724310875, + 0.1590149849653244, + 0.029083866626024246, + 0.04565495625138283, + 0.04510516673326492 + ], + "credit_map": [ + 0.044463757425546646, + 0.055737193673849106, + 0.012151580303907394, + 0.010877130553126335, + 0.019526278600096703, + 0.009835970588028431, + 0.015016884543001652, + 0.012267202138900757, + 0.04308735951781273, + 0.04671654850244522, + 0.024120425805449486, + 0.01806814596056938, + 0.01699412614107132, + 0.009651998989284039, + 0.013043244369328022, + 0.008499369956552982, + 0.025737294927239418, + 0.03962085396051407, + 0.022145044058561325, + 0.04687248170375824, + 0.022610191255807877, + 0.010373995639383793, + 0.01853848248720169, + 0.012194554321467876, + 0.0035215746611356735, + 0.0039081089198589325, + 0.0060684094205498695, + 0.004781574010848999, + 0.0037343022413551807, + 0.0031560328789055347, + 0.004179820418357849, + 0.0028368632774800062, + 0.006399128120392561, + 0.010470922105014324, + 0.005090935621410608, + 0.0090515511110425, + 0.015390782617032528, + 0.004499780014157295, + 0.006866004317998886, + 0.0038970145396888256, + 0.006486274302005768, + 0.013834762386977673, + 0.005025445017963648, + 0.010253791697323322, + 0.015361836180090904, + 0.006626902148127556, + 0.006283457390964031, + 0.005265601444989443, + 0.01621861755847931, + 0.03002256341278553, + 0.022370437160134315, + 0.030168144032359123, + 0.022575613111257553, + 0.012091430835425854, + 0.014135457575321198, + 0.011432711966335773, + 0.002665985142812133, + 0.004476197063922882, + 0.002874745288863778, + 0.0056633842177689075, + 0.00501695042476058, + 0.002038691658526659, + 0.003895081579685211, + 0.0024528303183615208, + 0.004903452005237341, + 0.007124538067728281, + 0.005323347635567188, + 0.0059690107591450214, + 0.008061419241130352, + 0.003848171327263117, + 0.00487866997718811, + 0.005546339321881533, + 0.004952500108629465, + 0.006038186140358448, + 0.0062751732766628265, + 0.007718816865235567, + 0.006462922785431147, + 0.0036101662553846836, + 0.006461901590228081, + 0.0035854969173669815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0046377950347959995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_tensor([904, 904])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_tensor([904, 904])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc9c8a96bc6b95becad6866d5434e50fbe9b4ff7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_tensor([904, 904])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a static, ultra-high-definition, vertical medium shot shot from a slight upward angle, the delicate brushstrokes of Japanese theatrical animation are brought to full effect. A plump wax apple fruit, transformed into a spirit, stands gracefully amidst a vast, emerald lotus leaf with distinct veins. Her figure is elegant and slender, her skin embodying the wax apple's characteristic luminous hue, gradating from ivory white at the top to ruby red at the base. Her smooth, porcelain-like surface reflects a soft luster. Her face is delicate and ethereal, her eyes gently closed, her expression serene. Her hair is artfully shaped to resemble the calyx of the wax apple, seamlessly blending with the fruit itself. A soft, pearly, warm aura emanates from her, enveloping her and imbuing the scene with a sacred and dreamy aura. The background is a tranquil lake shrouded in morning mist. Soft, diffused light from the sky penetrates the fog, intertwining with the spirit's own radiance, casting a faint, blurred shadow on the lotus leaf. A large, crystalline drop of water slowly glides down the graceful curve of her side, subtly distorting the colors beneath it as it moves, leaving a fleeting trail of moisture before silently landing on the surface of a lotus leaf, creating a circle of tiny, perfectly formed ripples. The entire image is filled with exquisite detail, from the translucent texture of the elf's skin to the velvety texture of the lotus leaf. Each frame resembles a meticulously crafted illustration, creating an atmosphere of profound tranquility and pure poetry.", + "seed": 24632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_904_seed_24632_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1002559281885624, + "vlm_logprob": -0.1002559281885624, + "vlm_yes_no_margin": 2.949999976158142, + "vlm_token_credit": -0.1002559281885624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08222024887800217 + ], + "opposite_logprob_tokens": [ + -2.8322203159332275 + ], + "logprob_reward": -0.08222024887800217, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08222024887800217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17869843542575836, + 0.17330379784107208, + 0.03389348089694977, + 0.07363273203372955, + 0.11407741904258728, + 0.06749124079942703, + 0.18962064385414124, + 0.04895531013607979, + 0.05263856053352356, + 0.06768838316202164 + ], + "credit_map": [ + 0.0384102538228035, + 0.015288717113435268, + 0.01732272282242775, + 0.018925905227661133, + 0.015567246824502945, + 0.04285018891096115, + 0.013129787519574165, + 0.017203621566295624, + 0.032705292105674744, + 0.015436536632478237, + 0.013905804604291916, + 0.01244295109063387, + 0.02596483752131462, + 0.03860534355044365, + 0.01616636849939823, + 0.018076667562127113, + 0.006063634064048529, + 0.002423515310510993, + 0.003099437104538083, + 0.003367404919117689, + 0.0036869870964437723, + 0.00793614611029625, + 0.0029510853346437216, + 0.004365270026028156, + 0.009337348863482475, + 0.007696274667978287, + 0.005689850077033043, + 0.013532803393900394, + 0.006726917810738087, + 0.016740478575229645, + 0.006692038848996162, + 0.007217014208436012, + 0.011907432228326797, + 0.01090423483401537, + 0.014955108985304832, + 0.025844134390354156, + 0.008076047524809837, + 0.018860580399632454, + 0.010056954808533192, + 0.013472926802933216, + 0.009238575585186481, + 0.013575307093560696, + 0.005632458254694939, + 0.010921730659902096, + 0.007134787738323212, + 0.009996899403631687, + 0.005566772539168596, + 0.005424708593636751, + 0.016371499747037888, + 0.010320090688765049, + 0.05248700827360153, + 0.01592312753200531, + 0.00963195413351059, + 0.05183994024991989, + 0.00930881965905428, + 0.023738203570246696, + 0.004909936338663101, + 0.005562555976212025, + 0.006566507276147604, + 0.007881599478423595, + 0.005346296355128288, + 0.0057617430575191975, + 0.0064867720939219, + 0.006439899560064077, + 0.0053162709809839725, + 0.003996167331933975, + 0.007395569700747728, + 0.003964812494814396, + 0.004148554522544146, + 0.00905295554548502, + 0.006103545892983675, + 0.012660683132708073, + 0.006551950704306364, + 0.006375300232321024, + 0.009510092437267303, + 0.006925769150257111, + 0.005973129533231258, + 0.012651201337575912, + 0.008594161830842495, + 0.01110677607357502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.08222024887800217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810801595449448 + ], + "opposite_logprob_tokens": [ + -4.634810924530029 + ], + "logprob_reward": -0.009810801595449448, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810801595449448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33003777265548706, + 0.15327906608581543, + 0.06796875596046448, + 0.048524994403123856, + 0.09828151762485504, + 0.043932847678661346, + 0.11595182120800018, + 0.05570470914244652, + 0.04417463019490242, + 0.04214392974972725 + ], + "credit_map": [ + 0.05989318713545799, + 0.024314148351550102, + 0.03018750622868538, + 0.031831201165914536, + 0.02840953692793846, + 0.07734600454568863, + 0.02311556413769722, + 0.05494062602519989, + 0.023916101083159447, + 0.02081781066954136, + 0.014164377003908157, + 0.022526394575834274, + 0.020675426349043846, + 0.0283493772149086, + 0.011004979722201824, + 0.011824592016637325, + 0.009322130121290684, + 0.0077461558394134045, + 0.00720260851085186, + 0.012116718105971813, + 0.009610834531486034, + 0.010442838072776794, + 0.005311279557645321, + 0.006216196808964014, + 0.005180460400879383, + 0.005500953644514084, + 0.0045015388168394566, + 0.009296521544456482, + 0.0042612566612660885, + 0.009342395700514317, + 0.004285573493689299, + 0.006156293675303459, + 0.018411632627248764, + 0.011460302397608757, + 0.008539166301488876, + 0.01596618816256523, + 0.008274421095848083, + 0.01810450293123722, + 0.007402471732348204, + 0.010122835636138916, + 0.006580248940736055, + 0.004500256851315498, + 0.00584504846483469, + 0.006848991848528385, + 0.004131315741688013, + 0.0063751451671123505, + 0.004314228426665068, + 0.00533761503174901, + 0.013370228931307793, + 0.006697242613881826, + 0.02816978096961975, + 0.009164712391793728, + 0.008140823803842068, + 0.025080449879169464, + 0.007984603755176067, + 0.017343977466225624, + 0.005087885539978743, + 0.005871245171874762, + 0.00780000863596797, + 0.007585679180920124, + 0.004712920635938644, + 0.007160824723541737, + 0.005967719480395317, + 0.011518428102135658, + 0.00457254983484745, + 0.0038784730713814497, + 0.006432404741644859, + 0.004180624149739742, + 0.003933988511562347, + 0.006949273869395256, + 0.0049012210220098495, + 0.009326099418103695, + 0.0029397294856607914, + 0.005493101663887501, + 0.005211796145886183, + 0.0058046248741447926, + 0.0045131295919418335, + 0.005868143402040005, + 0.006752211134880781, + 0.005561193451285362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009810801595449448 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521834671497345 + ], + "opposite_logprob_tokens": [ + -1.502183437347412 + ], + "logprob_reward": -0.2521834671497345, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521834671497345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11719787120819092, + 0.09760299324989319, + 0.06449709832668304, + 0.03381052613258362, + 0.042399968951940536, + 0.024532558396458626, + 0.5469974279403687, + 0.022652586922049522, + 0.0214205551892519, + 0.02888842485845089 + ], + "credit_map": [ + 0.020103774964809418, + 0.01221194863319397, + 0.013921298086643219, + 0.014453884214162827, + 0.014513708651065826, + 0.018080011010169983, + 0.010747166350483894, + 0.013166072778403759, + 0.025725416839122772, + 0.006521258968859911, + 0.005743308458477259, + 0.006573045626282692, + 0.007581210229545832, + 0.03307836875319481, + 0.004072268959134817, + 0.008308121003210545, + 0.014823478646576405, + 0.003058601403608918, + 0.005097543820738792, + 0.003097143489867449, + 0.0038507282733917236, + 0.023164447396993637, + 0.00364640261977911, + 0.00775875011458993, + 0.00687787588685751, + 0.001895866822451353, + 0.0026404832024127245, + 0.002672165399417281, + 0.002123431768268347, + 0.011899675242602825, + 0.0021085659973323345, + 0.003592464840039611, + 0.007017769385129213, + 0.002152811037376523, + 0.0042031933553516865, + 0.006471571978181601, + 0.002283622045069933, + 0.013715841807425022, + 0.0028460535686463118, + 0.0037091108970344067, + 0.0030831415206193924, + 0.0025974358431994915, + 0.0040838154964149, + 0.0034832521341741085, + 0.0022914293222129345, + 0.002916318830102682, + 0.0023847182746976614, + 0.0036924469750374556, + 0.043016206473112106, + 0.021062936633825302, + 0.19966062903404236, + 0.027983209118247032, + 0.017568660899996758, + 0.15285295248031616, + 0.013589725829660892, + 0.07126309722661972, + 0.0030440445989370346, + 0.002158333780243993, + 0.0030049378983676434, + 0.001559391850605607, + 0.0020095184445381165, + 0.004122852347791195, + 0.002520732581615448, + 0.004232774488627911, + 0.0024918930139392614, + 0.00157470116391778, + 0.0035233127418905497, + 0.0017797197215259075, + 0.0016731146024540067, + 0.004386679735034704, + 0.00245142774656415, + 0.003539706114679575, + 0.0035058297216892242, + 0.00212594261392951, + 0.004345754161477089, + 0.0029511675238609314, + 0.0027805990539491177, + 0.005472130607813597, + 0.004342495929449797, + 0.0033645043149590492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.2521834671497345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.142861008644104 + ], + "opposite_logprob_tokens": [ + -2.0178608894348145 + ], + "logprob_reward": -0.142861008644104, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.142861008644104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.176727756857872, + 0.13584959506988525, + 0.02766253985464573, + 0.036360785365104675, + 0.09126077592372894, + 0.057585425674915314, + 0.38134512305259705, + 0.033023785799741745, + 0.025937527418136597, + 0.03424672409892082 + ], + "credit_map": [ + 0.03840356692671776, + 0.011093362234532833, + 0.01499997265636921, + 0.012599238194525242, + 0.01352831069380045, + 0.051159992814064026, + 0.012762047350406647, + 0.022181252017617226, + 0.02905563823878765, + 0.010489414446055889, + 0.0074045779183506966, + 0.0072350711561739445, + 0.0170613843947649, + 0.041094787418842316, + 0.00970599427819252, + 0.013802711851894855, + 0.0035713587421923876, + 0.002025234280154109, + 0.0038893434684723616, + 0.0038358536548912525, + 0.004125112667679787, + 0.003908255137503147, + 0.0028007621876895428, + 0.0035066199488937855, + 0.00496595399454236, + 0.006884334608912468, + 0.003183447290211916, + 0.00568906357511878, + 0.0032740598544478416, + 0.005441967397928238, + 0.003410483244806528, + 0.0035114747006446123, + 0.013124363496899605, + 0.0059189326129853725, + 0.007842626422643661, + 0.017769962549209595, + 0.005194494500756264, + 0.02706357091665268, + 0.006402459926903248, + 0.007944358512759209, + 0.008540069684386253, + 0.007618851959705353, + 0.0056051346473395824, + 0.008116976357996464, + 0.0051755174063146114, + 0.011722791939973831, + 0.005837723612785339, + 0.0049683586694300175, + 0.03041481040418148, + 0.014657438732683659, + 0.13962554931640625, + 0.018328579142689705, + 0.012465232983231544, + 0.10403823852539062, + 0.010608497075736523, + 0.05120677873492241, + 0.0035223059821873903, + 0.0029881461523473263, + 0.004356435500085354, + 0.0042396788485348225, + 0.0029376973398029804, + 0.005872558802366257, + 0.0031136348843574524, + 0.005993327125906944, + 0.002385178115218878, + 0.0018056565895676613, + 0.003768322989344597, + 0.0025853589177131653, + 0.002155735157430172, + 0.004598182626068592, + 0.0030689891427755356, + 0.0055701048113405704, + 0.002960572950541973, + 0.0024318343494087458, + 0.005037845578044653, + 0.0025593675673007965, + 0.002728092484176159, + 0.0058804224245250225, + 0.005133446771651506, + 0.0075151450000703335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.142861008644104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420411467552185 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.01420411467552185, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420411467552185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18260648846626282, + 0.06490179151296616, + 0.0923398807644844, + 0.04974091053009033, + 0.08627357333898544, + 0.05742758512496948, + 0.3751012682914734, + 0.04838204383850098, + 0.014315946027636528, + 0.028910517692565918 + ], + "credit_map": [ + 0.031563837081193924, + 0.011883093044161797, + 0.02065558172762394, + 0.015145369805395603, + 0.012174896895885468, + 0.045287784188985825, + 0.011013124138116837, + 0.0348827987909317, + 0.008362793363630772, + 0.007192568387836218, + 0.009156138636171818, + 0.007469716016203165, + 0.008388129994273186, + 0.008659152314066887, + 0.006904118228703737, + 0.008769174106419086, + 0.011342613026499748, + 0.011769686825573444, + 0.010236699134111404, + 0.011833888478577137, + 0.009179850108921528, + 0.01593303307890892, + 0.008859897032380104, + 0.01318422146141529, + 0.0056892214342951775, + 0.0066997697576880455, + 0.005767433438450098, + 0.01012767106294632, + 0.0051977746188640594, + 0.0059201414696872234, + 0.004577817860990763, + 0.005761082284152508, + 0.013844688422977924, + 0.00856983382254839, + 0.006229907739907503, + 0.012324566021561623, + 0.007242564111948013, + 0.023455161601305008, + 0.007729297038167715, + 0.006877552717924118, + 0.007683837320655584, + 0.010548464953899384, + 0.004730597138404846, + 0.009744768030941486, + 0.005856935866177082, + 0.009428896941244602, + 0.005160696338862181, + 0.00427338806912303, + 0.029085006564855576, + 0.015746813267469406, + 0.1346922069787979, + 0.020436706021428108, + 0.013051465153694153, + 0.10438167303800583, + 0.009504047222435474, + 0.04820334538817406, + 0.006183224264532328, + 0.003845751052722335, + 0.007277465891093016, + 0.004717939533293247, + 0.0033262253273278475, + 0.010491248220205307, + 0.003744848072528839, + 0.008795341476798058, + 0.0015288629801943898, + 0.0013970177387818694, + 0.0018703874666243792, + 0.0019016613950952888, + 0.0015406158054247499, + 0.0023424155078828335, + 0.0017839653883129358, + 0.0019510206766426563, + 0.0028326879255473614, + 0.0026189505588263273, + 0.003762022592127323, + 0.0025836543645709753, + 0.0024390884209424257, + 0.006310253404080868, + 0.003580533666536212, + 0.004783324897289276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.01420411467552185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, ultra-high-definition, vertical medium shot shot from a slight upward angle, the delicate brushstrokes of Japanese theatrical animation are brought to full effect. A plump wax apple fruit, transformed into a spirit, stands gracefully amidst a vast, emerald lotus leaf with distinct veins. Her figure is elegant and slender, her skin embodying the wax apple's characteristic luminous hue, gradating from ivory white at the top to ruby red at the base. Her smooth, porcelain-like surface reflects a soft luster. Her face is delicate and ethereal, her eyes gently closed, her expression serene. Her hair is artfully shaped to resemble the calyx of the wax apple, seamlessly blending with the fruit itself. A soft, pearly, warm aura emanates from her, enveloping her and imbuing the scene with a sacred and dreamy aura. The background is a tranquil lake shrouded in morning mist. Soft, diffused light from the sky penetrates the fog, intertwining with the spirit's own radiance, casting a faint, blurred shadow on the lotus leaf. A large, crystalline drop of water slowly glides down the graceful curve of her side, subtly distorting the colors beneath it as it moves, leaving a fleeting trail of moisture before silently landing on the surface of a lotus leaf, creating a circle of tiny, perfectly formed ripples. The entire image is filled with exquisite detail, from the translucent texture of the elf's skin to the velvety texture of the lotus leaf. Each frame resembles a meticulously crafted illustration, creating an atmosphere of profound tranquility and pure poetry.", + "seed": 24633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_904_seed_24633_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11834046021103858, + "vlm_logprob": -0.11834046021103858, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.11834046021103858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06591203808784485 + ], + "opposite_logprob_tokens": [ + -2.9409120082855225 + ], + "logprob_reward": -0.06591203808784485, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06591203808784485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33034226298332214, + 0.0894651785492897, + 0.10136345028877258, + 0.08828587085008621, + 0.11664368957281113, + 0.09555797278881073, + 0.037633366882801056, + 0.030103936791419983, + 0.0423225536942482, + 0.06828173249959946 + ], + "credit_map": [ + 0.03494550660252571, + 0.04648953303694725, + 0.05935264751315117, + 0.027004513889551163, + 0.034057632088661194, + 0.05976743623614311, + 0.0330699160695076, + 0.035655055195093155, + 0.009776092134416103, + 0.011977753601968288, + 0.014715118333697319, + 0.007100473158061504, + 0.010560225695371628, + 0.012319735251367092, + 0.010956293903291225, + 0.012059483677148819, + 0.015378706157207489, + 0.009085464291274548, + 0.015402114018797874, + 0.005916636437177658, + 0.01024889200925827, + 0.026075275614857674, + 0.007116404362022877, + 0.012139946222305298, + 0.010271883569657803, + 0.009419344365596771, + 0.01632886566221714, + 0.005784207955002785, + 0.007419626694172621, + 0.011814448051154613, + 0.008820140734314919, + 0.018427353352308273, + 0.014246908016502857, + 0.015072454698383808, + 0.018512016162276268, + 0.011159085668623447, + 0.01427351776510477, + 0.01576075330376625, + 0.012898261658847332, + 0.014720688574016094, + 0.01165340468287468, + 0.010668403469026089, + 0.013325950130820274, + 0.009478411637246609, + 0.007998021319508553, + 0.019725313410162926, + 0.0072998651303350925, + 0.0154085922986269, + 0.004431808367371559, + 0.004346924833953381, + 0.005933735985308886, + 0.004059225786477327, + 0.004988358821719885, + 0.005773214623332024, + 0.003709483426064253, + 0.004390612710267305, + 0.003881203942000866, + 0.004238621331751347, + 0.004733857233077288, + 0.0030314484611153603, + 0.00431012362241745, + 0.002916814526543021, + 0.00280868005938828, + 0.0041831862181425095, + 0.004992592614144087, + 0.004197771195322275, + 0.0075590298511087894, + 0.0029872292652726173, + 0.004148622043430805, + 0.007354022469371557, + 0.004061933606863022, + 0.007021349389106035, + 0.009571352042257786, + 0.007852507755160332, + 0.015084002166986465, + 0.00575998006388545, + 0.005186944734305143, + 0.008726111613214016, + 0.005969468969851732, + 0.01013135351240635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.06591203808784485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029830025508999825 + ], + "opposite_logprob_tokens": [ + -3.529829978942871 + ], + "logprob_reward": -0.029830025508999825, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029830025508999825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2741357684135437, + 0.1319131702184677, + 0.08365319669246674, + 0.0732712373137474, + 0.10620909929275513, + 0.08971080929040909, + 0.1028975322842598, + 0.026151342317461967, + 0.07099834084510803, + 0.04105949029326439 + ], + "credit_map": [ + 0.03685743734240532, + 0.042945779860019684, + 0.029519427567720413, + 0.02616187557578087, + 0.045822031795978546, + 0.04644688963890076, + 0.023194247856736183, + 0.023188084363937378, + 0.014300563372671604, + 0.02301921881735325, + 0.015206445939838886, + 0.015301293693482876, + 0.02006141096353531, + 0.016926534473896027, + 0.013641906902194023, + 0.013455797918140888, + 0.010255162604153156, + 0.010861294344067574, + 0.00806297454982996, + 0.0073626842349767685, + 0.008645493537187576, + 0.02107606828212738, + 0.005569345783442259, + 0.011820175684988499, + 0.006910179276019335, + 0.010654633864760399, + 0.01063076127320528, + 0.007432777900248766, + 0.01038299035280943, + 0.00818632636219263, + 0.007691261358559132, + 0.011382309719920158, + 0.014031929895281792, + 0.016870955005288124, + 0.014475807547569275, + 0.01139388233423233, + 0.013079985976219177, + 0.01310558058321476, + 0.009565599262714386, + 0.013685357756912708, + 0.011007930152118206, + 0.010842557065188885, + 0.011816225945949554, + 0.009915817528963089, + 0.008850710466504097, + 0.016116594895720482, + 0.007747790776193142, + 0.013413180597126484, + 0.012060732580721378, + 0.01057768240571022, + 0.01647871732711792, + 0.009030166082084179, + 0.011464904062449932, + 0.022984934970736504, + 0.008081530220806599, + 0.012218865565955639, + 0.0024299235083162785, + 0.0032872874289751053, + 0.005143787246197462, + 0.00208954606205225, + 0.003548292675986886, + 0.004711155779659748, + 0.0019578090868890285, + 0.002983539132401347, + 0.008331929333508015, + 0.012892688624560833, + 0.007632634602487087, + 0.00779512757435441, + 0.010467379353940487, + 0.009438223205506802, + 0.008160516619682312, + 0.006279841996729374, + 0.005448271986097097, + 0.004420699551701546, + 0.007941938936710358, + 0.0028851572424173355, + 0.00441724993288517, + 0.005527242086827755, + 0.004601082764565945, + 0.005817847326397896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.029830025508999825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562631964683533 + ], + "opposite_logprob_tokens": [ + -1.6006263494491577 + ], + "logprob_reward": -0.22562631964683533, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562631964683533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30736804008483887, + 0.08848363161087036, + 0.045119501650333405, + 0.032667335122823715, + 0.0460251122713089, + 0.11726014316082001, + 0.12496607005596161, + 0.04537476599216461, + 0.11777906864881516, + 0.07495630532503128 + ], + "credit_map": [ + 0.03471117466688156, + 0.028208063915371895, + 0.049858447164297104, + 0.02460264042019844, + 0.028929496183991432, + 0.06815175712108612, + 0.026987804099917412, + 0.04591868072748184, + 0.012786277569830418, + 0.010195793583989143, + 0.015546760521829128, + 0.00774078955873847, + 0.009960290044546127, + 0.011345501989126205, + 0.008250595070421696, + 0.012657620012760162, + 0.008407633751630783, + 0.00450688973069191, + 0.007365371100604534, + 0.003717632731422782, + 0.004712726455181837, + 0.006600529421120882, + 0.0036527435295283794, + 0.006155975628644228, + 0.0041137454099953175, + 0.003129107877612114, + 0.005196092184633017, + 0.001890449202619493, + 0.002680349862203002, + 0.0074007255025208, + 0.0033215272706001997, + 0.004935337696224451, + 0.004946987610310316, + 0.005405713338404894, + 0.00754186138510704, + 0.0036490068305283785, + 0.004723539110273123, + 0.007641748525202274, + 0.00436527980491519, + 0.007750974502414465, + 0.012142480351030827, + 0.012111900374293327, + 0.014754565432667732, + 0.010019266977906227, + 0.015622026287019253, + 0.0269290991127491, + 0.00747363455593586, + 0.01820717379450798, + 0.013047316111624241, + 0.015303881838917732, + 0.015754466876387596, + 0.011583119630813599, + 0.018307838588953018, + 0.025342607870697975, + 0.010086849331855774, + 0.015539993532001972, + 0.005637752823531628, + 0.006569774821400642, + 0.006798348389565945, + 0.004222868010401726, + 0.005746247246861458, + 0.008713982068002224, + 0.002888035262003541, + 0.0047977580688893795, + 0.01430053822696209, + 0.012504401616752148, + 0.020362360402941704, + 0.009588757529854774, + 0.012939601205289364, + 0.016208982095122337, + 0.015012973919510841, + 0.016861455515027046, + 0.008137171156704426, + 0.009384832344949245, + 0.011314046569168568, + 0.006786117795854807, + 0.007582932244986296, + 0.008097880519926548, + 0.014767429791390896, + 0.008885888382792473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.22562631964683533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213584303855896 + ], + "opposite_logprob_tokens": [ + -1.5021358728408813 + ], + "logprob_reward": -0.25213584303855896, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213584303855896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1221679225564003, + 0.09419732540845871, + 0.11242502182722092, + 0.13271953165531158, + 0.16616204380989075, + 0.16916167736053467, + 0.05252295732498169, + 0.028458688408136368, + 0.05060572922229767, + 0.07157915085554123 + ], + "credit_map": [ + 0.019952338188886642, + 0.01288903784006834, + 0.02217055670917034, + 0.009805000387132168, + 0.01365107111632824, + 0.015695977956056595, + 0.008370880968868732, + 0.01963304914534092, + 0.011557480320334435, + 0.016155816614627838, + 0.011087911203503609, + 0.010287165641784668, + 0.012843694537878036, + 0.011426189914345741, + 0.009833060204982758, + 0.01100599393248558, + 0.011931450106203556, + 0.020960280671715736, + 0.0104523291811347, + 0.012876655906438828, + 0.016318416222929955, + 0.018021846190094948, + 0.010123693384230137, + 0.011740339919924736, + 0.018940288573503494, + 0.011850194074213505, + 0.023709213361144066, + 0.008533581160008907, + 0.009736436419188976, + 0.034331198781728745, + 0.00823442917317152, + 0.017384188249707222, + 0.022663312032818794, + 0.02992907166481018, + 0.02105473168194294, + 0.018805069848895073, + 0.020467622205615044, + 0.023110637441277504, + 0.014023153111338615, + 0.016108427196741104, + 0.019459156319499016, + 0.027859119698405266, + 0.02521628327667713, + 0.02157132886350155, + 0.016492152586579323, + 0.022467633709311485, + 0.015013415366411209, + 0.0210825614631176, + 0.005875718779861927, + 0.004801091272383928, + 0.011217025108635426, + 0.0036015301011502743, + 0.004510478116571903, + 0.011395038105547428, + 0.003949481062591076, + 0.007172590121626854, + 0.003449150128290057, + 0.003590598003938794, + 0.0053251637145876884, + 0.0021774566266685724, + 0.003135115373879671, + 0.004514678847044706, + 0.002450293628498912, + 0.0038162297569215298, + 0.006006036885082722, + 0.004866491071879864, + 0.011250823736190796, + 0.003913426771759987, + 0.003642421681433916, + 0.007446236442774534, + 0.00566489901393652, + 0.007815389893949032, + 0.009223349392414093, + 0.009222697466611862, + 0.012179077602922916, + 0.0062378281727433205, + 0.008177359588444233, + 0.008501018397510052, + 0.008585785515606403, + 0.009452025406062603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.25213584303855896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198074772953987 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018198074772953987, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198074772953987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21686996519565582, + 0.15902741253376007, + 0.08014030009508133, + 0.11618509888648987, + 0.10361642390489578, + 0.07779703289270401, + 0.09269660711288452, + 0.03865260258316994, + 0.020301807671785355, + 0.0947127491235733 + ], + "credit_map": [ + 0.024522405117750168, + 0.030546939000487328, + 0.03634503111243248, + 0.017631931230425835, + 0.026204321533441544, + 0.04505566135048866, + 0.016961539164185524, + 0.019602138549089432, + 0.020995600149035454, + 0.024211030453443527, + 0.021207038313150406, + 0.017768967896699905, + 0.019723698496818542, + 0.021316761150956154, + 0.015280679799616337, + 0.018523626029491425, + 0.009610461071133614, + 0.012451532296836376, + 0.007218989543616772, + 0.00992508139461279, + 0.010268881916999817, + 0.013958466239273548, + 0.006809813901782036, + 0.009897071868181229, + 0.012132717296481133, + 0.014235666953027248, + 0.022295232862234116, + 0.008520486764609814, + 0.014367024414241314, + 0.018842972815036774, + 0.009722518734633923, + 0.0160684734582901, + 0.011167819611728191, + 0.017052600160241127, + 0.0106132123619318, + 0.012512111105024815, + 0.01886371150612831, + 0.012263131327927113, + 0.010905148461461067, + 0.010238686576485634, + 0.008966581895947456, + 0.012015368789434433, + 0.009106501005589962, + 0.008004549890756607, + 0.014006322249770164, + 0.009195440448820591, + 0.007637939881533384, + 0.008864332921802998, + 0.011465823277831078, + 0.008553393185138702, + 0.016876643523573875, + 0.007760720327496529, + 0.006959688384085894, + 0.02148936502635479, + 0.007510223425924778, + 0.012080756947398186, + 0.004784503951668739, + 0.006029940210282803, + 0.004491978324949741, + 0.0033859671093523502, + 0.006215957924723625, + 0.006569953635334969, + 0.0034592770971357822, + 0.0037150259595364332, + 0.0020103449933230877, + 0.002564850263297558, + 0.0034132127184420824, + 0.0019278300460428, + 0.0028320651035755873, + 0.0025970980059355497, + 0.002870753174647689, + 0.0020856535993516445, + 0.010822237469255924, + 0.019306587055325508, + 0.009079146198928356, + 0.011837424710392952, + 0.014361847192049026, + 0.01156532671302557, + 0.011027132160961628, + 0.0067130490206182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.018198074772953987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, ultra-high-definition, vertical medium shot shot from a slight upward angle, the delicate brushstrokes of Japanese theatrical animation are brought to full effect. A plump wax apple fruit, transformed into a spirit, stands gracefully amidst a vast, emerald lotus leaf with distinct veins. Her figure is elegant and slender, her skin embodying the wax apple's characteristic luminous hue, gradating from ivory white at the top to ruby red at the base. Her smooth, porcelain-like surface reflects a soft luster. Her face is delicate and ethereal, her eyes gently closed, her expression serene. Her hair is artfully shaped to resemble the calyx of the wax apple, seamlessly blending with the fruit itself. A soft, pearly, warm aura emanates from her, enveloping her and imbuing the scene with a sacred and dreamy aura. The background is a tranquil lake shrouded in morning mist. Soft, diffused light from the sky penetrates the fog, intertwining with the spirit's own radiance, casting a faint, blurred shadow on the lotus leaf. A large, crystalline drop of water slowly glides down the graceful curve of her side, subtly distorting the colors beneath it as it moves, leaving a fleeting trail of moisture before silently landing on the surface of a lotus leaf, creating a circle of tiny, perfectly formed ripples. The entire image is filled with exquisite detail, from the translucent texture of the elf's skin to the velvety texture of the lotus leaf. Each frame resembles a meticulously crafted illustration, creating an atmosphere of profound tranquility and pure poetry.", + "seed": 24634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_904_seed_24634_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07272227499634028, + "vlm_logprob": -0.07272227499634028, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.07272227499634028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049697741866111755 + ], + "opposite_logprob_tokens": [ + -3.4246976375579834 + ], + "logprob_reward": -0.049697741866111755, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.049697741866111755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.260309100151062, + 0.08983057737350464, + 0.10402380675077438, + 0.030698228627443314, + 0.06965610384941101, + 0.061757300049066544, + 0.08069653064012527, + 0.08974242210388184, + 0.16566060483455658, + 0.047625310719013214 + ], + "credit_map": [ + 0.03215798735618591, + 0.01990218088030815, + 0.04245675727725029, + 0.03560235723853111, + 0.023625098168849945, + 0.02976118214428425, + 0.02959394082427025, + 0.04720957949757576, + 0.016286658123135567, + 0.00971705000847578, + 0.011646650731563568, + 0.0073511796072125435, + 0.008183002471923828, + 0.01841598004102707, + 0.0075674415566027164, + 0.010662601329386234, + 0.030233480036258698, + 0.007266554515808821, + 0.006535789929330349, + 0.005084038246423006, + 0.006004902068525553, + 0.03600078448653221, + 0.006138627417385578, + 0.006759622599929571, + 0.003627278609201312, + 0.0034533157013356686, + 0.004982649348676205, + 0.004677190911024809, + 0.0022194995544850826, + 0.00400773249566555, + 0.002855919301509857, + 0.004874639678746462, + 0.008161094039678574, + 0.009050148539245129, + 0.00936385989189148, + 0.01138473954051733, + 0.007265114225447178, + 0.006389898248016834, + 0.007033917587250471, + 0.011007326655089855, + 0.007150983437895775, + 0.008704339154064655, + 0.005988960154354572, + 0.008913647383451462, + 0.007092972286045551, + 0.006961297709494829, + 0.00815120991319418, + 0.008793882094323635, + 0.007150128949433565, + 0.011760915629565716, + 0.00978265330195427, + 0.014134359546005726, + 0.007516095880419016, + 0.00661104591563344, + 0.010598067194223404, + 0.013143252581357956, + 0.012335260398685932, + 0.010294129140675068, + 0.014565880410373211, + 0.01283544022589922, + 0.010568087920546532, + 0.007494194433093071, + 0.010590381920337677, + 0.01105903647840023, + 0.016976190730929375, + 0.020128119736909866, + 0.015383267775177956, + 0.017443224787712097, + 0.028433019295334816, + 0.02123088389635086, + 0.025767570361495018, + 0.020298315212130547, + 0.004980864003300667, + 0.005074840504676104, + 0.005277653224766254, + 0.005816312041133642, + 0.005017294082790613, + 0.006396582815796137, + 0.006004529073834419, + 0.009057228453457355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.049697741866111755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820075258612633 + ], + "opposite_logprob_tokens": [ + -3.529819965362549 + ], + "logprob_reward": -0.029820075258612633, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820075258612633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24936231970787048, + 0.10532750934362411, + 0.12971681356430054, + 0.04038805514574051, + 0.031655583530664444, + 0.11907970160245895, + 0.04181738570332527, + 0.04914639890193939, + 0.040496934205293655, + 0.19300925731658936 + ], + "credit_map": [ + 0.029594095423817635, + 0.027076443657279015, + 0.03560231253504753, + 0.021092090755701065, + 0.052309777587652206, + 0.028319979086518288, + 0.028876645490527153, + 0.02649099752306938, + 0.01783096417784691, + 0.012569084763526917, + 0.013916581869125366, + 0.010024547576904297, + 0.006319859065115452, + 0.0190461166203022, + 0.009115606546401978, + 0.01650475151836872, + 0.03671160712838173, + 0.009049909189343452, + 0.009157903492450714, + 0.006602328736335039, + 0.008575630374252796, + 0.04396140202879906, + 0.006670392584055662, + 0.008987664245069027, + 0.0048452625051140785, + 0.0052897315472364426, + 0.004734434653073549, + 0.004752344451844692, + 0.004584087990224361, + 0.00569493230432272, + 0.0038587572053074837, + 0.006628507748246193, + 0.0042760842479765415, + 0.0037459188606590033, + 0.004740030039101839, + 0.003802412422373891, + 0.002143594203516841, + 0.0070391385816037655, + 0.0021181569900363684, + 0.0037902495823800564, + 0.00847457256168127, + 0.011370431631803513, + 0.02337944135069847, + 0.03754838556051254, + 0.008613516576588154, + 0.008182359859347343, + 0.0065727452747523785, + 0.014938252046704292, + 0.0038219382986426353, + 0.005522613413631916, + 0.0043724109418690205, + 0.006983589380979538, + 0.005295158829540014, + 0.0033118571154773235, + 0.006577868480235338, + 0.005931952968239784, + 0.00446130009368062, + 0.006082294508814812, + 0.006213821470737457, + 0.00802343524992466, + 0.006327962502837181, + 0.004074908792972565, + 0.005846111103892326, + 0.008116573095321655, + 0.004609791096299887, + 0.004983535036444664, + 0.004622759763151407, + 0.003948837518692017, + 0.005443825852125883, + 0.005419542081654072, + 0.0067571671679615974, + 0.004711478017270565, + 0.02254301868379116, + 0.03972870483994484, + 0.022243931889533997, + 0.02103879489004612, + 0.01690336875617504, + 0.03124750219285488, + 0.018259473145008087, + 0.021044494584202766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.029820075258612633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282608032226562 + ], + "opposite_logprob_tokens": [ + -2.0178260803222656 + ], + "logprob_reward": -0.14282608032226562, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282608032226562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11649656295776367, + 0.10495918989181519, + 0.29987436532974243, + 0.05007429048418999, + 0.0437830425798893, + 0.07826953381299973, + 0.07769212871789932, + 0.04168548062443733, + 0.13133934140205383, + 0.055826105177402496 + ], + "credit_map": [ + 0.014721817336976528, + 0.011611373163759708, + 0.024758005514740944, + 0.01509716734290123, + 0.011467326432466507, + 0.012973904609680176, + 0.009078326635062695, + 0.016788633540272713, + 0.015612229704856873, + 0.013079045340418816, + 0.013976878486573696, + 0.010787013918161392, + 0.008674196898937225, + 0.017182568088173866, + 0.010105478577315807, + 0.015541774220764637, + 0.09992299973964691, + 0.01639746129512787, + 0.013995335437357426, + 0.011222589761018753, + 0.01049873884767294, + 0.12254133075475693, + 0.008819583803415298, + 0.016476299613714218, + 0.006784382276237011, + 0.005657856818288565, + 0.006913344841450453, + 0.005848857574164867, + 0.004933372139930725, + 0.008853813633322716, + 0.003933794796466827, + 0.0071488674730062485, + 0.006242593750357628, + 0.004086155444383621, + 0.005192390643060207, + 0.0060922228731215, + 0.0032264033798128366, + 0.010835097171366215, + 0.003354431129992008, + 0.004753742832690477, + 0.009230043739080429, + 0.009775187820196152, + 0.012987256981432438, + 0.015853120014071465, + 0.004877571482211351, + 0.006479450035840273, + 0.0073084840551018715, + 0.011758417822420597, + 0.008664011023938656, + 0.017259472981095314, + 0.009381269104778767, + 0.010617148131132126, + 0.005614243447780609, + 0.009890783578157425, + 0.0051583400927484035, + 0.011106854304671288, + 0.004451049491763115, + 0.008803395554423332, + 0.004479659721255302, + 0.008150197565555573, + 0.0036849002353847027, + 0.004499305505305529, + 0.0035305896308273077, + 0.00408638222143054, + 0.012614108622074127, + 0.0139791714027524, + 0.014819437637925148, + 0.014565395191311836, + 0.015839314088225365, + 0.019976424053311348, + 0.017575036734342575, + 0.021970441564917564, + 0.005525322165340185, + 0.01002186443656683, + 0.006406689062714577, + 0.0049456264823675156, + 0.0049850125797092915, + 0.009093028493225574, + 0.005538990255445242, + 0.009309567511081696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.14282608032226562 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270674765110016 + ], + "opposite_logprob_tokens": [ + -2.1270675659179688 + ], + "logprob_reward": -0.1270674765110016, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270674765110016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15884964168071747, + 0.06900383532047272, + 0.029447825625538826, + 0.020778952166438103, + 0.04933458939194679, + 0.03922474756836891, + 0.0853797122836113, + 0.07465376704931259, + 0.08978185802698135, + 0.38354507088661194 + ], + "credit_map": [ + 0.02065659500658512, + 0.013585156761109829, + 0.022444477304816246, + 0.020505348220467567, + 0.019797896966338158, + 0.01529776118695736, + 0.015635699033737183, + 0.03092670440673828, + 0.007081936113536358, + 0.007491451688110828, + 0.008741620928049088, + 0.00667697936296463, + 0.014831501059234142, + 0.009163914248347282, + 0.006883693393319845, + 0.00813273899257183, + 0.005817013792693615, + 0.002799996407702565, + 0.0035095643252134323, + 0.0028265146538615227, + 0.0022437393199652433, + 0.007318846881389618, + 0.0022703225258737803, + 0.0026618284173309803, + 0.002238675020635128, + 0.003182200714945793, + 0.0028948725666850805, + 0.002357920166105032, + 0.0024870706256479025, + 0.0024468251504004, + 0.0022063658107072115, + 0.0029650223441421986, + 0.007287605665624142, + 0.0044483039528131485, + 0.00642037857323885, + 0.007232367526739836, + 0.0023787561804056168, + 0.013535593636333942, + 0.0033047397155314684, + 0.00472684483975172, + 0.002818385837599635, + 0.006089922972023487, + 0.006054542493075132, + 0.007731501013040543, + 0.0031958927866071463, + 0.004433225840330124, + 0.0030284621752798557, + 0.005872813053429127, + 0.007884737104177475, + 0.009528089314699173, + 0.014533106237649918, + 0.01595509611070156, + 0.00803409144282341, + 0.006154716480523348, + 0.008285065181553364, + 0.015004807151854038, + 0.00904221273958683, + 0.008915519341826439, + 0.010549211874604225, + 0.012051161378622055, + 0.009102492593228817, + 0.005596908275038004, + 0.009352763183414936, + 0.010043495334684849, + 0.006895148195326328, + 0.012610488571226597, + 0.00926491990685463, + 0.01013491302728653, + 0.01512869168072939, + 0.011949124746024609, + 0.014463844709098339, + 0.009334735572338104, + 0.04566817358136177, + 0.08468936383724213, + 0.043140050023794174, + 0.04680212587118149, + 0.03288206458091736, + 0.05754748731851578, + 0.036366526037454605, + 0.03644929453730583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.1270674765110016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200001023709774 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200001023709774, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200001023709774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17528016865253448, + 0.0897873193025589, + 0.1399587243795395, + 0.03710297867655754, + 0.04258124157786369, + 0.14450697600841522, + 0.05557216703891754, + 0.06452332437038422, + 0.10422158986330032, + 0.1464655101299286 + ], + "credit_map": [ + 0.01764388009905815, + 0.020272191613912582, + 0.021009691059589386, + 0.016721036285161972, + 0.03568534553050995, + 0.022481851279735565, + 0.02159101888537407, + 0.019875144585967064, + 0.017071528360247612, + 0.008817372843623161, + 0.010485977865755558, + 0.008331838063895702, + 0.005608390551060438, + 0.017618628218770027, + 0.007560875732451677, + 0.014292706735432148, + 0.04674177244305611, + 0.007712909020483494, + 0.006399629637598991, + 0.0054070730693638325, + 0.005154078360646963, + 0.057137321680784225, + 0.004017834551632404, + 0.007388090714812279, + 0.003976298496127129, + 0.00464096711948514, + 0.003453719662502408, + 0.005547146778553724, + 0.004641905892640352, + 0.005228355992585421, + 0.003581603290513158, + 0.006032982841134071, + 0.006524277850985527, + 0.0035746623761951923, + 0.00592066952958703, + 0.004913025069981813, + 0.0024716886691749096, + 0.011270948685705662, + 0.003181618172675371, + 0.004724346566945314, + 0.010693544521927834, + 0.01715807430446148, + 0.026046285405755043, + 0.03783954679965973, + 0.016961228102445602, + 0.010760214179754257, + 0.009616807103157043, + 0.015431278385221958, + 0.005705042742192745, + 0.011066636070609093, + 0.006899470463395119, + 0.009367985650897026, + 0.004727199207991362, + 0.004528714809566736, + 0.005464147310703993, + 0.007812965661287308, + 0.009521322324872017, + 0.009579053148627281, + 0.00942469947040081, + 0.011133438907563686, + 0.006100740749388933, + 0.006235050968825817, + 0.005811840295791626, + 0.00671717245131731, + 0.00886671245098114, + 0.010609795339405537, + 0.01282364223152399, + 0.010713702999055386, + 0.011059294454753399, + 0.018166178837418556, + 0.013449016027152538, + 0.018533244729042053, + 0.014916063286364079, + 0.03172103315591812, + 0.01503695547580719, + 0.0187046118080616, + 0.012851737439632416, + 0.02149377018213272, + 0.012810270301997662, + 0.018931064754724503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.014200001023709774 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, ultra-high-definition, vertical medium shot shot from a slight upward angle, the delicate brushstrokes of Japanese theatrical animation are brought to full effect. A plump wax apple fruit, transformed into a spirit, stands gracefully amidst a vast, emerald lotus leaf with distinct veins. Her figure is elegant and slender, her skin embodying the wax apple's characteristic luminous hue, gradating from ivory white at the top to ruby red at the base. Her smooth, porcelain-like surface reflects a soft luster. Her face is delicate and ethereal, her eyes gently closed, her expression serene. Her hair is artfully shaped to resemble the calyx of the wax apple, seamlessly blending with the fruit itself. A soft, pearly, warm aura emanates from her, enveloping her and imbuing the scene with a sacred and dreamy aura. The background is a tranquil lake shrouded in morning mist. Soft, diffused light from the sky penetrates the fog, intertwining with the spirit's own radiance, casting a faint, blurred shadow on the lotus leaf. A large, crystalline drop of water slowly glides down the graceful curve of her side, subtly distorting the colors beneath it as it moves, leaving a fleeting trail of moisture before silently landing on the surface of a lotus leaf, creating a circle of tiny, perfectly formed ripples. The entire image is filled with exquisite detail, from the translucent texture of the elf's skin to the velvety texture of the lotus leaf. Each frame resembles a meticulously crafted illustration, creating an atmosphere of profound tranquility and pure poetry.", + "seed": 24635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_12/index_904_seed_24635_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03873172737658024, + "vlm_logprob": -0.03873172737658024, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03873172737658024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03960505127906799 + ], + "opposite_logprob_tokens": [ + -3.539605140686035 + ], + "logprob_reward": -0.03960505127906799, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03960505127906799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2873399555683136, + 0.07111803442239761, + 0.05687876045703888, + 0.10327593982219696, + 0.06261100620031357, + 0.06912422925233841, + 0.0909019336104393, + 0.11224041134119034, + 0.08463960140943527, + 0.061870116740465164 + ], + "credit_map": [ + 0.03505712002515793, + 0.03326280787587166, + 0.03616233170032501, + 0.04061449319124222, + 0.038929302245378494, + 0.04108178988099098, + 0.031216522678732872, + 0.031015580520033836, + 0.008351508527994156, + 0.009529106318950653, + 0.0076060486026108265, + 0.008837815374135971, + 0.010611381381750107, + 0.007894093170762062, + 0.008171661756932735, + 0.010116422548890114, + 0.005104692652821541, + 0.007164558861404657, + 0.007337320130318403, + 0.00859400536864996, + 0.007362029515206814, + 0.007091606501489878, + 0.007458321750164032, + 0.006766229402273893, + 0.010424437001347542, + 0.008153749629855156, + 0.010737916454672813, + 0.010408051311969757, + 0.009777610190212727, + 0.013936026021838188, + 0.015619820915162563, + 0.02421833574771881, + 0.005882893223315477, + 0.0069199311546981335, + 0.004306924529373646, + 0.005165138281881809, + 0.010739032179117203, + 0.013050397858023643, + 0.007566914428025484, + 0.008979780599474907, + 0.00990122091025114, + 0.011330770328640938, + 0.0056933145970106125, + 0.009938173927366734, + 0.009708181954920292, + 0.01036404725164175, + 0.00506163714453578, + 0.007126886863261461, + 0.010288420133292675, + 0.012321689166128635, + 0.007907552644610405, + 0.014746041968464851, + 0.012875131331384182, + 0.012455272488296032, + 0.007596041541546583, + 0.012711786665022373, + 0.01295305136591196, + 0.009577682241797447, + 0.01060295570641756, + 0.010228460654616356, + 0.013979746960103512, + 0.023576177656650543, + 0.011226652190089226, + 0.020095689222216606, + 0.010485871694982052, + 0.01116194948554039, + 0.007752296980470419, + 0.011317173019051552, + 0.011449531652033329, + 0.013906575739383698, + 0.007978738285601139, + 0.010587461292743683, + 0.007686784956604242, + 0.008385836146771908, + 0.006505015306174755, + 0.008979137055575848, + 0.008110501803457737, + 0.010270301252603531, + 0.004807921592146158, + 0.007124613504856825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.03960505127906799 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032100360840559006 + ], + "opposite_logprob_tokens": [ + -5.753210067749023 + ], + "logprob_reward": -0.0032100360840559006, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032100360840559006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3144548535346985, + 0.12279267609119415, + 0.0663132593035698, + 0.12681883573532104, + 0.06030606850981712, + 0.0540660060942173, + 0.053784001618623734, + 0.07856473326683044, + 0.07570885121822357, + 0.04719071835279465 + ], + "credit_map": [ + 0.03225146234035492, + 0.04806489124894142, + 0.02546882815659046, + 0.04541298374533653, + 0.046822283416986465, + 0.05075211077928543, + 0.031166190281510353, + 0.03451608493924141, + 0.013706421479582787, + 0.019395537674427032, + 0.012119144201278687, + 0.017865708097815514, + 0.015171892009675503, + 0.015215128660202026, + 0.013008607551455498, + 0.016310235485434532, + 0.0071465810760855675, + 0.011715295724570751, + 0.008036735467612743, + 0.008486345410346985, + 0.009087320417165756, + 0.0072242519818246365, + 0.006188052706420422, + 0.008428677916526794, + 0.009558046236634254, + 0.012551429681479931, + 0.012266963720321655, + 0.01280795969069004, + 0.013922834768891335, + 0.016651799902319908, + 0.021422870457172394, + 0.02763693407177925, + 0.005961396731436253, + 0.0048818704672157764, + 0.009411387145519257, + 0.00628425320610404, + 0.0055681332014501095, + 0.007709375116974115, + 0.009932480752468109, + 0.010557170957326889, + 0.004639028571546078, + 0.006660113111138344, + 0.004828073550015688, + 0.0071438574232161045, + 0.006448309402912855, + 0.00851711630821228, + 0.006089632865041494, + 0.009739875793457031, + 0.006926287431269884, + 0.006749358493834734, + 0.005729269701987505, + 0.006017394829541445, + 0.006697183474898338, + 0.01002813782542944, + 0.003956293221563101, + 0.007680077105760574, + 0.009984114207327366, + 0.010148759000003338, + 0.005979870446026325, + 0.00900842435657978, + 0.011429385282099247, + 0.014960380271077156, + 0.006901606451719999, + 0.010152184404432774, + 0.010107985697686672, + 0.009344611316919327, + 0.007206518203020096, + 0.010933414101600647, + 0.010096602141857147, + 0.009987134486436844, + 0.007730392273515463, + 0.010302196256816387, + 0.007021607365459204, + 0.005861101672053337, + 0.005023217294365168, + 0.005919917486608028, + 0.005853399634361267, + 0.006517808884382248, + 0.004868441727012396, + 0.006125224754214287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0032100360840559006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206217780709267 + ], + "opposite_logprob_tokens": [ + -2.8120622634887695 + ], + "logprob_reward": -0.06206217780709267, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206217780709267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3388453722000122, + 0.09197845309972763, + 0.04461465775966644, + 0.05609771981835365, + 0.06655003875494003, + 0.055776193737983704, + 0.08716004341840744, + 0.058298077434301376, + 0.11221057921648026, + 0.08846885710954666 + ], + "credit_map": [ + 0.03620634600520134, + 0.025125889107584953, + 0.027797246351838112, + 0.0339793860912323, + 0.034434396773576736, + 0.05239633470773697, + 0.07275093346834183, + 0.05615485832095146, + 0.008321288973093033, + 0.010051597841084003, + 0.008008800446987152, + 0.010455229319632053, + 0.012159345671534538, + 0.012904015369713306, + 0.014694151468575, + 0.015384032391011715, + 0.004193670116364956, + 0.004977404605597258, + 0.003954591229557991, + 0.0060293106362223625, + 0.00615676399320364, + 0.006925382651388645, + 0.005170559510588646, + 0.007206977345049381, + 0.005591735243797302, + 0.005129462108016014, + 0.005601771641522646, + 0.007300001103430986, + 0.004512253683060408, + 0.009670095518231392, + 0.008432842791080475, + 0.009859560988843441, + 0.006472320761531591, + 0.005628589540719986, + 0.00894287321716547, + 0.007923581637442112, + 0.00680188974365592, + 0.009273828007280827, + 0.009882966987788677, + 0.011623991653323174, + 0.007141308858990669, + 0.0052438294515013695, + 0.00679527223110199, + 0.006148179993033409, + 0.004706376697868109, + 0.008151211775839329, + 0.0066756815649569035, + 0.010914338752627373, + 0.010364270769059658, + 0.009894429706037045, + 0.011252674274146557, + 0.008424092084169388, + 0.007630941923707724, + 0.012378489598631859, + 0.010098576545715332, + 0.017116572707891464, + 0.006618737708777189, + 0.006332258228212595, + 0.007342629134654999, + 0.006811851169914007, + 0.00491031538695097, + 0.0077269794419407845, + 0.0076015484519302845, + 0.01095376256853342, + 0.014622263610363007, + 0.011515343561768532, + 0.012917573563754559, + 0.011269506067037582, + 0.009057101793587208, + 0.016120098531246185, + 0.012669425457715988, + 0.02403927966952324, + 0.013333503156900406, + 0.009924628771841526, + 0.008365155197679996, + 0.010705683380365372, + 0.0077634756453335285, + 0.014254868030548096, + 0.009686782024800777, + 0.014434774406254292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.06206217780709267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899748533964157 + ], + "opposite_logprob_tokens": [ + -2.5789973735809326 + ], + "logprob_reward": -0.07899748533964157, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899748533964157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37948521971702576, + 0.07119438797235489, + 0.06948664039373398, + 0.09528598934412003, + 0.089058518409729, + 0.03706476837396622, + 0.04923352599143982, + 0.06016634777188301, + 0.044538989663124084, + 0.10448556393384933 + ], + "credit_map": [ + 0.0392959788441658, + 0.0522022470831871, + 0.03554539382457733, + 0.05218673497438431, + 0.06130828708410263, + 0.05476309359073639, + 0.04014447331428528, + 0.04403906315565109, + 0.008137173019349575, + 0.012712235562503338, + 0.006307275965809822, + 0.010100966319441795, + 0.011104951612651348, + 0.008820165880024433, + 0.006634629797190428, + 0.007376990281045437, + 0.007756520062685013, + 0.007744087837636471, + 0.010790837928652763, + 0.006707875523716211, + 0.006398439407348633, + 0.009851807728409767, + 0.009476260282099247, + 0.010760818608105183, + 0.008666490204632282, + 0.0087371114641428, + 0.009886744432151318, + 0.010454977862536907, + 0.010909331031143665, + 0.012943189591169357, + 0.01533152163028717, + 0.01835663616657257, + 0.01006864570081234, + 0.007980790920555592, + 0.007956008426845074, + 0.008547713048756123, + 0.011366557329893112, + 0.012483150698244572, + 0.015751656144857407, + 0.014903995208442211, + 0.00530065456405282, + 0.003901965683326125, + 0.00429029343649745, + 0.005371459759771824, + 0.004126694053411484, + 0.004585184622555971, + 0.0033608118537813425, + 0.006127707660198212, + 0.005574783775955439, + 0.004167947452515364, + 0.0044623007997870445, + 0.00775200454518199, + 0.006457984913140535, + 0.007739424239844084, + 0.004783256910741329, + 0.00829582754522562, + 0.00725035835057497, + 0.005440175533294678, + 0.004822436720132828, + 0.006856685969978571, + 0.006154701113700867, + 0.014379854314029217, + 0.00572547409683466, + 0.009536669589579105, + 0.0056381928734481335, + 0.005906099919229746, + 0.0037680985406041145, + 0.005140454508364201, + 0.007023321930319071, + 0.006073717027902603, + 0.004242133814841509, + 0.006746972445398569, + 0.012777963653206825, + 0.013690483756363392, + 0.010388489812612534, + 0.013453788124024868, + 0.013423525728285313, + 0.014885587617754936, + 0.009647680446505547, + 0.01621805690228939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.07899748533964157 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a vertical medium shot taken from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main character a spirit resembling a plump wax apple with skin gradating from ivory white to ruby red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on a vast emerald lotus leaf with distinct veins in a tranquil lake shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large crystalline drop of water slowly glide down the character's side and land on a lotus leaf creating ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's hair artfully shaped to resemble the calyx of a wax apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16127340495586395, + 0.11907459050416946, + 0.11007406562566757, + 0.210018590092659, + 0.11261541396379471, + 0.05536910146474838, + 0.06200920045375824, + 0.06259658187627792, + 0.03613226115703583, + 0.07083682715892792 + ], + "credit_map": [ + 0.017514660954475403, + 0.0265390332788229, + 0.009857541881501675, + 0.024769369512796402, + 0.022165412083268166, + 0.02171807922422886, + 0.018522348254919052, + 0.02018696814775467, + 0.013190532103180885, + 0.018952051177620888, + 0.011070078238844872, + 0.017014937475323677, + 0.012438544072210789, + 0.014080692082643509, + 0.015157114714384079, + 0.01717063970863819, + 0.011685917153954506, + 0.01714622974395752, + 0.01824764534831047, + 0.015429126098752022, + 0.013541285879909992, + 0.012981285341084003, + 0.008743788115680218, + 0.012298787012696266, + 0.0178376417607069, + 0.016303757205605507, + 0.02260168269276619, + 0.02000325173139572, + 0.019290437921881676, + 0.029490258544683456, + 0.03719939664006233, + 0.047292158007621765, + 0.009607917629182339, + 0.007938590832054615, + 0.010902090929448605, + 0.009958631359040737, + 0.011601361446082592, + 0.01562538556754589, + 0.022771582007408142, + 0.024209851399064064, + 0.004592968616634607, + 0.005295188166201115, + 0.004757785703986883, + 0.0063585080206394196, + 0.0051139118149876595, + 0.007794464007019997, + 0.007984239608049393, + 0.013472036458551884, + 0.007167029194533825, + 0.0041175526566803455, + 0.006366298533976078, + 0.005063281394541264, + 0.005251196213066578, + 0.015757868066430092, + 0.006293410435318947, + 0.011992563493549824, + 0.007984787225723267, + 0.00812850333750248, + 0.004525108728557825, + 0.01225331425666809, + 0.008349515497684479, + 0.008914265781641006, + 0.005292899440973997, + 0.0071481866762042046, + 0.00472834100946784, + 0.003435231279581785, + 0.0038945560809224844, + 0.004960705991834402, + 0.0045986720360815525, + 0.004699917044490576, + 0.003797117853537202, + 0.006017718929797411, + 0.008127491921186447, + 0.007020033895969391, + 0.007788319606333971, + 0.009320965968072414, + 0.006553473882377148, + 0.009206821210682392, + 0.006965005770325661, + 0.015854716300964355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00978388637304306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_tensor([2062, 2062])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_tensor([2062, 2062])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f7feca8c39d067a3a790267133262e97e9241073 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_tensor([2062, 2062])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "grand, long, static shot places us on a desolate mountaintop shrouded in a torrential downpour. Shot in a stunning slow motion, the world seems frozen in time. At the center of the frame, a red-clad swordsman stands alone on the rugged, slippery rocks. He wears a rain-soaked crimson suit, the damp fabric clinging to his muscular frame, outlining his defined muscles. His long, dark hair, dampened by the rain, tangled strands plastered to his forehead and cheeks, highlighting his hawk-like eyes and furrowed brow, a fierce, untouchable aura woven between them. With a calm and resolute gesture, he slowly raises his arm, draws a gleaming ancient sword from his waist, and holds it firmly across his chest. During this process, every massive raindrop falling from the sky is clearly visible, like suspended transparent crystals, reflecting a faint light against the gray sky. Some slowly streak across the camera, while others collide and explode on his resolute face, scarlet clothes, and the mirror-like blade of his sword, splashing out circles of miniature, crystal-clear water crowns. The entire scene is gloomy, with only that hint of red, like a burning flame, forming a sharp contrast with the cold rain and gray rocks, creating a lonely, tense, and epic atmosphere of fate.", + "seed": 32649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_2062_seed_32649_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1210518459789455, + "vlm_logprob": -0.1210518459789455, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.1210518459789455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24918974936008453 + ], + "opposite_logprob_tokens": [ + -1.6241897344589233 + ], + "logprob_reward": -0.24918974936008453, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24918974936008453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24793340265750885, + 0.20727482438087463, + 0.033952344208955765, + 0.03995748236775398, + 0.12598760426044464, + 0.052484527230262756, + 0.08396553248167038, + 0.03344261273741722, + 0.060285910964012146, + 0.11471572518348694 + ], + "credit_map": [ + 0.03354797884821892, + 0.029608631506562233, + 0.04098935052752495, + 0.028477922081947327, + 0.02518521249294281, + 0.03170805796980858, + 0.028020549565553665, + 0.030395695939660072, + 0.036721061915159225, + 0.01746838539838791, + 0.037121471017599106, + 0.012676111422479153, + 0.02093547210097313, + 0.03444612771272659, + 0.017064394429326057, + 0.030841805040836334, + 0.005476900842040777, + 0.002993162954226136, + 0.006102718412876129, + 0.0018677664920687675, + 0.00476084416732192, + 0.006473831832408905, + 0.0030117796268314123, + 0.003265340579673648, + 0.0068357461132109165, + 0.002600644947960973, + 0.0054932101629674435, + 0.0023277972359210253, + 0.005385828670114279, + 0.003838044125586748, + 0.0038354673888534307, + 0.00964074395596981, + 0.0252820011228323, + 0.011082694865763187, + 0.020727217197418213, + 0.006402722559869289, + 0.014122850261628628, + 0.019161932170391083, + 0.00981489010155201, + 0.019393296912312508, + 0.008700666949152946, + 0.0062032341957092285, + 0.007093339692801237, + 0.0036761860828846693, + 0.007995801977813244, + 0.007576765492558479, + 0.003663646522909403, + 0.007574883289635181, + 0.015032276511192322, + 0.0070409891195595264, + 0.009090665727853775, + 0.005765774752944708, + 0.009843438863754272, + 0.009356923401355743, + 0.007632492110133171, + 0.020202968269586563, + 0.005784772802144289, + 0.0029197006952017546, + 0.0034983521327376366, + 0.003930894657969475, + 0.0042143031023442745, + 0.0025659685488790274, + 0.0035838705953210592, + 0.0069447532296180725, + 0.008211650885641575, + 0.006006893236190081, + 0.008131562732160091, + 0.008027770556509495, + 0.008471330627799034, + 0.006089175119996071, + 0.007154267746955156, + 0.008193258196115494, + 0.01826101914048195, + 0.013212730176746845, + 0.011223705485463142, + 0.009401955641806126, + 0.01900247484445572, + 0.014294425956904888, + 0.010723702609539032, + 0.018595702946186066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.24918974936008453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076702362857759 + ], + "opposite_logprob_tokens": [ + -4.8826704025268555 + ], + "logprob_reward": -0.0076702362857759, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076702362857759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4738486707210541, + 0.05771629139780998, + 0.09737979620695114, + 0.08689890056848526, + 0.09276826679706573, + 0.035836122930049896, + 0.03951634094119072, + 0.02844637632369995, + 0.04680928960442543, + 0.040779873728752136 + ], + "credit_map": [ + 0.03015870414674282, + 0.01784089207649231, + 0.03511638566851616, + 0.01658730022609234, + 0.021250246092677116, + 0.1717623919248581, + 0.02502373792231083, + 0.15610907971858978, + 0.00959891639649868, + 0.0058639636263251305, + 0.00892801582813263, + 0.004492603242397308, + 0.005271408706903458, + 0.009507662616670132, + 0.0062268259935081005, + 0.007826905697584152, + 0.011144928634166718, + 0.007372090592980385, + 0.013890864327549934, + 0.004418340977281332, + 0.012559546157717705, + 0.022295381873846054, + 0.00854537170380354, + 0.01715327985584736, + 0.01454325020313263, + 0.0054545216262340546, + 0.01343314629048109, + 0.0038825611118227243, + 0.008442671969532967, + 0.008771021850407124, + 0.0083819180727005, + 0.02398982085287571, + 0.014827496372163296, + 0.008087608963251114, + 0.020632443949580193, + 0.004729022271931171, + 0.010641840286552906, + 0.009507842361927032, + 0.005566086620092392, + 0.018775932490825653, + 0.007084473501890898, + 0.0038106669671833515, + 0.0040165432728827, + 0.0030797102954238653, + 0.005610877647995949, + 0.003387563629075885, + 0.0029410189017653465, + 0.005905274301767349, + 0.0060657006688416, + 0.003633492859080434, + 0.0029785064980387688, + 0.0024155585560947657, + 0.009500980377197266, + 0.006451116409152746, + 0.002743061864748597, + 0.005727926269173622, + 0.005014814902096987, + 0.003829293418675661, + 0.004696417134255171, + 0.003162989392876625, + 0.0037249987944960594, + 0.002419769298285246, + 0.0027925975155085325, + 0.002805497730150819, + 0.007853088900446892, + 0.004657156299799681, + 0.005082025658339262, + 0.006633454002439976, + 0.005034982692450285, + 0.005120903719216585, + 0.0037408405914902687, + 0.008686847984790802, + 0.007735558785498142, + 0.003658673260360956, + 0.0037427926436066628, + 0.003335970686748624, + 0.006097086239606142, + 0.006639162544161081, + 0.0030518423300236464, + 0.006518789567053318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0076702362857759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906377851963043 + ], + "opposite_logprob_tokens": [ + -2.464063882827759 + ], + "logprob_reward": -0.08906377851963043, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906377851963043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264861136674881, + 0.04828233644366264, + 0.0784599557518959, + 0.19147033989429474, + 0.03582547605037689, + 0.028501685708761215, + 0.04825235903263092, + 0.0725880116224289, + 0.15115956962108612, + 0.08059919625520706 + ], + "credit_map": [ + 0.015365923754870892, + 0.01021665707230568, + 0.023137694224715233, + 0.011682347394526005, + 0.012393319979310036, + 0.09523594379425049, + 0.01645204797387123, + 0.08037719130516052, + 0.004987665917724371, + 0.0052407970651984215, + 0.008970656432211399, + 0.0035678872372955084, + 0.006160138640552759, + 0.007803278975188732, + 0.004927680362015963, + 0.006624231114983559, + 0.013577270321547985, + 0.005596781149506569, + 0.012834586203098297, + 0.004925164394080639, + 0.009230434894561768, + 0.01146224420517683, + 0.007015973795205355, + 0.013817504979670048, + 0.03255750611424446, + 0.01166517473757267, + 0.02965620532631874, + 0.00838439166545868, + 0.017246490344405174, + 0.01951020583510399, + 0.019190166145563126, + 0.053260188549757004, + 0.004419532138854265, + 0.003391608130186796, + 0.0040976605378091335, + 0.0030372722540050745, + 0.005127194337546825, + 0.007410907652229071, + 0.0037741728592664003, + 0.00456712581217289, + 0.005046268925070763, + 0.0022455353755503893, + 0.0036194813437759876, + 0.0020924261771142483, + 0.0037517070304602385, + 0.002852166537195444, + 0.0030608216766268015, + 0.005833278875797987, + 0.011175273917615414, + 0.00502345897257328, + 0.008230920881032944, + 0.0044637941755354404, + 0.007003293372690678, + 0.0034644627012312412, + 0.0033935504034161568, + 0.005497602745890617, + 0.009109214879572392, + 0.006113800220191479, + 0.007729383651167154, + 0.010059473104774952, + 0.011896704323589802, + 0.007312838453799486, + 0.00968900416046381, + 0.010677590034902096, + 0.02340923622250557, + 0.017090925946831703, + 0.015828529372811317, + 0.014504997991025448, + 0.018422573804855347, + 0.016248514875769615, + 0.01592131517827511, + 0.02973347343504429, + 0.012440232560038567, + 0.009310655295848846, + 0.011407415382564068, + 0.008896451443433762, + 0.013731208629906178, + 0.007588726002722979, + 0.008297588676214218, + 0.008926916867494583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.08906377851963043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371020406484604 + ], + "opposite_logprob_tokens": [ + -3.408710241317749 + ], + "logprob_reward": -0.03371020406484604, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371020406484604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6079998016357422, + 0.15464581549167633, + 0.014231829904019833, + 0.06330692023038864, + 0.03212398663163185, + 0.027499664574861526, + 0.013742504641413689, + 0.03458564355969429, + 0.032106246799230576, + 0.01975771225988865 + ], + "credit_map": [ + 0.03974564000964165, + 0.022160371765494347, + 0.052480876445770264, + 0.02343665435910225, + 0.027486179023981094, + 0.2166127860546112, + 0.031196502968668938, + 0.1948806792497635, + 0.024758175015449524, + 0.014103634282946587, + 0.029948875308036804, + 0.01414361409842968, + 0.014269763603806496, + 0.018589012324810028, + 0.013622578233480453, + 0.025210142135620117, + 0.0019834651611745358, + 0.001368574914522469, + 0.002529565943405032, + 0.0010284589370712638, + 0.002176305977627635, + 0.0015181644121184945, + 0.0013504117960110307, + 0.002276881132274866, + 0.010716547258198261, + 0.004091428127139807, + 0.010039083659648895, + 0.002927567111328244, + 0.005428109783679247, + 0.006534055806696415, + 0.005982537753880024, + 0.017587585374712944, + 0.004871909040957689, + 0.003201827174052596, + 0.004813095089048147, + 0.0016822601901367307, + 0.0038355644792318344, + 0.005566785577684641, + 0.002153201727196574, + 0.005999342072755098, + 0.004279442597180605, + 0.0019050880800932646, + 0.0034051144029945135, + 0.0015494110994040966, + 0.003861145582050085, + 0.0044778380542993546, + 0.002236882457509637, + 0.005784739274531603, + 0.0034264782443642616, + 0.0015537380240857601, + 0.0019148709252476692, + 0.0011109061306342483, + 0.0017991448985412717, + 0.001258120290003717, + 0.0008948021568357944, + 0.00178444292396307, + 0.007970980368554592, + 0.0029458345379680395, + 0.003999064210802317, + 0.004130689427256584, + 0.0039732856675982475, + 0.003279173281043768, + 0.0034117563627660275, + 0.0048748585395514965, + 0.004975603893399239, + 0.003882960882037878, + 0.003410952165722847, + 0.004130202811211348, + 0.0038088862784206867, + 0.0032367627136409283, + 0.0034860027953982353, + 0.005174870602786541, + 0.0034155254252254963, + 0.002284602029249072, + 0.002153711160644889, + 0.001721857232041657, + 0.0029127539601176977, + 0.0027291581500321627, + 0.0014978621620684862, + 0.003042238997295499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03371020406484604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562526166439056 + ], + "opposite_logprob_tokens": [ + -1.6006252765655518 + ], + "logprob_reward": -0.22562526166439056, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562526166439056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.309060275554657, + 0.08064314723014832, + 0.08838260918855667, + 0.2271600067615509, + 0.0370417945086956, + 0.0590486116707325, + 0.026318781077861786, + 0.06026967987418175, + 0.05941418185830116, + 0.052660949528217316 + ], + "credit_map": [ + 0.017377063632011414, + 0.011853904463350773, + 0.02362373284995556, + 0.012266425415873528, + 0.015819896012544632, + 0.11205030232667923, + 0.0175346489995718, + 0.09853428602218628, + 0.01111258752644062, + 0.007552041206508875, + 0.01619081199169159, + 0.006168832071125507, + 0.007528417743742466, + 0.012882525101304054, + 0.006524837110191584, + 0.012683093547821045, + 0.008306745439767838, + 0.005708799231797457, + 0.012794923968613148, + 0.005127251613885164, + 0.009713138453662395, + 0.018080756068229675, + 0.008836764842271805, + 0.019814228639006615, + 0.03633331134915352, + 0.014325768686830997, + 0.036802392452955246, + 0.010202687233686447, + 0.020063545554876328, + 0.023085657507181168, + 0.02248450741171837, + 0.063862144947052, + 0.007301415782421827, + 0.0030937823466956615, + 0.005904327612370253, + 0.002143570687621832, + 0.003960566129535437, + 0.003532842267304659, + 0.003171932650730014, + 0.007933358661830425, + 0.008200637064874172, + 0.006592171266674995, + 0.005590766668319702, + 0.006661472376435995, + 0.009954506531357765, + 0.004724678583443165, + 0.01013230625540018, + 0.007192071992903948, + 0.0034103214275091887, + 0.003097917651757598, + 0.002571428893133998, + 0.0033978300634771585, + 0.004171473439782858, + 0.002324267290532589, + 0.0030108375940471888, + 0.004334703087806702, + 0.007422609720379114, + 0.005089647136628628, + 0.008357629179954529, + 0.008441824465990067, + 0.010786768980324268, + 0.008082189597189426, + 0.006299303378909826, + 0.005789705086499453, + 0.011946147307753563, + 0.007199390325695276, + 0.0056955269537866116, + 0.005797254852950573, + 0.0069389138370752335, + 0.004927671514451504, + 0.006877334322780371, + 0.010031946003437042, + 0.007725371979176998, + 0.007822121493518353, + 0.006090486887842417, + 0.004887285176664591, + 0.007770676165819168, + 0.005839421413838863, + 0.00508112832903862, + 0.00744446087628603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.22562526166439056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand, long, static shot places us on a desolate mountaintop shrouded in a torrential downpour. Shot in a stunning slow motion, the world seems frozen in time. At the center of the frame, a red-clad swordsman stands alone on the rugged, slippery rocks. He wears a rain-soaked crimson suit, the damp fabric clinging to his muscular frame, outlining his defined muscles. His long, dark hair, dampened by the rain, tangled strands plastered to his forehead and cheeks, highlighting his hawk-like eyes and furrowed brow, a fierce, untouchable aura woven between them. With a calm and resolute gesture, he slowly raises his arm, draws a gleaming ancient sword from his waist, and holds it firmly across his chest. During this process, every massive raindrop falling from the sky is clearly visible, like suspended transparent crystals, reflecting a faint light against the gray sky. Some slowly streak across the camera, while others collide and explode on his resolute face, scarlet clothes, and the mirror-like blade of his sword, splashing out circles of miniature, crystal-clear water crowns. The entire scene is gloomy, with only that hint of red, like a burning flame, forming a sharp contrast with the cold rain and gray rocks, creating a lonely, tense, and epic atmosphere of fate.", + "seed": 32650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_2062_seed_32650_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1097621601074934, + "vlm_logprob": -0.1097621601074934, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.1097621601074934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05542391911149025 + ], + "opposite_logprob_tokens": [ + -3.1804239749908447 + ], + "logprob_reward": -0.05542391911149025, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05542391911149025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2984433174133301, + 0.20878510177135468, + 0.029137803241610527, + 0.23834030330181122, + 0.032567791640758514, + 0.08507907390594482, + 0.030621137470006943, + 0.02002747729420662, + 0.0102016506716609, + 0.04679636284708977 + ], + "credit_map": [ + 0.031226420775055885, + 0.0281870998442173, + 0.0461200475692749, + 0.04557177796959877, + 0.03519385680556297, + 0.04245106503367424, + 0.03976861760020256, + 0.029924437403678894, + 0.031925808638334274, + 0.05288269743323326, + 0.03586835041642189, + 0.01145122665911913, + 0.015379783697426319, + 0.028116459026932716, + 0.011144120246171951, + 0.022016650065779686, + 0.005842468701303005, + 0.007358113769441843, + 0.003727350849658251, + 0.001964653842151165, + 0.0019843634217977524, + 0.0038202719297260046, + 0.0016338317655026913, + 0.002806749427691102, + 0.0407131053507328, + 0.06670378893613815, + 0.05605832114815712, + 0.011393709108233452, + 0.012288833037018776, + 0.026887932792305946, + 0.014198335818946362, + 0.010096287354826927, + 0.004069107118993998, + 0.00305793434381485, + 0.0064669460989534855, + 0.0027778921648859978, + 0.0032600576523691416, + 0.003953519742935896, + 0.004910481162369251, + 0.004071854054927826, + 0.007033031899482012, + 0.0071562123484909534, + 0.019537122920155525, + 0.00953996554017067, + 0.015029648318886757, + 0.010494887828826904, + 0.010189889930188656, + 0.00609831465408206, + 0.003997066058218479, + 0.0025114179588854313, + 0.0076005347073078156, + 0.0031722031999379396, + 0.003670627484098077, + 0.0039926208555698395, + 0.002345036016777158, + 0.003331630490720272, + 0.002059439430013299, + 0.0018844179576262832, + 0.003050171537324786, + 0.002051394432783127, + 0.002172761596739292, + 0.003016527509316802, + 0.0018919457215815783, + 0.00390081899240613, + 0.0012928115902468562, + 0.0013158483197912574, + 0.0024440253619104624, + 0.0013441952178254724, + 0.0012552787084132433, + 0.001078094937838614, + 0.0008142231963574886, + 0.0006571743288077414, + 0.008014261722564697, + 0.0047070542350411415, + 0.009268763475120068, + 0.0037954410072416067, + 0.003992154262959957, + 0.010130542330443859, + 0.0026423344388604164, + 0.004245806019753218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.05542391911149025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009815169498324394 + ], + "opposite_logprob_tokens": [ + -4.634815216064453 + ], + "logprob_reward": -0.009815169498324394, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009815169498324394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2580917775630951, + 0.11568930000066757, + 0.06114265322685242, + 0.0767584964632988, + 0.08289281278848648, + 0.04575997591018677, + 0.09576546400785446, + 0.09658792614936829, + 0.026803741231560707, + 0.14050786197185516 + ], + "credit_map": [ + 0.022205205634236336, + 0.023197876289486885, + 0.03556469827890396, + 0.02133406698703766, + 0.06782014667987823, + 0.031442150473594666, + 0.027198180556297302, + 0.02932947687804699, + 0.023261642083525658, + 0.02920757234096527, + 0.010593838058412075, + 0.004715998191386461, + 0.014376260340213776, + 0.0159431342035532, + 0.005959393922239542, + 0.011631462723016739, + 0.005893093533813953, + 0.008104577660560608, + 0.0073376731015741825, + 0.005017613526433706, + 0.00791222508996725, + 0.010973747819662094, + 0.0067361886613070965, + 0.009167534299194813, + 0.014559087343513966, + 0.013911855407059193, + 0.02096669375896454, + 0.004557599779218435, + 0.0047840578481554985, + 0.006277957931160927, + 0.006070880685001612, + 0.005630356725305319, + 0.009038046002388, + 0.0065688686445355415, + 0.015384991653263569, + 0.008103965781629086, + 0.013686362653970718, + 0.010244714096188545, + 0.009253895841538906, + 0.010611958801746368, + 0.0034060149919241667, + 0.003151969751343131, + 0.012969386763870716, + 0.00530533492565155, + 0.006880939472466707, + 0.004958157427608967, + 0.005243171937763691, + 0.0038450018037110567, + 0.007113755214959383, + 0.007421783171594143, + 0.011901912279427052, + 0.016330622136592865, + 0.01433977484703064, + 0.012047597207129002, + 0.01665978692471981, + 0.00995023362338543, + 0.00847054086625576, + 0.01028294488787651, + 0.0136598264798522, + 0.009646744467318058, + 0.011010165326297283, + 0.015022623352706432, + 0.009498797357082367, + 0.018996287137269974, + 0.0028192359022796154, + 0.0029328002128750086, + 0.004295152146369219, + 0.00313210510648787, + 0.004931899718940258, + 0.0026609590277075768, + 0.003380333539098501, + 0.002651255577802658, + 0.019348157569766045, + 0.012270369566977024, + 0.024809332564473152, + 0.01299168448895216, + 0.014624656178057194, + 0.03032532148063183, + 0.01597527414560318, + 0.010163058526813984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.009815169498324394 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215864181518555 + ], + "opposite_logprob_tokens": [ + -1.5021586418151855 + ], + "logprob_reward": -0.25215864181518555, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215864181518555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47036296129226685, + 0.09247174113988876, + 0.06733663380146027, + 0.06528952717781067, + 0.04630083218216896, + 0.030549274757504463, + 0.0606001541018486, + 0.08068375289440155, + 0.026226073503494263, + 0.06017911061644554 + ], + "credit_map": [ + 0.03161188215017319, + 0.03804238885641098, + 0.06070048362016678, + 0.08877741545438766, + 0.0803288146853447, + 0.04483896493911743, + 0.07573440670967102, + 0.050328563898801804, + 0.021744471043348312, + 0.02967938967049122, + 0.013626246713101864, + 0.004393329843878746, + 0.006366192828863859, + 0.00701222475618124, + 0.0046605332754552364, + 0.004989343229681253, + 0.012376177124679089, + 0.01683998852968216, + 0.009748201817274094, + 0.00438978336751461, + 0.006061243358999491, + 0.00769878039136529, + 0.004375077318400145, + 0.005847375839948654, + 0.013604158535599709, + 0.012520693242549896, + 0.014942330308258533, + 0.00314415548928082, + 0.00598340667784214, + 0.00466790609061718, + 0.005001765210181475, + 0.005425100214779377, + 0.004134793765842915, + 0.0041682906448841095, + 0.006903013214468956, + 0.004662476480007172, + 0.009837199933826923, + 0.0048018391244113445, + 0.005656043998897076, + 0.00613716896623373, + 0.0028804470784962177, + 0.0028222007676959038, + 0.005335958208888769, + 0.003294284688308835, + 0.005097239278256893, + 0.003891716478392482, + 0.0036458191461861134, + 0.0035816049203276634, + 0.00512774009257555, + 0.005946927238255739, + 0.008017836138606071, + 0.008345761336386204, + 0.011137767694890499, + 0.007442882284522057, + 0.008006039075553417, + 0.006575194653123617, + 0.007301468402147293, + 0.008341831155121326, + 0.01092447992414236, + 0.007344778161495924, + 0.007765092886984348, + 0.01302389893680811, + 0.007628095336258411, + 0.018354099243879318, + 0.003359683323651552, + 0.002503699855878949, + 0.004644532222300768, + 0.0023900424130260944, + 0.005293305963277817, + 0.0031100211199373007, + 0.002708521904423833, + 0.002216263208538294, + 0.006386748980730772, + 0.004428197164088488, + 0.007775450125336647, + 0.0070203556679189205, + 0.008344926871359348, + 0.008426813408732414, + 0.012010088190436363, + 0.005786525551229715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.25215864181518555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801908880472183 + ], + "opposite_logprob_tokens": [ + -3.529801845550537 + ], + "logprob_reward": -0.029801908880472183, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801908880472183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2749764323234558, + 0.07600908726453781, + 0.06851861625909805, + 0.3344114422798157, + 0.06056518480181694, + 0.08927758038043976, + 0.032199256122112274, + 0.017324386164546013, + 0.010404913686215878, + 0.03631310909986496 + ], + "credit_map": [ + 0.01957838423550129, + 0.021951228380203247, + 0.03877168521285057, + 0.05656919628381729, + 0.030497459694743156, + 0.027197450399398804, + 0.04936918243765831, + 0.031041834503412247, + 0.007604567799717188, + 0.01389715913683176, + 0.012255094945430756, + 0.004827848169952631, + 0.010745804756879807, + 0.011665957048535347, + 0.004976850003004074, + 0.01003580167889595, + 0.016229549422860146, + 0.020881641656160355, + 0.00937403179705143, + 0.00418076291680336, + 0.004445013590157032, + 0.005415631923824549, + 0.003394505474716425, + 0.004597481340169907, + 0.05915915220975876, + 0.09332442283630371, + 0.07629824429750443, + 0.016288653016090393, + 0.019670257344841957, + 0.036587949842214584, + 0.018973954021930695, + 0.014108785428106785, + 0.008565550670027733, + 0.005042044445872307, + 0.01921776682138443, + 0.004353180527687073, + 0.005392668303102255, + 0.008554727770388126, + 0.004347234033048153, + 0.005092010833323002, + 0.00839226134121418, + 0.007893872447311878, + 0.01803750731050968, + 0.010073897428810596, + 0.014790360815823078, + 0.009639120660722256, + 0.01006094180047512, + 0.010389617644250393, + 0.003560838522389531, + 0.002887657843530178, + 0.006306438706815243, + 0.0032017447520047426, + 0.003954887390136719, + 0.004629985895007849, + 0.0029004481621086597, + 0.004757254850119352, + 0.0018045847536996007, + 0.001527532353065908, + 0.0032186375465244055, + 0.001477338606491685, + 0.0016135552432388067, + 0.0026211501099169254, + 0.0015712620224803686, + 0.003490326227620244, + 0.0009362320997752249, + 0.0012002171715721488, + 0.0015153420390561223, + 0.0010646289447322488, + 0.002009319607168436, + 0.0016242151614278555, + 0.0011697744484990835, + 0.0008851843886077404, + 0.0049749705940485, + 0.0026783680077642202, + 0.004758809227496386, + 0.003152061253786087, + 0.003639696864411235, + 0.007909610867500305, + 0.005276632960885763, + 0.003922958858311176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.029801908880472183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161116123199463 + ], + "opposite_logprob_tokens": [ + -1.7016111612319946 + ], + "logprob_reward": -0.20161116123199463, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161116123199463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1525862067937851, + 0.194367453455925, + 0.08787581324577332, + 0.07809397578239441, + 0.09440475702285767, + 0.08547060191631317, + 0.08994404971599579, + 0.06570807099342346, + 0.031626131385564804, + 0.11992298066616058 + ], + "credit_map": [ + 0.019416088238358498, + 0.017215734347701073, + 0.01886766403913498, + 0.02130940370261669, + 0.025829430669546127, + 0.015454669483006, + 0.01743442937731743, + 0.017058782279491425, + 0.03598320856690407, + 0.05441911891102791, + 0.029393162578344345, + 0.011358908377587795, + 0.012462206184864044, + 0.022397616878151894, + 0.011055097915232182, + 0.017298128455877304, + 0.01079969946295023, + 0.01435085479170084, + 0.013998817652463913, + 0.00880470871925354, + 0.011361392214894295, + 0.011581089347600937, + 0.009422813542187214, + 0.007556432858109474, + 0.014968348667025566, + 0.017614277079701424, + 0.010146616026759148, + 0.006586561445146799, + 0.00850072130560875, + 0.005287142936140299, + 0.007643511053174734, + 0.007346798200160265, + 0.00893692672252655, + 0.008175588212907314, + 0.008510536514222622, + 0.009442154318094254, + 0.021200494840741158, + 0.014194883406162262, + 0.012701597064733505, + 0.011242570355534554, + 0.006199102383106947, + 0.009039867669343948, + 0.012647532857954502, + 0.008038575761020184, + 0.01950794644653797, + 0.012562097050249577, + 0.009327675215899944, + 0.008147800341248512, + 0.007081582676619291, + 0.008157599717378616, + 0.016439810395240784, + 0.01301539782434702, + 0.01498230081051588, + 0.009906920604407787, + 0.012165762484073639, + 0.008194671012461185, + 0.0057268766686320305, + 0.006202608812600374, + 0.010525936260819435, + 0.005722971633076668, + 0.006404642481356859, + 0.010792583227157593, + 0.005807178560644388, + 0.014525269158184528, + 0.002583779627457261, + 0.0029554462525993586, + 0.005168393719941378, + 0.004701996687799692, + 0.00635420810431242, + 0.003505089320242405, + 0.003919437061995268, + 0.0024377796798944473, + 0.014170324429869652, + 0.008072363212704659, + 0.014972744509577751, + 0.010667433962225914, + 0.013488298282027245, + 0.026272058486938477, + 0.01827191188931465, + 0.014007851481437683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.20161116123199463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand, long, static shot places us on a desolate mountaintop shrouded in a torrential downpour. Shot in a stunning slow motion, the world seems frozen in time. At the center of the frame, a red-clad swordsman stands alone on the rugged, slippery rocks. He wears a rain-soaked crimson suit, the damp fabric clinging to his muscular frame, outlining his defined muscles. His long, dark hair, dampened by the rain, tangled strands plastered to his forehead and cheeks, highlighting his hawk-like eyes and furrowed brow, a fierce, untouchable aura woven between them. With a calm and resolute gesture, he slowly raises his arm, draws a gleaming ancient sword from his waist, and holds it firmly across his chest. During this process, every massive raindrop falling from the sky is clearly visible, like suspended transparent crystals, reflecting a faint light against the gray sky. Some slowly streak across the camera, while others collide and explode on his resolute face, scarlet clothes, and the mirror-like blade of his sword, splashing out circles of miniature, crystal-clear water crowns. The entire scene is gloomy, with only that hint of red, like a burning flame, forming a sharp contrast with the cold rain and gray rocks, creating a lonely, tense, and epic atmosphere of fate.", + "seed": 32651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_2062_seed_32651_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11785979140549899, + "vlm_logprob": -0.11785979140549899, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.11785979140549899, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11413015425205231 + ], + "opposite_logprob_tokens": [ + -2.3641302585601807 + ], + "logprob_reward": -0.11413015425205231, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11413015425205231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27747732400894165, + 0.20558056235313416, + 0.10315471142530441, + 0.05820028856396675, + 0.10255935043096542, + 0.04839338734745979, + 0.04695693776011467, + 0.052648819983005524, + 0.08155263960361481, + 0.023475930094718933 + ], + "credit_map": [ + 0.05709749832749367, + 0.02644745633006096, + 0.0408891923725605, + 0.014939991757273674, + 0.02004925347864628, + 0.07980375736951828, + 0.015574995428323746, + 0.02267519384622574, + 0.026462282985448837, + 0.0208063293248415, + 0.031244290992617607, + 0.021862903609871864, + 0.014579631388187408, + 0.04363701492547989, + 0.025689005851745605, + 0.021299105137586594, + 0.009118834510445595, + 0.010287052020430565, + 0.013175551779568195, + 0.014850394800305367, + 0.00939708761870861, + 0.013055562041699886, + 0.021901177242398262, + 0.01136904675513506, + 0.009197207167744637, + 0.007589773274958134, + 0.007973377592861652, + 0.003588871331885457, + 0.0037349944468587637, + 0.017829960212111473, + 0.004613554570823908, + 0.003672546474263072, + 0.009593671187758446, + 0.009237509220838547, + 0.011198556050658226, + 0.011466722004115582, + 0.006940260995179415, + 0.015127629972994328, + 0.011618489399552345, + 0.027376510202884674, + 0.0059940023347735405, + 0.008112233132123947, + 0.008365449495613575, + 0.006176866125315428, + 0.0043909307569265366, + 0.007151726167649031, + 0.004516526125371456, + 0.0036856518127024174, + 0.009953675791621208, + 0.008755892515182495, + 0.008737349882721901, + 0.004246854223310947, + 0.004320865031331778, + 0.005812642630189657, + 0.0021433632355183363, + 0.002986293751746416, + 0.005824677646160126, + 0.006336551159620285, + 0.009369437582790852, + 0.004126652609556913, + 0.005741991568356752, + 0.00929658580571413, + 0.0069014024920761585, + 0.005051520653069019, + 0.018109431490302086, + 0.01949368417263031, + 0.007638335693627596, + 0.007018734700977802, + 0.005010739900171757, + 0.00969653483480215, + 0.006555641070008278, + 0.008029534481465816, + 0.0034412723034620285, + 0.006518910173326731, + 0.0016634257044643164, + 0.002161359181627631, + 0.0015309080481529236, + 0.003527424531057477, + 0.0014731454430148005, + 0.0031594857573509216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11413015425205231 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014233966358006 + ], + "opposite_logprob_tokens": [ + -4.2642340660095215 + ], + "logprob_reward": -0.014233966358006, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014233966358006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3196713626384735, + 0.17720210552215576, + 0.13301457464694977, + 0.03004349209368229, + 0.033222127705812454, + 0.03871811926364899, + 0.04121501371264458, + 0.04596662521362305, + 0.15510039031505585, + 0.025846289470791817 + ], + "credit_map": [ + 0.048516903072595596, + 0.04173149913549423, + 0.035853683948516846, + 0.017748117446899414, + 0.023249417543411255, + 0.11198883503675461, + 0.018681615591049194, + 0.02190125361084938, + 0.02781839668750763, + 0.017099546268582344, + 0.017891334369778633, + 0.018732229247689247, + 0.012467260472476482, + 0.04207511991262436, + 0.024816490709781647, + 0.01630169153213501, + 0.039304908365011215, + 0.010409357026219368, + 0.009454946964979172, + 0.004576149396598339, + 0.00666273245587945, + 0.04641575366258621, + 0.00639704754576087, + 0.009793657809495926, + 0.003118915017694235, + 0.004488024394959211, + 0.004332094918936491, + 0.002267814939841628, + 0.004235669504851103, + 0.00418876763433218, + 0.004137152805924416, + 0.0032750461250543594, + 0.0034521734341979027, + 0.004664410371333361, + 0.0037751428317278624, + 0.003743685083463788, + 0.004717116244137287, + 0.00390787934884429, + 0.003404032438993454, + 0.005557679105550051, + 0.005000315140932798, + 0.004790545906871557, + 0.004746967926621437, + 0.006224273703992367, + 0.004207995720207691, + 0.004253723192960024, + 0.005601080600172281, + 0.003893210319802165, + 0.00495189381763339, + 0.0048429532907903194, + 0.004889673087745905, + 0.005333381239324808, + 0.005931254010647535, + 0.0038867145776748657, + 0.00519799068570137, + 0.006181143689900637, + 0.003650016849860549, + 0.006178747862577438, + 0.007181553635746241, + 0.005325915291905403, + 0.0035581551492214203, + 0.00622095912694931, + 0.008112683892250061, + 0.0057385824620723724, + 0.04079866036772728, + 0.0379292294383049, + 0.012803073041141033, + 0.01360852736979723, + 0.010345238260924816, + 0.017442183569073677, + 0.010137446224689484, + 0.01203599851578474, + 0.0038702914025634527, + 0.0043932306580245495, + 0.0020709424279630184, + 0.0033186248037964106, + 0.0020917642395943403, + 0.002974988427013159, + 0.002568438882008195, + 0.004558002110570669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014233966358006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256317436695099 + ], + "opposite_logprob_tokens": [ + -1.6006317138671875 + ], + "logprob_reward": -0.2256317436695099, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256317436695099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08461882919073105, + 0.16043807566165924, + 0.20339460670948029, + 0.07110432535409927, + 0.22264377772808075, + 0.06076117604970932, + 0.03747136890888214, + 0.03969678282737732, + 0.04230832681059837, + 0.07756271213293076 + ], + "credit_map": [ + 0.009839357808232307, + 0.01372026838362217, + 0.00933851022273302, + 0.009888826869428158, + 0.010128650814294815, + 0.01039633248001337, + 0.009956231340765953, + 0.011350654065608978, + 0.014188911765813828, + 0.01688651368021965, + 0.01130533404648304, + 0.02233259193599224, + 0.014909845776855946, + 0.03034009411931038, + 0.02841932512819767, + 0.022055452689528465, + 0.05644877254962921, + 0.013061697594821453, + 0.01207047514617443, + 0.011235876940190792, + 0.010844497941434383, + 0.07114428281784058, + 0.013924086466431618, + 0.014664924703538418, + 0.010736390948295593, + 0.007958116941154003, + 0.006379774305969477, + 0.005325976759195328, + 0.007349073421210051, + 0.017717823386192322, + 0.009227285161614418, + 0.0064098830334842205, + 0.017367620021104813, + 0.01957242377102375, + 0.024036791175603867, + 0.018342023715376854, + 0.022197989746928215, + 0.045213568955659866, + 0.023083070293068886, + 0.0528302863240242, + 0.007694351952522993, + 0.008825695142149925, + 0.007002761587500572, + 0.010025128722190857, + 0.0053692650981247425, + 0.007949101738631725, + 0.007695435080677271, + 0.0061994330026209354, + 0.0067327581346035, + 0.004759857431054115, + 0.0038601935375481844, + 0.004073930438607931, + 0.0036435879301279783, + 0.003044914221391082, + 0.006238551344722509, + 0.005117574241012335, + 0.00450171297416091, + 0.0062630693428218365, + 0.005839673336595297, + 0.0031918594613671303, + 0.0025267149321734905, + 0.010118922218680382, + 0.0039420654065907, + 0.0033127651549875736, + 0.009906861931085587, + 0.010194284841418266, + 0.0039370679296553135, + 0.004024955444037914, + 0.0025898499879986048, + 0.004645755980163813, + 0.003120065201073885, + 0.00388948991894722, + 0.008098402060568333, + 0.01194045040756464, + 0.006979237776249647, + 0.010260886512696743, + 0.0069583626464009285, + 0.009948372840881348, + 0.007430089637637138, + 0.015946917235851288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.2256317436695099 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370674327015877 + ], + "opposite_logprob_tokens": [ + -3.4087066650390625 + ], + "logprob_reward": -0.03370674327015877, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370674327015877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2772677540779114, + 0.12423577904701233, + 0.08337350934743881, + 0.02751413732767105, + 0.1685897260904312, + 0.050042372196912766, + 0.031019607558846474, + 0.15136705338954926, + 0.05720175430178642, + 0.0293883066624403 + ], + "credit_map": [ + 0.019300337880849838, + 0.06309404224157333, + 0.05003896355628967, + 0.014191652648150921, + 0.017888719215989113, + 0.07665034383535385, + 0.017237026244401932, + 0.018866674974560738, + 0.011067964136600494, + 0.01901252008974552, + 0.018575254827737808, + 0.019172444939613342, + 0.011094729416072369, + 0.010282323695719242, + 0.022017808631062508, + 0.013012719340622425, + 0.006248256657272577, + 0.008889462798833847, + 0.007092549465596676, + 0.013039848767220974, + 0.009937007911503315, + 0.007672999519854784, + 0.02022281102836132, + 0.010270572267472744, + 0.0035709792282432318, + 0.0034528158139437437, + 0.003802892053499818, + 0.003321864176541567, + 0.002650084439665079, + 0.0036984048783779144, + 0.0033852134365588427, + 0.0036318812053650618, + 0.016407936811447144, + 0.016228241845965385, + 0.02498486265540123, + 0.015535936690866947, + 0.016168609261512756, + 0.02641887404024601, + 0.019138064235448837, + 0.03370720148086548, + 0.007192795630544424, + 0.00837932713329792, + 0.007764954585582018, + 0.005711456760764122, + 0.004777601920068264, + 0.007657094858586788, + 0.004780345130711794, + 0.0037787947803735733, + 0.005669682752341032, + 0.004919972270727158, + 0.004578308202326298, + 0.0033509822096675634, + 0.002524618525058031, + 0.003482850966975093, + 0.0037229680456221104, + 0.002770223654806614, + 0.020470881834626198, + 0.025131136178970337, + 0.029437771067023277, + 0.010993370786309242, + 0.008575692772865295, + 0.029962662607431412, + 0.012056910432875156, + 0.014738617464900017, + 0.013094538822770119, + 0.013402448035776615, + 0.005347175989300013, + 0.00471523217856884, + 0.0036005384754389524, + 0.006955885794013739, + 0.004106282256543636, + 0.005979649722576141, + 0.005934315267950296, + 0.006613986101001501, + 0.0020361351780593395, + 0.0018523904727771878, + 0.0026229817885905504, + 0.003488128539174795, + 0.0030365698039531708, + 0.003803797299042344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03370674327015877 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159634947776794 + ], + "opposite_logprob_tokens": [ + -1.7015963792800903 + ], + "logprob_reward": -0.20159634947776794, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159634947776794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19152048230171204, + 0.08288758248090744, + 0.04981309548020363, + 0.04799254238605499, + 0.2795237600803375, + 0.07005693763494492, + 0.03440441936254501, + 0.027070092037320137, + 0.1513802707195282, + 0.06535080075263977 + ], + "credit_map": [ + 0.03614475578069687, + 0.015449102967977524, + 0.019279275089502335, + 0.009675748646259308, + 0.014316694810986519, + 0.06644092500209808, + 0.011132663115859032, + 0.01908130571246147, + 0.009937272407114506, + 0.00973193347454071, + 0.0104673495516181, + 0.007375792600214481, + 0.007204395718872547, + 0.020887916907668114, + 0.009514614939689636, + 0.00776830967515707, + 0.007271704729646444, + 0.004566948860883713, + 0.004986880347132683, + 0.00493375351652503, + 0.005682135466486216, + 0.012686331756412983, + 0.005533271003514528, + 0.004152069799602032, + 0.008858159184455872, + 0.004023067187517881, + 0.006038137711584568, + 0.004407960455864668, + 0.0032164272852241993, + 0.010887178592383862, + 0.005710171069949865, + 0.004851441830396652, + 0.03287109360098839, + 0.03199256211519241, + 0.052344806492328644, + 0.02702588215470314, + 0.022620445117354393, + 0.03397713601589203, + 0.035138264298439026, + 0.043553560972213745, + 0.00906719733029604, + 0.010352428071200848, + 0.009030431509017944, + 0.009401029907166958, + 0.007287449669092894, + 0.010463968850672245, + 0.008231541141867638, + 0.006222892086952925, + 0.005212221294641495, + 0.003974569495767355, + 0.004274246282875538, + 0.0041969213634729385, + 0.0034491929691284895, + 0.003267209976911545, + 0.0058211395516991615, + 0.004208915401250124, + 0.0030482939910143614, + 0.003052100772038102, + 0.0042242370545864105, + 0.004311542958021164, + 0.0033167190849781036, + 0.003542775521054864, + 0.0031588806305080652, + 0.002415541559457779, + 0.037741776555776596, + 0.036314379423856735, + 0.013166380114853382, + 0.011635709553956985, + 0.009813886135816574, + 0.017986595630645752, + 0.010363132692873478, + 0.014358419924974442, + 0.008714697323739529, + 0.01052790880203247, + 0.005588700529187918, + 0.008374160155653954, + 0.005912856198847294, + 0.008384132757782936, + 0.006009846460074186, + 0.011838499456644058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.20159634947776794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand, long, static shot places us on a desolate mountaintop shrouded in a torrential downpour. Shot in a stunning slow motion, the world seems frozen in time. At the center of the frame, a red-clad swordsman stands alone on the rugged, slippery rocks. He wears a rain-soaked crimson suit, the damp fabric clinging to his muscular frame, outlining his defined muscles. His long, dark hair, dampened by the rain, tangled strands plastered to his forehead and cheeks, highlighting his hawk-like eyes and furrowed brow, a fierce, untouchable aura woven between them. With a calm and resolute gesture, he slowly raises his arm, draws a gleaming ancient sword from his waist, and holds it firmly across his chest. During this process, every massive raindrop falling from the sky is clearly visible, like suspended transparent crystals, reflecting a faint light against the gray sky. Some slowly streak across the camera, while others collide and explode on his resolute face, scarlet clothes, and the mirror-like blade of his sword, splashing out circles of miniature, crystal-clear water crowns. The entire scene is gloomy, with only that hint of red, like a burning flame, forming a sharp contrast with the cold rain and gray rocks, creating a lonely, tense, and epic atmosphere of fate.", + "seed": 32652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_13/index_2062_seed_32652_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11095003476366401, + "vlm_logprob": -0.11095003476366401, + "vlm_yes_no_margin": 3.249999988079071, + "vlm_token_credit": -0.11095003476366401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026745764538645744 + ], + "opposite_logprob_tokens": [ + -4.401745796203613 + ], + "logprob_reward": -0.026745764538645744, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.026745764538645744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11010165512561798, + 0.1251147836446762, + 0.050879161804914474, + 0.09174320101737976, + 0.10013020038604736, + 0.10219649970531464, + 0.08023519814014435, + 0.23296038806438446, + 0.09064918011426926, + 0.01598973199725151 + ], + "credit_map": [ + 0.010612751357257366, + 0.012346896342933178, + 0.011083897203207016, + 0.010167398490011692, + 0.01450706273317337, + 0.014338785782456398, + 0.020006868988275528, + 0.01703799143433571, + 0.04068781062960625, + 0.007371163927018642, + 0.007572981528937817, + 0.006822915747761726, + 0.0438237190246582, + 0.009132574312388897, + 0.006166727282106876, + 0.0035368874669075012, + 0.006038499530404806, + 0.00977631751447916, + 0.004885934758931398, + 0.007114844396710396, + 0.007527979556471109, + 0.007205598056316376, + 0.0037059402093291283, + 0.004624047316610813, + 0.013109525665640831, + 0.008991197682917118, + 0.007645336911082268, + 0.011903280392289162, + 0.011369402520358562, + 0.020562656223773956, + 0.006999828387051821, + 0.011161974631249905, + 0.011020294390618801, + 0.01194295845925808, + 0.015473439358174801, + 0.012971789576113224, + 0.00784561038017273, + 0.012482119724154472, + 0.010469886474311352, + 0.017924098297953606, + 0.010201453231275082, + 0.019149459898471832, + 0.01763743720948696, + 0.012330209836363792, + 0.010399767197668552, + 0.009169243276119232, + 0.01245418842881918, + 0.010854740627110004, + 0.009755968116223812, + 0.009261290542781353, + 0.007307152263820171, + 0.00988653115928173, + 0.007609844673424959, + 0.013105541467666626, + 0.010872245766222477, + 0.012436628341674805, + 0.026347370818257332, + 0.034445084631443024, + 0.04080543667078018, + 0.023954883217811584, + 0.022333230823278427, + 0.03291850909590721, + 0.025553051382303238, + 0.026602819561958313, + 0.010540447197854519, + 0.013171403668820858, + 0.009345220401883125, + 0.010555650107562542, + 0.008537821471691132, + 0.013310219161212444, + 0.009331988170742989, + 0.015856429934501648, + 0.0018242240184918046, + 0.0021590201649814844, + 0.00231703813187778, + 0.002049206057563424, + 0.0017107379389926791, + 0.0019649409223347902, + 0.0017146175960078835, + 0.002249948913231492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.026745764538645744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005974648054689169 + ], + "opposite_logprob_tokens": [ + -5.130974769592285 + ], + "logprob_reward": -0.005974648054689169, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005974648054689169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42029789090156555, + 0.05338302627205849, + 0.04665389284491539, + 0.06004945561289787, + 0.05869944393634796, + 0.08042196184396744, + 0.07852800190448761, + 0.10739129781723022, + 0.06890547275543213, + 0.02566959522664547 + ], + "credit_map": [ + 0.04400402307510376, + 0.028842782601714134, + 0.05046491697430611, + 0.02465851604938507, + 0.05965464562177658, + 0.05777479335665703, + 0.07990022003650665, + 0.07499799132347107, + 0.009615080431103706, + 0.0066619920544326305, + 0.005016092211008072, + 0.0052088056690990925, + 0.011323816142976284, + 0.006265038624405861, + 0.005516400095075369, + 0.0037757987156510353, + 0.0060575492680072784, + 0.00726457079872489, + 0.004897051490843296, + 0.006704071070998907, + 0.006762494333088398, + 0.0065544298849999905, + 0.003939162939786911, + 0.004474562592804432, + 0.00860280729830265, + 0.0083506153896451, + 0.005943506024777889, + 0.009783330373466015, + 0.007040757220238447, + 0.008887728676199913, + 0.005126243457198143, + 0.006314465310424566, + 0.006796383764594793, + 0.006195738911628723, + 0.009206080809235573, + 0.007663709577172995, + 0.004832362290471792, + 0.007853547111153603, + 0.006094449665397406, + 0.010057173669338226, + 0.008440621197223663, + 0.01276220753788948, + 0.012614664621651173, + 0.008538279682397842, + 0.008607972413301468, + 0.01116039790213108, + 0.006560901645570993, + 0.011736920103430748, + 0.009414722211658955, + 0.010014781728386879, + 0.009195907972753048, + 0.00915420800447464, + 0.007805544883012772, + 0.009245696477591991, + 0.01071048155426979, + 0.012986660934984684, + 0.011908276006579399, + 0.01593969762325287, + 0.019256126135587692, + 0.010549440048635006, + 0.009155236184597015, + 0.014989493414759636, + 0.012369569391012192, + 0.013223459012806416, + 0.008184457197785378, + 0.009999310597777367, + 0.008136574178934097, + 0.008800766430795193, + 0.006585362832993269, + 0.00955976266413927, + 0.007845788262784481, + 0.009793453849852085, + 0.0033693548757582903, + 0.00470805075019598, + 0.002308104420080781, + 0.002505657961592078, + 0.002412983449175954, + 0.004354372154921293, + 0.002304587047547102, + 0.0037064850330352783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.005974648054689169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497122183442116 + ], + "opposite_logprob_tokens": [ + -2.929971218109131 + ], + "logprob_reward": -0.05497122183442116, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497122183442116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08693625032901764, + 0.09163142740726471, + 0.06760674715042114, + 0.07578140497207642, + 0.09601419419050217, + 0.08362390846014023, + 0.1118050292134285, + 0.17642834782600403, + 0.10262008756399155, + 0.10755255073308945 + ], + "credit_map": [ + 0.010381643660366535, + 0.010370076633989811, + 0.015006140805780888, + 0.010402316227555275, + 0.012501273304224014, + 0.007460121531039476, + 0.014954226091504097, + 0.005860462319105864, + 0.010312463156878948, + 0.013292364776134491, + 0.015497773885726929, + 0.015373935922980309, + 0.011039976961910725, + 0.009546576999127865, + 0.01028378028422594, + 0.006284561473876238, + 0.009100562892854214, + 0.010788465850055218, + 0.008251570165157318, + 0.012161854654550552, + 0.008150303736329079, + 0.008031207136809826, + 0.005151465069502592, + 0.005971323698759079, + 0.009156538173556328, + 0.012014983221888542, + 0.008556527085602283, + 0.013989162631332874, + 0.006352008320391178, + 0.009063090197741985, + 0.007543077692389488, + 0.00910602230578661, + 0.01092240959405899, + 0.016930021345615387, + 0.012567365542054176, + 0.014336904510855675, + 0.007754983380436897, + 0.009963287971913815, + 0.009538429789245129, + 0.014000806957483292, + 0.008948070928454399, + 0.009799715131521225, + 0.015088671818375587, + 0.012689169496297836, + 0.0060317483730614185, + 0.011320983059704304, + 0.010836745612323284, + 0.008908807300031185, + 0.012268098071217537, + 0.013137347996234894, + 0.023947738111019135, + 0.013492758385837078, + 0.008478167466819286, + 0.012347925454378128, + 0.013333084993064404, + 0.014799910597503185, + 0.021104227751493454, + 0.025309566408395767, + 0.036670997738838196, + 0.018639657646417618, + 0.01601584255695343, + 0.02481473796069622, + 0.01599395088851452, + 0.01787937432527542, + 0.010875820182263851, + 0.015944838523864746, + 0.016730211675167084, + 0.014058184809982777, + 0.008781356737017632, + 0.010508030652999878, + 0.011077906005084515, + 0.01464375201612711, + 0.013231252320110798, + 0.018824860453605652, + 0.018232621252536774, + 0.012632524594664574, + 0.008319953456521034, + 0.012236603535711765, + 0.0102933868765831, + 0.0137813501060009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.05497122183442116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099825382232666 + ], + "opposite_logprob_tokens": [ + -3.288099765777588 + ], + "logprob_reward": -0.038099825382232666, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099825382232666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47282952070236206, + 0.04484853520989418, + 0.03388872370123863, + 0.05566255748271942, + 0.06668348610401154, + 0.0669483169913292, + 0.04859209060668945, + 0.14490406215190887, + 0.049024712294340134, + 0.016617897897958755 + ], + "credit_map": [ + 0.05137468874454498, + 0.03195801004767418, + 0.04802193120121956, + 0.026424819603562355, + 0.0654585212469101, + 0.07260598987340927, + 0.08614198118448257, + 0.09084359556436539, + 0.004435384646058083, + 0.004670368507504463, + 0.005149343051016331, + 0.004428461194038391, + 0.006102303974330425, + 0.007390794809907675, + 0.006463997066020966, + 0.006207884289324284, + 0.003747160080820322, + 0.0064035323448479176, + 0.0030512078665196896, + 0.004045908339321613, + 0.004217901267111301, + 0.0047836084850132465, + 0.0025121080689132214, + 0.005127298645675182, + 0.007704444229602814, + 0.005320953205227852, + 0.004705536644905806, + 0.007182395551353693, + 0.006445724982768297, + 0.012907193042337894, + 0.00423257565125823, + 0.007163732312619686, + 0.008121898397803307, + 0.009673652239143848, + 0.009481650777161121, + 0.007939421571791172, + 0.0055776238441467285, + 0.008584145456552505, + 0.006101617589592934, + 0.011203477159142494, + 0.006779036484658718, + 0.011629668064415455, + 0.010365298017859459, + 0.009068740531802177, + 0.006408588495105505, + 0.00619627395644784, + 0.009286908432841301, + 0.007213803939521313, + 0.005947279278188944, + 0.006379827857017517, + 0.005207800772041082, + 0.005644896067678928, + 0.004201310686767101, + 0.008350178599357605, + 0.006429008673876524, + 0.0064317891374230385, + 0.016487542539834976, + 0.022155139595270157, + 0.025202879682183266, + 0.015320256352424622, + 0.013926573097705841, + 0.020398568361997604, + 0.015490075573325157, + 0.015923036262392998, + 0.005782119929790497, + 0.00734926899895072, + 0.004466303624212742, + 0.005654372274875641, + 0.004589225631207228, + 0.006655639968812466, + 0.006104643456637859, + 0.008423139341175556, + 0.001845210324972868, + 0.002673251088708639, + 0.0029583382420241833, + 0.002272426150739193, + 0.0015236365143209696, + 0.0013438506284728646, + 0.0019266351591795683, + 0.002074550138786435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.038099825382232666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountaintop with rugged, slippery rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a red-clad swordsman standing alone in the center of the frame wearing a rain-soaked crimson suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman slowly raise his arm to draw a gleaming ancient sword from his waist and hold it across his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive raindrops clearly visible as suspended transparent crystals reflecting light against a gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static frame captured in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289587140083313 + ], + "opposite_logprob_tokens": [ + -1.0539586544036865 + ], + "logprob_reward": -0.4289587140083313, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4289587140083313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37872153520584106, + 0.08514844626188278, + 0.10249944776296616, + 0.05349113419651985, + 0.08414018899202347, + 0.0877358540892601, + 0.07308263331651688, + 0.04461114481091499, + 0.04847702756524086, + 0.04209257662296295 + ], + "credit_map": [ + 0.04158596694469452, + 0.02214723825454712, + 0.04273120313882828, + 0.022004317492246628, + 0.04762314632534981, + 0.05154295265674591, + 0.06938263028860092, + 0.08170408010482788, + 0.01722870208323002, + 0.007069517858326435, + 0.007959386333823204, + 0.010009446181356907, + 0.019929280504584312, + 0.008670851588249207, + 0.006463769357651472, + 0.007817494682967663, + 0.012632720172405243, + 0.012073182500898838, + 0.009756102226674557, + 0.012853247113525867, + 0.01378792803734541, + 0.013484410010278225, + 0.008752845227718353, + 0.019159015268087387, + 0.006812988314777613, + 0.005187263712286949, + 0.0074820625595748425, + 0.007278681267052889, + 0.004967698361724615, + 0.011348460800945759, + 0.004048993811011314, + 0.006364981643855572, + 0.010653182864189148, + 0.017075151205062866, + 0.010826301760971546, + 0.01262350007891655, + 0.00660787895321846, + 0.010416852310299873, + 0.00805665273219347, + 0.00788067001849413, + 0.008612677454948425, + 0.01729981042444706, + 0.015047854743897915, + 0.0104313800111413, + 0.009047889150679111, + 0.006981891114264727, + 0.010923716239631176, + 0.009390640072524548, + 0.007197525352239609, + 0.007936877198517323, + 0.011838079430162907, + 0.010042917914688587, + 0.005759581923484802, + 0.011130021885037422, + 0.009772853925824165, + 0.009404770098626614, + 0.005062548443675041, + 0.008554534986615181, + 0.005184251349419355, + 0.004738485906273127, + 0.005393440835177898, + 0.004494192078709602, + 0.0057558780536055565, + 0.005427812226116657, + 0.005775194149464369, + 0.007988076657056808, + 0.008394209668040276, + 0.006408065091818571, + 0.0036497339606285095, + 0.005759196821600199, + 0.004636544734239578, + 0.005866006948053837, + 0.005096472334116697, + 0.007944950833916664, + 0.0061216168105602264, + 0.004693044815212488, + 0.004309775307774544, + 0.004455979447811842, + 0.0047391438856720924, + 0.00473159272223711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.4289587140083313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_tensor([767, 767])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_tensor([767, 767])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..687ce8a9f3648ba23aa48491e7eaca835a274aed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_tensor([767, 767])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures the emptiness of a late-night grocery store aisle. The floor, made of highly polished white linoleum, reflects the cold, buzzing green-white glare of the overhead fluorescent lights, creating a sterile and unsettling environment. Towering shelves line the aisles, perfectly stocked with generic, unlabeled boxes and cans, casting long, dark shadows that stretch into the deep background. In the very center of the frame sits a lone, slightly worn metal shopping cart, its chrome wire basket gleaming in the harsh light, its red plastic handle faded. Initially completely still, the cart begins to tremble slightly, a faint vibration that develops into a violent, convulsive shudder. Its entire metal frame wobbles unsteadily, and its small, black wheels begin to rattle violently and loudly against the polished floor. The sharp, rattling sound echoes uncomfortably throughout the desolate space, creating a tense and deeply unsettling atmosphere of unseen forces at work.", + "seed": 10782, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_767_seed_10782_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06181514938361943, + "vlm_logprob": -0.06181514938361943, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06181514938361943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051449261605739594 + ], + "opposite_logprob_tokens": [ + -3.0514492988586426 + ], + "logprob_reward": -0.051449261605739594, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051449261605739594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20910672843456268, + 0.05420808866620064, + 0.01533025223761797, + 0.011219603940844536, + 0.007176118902862072, + 0.010979895479977131, + 0.015334822237491608, + 0.0061895884573459625, + 0.6034113764762878, + 0.06704340875148773 + ], + "credit_map": [ + 0.015512622892856598, + 0.01350528933107853, + 0.012795304879546165, + 0.056973960250616074, + 0.011623105965554714, + 0.012342204339802265, + 0.022739028558135033, + 0.06361524015665054, + 0.003979808650910854, + 0.0034960941411554813, + 0.0029565836302936077, + 0.016259312629699707, + 0.0027340431697666645, + 0.0033348745200783014, + 0.006756535265594721, + 0.014690844342112541, + 0.0012454009847715497, + 0.0008821429219096899, + 0.0009165247320197523, + 0.0044667585752904415, + 0.0012018440756946802, + 0.0009716570493765175, + 0.0019055928569287062, + 0.0037403320893645287, + 0.0010689999908208847, + 0.0008072350756265223, + 0.0011097653768956661, + 0.003088048193603754, + 0.0009284931002184749, + 0.0006797875976189971, + 0.001114560873247683, + 0.0024227146059274673, + 0.0007501904619857669, + 0.00044457323383539915, + 0.0005603809840977192, + 0.0018534815171733499, + 0.000595846155192703, + 0.0004945601685903966, + 0.000893009768333286, + 0.0015840771375223994, + 0.001019696006551385, + 0.0008086478337645531, + 0.0007959460490383208, + 0.003372553735971451, + 0.0007684995071031153, + 0.0006497386493720114, + 0.0012970786774531007, + 0.002267735544592142, + 0.0019683002028614283, + 0.001372462953440845, + 0.0015480808215215802, + 0.0027845625299960375, + 0.0012207092950120568, + 0.0011252930853515863, + 0.001777638797648251, + 0.0035377757158130407, + 0.0005436017527244985, + 0.000529989309143275, + 0.0008461844408884645, + 0.001275650691241026, + 0.0009844357846304774, + 0.00044464762322604656, + 0.0009969411185011268, + 0.0005681382026523352, + 0.04743652045726776, + 0.03157978504896164, + 0.029860449954867363, + 0.20836353302001953, + 0.01572176069021225, + 0.033148519694805145, + 0.06915847957134247, + 0.1681423932313919, + 0.007967124693095684, + 0.007008884567767382, + 0.007047970313578844, + 0.007740252651274204, + 0.008674081414937973, + 0.007010361645370722, + 0.011200807057321072, + 0.010393929667770863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.051449261605739594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060040831565857 + ], + "opposite_logprob_tokens": [ + -3.8956003189086914 + ], + "logprob_reward": -0.02060040831565857, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060040831565857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18996194005012512, + 0.019759293645620346, + 0.013645973056554794, + 0.009385828860104084, + 0.02044421248137951, + 0.015986494719982147, + 0.015496560372412205, + 0.025447579100728035, + 0.5920478701591492, + 0.09782420843839645 + ], + "credit_map": [ + 0.015095728449523449, + 0.011536397971212864, + 0.009575334377586842, + 0.06148485094308853, + 0.007718218956142664, + 0.011597630567848682, + 0.021717548370361328, + 0.05123624950647354, + 0.001670078025199473, + 0.0018439505947753787, + 0.00174629723187536, + 0.002915944205597043, + 0.002365299966186285, + 0.0013010927941650152, + 0.0026838607154786587, + 0.005232768598943949, + 0.0010822993936017156, + 0.0010376775171607733, + 0.0010383592452853918, + 0.003961156588047743, + 0.0011087656021118164, + 0.0009362984565086663, + 0.0021468389313668013, + 0.0023345774970948696, + 0.0008968039182946086, + 0.000658712291624397, + 0.0012587736127898097, + 0.0021309168078005314, + 0.0007958552450872958, + 0.0007113287574611604, + 0.0011242963373661041, + 0.0018091420643031597, + 0.0015923460014164448, + 0.0011326585663482547, + 0.001015635789372027, + 0.006595568731427193, + 0.001066292985342443, + 0.0010457157623022795, + 0.0021558331791311502, + 0.005840160418301821, + 0.0013469423865899444, + 0.0009311236790381372, + 0.001158243860118091, + 0.004961305297911167, + 0.0006907478091306984, + 0.0008946425514295697, + 0.0016718135448172688, + 0.004331675358116627, + 0.0019599914085119963, + 0.0014348655240610242, + 0.001388802076689899, + 0.0030330687295645475, + 0.0024269723799079657, + 0.0010730235371738672, + 0.0026777861639857292, + 0.0015020499704405665, + 0.0019207221921533346, + 0.0016916419845074415, + 0.0015843171859160066, + 0.006193140521645546, + 0.0021364495623856783, + 0.001597442664206028, + 0.0028945691883563995, + 0.007429298013448715, + 0.04654514789581299, + 0.03068976290524006, + 0.028979647904634476, + 0.202813521027565, + 0.01706334762275219, + 0.032543741166591644, + 0.06710349768400192, + 0.16630919277668, + 0.013834971003234386, + 0.008317497558891773, + 0.022199071943759918, + 0.012154508382081985, + 0.009071802720427513, + 0.005791477859020233, + 0.01725393533706665, + 0.009200945496559143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02060040831565857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824367256835103 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.002824367256835103, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824367256835103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3124704360961914, + 0.12477310001850128, + 0.16504345834255219, + 0.043611105531454086, + 0.04779189079999924, + 0.04527617245912552, + 0.03778783231973648, + 0.032506220042705536, + 0.11063998192548752, + 0.08009982854127884 + ], + "credit_map": [ + 0.025939973071217537, + 0.024895718321204185, + 0.026658134534955025, + 0.10470965504646301, + 0.041387300938367844, + 0.019207263365387917, + 0.04242968559265137, + 0.02724270522594452, + 0.012342752888798714, + 0.01587957888841629, + 0.010627022944390774, + 0.0194461140781641, + 0.02689177356660366, + 0.008607489988207817, + 0.020409947261214256, + 0.010568412952125072, + 0.01275184378027916, + 0.014828525483608246, + 0.010112335905432701, + 0.039530619978904724, + 0.02009165845811367, + 0.00865611620247364, + 0.015946006402373314, + 0.04312635213136673, + 0.005242763087153435, + 0.004576660227030516, + 0.0036738705821335316, + 0.006723922211676836, + 0.007261019200086594, + 0.003331756452098489, + 0.006431250367313623, + 0.006369860842823982, + 0.003935799468308687, + 0.00441681919619441, + 0.0043082269839942455, + 0.01443670503795147, + 0.0059855347499251366, + 0.002851746743544936, + 0.006044850684702396, + 0.005812210962176323, + 0.004370800219476223, + 0.004310045391321182, + 0.0030806604772806168, + 0.010334626771509647, + 0.005447571165859699, + 0.00260805687867105, + 0.005338608752936125, + 0.00978580117225647, + 0.003309143939986825, + 0.0034266039729118347, + 0.0032755043357610703, + 0.010322830639779568, + 0.004019912797957659, + 0.0019309523049741983, + 0.004582234658300877, + 0.0069206515327095985, + 0.0037358496338129044, + 0.0050583407282829285, + 0.0028395738918334246, + 0.004855820443481207, + 0.00652301637455821, + 0.0021313035394996405, + 0.0047991047613322735, + 0.00256321276538074, + 0.00788603350520134, + 0.007036007475107908, + 0.007081602234393358, + 0.030510760843753815, + 0.01000999566167593, + 0.006455632857978344, + 0.01566765271127224, + 0.025992291048169136, + 0.007806545123457909, + 0.00807330384850502, + 0.013473291881382465, + 0.010461481288075447, + 0.010217657312750816, + 0.0063873110339045525, + 0.01159743033349514, + 0.01208280399441719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002824367256835103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556638717651367 + ], + "opposite_logprob_tokens": [ + -1.6005663871765137 + ], + "logprob_reward": -0.22556638717651367, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556638717651367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03671807050704956, + 0.0240931436419487, + 0.008900951594114304, + 0.00356182805262506, + 0.01332364697009325, + 0.003010704182088375, + 0.012366901151835918, + 0.014343169517815113, + 0.8698190450668335, + 0.013862558640539646 + ], + "credit_map": [ + 0.0034778332337737083, + 0.003235507756471634, + 0.0046362061984837055, + 0.011049345135688782, + 0.0042584543116390705, + 0.002537192078307271, + 0.00505834212526679, + 0.0024651868734508753, + 0.0019191940082237124, + 0.0016685917507857084, + 0.0014469749294221401, + 0.0073234280571341515, + 0.002105498919263482, + 0.0014775757445022464, + 0.0029316397849470377, + 0.0052202376537024975, + 0.0007316841511055827, + 0.0007476447499357164, + 0.0006019067368470132, + 0.002066939603537321, + 0.0014527294551953673, + 0.0006170974229462445, + 0.0015272287419065833, + 0.0011557203251868486, + 0.00034257693914696574, + 0.0003490695671644062, + 0.00025147784617729485, + 0.0005490314797498286, + 0.0007064402452670038, + 0.0003200875362381339, + 0.0005449890159070492, + 0.000498155364766717, + 0.0011278896126896143, + 0.0007471950375474989, + 0.0007600323879159987, + 0.004301633685827255, + 0.000768986705224961, + 0.0007849943940527737, + 0.0015315687051042914, + 0.0033013455104082823, + 0.00033961268491111696, + 0.0003426160546950996, + 0.0002849512093234807, + 0.00038560948451049626, + 0.0005179739673621953, + 0.00022735680977348238, + 0.000493241532240063, + 0.0004193421336822212, + 0.0010317215928807855, + 0.0008007222204469144, + 0.0008133035735227168, + 0.0029281110037118196, + 0.0013031124835833907, + 0.0005760190542787313, + 0.0013212555786594748, + 0.0035926541313529015, + 0.0012365401489660144, + 0.0007288297056220472, + 0.0007544805994257331, + 0.004372026305645704, + 0.0007441097986884415, + 0.0007485975511372089, + 0.0014870184240862727, + 0.0042715673334896564, + 0.06810307502746582, + 0.04539884254336357, + 0.04271046444773674, + 0.299019455909729, + 0.0232856422662735, + 0.04792725294828415, + 0.09933032840490341, + 0.24404391646385193, + 0.0015860272105783224, + 0.0011733026476576924, + 0.0024121892638504505, + 0.002000631997361779, + 0.0015981647884473205, + 0.0011950032785534859, + 0.002411419991403818, + 0.0014858184149488807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.22556638717651367 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09600090235471725, + 0.019904309883713722, + 0.037427809089422226, + 0.013436106964945793, + 0.010335158556699753, + 0.010640249587595463, + 0.02706979215145111, + 0.006654972676187754, + 0.7170876860618591, + 0.061443012207746506 + ], + "credit_map": [ + 0.011706167832016945, + 0.007521662395447493, + 0.008155498653650284, + 0.020648831501603127, + 0.006989206187427044, + 0.006097767502069473, + 0.011692013591527939, + 0.023189758881926537, + 0.002310725161805749, + 0.0022622467949986458, + 0.0024507702328264713, + 0.003037598915398121, + 0.0023051337338984013, + 0.0013976584887132049, + 0.0031364255119115114, + 0.003003752091899514, + 0.0029328290838748217, + 0.0023559625260531902, + 0.0022800806909799576, + 0.010550969280302525, + 0.001905833138152957, + 0.002445884281769395, + 0.004041744861751795, + 0.010914506390690804, + 0.001018580631352961, + 0.0008789767161943018, + 0.0009026207262650132, + 0.0035972571931779385, + 0.0009596634772606194, + 0.0007881724159233272, + 0.0015209292760118842, + 0.0037699067033827305, + 0.00084814423462376, + 0.0011259904131293297, + 0.0008311865967698395, + 0.002409892389550805, + 0.0010173283517360687, + 0.0005949159967713058, + 0.0017650073859840631, + 0.0017426928970962763, + 0.0011050842003896832, + 0.001250474713742733, + 0.0012318193912506104, + 0.0027931886725127697, + 0.0009611492278054357, + 0.0006915656849741936, + 0.0012601420748978853, + 0.0013468251563608646, + 0.002093348652124405, + 0.00164107047021389, + 0.0019124987302348018, + 0.008732784539461136, + 0.001317647285759449, + 0.0014509001048281789, + 0.003093510866165161, + 0.006828032433986664, + 0.0006227839039638638, + 0.00099810806568712, + 0.0006394365918822587, + 0.0008699385216459632, + 0.0010395674034953117, + 0.00042309961281716824, + 0.0013166309799998999, + 0.0007454071310348809, + 0.05598617345094681, + 0.03797003626823425, + 0.03526056557893753, + 0.24723677337169647, + 0.018990829586982727, + 0.03958003968000412, + 0.08296497166156769, + 0.19909827411174774, + 0.00619560107588768, + 0.0065242392010986805, + 0.011281011626124382, + 0.006520859897136688, + 0.009597640484571457, + 0.00463165482506156, + 0.010647204704582691, + 0.006044798064976931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0086353225633502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the emptiness of a late-night grocery store aisle. The floor, made of highly polished white linoleum, reflects the cold, buzzing green-white glare of the overhead fluorescent lights, creating a sterile and unsettling environment. Towering shelves line the aisles, perfectly stocked with generic, unlabeled boxes and cans, casting long, dark shadows that stretch into the deep background. In the very center of the frame sits a lone, slightly worn metal shopping cart, its chrome wire basket gleaming in the harsh light, its red plastic handle faded. Initially completely still, the cart begins to tremble slightly, a faint vibration that develops into a violent, convulsive shudder. Its entire metal frame wobbles unsteadily, and its small, black wheels begin to rattle violently and loudly against the polished floor. The sharp, rattling sound echoes uncomfortably throughout the desolate space, creating a tense and deeply unsettling atmosphere of unseen forces at work.", + "seed": 10783, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_767_seed_10783_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1551086745224893, + "vlm_logprob": -0.1551086745224893, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.1551086745224893, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05326610058546066 + ], + "opposite_logprob_tokens": [ + -3.0532660484313965 + ], + "logprob_reward": -0.05326610058546066, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05326610058546066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4378073513507843, + 0.04675714299082756, + 0.03453155979514122, + 0.030300015583634377, + 0.03804996982216835, + 0.13844899833202362, + 0.03741941973567009, + 0.10798788070678711, + 0.08299458026885986, + 0.045703064650297165 + ], + "credit_map": [ + 0.040403008460998535, + 0.010829085484147072, + 0.014208273962140083, + 0.08602017909288406, + 0.018780285492539406, + 0.017401963472366333, + 0.16661210358142853, + 0.08355248719453812, + 0.005558253265917301, + 0.0030088683124631643, + 0.005320530850440264, + 0.006861575413495302, + 0.0027004044968634844, + 0.003810723777860403, + 0.007646349724382162, + 0.011850436218082905, + 0.0028085075318813324, + 0.0020209145732223988, + 0.002934232586994767, + 0.008018054999411106, + 0.0019370559602975845, + 0.002239583758637309, + 0.0043272459879517555, + 0.010245965793728828, + 0.0033650831319391727, + 0.0017929138848558068, + 0.0018991880351677537, + 0.0042533366940915585, + 0.00208150502294302, + 0.001600741408765316, + 0.007407601457089186, + 0.007899644784629345, + 0.00585746206343174, + 0.0030659339390695095, + 0.0037713327910751104, + 0.004318871069699526, + 0.002636759774759412, + 0.004737251438200474, + 0.005312703084200621, + 0.008349654264748096, + 0.015304862521588802, + 0.006078279111534357, + 0.009599581360816956, + 0.022085176780819893, + 0.006602768320590258, + 0.0102616548538208, + 0.04218200966715813, + 0.02633466012775898, + 0.002776532666757703, + 0.0015867369947955012, + 0.0016332565573975444, + 0.013447325676679611, + 0.00169279333204031, + 0.0013005777727812529, + 0.012587934732437134, + 0.0023942606057971716, + 0.007438404951244593, + 0.004654645454138517, + 0.007876889780163765, + 0.03510985150933266, + 0.004480138886719942, + 0.0058655147440731525, + 0.036927785724401474, + 0.00563464593142271, + 0.010215419344604015, + 0.01306856982409954, + 0.00918467715382576, + 0.013804150745272636, + 0.00992983765900135, + 0.009171062149107456, + 0.008397046476602554, + 0.009223819710314274, + 0.005547506734728813, + 0.005102552939206362, + 0.005674763582646847, + 0.006731023546308279, + 0.007645995356142521, + 0.005915675312280655, + 0.00474507687613368, + 0.004340472165495157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.05326610058546066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102093383669853 + ], + "opposite_logprob_tokens": [ + -4.511102199554443 + ], + "logprob_reward": -0.011102093383669853, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102093383669853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3584921956062317, + 0.08175881952047348, + 0.07211556285619736, + 0.010651332326233387, + 0.03142646700143814, + 0.12038497626781464, + 0.024288540706038475, + 0.1948562115430832, + 0.07636356353759766, + 0.029662346467375755 + ], + "credit_map": [ + 0.031894370913505554, + 0.0111448485404253, + 0.015444565564393997, + 0.07516870647668839, + 0.016666168347001076, + 0.013578999787569046, + 0.13249561190605164, + 0.06209892779588699, + 0.008629908785223961, + 0.007701868191361427, + 0.007889802567660809, + 0.018103817477822304, + 0.006997114513069391, + 0.007286002393811941, + 0.014849931932985783, + 0.010300373658537865, + 0.006528788711875677, + 0.003362552262842655, + 0.004947736859321594, + 0.02414991520345211, + 0.004134104121476412, + 0.003645164193585515, + 0.02053474262356758, + 0.004812560975551605, + 0.0011918970849364996, + 0.0008339884807355702, + 0.0008899233653210104, + 0.0018646171083673835, + 0.0009146296652033925, + 0.0008552651852369308, + 0.002869816031306982, + 0.0012311951722949743, + 0.0021212357096374035, + 0.0028761394787579775, + 0.0025149553548544645, + 0.008963558822870255, + 0.0027454434894025326, + 0.0022336686961352825, + 0.007280851248651743, + 0.0026906125713139772, + 0.011634191498160362, + 0.0031585616525262594, + 0.003511378075927496, + 0.02668202668428421, + 0.005076105706393719, + 0.004855647217482328, + 0.04590315744280815, + 0.019563905894756317, + 0.0018502917373552918, + 0.0013376158894971013, + 0.0012050479417666793, + 0.008001267910003662, + 0.0012650745920836926, + 0.0014384808018803596, + 0.005682365968823433, + 0.0035083952825516462, + 0.009824358858168125, + 0.008798349648714066, + 0.01780254952609539, + 0.062217894941568375, + 0.008055443875491619, + 0.007921360433101654, + 0.06797249615192413, + 0.012263762764632702, + 0.007963508367538452, + 0.007717340253293514, + 0.008653827011585236, + 0.015116152353584766, + 0.009438879787921906, + 0.007640884257853031, + 0.01134017389267683, + 0.008492802269756794, + 0.0022843272890895605, + 0.0043868571519851685, + 0.002897827187553048, + 0.005147148855030537, + 0.007254311349242926, + 0.0027962299063801765, + 0.0025602378882467747, + 0.0023354075383394957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.011102093383669853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086059734225273 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011086059734225273, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086059734225273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41332101821899414, + 0.04628818854689598, + 0.01822040230035782, + 0.007727371528744698, + 0.021828899160027504, + 0.12146890163421631, + 0.012590060941874981, + 0.22507402300834656, + 0.06776432693004608, + 0.06571677327156067 + ], + "credit_map": [ + 0.03634024038910866, + 0.018008628860116005, + 0.023930130526423454, + 0.07788418233394623, + 0.019950689747929573, + 0.01972643844783306, + 0.1373002678155899, + 0.08018049597740173, + 0.00561858294531703, + 0.0037110422272235155, + 0.0046551828272640705, + 0.006674493663012981, + 0.004173900466412306, + 0.003801092505455017, + 0.006842533592134714, + 0.010811365209519863, + 0.001788145862519741, + 0.0020325123332440853, + 0.0014328050892800093, + 0.003963687922805548, + 0.0017110409680753946, + 0.0013963201781734824, + 0.0022855098359286785, + 0.0036103830207139254, + 0.0008598953718319535, + 0.00072250742232427, + 0.0008779229829087853, + 0.0013836444122716784, + 0.000731474079657346, + 0.0009193156729452312, + 0.001206509186886251, + 0.0010261022252961993, + 0.0034238439984619617, + 0.0020756302401423454, + 0.0022624912671744823, + 0.0029473784379661083, + 0.0016809663502499461, + 0.0024384907446801662, + 0.0032808806281536818, + 0.0037192190065979958, + 0.012110697105526924, + 0.002975888317450881, + 0.0034512714482843876, + 0.0263929832726717, + 0.004825356416404247, + 0.005017566494643688, + 0.047463130205869675, + 0.019232019782066345, + 0.0011653808178380132, + 0.000620788661763072, + 0.0009235739125870168, + 0.003146624658256769, + 0.0006952897529117763, + 0.0007626087754033506, + 0.001583258854225278, + 0.003692536847665906, + 0.011737876571714878, + 0.010001595132052898, + 0.02068363130092621, + 0.07062884420156479, + 0.009305806830525398, + 0.009600135497748852, + 0.07841552048921585, + 0.014700623229146004, + 0.007963226176798344, + 0.00502986554056406, + 0.006461859215050936, + 0.008352374657988548, + 0.007256383541971445, + 0.008211539126932621, + 0.010275573469698429, + 0.01421351358294487, + 0.004692030604928732, + 0.009482230991125107, + 0.0057290056720376015, + 0.009554876945912838, + 0.019470246508717537, + 0.005721624940633774, + 0.006070781033486128, + 0.0049959891475737095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.011086059734225273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933517456054688 + ], + "opposite_logprob_tokens": [ + -0.6933517456054688 + ], + "logprob_reward": -0.6933517456054688, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933517456054688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6203206181526184, + 0.02711571380496025, + 0.021292567253112793, + 0.015857994556427002, + 0.018282467499375343, + 0.13012543320655823, + 0.020287301391363144, + 0.04935861751437187, + 0.05982436239719391, + 0.0375349260866642 + ], + "credit_map": [ + 0.055380068719387054, + 0.018046673387289047, + 0.02638273499906063, + 0.12439727783203125, + 0.024725589901208878, + 0.025836974382400513, + 0.22307197749614716, + 0.12247928231954575, + 0.003518085926771164, + 0.0023501026444137096, + 0.0030799084343016148, + 0.0031323153525590897, + 0.0023780823685228825, + 0.002362787025049329, + 0.0035807983949780464, + 0.006713631097227335, + 0.0021878674160689116, + 0.0014512697234749794, + 0.0020835178438574076, + 0.0038126136641949415, + 0.0013761253794655204, + 0.0015341975959017873, + 0.005151332821696997, + 0.003695641877129674, + 0.0021487418562173843, + 0.0011912243207916617, + 0.0021480638533830643, + 0.0030281199142336845, + 0.0012661244254559278, + 0.0014338316395878792, + 0.002967948792502284, + 0.0016739378916099668, + 0.002782379509881139, + 0.0019284537993371487, + 0.002177455695345998, + 0.0031327686738222837, + 0.0017796705942600965, + 0.002140727825462818, + 0.0027552915271371603, + 0.001585719408467412, + 0.012943744659423828, + 0.0030243447981774807, + 0.0038387721870094538, + 0.027012597769498825, + 0.005208236165344715, + 0.005259941332042217, + 0.05080061033368111, + 0.02203718014061451, + 0.002194356871768832, + 0.0010161260142922401, + 0.0015205743256956339, + 0.0031299057882279158, + 0.000998020637780428, + 0.001304818200878799, + 0.003925356548279524, + 0.0061981407925486565, + 0.003194408258423209, + 0.0038853357546031475, + 0.004220982547849417, + 0.01506119966506958, + 0.003152727149426937, + 0.0020363102667033672, + 0.014743820764124393, + 0.003063831478357315, + 0.005463617853820324, + 0.007037723436951637, + 0.0059061190113425255, + 0.012201627716422081, + 0.007327488157898188, + 0.00729535473510623, + 0.008694969117641449, + 0.005897461902350187, + 0.003126532305032015, + 0.005008883308619261, + 0.004201383795589209, + 0.006763911340385675, + 0.00760639226064086, + 0.0035840249620378017, + 0.0037152061704546213, + 0.0035285893827676773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.6933517456054688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737373303622007 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737373303622007, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737373303622007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23550766706466675, + 0.0726637989282608, + 0.07885795831680298, + 0.05470261350274086, + 0.028718503192067146, + 0.1544807404279709, + 0.05822306126356125, + 0.11998091638088226, + 0.15476252138614655, + 0.04210224375128746 + ], + "credit_map": [ + 0.02547350339591503, + 0.024963822215795517, + 0.031231334432959557, + 0.0434635691344738, + 0.01761498488485813, + 0.022972483187913895, + 0.04594430699944496, + 0.02384364791214466, + 0.005324718076735735, + 0.005930568091571331, + 0.005751113872975111, + 0.01648961380124092, + 0.005670059937983751, + 0.004995095077902079, + 0.013190463185310364, + 0.015312165021896362, + 0.004396332893520594, + 0.006203120574355125, + 0.0038346480578184128, + 0.023396525532007217, + 0.004626929294317961, + 0.004998920951038599, + 0.013124431483447552, + 0.01827704906463623, + 0.006613493897020817, + 0.0035493099130690098, + 0.004973673727363348, + 0.00767144002020359, + 0.0038042617961764336, + 0.004382925573736429, + 0.010942401364445686, + 0.012765109539031982, + 0.0028518864419311285, + 0.0024916573893278837, + 0.0029943750705569983, + 0.005633045919239521, + 0.0028004865162074566, + 0.002085968619212508, + 0.004593800287693739, + 0.005267285276204348, + 0.014150612987577915, + 0.00862938817590475, + 0.009249182417988777, + 0.028442727401852608, + 0.010532139800488949, + 0.008417697623372078, + 0.04923452064394951, + 0.025824476033449173, + 0.005953773856163025, + 0.0029842329677194357, + 0.004381986800581217, + 0.007610097527503967, + 0.003239591605961323, + 0.003997150808572769, + 0.01232085283845663, + 0.017735380679368973, + 0.008625129237771034, + 0.006202631630003452, + 0.009082193486392498, + 0.03928668424487114, + 0.006405336782336235, + 0.005155076272785664, + 0.039318423718214035, + 0.005905444268137217, + 0.014010163024067879, + 0.012748455628752708, + 0.01623416878283024, + 0.022486379370093346, + 0.01837293989956379, + 0.017699379473924637, + 0.022946378216147423, + 0.030264664441347122, + 0.0028933663852512836, + 0.004940789192914963, + 0.00376315345056355, + 0.007251271512359381, + 0.00969682913273573, + 0.004476227331906557, + 0.005571257788687944, + 0.00350934942252934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.006737373303622007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the emptiness of a late-night grocery store aisle. The floor, made of highly polished white linoleum, reflects the cold, buzzing green-white glare of the overhead fluorescent lights, creating a sterile and unsettling environment. Towering shelves line the aisles, perfectly stocked with generic, unlabeled boxes and cans, casting long, dark shadows that stretch into the deep background. In the very center of the frame sits a lone, slightly worn metal shopping cart, its chrome wire basket gleaming in the harsh light, its red plastic handle faded. Initially completely still, the cart begins to tremble slightly, a faint vibration that develops into a violent, convulsive shudder. Its entire metal frame wobbles unsteadily, and its small, black wheels begin to rattle violently and loudly against the polished floor. The sharp, rattling sound echoes uncomfortably throughout the desolate space, creating a tense and deeply unsettling atmosphere of unseen forces at work.", + "seed": 10784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_767_seed_10784_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09192823069170117, + "vlm_logprob": -0.09192823069170117, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.09192823069170117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03961421921849251 + ], + "opposite_logprob_tokens": [ + -3.414614200592041 + ], + "logprob_reward": -0.03961421921849251, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03961421921849251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30350998044013977, + 0.048890914767980576, + 0.03015735000371933, + 0.03954889252781868, + 0.0349467471241951, + 0.06235787644982338, + 0.15400460362434387, + 0.05471659079194069, + 0.06317377090454102, + 0.2086932361125946 + ], + "credit_map": [ + 0.039668504148721695, + 0.0679270476102829, + 0.03478144109249115, + 0.03498373553156853, + 0.03226906806230545, + 0.030499741435050964, + 0.027216056361794472, + 0.036164380609989166, + 0.006428187247365713, + 0.008378679864108562, + 0.0061067622154951096, + 0.006961044855415821, + 0.005891994573175907, + 0.0043751937337219715, + 0.0052552069537341595, + 0.005493844859302044, + 0.0033356754574924707, + 0.005856866016983986, + 0.0038703421596437693, + 0.003612743690609932, + 0.003483153646811843, + 0.002328674541786313, + 0.003824772546067834, + 0.003845121944323182, + 0.003422460285946727, + 0.004978744778782129, + 0.005182019434869289, + 0.006178003735840321, + 0.003973003011196852, + 0.0029597526881843805, + 0.00735511677339673, + 0.005499789025634527, + 0.003385810414329171, + 0.004963092040270567, + 0.004197410307824612, + 0.0038012550212442875, + 0.008204159326851368, + 0.002524455077946186, + 0.004519996233284473, + 0.0033505673054605722, + 0.00603359704837203, + 0.006540579721331596, + 0.008292867802083492, + 0.010589457117021084, + 0.008136202581226826, + 0.00551565270870924, + 0.00876699760556221, + 0.008482521399855614, + 0.018557948991656303, + 0.02197987213730812, + 0.01825808919966221, + 0.022777575999498367, + 0.020002201199531555, + 0.011856972239911556, + 0.02166512794792652, + 0.01890682801604271, + 0.005567867308855057, + 0.006434895098209381, + 0.007581475656479597, + 0.007859451696276665, + 0.01020023226737976, + 0.0046982779167592525, + 0.007006548345088959, + 0.005367845296859741, + 0.00697257649153471, + 0.008195274509489536, + 0.008222643285989761, + 0.008598345331847668, + 0.009042545221745968, + 0.00464596739038825, + 0.011125832796096802, + 0.006370586808770895, + 0.012330419383943081, + 0.03683827444911003, + 0.04934956878423691, + 0.01851564832031727, + 0.02169315330684185, + 0.02886434830725193, + 0.01306237280368805, + 0.028039459139108658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.03961421921849251 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422632671892643 + ], + "opposite_logprob_tokens": [ + -4.26422643661499 + ], + "logprob_reward": -0.01422632671892643, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422632671892643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12352262437343597, + 0.052692171186208725, + 0.031167635694146156, + 0.01417996920645237, + 0.04086344316601753, + 0.10736874490976334, + 0.07698960602283478, + 0.04425755888223648, + 0.1247752457857132, + 0.38418295979499817 + ], + "credit_map": [ + 0.01470906101167202, + 0.016470149159431458, + 0.021471425890922546, + 0.01928054168820381, + 0.015856817364692688, + 0.009078738279640675, + 0.016048328951001167, + 0.01060757227241993, + 0.006470552645623684, + 0.0067991032265126705, + 0.008900200016796589, + 0.007574396673589945, + 0.006783195771276951, + 0.004567967727780342, + 0.006093658972531557, + 0.0055031017400324345, + 0.002928153146058321, + 0.003573151072487235, + 0.004626684356480837, + 0.005967388860881329, + 0.0038009423296898603, + 0.001955232582986355, + 0.005389107391238213, + 0.0029269768856465816, + 0.001276021939702332, + 0.0022215694189071655, + 0.0018922419985756278, + 0.0019841070752590895, + 0.0019841298926621675, + 0.0012896929401904345, + 0.0018212905852124095, + 0.0017109165200963616, + 0.0033254441805183887, + 0.005836721044033766, + 0.004114401992410421, + 0.007420136593282223, + 0.004913210403174162, + 0.0027721100486814976, + 0.008444047532975674, + 0.0040373774245381355, + 0.006966554559767246, + 0.012590954080224037, + 0.011933739297091961, + 0.023402245715260506, + 0.009776131249964237, + 0.006849323865026236, + 0.02537522464990616, + 0.01047457940876484, + 0.005946039222180843, + 0.012669491581618786, + 0.01019451953470707, + 0.014293267391622066, + 0.00684020621702075, + 0.003938241861760616, + 0.017500100657343864, + 0.00560775026679039, + 0.003572283312678337, + 0.006251636426895857, + 0.005078731570392847, + 0.008277680724859238, + 0.006192635744810104, + 0.003370732069015503, + 0.007709282450377941, + 0.0038045793771743774, + 0.01022258773446083, + 0.020982416346669197, + 0.014057882130146027, + 0.021858032792806625, + 0.013938779011368752, + 0.006736616138368845, + 0.027294639497995377, + 0.009684296324849129, + 0.01838500052690506, + 0.1020587682723999, + 0.02067192830145359, + 0.029495850205421448, + 0.03779825195670128, + 0.047967396676540375, + 0.02001047506928444, + 0.10779532045125961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.01422632671892643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639123592525721 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639123592525721, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639123592525721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0774223729968071, + 0.018400441855192184, + 0.017791662365198135, + 0.02315131202340126, + 0.021993931382894516, + 0.018541032448410988, + 0.05429726094007492, + 0.021439900621771812, + 0.0670938789844513, + 0.6798682808876038 + ], + "credit_map": [ + 0.004714182578027248, + 0.010425474494695663, + 0.011267771944403648, + 0.01177612878382206, + 0.010442364029586315, + 0.005256920587271452, + 0.013453519903123379, + 0.01008601114153862, + 0.001673107035458088, + 0.002306103240698576, + 0.0027333369944244623, + 0.002939701545983553, + 0.0019909341353923082, + 0.0016576345078647137, + 0.0030472525395452976, + 0.002052370924502611, + 0.0017517461674287915, + 0.002127818763256073, + 0.0027130243834108114, + 0.003277997486293316, + 0.0019402066245675087, + 0.0013516542967408895, + 0.0029680749867111444, + 0.0016611400060355663, + 0.001956119667738676, + 0.0021633619908243418, + 0.0030257992912083864, + 0.004127432592213154, + 0.002297781640663743, + 0.0019740446005016565, + 0.004767176229506731, + 0.002839594380930066, + 0.002018026076257229, + 0.002546505071222782, + 0.0025487260427325964, + 0.003821413964033127, + 0.0036746468394994736, + 0.0015499363653361797, + 0.002561591099947691, + 0.00327308545820415, + 0.0022046673111617565, + 0.0033005159348249435, + 0.0020104048307985067, + 0.0024508836213499308, + 0.0017590420320630074, + 0.0018223576480522752, + 0.002945750718936324, + 0.0020474097691476345, + 0.008635777980089188, + 0.006244454998522997, + 0.006526048760861158, + 0.007756202947348356, + 0.005699578206986189, + 0.004609760828316212, + 0.009027770720422268, + 0.0057976641692221165, + 0.001822226564399898, + 0.0021278122439980507, + 0.003439258551225066, + 0.003590591484680772, + 0.003516069147735834, + 0.0015022007282823324, + 0.0032656227704137564, + 0.0021761185489594936, + 0.008476627990603447, + 0.007986876182258129, + 0.007594426162540913, + 0.010579830035567284, + 0.008159605786204338, + 0.00581350876018405, + 0.009864848107099533, + 0.008618155494332314, + 0.034689079970121384, + 0.22148019075393677, + 0.0762583315372467, + 0.08184730261564255, + 0.0670967772603035, + 0.1355908215045929, + 0.025804227218031883, + 0.03710150718688965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.007639123592525721 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870834410190582 + ], + "opposite_logprob_tokens": [ + -1.1370834112167358 + ], + "logprob_reward": -0.3870834410190582, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870834410190582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04779234528541565, + 0.028431259095668793, + 0.017156576737761497, + 0.011747035197913647, + 0.00780909787863493, + 0.0067488583736121655, + 0.035683706402778625, + 0.010924038477241993, + 0.018148556351661682, + 0.8155584931373596 + ], + "credit_map": [ + 0.0036880483385175467, + 0.006148314103484154, + 0.005354774184525013, + 0.008316892199218273, + 0.00669987965375185, + 0.0034826670307666063, + 0.009594546630978584, + 0.004507227335125208, + 0.002348351990804076, + 0.0026024929247796535, + 0.0037553845904767513, + 0.0052468376234173775, + 0.002931906608864665, + 0.0021602732595056295, + 0.006299885921180248, + 0.0030861282721161842, + 0.001448573311790824, + 0.001691194367595017, + 0.0022299648262560368, + 0.0032282427418977022, + 0.0017498595407232642, + 0.0013417098671197891, + 0.0033991874661296606, + 0.002067845780402422, + 0.0009666040423326194, + 0.0011698168236762285, + 0.0016211058245971799, + 0.0021625575609505177, + 0.0012137414887547493, + 0.0009860600112006068, + 0.0023294223938137293, + 0.0012977266451343894, + 0.0006520770839415491, + 0.0008104183943942189, + 0.001005127327516675, + 0.0012667360715568066, + 0.0015355694340541959, + 0.0005850756424479187, + 0.0008502581622451544, + 0.0011038370430469513, + 0.0007732359226793051, + 0.0008189028012566268, + 0.0009660052019171417, + 0.0009767343290150166, + 0.0008720536134205759, + 0.000699471915140748, + 0.000668071152176708, + 0.0009743836126290262, + 0.0030238195322453976, + 0.005671303253620863, + 0.00504162535071373, + 0.007209529168903828, + 0.0030542085878551006, + 0.0017974479123950005, + 0.007166578434407711, + 0.0027191946282982826, + 0.001050072954967618, + 0.0010778651339933276, + 0.0013070636196061969, + 0.0019764467142522335, + 0.0015529763186350465, + 0.0009771492332220078, + 0.0019277583342045546, + 0.001054706983268261, + 0.0019186374265700579, + 0.0026311061810702085, + 0.002158713061362505, + 0.0028127445839345455, + 0.0022008807864040136, + 0.0009218164486810565, + 0.003494933946058154, + 0.002009725896641612, + 0.0426974892616272, + 0.2925804853439331, + 0.05052246153354645, + 0.10432575643062592, + 0.08227837085723877, + 0.15747404098510742, + 0.029082562774419785, + 0.056597352027893066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.3870834410190582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078042909502983 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.011078042909502983, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078042909502983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19615446031093597, + 0.0378272719681263, + 0.02979099377989769, + 0.023865893483161926, + 0.03695744648575783, + 0.02280556783080101, + 0.04488464072346687, + 0.03531422093510628, + 0.03594556450843811, + 0.5364539623260498 + ], + "credit_map": [ + 0.020176328718662262, + 0.026064855977892876, + 0.032069552689790726, + 0.03132053464651108, + 0.01719844527542591, + 0.013552493415772915, + 0.034451957792043686, + 0.021320302039384842, + 0.0045574018731713295, + 0.004691488109529018, + 0.006177640985697508, + 0.005853601731359959, + 0.005314202513545752, + 0.0028041978366672993, + 0.005289569962769747, + 0.003139167558401823, + 0.0026848737616091967, + 0.003812044160440564, + 0.004447453189641237, + 0.005155827850103378, + 0.003029256360605359, + 0.0016231891931965947, + 0.006742059253156185, + 0.0022962891962379217, + 0.0021348160225898027, + 0.0026551890186965466, + 0.0029109755996614695, + 0.003714973106980324, + 0.004433745052665472, + 0.0018647079123184085, + 0.0033011455088853836, + 0.0028503411449491978, + 0.0034868046641349792, + 0.004190093372017145, + 0.005301185417920351, + 0.005588048603385687, + 0.006828377489000559, + 0.0020730113610625267, + 0.006479096598923206, + 0.003010825254023075, + 0.002177223563194275, + 0.002701090881600976, + 0.0026529915630817413, + 0.0034872384276241064, + 0.0036778850480914116, + 0.0021277274936437607, + 0.0025865354109555483, + 0.0033948749769479036, + 0.003485843539237976, + 0.006200186908245087, + 0.0051248567178845406, + 0.00844939798116684, + 0.0038952084723860025, + 0.0029112279880791903, + 0.011701356619596481, + 0.0031165629625320435, + 0.00594584783539176, + 0.004479145631194115, + 0.00596975302323699, + 0.003700637025758624, + 0.0033838378731161356, + 0.0025827912613749504, + 0.003117682645097375, + 0.0061345272697508335, + 0.004380918573588133, + 0.00485253008082509, + 0.004474313464015722, + 0.00501166982576251, + 0.006273204926401377, + 0.003646178636699915, + 0.00415801489725709, + 0.003148732939735055, + 0.027652891352772713, + 0.12911570072174072, + 0.09787969291210175, + 0.059463467448949814, + 0.04651861637830734, + 0.1194252073764801, + 0.022987712174654007, + 0.03341062366962433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.011078042909502983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures the emptiness of a late-night grocery store aisle. The floor, made of highly polished white linoleum, reflects the cold, buzzing green-white glare of the overhead fluorescent lights, creating a sterile and unsettling environment. Towering shelves line the aisles, perfectly stocked with generic, unlabeled boxes and cans, casting long, dark shadows that stretch into the deep background. In the very center of the frame sits a lone, slightly worn metal shopping cart, its chrome wire basket gleaming in the harsh light, its red plastic handle faded. Initially completely still, the cart begins to tremble slightly, a faint vibration that develops into a violent, convulsive shudder. Its entire metal frame wobbles unsteadily, and its small, black wheels begin to rattle violently and loudly against the polished floor. The sharp, rattling sound echoes uncomfortably throughout the desolate space, creating a tense and deeply unsettling atmosphere of unseen forces at work.", + "seed": 10785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_14/index_767_seed_10785_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12044811528176069, + "vlm_logprob": -0.12044811528176069, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.12044811528176069, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07454627752304077 + ], + "opposite_logprob_tokens": [ + -2.6995463371276855 + ], + "logprob_reward": -0.07454627752304077, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07454627752304077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1937706470489502, + 0.12012825161218643, + 0.030198000371456146, + 0.04406506568193436, + 0.03466268628835678, + 0.036793302744627, + 0.05279793217778206, + 0.021359536796808243, + 0.037737999111413956, + 0.42848649621009827 + ], + "credit_map": [ + 0.04046856239438057, + 0.020790209993720055, + 0.02826111949980259, + 0.018200958147644997, + 0.02541969157755375, + 0.019541222602128983, + 0.02076948992908001, + 0.020319383591413498, + 0.038454074412584305, + 0.009586402215063572, + 0.013703779317438602, + 0.006407565902918577, + 0.0075037237256765366, + 0.008375594392418861, + 0.008585309609770775, + 0.027511801570653915, + 0.0030922754667699337, + 0.004565026611089706, + 0.0046674818731844425, + 0.0025580227375030518, + 0.00450829928740859, + 0.002841007662937045, + 0.004332156851887703, + 0.0036337277851998806, + 0.005166139919310808, + 0.0026624861638993025, + 0.00932582188397646, + 0.004459341522306204, + 0.004087003413587809, + 0.0030223361682146788, + 0.004483044613152742, + 0.010858896188437939, + 0.0038403652142733335, + 0.0034059337340295315, + 0.005144746508449316, + 0.002884348388761282, + 0.004305981565266848, + 0.0036283486988395452, + 0.0035575509537011385, + 0.007895410992205143, + 0.0037580253556370735, + 0.0033154741395264864, + 0.006438753567636013, + 0.0044365208595991135, + 0.0047083767130970955, + 0.0032084903214126825, + 0.0043974388390779495, + 0.0065302252769470215, + 0.008323402144014835, + 0.005283951759338379, + 0.005094441119581461, + 0.007006375584751368, + 0.006324772723019123, + 0.004892840515822172, + 0.006548155564814806, + 0.009323990903794765, + 0.002694885479286313, + 0.0014244085177779198, + 0.0034091812558472157, + 0.0019697125535458326, + 0.0026346570812165737, + 0.002126481384038925, + 0.0021159304305911064, + 0.004984278231859207, + 0.009436455555260181, + 0.0027103261090815067, + 0.0063977958634495735, + 0.005660288501530886, + 0.0031784342136234045, + 0.0027883725706487894, + 0.0040135751478374004, + 0.003552752546966076, + 0.03366407752037048, + 0.024864351376891136, + 0.0175570510327816, + 0.16914089024066925, + 0.027312900871038437, + 0.03973424807190895, + 0.056654732674360275, + 0.05955827608704567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.07454627752304077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371550515294075 + ], + "opposite_logprob_tokens": [ + -3.4087154865264893 + ], + "logprob_reward": -0.03371550515294075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371550515294075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26425302028656006, + 0.16260944306850433, + 0.0896468311548233, + 0.04408479109406471, + 0.021756300702691078, + 0.09470007568597794, + 0.02255362831056118, + 0.02533000148832798, + 0.036392200738191605, + 0.23867379128932953 + ], + "credit_map": [ + 0.08511310815811157, + 0.019024336710572243, + 0.03883485496044159, + 0.01732884719967842, + 0.019198725000023842, + 0.019816547632217407, + 0.024076160043478012, + 0.04086044430732727, + 0.05636049807071686, + 0.00947424303740263, + 0.02053147181868553, + 0.007625148631632328, + 0.008271614089608192, + 0.010551421903073788, + 0.0100082503631711, + 0.039786793291568756, + 0.03128668665885925, + 0.005582829937338829, + 0.011124121025204659, + 0.004030565731227398, + 0.005201464518904686, + 0.005991487298160791, + 0.005839263554662466, + 0.020590418949723244, + 0.00812667328864336, + 0.0029116154182702303, + 0.006566298194229603, + 0.0038113535847514868, + 0.0037722259294241667, + 0.003427286632359028, + 0.0035854021552950144, + 0.011883934959769249, + 0.002599973464384675, + 0.002787564415484667, + 0.003992307465523481, + 0.002536909654736519, + 0.0026535787619650364, + 0.002232664031907916, + 0.0025695948861539364, + 0.002383706858381629, + 0.017904818058013916, + 0.005919460207223892, + 0.017567124217748642, + 0.008063658140599728, + 0.007554374169558287, + 0.00531771732494235, + 0.006301005836576223, + 0.02607192099094391, + 0.0052306195721030235, + 0.0018301720265299082, + 0.0022430175449699163, + 0.0018418347463011742, + 0.002464433666318655, + 0.0017529797041788697, + 0.0018080235458910465, + 0.005382546689361334, + 0.004478029441088438, + 0.002414897782728076, + 0.0040324074216187, + 0.0022190145682543516, + 0.0028114719316363335, + 0.001890511717647314, + 0.0023886235430836678, + 0.005095046013593674, + 0.009099573828279972, + 0.003152637043967843, + 0.004971098154783249, + 0.002901222091168165, + 0.0037265613209456205, + 0.0024658069014549255, + 0.002824153983965516, + 0.007251149974763393, + 0.017228296026587486, + 0.015849370509386063, + 0.010481145232915878, + 0.09154403209686279, + 0.014378984458744526, + 0.020629359409213066, + 0.03463616967201233, + 0.03392643481492996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03371550515294075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077570728957653 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077570728957653, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077570728957653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19021190702915192, + 0.22934867441654205, + 0.10266327857971191, + 0.045251645147800446, + 0.01902458630502224, + 0.06554671376943588, + 0.0494513064622879, + 0.01499693188816309, + 0.025521384552121162, + 0.25798359513282776 + ], + "credit_map": [ + 0.028698135167360306, + 0.018976079300045967, + 0.02550097554922104, + 0.018935300409793854, + 0.01978270336985588, + 0.015148229897022247, + 0.019138207659125328, + 0.044032275676727295, + 0.07881393283605576, + 0.013273322954773903, + 0.028087157756090164, + 0.01062999852001667, + 0.0121378842741251, + 0.015136796049773693, + 0.014892545528709888, + 0.05637703090906143, + 0.03330526873469353, + 0.006341714411973953, + 0.012344060465693474, + 0.004980777855962515, + 0.007609756663441658, + 0.0077170575968921185, + 0.007518326863646507, + 0.02284630760550499, + 0.011879902333021164, + 0.002758494345471263, + 0.0038114027120172977, + 0.0028960383497178555, + 0.003514024894684553, + 0.0034641725942492485, + 0.003644452430307865, + 0.013283156789839268, + 0.002360440557822585, + 0.0021610732655972242, + 0.0026265138294547796, + 0.001751032192260027, + 0.0026780064217746258, + 0.0018479392165318131, + 0.002517741173505783, + 0.0030818390659987926, + 0.013395520858466625, + 0.0050832368433475494, + 0.015216858126223087, + 0.006070015486329794, + 0.006596237886697054, + 0.005307287909090519, + 0.006953578907996416, + 0.006923968438059092, + 0.009521463885903358, + 0.0034357584081590176, + 0.00601968914270401, + 0.005033387336879969, + 0.004287057090550661, + 0.0035879947245121002, + 0.004660939332097769, + 0.01290501281619072, + 0.0027646964881569147, + 0.0016722272848710418, + 0.0018821746343746781, + 0.0012303260155022144, + 0.0016117370687425137, + 0.0011818392667919397, + 0.001991085708141327, + 0.0026628451887518167, + 0.006055319681763649, + 0.003157402854412794, + 0.0030080494470894337, + 0.001965930685400963, + 0.002959386445581913, + 0.0024894345551729202, + 0.0023570542689412832, + 0.003528805449604988, + 0.017491400241851807, + 0.017588753253221512, + 0.010219932533800602, + 0.10367563366889954, + 0.015314435586333275, + 0.020016290247440338, + 0.04194028303027153, + 0.031736865639686584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.011077570728957653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742656648159027 + ], + "opposite_logprob_tokens": [ + -0.9742656946182251 + ], + "logprob_reward": -0.4742656648159027, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742656648159027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23096638917922974, + 0.11950128525495529, + 0.036801550537347794, + 0.07393094152212143, + 0.022554129362106323, + 0.06267663091421127, + 0.06393497437238693, + 0.015487782657146454, + 0.03456934168934822, + 0.33957695960998535 + ], + "credit_map": [ + 0.08248385787010193, + 0.01570124551653862, + 0.04209711775183678, + 0.013113608583807945, + 0.0124009158462286, + 0.013331645168364048, + 0.014400923624634743, + 0.0374370813369751, + 0.04544976353645325, + 0.007039652671664953, + 0.017668426036834717, + 0.005527859553694725, + 0.004902975633740425, + 0.006773245986551046, + 0.006689779926091433, + 0.025449592620134354, + 0.009320128709077835, + 0.0028330599889159203, + 0.008120960555970669, + 0.002922262530773878, + 0.003999774344265461, + 0.0032731967512518167, + 0.002876178128644824, + 0.003455988597124815, + 0.026445044204592705, + 0.00499049061909318, + 0.008399056270718575, + 0.003560124896466732, + 0.003633908461779356, + 0.0044975923374295235, + 0.004092068877071142, + 0.018312659114599228, + 0.002389446599408984, + 0.0022008444648236036, + 0.0027293232269585133, + 0.00231773778796196, + 0.004447856452316046, + 0.0023875355254858732, + 0.0026448105927556753, + 0.003436577506363392, + 0.010355247184634209, + 0.004474459681659937, + 0.014128569513559341, + 0.006809350568801165, + 0.0062515875324606895, + 0.005006186198443174, + 0.005865083541721106, + 0.009786148555576801, + 0.018381845206022263, + 0.004719333723187447, + 0.0060752821154892445, + 0.0051182107999920845, + 0.00420005526393652, + 0.003961828537285328, + 0.004988840315490961, + 0.016489583998918533, + 0.0028364078607410192, + 0.000962933583650738, + 0.002425628714263439, + 0.0012279629008844495, + 0.0011053753551095724, + 0.001115923747420311, + 0.0010104042012244463, + 0.004803146701306105, + 0.009735879488289356, + 0.0021815760992467403, + 0.0023043211549520493, + 0.0023849671706557274, + 0.0024738425854593515, + 0.0022028200328350067, + 0.0023762346245348454, + 0.01090970728546381, + 0.022559190168976784, + 0.02147536538541317, + 0.013240509666502476, + 0.14091280102729797, + 0.020013203844428062, + 0.02792009897530079, + 0.05486050993204117, + 0.0385952927172184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.4742656648159027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an empty late-night grocery store aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of highly polished white linoleum reflecting green-white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone, slightly worn metal shopping cart with a red plastic handle visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shopping cart begin to tremble and shudder violently while its wheels rattle against the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635558187961578 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635558187961578, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635558187961578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22894687950611115, + 0.05809945613145828, + 0.22130417823791504, + 0.08930090069770813, + 0.03307514637708664, + 0.06791568547487259, + 0.05014365538954735, + 0.0520905926823616, + 0.09024827927350998, + 0.10887523740530014 + ], + "credit_map": [ + 0.0707857683300972, + 0.01807270385324955, + 0.03412381932139397, + 0.011531258933246136, + 0.015950169414281845, + 0.01594620756804943, + 0.01581556908786297, + 0.04672137275338173, + 0.006079505663365126, + 0.005273128394037485, + 0.012366586364805698, + 0.00666949525475502, + 0.005478380247950554, + 0.004625599831342697, + 0.005658646579831839, + 0.01194811426103115, + 0.0734245777130127, + 0.01592942513525486, + 0.018219545483589172, + 0.011413577012717724, + 0.010495856404304504, + 0.013200612738728523, + 0.012697353027760983, + 0.06592323631048203, + 0.018298866227269173, + 0.009643028490245342, + 0.008090133778750896, + 0.010671166703104973, + 0.009483976289629936, + 0.005960674490779638, + 0.00920338649302721, + 0.017949674278497696, + 0.0031605963595211506, + 0.0038577576633542776, + 0.0058504859916865826, + 0.0033778739161789417, + 0.0037227405700832605, + 0.002992358524352312, + 0.003561906749382615, + 0.006551425438374281, + 0.023185059428215027, + 0.0042554172687232494, + 0.014909732155501842, + 0.00496270228177309, + 0.004826673306524754, + 0.004805061034858227, + 0.004556473810225725, + 0.006414563395082951, + 0.010681386105716228, + 0.003642465453594923, + 0.004187766462564468, + 0.0049966685473918915, + 0.004881412256509066, + 0.003837188007310033, + 0.00439057033509016, + 0.013526197522878647, + 0.007624464109539986, + 0.005672436207532883, + 0.007099660113453865, + 0.00420319614931941, + 0.005398777313530445, + 0.0037638258654624224, + 0.005606541875749826, + 0.012721693143248558, + 0.010834812186658382, + 0.007872960530221462, + 0.010327025316655636, + 0.011023388244211674, + 0.009724387899041176, + 0.006294793915003538, + 0.010245881974697113, + 0.023925036191940308, + 0.011399801820516586, + 0.007510680705308914, + 0.007373468950390816, + 0.024382570758461952, + 0.010433689691126347, + 0.00997935701161623, + 0.011943168938159943, + 0.02585250698029995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.008635558187961578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_tensor([6735, 6735])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_tensor([6735, 6735])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a5c462ba41582affa789d8f038c8fad719a657a1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_tensor([6735, 6735])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning, static macro shot immerses the viewer in a scene of culinary artistry, focusing on a whole braised grouper presented in an exotic, high-end setting. The fish, its form intact from head to tail, rests on a minimalist, dark gray stone platter. Under a warm, focused spotlight, its skin gleams, its deep red soy glaze so smooth it reflects the light like polished varnish. Below the glaze, the flesh is pearlescent, delicate and tender, its jagged veins visible in minute detail. From just out of the frame, the curved bowl of a dark ceramic spoon enters the shot, steadily and deliberately tilted, pouring a fresh layer of rich sauce. The thick, syrupy liquid cascades down, creating slow, mesmerizing ripples that expand in concentric circles across the fish's shimmering surface and flow lazily along its sides. Set against a softly blurred background that suggests a luxurious dining setting, the entire scene evokes a powerful sense of sensual indulgence and dedicated craftsmanship, climaxing with the hypnotic movement of the sauce coating the pearly flesh before the visual sequence abruptly ends.", + "seed": 18145, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_6735_seed_18145_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04560407737735659, + "vlm_logprob": -0.04560407737735659, + "vlm_yes_no_margin": 4.000000023841858, + "vlm_token_credit": -0.04560407737735659, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0596538707613945 + ], + "opposite_logprob_tokens": [ + -3.0596537590026855 + ], + "logprob_reward": -0.0596538707613945, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0596538707613945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10590484738349915, + 0.08592073619365692, + 0.027766181156039238, + 0.02038806676864624, + 0.03715444356203079, + 0.2357242852449417, + 0.033841777592897415, + 0.020131466910243034, + 0.3619444668292999, + 0.07122369855642319 + ], + "credit_map": [ + 0.010649402625858784, + 0.007265719119459391, + 0.004501948133111, + 0.007287759333848953, + 0.007505680900067091, + 0.007508516777306795, + 0.04668721184134483, + 0.014498606324195862, + 0.015085089951753616, + 0.005456425715237856, + 0.002716873772442341, + 0.005238067824393511, + 0.0061081270687282085, + 0.0051569039933383465, + 0.03577436879277229, + 0.0103848772123456, + 0.005564917344599962, + 0.002341323299333453, + 0.0023831925354897976, + 0.002077514538541436, + 0.003045252989977598, + 0.0021856713574379683, + 0.0035618303809314966, + 0.006606479175388813, + 0.0023821000941097736, + 0.002246131421998143, + 0.001572284265421331, + 0.002552342601120472, + 0.0030346920248121023, + 0.0029931291937828064, + 0.0030067802872508764, + 0.002600606996566057, + 0.005813501309603453, + 0.002403754275292158, + 0.003457332029938698, + 0.0043504745699465275, + 0.00493974145501852, + 0.0046069868840277195, + 0.006906347349286079, + 0.004676307085901499, + 0.024651719257235527, + 0.01353470142930746, + 0.01388645265251398, + 0.02245916984975338, + 0.027241481468081474, + 0.01966608129441738, + 0.060991864651441574, + 0.05329282209277153, + 0.0018321109237149358, + 0.0009708640864118934, + 0.0012527721701189876, + 0.002057461068034172, + 0.0027144053019583225, + 0.002933282172307372, + 0.013573645614087582, + 0.008507235907018185, + 0.0026657998096197844, + 0.0018325593555346131, + 0.0013183294795453548, + 0.002092538168653846, + 0.0032885700929909945, + 0.0016180722741410136, + 0.004007526207715273, + 0.0033080715220421553, + 0.014666223898530006, + 0.00742311729118228, + 0.006858596578240395, + 0.011628330685198307, + 0.12113387882709503, + 0.020217111334204674, + 0.15911272168159485, + 0.020904501900076866, + 0.010083744302392006, + 0.005191582255065441, + 0.0028918730095028877, + 0.006645573768764734, + 0.012076483108103275, + 0.011956515721976757, + 0.012578895315527916, + 0.009799030609428883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.0596538707613945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279787242412567 + ], + "opposite_logprob_tokens": [ + -2.0177979469299316 + ], + "logprob_reward": -0.14279787242412567, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14279787242412567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2688526511192322, + 0.13639423251152039, + 0.08231634646654129, + 0.05597258731722832, + 0.023778926581144333, + 0.04703532159328461, + 0.18386119604110718, + 0.02033979631960392, + 0.13510958850383759, + 0.046339355409145355 + ], + "credit_map": [ + 0.050163354724645615, + 0.0184202641248703, + 0.03061761148273945, + 0.02578987367451191, + 0.019633963704109192, + 0.02571270614862442, + 0.06204679235816002, + 0.03646806254982948, + 0.03273487091064453, + 0.007138892076909542, + 0.008039194159209728, + 0.010614139959216118, + 0.009362692944705486, + 0.010606284253299236, + 0.044279277324676514, + 0.013618885539472103, + 0.014993807300925255, + 0.008975912816822529, + 0.007937428541481495, + 0.007532643619924784, + 0.010104990564286709, + 0.009192319586873055, + 0.00812873151153326, + 0.01545051857829094, + 0.01145084947347641, + 0.0057113696821033955, + 0.003641311312094331, + 0.0054549905471503735, + 0.008096004836261272, + 0.006288443226367235, + 0.007352064363658428, + 0.007977556437253952, + 0.0047758640721440315, + 0.0018896707333624363, + 0.0019098035991191864, + 0.0028123243246227503, + 0.003206235356628895, + 0.0033044854644685984, + 0.0028005684725940228, + 0.003079974092543125, + 0.007843263447284698, + 0.0042197443544864655, + 0.003581513185054064, + 0.005294031463563442, + 0.0061944229528307915, + 0.00807549525052309, + 0.005266641266644001, + 0.006560205947607756, + 0.0061537656001746655, + 0.006633981596678495, + 0.0036891784984618425, + 0.0064962743781507015, + 0.013220284134149551, + 0.009732455015182495, + 0.08682987838983536, + 0.05110538378357887, + 0.0031050979159772396, + 0.001680241315625608, + 0.0013309482019394636, + 0.0015289993025362492, + 0.0034879364538937807, + 0.0018962281756103039, + 0.0038536733482033014, + 0.0034566703252494335, + 0.011348729953169823, + 0.007443924434483051, + 0.005034708417952061, + 0.008440005593001842, + 0.0357651561498642, + 0.011136608198285103, + 0.046651650220155716, + 0.00928878877311945, + 0.004656351637095213, + 0.0036504543386399746, + 0.00240016239695251, + 0.0044449372217059135, + 0.008969255723059177, + 0.005590083077549934, + 0.009989061392843723, + 0.006639048457145691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.14279787242412567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084998026490211 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011084998026490211, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084998026490211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08112252503633499, + 0.038599785417318344, + 0.06262939423322678, + 0.029687220230698586, + 0.02656179666519165, + 0.06548724323511124, + 0.0954044908285141, + 0.06766863912343979, + 0.3298724591732025, + 0.20296649634838104 + ], + "credit_map": [ + 0.01620301604270935, + 0.004761624149978161, + 0.009631977416574955, + 0.008981310762465, + 0.009291051886975765, + 0.006576707120984793, + 0.015354294329881668, + 0.010322547517716885, + 0.00501669105142355, + 0.004161463584750891, + 0.0019302035216242075, + 0.0028651957400143147, + 0.004304693546146154, + 0.004895188380032778, + 0.008553056046366692, + 0.0068732937797904015, + 0.006838100031018257, + 0.005392284132540226, + 0.004691407084465027, + 0.006577143911272287, + 0.009532577358186245, + 0.009787828661501408, + 0.009922015480697155, + 0.009888042695820332, + 0.003911527339369059, + 0.0025655473582446575, + 0.00303953280672431, + 0.003996396437287331, + 0.003510520327836275, + 0.004092979710549116, + 0.0043118055909872055, + 0.004258910194039345, + 0.0043175108730793, + 0.0019571008160710335, + 0.002289296593517065, + 0.0032288182992488146, + 0.0027877960819751024, + 0.003641982562839985, + 0.0042835320346057415, + 0.004055761732161045, + 0.009981492534279823, + 0.005347718950361013, + 0.00588428508490324, + 0.005290857516229153, + 0.0075185722671449184, + 0.008167983964085579, + 0.013928615488111973, + 0.009367714636027813, + 0.004633474163711071, + 0.0042769345454871655, + 0.002436324954032898, + 0.004149453714489937, + 0.007102985866367817, + 0.009436668828129768, + 0.04028627276420593, + 0.02308238111436367, + 0.013578428886830807, + 0.0051996526308357716, + 0.0031799019780009985, + 0.004867764189839363, + 0.015078571625053883, + 0.004842431750148535, + 0.010857753455638885, + 0.010064131580293179, + 0.013465095311403275, + 0.0073697809129953384, + 0.0065655941143631935, + 0.011654278263449669, + 0.11006678640842438, + 0.018114008009433746, + 0.1444859653711319, + 0.018150940537452698, + 0.020244039595127106, + 0.021579084917902946, + 0.012439808808267117, + 0.01549746934324503, + 0.037748873233795166, + 0.045839227735996246, + 0.026804620400071144, + 0.022813349962234497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.011084998026490211 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542181648314 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542181648314, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542181648314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13355594873428345, + 0.09348685294389725, + 0.20277228951454163, + 0.027115944772958755, + 0.04515964910387993, + 0.08776237070560455, + 0.10025332123041153, + 0.042363446205854416, + 0.16405607759952545, + 0.10347400605678558 + ], + "credit_map": [ + 0.0201418437063694, + 0.011060092598199844, + 0.015579016879200935, + 0.013578448444604874, + 0.01625085063278675, + 0.018045861274003983, + 0.019259072840213776, + 0.01964077167212963, + 0.019940614700317383, + 0.007250407710671425, + 0.006385770160704851, + 0.00973652396351099, + 0.009340223856270313, + 0.007440734654664993, + 0.023261358961462975, + 0.010131212882697582, + 0.02570210210978985, + 0.01899753324687481, + 0.01605367660522461, + 0.023569248616695404, + 0.031135238707065582, + 0.025452956557273865, + 0.0352705679833889, + 0.02659096010029316, + 0.0031647731084376574, + 0.002918792422860861, + 0.0018915800610557199, + 0.003974841441959143, + 0.0038282277528196573, + 0.0034163068048655987, + 0.004420243203639984, + 0.0035011807922273874, + 0.007272524293512106, + 0.003898364957422018, + 0.0031948876567184925, + 0.0061318171210587025, + 0.007817351259291172, + 0.005582555662840605, + 0.006141205318272114, + 0.005120942834764719, + 0.009271782822906971, + 0.0064181177876889706, + 0.005654699169099331, + 0.008431380614638329, + 0.009479050524532795, + 0.021605035290122032, + 0.01471487432718277, + 0.012187435291707516, + 0.003788137575611472, + 0.004385271109640598, + 0.0023154201917350292, + 0.0036874974612146616, + 0.006945296656340361, + 0.006551028229296207, + 0.0460621602833271, + 0.026518506929278374, + 0.0051285154186189175, + 0.0028419559821486473, + 0.0015581371262669563, + 0.00298227253369987, + 0.01040854211896658, + 0.0050432621501386166, + 0.007740567903965712, + 0.006660195533186197, + 0.011126413010060787, + 0.006604465190321207, + 0.0056844973005354404, + 0.008094394579529762, + 0.04869181662797928, + 0.010411555878818035, + 0.062123432755470276, + 0.011319510638713837, + 0.013489863835275173, + 0.00827860552817583, + 0.0066649289801716805, + 0.01277368888258934, + 0.018611110746860504, + 0.02178347110748291, + 0.010572942905128002, + 0.011299402453005314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.012542181648314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019414640264585614 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019414640264585614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019414640264585614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3102656900882721, + 0.02739705704152584, + 0.04086681455373764, + 0.017451567575335503, + 0.019613798707723618, + 0.04482544958591461, + 0.3562581241130829, + 0.04849237576127052, + 0.11474602669477463, + 0.02008311077952385 + ], + "credit_map": [ + 0.05925748497247696, + 0.01435254979878664, + 0.012696556746959686, + 0.015126153826713562, + 0.024528583511710167, + 0.022048572078347206, + 0.13247406482696533, + 0.029781736433506012, + 0.004468251951038837, + 0.0018410889897495508, + 0.0021842315327376127, + 0.0021340521052479744, + 0.0026063674595206976, + 0.0039155264385044575, + 0.005929833743721247, + 0.004317706450819969, + 0.008772134780883789, + 0.003956547938287258, + 0.0027093589305877686, + 0.002913046395406127, + 0.007132772356271744, + 0.003894312772899866, + 0.004307165741920471, + 0.007181480526924133, + 0.0030826274305582047, + 0.0018897498957812786, + 0.0012592881685122848, + 0.0018364209681749344, + 0.002554110949859023, + 0.0022393646650016308, + 0.0022896300069987774, + 0.0023003758396953344, + 0.004380221478641033, + 0.001808827742934227, + 0.001651781378313899, + 0.0015968334628269076, + 0.0027184353675693274, + 0.0022395600099116564, + 0.0018730651354417205, + 0.0033450748305767775, + 0.004372944589704275, + 0.0030406974256038666, + 0.0019881967455148697, + 0.0027898673433810472, + 0.004161464050412178, + 0.005441966000944376, + 0.012912202626466751, + 0.010118113830685616, + 0.011452644132077694, + 0.013166004791855812, + 0.007341590244323015, + 0.012327779084444046, + 0.025964438915252686, + 0.018550505861639977, + 0.17035198211669922, + 0.09710317850112915, + 0.003452167846262455, + 0.005006993189454079, + 0.0012315585045143962, + 0.002590612042695284, + 0.010711057111620903, + 0.005637879949063063, + 0.011122211813926697, + 0.008739895187318325, + 0.0053075216710567474, + 0.0033297904301434755, + 0.002247102791443467, + 0.004128717351704836, + 0.038825538009405136, + 0.006148874759674072, + 0.04838874191045761, + 0.006369731854647398, + 0.0022047541569918394, + 0.0019551883451640606, + 0.0011323924409225583, + 0.0021886792965233326, + 0.0037006400525569916, + 0.0035103780683130026, + 0.002860855311155319, + 0.002530223922803998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0019414640264585614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, static macro shot immerses the viewer in a scene of culinary artistry, focusing on a whole braised grouper presented in an exotic, high-end setting. The fish, its form intact from head to tail, rests on a minimalist, dark gray stone platter. Under a warm, focused spotlight, its skin gleams, its deep red soy glaze so smooth it reflects the light like polished varnish. Below the glaze, the flesh is pearlescent, delicate and tender, its jagged veins visible in minute detail. From just out of the frame, the curved bowl of a dark ceramic spoon enters the shot, steadily and deliberately tilted, pouring a fresh layer of rich sauce. The thick, syrupy liquid cascades down, creating slow, mesmerizing ripples that expand in concentric circles across the fish's shimmering surface and flow lazily along its sides. Set against a softly blurred background that suggests a luxurious dining setting, the entire scene evokes a powerful sense of sensual indulgence and dedicated craftsmanship, climaxing with the hypnotic movement of the sauce coating the pearly flesh before the visual sequence abruptly ends.", + "seed": 18146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_6735_seed_18146_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07824778519570827, + "vlm_logprob": -0.07824778519570827, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.07824778519570827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08999072015285492 + ], + "opposite_logprob_tokens": [ + -2.5899906158447266 + ], + "logprob_reward": -0.08999072015285492, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08999072015285492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17855405807495117, + 0.2904497981071472, + 0.05381099134683609, + 0.03569730371236801, + 0.030552128329873085, + 0.09883853793144226, + 0.03182801976799965, + 0.04941831901669502, + 0.0303863063454628, + 0.20046457648277283 + ], + "credit_map": [ + 0.025343433022499084, + 0.02722250111401081, + 0.012790396809577942, + 0.01594967022538185, + 0.02809472382068634, + 0.02818010002374649, + 0.019037360325455666, + 0.02193586342036724, + 0.017392458394169807, + 0.0972653478384018, + 0.007732650265097618, + 0.014529142528772354, + 0.025544559583067894, + 0.09950922429561615, + 0.016414405778050423, + 0.012061982415616512, + 0.007722853217273951, + 0.005912576802074909, + 0.00966794416308403, + 0.003823433071374893, + 0.0062294602394104, + 0.009431344456970692, + 0.004694622941315174, + 0.006328747607767582, + 0.003400499699637294, + 0.004011569079011679, + 0.003137106541544199, + 0.0032672863453626633, + 0.007639954797923565, + 0.006337177939713001, + 0.004463541321456432, + 0.0034401612356305122, + 0.0038646026514470577, + 0.0037680682726204395, + 0.0019796111155301332, + 0.0036670167464762926, + 0.006304686423391104, + 0.004249948542565107, + 0.003554492723196745, + 0.0031636981293559074, + 0.01600940339267254, + 0.01956327073276043, + 0.007928089238703251, + 0.0076825679279863834, + 0.011605216190218925, + 0.0118626169860363, + 0.01223788969218731, + 0.011949467472732067, + 0.003526658983901143, + 0.0046214149333536625, + 0.002392431953921914, + 0.004432323854416609, + 0.004257578402757645, + 0.003656212007626891, + 0.00517529109492898, + 0.003766106441617012, + 0.0044949897564947605, + 0.004288701806217432, + 0.004388380795717239, + 0.004185302648693323, + 0.011590276844799519, + 0.0065818894654512405, + 0.005638042464852333, + 0.008250733837485313, + 0.002502980176359415, + 0.006032878067344427, + 0.0021094477269798517, + 0.0016883527860045433, + 0.007476406171917915, + 0.004563210066407919, + 0.002801252994686365, + 0.003211773931980133, + 0.015877658501267433, + 0.015190120786428452, + 0.013101556338369846, + 0.026977431029081345, + 0.02062358893454075, + 0.022677816450595856, + 0.05413871258497238, + 0.03187766298651695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.08999072015285492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520960867404938 + ], + "opposite_logprob_tokens": [ + -1.5020960569381714 + ], + "logprob_reward": -0.2520960867404938, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520960867404938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1260070651769638, + 0.15844368934631348, + 0.08536850661039352, + 0.059107743203639984, + 0.10220491886138916, + 0.08834289759397507, + 0.0701570063829422, + 0.06098706275224686, + 0.12204188853502274, + 0.12733925879001617 + ], + "credit_map": [ + 0.020531680434942245, + 0.018146680667996407, + 0.010366496630012989, + 0.012982131913304329, + 0.02257685549557209, + 0.01952894777059555, + 0.012991727329790592, + 0.008882555179297924, + 0.025397157296538353, + 0.016273878514766693, + 0.01684589684009552, + 0.015985412523150444, + 0.025306079536676407, + 0.025848137214779854, + 0.017878569662570953, + 0.014908560551702976, + 0.009220043197274208, + 0.0076997424475848675, + 0.01075035985559225, + 0.009303326718509197, + 0.014420675113797188, + 0.012297684326767921, + 0.011946121230721474, + 0.009730548597872257, + 0.005871197208762169, + 0.007725628092885017, + 0.005556467920541763, + 0.00452400604262948, + 0.012100903317332268, + 0.009493356570601463, + 0.007130470126867294, + 0.006705712527036667, + 0.00996125303208828, + 0.014298960566520691, + 0.006771234795451164, + 0.017714712768793106, + 0.014478471130132675, + 0.013000515289604664, + 0.012806492857635021, + 0.013173283077776432, + 0.008188406005501747, + 0.009604526683688164, + 0.006910054013133049, + 0.010882717557251453, + 0.01048178318887949, + 0.010165330953896046, + 0.017931032925844193, + 0.014179042540490627, + 0.00969610083848238, + 0.010206474922597408, + 0.005382695235311985, + 0.00681485328823328, + 0.012317823246121407, + 0.00883983913809061, + 0.00763519573956728, + 0.009264025837182999, + 0.013069386593997478, + 0.007133204024285078, + 0.0034846412017941475, + 0.005059855990111828, + 0.012019237503409386, + 0.008041255176067352, + 0.0057312482967972755, + 0.006448236294090748, + 0.014366229996085167, + 0.011260696686804295, + 0.011488018557429314, + 0.012471955269575119, + 0.017448434606194496, + 0.01781000755727291, + 0.016572456806898117, + 0.02062409557402134, + 0.01597452536225319, + 0.015596365556120872, + 0.008329039439558983, + 0.014640131033957005, + 0.023060392588377, + 0.019349265843629837, + 0.014190573245286942, + 0.016198961064219475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.2520960867404938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078924760222435 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078924760222435, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078924760222435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04759087786078453, + 0.4078030586242676, + 0.029892463237047195, + 0.029586708173155785, + 0.05335446447134018, + 0.2974761724472046, + 0.04735691472887993, + 0.00882072001695633, + 0.03538282960653305, + 0.04273570701479912 + ], + "credit_map": [ + 0.0077889892272651196, + 0.006719024386256933, + 0.0024651868734508753, + 0.003478017635643482, + 0.007365735713392496, + 0.006794983521103859, + 0.008542854338884354, + 0.0044360896572470665, + 0.021651126444339752, + 0.15042127668857574, + 0.007777170743793249, + 0.01938757114112377, + 0.03147295117378235, + 0.1523691564798355, + 0.012503047473728657, + 0.012220787815749645, + 0.004065286833792925, + 0.0029490068554878235, + 0.0040874965488910675, + 0.0025075313169509172, + 0.0041189044713974, + 0.004856066778302193, + 0.003412865102291107, + 0.003895306959748268, + 0.0027580871246755123, + 0.002491051098331809, + 0.0019126269035041332, + 0.0017648075008764863, + 0.006670511793345213, + 0.004922992549836636, + 0.0035157413221895695, + 0.005550890229642391, + 0.0057350643910467625, + 0.006898820865899324, + 0.0035259025171399117, + 0.003027440281584859, + 0.014132578857243061, + 0.008066114038228989, + 0.005060067865997553, + 0.006908477749675512, + 0.048605356365442276, + 0.07713490724563599, + 0.024245847016572952, + 0.020037272945046425, + 0.03939531743526459, + 0.04306473210453987, + 0.02248389832675457, + 0.02250884659588337, + 0.005448448937386274, + 0.0059999730437994, + 0.0045578801073133945, + 0.004345193039625883, + 0.010319162160158157, + 0.006852617487311363, + 0.0047765797935426235, + 0.005057063419371843, + 0.0013764039613306522, + 0.0009389426559209824, + 0.0008846352575346828, + 0.0009562553022988141, + 0.0008340057684108615, + 0.001216443837620318, + 0.0011222754837945104, + 0.0014917586231604218, + 0.0026841291692107916, + 0.007948343642055988, + 0.002973317401483655, + 0.002803976647555828, + 0.008266986347734928, + 0.004401621408760548, + 0.00218181568197906, + 0.004122644662857056, + 0.0032726547215133905, + 0.003454351332038641, + 0.003059323178604245, + 0.0044379401952028275, + 0.00599210150539875, + 0.005962887778878212, + 0.01044523622840643, + 0.006111215800046921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.016078924760222435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292848840355873 + ], + "opposite_logprob_tokens": [ + -3.7732927799224854 + ], + "logprob_reward": -0.023292848840355873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292848840355873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19358369708061218, + 0.1197907030582428, + 0.055968448519706726, + 0.06395618617534637, + 0.08659237623214722, + 0.10203465819358826, + 0.10250859707593918, + 0.07105422765016556, + 0.03307068347930908, + 0.17144042253494263 + ], + "credit_map": [ + 0.027622688561677933, + 0.03232704848051071, + 0.0126790227368474, + 0.016777964308857918, + 0.025786448270082474, + 0.027994295582175255, + 0.030817613005638123, + 0.01957862079143524, + 0.010744728147983551, + 0.03324767202138901, + 0.004090058151632547, + 0.007222057320177555, + 0.014337878674268723, + 0.0334598608314991, + 0.007172955200076103, + 0.009515494108200073, + 0.007976757362484932, + 0.005692740902304649, + 0.009441288188099861, + 0.0051268027164042, + 0.006294182036072016, + 0.008548595011234283, + 0.006725681480020285, + 0.006162404548376799, + 0.006840710993856192, + 0.005985992029309273, + 0.004340320825576782, + 0.0036407336592674255, + 0.016499804332852364, + 0.011518901214003563, + 0.0064529115334153175, + 0.008676806464791298, + 0.010167805477976799, + 0.009940427727997303, + 0.005866116378456354, + 0.006818158086389303, + 0.018578389659523964, + 0.012046732008457184, + 0.011011889204382896, + 0.012162866070866585, + 0.011417511850595474, + 0.01874697394669056, + 0.006088524591177702, + 0.009837324731051922, + 0.011835298500955105, + 0.015117895789444447, + 0.01678663305938244, + 0.012204495258629322, + 0.0086023835465312, + 0.01272927038371563, + 0.007426577154546976, + 0.011994251981377602, + 0.02323649823665619, + 0.015318896621465683, + 0.01371247973293066, + 0.009488238953053951, + 0.007010370958596468, + 0.009570439346134663, + 0.007832157425582409, + 0.009637914597988129, + 0.004984356462955475, + 0.008637427352368832, + 0.009601753205060959, + 0.013779806904494762, + 0.0025721611455082893, + 0.006716648116707802, + 0.0028698963578790426, + 0.0031728744506835938, + 0.006891191005706787, + 0.004086571745574474, + 0.0025471372064203024, + 0.004214203916490078, + 0.015528975985944271, + 0.016058776527643204, + 0.012560777366161346, + 0.027014905586838722, + 0.018047992140054703, + 0.02081977017223835, + 0.03452937677502632, + 0.026879865676164627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.023292848840355873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10791359096765518, + 0.3063660264015198, + 0.06286238133907318, + 0.05287182703614235, + 0.03173103556036949, + 0.20634889602661133, + 0.044771477580070496, + 0.04966045916080475, + 0.07732617110013962, + 0.06014819070696831 + ], + "credit_map": [ + 0.016874926164746284, + 0.01629793643951416, + 0.006824717856943607, + 0.013181467540562153, + 0.011049093678593636, + 0.014627254568040371, + 0.019002512097358704, + 0.010055671446025372, + 0.015665573999285698, + 0.11519118398427963, + 0.005449902266263962, + 0.011513891629874706, + 0.023514464497566223, + 0.11719562113285065, + 0.00913154985755682, + 0.00870380736887455, + 0.008234485983848572, + 0.006042397581040859, + 0.01272936724126339, + 0.005893244408071041, + 0.0074037592858076096, + 0.008326515555381775, + 0.0074477908201515675, + 0.006784820929169655, + 0.00505690136924386, + 0.004969172179698944, + 0.0027928000781685114, + 0.0030149188823997974, + 0.013674404472112656, + 0.01019203383475542, + 0.004732169210910797, + 0.008439420722424984, + 0.0030646738596260548, + 0.004121800884604454, + 0.0037189184222370386, + 0.0022209160961210728, + 0.005053447559475899, + 0.003593962173908949, + 0.005638085771352053, + 0.004319228231906891, + 0.03273537755012512, + 0.05019845813512802, + 0.01597418077290058, + 0.01472108718007803, + 0.026507116854190826, + 0.02906830422580242, + 0.01733441650867462, + 0.01980992965400219, + 0.0036119180731475353, + 0.00616027694195509, + 0.0032624048180878162, + 0.003922334406524897, + 0.011483128182590008, + 0.007115100044757128, + 0.0050047338008880615, + 0.004211581312119961, + 0.009676826186478138, + 0.006242557894438505, + 0.004936894401907921, + 0.004505757708102465, + 0.006579514592885971, + 0.007554329000413418, + 0.0035792160779237747, + 0.006585360504686832, + 0.0067240893840789795, + 0.007957153022289276, + 0.0055154492147266865, + 0.008608466945588589, + 0.01031953003257513, + 0.014269421808421612, + 0.009912691079080105, + 0.014019359834492207, + 0.00562835531309247, + 0.005946479272097349, + 0.003396386280655861, + 0.0073408628813922405, + 0.011095136404037476, + 0.01079560536891222, + 0.008448641747236252, + 0.007496719714254141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.009780345484614372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, static macro shot immerses the viewer in a scene of culinary artistry, focusing on a whole braised grouper presented in an exotic, high-end setting. The fish, its form intact from head to tail, rests on a minimalist, dark gray stone platter. Under a warm, focused spotlight, its skin gleams, its deep red soy glaze so smooth it reflects the light like polished varnish. Below the glaze, the flesh is pearlescent, delicate and tender, its jagged veins visible in minute detail. From just out of the frame, the curved bowl of a dark ceramic spoon enters the shot, steadily and deliberately tilted, pouring a fresh layer of rich sauce. The thick, syrupy liquid cascades down, creating slow, mesmerizing ripples that expand in concentric circles across the fish's shimmering surface and flow lazily along its sides. Set against a softly blurred background that suggests a luxurious dining setting, the entire scene evokes a powerful sense of sensual indulgence and dedicated craftsmanship, climaxing with the hypnotic movement of the sauce coating the pearly flesh before the visual sequence abruptly ends.", + "seed": 18147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_6735_seed_18147_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07848920784890652, + "vlm_logprob": -0.07848920784890652, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.07848920784890652, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0374426506459713 + ], + "opposite_logprob_tokens": [ + -3.537442684173584 + ], + "logprob_reward": -0.0374426506459713, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0374426506459713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24579985439777374, + 0.06765619665384293, + 0.05343922972679138, + 0.10181190818548203, + 0.11442892253398895, + 0.07508230954408646, + 0.05504674091935158, + 0.05313229188323021, + 0.1446371227502823, + 0.08896540105342865 + ], + "credit_map": [ + 0.048136621713638306, + 0.035142023116350174, + 0.023033320903778076, + 0.020883526653051376, + 0.03376318886876106, + 0.023266032338142395, + 0.04088909178972244, + 0.020686060190200806, + 0.004094902891665697, + 0.013737799599766731, + 0.005137465428560972, + 0.004513792227953672, + 0.010266070254147053, + 0.007361949421465397, + 0.0179348923265934, + 0.004609334282577038, + 0.00861883070319891, + 0.00442583067342639, + 0.006954172160476446, + 0.004619223531335592, + 0.008137155324220657, + 0.005114384926855564, + 0.008946992456912994, + 0.006622642278671265, + 0.017038114368915558, + 0.008764603175222874, + 0.01789993792772293, + 0.010294470936059952, + 0.013584702275693417, + 0.012591482140123844, + 0.012480497360229492, + 0.00915810652077198, + 0.015699118375778198, + 0.01486227661371231, + 0.013681191019713879, + 0.011184921488165855, + 0.016744041815400124, + 0.017216993495821953, + 0.014067078940570354, + 0.01097330916672945, + 0.011111358180642128, + 0.0093863345682621, + 0.008970262482762337, + 0.008083772845566273, + 0.010965283028781414, + 0.008170132525265217, + 0.012148904614150524, + 0.006246267352253199, + 0.010343533009290695, + 0.00639829458668828, + 0.007153413258492947, + 0.004757438786327839, + 0.010285357013344765, + 0.006863377057015896, + 0.005663137882947922, + 0.003582195146009326, + 0.007305908016860485, + 0.0063203782774508, + 0.00868294294923544, + 0.006224013399332762, + 0.005423796363174915, + 0.0057314191944897175, + 0.0068467543460428715, + 0.006597083993256092, + 0.018246663734316826, + 0.02300000935792923, + 0.014836287125945091, + 0.021114317700266838, + 0.019787022843956947, + 0.0158548541367054, + 0.015576368197798729, + 0.016221603378653526, + 0.018277091905474663, + 0.014429710805416107, + 0.011300786398351192, + 0.007705942261964083, + 0.010490662418305874, + 0.009110447019338608, + 0.011763479560613632, + 0.005887283477932215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0374426506459713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707126140594482 + ], + "opposite_logprob_tokens": [ + -2.1270713806152344 + ], + "logprob_reward": -0.12707126140594482, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707126140594482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15867312252521515, + 0.07989263534545898, + 0.054187923669815063, + 0.10087380558252335, + 0.2765699625015259, + 0.021742645651102066, + 0.06300996243953705, + 0.033583179116249084, + 0.055436648428440094, + 0.1560300886631012 + ], + "credit_map": [ + 0.03200329467654228, + 0.014435476623475552, + 0.020678022876381874, + 0.017597150057554245, + 0.021381568163633347, + 0.016407307237386703, + 0.022635139524936676, + 0.013535158708691597, + 0.007191282697021961, + 0.01713334023952484, + 0.006699186749756336, + 0.006459988187998533, + 0.005522352177649736, + 0.004441591911017895, + 0.026686156168580055, + 0.005758743733167648, + 0.011032923124730587, + 0.005214432720094919, + 0.005929103121161461, + 0.0033641206100583076, + 0.005117983091622591, + 0.0038953046314418316, + 0.007472819648683071, + 0.012161238119006157, + 0.013153351843357086, + 0.008272391743957996, + 0.015988532453775406, + 0.014530617743730545, + 0.013781988993287086, + 0.011929765343666077, + 0.013838684186339378, + 0.009378469549119473, + 0.06948493421077728, + 0.041809555143117905, + 0.030301399528980255, + 0.014798982068896294, + 0.04322267696261406, + 0.03417892009019852, + 0.02382292039692402, + 0.01895054429769516, + 0.0022624866105616093, + 0.002343440195545554, + 0.002341645071282983, + 0.0017845085822045803, + 0.006339827086776495, + 0.0024745012633502483, + 0.0020267991349101067, + 0.002169437939301133, + 0.014668031595647335, + 0.007660394534468651, + 0.009004545398056507, + 0.004977592267096043, + 0.012788821943104267, + 0.005867587868124247, + 0.0037819526623934507, + 0.004261034075170755, + 0.006064808927476406, + 0.0032558105885982513, + 0.004436455201357603, + 0.0033000472467392683, + 0.0057237050496041775, + 0.00420145271345973, + 0.003423411166295409, + 0.0031774896197021008, + 0.006862308364361525, + 0.005651337560266256, + 0.0059856511652469635, + 0.0057911621406674385, + 0.012560606002807617, + 0.008421098813414574, + 0.005021371878683567, + 0.0051431115716695786, + 0.04050969332456589, + 0.018727563321590424, + 0.018974458798766136, + 0.013003832660615444, + 0.01868532970547676, + 0.016444336622953415, + 0.019183414056897163, + 0.010501464828848839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.12707126140594482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993177473545074 + ], + "opposite_logprob_tokens": [ + -1.8049317598342896 + ], + "logprob_reward": -0.17993177473545074, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993177473545074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1649942398071289, + 0.13712187111377716, + 0.1376025527715683, + 0.10605672746896744, + 0.1105879619717598, + 0.035810552537441254, + 0.09489592909812927, + 0.05172153562307358, + 0.04630858078598976, + 0.11490005254745483 + ], + "credit_map": [ + 0.0254194512963295, + 0.022214647382497787, + 0.01771455444395542, + 0.022089766338467598, + 0.02094639092683792, + 0.01944689266383648, + 0.0245486032217741, + 0.012613925151526928, + 0.011296696029603481, + 0.02822164259850979, + 0.009857933036983013, + 0.011870094574987888, + 0.013153382577002048, + 0.010672548785805702, + 0.04068667069077492, + 0.011362908408045769, + 0.03728416562080383, + 0.014545449987053871, + 0.016008654609322548, + 0.013496983796358109, + 0.012365780770778656, + 0.008530579507350922, + 0.01659587025642395, + 0.01877507008612156, + 0.014718280173838139, + 0.009087488986551762, + 0.017919335514307022, + 0.011036088690161705, + 0.011361452750861645, + 0.014348102733492851, + 0.014299023896455765, + 0.01328695472329855, + 0.02882852405309677, + 0.015979956835508347, + 0.012831679545342922, + 0.008055135607719421, + 0.015329950489103794, + 0.014191619120538235, + 0.00859911646693945, + 0.0067719826474785805, + 0.004616950172930956, + 0.003721949178725481, + 0.006677205208688974, + 0.0033005643635988235, + 0.0069096689112484455, + 0.003968167118728161, + 0.004264426417648792, + 0.0023516234941780567, + 0.0223037451505661, + 0.011489094235002995, + 0.01135365478694439, + 0.006691218353807926, + 0.01734733395278454, + 0.01281097810715437, + 0.007568766828626394, + 0.005331139080226421, + 0.0075175222009420395, + 0.004700362216681242, + 0.0067818425595760345, + 0.006466380786150694, + 0.008224934339523315, + 0.0067026750184595585, + 0.006798825692385435, + 0.004528992809355259, + 0.005115585867315531, + 0.004704552236944437, + 0.004741563927382231, + 0.006594366393983364, + 0.00787245761603117, + 0.005769482348114252, + 0.007498661056160927, + 0.0040119122713804245, + 0.024296848103404045, + 0.014979391358792782, + 0.015937456861138344, + 0.009325377643108368, + 0.013052094727754593, + 0.011458449065685272, + 0.01608891971409321, + 0.00976150669157505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.17993177473545074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810355976223946 + ], + "opposite_logprob_tokens": [ + -3.5298104286193848 + ], + "logprob_reward": -0.029810355976223946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810355976223946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17695416510105133, + 0.10419823974370956, + 0.06626566499471664, + 0.07660049200057983, + 0.23712025582790375, + 0.051306676119565964, + 0.07692936807870865, + 0.04347030445933342, + 0.036797814071178436, + 0.13035701215267181 + ], + "credit_map": [ + 0.025707757100462914, + 0.022940224036574364, + 0.016456052660942078, + 0.016725672408938408, + 0.028723707422614098, + 0.0189389456063509, + 0.031589969992637634, + 0.01587183028459549, + 0.008013845421373844, + 0.02562767080962658, + 0.007551116868853569, + 0.007691826205700636, + 0.00709066865965724, + 0.00695255771279335, + 0.03498835489153862, + 0.0062821959145367146, + 0.013823254033923149, + 0.00582084571942687, + 0.0083360206335783, + 0.005988913122564554, + 0.006631302181631327, + 0.004638449754565954, + 0.013134031556546688, + 0.007892847061157227, + 0.011954429559409618, + 0.007111760787665844, + 0.010550879873335361, + 0.0059597729705274105, + 0.011600055731832981, + 0.011228828690946102, + 0.010191941633820534, + 0.008002816699445248, + 0.07041700929403305, + 0.034744665026664734, + 0.028864827007055283, + 0.011868485249578953, + 0.0314311608672142, + 0.030917787924408913, + 0.01762789487838745, + 0.01124842744320631, + 0.006778282579034567, + 0.005205611698329449, + 0.007681651506572962, + 0.004537881817668676, + 0.009654367342591286, + 0.006093330681324005, + 0.007570376619696617, + 0.0037851755041629076, + 0.018709711730480194, + 0.00963403470814228, + 0.010489776730537415, + 0.0058783371932804585, + 0.01181662455201149, + 0.01133769191801548, + 0.004919261205941439, + 0.004143929108977318, + 0.007225856650620699, + 0.005993752274662256, + 0.005833675619214773, + 0.004000566899776459, + 0.006001414731144905, + 0.004883807618170977, + 0.004757032264024019, + 0.004774199798703194, + 0.004425216931849718, + 0.003871674183756113, + 0.0036804238334298134, + 0.004025213420391083, + 0.0063177491538226604, + 0.006428136024624109, + 0.004589902237057686, + 0.003459499217569828, + 0.03753155097365379, + 0.01711556874215603, + 0.019273946061730385, + 0.01024384330958128, + 0.012782327830791473, + 0.014124283567070961, + 0.01188107393682003, + 0.007404427509754896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.029810355976223946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10251877456903458, + 0.07283265888690948, + 0.11305811256170273, + 0.30523863434791565, + 0.12034454196691513, + 0.02937954105436802, + 0.04698764905333519, + 0.12704996764659882, + 0.05577879026532173, + 0.026811307296156883 + ], + "credit_map": [ + 0.01836078241467476, + 0.011387638747692108, + 0.011017831973731518, + 0.009774118661880493, + 0.013115821406245232, + 0.01239957194775343, + 0.015353311784565449, + 0.01110970601439476, + 0.00748447747901082, + 0.016149774193763733, + 0.006239497568458319, + 0.007786451373249292, + 0.009078862145543098, + 0.008994882926344872, + 0.009863153100013733, + 0.007235564757138491, + 0.030945658683776855, + 0.011314493604004383, + 0.014430750161409378, + 0.008313505910336971, + 0.009510733187198639, + 0.007259581238031387, + 0.010632088407874107, + 0.020651301369071007, + 0.03604305908083916, + 0.022335100919008255, + 0.04091641306877136, + 0.03753100708127022, + 0.039479587227106094, + 0.049566689878702164, + 0.03935599699616432, + 0.04001077264547348, + 0.031617265194654465, + 0.016418002545833588, + 0.015542371198534966, + 0.0064310128800570965, + 0.019577214494347572, + 0.01603204756975174, + 0.007288522552698851, + 0.007438106928020716, + 0.003951702266931534, + 0.0032380816992372274, + 0.004128649830818176, + 0.003362133400514722, + 0.006772481836378574, + 0.00320358294993639, + 0.002823687857016921, + 0.0018992215627804399, + 0.006777492351830006, + 0.00513344444334507, + 0.0063762483187019825, + 0.004005658905953169, + 0.009474308229982853, + 0.006298636086285114, + 0.004511117935180664, + 0.00441074138507247, + 0.01970718614757061, + 0.01416393555700779, + 0.014317180030047894, + 0.011467287316918373, + 0.027546998113393784, + 0.01661958359181881, + 0.01383538730442524, + 0.009392405860126019, + 0.006790895480662584, + 0.007743884343653917, + 0.0057303630746901035, + 0.00724228098988533, + 0.01086051668971777, + 0.006681892089545727, + 0.005018665920943022, + 0.005710290744900703, + 0.0038067668210715055, + 0.002739405957981944, + 0.002467287937179208, + 0.0027682569343596697, + 0.007494218181818724, + 0.003184203989803791, + 0.0029340763576328754, + 0.0014170907670632005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.018189996480941772 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, static macro shot immerses the viewer in a scene of culinary artistry, focusing on a whole braised grouper presented in an exotic, high-end setting. The fish, its form intact from head to tail, rests on a minimalist, dark gray stone platter. Under a warm, focused spotlight, its skin gleams, its deep red soy glaze so smooth it reflects the light like polished varnish. Below the glaze, the flesh is pearlescent, delicate and tender, its jagged veins visible in minute detail. From just out of the frame, the curved bowl of a dark ceramic spoon enters the shot, steadily and deliberately tilted, pouring a fresh layer of rich sauce. The thick, syrupy liquid cascades down, creating slow, mesmerizing ripples that expand in concentric circles across the fish's shimmering surface and flow lazily along its sides. Set against a softly blurred background that suggests a luxurious dining setting, the entire scene evokes a powerful sense of sensual indulgence and dedicated craftsmanship, climaxing with the hypnotic movement of the sauce coating the pearly flesh before the visual sequence abruptly ends.", + "seed": 18148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_15/index_6735_seed_18148_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05290274061262608, + "vlm_logprob": -0.05290274061262608, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.05290274061262608, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06962349265813828 + ], + "opposite_logprob_tokens": [ + -2.8196234703063965 + ], + "logprob_reward": -0.06962349265813828, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06962349265813828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09989365935325623, + 0.043028756976127625, + 0.26255542039871216, + 0.016457369551062584, + 0.02142656035721302, + 0.04604576900601387, + 0.2234477996826172, + 0.1852952539920807, + 0.08111422508955002, + 0.02073514461517334 + ], + "credit_map": [ + 0.016330987215042114, + 0.01311967521905899, + 0.010387725196778774, + 0.011116374284029007, + 0.017191268503665924, + 0.008414436131715775, + 0.011244885623455048, + 0.012088305316865444, + 0.004309515934437513, + 0.004611365031450987, + 0.0051418328657746315, + 0.005660675000399351, + 0.00993500929325819, + 0.004485278856009245, + 0.005293597001582384, + 0.0035914829932153225, + 0.018908068537712097, + 0.014871960505843163, + 0.06623674929141998, + 0.015088832937180996, + 0.10190775245428085, + 0.03020785190165043, + 0.008193837478756905, + 0.007140366826206446, + 0.002175087807700038, + 0.002368470886722207, + 0.002021623309701681, + 0.002186681842431426, + 0.0021077736746519804, + 0.001556298229843378, + 0.0017587083857506514, + 0.0022827263455837965, + 0.0028552482835948467, + 0.0031937793828547, + 0.0018983129411935806, + 0.0023644736502319574, + 0.00535730691626668, + 0.001961508532986045, + 0.0022191633470356464, + 0.0015767676522955298, + 0.012745707295835018, + 0.006595220882445574, + 0.004184412304311991, + 0.005119060631841421, + 0.0049166553653776646, + 0.0034766022581607103, + 0.003340156050398946, + 0.005667951423674822, + 0.03796538710594177, + 0.04955356568098068, + 0.014826945029199123, + 0.014983787201344967, + 0.052027225494384766, + 0.021494058892130852, + 0.011197904124855995, + 0.021398931741714478, + 0.060227230191230774, + 0.010320784524083138, + 0.007044597994536161, + 0.018244825303554535, + 0.054869841784238815, + 0.010598383843898773, + 0.015049485489726067, + 0.008940105326473713, + 0.014078138396143913, + 0.0061819422990083694, + 0.0039916010573506355, + 0.006934884004294872, + 0.027811789885163307, + 0.012958409264683723, + 0.0042397212237119675, + 0.004917740356177092, + 0.003987753298133612, + 0.0028679226525127888, + 0.0020131561905145645, + 0.0029812315478920937, + 0.0015869446797296405, + 0.002305686706677079, + 0.0016479295445606112, + 0.0033445185981690884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.06962349265813828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031970590353012 + ], + "opposite_logprob_tokens": [ + -2.3503196239471436 + ], + "logprob_reward": -0.10031970590353012, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031970590353012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18424144387245178, + 0.03399228677153587, + 0.19286799430847168, + 0.06337597966194153, + 0.1603633165359497, + 0.020184632390737534, + 0.08402560651302338, + 0.0585080161690712, + 0.17873769998550415, + 0.0237029567360878 + ], + "credit_map": [ + 0.02263355813920498, + 0.024434857070446014, + 0.024532729759812355, + 0.02392042987048626, + 0.01957232877612114, + 0.01845916360616684, + 0.02529909834265709, + 0.025389298796653748, + 0.006982715800404549, + 0.004737902898341417, + 0.004586212802678347, + 0.00330670946277678, + 0.004457708448171616, + 0.003812455339357257, + 0.0035666893236339092, + 0.0025418936274945736, + 0.016141364350914955, + 0.015009988099336624, + 0.047001853585243225, + 0.01093695592135191, + 0.0703229084610939, + 0.020630059763789177, + 0.006221705581992865, + 0.006603177171200514, + 0.010167754255235195, + 0.012830013409256935, + 0.0074342405423521996, + 0.007971511222422123, + 0.007807524874806404, + 0.005199651699513197, + 0.005202278029173613, + 0.006763012148439884, + 0.024678166955709457, + 0.021321367472410202, + 0.012336411513388157, + 0.016588838770985603, + 0.033553410321474075, + 0.01650420017540455, + 0.017653316259384155, + 0.01772761531174183, + 0.0059920065104961395, + 0.0030499882996082306, + 0.0014314798172563314, + 0.0017077381489798427, + 0.0023433994501829147, + 0.0014212847454473376, + 0.001399524393491447, + 0.002839213702827692, + 0.01568876951932907, + 0.019629579037427902, + 0.005490598268806934, + 0.007403737865388393, + 0.017894405871629715, + 0.008280569687485695, + 0.0036627287045121193, + 0.005975218955427408, + 0.015569579787552357, + 0.00398613978177309, + 0.0026418722700327635, + 0.007806722540408373, + 0.01696496456861496, + 0.003667691256850958, + 0.004952391143888235, + 0.0029186613392084837, + 0.04184625670313835, + 0.011953245848417282, + 0.005979020614176989, + 0.0070685213431715965, + 0.06801556050777435, + 0.024920126423239708, + 0.006888872012495995, + 0.012066123075783253, + 0.005533999297767878, + 0.002799784066155553, + 0.002641404513269663, + 0.0021741732489317656, + 0.002062183106318116, + 0.0032378521282225847, + 0.0020030844025313854, + 0.0032504808623343706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.10031970590353012 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498193949460983 + ], + "opposite_logprob_tokens": [ + -2.9299819469451904 + ], + "logprob_reward": -0.05498193949460983, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498193949460983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030912747606635094, + 0.020959265530109406, + 0.04130251705646515, + 0.008655387908220291, + 0.015079787001013756, + 0.03367926925420761, + 0.7737393975257874, + 0.05374026298522949, + 0.010183770209550858, + 0.01174758281558752 + ], + "credit_map": [ + 0.003807272994890809, + 0.0031797296833246946, + 0.004370738752186298, + 0.003481751075014472, + 0.004564197268337011, + 0.0037120007909834385, + 0.003555653616786003, + 0.004241407848894596, + 0.0030307690612971783, + 0.003205913119018078, + 0.0026536667719483376, + 0.0033376559149473906, + 0.003306852886453271, + 0.002104835817590356, + 0.0018330691382288933, + 0.0014865038683637977, + 0.0030970468651503325, + 0.0025781646836549044, + 0.010322324000298977, + 0.002533875871449709, + 0.015470104292035103, + 0.004580629989504814, + 0.0012941774912178516, + 0.0014261967735365033, + 0.001491344184614718, + 0.0013223369605839252, + 0.0011139288544654846, + 0.0009400758426636457, + 0.0012588464887812734, + 0.0010050309356302023, + 0.0006714256596751511, + 0.0008523999131284654, + 0.002172997687011957, + 0.001863455749116838, + 0.0017532544443383813, + 0.0019807731732726097, + 0.0018696823390200734, + 0.0014561641728505492, + 0.0021713185124099255, + 0.0018121417379006743, + 0.010450326837599277, + 0.004868840333074331, + 0.002728403080254793, + 0.002314232522621751, + 0.004382128827273846, + 0.0021237796172499657, + 0.0021789188031107187, + 0.004632643889635801, + 0.13786379992961884, + 0.17141376435756683, + 0.051327869296073914, + 0.049410536885261536, + 0.18033330142498016, + 0.07296477258205414, + 0.03975256159901619, + 0.07067286968231201, + 0.015212968923151493, + 0.0027279185596853495, + 0.002444429323077202, + 0.007021868135780096, + 0.01618700660765171, + 0.0027019900735467672, + 0.004907721653580666, + 0.002536362735554576, + 0.0027677572797983885, + 0.0009133358835242689, + 0.0006298389052972198, + 0.0007389036472886801, + 0.0024244028609246016, + 0.0009375268709845841, + 0.0007543431129306555, + 0.0010176629293709993, + 0.0018688702257350087, + 0.0016498479526489973, + 0.0010169559391215444, + 0.0023967954330146313, + 0.001276402617804706, + 0.0012919364962726831, + 0.000864896341226995, + 0.0013818797888234258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.05498193949460983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805032536387444 + ], + "opposite_logprob_tokens": [ + -3.5298049449920654 + ], + "logprob_reward": -0.029805032536387444, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805032536387444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047260478138923645, + 0.036855921149253845, + 0.0893285721540451, + 0.03941464424133301, + 0.1072327122092247, + 0.07301950454711914, + 0.26750001311302185, + 0.20722928643226624, + 0.03639639541506767, + 0.0957624688744545 + ], + "credit_map": [ + 0.008122297935187817, + 0.007381591480225325, + 0.0043719494715332985, + 0.004308545496314764, + 0.007867605425417423, + 0.003523317165672779, + 0.007402308285236359, + 0.004282860551029444, + 0.004813545849174261, + 0.0038131484761834145, + 0.005018967669457197, + 0.005443967878818512, + 0.0065894341096282005, + 0.0034740285482257605, + 0.0042371307499706745, + 0.003465695772320032, + 0.006320120766758919, + 0.005424289498478174, + 0.021175449714064598, + 0.005538106895983219, + 0.03392358869314194, + 0.010145347565412521, + 0.003851929446682334, + 0.0029497358482331038, + 0.005491634830832481, + 0.004747804719954729, + 0.004714448470622301, + 0.004205908626317978, + 0.008124327287077904, + 0.003398637054488063, + 0.0036400938406586647, + 0.005091789644211531, + 0.01191919855773449, + 0.01630529761314392, + 0.007893397472798824, + 0.009019059129059315, + 0.02767820842564106, + 0.009953032247722149, + 0.016116054728627205, + 0.008348457515239716, + 0.016248831525444984, + 0.011117947287857533, + 0.004877849947661161, + 0.006275958847254515, + 0.01523363497108221, + 0.004902465268969536, + 0.005172949284315109, + 0.009189873933792114, + 0.0477481372654438, + 0.055122260004282, + 0.01958758756518364, + 0.01944473572075367, + 0.05813010409474373, + 0.02353973686695099, + 0.01651221141219139, + 0.027415255084633827, + 0.06101326271891594, + 0.011018479242920876, + 0.008994650095701218, + 0.025180811062455177, + 0.06399666517972946, + 0.01114791538566351, + 0.017419418320059776, + 0.00845807883888483, + 0.005754430778324604, + 0.0037666873540729284, + 0.0027581381145864725, + 0.003145574824884534, + 0.007303341757506132, + 0.005421716254204512, + 0.003863647347316146, + 0.004382862243801355, + 0.012425784021615982, + 0.01012600027024746, + 0.011439185589551926, + 0.02555662766098976, + 0.011179148219525814, + 0.009593192487955093, + 0.007178680505603552, + 0.008263854309916496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.029805032536387444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a whole braised grouper with its form intact from head to tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grouper resting on a minimalist, dark gray stone platter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark ceramic spoon enter the frame from the side to pour a fresh layer of rich sauce?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the poured sauce create slow, mesmerizing ripples that expand in concentric circles across the fish's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred to suggest a luxurious dining setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783532470464706 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783532470464706, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783532470464706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08195972442626953, + 0.018675798550248146, + 0.192373126745224, + 0.019724547863006592, + 0.04690791293978691, + 0.0763951763510704, + 0.13289542496204376, + 0.23952807486057281, + 0.11605328321456909, + 0.07548696547746658 + ], + "credit_map": [ + 0.012658081948757172, + 0.01393836084753275, + 0.00709095224738121, + 0.007564641535282135, + 0.018400944769382477, + 0.009075572714209557, + 0.0062903743237257, + 0.006940794643014669, + 0.0025419804733246565, + 0.0020637670531868935, + 0.0019043185748159885, + 0.0019639511592686176, + 0.004149172455072403, + 0.0019096701871603727, + 0.002393473871052265, + 0.001749466173350811, + 0.015584439039230347, + 0.012736529111862183, + 0.0467485710978508, + 0.010695702396333218, + 0.07208094000816345, + 0.021522166207432747, + 0.0070731365121901035, + 0.00593162514269352, + 0.0027819829992949963, + 0.002355844946578145, + 0.002212239895015955, + 0.0015026368200778961, + 0.004276618827134371, + 0.002291919896379113, + 0.0018872206564992666, + 0.0024160826578736305, + 0.007574955001473427, + 0.00839291512966156, + 0.0031080699991434813, + 0.003317801747471094, + 0.010793914087116718, + 0.003969131037592888, + 0.005445533897727728, + 0.00430558854714036, + 0.02302326075732708, + 0.011508406139910221, + 0.0041971588507294655, + 0.005030421540141106, + 0.011296343989670277, + 0.006029633805155754, + 0.005291359964758158, + 0.010018592700362206, + 0.02244165539741516, + 0.028339629992842674, + 0.008616787381470203, + 0.008803926408290863, + 0.03135650232434273, + 0.013120235875248909, + 0.006968311499804258, + 0.013248364441096783, + 0.0706181526184082, + 0.012255745008587837, + 0.009951701387763023, + 0.02903067134320736, + 0.07503405213356018, + 0.012571167200803757, + 0.020339863374829292, + 0.009726718068122864, + 0.03898424655199051, + 0.008325382135808468, + 0.0050493888556957245, + 0.007555583957582712, + 0.03602275997400284, + 0.007394639775156975, + 0.003993767313659191, + 0.008727514185011387, + 0.014713325537741184, + 0.007075965870171785, + 0.008228084072470665, + 0.018709752708673477, + 0.0070862481370568275, + 0.006705909501761198, + 0.0060879020020365715, + 0.006879773922264576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.009783532470464706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_tensor([4872, 4872])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_tensor([4872, 4872])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d0d8ee3c0cc0778bce8a12020fa6554cb992e51d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_tensor([4872, 4872])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot captures a vast, sun-scorched landscape as the camera begins a slow, deliberate movement forward, tracking a lone figure on their journey. Their features are obscured by the deep hood of a heavy, forest-green wool cloak, its frayed hems, as they walk away from the camera with a steady, determined gait. They hold a rough wooden staff in one hand and carry a worn leather bag slung across their body, foreshadowing the long journey ahead. They are approaching a massive, ancient stone archway, constructed from massive blocks of weathered gray granite, carved with faded, intricate carvings and dotted with patches of green moss. The archway towers above the statues, silhouetted against a brilliant, clear blue sky. The entire scene is drenched in the bright, cinematic sunlight characteristic of the late afternoon golden hour, casting the travelers' long, sharp shadows across the dusty, ochre-colored path. As the figure steadily makes its way through the vast opening, their body briefly silhouetted against the light, falling into the cool, deep shadows of the archway before they emerge on the other side, bathed once more in warm, brilliant light, creating a powerful, epic atmosphere of the beginning of a momentous quest.", + "seed": 28177, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_4872_seed_28177_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03752166144549847, + "vlm_logprob": -0.03752166144549847, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.03752166144549847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0529804453253746 + ], + "opposite_logprob_tokens": [ + -3.552980422973633 + ], + "logprob_reward": -0.0529804453253746, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0529804453253746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08519838005304337, + 0.35702094435691833, + 0.049107763916254044, + 0.04084355756640434, + 0.019330544397234917, + 0.016494419425725937, + 0.23192794620990753, + 0.07020698487758636, + 0.05881426855921745, + 0.07105518132448196 + ], + "credit_map": [ + 0.013043870218098164, + 0.015286596491932869, + 0.011122283525764942, + 0.015538292936980724, + 0.009858261793851852, + 0.0038228468038141727, + 0.008861004374921322, + 0.007665228098630905, + 0.05320374667644501, + 0.02327895723283291, + 0.054253287613391876, + 0.15297387540340424, + 0.012185871601104736, + 0.0067007518373429775, + 0.018414754420518875, + 0.03600972518324852, + 0.004245575983077288, + 0.005772995762526989, + 0.0066500152461230755, + 0.007142744027078152, + 0.007481340318918228, + 0.0037766885943710804, + 0.0064608994871377945, + 0.007577506359666586, + 0.0027587497606873512, + 0.003574596019461751, + 0.015065588988363743, + 0.0037423858884721994, + 0.0027515992987900972, + 0.0033178520388901234, + 0.006574556231498718, + 0.0030582300387322903, + 0.0016014832071959972, + 0.0019475050503388047, + 0.0037907303776592016, + 0.0017927114386111498, + 0.003280038246884942, + 0.0012233748566359282, + 0.0033338244538754225, + 0.002360879210755229, + 0.002499775495380163, + 0.002875623060390353, + 0.0025146673433482647, + 0.0017492052866145968, + 0.0017038271762430668, + 0.0009945600759238005, + 0.002375770127400756, + 0.0017809903947636485, + 0.02398214489221573, + 0.03198939561843872, + 0.03807561844587326, + 0.022444117814302444, + 0.027364840731024742, + 0.01947334036231041, + 0.03384559974074364, + 0.03475290164351463, + 0.006864179857075214, + 0.012086199596524239, + 0.01515554916113615, + 0.006364264525473118, + 0.007420070935040712, + 0.005452032666653395, + 0.00910948496311903, + 0.007755210157483816, + 0.006915183737874031, + 0.009929586201906204, + 0.009144806303083897, + 0.005107977893203497, + 0.007989269681274891, + 0.00409106258302927, + 0.006304805167019367, + 0.009331581182777882, + 0.01080380193889141, + 0.009674638509750366, + 0.01074423361569643, + 0.004339417442679405, + 0.0076652769930660725, + 0.0075197117403149605, + 0.007142592687159777, + 0.013165514916181564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0529804453253746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029835810884833336 + ], + "opposite_logprob_tokens": [ + -3.5298357009887695 + ], + "logprob_reward": -0.029835810884833336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029835810884833336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055704254657030106, + 0.20474964380264282, + 0.054679024964571, + 0.04101545363664627, + 0.03676488250494003, + 0.02519548498094082, + 0.11130645126104355, + 0.26369085907936096, + 0.07492829859256744, + 0.13196571171283722 + ], + "credit_map": [ + 0.008235380053520203, + 0.011745301075279713, + 0.009107072837650776, + 0.0065122563391923904, + 0.0062501560896635056, + 0.0027709114365279675, + 0.005982950795441866, + 0.005100218113511801, + 0.030885785818099976, + 0.01564319245517254, + 0.02878461591899395, + 0.08213616162538528, + 0.01145201362669468, + 0.004727857653051615, + 0.011563095264136791, + 0.019556893035769463, + 0.00463943462818861, + 0.004589347634464502, + 0.018047360703349113, + 0.0042901188135147095, + 0.005316191352903843, + 0.005350176244974136, + 0.0077376295812428, + 0.004708764608949423, + 0.002714931732043624, + 0.0025563279632478952, + 0.0160492192953825, + 0.003189139999449253, + 0.0027725256513804197, + 0.002505234908312559, + 0.0068444618955254555, + 0.004383611027151346, + 0.0034675728529691696, + 0.004773886874318123, + 0.00663499953225255, + 0.003227340755984187, + 0.003933899570256472, + 0.004439108539372683, + 0.005499301478266716, + 0.004788768477737904, + 0.0020873346365988255, + 0.002808112185448408, + 0.0027060983702540398, + 0.003370846388861537, + 0.00306289573200047, + 0.0022937266621738672, + 0.004537356086075306, + 0.004329112358391285, + 0.010908047668635845, + 0.011844494380056858, + 0.017564058303833008, + 0.007552420254796743, + 0.012052166275680065, + 0.0135151082649827, + 0.018154630437493324, + 0.019715500995516777, + 0.022675884887576103, + 0.03091558627784252, + 0.04241780564188957, + 0.01670229062438011, + 0.04176412522792816, + 0.015446772798895836, + 0.05193265900015831, + 0.0418357327580452, + 0.006771520711481571, + 0.010680616833269596, + 0.010837172158062458, + 0.0062162866815924644, + 0.01088809221982956, + 0.005225751549005508, + 0.011153165251016617, + 0.013155683875083923, + 0.015566115267574787, + 0.015509631484746933, + 0.023630158975720406, + 0.00606285547837615, + 0.015114769339561462, + 0.017071876674890518, + 0.01151992753148079, + 0.027490362524986267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.029835810884833336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981428988277912 + ], + "opposite_logprob_tokens": [ + -3.5298142433166504 + ], + "logprob_reward": -0.02981428988277912, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981428988277912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1009691134095192, + 0.1214333176612854, + 0.08380072563886642, + 0.11653350293636322, + 0.09828216582536697, + 0.057137977331876755, + 0.0988287627696991, + 0.13366936147212982, + 0.07794652134180069, + 0.11139853298664093 + ], + "credit_map": [ + 0.012611474841833115, + 0.021193604916334152, + 0.012806751765310764, + 0.010880262590944767, + 0.018287362530827522, + 0.006205817684531212, + 0.010269585065543652, + 0.00871425960212946, + 0.021387897431850433, + 0.016791095957159996, + 0.012908604927361012, + 0.03755280748009682, + 0.010582557879388332, + 0.004642216954380274, + 0.008092661388218403, + 0.00947549194097519, + 0.008490405045449734, + 0.010125218890607357, + 0.020798644050955772, + 0.00933560635894537, + 0.011888623237609863, + 0.005433797836303711, + 0.010154510848224163, + 0.007573931012302637, + 0.0074187982827425, + 0.008566364645957947, + 0.04463690146803856, + 0.008887975476682186, + 0.007904098369181156, + 0.010436534881591797, + 0.01744764856994152, + 0.011235188692808151, + 0.009440034627914429, + 0.013302161358296871, + 0.01885215938091278, + 0.007260543294250965, + 0.013418006710708141, + 0.011741462163627148, + 0.01382697094231844, + 0.010440830141305923, + 0.004196844529360533, + 0.006998047232627869, + 0.006071846932172775, + 0.006677843630313873, + 0.00839459989219904, + 0.007919461466372013, + 0.007796839345246553, + 0.009082497097551823, + 0.007274516858160496, + 0.007960759103298187, + 0.025334835052490234, + 0.004809081554412842, + 0.0076407743617892265, + 0.007562164682894945, + 0.011790206655859947, + 0.026456430554389954, + 0.014297195710241795, + 0.016221681609749794, + 0.022036809474229813, + 0.011048058979213238, + 0.02085353247821331, + 0.007451738230884075, + 0.020429085940122604, + 0.021331261843442917, + 0.00673731230199337, + 0.009878560900688171, + 0.01072773989289999, + 0.008896597661077976, + 0.013242618180811405, + 0.005908442195504904, + 0.010905089788138866, + 0.011650157161056995, + 0.011585895903408527, + 0.015359184704720974, + 0.020801017060875893, + 0.009711980819702148, + 0.012167164124548435, + 0.012174342758953571, + 0.008561731316149235, + 0.021037224680185318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.02981428988277912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634398825466633 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634398825466633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634398825466633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06479737907648087, + 0.10022707283496857, + 0.03028206154704094, + 0.01937982626259327, + 0.023976774886250496, + 0.025671951472759247, + 0.4237460196018219, + 0.09037322551012039, + 0.04370132461190224, + 0.1778443455696106 + ], + "credit_map": [ + 0.0074842520989477634, + 0.019284624606370926, + 0.010140301659703255, + 0.009253568947315216, + 0.005555708426982164, + 0.0031051444821059704, + 0.004800861235707998, + 0.005172926466912031, + 0.011193382553756237, + 0.010352634824812412, + 0.017746947705745697, + 0.03940272703766823, + 0.004577255342155695, + 0.0021059203427284956, + 0.006011904217302799, + 0.00883630383759737, + 0.002738297451287508, + 0.0036934835370630026, + 0.005891166627407074, + 0.004159857053309679, + 0.004214625805616379, + 0.0023771498817950487, + 0.004503475967794657, + 0.0027040073182433844, + 0.001659720903262496, + 0.0021817213855683804, + 0.0050757271237671375, + 0.0023049507290124893, + 0.002175829838961363, + 0.0009980004979297519, + 0.0034017134457826614, + 0.0015821645502001047, + 0.002714649774134159, + 0.003684403607621789, + 0.003649577498435974, + 0.001941887428984046, + 0.00309390970505774, + 0.002631962299346924, + 0.004185585770756006, + 0.0020748006645590067, + 0.002517031505703926, + 0.0027563832700252533, + 0.002808586461469531, + 0.0022481936030089855, + 0.0027498977724462748, + 0.004433885682374239, + 0.0038433854933828115, + 0.0043145897798240185, + 0.024870365858078003, + 0.02542191371321678, + 0.15556298196315765, + 0.03536486253142357, + 0.026328137144446373, + 0.034246936440467834, + 0.06372067332267761, + 0.05823017284274101, + 0.010077029466629028, + 0.011010348796844482, + 0.01974150352180004, + 0.005668561905622482, + 0.007061331998556852, + 0.00640373770147562, + 0.011031168513000011, + 0.019379550591111183, + 0.003875617403537035, + 0.006425665225833654, + 0.006549656391143799, + 0.004566861316561699, + 0.006423582788556814, + 0.003021468408405781, + 0.005204630549997091, + 0.007633845787495375, + 0.022660505026578903, + 0.02158506214618683, + 0.031869057565927505, + 0.0073128980584442616, + 0.020900128409266472, + 0.022096551954746246, + 0.012594161555171013, + 0.03882598876953125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.02634398825466633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048633772879838943 + ], + "opposite_logprob_tokens": [ + -3.0486338138580322 + ], + "logprob_reward": -0.048633772879838943, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048633772879838943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19128714501857758, + 0.08719545602798462, + 0.04481065273284912, + 0.014772356487810612, + 0.02188790962100029, + 0.04194772243499756, + 0.2423178255558014, + 0.061884552240371704, + 0.06072014197707176, + 0.23317617177963257 + ], + "credit_map": [ + 0.025093406438827515, + 0.03253878653049469, + 0.026648594066500664, + 0.028148092329502106, + 0.028737014159560204, + 0.010093743912875652, + 0.026065057143568993, + 0.013962441124022007, + 0.007607778999954462, + 0.007747521158307791, + 0.014017214067280293, + 0.02975328266620636, + 0.008292827755212784, + 0.002648663939908147, + 0.00889176968485117, + 0.008236391469836235, + 0.004624133929610252, + 0.0036714619491249323, + 0.015649499371647835, + 0.004035411402583122, + 0.0036001040134578943, + 0.0035015838220715523, + 0.005435433704406023, + 0.004293026868253946, + 0.0013573926407843828, + 0.0016387909417971969, + 0.0022071260027587414, + 0.001533066388219595, + 0.0018468094058334827, + 0.001919097383506596, + 0.00234428932890296, + 0.0019257845124229789, + 0.0023723281919956207, + 0.0025807046331465244, + 0.003615537192672491, + 0.0015448409831151366, + 0.003409987548366189, + 0.0012903142487630248, + 0.004340317100286484, + 0.0027338811196386814, + 0.0035954490303993225, + 0.00458786403760314, + 0.0052671427838504314, + 0.0068636490032076836, + 0.004910117946565151, + 0.003371569560840726, + 0.007718214765191078, + 0.005633717868477106, + 0.014374501071870327, + 0.014999529346823692, + 0.08862346410751343, + 0.019685031846165657, + 0.013024221174418926, + 0.025524508208036423, + 0.03870979696512222, + 0.027376802638173103, + 0.007021784782409668, + 0.00810654554516077, + 0.01188800297677517, + 0.005198637489229441, + 0.008270157501101494, + 0.004124340135604143, + 0.008468111976981163, + 0.008806976489722729, + 0.007224936969578266, + 0.009649183601140976, + 0.008586740121245384, + 0.004573015030473471, + 0.00990165676921606, + 0.00457638967782259, + 0.0054844459518790245, + 0.010723770596086979, + 0.029117681086063385, + 0.03846094012260437, + 0.03435027599334717, + 0.017054764553904533, + 0.03482823073863983, + 0.018137572333216667, + 0.025675062090158463, + 0.035551656037569046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.048633772879838943 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot captures a vast, sun-scorched landscape as the camera begins a slow, deliberate movement forward, tracking a lone figure on their journey. Their features are obscured by the deep hood of a heavy, forest-green wool cloak, its frayed hems, as they walk away from the camera with a steady, determined gait. They hold a rough wooden staff in one hand and carry a worn leather bag slung across their body, foreshadowing the long journey ahead. They are approaching a massive, ancient stone archway, constructed from massive blocks of weathered gray granite, carved with faded, intricate carvings and dotted with patches of green moss. The archway towers above the statues, silhouetted against a brilliant, clear blue sky. The entire scene is drenched in the bright, cinematic sunlight characteristic of the late afternoon golden hour, casting the travelers' long, sharp shadows across the dusty, ochre-colored path. As the figure steadily makes its way through the vast opening, their body briefly silhouetted against the light, falling into the cool, deep shadows of the archway before they emerge on the other side, bathed once more in warm, brilliant light, creating a powerful, epic atmosphere of the beginning of a momentous quest.", + "seed": 28178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_4872_seed_28178_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04878315143287182, + "vlm_logprob": -0.04878315143287182, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.04878315143287182, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04377474635839462 + ], + "opposite_logprob_tokens": [ + -3.9187748432159424 + ], + "logprob_reward": -0.04377474635839462, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04377474635839462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19500355422496796, + 0.12094389647245407, + 0.09210517257452011, + 0.09864653646945953, + 0.046492885798215866, + 0.03991067782044411, + 0.061210233718156815, + 0.07188752293586731, + 0.04838719218969345, + 0.22541233897209167 + ], + "credit_map": [ + 0.014150022529065609, + 0.022533804178237915, + 0.016012044623494148, + 0.029182951897382736, + 0.027261195704340935, + 0.025486236438155174, + 0.04121246188879013, + 0.01916481927037239, + 0.01197978388518095, + 0.013587835244834423, + 0.014200950041413307, + 0.01940893940627575, + 0.016503531485795975, + 0.009078949689865112, + 0.02061762847006321, + 0.01556627918034792, + 0.005500535015016794, + 0.014062261208891869, + 0.006006029900163412, + 0.016300804913043976, + 0.013257335871458054, + 0.005236152559518814, + 0.014721106737852097, + 0.017020946368575096, + 0.009192153811454773, + 0.011807947419583797, + 0.010085183195769787, + 0.013511714525520802, + 0.016356470063328743, + 0.008781448006629944, + 0.0177933182567358, + 0.01111829373985529, + 0.006408282555639744, + 0.00410552229732275, + 0.005576409865170717, + 0.004366814158856869, + 0.007329766172915697, + 0.005411169957369566, + 0.006753423251211643, + 0.006541499402374029, + 0.003997578285634518, + 0.0029459588695317507, + 0.003830595640465617, + 0.003672979539260268, + 0.006813094951212406, + 0.00882834941148758, + 0.004102334380149841, + 0.005719788838177919, + 0.007446751464158297, + 0.005306739825755358, + 0.006909272633492947, + 0.009047694504261017, + 0.008905851282179356, + 0.010096057318150997, + 0.007663577329367399, + 0.005834287032485008, + 0.007544974330812693, + 0.005493205040693283, + 0.006042009685188532, + 0.010866489261388779, + 0.0065165068954229355, + 0.011266220360994339, + 0.01504729688167572, + 0.009110820479691029, + 0.008239221759140491, + 0.005218712612986565, + 0.008530721999704838, + 0.005071968771517277, + 0.0034253355115652084, + 0.0075173950754106045, + 0.005293870810419321, + 0.005089963786303997, + 0.03637763857841492, + 0.018243223428726196, + 0.02329554222524166, + 0.020336272194981575, + 0.017034074291586876, + 0.058186545968055725, + 0.017920101061463356, + 0.03401894122362137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04377474635839462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498826131224632 + ], + "opposite_logprob_tokens": [ + -2.929988145828247 + ], + "logprob_reward": -0.05498826131224632, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498826131224632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4340745508670807, + 0.03563889116048813, + 0.07470597326755524, + 0.15100175142288208, + 0.10023999959230423, + 0.047711391001939774, + 0.04922574386000633, + 0.04266924783587456, + 0.02125615067780018, + 0.043476205319166183 + ], + "credit_map": [ + 0.02457813173532486, + 0.04747480899095535, + 0.025648798793554306, + 0.085377536714077, + 0.07991601526737213, + 0.08411338180303574, + 0.055350564420223236, + 0.031615372747182846, + 0.002701947931200266, + 0.003785393200814724, + 0.0031990339048206806, + 0.005061253439635038, + 0.005098237190395594, + 0.0032406693790107965, + 0.007919685915112495, + 0.004632674623280764, + 0.005177033133804798, + 0.011733208782970905, + 0.005119973793625832, + 0.01352175697684288, + 0.011194533668458462, + 0.0040820445865392685, + 0.011640588752925396, + 0.012236842885613441, + 0.015439213253557682, + 0.016318462789058685, + 0.014947052113711834, + 0.02067948691546917, + 0.029651297256350517, + 0.009831434115767479, + 0.024854762479662895, + 0.019280051812529564, + 0.009376860223710537, + 0.006319839507341385, + 0.009356643073260784, + 0.011428548023104668, + 0.016463391482830048, + 0.01606825552880764, + 0.018162749707698822, + 0.013063712045550346, + 0.003244684310629964, + 0.002791518112644553, + 0.004536043386906385, + 0.004470049869269133, + 0.00824450608342886, + 0.011572104878723621, + 0.005503419786691666, + 0.007349070161581039, + 0.006464289966970682, + 0.003987145144492388, + 0.005550427827984095, + 0.004763927776366472, + 0.0065439133904874325, + 0.010343135334551334, + 0.0062499623745679855, + 0.005322947632521391, + 0.004336372949182987, + 0.0031121280044317245, + 0.0037310754414647818, + 0.004423015285283327, + 0.006761597003787756, + 0.009389856830239296, + 0.005094980355352163, + 0.0058202240616083145, + 0.003794451244175434, + 0.0017300312174484134, + 0.0034276358783245087, + 0.0017502875998616219, + 0.0017718843882903457, + 0.004267510492354631, + 0.0019470307743176818, + 0.0025673198979347944, + 0.004069757182151079, + 0.0029294234700500965, + 0.003516833996400237, + 0.006460142321884632, + 0.0028801553416997194, + 0.010389716364443302, + 0.006933750584721565, + 0.006296431180089712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.05498826131224632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263564120978117 + ], + "opposite_logprob_tokens": [ + -3.6513564586639404 + ], + "logprob_reward": -0.0263564120978117, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263564120978117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13722585141658783, + 0.08118592202663422, + 0.19802731275558472, + 0.16050028800964355, + 0.1641356199979782, + 0.04409845173358917, + 0.0459587536752224, + 0.06613346934318542, + 0.029917586594820023, + 0.07281673699617386 + ], + "credit_map": [ + 0.01494957972317934, + 0.013101756572723389, + 0.013794109225273132, + 0.01894141361117363, + 0.02605344168841839, + 0.022184571251273155, + 0.014498962089419365, + 0.01370201539248228, + 0.010432434268295765, + 0.012546595185995102, + 0.009291102178394794, + 0.013055229559540749, + 0.007247090805321932, + 0.005671574734151363, + 0.012595797888934612, + 0.01034610066562891, + 0.011302832514047623, + 0.028043339028954506, + 0.0130003048107028, + 0.0317789651453495, + 0.034111205488443375, + 0.012402246706187725, + 0.03203880041837692, + 0.03534962236881256, + 0.013843530789017677, + 0.012120258994400501, + 0.016328200697898865, + 0.019637010991573334, + 0.03705314174294472, + 0.015268933959305286, + 0.024212533608078957, + 0.022036686539649963, + 0.018814990296959877, + 0.009539125487208366, + 0.017711808905005455, + 0.017409509047865868, + 0.023254157975316048, + 0.024019835516810417, + 0.030911602079868317, + 0.022474585101008415, + 0.0033828529994934797, + 0.0035615426022559404, + 0.0062850466929376125, + 0.0040139672346413136, + 0.005474252626299858, + 0.009343585930764675, + 0.005644934717565775, + 0.006392271723598242, + 0.006357169710099697, + 0.0032879291102290154, + 0.005233513657003641, + 0.005174897611141205, + 0.00589204765856266, + 0.008988309651613235, + 0.0057007805444300175, + 0.005324107129126787, + 0.006050740368664265, + 0.003908147104084492, + 0.005665384232997894, + 0.006080139894038439, + 0.007507102098315954, + 0.012900402769446373, + 0.014595430344343185, + 0.009426121599972248, + 0.004186568781733513, + 0.0018334470223635435, + 0.0026344151701778173, + 0.00335322436876595, + 0.0020760735496878624, + 0.008224237710237503, + 0.0032688819337636232, + 0.00434073805809021, + 0.006818712688982487, + 0.004318893887102604, + 0.008263146504759789, + 0.011623721569776535, + 0.005125478375703096, + 0.015744753181934357, + 0.01106635108590126, + 0.009855680167675018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.0263564120978117 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790569096803665 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790569096803665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790569096803665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2772035002708435, + 0.027022354304790497, + 0.19131098687648773, + 0.0731186717748642, + 0.05751064047217369, + 0.09387484192848206, + 0.09472368657588959, + 0.11132500320672989, + 0.022908015176653862, + 0.05100230127573013 + ], + "credit_map": [ + 0.013451775535941124, + 0.0255888681858778, + 0.01648780144751072, + 0.06266133487224579, + 0.04672976955771446, + 0.06397200375795364, + 0.03199700638651848, + 0.016314949840307236, + 0.003487707581371069, + 0.003046647412702441, + 0.0042617786675691605, + 0.0035064329858869314, + 0.0025237288791686296, + 0.0019847415387630463, + 0.004687066189944744, + 0.003524250350892544, + 0.013740304857492447, + 0.028005510568618774, + 0.013504684902727604, + 0.029985224828124046, + 0.03241554647684097, + 0.011229179799556732, + 0.031621068716049194, + 0.030809462070465088, + 0.008642029948532581, + 0.008512044325470924, + 0.007807644549757242, + 0.007003488019108772, + 0.01785394363105297, + 0.005072839558124542, + 0.009328645654022694, + 0.00889803096652031, + 0.005326626822352409, + 0.005002129357308149, + 0.004588328301906586, + 0.007802718784660101, + 0.007973873056471348, + 0.010502126067876816, + 0.009027831256389618, + 0.0072870030999183655, + 0.008357769809663296, + 0.006463255267590284, + 0.00984780304133892, + 0.009773210622370243, + 0.014785355888307095, + 0.021146060898900032, + 0.009610001929104328, + 0.013891387730836868, + 0.010614839382469654, + 0.009171294048428535, + 0.007739448919892311, + 0.00814025104045868, + 0.01450938917696476, + 0.020416513085365295, + 0.013338572345674038, + 0.010793381370604038, + 0.008624399080872536, + 0.011290202848613262, + 0.009587623178958893, + 0.018843594938516617, + 0.011945492587983608, + 0.022488486021757126, + 0.014847364276647568, + 0.013697845861315727, + 0.00316965626552701, + 0.0014446908608078957, + 0.0021635661832988262, + 0.002302789594978094, + 0.001577510847710073, + 0.005763575900346041, + 0.0035345335491001606, + 0.002951692556962371, + 0.005862127523869276, + 0.004092057701200247, + 0.004852267913520336, + 0.00823728647083044, + 0.004189074505120516, + 0.010484431870281696, + 0.006582154892385006, + 0.006702898070216179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.029790569096803665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900576829910278 + ], + "opposite_logprob_tokens": [ + -2.464005708694458 + ], + "logprob_reward": -0.08900576829910278, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900576829910278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22874976694583893, + 0.06152411177754402, + 0.08711525797843933, + 0.16130194067955017, + 0.05813021957874298, + 0.07306236028671265, + 0.04978853464126587, + 0.14374980330467224, + 0.051203954964876175, + 0.08537406474351883 + ], + "credit_map": [ + 0.02149389497935772, + 0.029113849624991417, + 0.01864304207265377, + 0.02449488639831543, + 0.04135190322995186, + 0.02812078222632408, + 0.043478548526763916, + 0.022052865475416183, + 0.00591585598886013, + 0.007686594966799021, + 0.01025598868727684, + 0.009341862984001637, + 0.005635642446577549, + 0.005192744545638561, + 0.009595259092748165, + 0.007900167256593704, + 0.004626809619367123, + 0.007813692092895508, + 0.0072753108106553555, + 0.013941965065896511, + 0.00982852466404438, + 0.006764475721865892, + 0.01801760494709015, + 0.018846875056624413, + 0.016649924218654633, + 0.012454483658075333, + 0.012892971746623516, + 0.015875548124313354, + 0.030084913596510887, + 0.027993371710181236, + 0.021744878962635994, + 0.023605845868587494, + 0.00724794901907444, + 0.004011584911495447, + 0.005578103009611368, + 0.007089803460985422, + 0.008876865729689598, + 0.007604044396430254, + 0.009427925571799278, + 0.008293944410979748, + 0.005987564567476511, + 0.007920634001493454, + 0.008307798765599728, + 0.013058445416390896, + 0.010188164189457893, + 0.009704720228910446, + 0.00815813709050417, + 0.00973689928650856, + 0.004370827693492174, + 0.0037098308093845844, + 0.006078934762626886, + 0.004371742717921734, + 0.004923609551042318, + 0.011213662102818489, + 0.00900486670434475, + 0.0061150589026510715, + 0.011365976184606552, + 0.009530705399811268, + 0.01131567731499672, + 0.023647010326385498, + 0.014798387885093689, + 0.01898060366511345, + 0.035583700984716415, + 0.018527746200561523, + 0.008529671467840672, + 0.0029072691686451435, + 0.005280385725200176, + 0.004637866746634245, + 0.0035180868580937386, + 0.013083988800644875, + 0.005859512370079756, + 0.007387181278318167, + 0.007191958837211132, + 0.007159706205129623, + 0.0066042207181453705, + 0.01632390357553959, + 0.006743181962519884, + 0.013650910928845406, + 0.017591459676623344, + 0.01010872982442379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.08900576829910278 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot captures a vast, sun-scorched landscape as the camera begins a slow, deliberate movement forward, tracking a lone figure on their journey. Their features are obscured by the deep hood of a heavy, forest-green wool cloak, its frayed hems, as they walk away from the camera with a steady, determined gait. They hold a rough wooden staff in one hand and carry a worn leather bag slung across their body, foreshadowing the long journey ahead. They are approaching a massive, ancient stone archway, constructed from massive blocks of weathered gray granite, carved with faded, intricate carvings and dotted with patches of green moss. The archway towers above the statues, silhouetted against a brilliant, clear blue sky. The entire scene is drenched in the bright, cinematic sunlight characteristic of the late afternoon golden hour, casting the travelers' long, sharp shadows across the dusty, ochre-colored path. As the figure steadily makes its way through the vast opening, their body briefly silhouetted against the light, falling into the cool, deep shadows of the archway before they emerge on the other side, bathed once more in warm, brilliant light, creating a powerful, epic atmosphere of the beginning of a momentous quest.", + "seed": 28179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_4872_seed_28179_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0452385488897562, + "vlm_logprob": -0.0452385488897562, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.0452385488897562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05295319855213165 + ], + "opposite_logprob_tokens": [ + -3.5529532432556152 + ], + "logprob_reward": -0.05295319855213165, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05295319855213165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2668353021144867, + 0.051825981587171555, + 0.052407156676054, + 0.11395376920700073, + 0.1413121223449707, + 0.08166138827800751, + 0.06336013227701187, + 0.0753379613161087, + 0.07202238589525223, + 0.08128378540277481 + ], + "credit_map": [ + 0.03517834469676018, + 0.016607213765382767, + 0.03280283138155937, + 0.02660745568573475, + 0.04464064538478851, + 0.024475792422890663, + 0.04099449887871742, + 0.04552852362394333, + 0.00400738837197423, + 0.004052801989018917, + 0.0053105284459888935, + 0.0050049638375639915, + 0.011328143998980522, + 0.006210901774466038, + 0.007905128411948681, + 0.008006121963262558, + 0.004799692891538143, + 0.0035701182205229998, + 0.006894114427268505, + 0.005445713177323341, + 0.01089895237237215, + 0.005935754626989365, + 0.008304906077682972, + 0.006557905580848455, + 0.008434772491455078, + 0.003884045872837305, + 0.03715849667787552, + 0.00843964982777834, + 0.009113322012126446, + 0.007832949049770832, + 0.007382769137620926, + 0.0317077599465847, + 0.01586783677339554, + 0.01549141388386488, + 0.015160832554101944, + 0.012176292017102242, + 0.02854996919631958, + 0.014515077695250511, + 0.021809685975313187, + 0.017741011455655098, + 0.008767228573560715, + 0.006288212724030018, + 0.013188070617616177, + 0.006916870828717947, + 0.007628232706338167, + 0.007512898650020361, + 0.01922820508480072, + 0.012131673283874989, + 0.007355904672294855, + 0.006144121289253235, + 0.00757173914462328, + 0.007155439350754023, + 0.006771050859242678, + 0.009946110658347607, + 0.010174311697483063, + 0.008241456933319569, + 0.005091866943985224, + 0.009388161823153496, + 0.008043181151151657, + 0.004718627780675888, + 0.01295231468975544, + 0.007783082779496908, + 0.014381561428308487, + 0.012979166582226753, + 0.006434039678424597, + 0.004615155979990959, + 0.006367817986756563, + 0.003834967501461506, + 0.0059332107193768024, + 0.01606425642967224, + 0.02021149732172489, + 0.008561445400118828, + 0.00861331820487976, + 0.013600059784948826, + 0.008679176680743694, + 0.006918528117239475, + 0.012584825046360493, + 0.008142456412315369, + 0.014692193828523159, + 0.008053227327764034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.05295319855213165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813929855823517 + ], + "opposite_logprob_tokens": [ + -3.2881393432617188 + ], + "logprob_reward": -0.03813929855823517, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813929855823517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057342398911714554, + 0.12831178307533264, + 0.14196963608264923, + 0.2837439477443695, + 0.08120317757129669, + 0.08137364685535431, + 0.027036001905798912, + 0.04478532820940018, + 0.06290893256664276, + 0.09132517129182816 + ], + "credit_map": [ + 0.006501962896436453, + 0.005779178347438574, + 0.006776975467801094, + 0.005830755922943354, + 0.007862789556384087, + 0.006820813287049532, + 0.006747312843799591, + 0.011022602207958698, + 0.011861825361847878, + 0.0074607557617127895, + 0.014438441023230553, + 0.010719710029661655, + 0.026759173721075058, + 0.00992522295564413, + 0.029673190787434578, + 0.017473455518484116, + 0.012437734752893448, + 0.010609796270728111, + 0.013014580123126507, + 0.01357033010572195, + 0.03706861659884453, + 0.012191157788038254, + 0.02730819582939148, + 0.015769219025969505, + 0.023262983188033104, + 0.007418494205921888, + 0.10587706416845322, + 0.02000311389565468, + 0.016658127307891846, + 0.015526257455348969, + 0.01678515039384365, + 0.07821274548768997, + 0.009022449143230915, + 0.009332217276096344, + 0.007904917933046818, + 0.0062192995101213455, + 0.015388347208499908, + 0.009587183594703674, + 0.012041615322232246, + 0.011707139201462269, + 0.008763131685554981, + 0.005942437797784805, + 0.00811193510890007, + 0.004397082142531872, + 0.006911072414368391, + 0.008483394980430603, + 0.026049286127090454, + 0.012715301476418972, + 0.003346048528328538, + 0.003494828939437866, + 0.00421128561720252, + 0.0031905346550047398, + 0.0028146819677203894, + 0.0037416191771626472, + 0.0036302623338997364, + 0.0026067395228892565, + 0.0037055236753076315, + 0.0055365776643157005, + 0.008638073690235615, + 0.0024447436444461346, + 0.0047302148304879665, + 0.004714468028396368, + 0.00787392258644104, + 0.007141799200326204, + 0.006018443498760462, + 0.003671143436804414, + 0.0052546844817698, + 0.0028795357793569565, + 0.005672764964401722, + 0.018720507621765137, + 0.013962958008050919, + 0.006728886626660824, + 0.01033188495784998, + 0.008672088384628296, + 0.008316346444189548, + 0.0054028877057135105, + 0.007667019031941891, + 0.026371505111455917, + 0.01633584499359131, + 0.008227589540183544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.03813929855823517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352813467383385 + ], + "opposite_logprob_tokens": [ + -3.651352882385254 + ], + "logprob_reward": -0.026352813467383385, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352813467383385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07882729172706604, + 0.3344545066356659, + 0.06433770805597305, + 0.04876898229122162, + 0.12926049530506134, + 0.04929780960083008, + 0.048776768147945404, + 0.07256394624710083, + 0.08930782973766327, + 0.08440471440553665 + ], + "credit_map": [ + 0.00861317478120327, + 0.007919523864984512, + 0.010105689987540245, + 0.007593592628836632, + 0.00715245446190238, + 0.008743378333747387, + 0.00843423418700695, + 0.020265236496925354, + 0.02709648571908474, + 0.012947197072207928, + 0.04322499781847, + 0.01932438090443611, + 0.08295514434576035, + 0.023144058883190155, + 0.08557397872209549, + 0.040188245475292206, + 0.005522641818970442, + 0.007219928782433271, + 0.007514411583542824, + 0.010602794587612152, + 0.010540464892983437, + 0.006279775407165289, + 0.007605339866131544, + 0.009052351117134094, + 0.0035609803162515163, + 0.0034628338180482388, + 0.01332135684788227, + 0.004431699402630329, + 0.005074311979115009, + 0.0038689356297254562, + 0.004155838396400213, + 0.010893025435507298, + 0.01448576059192419, + 0.012845448218286037, + 0.013564460910856724, + 0.009755231440067291, + 0.02334900200366974, + 0.016609063372015953, + 0.01713111437857151, + 0.02152041345834732, + 0.0066120619885623455, + 0.004580026492476463, + 0.00728260725736618, + 0.004206918645650148, + 0.005327199120074511, + 0.006945173256099224, + 0.007472787983715534, + 0.006871035322546959, + 0.005247313529253006, + 0.004283027723431587, + 0.005279420875012875, + 0.00503797922283411, + 0.0039017004892230034, + 0.006421684753149748, + 0.008828210644423962, + 0.0097774313762784, + 0.005389753729104996, + 0.009724161587655544, + 0.013712295331060886, + 0.0038423577789217234, + 0.008265148848295212, + 0.007338728290051222, + 0.012471531517803669, + 0.01181996613740921, + 0.008892924524843693, + 0.005745448637753725, + 0.008870174176990986, + 0.0041902982629835606, + 0.0072837225161492825, + 0.028169874101877213, + 0.016983840614557266, + 0.009171545505523682, + 0.008992460556328297, + 0.009094713255763054, + 0.008225709199905396, + 0.005427923984825611, + 0.008316717110574245, + 0.02159961126744747, + 0.014014073647558689, + 0.008733500726521015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.026352813467383385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896079868078232 + ], + "opposite_logprob_tokens": [ + -2.57896089553833 + ], + "logprob_reward": -0.07896079868078232, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896079868078232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13392235338687897, + 0.04029402881860733, + 0.048808250576257706, + 0.046196095645427704, + 0.26588329672813416, + 0.08565171808004379, + 0.0790744349360466, + 0.0370042510330677, + 0.1693815290927887, + 0.09378401935100555 + ], + "credit_map": [ + 0.014879942871630192, + 0.02102448046207428, + 0.016218429431319237, + 0.014606992714107037, + 0.015597810968756676, + 0.013821041211485863, + 0.016847066581249237, + 0.020926598459482193, + 0.004842262715101242, + 0.004557678010314703, + 0.0037927983794361353, + 0.003994999453425407, + 0.00857449322938919, + 0.0034201520029455423, + 0.00629783421754837, + 0.004813815001398325, + 0.003673835890367627, + 0.005616714712232351, + 0.005897758528590202, + 0.007242766674607992, + 0.009664100594818592, + 0.003687961958348751, + 0.007219286635518074, + 0.0058058276772499084, + 0.003626507008448243, + 0.0034575925674289465, + 0.013314569368958473, + 0.0041027008555829525, + 0.004105325788259506, + 0.003998185973614454, + 0.003945343196392059, + 0.009645873680710793, + 0.03057999536395073, + 0.028544899076223373, + 0.026863975450396538, + 0.02093878574669361, + 0.04909847676753998, + 0.029703781008720398, + 0.03525763005018234, + 0.044895775616168976, + 0.010606316849589348, + 0.006851194892078638, + 0.014334123581647873, + 0.006291088182479143, + 0.012989813461899757, + 0.010678336955606937, + 0.011860555969178677, + 0.012040299363434315, + 0.007926169782876968, + 0.008092175237834454, + 0.010389580391347408, + 0.007317236624658108, + 0.008277486078441143, + 0.00812553334981203, + 0.013876610435545444, + 0.015069648623466492, + 0.0036973897367715836, + 0.0051717376336455345, + 0.004839036613702774, + 0.0037753565702587366, + 0.003862239420413971, + 0.0052561587654054165, + 0.005063582211732864, + 0.005338751245290041, + 0.016888882964849472, + 0.008022076450288296, + 0.016902362927794456, + 0.006550883408635855, + 0.011464009061455727, + 0.04469519481062889, + 0.04557637497782707, + 0.01928175613284111, + 0.010151159949600697, + 0.008217306807637215, + 0.009272265248000622, + 0.006027881987392902, + 0.00965514313429594, + 0.024310728535056114, + 0.019857486709952354, + 0.0062920572236180305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.07896079868078232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978663519024849 + ], + "opposite_logprob_tokens": [ + -3.5297865867614746 + ], + "logprob_reward": -0.02978663519024849, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978663519024849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09349234402179718, + 0.12249381095170975, + 0.054669089615345, + 0.1970057189464569, + 0.13390927016735077, + 0.11937171220779419, + 0.08913661539554596, + 0.07149118930101395, + 0.04315653815865517, + 0.07527372241020203 + ], + "credit_map": [ + 0.01189444214105606, + 0.00979921966791153, + 0.010111043229699135, + 0.008419135585427284, + 0.015422814525663853, + 0.009852133691310883, + 0.014116204343736172, + 0.01387734804302454, + 0.010240261442959309, + 0.00669488962739706, + 0.01573588140308857, + 0.008423813618719578, + 0.02598259411752224, + 0.007887888699769974, + 0.031528644263744354, + 0.015999827533960342, + 0.005233575589954853, + 0.003659695852547884, + 0.00503825256600976, + 0.005469826050102711, + 0.01405543927103281, + 0.005698422435671091, + 0.0090801902115345, + 0.006433685775846243, + 0.015392951667308807, + 0.005254070274531841, + 0.0750230997800827, + 0.011692422442138195, + 0.009507574141025543, + 0.009229031391441822, + 0.010215375572443008, + 0.06069119647145271, + 0.014570475555956364, + 0.011484536342322826, + 0.010137699544429779, + 0.012327900156378746, + 0.02224455587565899, + 0.012456074357032776, + 0.030863279476761818, + 0.019824739545583725, + 0.01507839560508728, + 0.0125807523727417, + 0.009995429776608944, + 0.009031192399561405, + 0.009357001632452011, + 0.010485777631402016, + 0.03609873726963997, + 0.016744423657655716, + 0.01039317436516285, + 0.008820775896310806, + 0.007612522225826979, + 0.008523717522621155, + 0.010236772708594799, + 0.012180604040622711, + 0.015990572050213814, + 0.015378470532596111, + 0.00429424736648798, + 0.008832527324557304, + 0.008476878516376019, + 0.005899013485759497, + 0.012809023261070251, + 0.007740156725049019, + 0.013734246604144573, + 0.00970509648323059, + 0.005871978122740984, + 0.003392755286768079, + 0.012132507748901844, + 0.0034693963825702667, + 0.0036273144651204348, + 0.0035896184854209423, + 0.007318840827792883, + 0.003754125442355871, + 0.008881935849785805, + 0.014995143748819828, + 0.006554868072271347, + 0.004717799369245768, + 0.013309359550476074, + 0.011863786727190018, + 0.009423144161701202, + 0.0055276863276958466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.02978663519024849 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot captures a vast, sun-scorched landscape as the camera begins a slow, deliberate movement forward, tracking a lone figure on their journey. Their features are obscured by the deep hood of a heavy, forest-green wool cloak, its frayed hems, as they walk away from the camera with a steady, determined gait. They hold a rough wooden staff in one hand and carry a worn leather bag slung across their body, foreshadowing the long journey ahead. They are approaching a massive, ancient stone archway, constructed from massive blocks of weathered gray granite, carved with faded, intricate carvings and dotted with patches of green moss. The archway towers above the statues, silhouetted against a brilliant, clear blue sky. The entire scene is drenched in the bright, cinematic sunlight characteristic of the late afternoon golden hour, casting the travelers' long, sharp shadows across the dusty, ochre-colored path. As the figure steadily makes its way through the vast opening, their body briefly silhouetted against the light, falling into the cool, deep shadows of the archway before they emerge on the other side, bathed once more in warm, brilliant light, creating a powerful, epic atmosphere of the beginning of a momentous quest.", + "seed": 28180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_2/index_4872_seed_28180_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06318073272705078, + "vlm_logprob": -0.06318073272705078, + "vlm_yes_no_margin": 3.0249999284744264, + "vlm_token_credit": -0.06318073272705078, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03423183783888817 + ], + "opposite_logprob_tokens": [ + -4.034231662750244 + ], + "logprob_reward": -0.03423183783888817, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03423183783888817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26175737380981445, + 0.025124741718173027, + 0.04755394533276558, + 0.07853135466575623, + 0.07161396741867065, + 0.05232645571231842, + 0.04384101554751396, + 0.084676094353199, + 0.20235629379749298, + 0.13221873342990875 + ], + "credit_map": [ + 0.02381153404712677, + 0.018791688606142998, + 0.061528392136096954, + 0.04837647080421448, + 0.011688683182001114, + 0.019437890499830246, + 0.028230899944901466, + 0.04989180713891983, + 0.002276989398524165, + 0.0030059083364903927, + 0.003299555042758584, + 0.003152142046019435, + 0.0021038036793470383, + 0.0020660231821238995, + 0.004835246596485376, + 0.004385074134916067, + 0.0039053233340382576, + 0.004698930308222771, + 0.008824716322124004, + 0.006768429651856422, + 0.0026073132175952196, + 0.0033872907515615225, + 0.008612644858658314, + 0.008749295026063919, + 0.004062855150550604, + 0.004835609346628189, + 0.026948995888233185, + 0.003994929138571024, + 0.0026072582695633173, + 0.004105763044208288, + 0.0058927955105900764, + 0.026083149015903473, + 0.0073443143628537655, + 0.020910531282424927, + 0.00878248829394579, + 0.0052663106471300125, + 0.004443870857357979, + 0.00832226313650608, + 0.004684661515057087, + 0.011859527789056301, + 0.006529978942126036, + 0.01562348660081625, + 0.004730910528451204, + 0.005731859244406223, + 0.003414852311834693, + 0.0038433659356087446, + 0.005705625284463167, + 0.006746373604983091, + 0.0046798475086688995, + 0.005741133354604244, + 0.011081638745963573, + 0.007948790676891804, + 0.002840927103534341, + 0.0023082171101123095, + 0.004017716273665428, + 0.0052227433770895, + 0.012937556952238083, + 0.014267834834754467, + 0.014978569000959396, + 0.006959951017051935, + 0.00706054363399744, + 0.005332577042281628, + 0.009738022461533546, + 0.013401040807366371, + 0.03004330024123192, + 0.04007495567202568, + 0.037941765040159225, + 0.022736895829439163, + 0.012283271178603172, + 0.015963014215230942, + 0.017465591430664062, + 0.02584749646484852, + 0.02051416039466858, + 0.01305339764803648, + 0.02570241317152977, + 0.016953671351075172, + 0.007481997832655907, + 0.01424587145447731, + 0.018184460699558258, + 0.016082763671875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03423183783888817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285047352313995 + ], + "opposite_logprob_tokens": [ + -2.017850399017334 + ], + "logprob_reward": -0.14285047352313995, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285047352313995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10562771558761597, + 0.023769257590174675, + 0.028224948793649673, + 0.032240819185972214, + 0.017854096367955208, + 0.13843654096126556, + 0.07675883919000626, + 0.11355860531330109, + 0.15426142513751984, + 0.30926769971847534 + ], + "credit_map": [ + 0.008524131961166859, + 0.009545117616653442, + 0.019431104883551598, + 0.01662621833384037, + 0.006981686223298311, + 0.00944396760314703, + 0.013123363256454468, + 0.02195213921368122, + 0.002667004242539406, + 0.0030491165816783905, + 0.003655657870694995, + 0.00347780529409647, + 0.001951126498170197, + 0.0023024543188512325, + 0.002955003408715129, + 0.0037110913544893265, + 0.0029186976607888937, + 0.002540083834901452, + 0.004412088543176651, + 0.004122336860746145, + 0.0022520574275404215, + 0.0026234558317810297, + 0.004822274670004845, + 0.0045339567586779594, + 0.0030995579436421394, + 0.005226959474384785, + 0.005960941314697266, + 0.0024598995223641396, + 0.0027256179600954056, + 0.004153439309448004, + 0.004519026260823011, + 0.004095383454114199, + 0.0026497640646994114, + 0.0025921445339918137, + 0.0032545379363000393, + 0.002083712024614215, + 0.001143819303251803, + 0.002240563742816448, + 0.0019392630783841014, + 0.0019502942450344563, + 0.012678371742367744, + 0.04699479416012764, + 0.008091602474451065, + 0.014022469520568848, + 0.0074114082381129265, + 0.008013184182345867, + 0.012198924086987972, + 0.029025811702013016, + 0.016645673662424088, + 0.015559131279587746, + 0.0232793390750885, + 0.005575668532401323, + 0.003906229045242071, + 0.0032098155934363604, + 0.00436191214248538, + 0.004221081268042326, + 0.016195226460695267, + 0.02007715217769146, + 0.020875390619039536, + 0.009968303143978119, + 0.0067427875474095345, + 0.008252612315118313, + 0.011615014635026455, + 0.019832132384181023, + 0.037668634206056595, + 0.0160890631377697, + 0.04606829211115837, + 0.01250991690903902, + 0.009484156034886837, + 0.009367709048092365, + 0.009400668554008007, + 0.013673003762960434, + 0.06253299862146378, + 0.03175199776887894, + 0.07664559781551361, + 0.035652294754981995, + 0.01585482805967331, + 0.02511485107243061, + 0.033143311738967896, + 0.02857184037566185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.14285047352313995 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620812252163887 + ], + "opposite_logprob_tokens": [ + -2.8120813369750977 + ], + "logprob_reward": -0.0620812252163887, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620812252163887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1619417369365692, + 0.0608225092291832, + 0.03028050623834133, + 0.055824510753154755, + 0.11912229657173157, + 0.2079821079969406, + 0.10066401958465576, + 0.0558956153690815, + 0.09576599299907684, + 0.11170069128274918 + ], + "credit_map": [ + 0.013952727429568768, + 0.014103165827691555, + 0.02818646840751171, + 0.03010585904121399, + 0.010957123711705208, + 0.01589072123169899, + 0.01647544465959072, + 0.032270219177007675, + 0.007959038950502872, + 0.007141788024455309, + 0.013586482964456081, + 0.007778890896588564, + 0.0038858489133417606, + 0.006147134583443403, + 0.0056676403619349, + 0.008655685931444168, + 0.0021252818405628204, + 0.002937775105237961, + 0.004549834877252579, + 0.004304084461182356, + 0.002800778718665242, + 0.0019280980341136456, + 0.006952937692403793, + 0.004681717604398727, + 0.004865415394306183, + 0.006689248140901327, + 0.010103466920554638, + 0.004153515677899122, + 0.0027779340744018555, + 0.00487139169126749, + 0.005135536193847656, + 0.017228007316589355, + 0.010175133123993874, + 0.018738863989710808, + 0.017876407131552696, + 0.015636194497346878, + 0.005540867801755667, + 0.008852912113070488, + 0.02125842310488224, + 0.021043486893177032, + 0.017892928794026375, + 0.0703420639038086, + 0.012936484068632126, + 0.018912039697170258, + 0.009952534921467304, + 0.011905570514500141, + 0.01563735119998455, + 0.050403136759996414, + 0.01768599823117256, + 0.023129042237997055, + 0.036716558039188385, + 0.006053281016647816, + 0.003721383633092046, + 0.0031631512101739645, + 0.004630096722394228, + 0.00556451128795743, + 0.006686387117952108, + 0.006872331257909536, + 0.0069337389431893826, + 0.005815926473587751, + 0.0036203996278345585, + 0.004766328260302544, + 0.007948894053697586, + 0.013251610100269318, + 0.01615246757864952, + 0.01664927415549755, + 0.01682060956954956, + 0.011363320052623749, + 0.0055925436317920685, + 0.006875948049128056, + 0.009479990229010582, + 0.012831845320761204, + 0.0189362820237875, + 0.021721435710787773, + 0.01922442950308323, + 0.020793283358216286, + 0.006844894960522652, + 0.010265757329761982, + 0.007357558701187372, + 0.006557059474289417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0620812252163887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369602560997009 + ], + "opposite_logprob_tokens": [ + -3.408695936203003 + ], + "logprob_reward": -0.03369602560997009, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369602560997009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23954127728939056, + 0.04118192195892334, + 0.031110169366002083, + 0.13330358266830444, + 0.033180080354213715, + 0.2554416060447693, + 0.08212712407112122, + 0.053032707422971725, + 0.07896570861339569, + 0.05211574211716652 + ], + "credit_map": [ + 0.02385893650352955, + 0.02352631837129593, + 0.05596998706459999, + 0.041307248175144196, + 0.011108431033790112, + 0.01598120480775833, + 0.025365572422742844, + 0.04242359846830368, + 0.004541271366178989, + 0.0041662282310426235, + 0.011012865230441093, + 0.005202184896916151, + 0.002499726600944996, + 0.002519146306440234, + 0.004853630438446999, + 0.0063868737779557705, + 0.0027093146927654743, + 0.003702903399243951, + 0.004720573779195547, + 0.0032216315157711506, + 0.00280786887742579, + 0.0028064665384590626, + 0.006057021673768759, + 0.005084393545985222, + 0.014506367035210133, + 0.015912719070911407, + 0.012921466492116451, + 0.009535073302686214, + 0.00768654840067029, + 0.012338674627244473, + 0.01571989618241787, + 0.04468285292387009, + 0.00350996688939631, + 0.003546925261616707, + 0.005051558371633291, + 0.0038170272018760443, + 0.002823290880769491, + 0.0038551075849682093, + 0.004602859728038311, + 0.005973348394036293, + 0.024362212046980858, + 0.0884566605091095, + 0.014300445094704628, + 0.025572817772626877, + 0.01402177195996046, + 0.014737256802618504, + 0.022343846037983894, + 0.0516466349363327, + 0.013561797328293324, + 0.020079655572772026, + 0.026394253596663475, + 0.005932222586125135, + 0.0038065786939114332, + 0.0030559152364730835, + 0.004156576469540596, + 0.005140137858688831, + 0.006060318555682898, + 0.0054204268380999565, + 0.006901953835040331, + 0.004997566342353821, + 0.0027872889768332243, + 0.005581763107329607, + 0.008690372109413147, + 0.012593025341629982, + 0.0157352052628994, + 0.007970917038619518, + 0.0291599091142416, + 0.00486082723364234, + 0.004932267591357231, + 0.004575738217681646, + 0.004750343970954418, + 0.006980505306273699, + 0.008918115869164467, + 0.0057059843093156815, + 0.011427399702370167, + 0.005852049682289362, + 0.0031437764409929514, + 0.006247001700103283, + 0.006018166895955801, + 0.004803250078111887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.03369602560997009 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy forest-green wool cloak with frayed hems and a deep hood obscuring their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rough wooden staff and carrying a worn leather bag slung across their body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a massive ancient stone archway made of weathered gray granite with faded carvings and green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and deliberately forward, tracking the figure as they walk away from the viewer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304410144686699 + ], + "opposite_logprob_tokens": [ + -3.168044090270996 + ], + "logprob_reward": -0.04304410144686699, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304410144686699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08267973363399506, + 0.043192874640226364, + 0.06380841881036758, + 0.057806432247161865, + 0.051828496158123016, + 0.07493507117033005, + 0.1886187046766281, + 0.14425861835479736, + 0.17440737783908844, + 0.11846424639225006 + ], + "credit_map": [ + 0.007601662538945675, + 0.00781870074570179, + 0.012472671456634998, + 0.008744627237319946, + 0.007764934096485376, + 0.006297554820775986, + 0.021243033930659294, + 0.010736549273133278, + 0.005129688885062933, + 0.005094362888485193, + 0.008370875380933285, + 0.005985264200717211, + 0.0031893502455204725, + 0.003998671192675829, + 0.004948065150529146, + 0.0064765955321490765, + 0.005766098387539387, + 0.005814474541693926, + 0.01094367727637291, + 0.01306873932480812, + 0.0034962687641382217, + 0.0058852736838161945, + 0.010199584066867828, + 0.008634301833808422, + 0.005213436204940081, + 0.006167043000459671, + 0.020405713468790054, + 0.004544631578028202, + 0.0031570896971970797, + 0.003208255860954523, + 0.006109196692705154, + 0.009001066908240318, + 0.005332891829311848, + 0.006383327767252922, + 0.007473925594240427, + 0.008550556376576424, + 0.004357633180916309, + 0.006704561412334442, + 0.006676181685179472, + 0.006349421106278896, + 0.0068004135973751545, + 0.011744014918804169, + 0.015227355994284153, + 0.008613322861492634, + 0.0060162777081131935, + 0.010706022381782532, + 0.009079336188733578, + 0.006748329848051071, + 0.019272416830062866, + 0.04898359999060631, + 0.06638214737176895, + 0.009062288329005241, + 0.011712650768458843, + 0.008938225917518139, + 0.011924401856958866, + 0.01234297826886177, + 0.016294686123728752, + 0.01750038005411625, + 0.02397899702191353, + 0.014823861420154572, + 0.017068855464458466, + 0.010335862636566162, + 0.016045762225985527, + 0.028210211545228958, + 0.042977459728717804, + 0.012003371492028236, + 0.05486642196774483, + 0.018271423876285553, + 0.009310895577073097, + 0.010836482979357243, + 0.012199359014630318, + 0.013941947370767593, + 0.01882326602935791, + 0.016005106270313263, + 0.023650988936424255, + 0.015580235980451107, + 0.009911867789924145, + 0.013721204362809658, + 0.010055864229798317, + 0.010715716518461704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.04304410144686699 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_tensor([2873, 2873])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_tensor([2873, 2873])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..175d9fb0c03c152d4e2af619a5eb9fd105a674a6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_tensor([2873, 2873])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic medium tracking shot follows the skier in a smooth, fluid motion, keeping pace with him. The skier is dressed in a vibrant royal blue ski suit, black ski pants, and a protective white helmet, the visor lenses reflecting a rainbow of colors in the sunlight. He is speeding down a vast, untouched slope, the deep snow shimmering in the bright winter sun. The skier shifts his weight from side to side with a powerful and beautiful rhythm. With each clean turn, the wide edges of his skis carve deep into the soft snow, unleashing a spectacular curtain of white powder behind him. The mist briefly fills the air, catching the sunlight and creating a translucent halo. As he accelerates downward, the mist he stirs up grows thicker, nearly engulfing his lower body. The entire scene is filled with a sense of speed, freedom, and exhilarating energy, creating the atmosphere of an ultimate, immersive skiing experience.", + "seed": 22350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_2873_seed_22350_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06788109922781586, + "vlm_logprob": -0.06788109922781586, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06788109922781586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1290019452571869 + ], + "opposite_logprob_tokens": [ + -2.3790018558502197 + ], + "logprob_reward": -0.1290019452571869, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1290019452571869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06525969505310059, + 0.32769638299942017, + 0.08625049889087677, + 0.08349993079900742, + 0.07898996025323868, + 0.1812548041343689, + 0.03544992208480835, + 0.028275838121771812, + 0.05822737514972687, + 0.05509553104639053 + ], + "credit_map": [ + 0.004776784218847752, + 0.009618205018341541, + 0.010183916427195072, + 0.007028987631201744, + 0.01479520183056593, + 0.007250158581882715, + 0.0051173376850783825, + 0.006489102728664875, + 0.0229363851249218, + 0.096063032746315, + 0.10129503160715103, + 0.03788389638066292, + 0.02411387301981449, + 0.020845167338848114, + 0.01098681055009365, + 0.013572199270129204, + 0.007049368694424629, + 0.019132763147354126, + 0.014595980755984783, + 0.007855918258428574, + 0.012878810055553913, + 0.008497086353600025, + 0.0099337762221694, + 0.006306793075054884, + 0.007299864664673805, + 0.023890139535069466, + 0.018442412838339806, + 0.009452811442315578, + 0.006577755324542522, + 0.007074043620377779, + 0.0057674371637403965, + 0.004995472729206085, + 0.00979885458946228, + 0.013942900113761425, + 0.010153441689908504, + 0.007074118126183748, + 0.009720461443066597, + 0.009513813070952892, + 0.010312573052942753, + 0.00847379956394434, + 0.01274033635854721, + 0.027585718780755997, + 0.034052710980176926, + 0.02674400433897972, + 0.02961454540491104, + 0.019030295312404633, + 0.018169626593589783, + 0.013317575678229332, + 0.004248843062669039, + 0.008540406823158264, + 0.0049577513709664345, + 0.004286563489586115, + 0.0032651387155056, + 0.003896111622452736, + 0.0030320053920149803, + 0.003223103005439043, + 0.0030451188795268536, + 0.004167924169450998, + 0.004503278993070126, + 0.004125668201595545, + 0.004295850172638893, + 0.003159848041832447, + 0.001666259835474193, + 0.003311889711767435, + 0.004895837977528572, + 0.013665402308106422, + 0.007888407446444035, + 0.008708380162715912, + 0.0077216667123138905, + 0.005725715309381485, + 0.004721729084849358, + 0.0049002389423549175, + 0.003621229901909828, + 0.008805380202829838, + 0.005675282794982195, + 0.0072454470209777355, + 0.009424650110304356, + 0.005605427548289299, + 0.00793833564966917, + 0.006779777351766825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.1290019452571869 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798926442861557 + ], + "opposite_logprob_tokens": [ + -1.804892659187317 + ], + "logprob_reward": -0.1798926442861557, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798926442861557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13040871918201447, + 0.2160123735666275, + 0.05889372155070305, + 0.12933997809886932, + 0.058171991258859634, + 0.0790533497929573, + 0.056933946907520294, + 0.1847091019153595, + 0.04665669798851013, + 0.03982015326619148 + ], + "credit_map": [ + 0.008560684509575367, + 0.01917402260005474, + 0.012709836475551128, + 0.01215409766882658, + 0.01952793262898922, + 0.031170766800642014, + 0.012932004407048225, + 0.014179367572069168, + 0.014700889587402344, + 0.06229470297694206, + 0.06747445464134216, + 0.025555286556482315, + 0.015940871089696884, + 0.011504978872835636, + 0.008980508893728256, + 0.009560639038681984, + 0.003978275693953037, + 0.008324654772877693, + 0.010694734752178192, + 0.0065577467903494835, + 0.009549512527883053, + 0.009302967227995396, + 0.005778837017714977, + 0.004706983920186758, + 0.010306373238563538, + 0.02693559043109417, + 0.03064034692943096, + 0.014126117341220379, + 0.01449596881866455, + 0.010914642363786697, + 0.013336417265236378, + 0.00858450960367918, + 0.007511703763157129, + 0.011118264868855476, + 0.007374954875558615, + 0.0054608723148703575, + 0.007693394552916288, + 0.008314402773976326, + 0.004801751580089331, + 0.005896641872823238, + 0.004886087961494923, + 0.008616802282631397, + 0.01622186414897442, + 0.011937177740037441, + 0.010902848094701767, + 0.005466110538691282, + 0.011310122907161713, + 0.009712334722280502, + 0.007331029511988163, + 0.010686849243938923, + 0.009267664514482021, + 0.006502558011561632, + 0.006194750312715769, + 0.006199265830218792, + 0.005356964655220509, + 0.005394854582846165, + 0.01811959780752659, + 0.0298566035926342, + 0.04086931049823761, + 0.03443775698542595, + 0.023228170350193977, + 0.015536582097411156, + 0.009531675837934017, + 0.013129387982189655, + 0.0028890492394566536, + 0.010709619149565697, + 0.005339621100574732, + 0.007144552189856768, + 0.007002439349889755, + 0.006238714326173067, + 0.0036330644506961107, + 0.003699634922668338, + 0.0027603651396930218, + 0.0060035791248083115, + 0.0040338910184800625, + 0.005255065858364105, + 0.0063973115757107735, + 0.005361939314752817, + 0.004612209741026163, + 0.005395785439759493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.1798926442861557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614442190155387 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.003614442190155387, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614442190155387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08647128194570541, + 0.20554231107234955, + 0.13560360670089722, + 0.10276389122009277, + 0.10961522161960602, + 0.14767612516880035, + 0.07110165059566498, + 0.065254807472229, + 0.03786011412739754, + 0.03811102733016014 + ], + "credit_map": [ + 0.007364985998719931, + 0.01274274941533804, + 0.011041923426091671, + 0.011696103028953075, + 0.016736386343836784, + 0.00944857019931078, + 0.009101232513785362, + 0.008339324966073036, + 0.013966026715934277, + 0.05826479196548462, + 0.06264673918485641, + 0.024245355278253555, + 0.015767334029078484, + 0.013158307410776615, + 0.007919460535049438, + 0.009574279189109802, + 0.011274365708231926, + 0.031147850677371025, + 0.031096283346414566, + 0.01976427249610424, + 0.0108925960958004, + 0.01302921399474144, + 0.008543082512915134, + 0.009855934418737888, + 0.009319168515503407, + 0.011482907459139824, + 0.020058266818523407, + 0.023028191179037094, + 0.009591486304998398, + 0.009660480543971062, + 0.008898613043129444, + 0.010724762454628944, + 0.01447070762515068, + 0.019939100369811058, + 0.013830985873937607, + 0.01331462524831295, + 0.01263717282563448, + 0.014841347932815552, + 0.009692458435893059, + 0.010888815857470036, + 0.01161368377506733, + 0.01941368542611599, + 0.02330373227596283, + 0.018396295607089996, + 0.02268066443502903, + 0.021544089540839195, + 0.019761253148317337, + 0.01096271350979805, + 0.006268375087529421, + 0.009655831381678581, + 0.013320562429726124, + 0.009233859367668629, + 0.009646274149417877, + 0.008628337644040585, + 0.005936233792454004, + 0.008412167429924011, + 0.005844517610967159, + 0.009995554573833942, + 0.007888764142990112, + 0.010676506906747818, + 0.010629734955728054, + 0.00772048020735383, + 0.004906744696199894, + 0.007592500653117895, + 0.0030841773841530085, + 0.008453428745269775, + 0.0063879783265292645, + 0.005954020190984011, + 0.004382571671158075, + 0.0033244530204683542, + 0.003177821170538664, + 0.0030956577975302935, + 0.003250954207032919, + 0.008997714146971703, + 0.005542654078453779, + 0.0041637420654296875, + 0.0050030965358018875, + 0.004271615296602249, + 0.003605600679293275, + 0.003275645198300481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.003614442190155387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614679677411914 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003614679677411914, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614679677411914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08390527218580246, + 0.23495376110076904, + 0.06932517886161804, + 0.09048353880643845, + 0.05697052925825119, + 0.22934319078922272, + 0.05450106039643288, + 0.039740003645420074, + 0.09408444911241531, + 0.046692993491888046 + ], + "credit_map": [ + 0.006577666383236647, + 0.015761038288474083, + 0.008679072372615337, + 0.006061714608222246, + 0.025470757856965065, + 0.008970503695309162, + 0.0064951214008033276, + 0.005889402702450752, + 0.01652964949607849, + 0.0706554502248764, + 0.07417123019695282, + 0.027747122570872307, + 0.017294591292738914, + 0.01297822967171669, + 0.007130759768188, + 0.008446725085377693, + 0.006646749563515186, + 0.016724344342947006, + 0.014935420826077461, + 0.006902194116264582, + 0.009199686348438263, + 0.006447566207498312, + 0.004355169367045164, + 0.004114051349461079, + 0.009415291249752045, + 0.026171071454882622, + 0.01483305636793375, + 0.009833081625401974, + 0.008963020518422127, + 0.008372336626052856, + 0.006575922481715679, + 0.00631976081058383, + 0.0058306423015892506, + 0.006939017679542303, + 0.006679379381239414, + 0.008617997169494629, + 0.008540819399058819, + 0.006188950501382351, + 0.007396342698484659, + 0.006777381058782339, + 0.017508938908576965, + 0.03719460219144821, + 0.039778947830200195, + 0.03188076987862587, + 0.024556690827012062, + 0.022953497245907784, + 0.023831481114029884, + 0.031638264656066895, + 0.0033490099012851715, + 0.005650168284773827, + 0.011363103985786438, + 0.011666163802146912, + 0.005644544027745724, + 0.00485354894772172, + 0.008359121158719063, + 0.003615402150899172, + 0.003437206381931901, + 0.006439459044486284, + 0.010054838843643665, + 0.008104716427624226, + 0.004267975687980652, + 0.002194940345361829, + 0.0021323668770492077, + 0.0031084988731890917, + 0.005179665517061949, + 0.014408967457711697, + 0.0096510611474514, + 0.008413875475525856, + 0.01735534332692623, + 0.016165897250175476, + 0.008723168633878231, + 0.014186477288603783, + 0.0032541160471737385, + 0.003516154130920768, + 0.006370941177010536, + 0.005546548869460821, + 0.009384808130562305, + 0.006210298277437687, + 0.006699537392705679, + 0.005710590165108442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.003614679677411914 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328178472816944 + ], + "opposite_logprob_tokens": [ + -3.7732818126678467 + ], + "logprob_reward": -0.02328178472816944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328178472816944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09017051756381989, + 0.070511095225811, + 0.0627593994140625, + 0.14773474633693695, + 0.09014155715703964, + 0.24256227910518646, + 0.05774892866611481, + 0.026223449036478996, + 0.06367961317300797, + 0.14846840500831604 + ], + "credit_map": [ + 0.007117872126400471, + 0.012239273637533188, + 0.013454162515699863, + 0.011643384583294392, + 0.020313432440161705, + 0.011619088239967823, + 0.008894873782992363, + 0.004888427909463644, + 0.006024295464158058, + 0.007399733643978834, + 0.009236467070877552, + 0.011188728734850883, + 0.00888277031481266, + 0.009890751913189888, + 0.007590294815599918, + 0.010298062115907669, + 0.006440525408834219, + 0.013400275260210037, + 0.012111611664295197, + 0.006336119491606951, + 0.008702940307557583, + 0.0060425191186368465, + 0.004408190026879311, + 0.005317213013768196, + 0.011851387098431587, + 0.02776637114584446, + 0.03774362802505493, + 0.018921364098787308, + 0.016255702823400497, + 0.012166805565357208, + 0.012115095742046833, + 0.010914377868175507, + 0.011030904948711395, + 0.013102659024298191, + 0.009382904507219791, + 0.009698993526399136, + 0.01620536483824253, + 0.01208422239869833, + 0.009251761250197887, + 0.00938475038856268, + 0.0147163150832057, + 0.02776995673775673, + 0.042986489832401276, + 0.033979494124650955, + 0.036779433488845825, + 0.021141473203897476, + 0.043749645352363586, + 0.02143947407603264, + 0.006496646907180548, + 0.014682436361908913, + 0.009041464887559414, + 0.0065731508657336235, + 0.005730103235691786, + 0.006318793166428804, + 0.004706451669335365, + 0.004199877381324768, + 0.0023242279421538115, + 0.0035387184470891953, + 0.005148133262991905, + 0.004886167589575052, + 0.0028633372858166695, + 0.0026231827214360237, + 0.0020251022651791573, + 0.0028145776595920324, + 0.004999299068003893, + 0.010976250283420086, + 0.014119063504040241, + 0.010132256895303726, + 0.005724584683775902, + 0.006861396133899689, + 0.005608612671494484, + 0.005258146673440933, + 0.009610275737941265, + 0.013383946381509304, + 0.014574004337191582, + 0.0204479917883873, + 0.027402056381106377, + 0.02125714346766472, + 0.022384537383913994, + 0.019408434629440308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.02328178472816944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium tracking shot follows the skier in a smooth, fluid motion, keeping pace with him. The skier is dressed in a vibrant royal blue ski suit, black ski pants, and a protective white helmet, the visor lenses reflecting a rainbow of colors in the sunlight. He is speeding down a vast, untouched slope, the deep snow shimmering in the bright winter sun. The skier shifts his weight from side to side with a powerful and beautiful rhythm. With each clean turn, the wide edges of his skis carve deep into the soft snow, unleashing a spectacular curtain of white powder behind him. The mist briefly fills the air, catching the sunlight and creating a translucent halo. As he accelerates downward, the mist he stirs up grows thicker, nearly engulfing his lower body. The entire scene is filled with a sense of speed, freedom, and exhilarating energy, creating the atmosphere of an ultimate, immersive skiing experience.", + "seed": 22351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_2873_seed_22351_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07788714347407222, + "vlm_logprob": -0.07788714347407222, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.07788714347407222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14977896213531494 + ], + "opposite_logprob_tokens": [ + -2.2747788429260254 + ], + "logprob_reward": -0.14977896213531494, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.14977896213531494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37122634053230286, + 0.05310594290494919, + 0.06353811919689178, + 0.06054198369383812, + 0.17222115397453308, + 0.0618010088801384, + 0.05834738165140152, + 0.04872431606054306, + 0.08171138167381287, + 0.028782347217202187 + ], + "credit_map": [ + 0.04259854927659035, + 0.08072733879089355, + 0.031020870432257652, + 0.032325100153684616, + 0.07839588820934296, + 0.03298022970557213, + 0.034454815089702606, + 0.03872354328632355, + 0.006907925009727478, + 0.008492803201079369, + 0.0045890421606600285, + 0.005839615128934383, + 0.007468736730515957, + 0.0069571067579090595, + 0.006532734259963036, + 0.006317976396530867, + 0.009188075549900532, + 0.005404496565461159, + 0.007568200584501028, + 0.006839737296104431, + 0.0072430954314768314, + 0.011676369234919548, + 0.00838551763445139, + 0.007232623640447855, + 0.010608028620481491, + 0.0048006996512413025, + 0.007052987348288298, + 0.010860729962587357, + 0.006475102622061968, + 0.00743445660918951, + 0.0067406040616333485, + 0.006569376215338707, + 0.019808536395430565, + 0.025456758216023445, + 0.021621178835630417, + 0.01846959814429283, + 0.02105695754289627, + 0.03557104244828224, + 0.015459861606359482, + 0.014777230098843575, + 0.006643554661422968, + 0.00960597861558199, + 0.007054605521261692, + 0.007897501811385155, + 0.00948114600032568, + 0.0075856223702430725, + 0.0075179534032940865, + 0.006014648824930191, + 0.0046287113800644875, + 0.006712642498314381, + 0.0055279177613556385, + 0.004734579473733902, + 0.014617892913520336, + 0.004758014343678951, + 0.011915281414985657, + 0.005452343728393316, + 0.004430057480931282, + 0.006524707656353712, + 0.00796534027904272, + 0.006780924275517464, + 0.008326697163283825, + 0.003986946307122707, + 0.005278376862406731, + 0.005431266035884619, + 0.015612244606018066, + 0.012337933294475079, + 0.011757781729102135, + 0.00797246117144823, + 0.012117229402065277, + 0.007875639013946056, + 0.0056070974096655846, + 0.008430995978415012, + 0.003230343572795391, + 0.0038839541375637054, + 0.0029514171183109283, + 0.0024460756685584784, + 0.0042849453166127205, + 0.002895982703194022, + 0.0057179625146090984, + 0.003371665719896555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.14977896213531494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158971846103668 + ], + "opposite_logprob_tokens": [ + -1.7015897035598755 + ], + "logprob_reward": -0.20158971846103668, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158971846103668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2560601830482483, + 0.08243706077337265, + 0.10137728601694107, + 0.09764828532934189, + 0.08827538043260574, + 0.09247954934835434, + 0.07058838754892349, + 0.1050787940621376, + 0.06217307597398758, + 0.043881986290216446 + ], + "credit_map": [ + 0.029297292232513428, + 0.05396849662065506, + 0.02601306512951851, + 0.01722196862101555, + 0.061819080263376236, + 0.02285115234553814, + 0.02107863686978817, + 0.023810477927327156, + 0.010846340097486973, + 0.015074913389980793, + 0.006674544885754585, + 0.008503911085426807, + 0.014432336203753948, + 0.010393154807388783, + 0.007225212175399065, + 0.009286654181778431, + 0.014851296320557594, + 0.011570462957024574, + 0.014807216823101044, + 0.009065266698598862, + 0.011257033795118332, + 0.017239073291420937, + 0.01046182494610548, + 0.012125110253691673, + 0.019448736682534218, + 0.011744153685867786, + 0.01307659037411213, + 0.008533002808690071, + 0.01535686757415533, + 0.010784693993628025, + 0.009941836819052696, + 0.008762395940721035, + 0.007557055447250605, + 0.009304151870310307, + 0.008611742407083511, + 0.01689429208636284, + 0.009560997597873211, + 0.01767894998192787, + 0.009812209755182266, + 0.008855978958308697, + 0.007842637598514557, + 0.013562140986323357, + 0.011952478438615799, + 0.01147451065480709, + 0.0194158386439085, + 0.009126364253461361, + 0.008559022098779678, + 0.010546553879976273, + 0.00532578956335783, + 0.009459616616368294, + 0.008638259954750538, + 0.005042444448918104, + 0.018242137506604195, + 0.005124005023390055, + 0.012370085343718529, + 0.006386046763509512, + 0.010566158220171928, + 0.016336211934685707, + 0.012033498845994473, + 0.009791900403797626, + 0.027562960982322693, + 0.01337958313524723, + 0.009154162369668484, + 0.006254314444959164, + 0.007492662873119116, + 0.007878044620156288, + 0.009865125641226768, + 0.005768319126218557, + 0.01055187825113535, + 0.00516185536980629, + 0.008378138765692711, + 0.007077050395309925, + 0.006505853030830622, + 0.0058596194721758366, + 0.004128547850996256, + 0.0032679112628102303, + 0.005305546801537275, + 0.00509630935266614, + 0.009028957225382328, + 0.0046892412938177586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.20158971846103668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628831081092358 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628831081092358, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628831081092358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2716282904148102, + 0.052980948239564896, + 0.0805409774184227, + 0.043369393795728683, + 0.22405391931533813, + 0.14443320035934448, + 0.05911901965737343, + 0.08077045530080795, + 0.02159125916659832, + 0.02151259034872055 + ], + "credit_map": [ + 0.03238322213292122, + 0.05293142423033714, + 0.027796074748039246, + 0.01908816583454609, + 0.06526010483503342, + 0.023438606411218643, + 0.025591179728507996, + 0.025139499455690384, + 0.005411046091467142, + 0.008951770141720772, + 0.005402460694313049, + 0.006189373321831226, + 0.00992546509951353, + 0.006773654837161303, + 0.004880049265921116, + 0.005447121802717447, + 0.013311265967786312, + 0.008115491829812527, + 0.008814697153866291, + 0.007623050827533007, + 0.008821600116789341, + 0.015781443566083908, + 0.009366162121295929, + 0.008707258850336075, + 0.005412585102021694, + 0.0051330747082829475, + 0.0046839783899486065, + 0.004524468909949064, + 0.005824980325996876, + 0.007550430484116077, + 0.005200948566198349, + 0.005038924515247345, + 0.025685541331768036, + 0.028723474591970444, + 0.02691209316253662, + 0.02421054057776928, + 0.03244807943701744, + 0.04485897347331047, + 0.020825384184718132, + 0.02038983814418316, + 0.01346042100340128, + 0.017148783430457115, + 0.022398727014660835, + 0.01871316134929657, + 0.02506202645599842, + 0.018953988328576088, + 0.014415975660085678, + 0.014280106872320175, + 0.004202011041343212, + 0.008272574283182621, + 0.006771749351173639, + 0.005151679739356041, + 0.014690634794533253, + 0.005007903091609478, + 0.00944102369248867, + 0.005581439007073641, + 0.006182953715324402, + 0.011493975296616554, + 0.012373887933790684, + 0.005651096347719431, + 0.019380871206521988, + 0.0076079401187598705, + 0.008540598675608635, + 0.009539121761918068, + 0.0021898727864027023, + 0.0028956420719623566, + 0.002759966067969799, + 0.0023296400904655457, + 0.003706898307427764, + 0.0017008084105327725, + 0.0026120487600564957, + 0.003396380227059126, + 0.0019718832336366177, + 0.002979125827550888, + 0.0022902125492691994, + 0.0022410547826439142, + 0.004819079302251339, + 0.0019055011216551065, + 0.0031105312518775463, + 0.0021951990202069283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.007628831081092358 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095145035535097 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095145035535097, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095145035535097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11672558635473251, + 0.05025216564536095, + 0.03086748532950878, + 0.04834497347474098, + 0.09799505025148392, + 0.4500649571418762, + 0.04085192084312439, + 0.04956997558474541, + 0.03765766695141792, + 0.07767017930746078 + ], + "credit_map": [ + 0.01146596111357212, + 0.016580339521169662, + 0.012465467676520348, + 0.013322197832167149, + 0.026614787057042122, + 0.013686759397387505, + 0.013080156408250332, + 0.009509929455816746, + 0.0046610780991613865, + 0.009078003466129303, + 0.004737477749586105, + 0.005988135933876038, + 0.007755593862384558, + 0.007951399311423302, + 0.004725745413452387, + 0.005354738794267178, + 0.0032884650863707066, + 0.004057596903294325, + 0.002912944881245494, + 0.004410551395267248, + 0.0035836100578308105, + 0.004476428497582674, + 0.004405016545206308, + 0.003732874756678939, + 0.007222066633403301, + 0.007494278252124786, + 0.005009578075259924, + 0.006302726920694113, + 0.0058609177358448505, + 0.006781987380236387, + 0.004980880301445723, + 0.00469254469498992, + 0.012940024957060814, + 0.012550776824355125, + 0.015487094409763813, + 0.010247203521430492, + 0.012670852243900299, + 0.014903081580996513, + 0.009276890195906162, + 0.009919135831296444, + 0.036738988012075424, + 0.04657109081745148, + 0.08913453668355942, + 0.058517251163721085, + 0.0735296905040741, + 0.05069691315293312, + 0.04445304349064827, + 0.05042346566915512, + 0.003524055704474449, + 0.005759843159466982, + 0.006500461138784885, + 0.004173333756625652, + 0.006596200633794069, + 0.005804687738418579, + 0.005107784178107977, + 0.0033855601213872433, + 0.0042998306453228, + 0.005861163605004549, + 0.009406769648194313, + 0.004233764484524727, + 0.009061275981366634, + 0.006729022599756718, + 0.004970471374690533, + 0.005007682368159294, + 0.004550682846456766, + 0.005920962430536747, + 0.0050309025682508945, + 0.0031372562516480684, + 0.007730296812951565, + 0.0030692305881530046, + 0.0038594333454966545, + 0.004358905367553234, + 0.007110057398676872, + 0.009973541833460331, + 0.007187047507613897, + 0.006597236730158329, + 0.01415829174220562, + 0.0073075625114142895, + 0.016755523160099983, + 0.00858092401176691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004095145035535097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634306065738201 + ], + "opposite_logprob_tokens": [ + -3.6513431072235107 + ], + "logprob_reward": -0.02634306065738201, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634306065738201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14130695164203644, + 0.06896364688873291, + 0.02542100101709366, + 0.02816452458500862, + 0.17003823816776276, + 0.3914854824542999, + 0.024097681045532227, + 0.03268647566437721, + 0.02125067263841629, + 0.09658540040254593 + ], + "credit_map": [ + 0.010968591086566448, + 0.02426169253885746, + 0.015878958627581596, + 0.009403825737535954, + 0.031920772045850754, + 0.017374787479639053, + 0.01215697918087244, + 0.019341329112648964, + 0.01040543895214796, + 0.012470117770135403, + 0.00713147409260273, + 0.00833205133676529, + 0.009532570838928223, + 0.00795081164687872, + 0.0059495773166418076, + 0.007191600743681192, + 0.003018971299752593, + 0.004436516668647528, + 0.002710570814087987, + 0.003538035089150071, + 0.0029166017193347216, + 0.002867409260943532, + 0.002781306393444538, + 0.0031515874434262514, + 0.003768540918827057, + 0.0040218764916062355, + 0.00452668871730566, + 0.0025432531256228685, + 0.003275314811617136, + 0.004381778184324503, + 0.0028729953337460756, + 0.0027740783989429474, + 0.02244659699499607, + 0.02426048181951046, + 0.026818029582500458, + 0.016940616071224213, + 0.020175278186798096, + 0.027043625712394714, + 0.016099290922284126, + 0.01625431701540947, + 0.0285183172672987, + 0.04610924422740936, + 0.07160361856222153, + 0.046501003205776215, + 0.06962068378925323, + 0.043472860008478165, + 0.03822200745344162, + 0.04743775725364685, + 0.0029998780228197575, + 0.00349489226937294, + 0.002608372364193201, + 0.001652065897360444, + 0.004648537840694189, + 0.003098432905972004, + 0.003568060928955674, + 0.0020274424459785223, + 0.003016936592757702, + 0.0038210737984627485, + 0.00596321327611804, + 0.003120174864307046, + 0.006467863917350769, + 0.003749171271920204, + 0.003358929418027401, + 0.003189112525433302, + 0.002908236114308238, + 0.0031331938225775957, + 0.0027154593262821436, + 0.0017193709500133991, + 0.0049821482971310616, + 0.0018924352480098605, + 0.0020900946110486984, + 0.001809735898859799, + 0.00951461587101221, + 0.01286276150494814, + 0.008317804895341396, + 0.008188591338694096, + 0.018504170700907707, + 0.00926682073622942, + 0.019298218190670013, + 0.010632405988872051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.02634306065738201 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium tracking shot follows the skier in a smooth, fluid motion, keeping pace with him. The skier is dressed in a vibrant royal blue ski suit, black ski pants, and a protective white helmet, the visor lenses reflecting a rainbow of colors in the sunlight. He is speeding down a vast, untouched slope, the deep snow shimmering in the bright winter sun. The skier shifts his weight from side to side with a powerful and beautiful rhythm. With each clean turn, the wide edges of his skis carve deep into the soft snow, unleashing a spectacular curtain of white powder behind him. The mist briefly fills the air, catching the sunlight and creating a translucent halo. As he accelerates downward, the mist he stirs up grows thicker, nearly engulfing his lower body. The entire scene is filled with a sense of speed, freedom, and exhilarating energy, creating the atmosphere of an ultimate, immersive skiing experience.", + "seed": 22352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_2873_seed_22352_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05888843252323568, + "vlm_logprob": -0.05888843252323568, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.05888843252323568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08054494857788086 + ], + "opposite_logprob_tokens": [ + -3.205544948577881 + ], + "logprob_reward": -0.08054494857788086, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.08054494857788086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09345497190952301, + 0.027228815481066704, + 0.3634711503982544, + 0.019744466990232468, + 0.0189255028963089, + 0.06385207921266556, + 0.042966265231370926, + 0.0454321950674057, + 0.2910633683204651, + 0.03386126458644867 + ], + "credit_map": [ + 0.008722721599042416, + 0.008244683966040611, + 0.02304544299840927, + 0.02441164292395115, + 0.005549469031393528, + 0.00582161033526063, + 0.011441061273217201, + 0.0062183355912566185, + 0.0018077425193041563, + 0.003935496788471937, + 0.004451303277164698, + 0.004933565389364958, + 0.0029277238063514233, + 0.004304205533117056, + 0.003136399434879422, + 0.0017323758220300078, + 0.01461892481893301, + 0.05221597105264664, + 0.11381299048662186, + 0.13713683187961578, + 0.007719909306615591, + 0.022654615342617035, + 0.009167690761387348, + 0.00614418089389801, + 0.0018874898087233305, + 0.00273073255084455, + 0.004017087165266275, + 0.003805827349424362, + 0.002301187254488468, + 0.0014799347845837474, + 0.0014427079586312175, + 0.002079497091472149, + 0.002646213863044977, + 0.0023488635197281837, + 0.0028636332135647535, + 0.002450279425829649, + 0.001969380071386695, + 0.0016441779443994164, + 0.002145303413271904, + 0.0028576492331922054, + 0.004150735680013895, + 0.004417960066348314, + 0.006828320678323507, + 0.009619983844459057, + 0.005474280100315809, + 0.004582105204463005, + 0.01646300032734871, + 0.012315683998167515, + 0.003302390454337001, + 0.004388834349811077, + 0.0039931549690663815, + 0.010752187110483646, + 0.0037737484090030193, + 0.003686655079945922, + 0.008293861523270607, + 0.004775425884872675, + 0.004065113607794046, + 0.007227031979709864, + 0.00694632763043046, + 0.00754177663475275, + 0.006930110044777393, + 0.0030701232608407736, + 0.005636864807456732, + 0.0040148403495550156, + 0.009957712143659592, + 0.009635018184781075, + 0.10730209201574326, + 0.018370769917964935, + 0.011179126799106598, + 0.0072396378964185715, + 0.11702409386634827, + 0.010354878380894661, + 0.004438304342329502, + 0.0063882796093821526, + 0.004131412599235773, + 0.0049725607968866825, + 0.003837245749309659, + 0.002705636201426387, + 0.004396798089146614, + 0.002991022542119026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.08054494857788086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015867978334427 + ], + "opposite_logprob_tokens": [ + -1.7015868425369263 + ], + "logprob_reward": -0.2015867978334427, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015867978334427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16255909204483032, + 0.026347840204834938, + 0.2576245665550232, + 0.016030319035053253, + 0.0871109664440155, + 0.04107454791665077, + 0.05235172435641289, + 0.020269248634576797, + 0.2928067147731781, + 0.04382489621639252 + ], + "credit_map": [ + 0.0124063016846776, + 0.009182754904031754, + 0.05185329169034958, + 0.04318442568182945, + 0.011898791417479515, + 0.01456402987241745, + 0.011859811842441559, + 0.007609694730490446, + 0.0023580691777169704, + 0.005502363201230764, + 0.003008752129971981, + 0.00292955175973475, + 0.0036165951751172543, + 0.005015604197978973, + 0.0023491906467825174, + 0.00156771345064044, + 0.010926305316388607, + 0.048213109374046326, + 0.09274483472108841, + 0.06542124599218369, + 0.008164380677044392, + 0.02064555510878563, + 0.004371216986328363, + 0.007137952838093042, + 0.001942112809047103, + 0.0019770513754338026, + 0.002310357289388776, + 0.002064669970422983, + 0.0026833650190383196, + 0.0013595159398391843, + 0.0021234231535345316, + 0.001569822896271944, + 0.005703647620975971, + 0.008390170522034168, + 0.015726186335086823, + 0.024431055411696434, + 0.006881379522383213, + 0.0056373863480985165, + 0.010599812492728233, + 0.00974133051931858, + 0.004654437769204378, + 0.005030149128288031, + 0.005489750299602747, + 0.004838109947741032, + 0.00391011917963624, + 0.004586207680404186, + 0.007404860574752092, + 0.005160913337022066, + 0.005353682674467564, + 0.006333667319267988, + 0.008335882797837257, + 0.009657781571149826, + 0.00369922979734838, + 0.0033557091373950243, + 0.007866713218390942, + 0.0077490597032010555, + 0.0016595604829490185, + 0.001902212854474783, + 0.004619313403964043, + 0.004952562041580677, + 0.0014206289779394865, + 0.0010577539214864373, + 0.002515378873795271, + 0.002141838427633047, + 0.010311039164662361, + 0.007641838397830725, + 0.10755066573619843, + 0.019744589924812317, + 0.010318572632968426, + 0.006788844708353281, + 0.11945431679487228, + 0.01099686324596405, + 0.005750822369009256, + 0.007525693159550428, + 0.006618575658649206, + 0.006657947786152363, + 0.005897579248994589, + 0.0034203033428639174, + 0.004872981458902359, + 0.0030809950549155474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.2015867978334427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488613361492753 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.002488613361492753, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488613361492753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04240128770470619, + 0.02385975793004036, + 0.565991222858429, + 0.05685538798570633, + 0.026079880073666573, + 0.03358754515647888, + 0.07717322558164597, + 0.03879641368985176, + 0.09591557085514069, + 0.039339687675237656 + ], + "credit_map": [ + 0.004833751358091831, + 0.0036204145289957523, + 0.005988501012325287, + 0.008378327824175358, + 0.0040318346582353115, + 0.003701354842633009, + 0.00732263782992959, + 0.004524464253336191, + 0.0030097579583525658, + 0.0027553699910640717, + 0.004924073815345764, + 0.005276955664157867, + 0.0023499745875597, + 0.0024046828038990498, + 0.0015418844996020198, + 0.001597058610059321, + 0.0236688069999218, + 0.0947926864027977, + 0.16222000122070312, + 0.20571617782115936, + 0.013459144160151482, + 0.041226375848054886, + 0.014431355521082878, + 0.010476690717041492, + 0.003241278463974595, + 0.0029804836958646774, + 0.021189993247389793, + 0.018777500838041306, + 0.0030976408161222935, + 0.0028333673253655434, + 0.0014759533805772662, + 0.0032591710332781076, + 0.002349262125790119, + 0.0024846079759299755, + 0.00475279288366437, + 0.007762466557323933, + 0.001754175522364676, + 0.001780992140993476, + 0.0025607766583561897, + 0.002634806325659156, + 0.004507873672991991, + 0.005590139422565699, + 0.0030225315131247044, + 0.0037606386467814445, + 0.003080780850723386, + 0.0032451890874654055, + 0.00602570828050375, + 0.004354683682322502, + 0.005742249079048634, + 0.006219957489520311, + 0.006018094718456268, + 0.017851054668426514, + 0.005348697770386934, + 0.004504585638642311, + 0.0247797891497612, + 0.006708801258355379, + 0.00470182066783309, + 0.0044029648415744305, + 0.004632280673831701, + 0.009725870564579964, + 0.004143459722399712, + 0.002671560738235712, + 0.004679330624639988, + 0.0038391242269426584, + 0.00413361145183444, + 0.005362843628972769, + 0.03154759481549263, + 0.007494902703911066, + 0.004110639914870262, + 0.00239888159558177, + 0.03656598553061485, + 0.004301103297621012, + 0.003998893313109875, + 0.008920026011765003, + 0.004243331495672464, + 0.0058219945058226585, + 0.004713225178420544, + 0.0032495218329131603, + 0.00558645511046052, + 0.0028062425553798676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.002488613361492753 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196639310568571 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.002196639310568571, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196639310568571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04571131616830826, + 0.027850013226270676, + 0.2666688561439514, + 0.012695515528321266, + 0.03907705843448639, + 0.029162224382162094, + 0.12275324761867523, + 0.10355601459741592, + 0.2607188820838928, + 0.09180685877799988 + ], + "credit_map": [ + 0.0032124067656695843, + 0.00307824625633657, + 0.012794299982488155, + 0.011548392474651337, + 0.0035766158252954483, + 0.003788126166909933, + 0.004105599131435156, + 0.003607627935707569, + 0.0014912522165104747, + 0.003771586110815406, + 0.005742937792092562, + 0.00677713006734848, + 0.0022274197544902563, + 0.004429402761161327, + 0.0017751698615029454, + 0.0016351157100871205, + 0.0176242608577013, + 0.09603238850831985, + 0.037074390798807144, + 0.03912612423300743, + 0.013941844925284386, + 0.04238651692867279, + 0.008847885765135288, + 0.011635439470410347, + 0.001396886887960136, + 0.00180809514131397, + 0.0021941675804555416, + 0.0017975950613617897, + 0.002002433640882373, + 0.001217503915540874, + 0.0009606231469660997, + 0.0013182094553485513, + 0.0023952003102749586, + 0.0032651852816343307, + 0.007856872864067554, + 0.0129843195900321, + 0.002598786959424615, + 0.0020628857892006636, + 0.003892585402354598, + 0.004021220840513706, + 0.003875122871249914, + 0.0035578799434006214, + 0.0017671777168288827, + 0.003606067970395088, + 0.0029787279199808836, + 0.0038621211424469948, + 0.0059020803309977055, + 0.003613047068938613, + 0.00808021705597639, + 0.0111152408644557, + 0.009490340948104858, + 0.03167389705777168, + 0.011036722920835018, + 0.01123405434191227, + 0.026118842884898186, + 0.014003937132656574, + 0.007367071229964495, + 0.013600146397948265, + 0.0123806893825531, + 0.029969805851578712, + 0.010309862904250622, + 0.006918296683579683, + 0.011427228339016438, + 0.011582914739847183, + 0.00903304759413004, + 0.00738773588091135, + 0.0957970842719078, + 0.017255282029509544, + 0.009358474053442478, + 0.006263106130063534, + 0.10589282959699631, + 0.009731329046189785, + 0.008643781766295433, + 0.018176663666963577, + 0.012184632942080498, + 0.014096108265221119, + 0.0120698232203722, + 0.007530993316322565, + 0.0125650679692626, + 0.006539787631481886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.002196639310568571 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625163532793522 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625163532793522, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625163532793522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052171844989061356, + 0.025648219510912895, + 0.27125030755996704, + 0.024090994149446487, + 0.0191398486495018, + 0.04519667848944664, + 0.08395062386989594, + 0.040556203573942184, + 0.34052979946136475, + 0.09746543318033218 + ], + "credit_map": [ + 0.003372877137735486, + 0.0028718248941004276, + 0.016891703009605408, + 0.014888882637023926, + 0.0027662543579936028, + 0.004726842045783997, + 0.003919940907508135, + 0.0027335199993103743, + 0.001745520974509418, + 0.004868562798947096, + 0.003780702129006386, + 0.0037534222938120365, + 0.0031380625441670418, + 0.004217447247356176, + 0.002182275988161564, + 0.001962227514013648, + 0.017527854070067406, + 0.09240851551294327, + 0.027736904099583626, + 0.061808615922927856, + 0.012982659041881561, + 0.04017367959022522, + 0.008862857706844807, + 0.009749217890202999, + 0.0015561418840661645, + 0.002582261571660638, + 0.007111931219696999, + 0.00664457818493247, + 0.0019070374546572566, + 0.0015735417837277055, + 0.0011678613955155015, + 0.0015476415865123272, + 0.001204628380946815, + 0.001845675054937601, + 0.0037556830793619156, + 0.006087699439376593, + 0.001471438561566174, + 0.0010548534337431192, + 0.002071989234536886, + 0.001647882629185915, + 0.0028845358174294233, + 0.0045456066727638245, + 0.0028416921850293875, + 0.006381040904670954, + 0.003775552613660693, + 0.0036940029822289944, + 0.012600637041032314, + 0.008473609574139118, + 0.0059187887236475945, + 0.008858676999807358, + 0.0072174277156591415, + 0.02121180109679699, + 0.0067205666564404964, + 0.007089011836796999, + 0.017710383981466293, + 0.009223973378539085, + 0.003455196740105748, + 0.008199275471270084, + 0.004942973144352436, + 0.007064474280923605, + 0.005428825970739126, + 0.002691356698051095, + 0.005418550688773394, + 0.0033555508125573397, + 0.013373100198805332, + 0.010373572818934917, + 0.12511159479618073, + 0.022528836503624916, + 0.012313781306147575, + 0.008272793143987656, + 0.1359427571296692, + 0.01261337473988533, + 0.00885283574461937, + 0.01458957139402628, + 0.013378857634961605, + 0.015373853966593742, + 0.016697492450475693, + 0.007259482052177191, + 0.013671955093741417, + 0.007641382049769163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.007625163532793522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium tracking shot follows the skier in a smooth, fluid motion, keeping pace with him. The skier is dressed in a vibrant royal blue ski suit, black ski pants, and a protective white helmet, the visor lenses reflecting a rainbow of colors in the sunlight. He is speeding down a vast, untouched slope, the deep snow shimmering in the bright winter sun. The skier shifts his weight from side to side with a powerful and beautiful rhythm. With each clean turn, the wide edges of his skis carve deep into the soft snow, unleashing a spectacular curtain of white powder behind him. The mist briefly fills the air, catching the sunlight and creating a translucent halo. As he accelerates downward, the mist he stirs up grows thicker, nearly engulfing his lower body. The entire scene is filled with a sense of speed, freedom, and exhilarating energy, creating the atmosphere of an ultimate, immersive skiing experience.", + "seed": 22353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_3/index_2873_seed_22353_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04808468725532293, + "vlm_logprob": -0.04808468725532293, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04808468725532293, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05591507628560066 + ], + "opposite_logprob_tokens": [ + -3.680915117263794 + ], + "logprob_reward": -0.05591507628560066, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05591507628560066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05458984151482582, + 0.04407039284706116, + 0.09645845741033554, + 0.09654735028743744, + 0.16484802961349487, + 0.33561527729034424, + 0.11602488160133362, + 0.038388047367334366, + 0.023237962275743484, + 0.030219756066799164 + ], + "credit_map": [ + 0.005876635666936636, + 0.0044638849794864655, + 0.0051476191729307175, + 0.005441351793706417, + 0.009859658777713776, + 0.008424538187682629, + 0.008157437667250633, + 0.007218713406473398, + 0.006806898862123489, + 0.010576722212135792, + 0.004177412949502468, + 0.004650024231523275, + 0.00509127601981163, + 0.004086730070412159, + 0.004455164074897766, + 0.004226163029670715, + 0.01277276873588562, + 0.007076258771121502, + 0.012876702472567558, + 0.007383429910987616, + 0.011151893064379692, + 0.006194662768393755, + 0.02079070545732975, + 0.018212029710412025, + 0.008248883299529552, + 0.007442096248269081, + 0.010096468962728977, + 0.008136950433254242, + 0.007357697002589703, + 0.010986404493451118, + 0.02196517214179039, + 0.0223136767745018, + 0.013682610355317593, + 0.017918964847922325, + 0.01756887510418892, + 0.02034531719982624, + 0.015965482220053673, + 0.013938634656369686, + 0.025086570531129837, + 0.0403415784239769, + 0.026267264038324356, + 0.03873984143137932, + 0.0524739995598793, + 0.06190882623195648, + 0.02642272785305977, + 0.03145213797688484, + 0.04780659079551697, + 0.050543904304504395, + 0.014276068657636642, + 0.03616341948509216, + 0.002782803028821945, + 0.0038034769240766764, + 0.017739513888955116, + 0.005853867623955011, + 0.015228087082505226, + 0.02017764374613762, + 0.0031171308364719152, + 0.004844202194362879, + 0.004364601336419582, + 0.0046653239987790585, + 0.00421465840190649, + 0.003876423230394721, + 0.008201583288609982, + 0.005104124080389738, + 0.0020625509787350893, + 0.004413853865116835, + 0.002760891802608967, + 0.003931852523237467, + 0.002042447915300727, + 0.0023115475196391344, + 0.0032426700927317142, + 0.0024721489753574133, + 0.0027832896448671818, + 0.004034736193716526, + 0.0031330916099250317, + 0.0036025228910148144, + 0.003502123523503542, + 0.005265153478831053, + 0.0030809524469077587, + 0.004817884881049395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.05591507628560066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037967801094055 + ], + "opposite_logprob_tokens": [ + -1.9103796482086182 + ], + "logprob_reward": -0.16037967801094055, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037967801094055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10182022303342819, + 0.0748041644692421, + 0.07320898771286011, + 0.08309748768806458, + 0.1488216519355774, + 0.30980384349823, + 0.049780990928411484, + 0.052006639540195465, + 0.07703246176242828, + 0.02962356247007847 + ], + "credit_map": [ + 0.012284901924431324, + 0.011593521572649479, + 0.01655278541147709, + 0.009921330027282238, + 0.011627241037786007, + 0.01563302055001259, + 0.010392015799880028, + 0.013815405778586864, + 0.008199521340429783, + 0.014305274933576584, + 0.008083319291472435, + 0.008355631493031979, + 0.008203072473406792, + 0.009012793190777302, + 0.010991167277097702, + 0.007653383072465658, + 0.008962256833910942, + 0.0056196581572294235, + 0.010980522260069847, + 0.0063315792940557, + 0.008034154772758484, + 0.006648267153650522, + 0.01470961607992649, + 0.011922936886548996, + 0.005670651327818632, + 0.006573269609361887, + 0.007358101662248373, + 0.0059213233180344105, + 0.00528740556910634, + 0.007353466469794512, + 0.02243902161717415, + 0.022494245320558548, + 0.014386851340532303, + 0.01855657622218132, + 0.01585300639271736, + 0.018786484375596046, + 0.01435166783630848, + 0.01186675950884819, + 0.02152421325445175, + 0.033496078103780746, + 0.02462696097791195, + 0.03642785921692848, + 0.049185365438461304, + 0.057099275290966034, + 0.024274485185742378, + 0.02869725227355957, + 0.0431685633957386, + 0.046324070543050766, + 0.006418580189347267, + 0.015987519174814224, + 0.00295975967310369, + 0.0038931958843022585, + 0.008109969086945057, + 0.0028451497200876474, + 0.003997770603746176, + 0.005569046828895807, + 0.004545731004327536, + 0.007774618454277515, + 0.0063871112652122974, + 0.010263740085065365, + 0.0042870924808084965, + 0.004419287666678429, + 0.00841020792722702, + 0.005918851587921381, + 0.00907543022185564, + 0.021397175267338753, + 0.006407322362065315, + 0.01117040403187275, + 0.005045605357736349, + 0.00605620164424181, + 0.00977469515055418, + 0.008105628192424774, + 0.0035933058243244886, + 0.004298743326216936, + 0.0035129589959979057, + 0.004267767537385225, + 0.0032224098686128855, + 0.0036160529125481844, + 0.0038211620412766933, + 0.003291158238425851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.16037967801094055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249999929219484 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249999929219484, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249999929219484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06279504299163818, + 0.03524014726281166, + 0.13627620041370392, + 0.04288609325885773, + 0.05231140926480293, + 0.07502815872430801, + 0.4668503403663635, + 0.05687076970934868, + 0.03944258391857147, + 0.03229927644133568 + ], + "credit_map": [ + 0.0063050538301467896, + 0.007749262265861034, + 0.008051014505326748, + 0.005337049253284931, + 0.010027434676885605, + 0.008736497722566128, + 0.007159878499805927, + 0.009428847581148148, + 0.007252417039126158, + 0.009046553634107113, + 0.0035063009709119797, + 0.0028094134759157896, + 0.003438686253502965, + 0.002751049818471074, + 0.0028637039940804243, + 0.003572023008018732, + 0.020462701097130775, + 0.008889238350093365, + 0.019502051174640656, + 0.010216387920081615, + 0.00935766939073801, + 0.009197789244353771, + 0.02658022567629814, + 0.03207012638449669, + 0.0025550946593284607, + 0.00457176985219121, + 0.004034868907183409, + 0.0036026767920702696, + 0.003657085355371237, + 0.004007709212601185, + 0.010438893921673298, + 0.010017994791269302, + 0.00500480318441987, + 0.008587948977947235, + 0.005881340242922306, + 0.006894638296216726, + 0.005731978453695774, + 0.005227435380220413, + 0.005738696549087763, + 0.009244569577276707, + 0.006160423159599304, + 0.008733837865293026, + 0.011892877519130707, + 0.014112654142081738, + 0.005729579366743565, + 0.00678859930485487, + 0.010768132284283638, + 0.010842056013643742, + 0.0596301443874836, + 0.15095984935760498, + 0.008705373853445053, + 0.012478962540626526, + 0.0743139460682869, + 0.023262737318873405, + 0.05817817896604538, + 0.07932113111019135, + 0.0050882757641375065, + 0.007580791600048542, + 0.004737407900393009, + 0.009714927524328232, + 0.008012836799025536, + 0.008651348762214184, + 0.008084191009402275, + 0.005000986624509096, + 0.004046858754009008, + 0.007153140380978584, + 0.004229499492794275, + 0.00495675066486001, + 0.004132263362407684, + 0.004368443042039871, + 0.006399075035005808, + 0.004156553652137518, + 0.002419647527858615, + 0.006100183818489313, + 0.003008123952895403, + 0.0035336092114448547, + 0.004227673169225454, + 0.0051932791247963905, + 0.003314198460429907, + 0.004502561874687672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.005249999929219484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028173536993563175 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.0028173536993563175, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028173536993563175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04336465150117874, + 0.022780241444706917, + 0.1659313291311264, + 0.029677825048565865, + 0.046169813722372055, + 0.045432910323143005, + 0.29754164814949036, + 0.22556857764720917, + 0.08392356336116791, + 0.03960944339632988 + ], + "credit_map": [ + 0.004920290317386389, + 0.0054097287356853485, + 0.004743344150483608, + 0.006363730411976576, + 0.005929779727011919, + 0.005058933049440384, + 0.0051839277148246765, + 0.005754923447966576, + 0.0025506671518087387, + 0.0034533124417066574, + 0.002269928576424718, + 0.0023907339200377464, + 0.003938727080821991, + 0.0023688501678407192, + 0.0031840000301599503, + 0.0026240248698741198, + 0.021697457879781723, + 0.011527573689818382, + 0.027893254533410072, + 0.012220657430589199, + 0.011702395975589752, + 0.008857166394591331, + 0.03271825984120369, + 0.039314571768045425, + 0.003938274458050728, + 0.0044182282872498035, + 0.0030232130084186792, + 0.004248464945703745, + 0.0024339936207979918, + 0.004065535496920347, + 0.004319882020354271, + 0.003230235306546092, + 0.0045207832008600235, + 0.005656372290104628, + 0.0055870055221021175, + 0.0060317013412714005, + 0.0048622433096170425, + 0.00415931548923254, + 0.005724215880036354, + 0.009628179483115673, + 0.003607516409829259, + 0.005502596031874418, + 0.007263863459229469, + 0.0072789606638252735, + 0.004193950444459915, + 0.004271949175745249, + 0.006528065074235201, + 0.006786012090742588, + 0.03734005615115166, + 0.09376221895217896, + 0.008454740047454834, + 0.011172948405146599, + 0.04640797898173332, + 0.01474729273468256, + 0.03567603975534439, + 0.04998040571808815, + 0.0194503553211689, + 0.016658199951052666, + 0.022630080580711365, + 0.02596946433186531, + 0.02003154158592224, + 0.02536635845899582, + 0.06077801063656807, + 0.03468457981944084, + 0.008094590157270432, + 0.01616811938583851, + 0.008926354348659515, + 0.01266767643392086, + 0.007121675182133913, + 0.007103400770574808, + 0.014870304614305496, + 0.008971459232270718, + 0.0031171292066574097, + 0.0079853730276227, + 0.003610198851674795, + 0.005098751746118069, + 0.00491798622533679, + 0.005895364563912153, + 0.0041502234525978565, + 0.004834420513361692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0028173536993563175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dynamic medium tracking shot that follows the skier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skier wearing a royal blue ski suit, black ski pants, and a white helmet with rainbow-reflecting visor lenses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skier speeding down a vast, untouched slope with deep snow shimmering in bright sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skier shift his weight from side to side while carving deep turns that create a curtain of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does mist briefly fill the air creating a translucent halo and grow thicker to nearly engulf the skier's lower body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606132835149765 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606132835149765, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606132835149765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04969211295247078, + 0.02293437160551548, + 0.11081994324922562, + 0.025015003979206085, + 0.05671008303761482, + 0.031904254108667374, + 0.2687593400478363, + 0.10835684835910797, + 0.2557915449142456, + 0.07001645117998123 + ], + "credit_map": [ + 0.0075940596871078014, + 0.00619129091501236, + 0.0053247492760419846, + 0.004865230061113834, + 0.00839037261903286, + 0.005984222982078791, + 0.00459224171936512, + 0.006749944761395454, + 0.0033014537766575813, + 0.0034028703812509775, + 0.0021731348242610693, + 0.0024000273551791906, + 0.004179062321782112, + 0.0018970337696373463, + 0.0032506841234862804, + 0.0023301062174141407, + 0.01746292971074581, + 0.010061673820018768, + 0.014270040206611156, + 0.007290251087397337, + 0.014444855973124504, + 0.00959089770913124, + 0.01619110070168972, + 0.021508201956748962, + 0.0025887666270136833, + 0.0036043510772287846, + 0.003165248781442642, + 0.004156751558184624, + 0.002126381965354085, + 0.002572189085185528, + 0.003519757417961955, + 0.0032815570011734962, + 0.005095051601529121, + 0.008374148048460484, + 0.006558814086019993, + 0.006938794162124395, + 0.006121925078332424, + 0.004205875098705292, + 0.007185778114944696, + 0.012229696847498417, + 0.0028194310143589973, + 0.003079218789935112, + 0.0061674113385379314, + 0.005899909418076277, + 0.0026635481044650078, + 0.003934942651540041, + 0.0036648015957325697, + 0.003674994222819805, + 0.03317426145076752, + 0.0840344950556755, + 0.007783289533108473, + 0.010645980015397072, + 0.04152284190058708, + 0.0132416607812047, + 0.032591238617897034, + 0.04576556384563446, + 0.013076146133244038, + 0.013594944030046463, + 0.009787065908312798, + 0.012300966307520866, + 0.012824793346226215, + 0.015499777160584927, + 0.019625112414360046, + 0.011648040264844894, + 0.022475875914096832, + 0.04993433505296707, + 0.02538122609257698, + 0.04180222749710083, + 0.018016532063484192, + 0.020978737622499466, + 0.047367338091135025, + 0.029835283756256104, + 0.010162418708205223, + 0.009364165365695953, + 0.008258913643658161, + 0.011084068566560745, + 0.00759890815243125, + 0.008317762985825539, + 0.007222083862870932, + 0.008008127100765705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.01606132835149765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_tensor([2597, 2597])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_tensor([2597, 2597])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b88ffdd1e13ee9170f369003cfa73ecfbae99172 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_tensor([2597, 2597])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, extreme close-up captures the rocket engine's massive bell-shaped nozzle, its black, soot-stained metal surface covered in a complex network of pipes and thermal insulation. For a brief moment, the scene is utterly silent. Then, a dazzling flash of pure white instantly erupts from the engine's throat, dramatically expanding into a supersonic plume of incandescent plasma. The core of the exhaust burns with a bluish-white electrical intensity, clearly revealing the repetitive diamond-shaped shockwaves known as Mach diamonds. This core is surrounded by a roiling, surging torrent of flame, its color shifting from golden yellow to a deep, fiery orange, tumbling downward and outward at a ferocious pace. The flames engulf the vast, charred concrete launch pad below, while the flood system simultaneously unleashes a massive spray, producing a vast, billowing cloud of white steam that violently clashes and intermingles with the fire and smoke. The entire frame trembles intensely with the sheer acoustic force of the ignition, creating a visceral, chaotic, and awe-inspiring atmosphere of unleashing immense raw power.", + "seed": 18180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_2597_seed_18180_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.044969369657337666, + "vlm_logprob": -0.044969369657337666, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.044969369657337666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05279931053519249 + ], + "opposite_logprob_tokens": [ + -3.0527992248535156 + ], + "logprob_reward": -0.05279931053519249, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05279931053519249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10894151031970978, + 0.08145809918642044, + 0.10023342818021774, + 0.08389735966920853, + 0.27605634927749634, + 0.05842160806059837, + 0.11970707774162292, + 0.04895005002617836, + 0.07697626203298569, + 0.04535820707678795 + ], + "credit_map": [ + 0.009659778326749802, + 0.014692725613713264, + 0.009791693650186062, + 0.013009345158934593, + 0.022471224889159203, + 0.017351793125271797, + 0.014570320025086403, + 0.007394632324576378, + 0.008728326298296452, + 0.009160102345049381, + 0.013096625916659832, + 0.009340502321720123, + 0.013130709528923035, + 0.011649718508124352, + 0.00769157987087965, + 0.008660536259412766, + 0.013298815116286278, + 0.01698797568678856, + 0.01471822988241911, + 0.012586042284965515, + 0.008405924774706364, + 0.013997334986925125, + 0.008157515898346901, + 0.012081587687134743, + 0.009984183125197887, + 0.009478840976953506, + 0.007044102996587753, + 0.014192488044500351, + 0.009344336576759815, + 0.010636038146913052, + 0.008562298491597176, + 0.014655070379376411, + 0.03323207050561905, + 0.03173463046550751, + 0.03153017535805702, + 0.021990276873111725, + 0.053829126060009, + 0.03766733035445213, + 0.035473328083753586, + 0.030599400401115417, + 0.005943110212683678, + 0.008544142358005047, + 0.009292121045291424, + 0.0059220800176262856, + 0.008083488792181015, + 0.006603106390684843, + 0.006635953672230244, + 0.007397605571895838, + 0.016008710488677025, + 0.016449786722660065, + 0.025105664506554604, + 0.010479400865733624, + 0.015724988654255867, + 0.013307676650583744, + 0.009516620077192783, + 0.01311422884464264, + 0.005595983937382698, + 0.004941581282764673, + 0.00632875831797719, + 0.004641020204871893, + 0.010609351098537445, + 0.008091806434094906, + 0.003458507126197219, + 0.005283038597553968, + 0.009596403688192368, + 0.005734466016292572, + 0.0110193882137537, + 0.005452161654829979, + 0.009806654416024685, + 0.011259965598583221, + 0.005915267858654261, + 0.018191959708929062, + 0.005905692931264639, + 0.005101731978356838, + 0.005943736061453819, + 0.0028622541576623917, + 0.0066453986801207066, + 0.006702837999910116, + 0.004793299362063408, + 0.007403251715004444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.05279931053519249 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06212323531508446 + ], + "opposite_logprob_tokens": [ + -2.8121232986450195 + ], + "logprob_reward": -0.06212323531508446, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06212323531508446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496385782957077, + 0.06245841458439827, + 0.11170335859060287, + 0.1081453189253807, + 0.2718352973461151, + 0.048134222626686096, + 0.07365089654922485, + 0.03636620566248894, + 0.08101346343755722, + 0.057054221630096436 + ], + "credit_map": [ + 0.013640894554555416, + 0.020130213350057602, + 0.015009110793471336, + 0.016258148476481438, + 0.023812154307961464, + 0.02253447286784649, + 0.02500925026834011, + 0.01324433647096157, + 0.006394552998244762, + 0.005869131535291672, + 0.010157332755625248, + 0.004798773676156998, + 0.011428603902459145, + 0.009539715945720673, + 0.006926666013896465, + 0.007343635894358158, + 0.015104454942047596, + 0.017389362677931786, + 0.016906602308154106, + 0.015073665417730808, + 0.009096791967749596, + 0.015233506448566914, + 0.009063977748155594, + 0.013834984041750431, + 0.014009286649525166, + 0.011513658799231052, + 0.009256982244551182, + 0.01592136360704899, + 0.013272850774228573, + 0.016359584406018257, + 0.009456449188292027, + 0.018355147913098335, + 0.031954146921634674, + 0.032665420323610306, + 0.03243376687169075, + 0.022723915055394173, + 0.0474475733935833, + 0.0353589802980423, + 0.03717602789402008, + 0.03207546845078468, + 0.005751464050263166, + 0.006727810483425856, + 0.00683253351598978, + 0.0050120484083890915, + 0.006909804884344339, + 0.005682703573256731, + 0.004843341186642647, + 0.006374517921358347, + 0.009795294143259525, + 0.010303322225809097, + 0.014812790788710117, + 0.006923469714820385, + 0.009753013961017132, + 0.008363955654203892, + 0.005875793751329184, + 0.007823264226317406, + 0.004239748697727919, + 0.003750399686396122, + 0.0068346611224114895, + 0.0037095416337251663, + 0.006135334726423025, + 0.004093775060027838, + 0.0032276930287480354, + 0.004375052638351917, + 0.012015164829790592, + 0.006679522804915905, + 0.014116013422608376, + 0.006755135953426361, + 0.009133716113865376, + 0.01097843423485756, + 0.007061463315039873, + 0.01427401416003704, + 0.007634331006556749, + 0.006879141554236412, + 0.007600974291563034, + 0.004625924862921238, + 0.00811351090669632, + 0.00802229717373848, + 0.0053901211358606815, + 0.008787923492491245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.06212323531508446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256160531193018 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.01256160531193018, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256160531193018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10255009680986404, + 0.060913048684597015, + 0.06921525299549103, + 0.20865152776241302, + 0.13888582587242126, + 0.052100736647844315, + 0.1491629034280777, + 0.06870897859334946, + 0.08992046862840652, + 0.059891227632761 + ], + "credit_map": [ + 0.010399683378636837, + 0.00986751914024353, + 0.011019247584044933, + 0.011425524950027466, + 0.020476320758461952, + 0.017813121899962425, + 0.014789328910410404, + 0.006759354379028082, + 0.00563555583357811, + 0.008648020215332508, + 0.0066571468487381935, + 0.004237258341163397, + 0.013130647130310535, + 0.010922200977802277, + 0.006483767181634903, + 0.005198458209633827, + 0.008990674279630184, + 0.014421516098082066, + 0.008901603519916534, + 0.008882252499461174, + 0.007012158632278442, + 0.009216692298650742, + 0.005177427548915148, + 0.006612925324589014, + 0.027340447530150414, + 0.023933028802275658, + 0.017261022701859474, + 0.02763451635837555, + 0.033856336027383804, + 0.029943671077489853, + 0.015905611217021942, + 0.03277689591050148, + 0.016154255717992783, + 0.018280543386936188, + 0.016066839918494225, + 0.011698235757648945, + 0.02381935715675354, + 0.01760181598365307, + 0.019798383116722107, + 0.015466389246284962, + 0.006295232567936182, + 0.00582248717546463, + 0.007260654587298632, + 0.005217617843300104, + 0.007451757322996855, + 0.006202585529536009, + 0.005097295623272657, + 0.008753107860684395, + 0.020692087709903717, + 0.020335910841822624, + 0.03124416619539261, + 0.013012723065912724, + 0.019572267308831215, + 0.01652505062520504, + 0.011606749147176743, + 0.016173947602510452, + 0.007587192580103874, + 0.00682968320325017, + 0.00823725201189518, + 0.007372952997684479, + 0.014228934422135353, + 0.011809222400188446, + 0.00474430900067091, + 0.00789943803101778, + 0.012617542408406734, + 0.007934421300888062, + 0.012900182045996189, + 0.00608103908598423, + 0.011343569494783878, + 0.016507109627127647, + 0.0061264801770448685, + 0.01641012541949749, + 0.00806110817939043, + 0.005708108190447092, + 0.00828672107309103, + 0.004065047483891249, + 0.009823703207075596, + 0.00977457594126463, + 0.005319139454513788, + 0.008852828294038773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.01256160531193018 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868816211819649 + ], + "opposite_logprob_tokens": [ + -3.0486881732940674 + ], + "logprob_reward": -0.04868816211819649, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868816211819649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09266006946563721, + 0.04752494394779205, + 0.06632460653781891, + 0.15085704624652863, + 0.07619736343622208, + 0.09365096688270569, + 0.23097319900989532, + 0.09996813535690308, + 0.0878896713256836, + 0.05395400524139404 + ], + "credit_map": [ + 0.008648966439068317, + 0.011572886258363724, + 0.009583456441760063, + 0.008647331967949867, + 0.014764618128538132, + 0.013888977468013763, + 0.01755060814321041, + 0.008003223687410355, + 0.004163993056863546, + 0.006085276138037443, + 0.004179979674518108, + 0.0030584970954805613, + 0.0126044237986207, + 0.009457984007894993, + 0.004259485751390457, + 0.0037153041921555996, + 0.00883380975574255, + 0.012764669954776764, + 0.00986617710441351, + 0.00787417683750391, + 0.005591599736362696, + 0.009328282438218594, + 0.005104768089950085, + 0.006961122155189514, + 0.01904633268713951, + 0.017516039311885834, + 0.011949839070439339, + 0.021452762186527252, + 0.025871021673083305, + 0.021225856617093086, + 0.011686046607792377, + 0.022109145298600197, + 0.00852966122329235, + 0.009649502113461494, + 0.009212620556354523, + 0.006349057424813509, + 0.013288467191159725, + 0.009893611073493958, + 0.011093449778854847, + 0.008181002922356129, + 0.013090141117572784, + 0.00981212966144085, + 0.016076266765594482, + 0.008995365351438522, + 0.011281461454927921, + 0.011436467058956623, + 0.007909887470304966, + 0.015049239620566368, + 0.032091785222291946, + 0.03155846148729324, + 0.0491989366710186, + 0.020214121788740158, + 0.030140772461891174, + 0.02546539343893528, + 0.017685579136013985, + 0.024618161842226982, + 0.011675553396344185, + 0.009762182831764221, + 0.011650399304926395, + 0.009511302225291729, + 0.021583959460258484, + 0.01786731369793415, + 0.006851120386272669, + 0.011066298931837082, + 0.013304103165864944, + 0.007577257696539164, + 0.015813766047358513, + 0.006839490495622158, + 0.010514461435377598, + 0.012634442187845707, + 0.006665829569101334, + 0.014540324918925762, + 0.006990219932049513, + 0.006045002490282059, + 0.007264455780386925, + 0.0035756248980760574, + 0.008120330981910229, + 0.00911739468574524, + 0.0052070277743041515, + 0.007633948232978582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.04868816211819649 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674535006284714 + ], + "opposite_logprob_tokens": [ + -3.0486745834350586 + ], + "logprob_reward": -0.048674535006284714, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674535006284714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12929357588291168, + 0.043866392225027084, + 0.094704270362854, + 0.19584965705871582, + 0.11804255098104477, + 0.05764113366603851, + 0.16738756000995636, + 0.07309597730636597, + 0.06953262537717819, + 0.050586238503456116 + ], + "credit_map": [ + 0.013534153811633587, + 0.012564262375235558, + 0.013439424335956573, + 0.009997311048209667, + 0.02448900043964386, + 0.02516968734562397, + 0.018414512276649475, + 0.011685221455991268, + 0.0036004497669637203, + 0.004713837988674641, + 0.004397811368107796, + 0.003845940111204982, + 0.009608019143342972, + 0.007269574794918299, + 0.005661543924361467, + 0.0047692167572677135, + 0.014091617427766323, + 0.014690493233501911, + 0.014380600303411484, + 0.010182630270719528, + 0.00777624873444438, + 0.015368196181952953, + 0.00757238082587719, + 0.01064209919422865, + 0.025707850232720375, + 0.021868914365768433, + 0.016862064599990845, + 0.028353264555335045, + 0.029187701642513275, + 0.026553701609373093, + 0.015651514753699303, + 0.031664665788412094, + 0.012909471988677979, + 0.016921211034059525, + 0.010555504821240902, + 0.00850977748632431, + 0.02610686421394348, + 0.016484595835208893, + 0.015659349039196968, + 0.010895774699747562, + 0.0074022794142365456, + 0.005801422521471977, + 0.01015805546194315, + 0.0057174526154994965, + 0.006987401749938726, + 0.006903874222189188, + 0.00451140571385622, + 0.010159237310290337, + 0.02351008728146553, + 0.02266971953213215, + 0.03535759449005127, + 0.014446492306888103, + 0.022129882127046585, + 0.01850517839193344, + 0.013024359941482544, + 0.01774424873292446, + 0.008414438925683498, + 0.0072813318111002445, + 0.008985606022179127, + 0.007800202816724777, + 0.014035739004611969, + 0.012835905887186527, + 0.005049494560807943, + 0.008693261072039604, + 0.010593709535896778, + 0.005733957514166832, + 0.013816230930387974, + 0.004926767665892839, + 0.008168255910277367, + 0.009153096936643124, + 0.006030736956745386, + 0.011109864339232445, + 0.006458960473537445, + 0.004895564168691635, + 0.0064131454564630985, + 0.002982992911711335, + 0.008689885959029198, + 0.008233994245529175, + 0.005542323924601078, + 0.007369371596723795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.048674535006284714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up captures the rocket engine's massive bell-shaped nozzle, its black, soot-stained metal surface covered in a complex network of pipes and thermal insulation. For a brief moment, the scene is utterly silent. Then, a dazzling flash of pure white instantly erupts from the engine's throat, dramatically expanding into a supersonic plume of incandescent plasma. The core of the exhaust burns with a bluish-white electrical intensity, clearly revealing the repetitive diamond-shaped shockwaves known as Mach diamonds. This core is surrounded by a roiling, surging torrent of flame, its color shifting from golden yellow to a deep, fiery orange, tumbling downward and outward at a ferocious pace. The flames engulf the vast, charred concrete launch pad below, while the flood system simultaneously unleashes a massive spray, producing a vast, billowing cloud of white steam that violently clashes and intermingles with the fire and smoke. The entire frame trembles intensely with the sheer acoustic force of the ignition, creating a visceral, chaotic, and awe-inspiring atmosphere of unleashing immense raw power.", + "seed": 18181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_2597_seed_18181_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037191997095942494, + "vlm_logprob": -0.037191997095942494, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.037191997095942494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07534404844045639 + ], + "opposite_logprob_tokens": [ + -2.9503440856933594 + ], + "logprob_reward": -0.07534404844045639, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07534404844045639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1198602169752121, + 0.0931505411863327, + 0.09612728655338287, + 0.10481304675340652, + 0.40606534481048584, + 0.03724716231226921, + 0.028812630102038383, + 0.0435347817838192, + 0.02264791540801525, + 0.047741085290908813 + ], + "credit_map": [ + 0.011202890425920486, + 0.008554797619581223, + 0.0160063486546278, + 0.026768459007143974, + 0.009317385032773018, + 0.01562531478703022, + 0.019344232976436615, + 0.013040775433182716, + 0.009400743059813976, + 0.00801230687648058, + 0.012195705436170101, + 0.014909675344824791, + 0.01310056634247303, + 0.012306307442486286, + 0.012042351067066193, + 0.011182887479662895, + 0.010208848863840103, + 0.0114965308457613, + 0.012056552805006504, + 0.012740684673190117, + 0.01638459786772728, + 0.011217660270631313, + 0.012517222203314304, + 0.009505189023911953, + 0.00928074587136507, + 0.010491151362657547, + 0.015433711931109428, + 0.025555720552802086, + 0.010314112529158592, + 0.009799228981137276, + 0.013779845088720322, + 0.010158524848520756, + 0.03445923700928688, + 0.05022704601287842, + 0.06436128169298172, + 0.060907986015081406, + 0.05056770145893097, + 0.03181101754307747, + 0.05734837427735329, + 0.0563826858997345, + 0.0036447288002818823, + 0.004696766845881939, + 0.0034121170174330473, + 0.005162655375897884, + 0.004369454458355904, + 0.003201985266059637, + 0.0031917474698275328, + 0.009567704983055592, + 0.002593844896182418, + 0.003253950271755457, + 0.0033157430589199066, + 0.006196548696607351, + 0.002746758982539177, + 0.003395207691937685, + 0.002821754664182663, + 0.0044888220727443695, + 0.005096048582345247, + 0.005167042370885611, + 0.007653448265045881, + 0.0061548384837806225, + 0.006503653712570667, + 0.00392690347507596, + 0.0044541326351463795, + 0.004578712861984968, + 0.0024527867790311575, + 0.0018368250457569957, + 0.00348023883998394, + 0.004150472581386566, + 0.002695563482120633, + 0.002741292119026184, + 0.002416405826807022, + 0.0028743320144712925, + 0.005831354763358831, + 0.0036113669630140066, + 0.007458277978003025, + 0.007927848026156425, + 0.003292514942586422, + 0.006530618295073509, + 0.007707131095230579, + 0.005381973460316658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07534404844045639 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01823541522026062 + ], + "opposite_logprob_tokens": [ + -4.018235206604004 + ], + "logprob_reward": -0.01823541522026062, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01823541522026062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19472934305667877, + 0.10415907949209213, + 0.1310451328754425, + 0.08718398958444595, + 0.060591161251068115, + 0.06274775415658951, + 0.07194092124700546, + 0.15481363236904144, + 0.07268837839365005, + 0.06010056659579277 + ], + "credit_map": [ + 0.014954022131860256, + 0.01825787127017975, + 0.020505744963884354, + 0.03276333585381508, + 0.018931947648525238, + 0.03127031400799751, + 0.024069298058748245, + 0.033976826816797256, + 0.010630257427692413, + 0.007991609163582325, + 0.013482478447258472, + 0.017560679465532303, + 0.008336174301803112, + 0.013767149299383163, + 0.014912327751517296, + 0.017478395253419876, + 0.015992090106010437, + 0.009847135283052921, + 0.012945094145834446, + 0.02533428557217121, + 0.013492594473063946, + 0.018944749608635902, + 0.014261391013860703, + 0.020227797329425812, + 0.006534371990710497, + 0.009504648856818676, + 0.012975217774510384, + 0.022442445158958435, + 0.008074573241174221, + 0.006812885869294405, + 0.012244007550179958, + 0.008595835417509079, + 0.005395099055022001, + 0.006483986973762512, + 0.00985202006995678, + 0.01354080531746149, + 0.00781784113496542, + 0.0038983055856078863, + 0.009428446181118488, + 0.0041746594943106174, + 0.010085020214319229, + 0.008027722127735615, + 0.007202593609690666, + 0.007429057266563177, + 0.008040742948651314, + 0.00454968586564064, + 0.006467098370194435, + 0.010945831425487995, + 0.006327190902084112, + 0.00861658900976181, + 0.010588100180029869, + 0.0127657949924469, + 0.00736651849001646, + 0.006094668060541153, + 0.00885765440762043, + 0.011324405670166016, + 0.016822371631860733, + 0.01690439134836197, + 0.027317846193909645, + 0.027384065091609955, + 0.0187425184994936, + 0.014998357743024826, + 0.01906842179596424, + 0.013575665652751923, + 0.007202928885817528, + 0.00928486604243517, + 0.0107862064614892, + 0.00975631270557642, + 0.01065722107887268, + 0.009467094205319881, + 0.007529844529926777, + 0.008003908209502697, + 0.006303656380623579, + 0.0057685598731040955, + 0.011695041321218014, + 0.008015706203877926, + 0.006105369422584772, + 0.0066123041324317455, + 0.008580098859965801, + 0.007019828539341688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.01823541522026062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796636179089546 + ], + "opposite_logprob_tokens": [ + -4.634796619415283 + ], + "logprob_reward": -0.009796636179089546, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796636179089546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16871322691440582, + 0.08652053773403168, + 0.3388608992099762, + 0.045736830681562424, + 0.09334594011306763, + 0.045881081372499466, + 0.0816996693611145, + 0.08420085161924362, + 0.03275013342499733, + 0.022290866822004318 + ], + "credit_map": [ + 0.009826232679188251, + 0.012513195164501667, + 0.02272673323750496, + 0.038702938705682755, + 0.013978078030049801, + 0.022188274189829826, + 0.03274941444396973, + 0.01602834090590477, + 0.008649302646517754, + 0.010070176795125008, + 0.009642088785767555, + 0.014826337806880474, + 0.008204451762139797, + 0.011284041218459606, + 0.010811313055455685, + 0.013032819144427776, + 0.03550804406404495, + 0.03813706338405609, + 0.03710784390568733, + 0.03169943019747734, + 0.0661054328083992, + 0.04369188845157623, + 0.04224061965942383, + 0.044370539486408234, + 0.0057873837649822235, + 0.003312992164865136, + 0.00546638946980238, + 0.013302914798259735, + 0.004292904399335384, + 0.0046495553106069565, + 0.005723428446799517, + 0.0032012579031288624, + 0.008983603678643703, + 0.011610732413828373, + 0.01582886464893818, + 0.013942627236247063, + 0.011930984444916248, + 0.008419889956712723, + 0.01160404458642006, + 0.011025184765458107, + 0.00392125966027379, + 0.005239189136773348, + 0.005182419903576374, + 0.006367863155901432, + 0.005687755066901445, + 0.0036588311195373535, + 0.004293297883123159, + 0.011530461721122265, + 0.007731746416538954, + 0.009994066320359707, + 0.012314492836594582, + 0.020115341991186142, + 0.00800047442317009, + 0.008641278371214867, + 0.009118827059864998, + 0.005783438682556152, + 0.009676545858383179, + 0.008723929524421692, + 0.014820529147982597, + 0.015691230073571205, + 0.00945178885012865, + 0.007768702693283558, + 0.010118757374584675, + 0.007949365302920341, + 0.003393324092030525, + 0.0032074979972094297, + 0.005991010926663876, + 0.004988706670701504, + 0.0030641420744359493, + 0.004537275992333889, + 0.004168923944234848, + 0.0033992489334195852, + 0.0023440858349204063, + 0.0024198382161557674, + 0.0038418753538280725, + 0.0036385057028383017, + 0.002663656137883663, + 0.0022843254264444113, + 0.0026236469857394695, + 0.0024749308358877897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009796636179089546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254606619477272 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.01254606619477272, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254606619477272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07989877462387085, + 0.07762127369642258, + 0.08281790465116501, + 0.07967960834503174, + 0.10530668497085571, + 0.16292540729045868, + 0.14987564086914062, + 0.13851015269756317, + 0.05476786941289902, + 0.06859666854143143 + ], + "credit_map": [ + 0.006655556615442038, + 0.007759031839668751, + 0.007739824242889881, + 0.013854684308171272, + 0.00609924178570509, + 0.012074926868081093, + 0.010339805856347084, + 0.015375697053968906, + 0.0065703196451067924, + 0.006698224693536758, + 0.008659844286739826, + 0.014638099819421768, + 0.007641430478543043, + 0.010623603127896786, + 0.01314691361039877, + 0.009642831049859524, + 0.007154503371566534, + 0.005885961465537548, + 0.009608960710465908, + 0.014572268351912498, + 0.00785901304334402, + 0.01217774860560894, + 0.010211666114628315, + 0.015347783453762531, + 0.007404633332043886, + 0.01592710241675377, + 0.009867376647889614, + 0.011831752955913544, + 0.0113050676882267, + 0.007331686560064554, + 0.007823496125638485, + 0.008188486099243164, + 0.007532583549618721, + 0.011862524785101414, + 0.013398589566349983, + 0.01861695386469364, + 0.01285482756793499, + 0.007165094837546349, + 0.0167643241584301, + 0.017111778259277344, + 0.029344119131565094, + 0.024522902444005013, + 0.021301982924342155, + 0.021837497130036354, + 0.025339806452393532, + 0.012606671079993248, + 0.014575468376278877, + 0.013396946713328362, + 0.01278454065322876, + 0.016186241060495377, + 0.021045658737421036, + 0.017048904672265053, + 0.01606622524559498, + 0.013853734359145164, + 0.018260715529322624, + 0.03462962806224823, + 0.01735505275428295, + 0.01760965585708618, + 0.0237326230853796, + 0.012495935894548893, + 0.025403395295143127, + 0.01252774428576231, + 0.015166970901191235, + 0.014218770898878574, + 0.004577066749334335, + 0.007125042844563723, + 0.006744859274476767, + 0.007109356112778187, + 0.01062751468271017, + 0.006164551712572575, + 0.006041923072189093, + 0.006377552170306444, + 0.009091650135815144, + 0.0071222721599042416, + 0.008854116313159466, + 0.009655390866100788, + 0.007400954142212868, + 0.007477339822798967, + 0.007768587209284306, + 0.011226356960833073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.01254606619477272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003781944513321 + ], + "opposite_logprob_tokens": [ + -2.695037841796875 + ], + "logprob_reward": -0.07003781944513321, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003781944513321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08262999355792999, + 0.03242947533726692, + 0.1577116698026657, + 0.05370428040623665, + 0.30404114723205566, + 0.06870698928833008, + 0.07118601351976395, + 0.0955626592040062, + 0.07949556410312653, + 0.0545322485268116 + ], + "credit_map": [ + 0.007264443673193455, + 0.008078829385340214, + 0.014485644176602364, + 0.014920255169272423, + 0.00644699577242136, + 0.010580779984593391, + 0.012502211146056652, + 0.008350824937224388, + 0.0030810702592134476, + 0.0037793016526848078, + 0.0032112128101289272, + 0.005194478202611208, + 0.0034179992508143187, + 0.004502426367253065, + 0.004128817934542894, + 0.0051141646690666676, + 0.01620805822312832, + 0.017057251185178757, + 0.017915980890393257, + 0.016786236315965652, + 0.028540290892124176, + 0.020141521468758583, + 0.02049226127564907, + 0.020570058375597, + 0.0045220558531582355, + 0.005429625511169434, + 0.0069154477678239346, + 0.011638420633971691, + 0.005851352587342262, + 0.0049622394144535065, + 0.009423687122762203, + 0.0049614496529102325, + 0.024279164150357246, + 0.03729648515582085, + 0.04470985010266304, + 0.04945196211338043, + 0.03868761286139488, + 0.01895182952284813, + 0.04751671478152275, + 0.0431474968791008, + 0.007780871819704771, + 0.011786806397140026, + 0.008112238720059395, + 0.012066042982041836, + 0.01019777450710535, + 0.004323073197156191, + 0.007731103803962469, + 0.00670907087624073, + 0.007057479582726955, + 0.0072709135711193085, + 0.009828765876591206, + 0.017405210062861443, + 0.00665636733174324, + 0.006861010566353798, + 0.008813935332000256, + 0.007292332127690315, + 0.011882465332746506, + 0.012402468360960484, + 0.01632031425833702, + 0.00935677345842123, + 0.017832741141319275, + 0.008112041279673576, + 0.010267958976328373, + 0.009387882426381111, + 0.008990597911179066, + 0.00956908892840147, + 0.01241055317223072, + 0.008439715020358562, + 0.014156046323478222, + 0.009445554576814175, + 0.0074067628011107445, + 0.009077237918972969, + 0.006128935609012842, + 0.006558967288583517, + 0.007272449787706137, + 0.006783415097743273, + 0.007356098387390375, + 0.0070855035446584225, + 0.0055004446767270565, + 0.00784643180668354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.07003781944513321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up captures the rocket engine's massive bell-shaped nozzle, its black, soot-stained metal surface covered in a complex network of pipes and thermal insulation. For a brief moment, the scene is utterly silent. Then, a dazzling flash of pure white instantly erupts from the engine's throat, dramatically expanding into a supersonic plume of incandescent plasma. The core of the exhaust burns with a bluish-white electrical intensity, clearly revealing the repetitive diamond-shaped shockwaves known as Mach diamonds. This core is surrounded by a roiling, surging torrent of flame, its color shifting from golden yellow to a deep, fiery orange, tumbling downward and outward at a ferocious pace. The flames engulf the vast, charred concrete launch pad below, while the flood system simultaneously unleashes a massive spray, producing a vast, billowing cloud of white steam that violently clashes and intermingles with the fire and smoke. The entire frame trembles intensely with the sheer acoustic force of the ignition, creating a visceral, chaotic, and awe-inspiring atmosphere of unleashing immense raw power.", + "seed": 18182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_2597_seed_18182_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016467363014817236, + "vlm_logprob": -0.016467363014817236, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.016467363014817236, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03659209981560707 + ], + "opposite_logprob_tokens": [ + -3.9115920066833496 + ], + "logprob_reward": -0.03659209981560707, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03659209981560707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20708338916301727, + 0.05396777018904686, + 0.16888728737831116, + 0.07875791937112808, + 0.062352605164051056, + 0.05990730598568916, + 0.12384161353111267, + 0.07051993161439896, + 0.08588586747646332, + 0.08879625052213669 + ], + "credit_map": [ + 0.03476650267839432, + 0.022072171792387962, + 0.037897683680057526, + 0.028575416654348373, + 0.01781962811946869, + 0.017849553376436234, + 0.033918216824531555, + 0.01418421883136034, + 0.006879251915961504, + 0.0050974516198039055, + 0.006222814321517944, + 0.008526970632374287, + 0.009954461827874184, + 0.00519130052998662, + 0.008090325631201267, + 0.004005194641649723, + 0.015320192091166973, + 0.021687442436814308, + 0.01772489584982395, + 0.02941993996500969, + 0.021793439984321594, + 0.01475557778030634, + 0.021196380257606506, + 0.02698943205177784, + 0.006146354600787163, + 0.013672561384737492, + 0.010609600692987442, + 0.01061230432242155, + 0.01380267646163702, + 0.007008764892816544, + 0.009555620141327381, + 0.007350048050284386, + 0.00533694913610816, + 0.006391642149537802, + 0.006978763733059168, + 0.01385638490319252, + 0.009168969467282295, + 0.004440756514668465, + 0.012081483379006386, + 0.004097659606486559, + 0.0059306686744093895, + 0.008629318326711655, + 0.004474416375160217, + 0.010640718974173069, + 0.009056170471012592, + 0.008114959113299847, + 0.008767680265009403, + 0.004293377045542002, + 0.005918329581618309, + 0.034542907029390335, + 0.0061678458005189896, + 0.014347776770591736, + 0.034682776778936386, + 0.007739085704088211, + 0.013895616866648197, + 0.00654727965593338, + 0.0073624528013169765, + 0.009847670793533325, + 0.007032713387161493, + 0.01074196957051754, + 0.007974687032401562, + 0.006458482705056667, + 0.015248874202370644, + 0.005853086709976196, + 0.01548042893409729, + 0.011154928244650364, + 0.0092680174857378, + 0.009017540141940117, + 0.007386540528386831, + 0.009208174422383308, + 0.0147307263687253, + 0.00963952112942934, + 0.010986795648932457, + 0.010089447721838951, + 0.010229429230093956, + 0.008608195930719376, + 0.010553011670708656, + 0.014053283259272575, + 0.009329748339951038, + 0.014946350827813148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.03659209981560707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014235729351639748 + ], + "opposite_logprob_tokens": [ + -4.264235496520996 + ], + "logprob_reward": -0.014235729351639748, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014235729351639748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3175867199897766, + 0.16349337995052338, + 0.05386270210146904, + 0.03676433861255646, + 0.02770458161830902, + 0.023410199210047722, + 0.2801697552204132, + 0.029110057279467583, + 0.03476250544190407, + 0.03313576802611351 + ], + "credit_map": [ + 0.04545266926288605, + 0.05010427162051201, + 0.06666244566440582, + 0.027630791068077087, + 0.03175777569413185, + 0.037663135677576065, + 0.03409614413976669, + 0.02421947568655014, + 0.020515263080596924, + 0.017023706808686256, + 0.013574795797467232, + 0.028140215203166008, + 0.031320709735155106, + 0.014964225701987743, + 0.025950271636247635, + 0.012004196643829346, + 0.004735281690955162, + 0.006833569146692753, + 0.005982280243188143, + 0.00771842198446393, + 0.006340058520436287, + 0.0048052179627120495, + 0.006971569266170263, + 0.010476303286850452, + 0.003809751244261861, + 0.0047233207151293755, + 0.005060351919382811, + 0.003958319779485464, + 0.006452322006225586, + 0.004165695980191231, + 0.004897647071629763, + 0.003696930129081011, + 0.0028783618472516537, + 0.004015043377876282, + 0.0025084565859287977, + 0.0037382482551038265, + 0.006546103395521641, + 0.002579344203695655, + 0.003504737513139844, + 0.0019342839950695634, + 0.0023544984869658947, + 0.004543250426650047, + 0.0017927750013768673, + 0.003172843484207988, + 0.004163064528256655, + 0.0033486345782876015, + 0.002281143795698881, + 0.0017539900727570057, + 0.01907307468354702, + 0.11602788418531418, + 0.00753282755613327, + 0.00845129694789648, + 0.10151989012956619, + 0.01124353613704443, + 0.008783501572906971, + 0.0075377668254077435, + 0.002786925295367837, + 0.0063753509894013405, + 0.0025328141637146473, + 0.00393768772482872, + 0.0038433466106653214, + 0.002844760660082102, + 0.004355523269623518, + 0.0024336499627679586, + 0.005466770846396685, + 0.003931089304387569, + 0.004220292903482914, + 0.004004421643912792, + 0.0025959075428545475, + 0.003877429524436593, + 0.006092838943004608, + 0.004573751240968704, + 0.0033213836140930653, + 0.004745314363390207, + 0.00393018638715148, + 0.003453927580267191, + 0.00350542482919991, + 0.004648089874535799, + 0.004295420367270708, + 0.005236019380390644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.014235729351639748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659668266773224 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.008659668266773224, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659668266773224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13411128520965576, + 0.22193673253059387, + 0.10970112681388855, + 0.054472651332616806, + 0.0306683462113142, + 0.04886721447110176, + 0.181459441781044, + 0.05260799452662468, + 0.07812704145908356, + 0.08804816752672195 + ], + "credit_map": [ + 0.018645986914634705, + 0.017039591446518898, + 0.02650066278874874, + 0.013478623703122139, + 0.013863918371498585, + 0.016757091507315636, + 0.015186693519353867, + 0.01263872068375349, + 0.028404468670487404, + 0.02232581563293934, + 0.01861714757978916, + 0.0420970544219017, + 0.04058786854147911, + 0.020245753228664398, + 0.03401167690753937, + 0.015646953135728836, + 0.014875398948788643, + 0.019299939274787903, + 0.010549246333539486, + 0.011436221189796925, + 0.022883204743266106, + 0.015889564529061317, + 0.008531122468411922, + 0.006236432585865259, + 0.004185455851256847, + 0.007091919891536236, + 0.00874803401529789, + 0.006852671038359404, + 0.009828035719692707, + 0.005109121557325125, + 0.006999935023486614, + 0.005657477770000696, + 0.0028033133130520582, + 0.004313741344958544, + 0.0039040865376591682, + 0.003529297187924385, + 0.005862542428076267, + 0.0033276895992457867, + 0.004563158378005028, + 0.0023645171895623207, + 0.004919322673231363, + 0.010029122233390808, + 0.005963271018117666, + 0.006015087943524122, + 0.006828715093433857, + 0.005052100867033005, + 0.005607088562101126, + 0.004452504217624664, + 0.012721492908895016, + 0.06903256475925446, + 0.009191462770104408, + 0.00707586295902729, + 0.05887967720627785, + 0.01107560284435749, + 0.007010492961853743, + 0.006472285371273756, + 0.005589745938777924, + 0.00804264098405838, + 0.005113484803587198, + 0.006898841820657253, + 0.006326623260974884, + 0.005315277725458145, + 0.00968911498785019, + 0.005632263608276844, + 0.009348457679152489, + 0.015054869465529919, + 0.008278912864625454, + 0.008999248966574669, + 0.00885474681854248, + 0.008576774038374424, + 0.00945571344345808, + 0.009558316320180893, + 0.010376502759754658, + 0.015826422721147537, + 0.008974401280283928, + 0.00928650051355362, + 0.01237497664988041, + 0.00948061142116785, + 0.012244312092661858, + 0.00948444101959467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.008659668266773224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645507782697678 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645507782697678, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645507782697678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12300457805395126, + 0.10430753976106644, + 0.06529723107814789, + 0.07838820666074753, + 0.03857195004820824, + 0.03968019410967827, + 0.23003898561000824, + 0.1143718957901001, + 0.08036640286445618, + 0.12597298622131348 + ], + "credit_map": [ + 0.017214909195899963, + 0.025555267930030823, + 0.02590320259332657, + 0.010820111259818077, + 0.018794672563672066, + 0.006809212267398834, + 0.011802146211266518, + 0.00610506022349, + 0.012732432223856449, + 0.009979740716516972, + 0.008850693702697754, + 0.017502907663583755, + 0.01939178816974163, + 0.00782232079654932, + 0.017002888023853302, + 0.011024761945009232, + 0.007690125610679388, + 0.010366948321461678, + 0.005733358673751354, + 0.009793204255402088, + 0.012981425039470196, + 0.0062810503877699375, + 0.007596052251756191, + 0.004855066072195768, + 0.007580825127661228, + 0.0138389365747571, + 0.00889536365866661, + 0.005817661993205547, + 0.014436415396630764, + 0.006824936252087355, + 0.008089441806077957, + 0.012904622592031956, + 0.003658460918813944, + 0.00533327367156744, + 0.004038524813950062, + 0.004447516985237598, + 0.009207991883158684, + 0.0038755848072469234, + 0.005535425152629614, + 0.002475170884281397, + 0.00524530466645956, + 0.0068457284942269325, + 0.0033172338735312223, + 0.004911560099571943, + 0.0051644230261445045, + 0.0060198549181222916, + 0.004324980545789003, + 0.003851107554510236, + 0.018705453723669052, + 0.09652060270309448, + 0.00490096490830183, + 0.007391700986772776, + 0.08171183615922928, + 0.009346071630716324, + 0.005753610748797655, + 0.005708772223442793, + 0.01174384355545044, + 0.02170610800385475, + 0.010434536263346672, + 0.015112302266061306, + 0.013908881694078445, + 0.012704207561910152, + 0.018877876922488213, + 0.009884148836135864, + 0.01435413770377636, + 0.00965913850814104, + 0.009296175092458725, + 0.010423841886222363, + 0.006373108830302954, + 0.0070834956131875515, + 0.013878917321562767, + 0.00929759070277214, + 0.015346101485192776, + 0.01121508702635765, + 0.013656508177518845, + 0.012174435891211033, + 0.012852020561695099, + 0.0214847382158041, + 0.014195646159350872, + 0.02504844404757023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.004645507782697678 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820380985736847 + ], + "opposite_logprob_tokens": [ + -4.0182037353515625 + ], + "logprob_reward": -0.01820380985736847, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820380985736847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23231291770935059, + 0.06867963820695877, + 0.05349447578191757, + 0.07526233792304993, + 0.03985738754272461, + 0.04161655157804489, + 0.3418305814266205, + 0.048082560300827026, + 0.05479072779417038, + 0.044072750955820084 + ], + "credit_map": [ + 0.030161665752530098, + 0.05430829897522926, + 0.033715758472681046, + 0.02227858267724514, + 0.03703664243221283, + 0.013470273464918137, + 0.028619403019547462, + 0.012722326442599297, + 0.008954844437539577, + 0.006531068589538336, + 0.005363800097256899, + 0.011577920988202095, + 0.014758302830159664, + 0.0059256176464259624, + 0.0105394646525383, + 0.005028625950217247, + 0.007983316667377949, + 0.005786776077002287, + 0.005810458678752184, + 0.010517360642552376, + 0.0061075203120708466, + 0.0061104390770196915, + 0.007975837215781212, + 0.0032027752604335546, + 0.005260465666651726, + 0.01065447460860014, + 0.010452358983457088, + 0.01087111048400402, + 0.010914742015302181, + 0.00614912761375308, + 0.011396491900086403, + 0.009563572704792023, + 0.0044136615470051765, + 0.00465302774682641, + 0.0049816216342151165, + 0.006969050504267216, + 0.006775425747036934, + 0.0036086440086364746, + 0.004252517130225897, + 0.004203441087156534, + 0.005039015784859657, + 0.010283662006258965, + 0.004098894074559212, + 0.004530703648924828, + 0.007854980416595936, + 0.0037234770134091377, + 0.003460705280303955, + 0.0026251196395605803, + 0.02261405624449253, + 0.14208321273326874, + 0.006827307865023613, + 0.01495711226016283, + 0.12295503169298172, + 0.013059578835964203, + 0.010854230262339115, + 0.008480100892484188, + 0.005077207460999489, + 0.007148089352995157, + 0.004934204276651144, + 0.006376891862601042, + 0.005221509840339422, + 0.005333964712917805, + 0.00888310931622982, + 0.005107589531689882, + 0.008727067150175571, + 0.009839344769716263, + 0.005934137385338545, + 0.004594389349222183, + 0.0062766182236373425, + 0.00534206023439765, + 0.007230401039123535, + 0.006846714299172163, + 0.005614293273538351, + 0.007158302236348391, + 0.005253979004919529, + 0.003623586380854249, + 0.005752387456595898, + 0.005263567436486483, + 0.00482154032215476, + 0.00658509973436594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.01820380985736847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up captures the rocket engine's massive bell-shaped nozzle, its black, soot-stained metal surface covered in a complex network of pipes and thermal insulation. For a brief moment, the scene is utterly silent. Then, a dazzling flash of pure white instantly erupts from the engine's throat, dramatically expanding into a supersonic plume of incandescent plasma. The core of the exhaust burns with a bluish-white electrical intensity, clearly revealing the repetitive diamond-shaped shockwaves known as Mach diamonds. This core is surrounded by a roiling, surging torrent of flame, its color shifting from golden yellow to a deep, fiery orange, tumbling downward and outward at a ferocious pace. The flames engulf the vast, charred concrete launch pad below, while the flood system simultaneously unleashes a massive spray, producing a vast, billowing cloud of white steam that violently clashes and intermingles with the fire and smoke. The entire frame trembles intensely with the sheer acoustic force of the ignition, creating a visceral, chaotic, and awe-inspiring atmosphere of unleashing immense raw power.", + "seed": 18183, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_4/index_2597_seed_18183_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05484068840742111, + "vlm_logprob": -0.05484068840742111, + "vlm_yes_no_margin": 3.324999976158142, + "vlm_token_credit": -0.05484068840742111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08416423201560974 + ], + "opposite_logprob_tokens": [ + -2.5841641426086426 + ], + "logprob_reward": -0.08416423201560974, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08416423201560974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09708120673894882, + 0.03201974928379059, + 0.033802662044763565, + 0.06508281081914902, + 0.05717935413122177, + 0.2563970386981964, + 0.10682258009910583, + 0.05208870768547058, + 0.1840554177761078, + 0.11547055840492249 + ], + "credit_map": [ + 0.015811216086149216, + 0.005680026952177286, + 0.025506503880023956, + 0.007100434973835945, + 0.005845807958394289, + 0.006093017291277647, + 0.011107007041573524, + 0.0199371837079525, + 0.003512781113386154, + 0.0029485959094017744, + 0.003349922364577651, + 0.0026892516762018204, + 0.006594421342015266, + 0.0029563824646174908, + 0.003938302863389254, + 0.0060300868935883045, + 0.0027048916090279818, + 0.007612152490764856, + 0.004263326991349459, + 0.003407204058021307, + 0.007454929873347282, + 0.002283253939822316, + 0.0034787082113325596, + 0.0025981913786381483, + 0.004046749323606491, + 0.015349352732300758, + 0.006516449619084597, + 0.005805487744510174, + 0.014046883210539818, + 0.004135634750127792, + 0.007714240811765194, + 0.007468007039278746, + 0.0046599023044109344, + 0.008799385279417038, + 0.005981918424367905, + 0.005133284255862236, + 0.016393668949604034, + 0.003677569329738617, + 0.005156733561307192, + 0.007376884575933218, + 0.01619848422706127, + 0.07463786751031876, + 0.020104026421904564, + 0.036276936531066895, + 0.05757872015237808, + 0.011022896505892277, + 0.024949464946985245, + 0.015628617256879807, + 0.007543925661593676, + 0.025095243006944656, + 0.015333394519984722, + 0.010963731445372105, + 0.018562277778983116, + 0.0074576931074261665, + 0.012194277718663216, + 0.009672033600509167, + 0.00427928427234292, + 0.009154940024018288, + 0.00697900727391243, + 0.007237979210913181, + 0.00910718459635973, + 0.004157484043389559, + 0.00494540948420763, + 0.006227413192391396, + 0.012369669042527676, + 0.01042501162737608, + 0.011776997707784176, + 0.017001532018184662, + 0.009400161914527416, + 0.041968509554862976, + 0.015194963663816452, + 0.06591855734586716, + 0.009587171487510204, + 0.015338070690631866, + 0.015841621905565262, + 0.02169160731136799, + 0.010229644365608692, + 0.019117509946227074, + 0.011023378930985928, + 0.012641540728509426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.08416423201560974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12716111540794373 + ], + "opposite_logprob_tokens": [ + -2.1271610260009766 + ], + "logprob_reward": -0.12716111540794373, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12716111540794373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17286844551563263, + 0.1718529760837555, + 0.051347751170396805, + 0.046931199729442596, + 0.0283329039812088, + 0.09399622678756714, + 0.01957997865974903, + 0.07306826114654541, + 0.17965692281723022, + 0.1623653918504715 + ], + "credit_map": [ + 0.02984217368066311, + 0.011510577984154224, + 0.0558064728975296, + 0.013870221562683582, + 0.011293454095721245, + 0.014654655009508133, + 0.01622045412659645, + 0.019670430570840836, + 0.017799776047468185, + 0.014002472162246704, + 0.017293298617005348, + 0.00881999172270298, + 0.0533585399389267, + 0.00897111278027296, + 0.015301382169127464, + 0.036306384950876236, + 0.0054248021915555, + 0.009907864034175873, + 0.008953266777098179, + 0.006803832016885281, + 0.006397407501935959, + 0.0027906575705856085, + 0.005486076232045889, + 0.005583844147622585, + 0.005291219335049391, + 0.01551294419914484, + 0.004351767711341381, + 0.004186703357845545, + 0.00651157833635807, + 0.002383842598646879, + 0.0039441389963030815, + 0.004749006126075983, + 0.0023080692626535892, + 0.0035608934704214334, + 0.007178483530879021, + 0.0021932872477918863, + 0.003891428466886282, + 0.0018205696251243353, + 0.0017199022695422173, + 0.005660271737724543, + 0.006085788365453482, + 0.03171415254473686, + 0.006874919403344393, + 0.01394942682236433, + 0.01761629804968834, + 0.0037619308568537235, + 0.008241206407546997, + 0.005752501543611288, + 0.0016755795804783702, + 0.0026561100967228413, + 0.0025806226767599583, + 0.0023438299540430307, + 0.0036753874737769365, + 0.001564485370181501, + 0.0025958665646612644, + 0.0024880957789719105, + 0.006507322192192078, + 0.010958094149827957, + 0.0107569033280015, + 0.01060225535184145, + 0.011087195947766304, + 0.006910725962370634, + 0.007577444892376661, + 0.008668321184813976, + 0.011403756216168404, + 0.010516307316720486, + 0.01273482944816351, + 0.01566154696047306, + 0.010881529189646244, + 0.047173161059617996, + 0.013709762133657932, + 0.05757603049278259, + 0.01482292078435421, + 0.02064453437924385, + 0.022817187011241913, + 0.03127414360642433, + 0.011907321400940418, + 0.026440368965268135, + 0.014503207057714462, + 0.019955696538090706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.12716111540794373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657658472657204 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.008657658472657204, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657658472657204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0700494721531868, + 0.07879514247179031, + 0.04166729003190994, + 0.04769203066825867, + 0.07119233161211014, + 0.16277453303337097, + 0.039938680827617645, + 0.064665287733078, + 0.24750632047653198, + 0.1757189929485321 + ], + "credit_map": [ + 0.00942689273506403, + 0.0035482763778418303, + 0.024622291326522827, + 0.0057970499619841576, + 0.003961300943046808, + 0.0082487091422081, + 0.004750120919197798, + 0.009694823063910007, + 0.021642718464136124, + 0.007127148099243641, + 0.005984088405966759, + 0.00570487417280674, + 0.004259629640728235, + 0.006220174487680197, + 0.005981335416436195, + 0.0218751672655344, + 0.0034662948455661535, + 0.006001453846693039, + 0.003559454809874296, + 0.0049006096087396145, + 0.010985173285007477, + 0.003138372441753745, + 0.00591521430760622, + 0.0037007129285484552, + 0.0034117652103304863, + 0.01114016491919756, + 0.005941164214164019, + 0.0047819847241044044, + 0.009211532771587372, + 0.003246860345825553, + 0.005514849908649921, + 0.0044437008909881115, + 0.004144219681620598, + 0.030726779252290726, + 0.003112510312348604, + 0.0034138131886720657, + 0.018315399065613747, + 0.0032052521128207445, + 0.003578711301088333, + 0.004695642739534378, + 0.01060324814170599, + 0.0451163686811924, + 0.012416682206094265, + 0.02300253137946129, + 0.03690854460000992, + 0.007315285503864288, + 0.01653757132589817, + 0.010874279774725437, + 0.0037414238322526217, + 0.005439717322587967, + 0.005404507275670767, + 0.005042624659836292, + 0.0071943216025829315, + 0.0033132564276456833, + 0.005962252616882324, + 0.0038405745290219784, + 0.005601086653769016, + 0.012302224524319172, + 0.006943068001419306, + 0.006898174528032541, + 0.012581207789480686, + 0.005428866017609835, + 0.006308696698397398, + 0.008601964451372623, + 0.015580380335450172, + 0.013735925778746605, + 0.012951629236340523, + 0.01820369064807892, + 0.014078366570174694, + 0.06310909986495972, + 0.019222063943743706, + 0.09062512964010239, + 0.015482804737985134, + 0.020503317937254906, + 0.02438344992697239, + 0.03239686042070389, + 0.014725226908922195, + 0.02853928878903389, + 0.018161842599511147, + 0.021526195108890533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.008657658472657204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016090890392661095 + ], + "opposite_logprob_tokens": [ + -4.1410908699035645 + ], + "logprob_reward": -0.016090890392661095, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016090890392661095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07256373763084412, + 0.10123893618583679, + 0.04456543177366257, + 0.027993135154247284, + 0.03384575992822647, + 0.10135392844676971, + 0.03565602749586105, + 0.11897970736026764, + 0.34466323256492615, + 0.11914011836051941 + ], + "credit_map": [ + 0.009746826253831387, + 0.0036442347336560488, + 0.013894747011363506, + 0.00687635550275445, + 0.005012123845517635, + 0.012632597237825394, + 0.006215325091034174, + 0.014541523531079292, + 0.02691044472157955, + 0.0077507877722382545, + 0.011256183497607708, + 0.006391242146492004, + 0.006412488408386707, + 0.006461921613663435, + 0.007787190843373537, + 0.028268665075302124, + 0.004188137594610453, + 0.007599325384944677, + 0.007855772040784359, + 0.0051017333753407, + 0.008102165535092354, + 0.002944303210824728, + 0.003788823029026389, + 0.004985167644917965, + 0.0025152319576591253, + 0.005521781276911497, + 0.0035615982487797737, + 0.00279790791682899, + 0.00468968041241169, + 0.0017507071606814861, + 0.003034385619685054, + 0.004121840465813875, + 0.0016498349141329527, + 0.012923556379973888, + 0.002051250310614705, + 0.0016249020118266344, + 0.010535178706049919, + 0.0015565817011520267, + 0.001685215043835342, + 0.0018192385323345661, + 0.0065771546214818954, + 0.025715015828609467, + 0.010071936063468456, + 0.013881230726838112, + 0.023827333003282547, + 0.004691660404205322, + 0.010050713084638119, + 0.006538867950439453, + 0.001545208739116788, + 0.004318356513977051, + 0.0048703779466450214, + 0.0035200798884034157, + 0.004095201846212149, + 0.0018640165217220783, + 0.011933757923543453, + 0.003509027184918523, + 0.011696616187691689, + 0.019782323390245438, + 0.019348880276083946, + 0.014295020140707493, + 0.015955280512571335, + 0.00911867618560791, + 0.017970066517591476, + 0.010812841355800629, + 0.02333381213247776, + 0.01757698878645897, + 0.024939239025115967, + 0.032383330166339874, + 0.01673627831041813, + 0.08044371008872986, + 0.028718534857034683, + 0.12053129076957703, + 0.009712534956634045, + 0.015075831674039364, + 0.015381217002868652, + 0.023249231278896332, + 0.01212102547287941, + 0.01848774403333664, + 0.010700102895498276, + 0.014412423595786095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.016090890392661095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, extreme close-up of a rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nozzle black, soot-stained, and covered in pipes and thermal insulation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the exhaust core show bluish-white intensity with visible diamond-shaped shockwaves (Mach diamonds)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a massive spray of white steam clashing with fire and smoke on a charred concrete launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame tremble intensely to simulate the acoustic force of the ignition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038129545748233795 + ], + "opposite_logprob_tokens": [ + -3.2881295680999756 + ], + "logprob_reward": -0.038129545748233795, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038129545748233795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1499006301164627, + 0.1067253053188324, + 0.03391493111848831, + 0.03955555707216263, + 0.02834801748394966, + 0.11699486523866653, + 0.08289298415184021, + 0.11740771681070328, + 0.1759272664785385, + 0.14833275973796844 + ], + "credit_map": [ + 0.048720523715019226, + 0.006898392457515001, + 0.011167628690600395, + 0.006769427563995123, + 0.005297500640153885, + 0.010313255712389946, + 0.008093329146504402, + 0.05264057219028473, + 0.02621157094836235, + 0.00857253186404705, + 0.00774329574778676, + 0.005597204901278019, + 0.030700668692588806, + 0.005707272794097662, + 0.01168446708470583, + 0.010508292354643345, + 0.002953475108370185, + 0.008548401296138763, + 0.004407570697367191, + 0.003464071312919259, + 0.004949949216097593, + 0.002063695341348648, + 0.003398012835532427, + 0.004129755310714245, + 0.003271214198321104, + 0.013315606862306595, + 0.003810326801612973, + 0.0036925217136740685, + 0.006110172253102064, + 0.00215482828207314, + 0.003945822827517986, + 0.0032550639007240534, + 0.002673936542123556, + 0.008212381042540073, + 0.0030337958596646786, + 0.001992213074117899, + 0.006031219847500324, + 0.0013554473407566547, + 0.001838778262026608, + 0.0032102447003126144, + 0.0076348669826984406, + 0.034258339554071426, + 0.008967339992523193, + 0.01595453917980194, + 0.023978473618626595, + 0.005662052426487207, + 0.012066091410815716, + 0.008473171852529049, + 0.005555714946240187, + 0.010417109355330467, + 0.010832286439836025, + 0.009210284799337387, + 0.01029018685221672, + 0.004157198593020439, + 0.023823807016015053, + 0.00860639102756977, + 0.010853258892893791, + 0.018158651888370514, + 0.01684938371181488, + 0.010356772691011429, + 0.021767476573586464, + 0.010987280867993832, + 0.012309974059462547, + 0.016124919056892395, + 0.01165609061717987, + 0.01122128963470459, + 0.015897611156105995, + 0.021506965160369873, + 0.012248699553310871, + 0.03544112667441368, + 0.016010234132409096, + 0.051945239305496216, + 0.014790559187531471, + 0.01544642448425293, + 0.01903274841606617, + 0.026921316981315613, + 0.011960485950112343, + 0.026220984756946564, + 0.014813723042607307, + 0.019146516919136047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.038129545748233795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_tensor([9855, 9855])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_tensor([9855, 9855])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f83a3e71c4461ece5cd3ec88a077d006dcbade32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_tensor([9855, 9855])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot begins from a low angle behind the sports car. The camera follows closely behind it as a sleek, liquid-metal silver sports car speeds at breakneck speed across a multi-level elevated overpass constructed of sleek carbon and transparent forcefield guardrails. Its blazing red LED taillights create long, light-filled trails with a subtle particle effect. As the camera steadily and smoothly ascends and slowly zooms out, the car's silhouette gradually shrinks, becoming a mere point of light darting across the complex network of roads. The camera then widens, revealing the bottomless urban canyon below. The headlights and taillights of countless autonomous vehicles and small aircraft converge into rivers of gold, white, and blue light, flowing continuously through the intricate interchange structure. Finally, the camera rises to a breathtakingly magnificent bird's-eye view, taking in the entire panoramic view of the futuristic city under the night sky - huge dynamic holographic advertisements are projected on the exterior walls of skyscrapers with strange structures that soar into the clouds. The entire picture is dotted with countless dynamic light sources under the cold cyberpunk blue and purple main colors, creating a futuristic atmosphere that is both high-speed and dynamic, magnificent and awe-inspiring.", + "seed": 16682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_9855_seed_16682_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014962460892274976, + "vlm_logprob": -0.014962460892274976, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.014962460892274976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04030168801546097 + ], + "opposite_logprob_tokens": [ + -3.29030179977417 + ], + "logprob_reward": -0.04030168801546097, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04030168801546097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24614959955215454, + 0.042892467230558395, + 0.09747173637151718, + 0.035362157970666885, + 0.02824859507381916, + 0.06925081461668015, + 0.30010083317756653, + 0.11250431090593338, + 0.024508308619260788, + 0.04351120442152023 + ], + "credit_map": [ + 0.026160461828112602, + 0.024843135848641396, + 0.025364182889461517, + 0.026009412482380867, + 0.02068142034113407, + 0.04958604648709297, + 0.04688933864235878, + 0.02661561220884323, + 0.006432011723518372, + 0.00697658397257328, + 0.006277779117226601, + 0.0057130069471895695, + 0.004764243494719267, + 0.005619185045361519, + 0.002880566520616412, + 0.0042290897108614445, + 0.008419104851782322, + 0.012952854856848717, + 0.012011981569230556, + 0.013765700161457062, + 0.009794794023036957, + 0.013131829909980297, + 0.012226889841258526, + 0.015168575569987297, + 0.002533641178160906, + 0.006910312920808792, + 0.0024618618190288544, + 0.0038496782071888447, + 0.0038394120056182146, + 0.004164214711636305, + 0.005974413827061653, + 0.005628622602671385, + 0.003085136879235506, + 0.005652041640132666, + 0.0026746243238449097, + 0.0022699336986988783, + 0.0028286934830248356, + 0.0030609257519245148, + 0.0031605905387550592, + 0.005516647361218929, + 0.0061956485733389854, + 0.0072298599407076836, + 0.013698216527700424, + 0.007840740494430065, + 0.008449389599263668, + 0.013659101910889149, + 0.006647098809480667, + 0.005530756898224354, + 0.01889650709927082, + 0.016605935990810394, + 0.0432233028113842, + 0.016252391040325165, + 0.12455964088439941, + 0.020299920812249184, + 0.01105543039739132, + 0.049207720905542374, + 0.00968917179852724, + 0.008552003651857376, + 0.017872026190161705, + 0.021703552454710007, + 0.015700876712799072, + 0.011628281325101852, + 0.010194350965321064, + 0.017164044082164764, + 0.0017169578932225704, + 0.002660572063177824, + 0.003882239107042551, + 0.003756122197955847, + 0.002723682438954711, + 0.002052322495728731, + 0.00521873589605093, + 0.0024976758286356926, + 0.007587787229567766, + 0.009416760876774788, + 0.004327575676143169, + 0.005331095773726702, + 0.005026493687182665, + 0.004586744122207165, + 0.0036368216387927532, + 0.0035979212261736393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.04030168801546097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216690324246883 + ], + "opposite_logprob_tokens": [ + -4.264216899871826 + ], + "logprob_reward": -0.014216690324246883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216690324246883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11650752276182175, + 0.04611721634864807, + 0.11226024478673935, + 0.21111412346363068, + 0.0914117619395256, + 0.10928958654403687, + 0.04648497328162193, + 0.12232084572315216, + 0.044610846787691116, + 0.0998828262090683 + ], + "credit_map": [ + 0.012079577893018723, + 0.01116978656500578, + 0.01112719438970089, + 0.011701872572302818, + 0.012378794141113758, + 0.02509196475148201, + 0.021794987842440605, + 0.011163362301886082, + 0.004702240228652954, + 0.007194281090050936, + 0.005900766234844923, + 0.005636389367282391, + 0.007554157637059689, + 0.005051242653280497, + 0.005989997647702694, + 0.004088142421096563, + 0.009124531410634518, + 0.021982908248901367, + 0.010301406495273113, + 0.013597194105386734, + 0.013162226416170597, + 0.020240401849150658, + 0.010876654647290707, + 0.012974935583770275, + 0.024226197972893715, + 0.06232096254825592, + 0.021290211006999016, + 0.01878478191792965, + 0.02294786088168621, + 0.023190950974822044, + 0.011854386888444424, + 0.026498792693018913, + 0.008220267482101917, + 0.030052777379751205, + 0.004529647994786501, + 0.004352979827672243, + 0.007753951009362936, + 0.006234386935830116, + 0.010973337106406689, + 0.019294418394565582, + 0.009470086544752121, + 0.008257253095507622, + 0.008420447818934917, + 0.003945535514503717, + 0.026080744341015816, + 0.034700773656368256, + 0.010346857830882072, + 0.008067895658314228, + 0.0064634415321052074, + 0.0074620661325752735, + 0.00596094224601984, + 0.004842447582632303, + 0.0048207733780145645, + 0.007586278952658176, + 0.0048268036916852, + 0.00452222116291523, + 0.010535463690757751, + 0.013128949329257011, + 0.019861482083797455, + 0.022670816630125046, + 0.014508984982967377, + 0.012392998673021793, + 0.014770791865885258, + 0.014451373368501663, + 0.0036624763160943985, + 0.005452477838844061, + 0.006023041438311338, + 0.005757506005465984, + 0.005094518419355154, + 0.0043356106616556644, + 0.008215293288230896, + 0.006069926079362631, + 0.009680191986262798, + 0.012649995274841785, + 0.017232302576303482, + 0.02308826893568039, + 0.011888041161000729, + 0.008196542039513588, + 0.008528683334589005, + 0.008618814870715141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.014216690324246883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036201435141265392 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036201435141265392, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036201435141265392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19476579129695892, + 0.04794156923890114, + 0.19633930921554565, + 0.10432793200016022, + 0.11070530861616135, + 0.10575632005929947, + 0.06873879581689835, + 0.04127640277147293, + 0.049962203949689865, + 0.0801863744854927 + ], + "credit_map": [ + 0.01914973370730877, + 0.01951434649527073, + 0.022307997569441795, + 0.021557271480560303, + 0.022799523547291756, + 0.03243974968791008, + 0.037250224500894547, + 0.019746940582990646, + 0.007396565284579992, + 0.007932371459901333, + 0.006515095476061106, + 0.005459016188979149, + 0.006395953241735697, + 0.0064505175687372684, + 0.0043491218239068985, + 0.0034429291263222694, + 0.014976654201745987, + 0.04561568796634674, + 0.01770137809216976, + 0.022334853187203407, + 0.023761313408613205, + 0.03045519068837166, + 0.018176622688770294, + 0.02331758849322796, + 0.012267766520380974, + 0.00882133562117815, + 0.023549707606434822, + 0.010579408146440983, + 0.008610757999122143, + 0.011544625274837017, + 0.014955280348658562, + 0.013999045826494694, + 0.011858779937028885, + 0.04548591747879982, + 0.00524916872382164, + 0.004496647976338863, + 0.01029844582080841, + 0.01022516842931509, + 0.007416201755404472, + 0.015674982219934464, + 0.010587381199002266, + 0.010991234332323074, + 0.022525297477841377, + 0.010880961082875729, + 0.012945031747221947, + 0.015919476747512817, + 0.012184715829789639, + 0.009722220711410046, + 0.004905200097709894, + 0.004941205028444529, + 0.007929956540465355, + 0.004100910387933254, + 0.025382302701473236, + 0.006141341757029295, + 0.0028024972416460514, + 0.012535377405583858, + 0.0036291053984314203, + 0.003983512055128813, + 0.0065615540370345116, + 0.006718469318002462, + 0.005655467510223389, + 0.004423751961439848, + 0.005271259695291519, + 0.005033283494412899, + 0.0037957860622555017, + 0.004130932968109846, + 0.007036477327346802, + 0.006373271346092224, + 0.0055642821826040745, + 0.004925220739096403, + 0.009726480580866337, + 0.008409750647842884, + 0.009945671074092388, + 0.011146881617605686, + 0.012276963330805302, + 0.013246526010334492, + 0.009231614880263805, + 0.007852756418287754, + 0.009097645990550518, + 0.007388316560536623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0036201435141265392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187190681695938 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.014187190681695938, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187190681695938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09982550889253616, + 0.04951255023479462, + 0.16372661292552948, + 0.05747320130467415, + 0.13558758795261383, + 0.11720766127109528, + 0.209786057472229, + 0.05239798128604889, + 0.051037341356277466, + 0.06344552338123322 + ], + "credit_map": [ + 0.012533348985016346, + 0.010929487645626068, + 0.010982305742800236, + 0.013274822384119034, + 0.010494344867765903, + 0.019868001341819763, + 0.011202958412468433, + 0.010540253482758999, + 0.007561717182397842, + 0.007138787303119898, + 0.00685378722846508, + 0.007850279100239277, + 0.005349254701286554, + 0.006331378594040871, + 0.0044252160005271435, + 0.004002128262072802, + 0.009286337532103062, + 0.05608528479933739, + 0.00807703286409378, + 0.011299277655780315, + 0.022331759333610535, + 0.01803845725953579, + 0.007944351062178612, + 0.030664121732115746, + 0.008061432279646397, + 0.009752451442182064, + 0.007933675311505795, + 0.006142891477793455, + 0.00581655977293849, + 0.006850770208984613, + 0.006439897231757641, + 0.006475529167801142, + 0.012531661428511143, + 0.05348724499344826, + 0.006350868847221136, + 0.005657112691551447, + 0.011864415369927883, + 0.009123751893639565, + 0.007815578021109104, + 0.02875695936381817, + 0.011822265572845936, + 0.012065843679010868, + 0.018435152247548103, + 0.007197291124612093, + 0.019506551325321198, + 0.02800559811294079, + 0.009201934561133385, + 0.01097303256392479, + 0.014835936948657036, + 0.008955075405538082, + 0.02978619746863842, + 0.011125645600259304, + 0.08998893946409225, + 0.016233209520578384, + 0.011369702406227589, + 0.02749137207865715, + 0.003947057295590639, + 0.005553404334932566, + 0.008514412678778172, + 0.009185570292174816, + 0.006972366478294134, + 0.006648307200521231, + 0.005172287579625845, + 0.006404579617083073, + 0.004075384698808193, + 0.004661239683628082, + 0.007804322987794876, + 0.006484673824161291, + 0.008933457545936108, + 0.0037557631731033325, + 0.010863810777664185, + 0.004458692390471697, + 0.006482958793640137, + 0.00738001661375165, + 0.01048026792705059, + 0.00756149971857667, + 0.008858425542712212, + 0.006030377931892872, + 0.0066876765340566635, + 0.009964309632778168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.014187190681695938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248659192584455 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.00248659192584455, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248659192584455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09958565980195999, + 0.049256473779678345, + 0.07773289829492569, + 0.03826287016272545, + 0.03820400312542915, + 0.06320621818304062, + 0.18135838210582733, + 0.13021215796470642, + 0.1553870290517807, + 0.1667943149805069 + ], + "credit_map": [ + 0.009126277640461922, + 0.008756116032600403, + 0.009514493867754936, + 0.009111867286264896, + 0.013993324711918831, + 0.03235273435711861, + 0.008412833325564861, + 0.008318008854985237, + 0.008472147397696972, + 0.00888952799141407, + 0.005889419931918383, + 0.005266128573566675, + 0.007386207114905119, + 0.004982345271855593, + 0.0040470315143466, + 0.004323665052652359, + 0.004481039475649595, + 0.026559974998235703, + 0.004511295352131128, + 0.005124673247337341, + 0.01046033576130867, + 0.007917769253253937, + 0.0034305904991924763, + 0.015247223898768425, + 0.003207782981917262, + 0.008957616053521633, + 0.003263546619564295, + 0.0044328090734779835, + 0.0038240605499595404, + 0.003107276977971196, + 0.00432279147207737, + 0.007146985735744238, + 0.005951936822384596, + 0.004003122914582491, + 0.0037728482857346535, + 0.0035421669017523527, + 0.004516162443906069, + 0.007374119479209185, + 0.004243992269039154, + 0.004799654241651297, + 0.006387557368725538, + 0.0083593325689435, + 0.003915156703442335, + 0.004397421143949032, + 0.012530220672488213, + 0.011131701059639454, + 0.010325409471988678, + 0.006159421056509018, + 0.01223491970449686, + 0.01078970916569233, + 0.026207393035292625, + 0.010473320260643959, + 0.07723432779312134, + 0.014939654618501663, + 0.007156556472182274, + 0.022322485223412514, + 0.012036735191941261, + 0.016341280192136765, + 0.022608313709497452, + 0.0191685501486063, + 0.019556891173124313, + 0.012715592049062252, + 0.015842381864786148, + 0.011942421086132526, + 0.01176824513822794, + 0.01281599048525095, + 0.02162710577249527, + 0.023769931867718697, + 0.019546089693903923, + 0.015361598692834377, + 0.02803860604763031, + 0.02245945855975151, + 0.02635168470442295, + 0.032804738730192184, + 0.02154809795320034, + 0.02409546822309494, + 0.020396368578076363, + 0.016557995229959488, + 0.011639765463769436, + 0.013400201685726643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.00248659192584455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot begins from a low angle behind the sports car. The camera follows closely behind it as a sleek, liquid-metal silver sports car speeds at breakneck speed across a multi-level elevated overpass constructed of sleek carbon and transparent forcefield guardrails. Its blazing red LED taillights create long, light-filled trails with a subtle particle effect. As the camera steadily and smoothly ascends and slowly zooms out, the car's silhouette gradually shrinks, becoming a mere point of light darting across the complex network of roads. The camera then widens, revealing the bottomless urban canyon below. The headlights and taillights of countless autonomous vehicles and small aircraft converge into rivers of gold, white, and blue light, flowing continuously through the intricate interchange structure. Finally, the camera rises to a breathtakingly magnificent bird's-eye view, taking in the entire panoramic view of the futuristic city under the night sky - huge dynamic holographic advertisements are projected on the exterior walls of skyscrapers with strange structures that soar into the clouds. The entire picture is dotted with countless dynamic light sources under the cold cyberpunk blue and purple main colors, creating a futuristic atmosphere that is both high-speed and dynamic, magnificent and awe-inspiring.", + "seed": 16683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_9855_seed_16683_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.036492179008200766, + "vlm_logprob": -0.036492179008200766, + "vlm_yes_no_margin": 4.2999999761581424, + "vlm_token_credit": -0.036492179008200766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12915952503681183 + ], + "opposite_logprob_tokens": [ + -2.129159450531006 + ], + "logprob_reward": -0.12915952503681183, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12915952503681183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23221717774868011, + 0.16263499855995178, + 0.05300668999552727, + 0.060670699924230576, + 0.03490927070379257, + 0.04504108056426048, + 0.1003151684999466, + 0.21571533381938934, + 0.0474906750023365, + 0.047998908907175064 + ], + "credit_map": [ + 0.037998735904693604, + 0.03709905222058296, + 0.01334371604025364, + 0.019867293536663055, + 0.036038838326931, + 0.030110113322734833, + 0.03254614770412445, + 0.02521328628063202, + 0.020578650757670403, + 0.020183058455586433, + 0.013982698321342468, + 0.012660989537835121, + 0.022658824920654297, + 0.018435172736644745, + 0.024774985387921333, + 0.02936062216758728, + 0.006304738577455282, + 0.006208666134625673, + 0.008238179609179497, + 0.004001200664788485, + 0.005832401569932699, + 0.010395325720310211, + 0.005709334742277861, + 0.006316846236586571, + 0.007467837538570166, + 0.009950339794158936, + 0.006480073556303978, + 0.010324395261704922, + 0.007595868781208992, + 0.0065965610556304455, + 0.0054342192597687244, + 0.006821402348577976, + 0.004231762140989304, + 0.004442189820110798, + 0.0036104607861489058, + 0.005570229142904282, + 0.005566096864640713, + 0.0038125861901789904, + 0.004401695914566517, + 0.0032742524053901434, + 0.004819189198315144, + 0.005197139922529459, + 0.005351497791707516, + 0.011559034697711468, + 0.004249295219779015, + 0.0051430100575089455, + 0.005854056216776371, + 0.002867857227101922, + 0.014046881347894669, + 0.016105245798826218, + 0.009023421443998814, + 0.005979260429739952, + 0.013052612543106079, + 0.022579187527298927, + 0.012415566481649876, + 0.007112992461770773, + 0.014141635037958622, + 0.06157216802239418, + 0.011945601552724838, + 0.008933988399803638, + 0.017583483830094337, + 0.06877265125513077, + 0.019898749887943268, + 0.012867063283920288, + 0.005698021501302719, + 0.003371582133695483, + 0.004187350627034903, + 0.0045181880705058575, + 0.003471740521490574, + 0.0064359866082668304, + 0.010519800707697868, + 0.009288002736866474, + 0.006242795381695032, + 0.005637196358293295, + 0.005707664415240288, + 0.007150138262659311, + 0.006063451059162617, + 0.005107554141432047, + 0.00527457008138299, + 0.006815535016357899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.12915952503681183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372011333703995 + ], + "opposite_logprob_tokens": [ + -3.408720016479492 + ], + "logprob_reward": -0.03372011333703995, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372011333703995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07051855325698853, + 0.09557834267616272, + 0.07686756551265717, + 0.32648274302482605, + 0.05505939945578575, + 0.06973497569561005, + 0.07258764654397964, + 0.0666160061955452, + 0.08902285248041153, + 0.07753190398216248 + ], + "credit_map": [ + 0.010354459285736084, + 0.010978072881698608, + 0.005556771066039801, + 0.005637794733047485, + 0.011897928081452847, + 0.01166957151144743, + 0.008590798825025558, + 0.005833160597831011, + 0.016627883538603783, + 0.012358319014310837, + 0.009891155175864697, + 0.0061395480297505856, + 0.00881414208561182, + 0.012765162624418736, + 0.012460565194487572, + 0.016521571204066277, + 0.009173701517283916, + 0.007909647189080715, + 0.015368464402854443, + 0.008626912720501423, + 0.0074859969317913055, + 0.012274333275854588, + 0.007385255768895149, + 0.00864326860755682, + 0.028839990496635437, + 0.06270740181207657, + 0.06009291484951973, + 0.038784727454185486, + 0.03050418570637703, + 0.037522055208683014, + 0.02714698016643524, + 0.04088449478149414, + 0.004526840522885323, + 0.005038945935666561, + 0.00999867357313633, + 0.009567362256348133, + 0.007200583349913359, + 0.006525970529764891, + 0.005404001567512751, + 0.006797025911509991, + 0.008306268602609634, + 0.008098714984953403, + 0.009586541913449764, + 0.017321547493338585, + 0.00512270350009203, + 0.00948646292090416, + 0.006199880037456751, + 0.005612868349999189, + 0.009772143326699734, + 0.009915895760059357, + 0.008017734624445438, + 0.004105396568775177, + 0.007256780751049519, + 0.01978568360209465, + 0.008635728619992733, + 0.005098293535411358, + 0.005202577915042639, + 0.01704823598265648, + 0.005068018566817045, + 0.0034071975387632847, + 0.00656190887093544, + 0.017474085092544556, + 0.0073410761542618275, + 0.0045129070058465, + 0.012340735644102097, + 0.006595105864107609, + 0.007698035798966885, + 0.009215116500854492, + 0.008534573018550873, + 0.01198051031678915, + 0.017183244228363037, + 0.015475545078516006, + 0.009176266379654408, + 0.0092406515032053, + 0.007073980290442705, + 0.008642734959721565, + 0.01288426574319601, + 0.010466906242072582, + 0.009476709179580212, + 0.01057039387524128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.03372011333703995 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031969649717211723 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.0031969649717211723, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031969649717211723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11438976973295212, + 0.15567809343338013, + 0.07625989615917206, + 0.2426750808954239, + 0.08331110328435898, + 0.07260134071111679, + 0.03454851359128952, + 0.12853974103927612, + 0.039135321974754333, + 0.05286116898059845 + ], + "credit_map": [ + 0.02499423548579216, + 0.013211985118687153, + 0.008359485305845737, + 0.011638762429356575, + 0.016134049743413925, + 0.013893106952309608, + 0.013990315608680248, + 0.012167826294898987, + 0.019896915182471275, + 0.017860088497400284, + 0.014581193216145039, + 0.016284607350826263, + 0.02248290739953518, + 0.0168442465364933, + 0.02343861199915409, + 0.024289516732096672, + 0.008774416521191597, + 0.008451083675026894, + 0.010862947441637516, + 0.011146295815706253, + 0.007636648137122393, + 0.012760120443999767, + 0.008058558218181133, + 0.008569817058742046, + 0.023403191938996315, + 0.047771260142326355, + 0.029975837096571922, + 0.04020420089364052, + 0.02317076176404953, + 0.027335625141859055, + 0.021839212626218796, + 0.028974970802664757, + 0.008014438673853874, + 0.010426613502204418, + 0.008511094376444817, + 0.013878969475626945, + 0.011684923432767391, + 0.008795259520411491, + 0.013728367164731026, + 0.008271429687738419, + 0.004833058454096317, + 0.0061187464743852615, + 0.010808488354086876, + 0.019723324105143547, + 0.005435976665467024, + 0.00773856695741415, + 0.011427531950175762, + 0.006515640299767256, + 0.004533277824521065, + 0.006151385139673948, + 0.0029656426049768925, + 0.0022189300507307053, + 0.006424696184694767, + 0.0038028473500162363, + 0.004025950096547604, + 0.00442578224465251, + 0.009962941519916058, + 0.03881389647722244, + 0.006508492864668369, + 0.0047842636704444885, + 0.009827270172536373, + 0.04361734166741371, + 0.008303147740662098, + 0.0067223841324448586, + 0.005064882338047028, + 0.002741612261161208, + 0.005872156471014023, + 0.004658611491322517, + 0.005219029262661934, + 0.004911674652248621, + 0.004105360247194767, + 0.006561990827322006, + 0.006551851984113455, + 0.005457740742713213, + 0.004902918357402086, + 0.005431007593870163, + 0.00805281288921833, + 0.006051693577319384, + 0.007892823778092861, + 0.008520320057868958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.0031969649717211723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188366010785103 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188366010785103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188366010785103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24752385914325714, + 0.07561983168125153, + 0.05699894204735756, + 0.05645323172211647, + 0.03595441207289696, + 0.039005495607852936, + 0.04732494056224823, + 0.0770392045378685, + 0.2581125497817993, + 0.10596754401922226 + ], + "credit_map": [ + 0.032491400837898254, + 0.04145331680774689, + 0.014344247058033943, + 0.02197945863008499, + 0.049180179834365845, + 0.03349746763706207, + 0.0314304456114769, + 0.023147327825427055, + 0.009898601099848747, + 0.00947515293955803, + 0.007159726228564978, + 0.007038107141852379, + 0.009860356338322163, + 0.00906557496637106, + 0.010346117429435253, + 0.012776194140315056, + 0.007311419118195772, + 0.00738187599927187, + 0.010790625587105751, + 0.0074646957218647, + 0.0059764189645648, + 0.006285518873482943, + 0.005590520799160004, + 0.0061978609301149845, + 0.0044929515570402145, + 0.01122521236538887, + 0.009052188135683537, + 0.007675869856029749, + 0.006316724698990583, + 0.006578770000487566, + 0.004895614460110664, + 0.00621589832007885, + 0.0036548494827002287, + 0.0029848353005945683, + 0.006345424801111221, + 0.006083441898226738, + 0.0044817086309194565, + 0.004711614456027746, + 0.003337149042636156, + 0.004355390090495348, + 0.005178708117455244, + 0.003814972471445799, + 0.004126775544136763, + 0.00635659322142601, + 0.0036776019260287285, + 0.007080062758177519, + 0.003964938689023256, + 0.004805843345820904, + 0.006528256461024284, + 0.008981820195913315, + 0.004368767142295837, + 0.002711037639528513, + 0.004421635996550322, + 0.013346970081329346, + 0.0043352218344807625, + 0.0026312321424484253, + 0.006051871459931135, + 0.017671747133135796, + 0.00559542141854763, + 0.004223785363137722, + 0.007288035470992327, + 0.019826699048280716, + 0.010716745629906654, + 0.005664896685630083, + 0.03007662110030651, + 0.01490611582994461, + 0.02042410895228386, + 0.031009424477815628, + 0.02369794435799122, + 0.030386600643396378, + 0.06191715598106384, + 0.04569457098841667, + 0.010873946361243725, + 0.015762437134981155, + 0.009451261721551418, + 0.010688433423638344, + 0.020263364538550377, + 0.012681897729635239, + 0.014529099687933922, + 0.011717105284333229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.014188366010785103 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195925684645772 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195925684645772, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195925684645772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3193162679672241, + 0.10351752489805222, + 0.05236468464136124, + 0.06075025349855423, + 0.04208388179540634, + 0.14344318211078644, + 0.054587263613939285, + 0.05624907463788986, + 0.05769066885113716, + 0.10999718308448792 + ], + "credit_map": [ + 0.044482797384262085, + 0.05261527746915817, + 0.023253247141838074, + 0.023503541946411133, + 0.060446783900260925, + 0.04800337925553322, + 0.03501441329717636, + 0.03199682757258415, + 0.01866627112030983, + 0.012051251716911793, + 0.008946448564529419, + 0.0077020227909088135, + 0.013101526536047459, + 0.014271457679569721, + 0.012808533385396004, + 0.015970012173056602, + 0.005436989478766918, + 0.005570190493017435, + 0.009905876591801643, + 0.004678974859416485, + 0.00608757883310318, + 0.010653599165380001, + 0.003923216834664345, + 0.006108255125582218, + 0.005312396213412285, + 0.006591222248971462, + 0.010976237244904041, + 0.0167169701308012, + 0.005815540440380573, + 0.004662389867007732, + 0.004905053414404392, + 0.005770443473011255, + 0.003726646536961198, + 0.0034361633006483316, + 0.005098142195492983, + 0.008117809891700745, + 0.005510734394192696, + 0.005167088937014341, + 0.006373003590852022, + 0.0046542938798666, + 0.012785252183675766, + 0.017392486333847046, + 0.017444344237446785, + 0.03771154209971428, + 0.01310703530907631, + 0.019453922286629677, + 0.012465267442166805, + 0.013083331286907196, + 0.006292670965194702, + 0.005460942629724741, + 0.0071543138474226, + 0.004747612401843071, + 0.005282636731863022, + 0.012800603173673153, + 0.005901349242776632, + 0.006947132758796215, + 0.006764340680092573, + 0.00921959150582552, + 0.006526305805891752, + 0.0034630487207323313, + 0.007971852086484432, + 0.01016157679259777, + 0.006361599545925856, + 0.005780757870525122, + 0.01042335107922554, + 0.005339656490832567, + 0.008626420982182026, + 0.006739032454788685, + 0.004463830031454563, + 0.007558369543403387, + 0.006860919296741486, + 0.0076790908351540565, + 0.011975253000855446, + 0.011925173923373222, + 0.010745198465883732, + 0.013497769832611084, + 0.016730420291423798, + 0.016156159341335297, + 0.013367811217904091, + 0.015599391423165798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.002195925684645772 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot begins from a low angle behind the sports car. The camera follows closely behind it as a sleek, liquid-metal silver sports car speeds at breakneck speed across a multi-level elevated overpass constructed of sleek carbon and transparent forcefield guardrails. Its blazing red LED taillights create long, light-filled trails with a subtle particle effect. As the camera steadily and smoothly ascends and slowly zooms out, the car's silhouette gradually shrinks, becoming a mere point of light darting across the complex network of roads. The camera then widens, revealing the bottomless urban canyon below. The headlights and taillights of countless autonomous vehicles and small aircraft converge into rivers of gold, white, and blue light, flowing continuously through the intricate interchange structure. Finally, the camera rises to a breathtakingly magnificent bird's-eye view, taking in the entire panoramic view of the futuristic city under the night sky - huge dynamic holographic advertisements are projected on the exterior walls of skyscrapers with strange structures that soar into the clouds. The entire picture is dotted with countless dynamic light sources under the cold cyberpunk blue and purple main colors, creating a futuristic atmosphere that is both high-speed and dynamic, magnificent and awe-inspiring.", + "seed": 16684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_9855_seed_16684_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016995724430307745, + "vlm_logprob": -0.016995724430307745, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.016995724430307745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05036255717277527 + ], + "opposite_logprob_tokens": [ + -3.0503625869750977 + ], + "logprob_reward": -0.05036255717277527, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05036255717277527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26543816924095154, + 0.041155457496643066, + 0.018518220633268356, + 0.3846553862094879, + 0.048960935324430466, + 0.07158263772726059, + 0.037449996918439865, + 0.05709706246852875, + 0.025950191542506218, + 0.04919186607003212 + ], + "credit_map": [ + 0.037099290639162064, + 0.0639580488204956, + 0.020299894735217094, + 0.018750006332993507, + 0.03881102055311203, + 0.03640987351536751, + 0.017852816730737686, + 0.03225722163915634, + 0.005828325171023607, + 0.011073293164372444, + 0.0027468190528452396, + 0.0025739131961017847, + 0.005843878258019686, + 0.004428703337907791, + 0.00362254586070776, + 0.005037982016801834, + 0.002862154273316264, + 0.0024345084093511105, + 0.0022016416769474745, + 0.0016667158342897892, + 0.001889046048745513, + 0.0026447121053934097, + 0.0023284258786588907, + 0.002491017570719123, + 0.021709315478801727, + 0.03067576326429844, + 0.07767307013273239, + 0.06831778585910797, + 0.02993306703865528, + 0.025084037333726883, + 0.07376798242330551, + 0.05749436095356941, + 0.006509166676551104, + 0.005560354795306921, + 0.008620453998446465, + 0.005477086175233126, + 0.007265259046107531, + 0.005788733717054129, + 0.004975642543286085, + 0.004764236509799957, + 0.008641643449664116, + 0.008787592872977257, + 0.007575510069727898, + 0.010447198525071144, + 0.006584678310900927, + 0.00938380416482687, + 0.012581314891576767, + 0.007580897770822048, + 0.0052072955295443535, + 0.005551845300942659, + 0.006104405969381332, + 0.0045739877969026566, + 0.004027542192488909, + 0.00487557565793395, + 0.0033608514349907637, + 0.0037484962958842516, + 0.00656814593821764, + 0.006710839923471212, + 0.007651360239833593, + 0.010835141874849796, + 0.0052764518186450005, + 0.006838019471615553, + 0.005705532152205706, + 0.007511570118367672, + 0.006167235318571329, + 0.003118601394817233, + 0.002327612368389964, + 0.002586052753031254, + 0.002129149390384555, + 0.0034059123136103153, + 0.0027654911391437054, + 0.0034501361660659313, + 0.007086409721523523, + 0.005904417019337416, + 0.007239639759063721, + 0.0042947386391460896, + 0.005808355752378702, + 0.00728959497064352, + 0.007788707036525011, + 0.003780003637075424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.05036255717277527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865340419113636 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.00865340419113636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865340419113636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14515690505504608, + 0.09861895442008972, + 0.08815412223339081, + 0.1897425353527069, + 0.08251437544822693, + 0.06361176818609238, + 0.16220957040786743, + 0.07356994599103928, + 0.06943226605653763, + 0.02698947675526142 + ], + "credit_map": [ + 0.01580933853983879, + 0.02855313941836357, + 0.013852881267666817, + 0.012386005371809006, + 0.020839491859078407, + 0.01678912714123726, + 0.0202341191470623, + 0.01669280230998993, + 0.014192361384630203, + 0.02122572436928749, + 0.006689704954624176, + 0.006621928419917822, + 0.012552648782730103, + 0.01634175144135952, + 0.00929740909487009, + 0.011697427369654179, + 0.020930519327521324, + 0.008701138198375702, + 0.0072023700922727585, + 0.007246823515743017, + 0.012238660827279091, + 0.015560965053737164, + 0.0059575727209448814, + 0.010316068306565285, + 0.011221787892282009, + 0.014756906777620316, + 0.04123507812619209, + 0.03164101392030716, + 0.014483819715678692, + 0.01155642606317997, + 0.03404071182012558, + 0.03080678917467594, + 0.009609458036720753, + 0.015111396089196205, + 0.013680611737072468, + 0.007378415670245886, + 0.01205032505095005, + 0.005912696477025747, + 0.007486283779144287, + 0.011285186745226383, + 0.011415093205869198, + 0.00887758657336235, + 0.006983680184930563, + 0.006677365396171808, + 0.008960223756730556, + 0.009871204383671284, + 0.005730337463319302, + 0.005096282344311476, + 0.031334251165390015, + 0.028512336313724518, + 0.0315525159239769, + 0.018967628479003906, + 0.013065798208117485, + 0.010968455113470554, + 0.01499925833195448, + 0.012809345498681068, + 0.009765303693711758, + 0.00857199914753437, + 0.014362447895109653, + 0.009587344713509083, + 0.005562508944422007, + 0.008474522270262241, + 0.007196844555437565, + 0.010048975236713886, + 0.008945910260081291, + 0.006589206866919994, + 0.006775210611522198, + 0.0057140979915857315, + 0.00850227102637291, + 0.016921866685152054, + 0.008016425184905529, + 0.007967282086610794, + 0.0031843166798353195, + 0.005234426353126764, + 0.0024888762272894382, + 0.002152771921828389, + 0.0031946240924298763, + 0.005070165265351534, + 0.002322777174413204, + 0.00334151997230947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.00865340419113636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200683578848839 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200683578848839, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200683578848839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08072847127914429, + 0.038360271602869034, + 0.04213229566812515, + 0.5089104771614075, + 0.10320291668176651, + 0.06298534572124481, + 0.03894611820578575, + 0.048551175743341446, + 0.03881623595952988, + 0.03736665099859238 + ], + "credit_map": [ + 0.01159241609275341, + 0.012065325863659382, + 0.009330899454653263, + 0.006378576625138521, + 0.01024678722023964, + 0.013260019943118095, + 0.008395830169320107, + 0.009458616375923157, + 0.0032756978180259466, + 0.0067727165296673775, + 0.005085602402687073, + 0.004406465217471123, + 0.004310109186917543, + 0.0032158582471311092, + 0.006635991856455803, + 0.004657830577343702, + 0.004439851269125938, + 0.0055970377288758755, + 0.003822031896561384, + 0.003912642132490873, + 0.0074690598994493484, + 0.004682130180299282, + 0.006563086528331041, + 0.005646454636007547, + 0.028818650171160698, + 0.04288897290825844, + 0.09473675489425659, + 0.09486428648233414, + 0.04111245647072792, + 0.03209482133388519, + 0.10279421508312225, + 0.07160035520792007, + 0.011906222440302372, + 0.014361622743308544, + 0.016709843650460243, + 0.012022126466035843, + 0.01144844014197588, + 0.007755936123430729, + 0.010790922679007053, + 0.018207799643278122, + 0.011764474213123322, + 0.00745415361598134, + 0.006005153059959412, + 0.008257625624537468, + 0.006198299583047628, + 0.009119640104472637, + 0.00813636090606451, + 0.006049639545381069, + 0.005167482420802116, + 0.007171661127358675, + 0.004826117306947708, + 0.005666397977620363, + 0.003440725151449442, + 0.0031899791210889816, + 0.006093316711485386, + 0.0033904388546943665, + 0.007054516579955816, + 0.0047211842611432076, + 0.005984388291835785, + 0.006869002245366573, + 0.0040056114085018635, + 0.005976540967822075, + 0.005097793415188789, + 0.008842140436172485, + 0.0068759070709347725, + 0.003922899253666401, + 0.003821493359282613, + 0.004162969999015331, + 0.005814423784613609, + 0.005131229292601347, + 0.004459541290998459, + 0.0046277702786028385, + 0.004915934521704912, + 0.004024446941912174, + 0.0038565443828701973, + 0.0033633008133620024, + 0.004391465336084366, + 0.005818339064717293, + 0.005405409261584282, + 0.005591210909187794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.002200683578848839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572498440742493 + ], + "opposite_logprob_tokens": [ + -3.8955724239349365 + ], + "logprob_reward": -0.020572498440742493, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572498440742493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09951464831829071, + 0.034063469618558884, + 0.03022133745253086, + 0.3878192603588104, + 0.06873916834592819, + 0.08031334728002548, + 0.05504601448774338, + 0.03879416361451149, + 0.15857237577438354, + 0.04691629484295845 + ], + "credit_map": [ + 0.011903288774192333, + 0.019541392102837563, + 0.007195222657173872, + 0.006904115900397301, + 0.016788125038146973, + 0.015777546912431717, + 0.008281880989670753, + 0.013123052194714546, + 0.004397011362016201, + 0.005876490380614996, + 0.0032034311443567276, + 0.003170877695083618, + 0.004830839112401009, + 0.004536795429885387, + 0.004322970751672983, + 0.003725046757608652, + 0.0033032558858394623, + 0.004537421744316816, + 0.003094204468652606, + 0.003195013152435422, + 0.005684984847903252, + 0.0026402862276881933, + 0.0037273638881742954, + 0.004038801416754723, + 0.023374874144792557, + 0.02998068556189537, + 0.08208723366260529, + 0.06571237742900848, + 0.030545992776751518, + 0.025852058082818985, + 0.069710873067379, + 0.0605551041662693, + 0.00707377539947629, + 0.011339211836457253, + 0.012745420448482037, + 0.006526861805468798, + 0.007924690842628479, + 0.0051423911936581135, + 0.004601267632097006, + 0.013385532423853874, + 0.012415076605975628, + 0.009565548039972782, + 0.007959251292049885, + 0.011002697981894016, + 0.007986629381775856, + 0.013314193114638329, + 0.010540809482336044, + 0.0075291176326572895, + 0.007610324304550886, + 0.008050380274653435, + 0.005879563745111227, + 0.009003669954836369, + 0.005865386687219143, + 0.0052027287892997265, + 0.008485397323966026, + 0.004948550369590521, + 0.0040205009281635284, + 0.005230875685811043, + 0.0056465850211679935, + 0.0023161228746175766, + 0.0027811205945909023, + 0.002990389708429575, + 0.00409406004473567, + 0.011714501306414604, + 0.022424301132559776, + 0.017554813995957375, + 0.017074501141905785, + 0.015015127137303352, + 0.020596658810973167, + 0.027360428124666214, + 0.018819432705640793, + 0.019727088510990143, + 0.005528483539819717, + 0.006813030689954758, + 0.004581224638968706, + 0.003506023669615388, + 0.005550736095756292, + 0.006240047048777342, + 0.010051482357084751, + 0.004645253997296095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.020572498440742493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031894787680357695 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.0031894787680357695, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031894787680357695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06122541427612305, + 0.030107684433460236, + 0.05521106347441673, + 0.5077932476997375, + 0.148503378033638, + 0.02173941768705845, + 0.05610746517777443, + 0.017961522564291954, + 0.06906527280807495, + 0.03228554502129555 + ], + "credit_map": [ + 0.007441266439855099, + 0.011768477037549019, + 0.005374870728701353, + 0.005469148512929678, + 0.009004024788737297, + 0.007231516297906637, + 0.007808254566043615, + 0.007127853576093912, + 0.004140091128647327, + 0.006017202511429787, + 0.0028305298183113337, + 0.0026230306830257177, + 0.0031796216499060392, + 0.0038175303488969803, + 0.004145454615354538, + 0.0033542218152433634, + 0.006802737712860107, + 0.005217762663960457, + 0.006039828527718782, + 0.006914745084941387, + 0.007134061772376299, + 0.007490823976695538, + 0.008919890969991684, + 0.00669121602550149, + 0.02906320057809353, + 0.04060308262705803, + 0.10343413054943085, + 0.0894315242767334, + 0.03968821093440056, + 0.03226502984762192, + 0.09559543430805206, + 0.07771264761686325, + 0.013605034910142422, + 0.02608385868370533, + 0.028688399121165276, + 0.018699361011385918, + 0.014760770834982395, + 0.00849918182939291, + 0.014780162833631039, + 0.02338661625981331, + 0.003307392355054617, + 0.002319223014637828, + 0.0025074605364352465, + 0.002605671528726816, + 0.0027908501215279102, + 0.003275967203080654, + 0.0026680161245167255, + 0.002264835638925433, + 0.011569243855774403, + 0.011283387430012226, + 0.006736929062753916, + 0.007964137010276318, + 0.004779187962412834, + 0.003972656559199095, + 0.005212440621107817, + 0.004589482676237822, + 0.0023418348282575607, + 0.0016761963488534093, + 0.00243512075394392, + 0.001710703014396131, + 0.001366051030345261, + 0.0015251293079927564, + 0.0025184955447912216, + 0.004387991968542337, + 0.013088105246424675, + 0.006238706409931183, + 0.006151563487946987, + 0.0030378529336303473, + 0.01011586468666792, + 0.016186360269784927, + 0.005499795079231262, + 0.00874701701104641, + 0.005371424835175276, + 0.004292669240385294, + 0.0023276833817362785, + 0.0017594483215361834, + 0.003920276183634996, + 0.0047861915081739426, + 0.006062374450266361, + 0.0037654764018952847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.0031894787680357695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot begins from a low angle behind the sports car. The camera follows closely behind it as a sleek, liquid-metal silver sports car speeds at breakneck speed across a multi-level elevated overpass constructed of sleek carbon and transparent forcefield guardrails. Its blazing red LED taillights create long, light-filled trails with a subtle particle effect. As the camera steadily and smoothly ascends and slowly zooms out, the car's silhouette gradually shrinks, becoming a mere point of light darting across the complex network of roads. The camera then widens, revealing the bottomless urban canyon below. The headlights and taillights of countless autonomous vehicles and small aircraft converge into rivers of gold, white, and blue light, flowing continuously through the intricate interchange structure. Finally, the camera rises to a breathtakingly magnificent bird's-eye view, taking in the entire panoramic view of the futuristic city under the night sky - huge dynamic holographic advertisements are projected on the exterior walls of skyscrapers with strange structures that soar into the clouds. The entire picture is dotted with countless dynamic light sources under the cold cyberpunk blue and purple main colors, creating a futuristic atmosphere that is both high-speed and dynamic, magnificent and awe-inspiring.", + "seed": 16685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_5/index_9855_seed_16685_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02857745923101902, + "vlm_logprob": -0.02857745923101902, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02857745923101902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10160303115844727 + ], + "opposite_logprob_tokens": [ + -2.3516030311584473 + ], + "logprob_reward": -0.10160303115844727, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10160303115844727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14305752515792847, + 0.10955323278903961, + 0.0403040312230587, + 0.08097554743289948, + 0.24969714879989624, + 0.06417364627122879, + 0.060697898268699646, + 0.062198664993047714, + 0.137759730219841, + 0.05158258229494095 + ], + "credit_map": [ + 0.015503725968301296, + 0.011738371104001999, + 0.013593433424830437, + 0.02372184954583645, + 0.012892888858914375, + 0.012603933922946453, + 0.02846505306661129, + 0.024538269266486168, + 0.01745685562491417, + 0.007937242276966572, + 0.03031894378364086, + 0.007940985262393951, + 0.009240999817848206, + 0.016746262088418007, + 0.007609161082655191, + 0.012302788905799389, + 0.005266470368951559, + 0.005746635142713785, + 0.005991569720208645, + 0.00591236911714077, + 0.006680582650005817, + 0.0027493038214743137, + 0.0056440141052007675, + 0.0023130872286856174, + 0.009887656196951866, + 0.009621500968933105, + 0.012601702474057674, + 0.012215915136039257, + 0.006357504986226559, + 0.00562977883964777, + 0.015414736233651638, + 0.009246747940778732, + 0.021670391783118248, + 0.0719301700592041, + 0.03427997604012489, + 0.040763311088085175, + 0.0220682080835104, + 0.012500342912971973, + 0.032390058040618896, + 0.014094674959778786, + 0.010048422031104565, + 0.013523552566766739, + 0.0070945098996162415, + 0.008997299708425999, + 0.005404759664088488, + 0.005669138859957457, + 0.00566891860216856, + 0.00776704354211688, + 0.00945480540394783, + 0.015274573117494583, + 0.0032224326860159636, + 0.0036965750623494387, + 0.006907025817781687, + 0.006182867102324963, + 0.005676301196217537, + 0.01028332021087408, + 0.009218135848641396, + 0.00826950091868639, + 0.010934568010270596, + 0.0056972200982272625, + 0.004861647728830576, + 0.009623992256820202, + 0.005855975206941366, + 0.007737623993307352, + 0.01807091198861599, + 0.014110448770225048, + 0.024314163252711296, + 0.01441306620836258, + 0.017293622717261314, + 0.018021324649453163, + 0.013775864616036415, + 0.01776033081114292, + 0.007858803495764732, + 0.01033747661858797, + 0.006080951541662216, + 0.005817946512252092, + 0.006417479366064072, + 0.005306784994900227, + 0.004135014954954386, + 0.005628128536045551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.10160303115844727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216454699635506 + ], + "opposite_logprob_tokens": [ + -4.264216423034668 + ], + "logprob_reward": -0.014216454699635506, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216454699635506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08400581777095795, + 0.05347438529133797, + 0.01543422881513834, + 0.04422859102487564, + 0.15987035632133484, + 0.059362687170505524, + 0.2318953424692154, + 0.20252837240695953, + 0.10160315036773682, + 0.047597069293260574 + ], + "credit_map": [ + 0.01912895403802395, + 0.007640848867595196, + 0.006175424437969923, + 0.0124835055321455, + 0.006354753393679857, + 0.007643483113497496, + 0.009157032705843449, + 0.015421819873154163, + 0.011531021445989609, + 0.00407556863501668, + 0.010266910307109356, + 0.003335045650601387, + 0.0036851235199719667, + 0.007344127632677555, + 0.004734782502055168, + 0.008501805365085602, + 0.00271578854881227, + 0.002428789623081684, + 0.0019408356165513396, + 0.001736353151500225, + 0.001969266450032592, + 0.0018267974955961108, + 0.0012733257608488202, + 0.0015430719358846545, + 0.007000119425356388, + 0.006160995457321405, + 0.008106742054224014, + 0.005563400220125914, + 0.003242833772674203, + 0.005059537477791309, + 0.004638759884983301, + 0.004456206690520048, + 0.01968136988580227, + 0.032640498131513596, + 0.02693833038210869, + 0.02690647542476654, + 0.015903454273939133, + 0.011818310245871544, + 0.016957849264144897, + 0.009024077095091343, + 0.008378059603273869, + 0.008773085661232471, + 0.004665272310376167, + 0.0054697077721357346, + 0.006150470115244389, + 0.010210619308054447, + 0.0062081594951450825, + 0.009507311508059502, + 0.08717348426580429, + 0.05588927119970322, + 0.007851054891943932, + 0.0053825960494577885, + 0.011262628249824047, + 0.018178511410951614, + 0.013697503134608269, + 0.0324602834880352, + 0.02583291567862034, + 0.021493112668395042, + 0.03414497151970863, + 0.014553078450262547, + 0.02126573957502842, + 0.03441968187689781, + 0.02511749230325222, + 0.025701379403471947, + 0.012826837599277496, + 0.011134997010231018, + 0.013090495020151138, + 0.008748802356421947, + 0.02127828076481819, + 0.013408062048256397, + 0.008080449886620045, + 0.013035224750638008, + 0.007485274691134691, + 0.00561748631298542, + 0.006202769465744495, + 0.005964115262031555, + 0.0060839117504656315, + 0.006728947628289461, + 0.004094627685844898, + 0.005419934634119272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.014216454699635506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255929194390774 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255929194390774, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255929194390774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12818218767642975, + 0.10429766774177551, + 0.06076245754957199, + 0.11689085513353348, + 0.15340371429920197, + 0.13115715980529785, + 0.06291408836841583, + 0.07968860119581223, + 0.08749237656593323, + 0.07521092146635056 + ], + "credit_map": [ + 0.02784847468137741, + 0.012428954243659973, + 0.011047294363379478, + 0.015948668122291565, + 0.009978977032005787, + 0.014283361844718456, + 0.011180068366229534, + 0.02546638622879982, + 0.011523643508553505, + 0.007779316045343876, + 0.027207275852560997, + 0.01143742073327303, + 0.007920220494270325, + 0.021987376734614372, + 0.0072596450336277485, + 0.009182765148580074, + 0.009737101383507252, + 0.01515086181461811, + 0.005998627282679081, + 0.008303064852952957, + 0.007211403921246529, + 0.005533201154321432, + 0.005423096939921379, + 0.0034050976391881704, + 0.018317123875021935, + 0.016561713069677353, + 0.022389981895685196, + 0.014304623007774353, + 0.008491430431604385, + 0.012010915204882622, + 0.013208601623773575, + 0.011606466956436634, + 0.01846993900835514, + 0.02470257133245468, + 0.02519453875720501, + 0.026365453377366066, + 0.0138181047514081, + 0.014798522926867008, + 0.02014409378170967, + 0.009910485707223415, + 0.01691872626543045, + 0.0178486630320549, + 0.009221876040101051, + 0.01272872556000948, + 0.013108540326356888, + 0.024879461154341698, + 0.014840281568467617, + 0.02161090075969696, + 0.009265453554689884, + 0.011196842417120934, + 0.004336754325777292, + 0.004312814213335514, + 0.007266915403306484, + 0.011634535156190395, + 0.006244382355362177, + 0.008656390011310577, + 0.012881394475698471, + 0.013941271230578423, + 0.010926741175353527, + 0.004656807519495487, + 0.007589777000248432, + 0.01007212046533823, + 0.0071785650216042995, + 0.012441925704479218, + 0.014012163504958153, + 0.008648303337395191, + 0.015299288555979729, + 0.00962794478982687, + 0.010095708072185516, + 0.009767048060894012, + 0.008531753905117512, + 0.01151016540825367, + 0.010397697798907757, + 0.00893464032560587, + 0.013432551175355911, + 0.00885845348238945, + 0.009701641276478767, + 0.01125464029610157, + 0.006364674307405949, + 0.0062666223384439945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.005255929194390774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187190681695938 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.014187190681695938, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187190681695938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08903535455465317, + 0.15507100522518158, + 0.037967439740896225, + 0.0674465149641037, + 0.05753371864557266, + 0.07881554961204529, + 0.16070084273815155, + 0.16250298917293549, + 0.08871661871671677, + 0.10220997035503387 + ], + "credit_map": [ + 0.016611577942967415, + 0.008080745115876198, + 0.007480528671294451, + 0.00800943560898304, + 0.007531007751822472, + 0.016082655638456345, + 0.012272452004253864, + 0.012966951355338097, + 0.035073649138212204, + 0.011805259622633457, + 0.02811550348997116, + 0.013205980882048607, + 0.011989187449216843, + 0.026632336899638176, + 0.011675057001411915, + 0.016574028879404068, + 0.005981458816677332, + 0.004645929671823978, + 0.004541575442999601, + 0.005829119123518467, + 0.005374337546527386, + 0.0039607323706150055, + 0.004684581886976957, + 0.0029497090727090836, + 0.007155817933380604, + 0.007174700498580933, + 0.010863288305699825, + 0.008432330563664436, + 0.006812738720327616, + 0.011588557623326778, + 0.008774596266448498, + 0.006644485052675009, + 0.004932193551212549, + 0.011675306595861912, + 0.008415523916482925, + 0.009845294989645481, + 0.004476969130337238, + 0.004497256595641375, + 0.008378765545785427, + 0.005312408320605755, + 0.009925208985805511, + 0.009238148108124733, + 0.007442038506269455, + 0.010275558568537235, + 0.007673123851418495, + 0.011351708322763443, + 0.010588178411126137, + 0.012321585789322853, + 0.05529451370239258, + 0.03803939372301102, + 0.006315781269222498, + 0.0050683217123150826, + 0.014481953345239162, + 0.01619230955839157, + 0.009022263810038567, + 0.016286306083202362, + 0.020001361146569252, + 0.019006511196494102, + 0.02887823060154915, + 0.010734153911471367, + 0.01749231666326523, + 0.03138642758131027, + 0.0168172437697649, + 0.018186746165156364, + 0.017052069306373596, + 0.007860166020691395, + 0.010130846872925758, + 0.008076196536421776, + 0.015143577940762043, + 0.010876915417611599, + 0.008754471316933632, + 0.01082236785441637, + 0.012161731719970703, + 0.016885703429579735, + 0.012621172703802586, + 0.013460158370435238, + 0.00984580721706152, + 0.013874649070203304, + 0.008816969580948353, + 0.014543785713613033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.014187190681695938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, liquid-metal silver sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place on a multi-level elevated overpass with carbon and transparent forcefield guardrails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's taillights appear as blazing red LED trails with a subtle particle effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle behind the car, ascend, zoom out, and finally rise to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic city at night with holographic advertisements and a cold cyberpunk blue and purple color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624690420925617 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624690420925617, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624690420925617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0782487541437149, + 0.05524223670363426, + 0.020644789561629295, + 0.08562473207712173, + 0.13031843304634094, + 0.29746031761169434, + 0.08500345051288605, + 0.05607392266392708, + 0.12986846268177032, + 0.06151486188173294 + ], + "credit_map": [ + 0.011029349640011787, + 0.005872892681509256, + 0.009298140183091164, + 0.008912693709135056, + 0.007370396982878447, + 0.009495123289525509, + 0.015230517834424973, + 0.011039639823138714, + 0.013232712633907795, + 0.005556121468544006, + 0.00830563623458147, + 0.004837975837290287, + 0.004344501066952944, + 0.009328353218734264, + 0.004558810498565435, + 0.0050781285390257835, + 0.0029850963037461042, + 0.003626391990110278, + 0.003170349868014455, + 0.002503237221390009, + 0.0029833735898137093, + 0.00177078100387007, + 0.0020310350228101015, + 0.0015745279379189014, + 0.016256581991910934, + 0.012504578568041325, + 0.013078849762678146, + 0.010789671912789345, + 0.007316240109503269, + 0.012896224856376648, + 0.0069546205922961235, + 0.005827971268445253, + 0.01800687238574028, + 0.022570669651031494, + 0.023505188524723053, + 0.020925646647810936, + 0.012196965515613556, + 0.010225299745798111, + 0.014024904929101467, + 0.008862907998263836, + 0.052943628281354904, + 0.05290444940328598, + 0.02428358793258667, + 0.02287529595196247, + 0.027490515261888504, + 0.04081651195883751, + 0.031222989782691002, + 0.04492335394024849, + 0.02268596552312374, + 0.013447398319840431, + 0.003399182576686144, + 0.004078746307641268, + 0.009871172718703747, + 0.00787653960287571, + 0.007660159841179848, + 0.01598428748548031, + 0.008141780272126198, + 0.0060863071121275425, + 0.009579604491591454, + 0.0042794253677129745, + 0.007247328758239746, + 0.008964519016444683, + 0.005294820759445429, + 0.0064801438711583614, + 0.023605206981301308, + 0.013570998795330524, + 0.01760001666843891, + 0.009594923816621304, + 0.015898730605840683, + 0.024101482704281807, + 0.009670580737292767, + 0.01582653634250164, + 0.00799848698079586, + 0.011882293969392776, + 0.009144201874732971, + 0.006731088273227215, + 0.007279608864337206, + 0.007737253326922655, + 0.004410665947943926, + 0.006331268232315779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.007624690420925617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_tensor([744, 744])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_tensor([744, 744])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..789b3bfa09169f4e14e3cca0577f918a9d86ab2d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_tensor([744, 744])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward tilt begins at a very low angle, capturing the sturdy lower body of a woman, perfectly balanced in a difficult handstand yoga pose. Her hands, fingers spread wide, rest firmly on the sunlit, light-polished wooden floor, the muscular contours of her arms defined and graceful. As the camera smoothly and steadily moves upward, the image gradually reveals her signature form-fitting tracksuit: a vibrant red top with simple, pure white stripes along the sides and arms. The highly stretchy fabric clings to her powerful curves. The background is a minimalist room, where large floor-to-ceiling windows slice the warm afternoon sunlight into brilliant beams, casting them onto the floor and her body, creating a stark contrast of light and shadow. Even tiny particles of dust can be seen dancing peacefully in the air. The camera continued to rise, passing over her black hair tied into a neat high ponytail, the hair glowing softly in the backlight, and finally, the camera slowly fixed on her calm and focused face, her eyes lightly closed, her facial muscles completely relaxed, with a faint smile of satisfaction at the corners of her mouth, and a few tiny beads of sweat oozing from her forehead crystal clear in the sunlight. The whole picture was filled with the Zen spirit of harmonious coexistence of strength and softness, exuding an atmosphere of extreme concentration and peace.", + "seed": 1535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_744_seed_1535_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03297764426097274, + "vlm_logprob": -0.03297764426097274, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03297764426097274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09285765886306763 + ], + "opposite_logprob_tokens": [ + -2.592857599258423 + ], + "logprob_reward": -0.09285765886306763, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09285765886306763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19229556620121002, + 0.09674422442913055, + 0.0430343896150589, + 0.030445849522948265, + 0.050484560430049896, + 0.11563453823328018, + 0.07703341543674469, + 0.06622916460037231, + 0.26474419236183167, + 0.06335413455963135 + ], + "credit_map": [ + 0.03393765538930893, + 0.015703219920396805, + 0.030283765867352486, + 0.039018500596284866, + 0.012574078515172005, + 0.02299867384135723, + 0.01154076773673296, + 0.02623889409005642, + 0.011182629503309727, + 0.009411776438355446, + 0.012181454338133335, + 0.011892717331647873, + 0.010997996665537357, + 0.02137068659067154, + 0.005797267891466618, + 0.013909687288105488, + 0.004137457814067602, + 0.004423410631716251, + 0.005474255885928869, + 0.006133376620709896, + 0.0068626790307462215, + 0.006452555768191814, + 0.003287650179117918, + 0.006262999027967453, + 0.004831917118281126, + 0.0039908564649522305, + 0.003931112587451935, + 0.005386293865740299, + 0.0022508164402097464, + 0.0027833296917378902, + 0.002512121805921197, + 0.004759399686008692, + 0.009496049024164677, + 0.010466214269399643, + 0.003785956185311079, + 0.0054907179437577724, + 0.004019905813038349, + 0.005607760511338711, + 0.0045182653702795506, + 0.007099688984453678, + 0.0068983943201601505, + 0.004372383002191782, + 0.04305719956755638, + 0.03659873455762863, + 0.004205719102174044, + 0.007766402326524258, + 0.006419805809855461, + 0.006315891165286303, + 0.0053789918310940266, + 0.009563738480210304, + 0.00821657944470644, + 0.011284120380878448, + 0.013964482583105564, + 0.015620134770870209, + 0.007235047873109579, + 0.005770313553512096, + 0.009445127099752426, + 0.008045895956456661, + 0.013929935172200203, + 0.01204051822423935, + 0.0068296329118311405, + 0.0045357802882790565, + 0.007751543540507555, + 0.0036507260520011187, + 0.08261063694953918, + 0.07183723896741867, + 0.032911933958530426, + 0.02030806988477707, + 0.024560697376728058, + 0.007716186810284853, + 0.016065523028373718, + 0.008733893744647503, + 0.00869568157941103, + 0.010931185446679592, + 0.008055589161813259, + 0.011112743988633156, + 0.007347107864916325, + 0.005865412298589945, + 0.006684137508273125, + 0.004662276711314917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.09285765886306763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076541476882994175 + ], + "opposite_logprob_tokens": [ + -4.882654190063477 + ], + "logprob_reward": -0.0076541476882994175, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076541476882994175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11983302235603333, + 0.07964204996824265, + 0.06084762513637543, + 0.08233713358640671, + 0.09299405664205551, + 0.04237202927470207, + 0.03519558161497116, + 0.0518551729619503, + 0.35588711500167847, + 0.07903619855642319 + ], + "credit_map": [ + 0.01781301014125347, + 0.01326213963329792, + 0.018770894035696983, + 0.0267439354211092, + 0.0062306043691933155, + 0.01388898678123951, + 0.008412874303758144, + 0.01471057627350092, + 0.010480069555342197, + 0.009776185266673565, + 0.010485837236046791, + 0.01348726637661457, + 0.007892457768321037, + 0.013037468306720257, + 0.005306183826178312, + 0.009176590479910374, + 0.004789764527231455, + 0.004792341962456703, + 0.015332668088376522, + 0.00637222733348608, + 0.006999426521360874, + 0.014764634892344475, + 0.0027796754147857428, + 0.005016888491809368, + 0.013117119669914246, + 0.013909225352108479, + 0.01034011971205473, + 0.01147378608584404, + 0.005011334549635649, + 0.011248537339270115, + 0.006942081265151501, + 0.010294931009411812, + 0.01365916058421135, + 0.01757597178220749, + 0.006709365174174309, + 0.010308588854968548, + 0.009052544832229614, + 0.0166988056153059, + 0.007891685701906681, + 0.011097930371761322, + 0.0026385863311588764, + 0.0023017663042992353, + 0.01593439094722271, + 0.013154075480997562, + 0.0016753203235566616, + 0.0029832252766937017, + 0.0017078794771805406, + 0.0019767875783145428, + 0.002357808407396078, + 0.004156137816607952, + 0.003325479105114937, + 0.004245026968419552, + 0.006996425800025463, + 0.008316729217767715, + 0.0031723163556307554, + 0.0026256563141942024, + 0.007786229718476534, + 0.005497993901371956, + 0.011167670600116253, + 0.011244316585361958, + 0.005344695411622524, + 0.0031296012457460165, + 0.00581028126180172, + 0.0018743863329291344, + 0.11763318628072739, + 0.10276281833648682, + 0.039177197962999344, + 0.023008622229099274, + 0.029849926009774208, + 0.009354358538985252, + 0.022323517128825188, + 0.011777481064200401, + 0.008843088522553444, + 0.00787084735929966, + 0.008364854380488396, + 0.016312383115291595, + 0.008393321186304092, + 0.004595980513840914, + 0.01916014961898327, + 0.005495570600032806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0076541476882994175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063171207904816 + ], + "opposite_logprob_tokens": [ + -3.168063163757324 + ], + "logprob_reward": -0.043063171207904816, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063171207904816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08733008056879044, + 0.036551788449287415, + 0.032704032957553864, + 0.027302416041493416, + 0.11022146791219711, + 0.029913613572716713, + 0.02023378759622574, + 0.02807488478720188, + 0.28446340560913086, + 0.3432045578956604 + ], + "credit_map": [ + 0.01475138682872057, + 0.007369345054030418, + 0.01349355187267065, + 0.017147457227110863, + 0.005859212018549442, + 0.009354432113468647, + 0.007132497150450945, + 0.012222200632095337, + 0.00484326109290123, + 0.00356224924325943, + 0.005597221199423075, + 0.007872761227190495, + 0.0027397803496569395, + 0.006109965965151787, + 0.0020965090952813625, + 0.0037300409749150276, + 0.0027006124146282673, + 0.002722588600590825, + 0.005056655500084162, + 0.004466527607291937, + 0.004450688604265451, + 0.007283080369234085, + 0.002053858246654272, + 0.003970020450651646, + 0.0035396863240748644, + 0.0031806088518351316, + 0.004875259008258581, + 0.004624500870704651, + 0.0015944789629429579, + 0.0031167755369096994, + 0.0017120252596214414, + 0.004659079946577549, + 0.016448987647891045, + 0.02406664378941059, + 0.0072902170941233635, + 0.009216026403009892, + 0.011600212194025517, + 0.021496115252375603, + 0.005836493335664272, + 0.014266774989664555, + 0.002157863462343812, + 0.0013539829524233937, + 0.010550658218562603, + 0.008600142784416676, + 0.0015074986731633544, + 0.00238406821154058, + 0.0016941871726885438, + 0.001665212563239038, + 0.0013241375563666224, + 0.0027449342887848616, + 0.002147249411791563, + 0.0024144183844327927, + 0.0033980240114033222, + 0.0042874617502093315, + 0.0018402296118438244, + 0.0020773308351635933, + 0.0030725114047527313, + 0.002731921151280403, + 0.005388479679822922, + 0.004810391925275326, + 0.0042667570523917675, + 0.0021996002178639174, + 0.00397131685167551, + 0.0016339060384780169, + 0.11327921599149704, + 0.09936325252056122, + 0.01694594882428646, + 0.011445698328316212, + 0.016540896147489548, + 0.006111601367592812, + 0.011952202767133713, + 0.008824601769447327, + 0.06584694981575012, + 0.03481178730726242, + 0.045080848038196564, + 0.06286312639713287, + 0.03365901857614517, + 0.017971472814679146, + 0.06601457297801971, + 0.016956759616732597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.043063171207904816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253438837826252 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253438837826252, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253438837826252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08450702577829361, + 0.06519700586795807, + 0.07011841982603073, + 0.09622092545032501, + 0.0710933580994606, + 0.17273028194904327, + 0.10102326422929764, + 0.05393175408244133, + 0.17400862276554108, + 0.11116936802864075 + ], + "credit_map": [ + 0.0097424928098917, + 0.009095922112464905, + 0.015440366230905056, + 0.021035123616456985, + 0.00716392369940877, + 0.009424361400306225, + 0.006282773334532976, + 0.006322057917714119, + 0.00881884340196848, + 0.007156627252697945, + 0.009233682416379452, + 0.011269818991422653, + 0.007230415940284729, + 0.012879771180450916, + 0.0037423132453113794, + 0.0048655252903699875, + 0.00643168156966567, + 0.005062793847173452, + 0.017611606046557426, + 0.005720164161175489, + 0.006521148141473532, + 0.017516091465950012, + 0.0029257668647915125, + 0.008329167030751705, + 0.011869529262185097, + 0.011358635500073433, + 0.015277707017958164, + 0.013344526290893555, + 0.006112218368798494, + 0.012445440515875816, + 0.00732930563390255, + 0.01848354935646057, + 0.010944083333015442, + 0.011081047356128693, + 0.0073082176968455315, + 0.007845492102205753, + 0.005681541282683611, + 0.012212902307510376, + 0.00697045773267746, + 0.009049610234797001, + 0.009453949518501759, + 0.00693147536367178, + 0.068987175822258, + 0.058825746178627014, + 0.005476863123476505, + 0.011026128195226192, + 0.005795671604573727, + 0.00623326376080513, + 0.006900354754179716, + 0.01223142072558403, + 0.010319625958800316, + 0.01589455083012581, + 0.018856434151530266, + 0.019790535792708397, + 0.010259084403514862, + 0.006771244574338198, + 0.007377961650490761, + 0.0074657658115029335, + 0.007314115762710571, + 0.005873272195458412, + 0.010444989427924156, + 0.0035852408036589622, + 0.008193034678697586, + 0.0036773669999092817, + 0.061690229922533035, + 0.058036766946315765, + 0.010421630926430225, + 0.008979087695479393, + 0.013795825652778149, + 0.005074783228337765, + 0.009727132506668568, + 0.00628315843641758, + 0.0186446662992239, + 0.013140896335244179, + 0.013254713267087936, + 0.017681635916233063, + 0.011840333230793476, + 0.009227979928255081, + 0.01991608738899231, + 0.007463047280907631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.005253438837826252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605980470776558 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.01605980470776558, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605980470776558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05453598499298096, + 0.044786300510168076, + 0.04682938754558563, + 0.0234841201454401, + 0.03520877659320831, + 0.13289181888103485, + 0.05915788933634758, + 0.06326903402805328, + 0.37528130412101746, + 0.16455544531345367 + ], + "credit_map": [ + 0.007734872866421938, + 0.0057308366522192955, + 0.010513558983802795, + 0.013996444642543793, + 0.003272942267358303, + 0.004824378993362188, + 0.004015584476292133, + 0.0044473642483353615, + 0.005430117249488831, + 0.004812503699213266, + 0.005858802702277899, + 0.007729485630989075, + 0.004780164919793606, + 0.008338904939591885, + 0.002861789893358946, + 0.004974530078470707, + 0.005687819328159094, + 0.004619336221367121, + 0.007658624090254307, + 0.005694778170436621, + 0.005097533110529184, + 0.008561205118894577, + 0.003543669590726495, + 0.005966415628790855, + 0.004057875368744135, + 0.003142711240798235, + 0.003017640672624111, + 0.003599243937060237, + 0.0014452607138082385, + 0.0024909554049372673, + 0.0017131264321506023, + 0.004017304629087448, + 0.004090579226613045, + 0.005421845708042383, + 0.003036120440810919, + 0.003073064610362053, + 0.0036490317434072495, + 0.007065752521157265, + 0.0031438914593309164, + 0.005728490650653839, + 0.007724544499069452, + 0.003349565900862217, + 0.05342213064432144, + 0.04435673728585243, + 0.004313590005040169, + 0.0096651092171669, + 0.005011144559830427, + 0.005048984661698341, + 0.004227550700306892, + 0.007219769060611725, + 0.007650002837181091, + 0.007567753549665213, + 0.008327769115567207, + 0.007166875060647726, + 0.00711397547274828, + 0.009884189814329147, + 0.007699564564973116, + 0.006183958146721125, + 0.013792765326797962, + 0.01394535694271326, + 0.007083155680447817, + 0.004072079434990883, + 0.0073758745566010475, + 0.0031162751838564873, + 0.13268285989761353, + 0.11744245886802673, + 0.033974580466747284, + 0.023648444563150406, + 0.030022654682397842, + 0.00929245539009571, + 0.016554677858948708, + 0.011663133278489113, + 0.02953294664621353, + 0.025894630700349808, + 0.021928809583187103, + 0.030951235443353653, + 0.015482679009437561, + 0.010258331894874573, + 0.021056728437542915, + 0.009450075216591358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.01605980470776558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward tilt begins at a very low angle, capturing the sturdy lower body of a woman, perfectly balanced in a difficult handstand yoga pose. Her hands, fingers spread wide, rest firmly on the sunlit, light-polished wooden floor, the muscular contours of her arms defined and graceful. As the camera smoothly and steadily moves upward, the image gradually reveals her signature form-fitting tracksuit: a vibrant red top with simple, pure white stripes along the sides and arms. The highly stretchy fabric clings to her powerful curves. The background is a minimalist room, where large floor-to-ceiling windows slice the warm afternoon sunlight into brilliant beams, casting them onto the floor and her body, creating a stark contrast of light and shadow. Even tiny particles of dust can be seen dancing peacefully in the air. The camera continued to rise, passing over her black hair tied into a neat high ponytail, the hair glowing softly in the backlight, and finally, the camera slowly fixed on her calm and focused face, her eyes lightly closed, her facial muscles completely relaxed, with a faint smile of satisfaction at the corners of her mouth, and a few tiny beads of sweat oozing from her forehead crystal clear in the sunlight. The whole picture was filled with the Zen spirit of harmonious coexistence of strength and softness, exuding an atmosphere of extreme concentration and peace.", + "seed": 1536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_744_seed_1536_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029201060300692916, + "vlm_logprob": -0.029201060300692916, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029201060300692916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08981606364250183 + ], + "opposite_logprob_tokens": [ + -2.589816093444824 + ], + "logprob_reward": -0.08981606364250183, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08981606364250183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20660577714443207, + 0.10269507020711899, + 0.267612487077713, + 0.06685913354158401, + 0.02201964519917965, + 0.04168141633272171, + 0.05712699145078659, + 0.11601090431213379, + 0.06322882324457169, + 0.05615980178117752 + ], + "credit_map": [ + 0.020396390929818153, + 0.01692776195704937, + 0.023558152839541435, + 0.021788133308291435, + 0.04217744991183281, + 0.017420168966054916, + 0.024646779522299767, + 0.03969092294573784, + 0.007923811674118042, + 0.00948932021856308, + 0.011623711325228214, + 0.00848189927637577, + 0.023084159940481186, + 0.009831602685153484, + 0.01070526521652937, + 0.02155529335141182, + 0.02956492453813553, + 0.017719490453600883, + 0.03508935868740082, + 0.03925604745745659, + 0.03444136306643486, + 0.01692488230764866, + 0.05169094726443291, + 0.04292546212673187, + 0.004309795796871185, + 0.005257778335362673, + 0.009766962379217148, + 0.009197733364999294, + 0.0045362478122115135, + 0.005315755028277636, + 0.01346740871667862, + 0.015007449313998222, + 0.002287806011736393, + 0.005216475576162338, + 0.0014640821609646082, + 0.0033203826751559973, + 0.0021573989652097225, + 0.001411736593581736, + 0.003078552894294262, + 0.003083206946030259, + 0.0035324953496456146, + 0.00736633874475956, + 0.003935656044632196, + 0.0052212439477443695, + 0.006604666355997324, + 0.005266088992357254, + 0.004775799810886383, + 0.004979120101779699, + 0.003467188449576497, + 0.010394798591732979, + 0.004648595117032528, + 0.007734899874776602, + 0.01091516762971878, + 0.007660095114260912, + 0.007103848271071911, + 0.005202393978834152, + 0.006025575567036867, + 0.034842729568481445, + 0.00729885371401906, + 0.010887053795158863, + 0.012297137640416622, + 0.007032263092696667, + 0.010889890603721142, + 0.026737390086054802, + 0.0036805281415581703, + 0.013959664851427078, + 0.0034101696219295263, + 0.011671766638755798, + 0.008408240973949432, + 0.004090361762791872, + 0.01278331782668829, + 0.005224769003689289, + 0.004458353389054537, + 0.008208302780985832, + 0.0047894977033138275, + 0.01050261128693819, + 0.0057239350862801075, + 0.004931318107992411, + 0.007614326197654009, + 0.009931448847055435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.08981606364250183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565961107611656 + ], + "opposite_logprob_tokens": [ + -4.387566089630127 + ], + "logprob_reward": -0.012565961107611656, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565961107611656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2935074269771576, + 0.09527676552534103, + 0.08787816017866135, + 0.06393343210220337, + 0.09756874293088913, + 0.07313890755176544, + 0.06368464231491089, + 0.08490289747714996, + 0.06293762475252151, + 0.07717136293649673 + ], + "credit_map": [ + 0.019488925114274025, + 0.016189629212021828, + 0.03067263960838318, + 0.02266545221209526, + 0.07636716961860657, + 0.022120477631688118, + 0.02897200919687748, + 0.07703114300966263, + 0.007321713026612997, + 0.011308732442557812, + 0.01255902647972107, + 0.010319801978766918, + 0.019245387986302376, + 0.01025170087814331, + 0.011292516253888607, + 0.012977887876331806, + 0.008685894310474396, + 0.007025363855063915, + 0.01114744320511818, + 0.012224067933857441, + 0.01067607942968607, + 0.009179544635117054, + 0.018008865416049957, + 0.010930904187262058, + 0.004280810710042715, + 0.005563370417803526, + 0.006873562466353178, + 0.010677014477550983, + 0.008744031190872192, + 0.0067820255644619465, + 0.011017254553735256, + 0.009995362721383572, + 0.007766662165522575, + 0.009790090844035149, + 0.0063486285507678986, + 0.010162660852074623, + 0.011294027790427208, + 0.009856801480054855, + 0.01645742729306221, + 0.025892438367009163, + 0.006003252696245909, + 0.011940118856728077, + 0.006659726146608591, + 0.010774929076433182, + 0.011121761985123158, + 0.01264522597193718, + 0.008143287152051926, + 0.005850605666637421, + 0.004622031003236771, + 0.008361918851733208, + 0.004356564022600651, + 0.010588021017611027, + 0.012578807771205902, + 0.010789130814373493, + 0.006398569326847792, + 0.005989601835608482, + 0.005011043976992369, + 0.014330817386507988, + 0.005687806289643049, + 0.008978002704679966, + 0.01736716739833355, + 0.011781357228755951, + 0.01294122077524662, + 0.008805474266409874, + 0.005091729108244181, + 0.010645417496562004, + 0.0039967759512364864, + 0.007746082730591297, + 0.006794946733862162, + 0.010036434046924114, + 0.013092138804495335, + 0.005534100811928511, + 0.008999147452414036, + 0.010914328508079052, + 0.005792059935629368, + 0.016600869596004486, + 0.007152264937758446, + 0.007062183693051338, + 0.008241209201514721, + 0.01240930613130331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.012565961107611656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699020743370056 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.033699020743370056, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699020743370056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26773783564567566, + 0.13157524168491364, + 0.11250052601099014, + 0.053665999323129654, + 0.1955588459968567, + 0.037929002195596695, + 0.07834376394748688, + 0.06329523772001266, + 0.029581381008028984, + 0.02981226146221161 + ], + "credit_map": [ + 0.04170706123113632, + 0.028844216838479042, + 0.037154097110033035, + 0.04195404052734375, + 0.03213127329945564, + 0.028322534635663033, + 0.03576511889696121, + 0.02185947261750698, + 0.010303154587745667, + 0.0056780255399644375, + 0.01339567732065916, + 0.007940578274428844, + 0.040058355778455734, + 0.009211741387844086, + 0.014180110767483711, + 0.03080756962299347, + 0.01069559808820486, + 0.005743632558733225, + 0.012310590595006943, + 0.012288582511246204, + 0.01809980347752571, + 0.008476543240249157, + 0.022840525954961777, + 0.022045236080884933, + 0.00632076570764184, + 0.004877452272921801, + 0.009659204632043839, + 0.00879315473139286, + 0.004406429361552, + 0.006141712423413992, + 0.006946221925318241, + 0.006521057337522507, + 0.014236331917345524, + 0.01736367493867874, + 0.013474879786372185, + 0.019756313413381577, + 0.02386331371963024, + 0.019009960815310478, + 0.03856407850980759, + 0.04929026961326599, + 0.003093200735747814, + 0.008871163241565228, + 0.003211478004232049, + 0.004607000853866339, + 0.005864100996404886, + 0.005851808935403824, + 0.004048739094287157, + 0.0023815042804926634, + 0.003244335763156414, + 0.01750664971768856, + 0.005773242097347975, + 0.013125925324857235, + 0.013874348253011703, + 0.012792984023690224, + 0.006294906605035067, + 0.005731360986828804, + 0.003677976783365011, + 0.01729768142104149, + 0.004385438282042742, + 0.0065896701999008656, + 0.008276354521512985, + 0.004883021116256714, + 0.005238837096840143, + 0.012946255505084991, + 0.0022950544953346252, + 0.00620653061196208, + 0.0016688451869413257, + 0.003619682276621461, + 0.004282316658645868, + 0.002692998154088855, + 0.005203405860811472, + 0.0036125441547483206, + 0.00296576926484704, + 0.005446422845125198, + 0.0026861487422138453, + 0.006930620409548283, + 0.00324814859777689, + 0.002700125565752387, + 0.003122151829302311, + 0.0027128723450005054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.033699020743370056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192092990502715 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003192092990502715, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192092990502715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2862405478954315, + 0.13570819795131683, + 0.0935181975364685, + 0.06901372969150543, + 0.04244426265358925, + 0.05757531523704529, + 0.06704764813184738, + 0.09091494232416153, + 0.09344516694545746, + 0.06409207731485367 + ], + "credit_map": [ + 0.016896236687898636, + 0.014192074537277222, + 0.022861208766698837, + 0.014774465002119541, + 0.09112440794706345, + 0.02098214253783226, + 0.023567453026771545, + 0.08184254914522171, + 0.009278020821511745, + 0.006864484399557114, + 0.013074232265353203, + 0.010294714011251926, + 0.037287674844264984, + 0.011763517744839191, + 0.016308730468153954, + 0.03083682432770729, + 0.012292073108255863, + 0.00396620761603117, + 0.012804119847714901, + 0.00986380036920309, + 0.011875840835273266, + 0.005471784621477127, + 0.01831742934882641, + 0.018926938995718956, + 0.00407536793500185, + 0.005986303556710482, + 0.007505797315388918, + 0.008999522775411606, + 0.005688162986189127, + 0.007426106370985508, + 0.012278909794986248, + 0.01705356128513813, + 0.0037902065087109804, + 0.004389114212244749, + 0.002868807874619961, + 0.00396746164187789, + 0.004910137504339218, + 0.004164565354585648, + 0.007782298605889082, + 0.010571667924523354, + 0.007249522488564253, + 0.009484641253948212, + 0.006440325174480677, + 0.006902001798152924, + 0.008573276922106743, + 0.007163444999605417, + 0.0058785188011825085, + 0.005883587524294853, + 0.004455676767975092, + 0.013508915901184082, + 0.0045431824401021, + 0.007486374583095312, + 0.01261207740753889, + 0.010111414827406406, + 0.005609642248600721, + 0.008720362558960915, + 0.00498722679913044, + 0.01813986524939537, + 0.0065809269435703754, + 0.008736502379179, + 0.014338929206132889, + 0.010034599341452122, + 0.01288341823965311, + 0.01521347463130951, + 0.005205146502703428, + 0.02477620728313923, + 0.004718744661659002, + 0.014602595008909702, + 0.009569907560944557, + 0.012041991576552391, + 0.013909948989748955, + 0.008620615117251873, + 0.004985090810805559, + 0.013014930300414562, + 0.0039468496106565, + 0.013264869339764118, + 0.009370272047817707, + 0.005295365583151579, + 0.0072307405062019825, + 0.0069839623756706715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.003192092990502715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732163019478321 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732163019478321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732163019478321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31394869089126587, + 0.08442952483892441, + 0.05551246553659439, + 0.06408920884132385, + 0.027768012136220932, + 0.0728721022605896, + 0.08808507025241852, + 0.11421200633049011, + 0.06962263584136963, + 0.10946024954319 + ], + "credit_map": [ + 0.029187751933932304, + 0.011218206956982613, + 0.020929327234625816, + 0.01989496313035488, + 0.10753114521503448, + 0.018416348844766617, + 0.012043903581798077, + 0.0947270616889, + 0.00830376148223877, + 0.0045326873660087585, + 0.005605810321867466, + 0.004854009486734867, + 0.02857998199760914, + 0.007029947359114885, + 0.006361191160976887, + 0.019162144511938095, + 0.007161187939345837, + 0.003021067474037409, + 0.007665726821869612, + 0.005627046339213848, + 0.008045264519751072, + 0.004865855444222689, + 0.01115393452346325, + 0.007972385734319687, + 0.0028258678503334522, + 0.0041411384008824825, + 0.008902381174266338, + 0.00744730606675148, + 0.0053462497889995575, + 0.005635702516883612, + 0.016954217106103897, + 0.012836351990699768, + 0.0025889906100928783, + 0.0042404490523040295, + 0.0022766455076634884, + 0.0020223434548825026, + 0.003808477893471718, + 0.003115925006568432, + 0.006121756508946419, + 0.0035934278275817633, + 0.0063021015375852585, + 0.01430295966565609, + 0.005667403806000948, + 0.008245355449616909, + 0.013070524670183659, + 0.011365019716322422, + 0.00838344544172287, + 0.005535298027098179, + 0.004130104091018438, + 0.018002623692154884, + 0.006423840764909983, + 0.020692747086286545, + 0.011825548484921455, + 0.014025535434484482, + 0.0070275552570819855, + 0.005957125220447779, + 0.007697620429098606, + 0.026907818391919136, + 0.008945256471633911, + 0.012590156868100166, + 0.01503722369670868, + 0.00865089613944292, + 0.013260718435049057, + 0.021122321486473083, + 0.005953120533376932, + 0.016843337565660477, + 0.0054071880877017975, + 0.009130755439400673, + 0.008685007691383362, + 0.0049178884364664555, + 0.013520525768399239, + 0.005164816044270992, + 0.01262493897229433, + 0.016610946506261826, + 0.009949548169970512, + 0.02415318787097931, + 0.012897814624011517, + 0.009099269285798073, + 0.014038045890629292, + 0.010086513124406338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.006732163019478321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward tilt begins at a very low angle, capturing the sturdy lower body of a woman, perfectly balanced in a difficult handstand yoga pose. Her hands, fingers spread wide, rest firmly on the sunlit, light-polished wooden floor, the muscular contours of her arms defined and graceful. As the camera smoothly and steadily moves upward, the image gradually reveals her signature form-fitting tracksuit: a vibrant red top with simple, pure white stripes along the sides and arms. The highly stretchy fabric clings to her powerful curves. The background is a minimalist room, where large floor-to-ceiling windows slice the warm afternoon sunlight into brilliant beams, casting them onto the floor and her body, creating a stark contrast of light and shadow. Even tiny particles of dust can be seen dancing peacefully in the air. The camera continued to rise, passing over her black hair tied into a neat high ponytail, the hair glowing softly in the backlight, and finally, the camera slowly fixed on her calm and focused face, her eyes lightly closed, her facial muscles completely relaxed, with a faint smile of satisfaction at the corners of her mouth, and a few tiny beads of sweat oozing from her forehead crystal clear in the sunlight. The whole picture was filled with the Zen spirit of harmonious coexistence of strength and softness, exuding an atmosphere of extreme concentration and peace.", + "seed": 1537, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_744_seed_1537_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06193013682495803, + "vlm_logprob": -0.06193013682495803, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.06193013682495803, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26676708459854126 + ], + "opposite_logprob_tokens": [ + -1.5167670249938965 + ], + "logprob_reward": -0.26676708459854126, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26676708459854126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13118666410446167, + 0.014259581454098225, + 0.02576485648751259, + 0.06969819962978363, + 0.018183348700404167, + 0.019363684579730034, + 0.6162779927253723, + 0.05075374245643616, + 0.032354407012462616, + 0.022157520055770874 + ], + "credit_map": [ + 0.012104030698537827, + 0.013744545169174671, + 0.020761987194418907, + 0.021201089024543762, + 0.017955627292394638, + 0.012267986312508583, + 0.015008727088570595, + 0.018142659217119217, + 0.0015067424392327666, + 0.001841314835473895, + 0.0016159664373844862, + 0.0020985323935747147, + 0.0024015556555241346, + 0.0018222788348793983, + 0.0014773145085200667, + 0.0014958760002627969, + 0.002542078262194991, + 0.0027004724834114313, + 0.0023480562958866358, + 0.004406937398016453, + 0.0037242602556943893, + 0.0033596160355955362, + 0.003768227295950055, + 0.0029152079951018095, + 0.007194531615823507, + 0.009209876880049706, + 0.007656567730009556, + 0.012776012532413006, + 0.010439646430313587, + 0.006632008124142885, + 0.0077671147882938385, + 0.008022437803447247, + 0.001925937132909894, + 0.0016826250357553363, + 0.001788092078641057, + 0.0017251280369237065, + 0.0016421906184405088, + 0.0038996925577521324, + 0.0015648109838366508, + 0.00395487155765295, + 0.0009533148258924484, + 0.0029559514950960875, + 0.0016819600714370608, + 0.0038291304372251034, + 0.002173700835555792, + 0.0016014300053939223, + 0.004845392890274525, + 0.0013228035531938076, + 0.020005667582154274, + 0.030456310138106346, + 0.030915727838873863, + 0.21312865614891052, + 0.04139194265007973, + 0.011878778226673603, + 0.225829616189003, + 0.0426713228225708, + 0.005975848529487848, + 0.005596475210040808, + 0.005131640005856752, + 0.008565317839384079, + 0.011034292168915272, + 0.004069621674716473, + 0.007170731667429209, + 0.0032098188530653715, + 0.005583711434155703, + 0.0038277292624115944, + 0.008115512318909168, + 0.003282587043941021, + 0.0035139319952577353, + 0.0034413861576467752, + 0.0022728487383574247, + 0.002316696336492896, + 0.0025567880365997553, + 0.0015748043078929186, + 0.0029532024636864662, + 0.0032152580097317696, + 0.002954831114038825, + 0.0034165491815656424, + 0.003014857182279229, + 0.002471228828653693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.26676708459854126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610391214489937 + ], + "opposite_logprob_tokens": [ + -4.141103744506836 + ], + "logprob_reward": -0.01610391214489937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610391214489937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04548982158303261, + 0.04224683344364166, + 0.02537771686911583, + 0.2377125471830368, + 0.07006386667490005, + 0.038813721388578415, + 0.46709591150283813, + 0.013632466085255146, + 0.02616707608103752, + 0.033399999141693115 + ], + "credit_map": [ + 0.004662743769586086, + 0.006181108299642801, + 0.006617097184062004, + 0.007560208439826965, + 0.006058233790099621, + 0.004276573657989502, + 0.004564172122627497, + 0.005569683853536844, + 0.004078708589076996, + 0.00435584457591176, + 0.003909807186573744, + 0.00609973818063736, + 0.0075669423677027225, + 0.005241448991000652, + 0.006354474928230047, + 0.004639868624508381, + 0.0017323811771348119, + 0.0032117096707224846, + 0.0027397817466408014, + 0.004474492743611336, + 0.0034672822803258896, + 0.003334159031510353, + 0.0033092587254941463, + 0.003108652075752616, + 0.025523975491523743, + 0.04544374719262123, + 0.039540186524391174, + 0.031287405639886856, + 0.029497317969799042, + 0.021277351304888725, + 0.02238054946064949, + 0.022762026637792587, + 0.007964315824210644, + 0.007899222895503044, + 0.008377066813409328, + 0.009419155307114124, + 0.006468129809945822, + 0.0126854432746768, + 0.005413923412561417, + 0.011836604215204716, + 0.0024664506781846285, + 0.005357447545975447, + 0.00602154852822423, + 0.008937950246036053, + 0.003925682976841927, + 0.0026836220640689135, + 0.0064825150184333324, + 0.0029385036323219538, + 0.015155519358813763, + 0.023358479142189026, + 0.02321338839828968, + 0.15970607101917267, + 0.032236386090517044, + 0.008893029764294624, + 0.17072762548923492, + 0.033805422484874725, + 0.0013324610190466046, + 0.0011904758866876364, + 0.00155737460590899, + 0.0021345322020351887, + 0.002050315961241722, + 0.0023680173326283693, + 0.002042282372713089, + 0.0009570078109391034, + 0.004741237964481115, + 0.0027597479056566954, + 0.0069580078125, + 0.002599887317046523, + 0.002863762667402625, + 0.001662969822064042, + 0.002439951291307807, + 0.002141510369256139, + 0.005355046596378088, + 0.0022071257699280977, + 0.004367804154753685, + 0.005068065132945776, + 0.0032378793694078922, + 0.005297227296978235, + 0.0033943094313144684, + 0.004472540691494942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.01610391214489937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419447734951973 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.01419447734951973, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419447734951973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19055168330669403, + 0.04513195529580116, + 0.04268842935562134, + 0.11114021390676498, + 0.023408927023410797, + 0.07735473662614822, + 0.09381476789712906, + 0.032019998878240585, + 0.22458364069461823, + 0.1593056470155716 + ], + "credit_map": [ + 0.01914675161242485, + 0.020800290629267693, + 0.02493896335363388, + 0.023921120911836624, + 0.02738688699901104, + 0.0202013086527586, + 0.02890041284263134, + 0.025255931541323662, + 0.0036806929856538773, + 0.005880666431039572, + 0.007349998690187931, + 0.006887324620038271, + 0.006291304714977741, + 0.006365172099322081, + 0.004880357999354601, + 0.0037964340299367905, + 0.0037691958714276552, + 0.0067793941125273705, + 0.0032856811303645372, + 0.005116394255310297, + 0.006254218984395266, + 0.006149873603135347, + 0.007883134298026562, + 0.0034505329094827175, + 0.011120224371552467, + 0.01850046217441559, + 0.017249980941414833, + 0.014081567525863647, + 0.011877656914293766, + 0.01174970529973507, + 0.010934427380561829, + 0.01562618650496006, + 0.0023515240754932165, + 0.0030411335173994303, + 0.0036671573761850595, + 0.0033698526676744223, + 0.002682490972802043, + 0.0030769393779337406, + 0.0023115789517760277, + 0.002908247523009777, + 0.006054578348994255, + 0.00875070784240961, + 0.009377293288707733, + 0.011297948658466339, + 0.009288105182349682, + 0.011814042925834656, + 0.011760890483856201, + 0.0090111643075943, + 0.0050977664068341255, + 0.006647254806011915, + 0.008668389171361923, + 0.02737666293978691, + 0.007148130796849728, + 0.0029765535145998, + 0.02632036991417408, + 0.00957963615655899, + 0.004371055401861668, + 0.003866450395435095, + 0.005644517485052347, + 0.0039692590944468975, + 0.004714182112365961, + 0.003006365615874529, + 0.003669199999421835, + 0.002778966212645173, + 0.04546762630343437, + 0.02191302552819252, + 0.06540130078792572, + 0.022023074328899384, + 0.020966826006770134, + 0.014180834405124187, + 0.017909055575728416, + 0.016721881926059723, + 0.03335683420300484, + 0.009924789890646935, + 0.023288661614060402, + 0.01805701106786728, + 0.01277951616793871, + 0.036367274820804596, + 0.012421360239386559, + 0.013110199943184853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.01419447734951973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015142414486035705 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015142414486035705, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015142414486035705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08966359496116638, + 0.05501851439476013, + 0.07850467413663864, + 0.12869802117347717, + 0.10772062093019485, + 0.03101334162056446, + 0.2360101193189621, + 0.02245640754699707, + 0.03795774281024933, + 0.21295705437660217 + ], + "credit_map": [ + 0.007186258211731911, + 0.008978833444416523, + 0.0160774327814579, + 0.013868491165339947, + 0.011187192052602768, + 0.008259058929979801, + 0.012494586408138275, + 0.011611737310886383, + 0.006986555177718401, + 0.006457371637225151, + 0.004930283408612013, + 0.007625803351402283, + 0.011796618811786175, + 0.00629012007266283, + 0.005945621989667416, + 0.004986134357750416, + 0.005652035120874643, + 0.010528155602514744, + 0.007696159649640322, + 0.015471060760319233, + 0.009247174486517906, + 0.010761274956166744, + 0.012280263006687164, + 0.006868541240692139, + 0.012996404431760311, + 0.019251827150583267, + 0.01520682591944933, + 0.01938600465655327, + 0.01648908294737339, + 0.014416535384953022, + 0.010252141393721104, + 0.020699186250567436, + 0.010082044638693333, + 0.012754950672388077, + 0.01125925499945879, + 0.01866520196199417, + 0.009049797430634499, + 0.01866435632109642, + 0.009395768865942955, + 0.017849233001470566, + 0.0016578494105488062, + 0.004392515402287245, + 0.002048854250460863, + 0.006165796425193548, + 0.004797849804162979, + 0.0022038002498447895, + 0.00701748114079237, + 0.0027291940059512854, + 0.008724652230739594, + 0.011757371947169304, + 0.013626332394778728, + 0.07780952006578445, + 0.016598517075181007, + 0.005582483485341072, + 0.08418236672878265, + 0.01772884838283062, + 0.0025326847098767757, + 0.002608883660286665, + 0.003152844961732626, + 0.0028075973968952894, + 0.0043852077797055244, + 0.002195174340158701, + 0.0029427038971334696, + 0.00183130765799433, + 0.006408883258700371, + 0.003966338466852903, + 0.009310644119977951, + 0.0035693154204636812, + 0.0036503602750599384, + 0.0036174794659018517, + 0.003828789107501507, + 0.0036059285048395395, + 0.04584953188896179, + 0.014944317750632763, + 0.031600333750247955, + 0.028119565919041634, + 0.015261420980095863, + 0.04707742854952812, + 0.013759605586528778, + 0.016344821080565453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.0015142414486035705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070968583226204 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011070968583226204, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070968583226204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03506199270486832, + 0.04134098440408707, + 0.03093554824590683, + 0.05223017930984497, + 0.030274897813796997, + 0.042916636914014816, + 0.03223038464784622, + 0.09354148060083389, + 0.21295621991157532, + 0.42851173877716064 + ], + "credit_map": [ + 0.0025148997083306313, + 0.004298689309507608, + 0.005783231928944588, + 0.005232721101492643, + 0.00408147694543004, + 0.0031032401602715254, + 0.0060019236989319324, + 0.004045804962515831, + 0.003778290469199419, + 0.005028698593378067, + 0.0037881319876760244, + 0.005208381451666355, + 0.00781448557972908, + 0.006025469861924648, + 0.005635976325720549, + 0.004061543848365545, + 0.002563069574534893, + 0.003710241289809346, + 0.0027870810590684414, + 0.005445215851068497, + 0.004923992790281773, + 0.004070613998919725, + 0.0043990169651806355, + 0.003036314155906439, + 0.005428631789982319, + 0.00680717034265399, + 0.007198958192020655, + 0.005672530736774206, + 0.007200193591415882, + 0.006154891103506088, + 0.004960616584867239, + 0.00880718044936657, + 0.003539517754688859, + 0.0032001249492168427, + 0.00439029186964035, + 0.004013686906546354, + 0.0035243008751422167, + 0.003975605592131615, + 0.0036375345662236214, + 0.0039938329719007015, + 0.0032150389160960913, + 0.005853438749909401, + 0.003410311648622155, + 0.007601624354720116, + 0.004761640448123217, + 0.006146291270852089, + 0.008062457665801048, + 0.0038658296689391136, + 0.0035829413682222366, + 0.005156319122761488, + 0.0033080601133406162, + 0.006452658213675022, + 0.0050027258694171906, + 0.0017977735260501504, + 0.004060348030179739, + 0.002869553631171584, + 0.012259016744792461, + 0.010533136315643787, + 0.009771487675607204, + 0.014098694548010826, + 0.019597282633185387, + 0.007291051559150219, + 0.013469668105244637, + 0.006521130912005901, + 0.04043245688080788, + 0.022919168695807457, + 0.05811919644474983, + 0.020540637895464897, + 0.02181134559214115, + 0.01368179265409708, + 0.019344884902238846, + 0.01610671915113926, + 0.08393540233373642, + 0.02983918972313404, + 0.05891027674078941, + 0.060904424637556076, + 0.04226337745785713, + 0.08739105612039566, + 0.033757928758859634, + 0.03151008114218712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.011070968583226204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward tilt begins at a very low angle, capturing the sturdy lower body of a woman, perfectly balanced in a difficult handstand yoga pose. Her hands, fingers spread wide, rest firmly on the sunlit, light-polished wooden floor, the muscular contours of her arms defined and graceful. As the camera smoothly and steadily moves upward, the image gradually reveals her signature form-fitting tracksuit: a vibrant red top with simple, pure white stripes along the sides and arms. The highly stretchy fabric clings to her powerful curves. The background is a minimalist room, where large floor-to-ceiling windows slice the warm afternoon sunlight into brilliant beams, casting them onto the floor and her body, creating a stark contrast of light and shadow. Even tiny particles of dust can be seen dancing peacefully in the air. The camera continued to rise, passing over her black hair tied into a neat high ponytail, the hair glowing softly in the backlight, and finally, the camera slowly fixed on her calm and focused face, her eyes lightly closed, her facial muscles completely relaxed, with a faint smile of satisfaction at the corners of her mouth, and a few tiny beads of sweat oozing from her forehead crystal clear in the sunlight. The whole picture was filled with the Zen spirit of harmonious coexistence of strength and softness, exuding an atmosphere of extreme concentration and peace.", + "seed": 1538, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_6/index_744_seed_1538_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0842093676328659, + "vlm_logprob": -0.0842093676328659, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.0842093676328659, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36287620663642883 + ], + "opposite_logprob_tokens": [ + -1.2378761768341064 + ], + "logprob_reward": -0.36287620663642883, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36287620663642883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15566116571426392, + 0.05937027931213379, + 0.06260988861322403, + 0.031321194022893906, + 0.053544625639915466, + 0.0923321470618248, + 0.120817631483078, + 0.28363296389579773, + 0.07001820206642151, + 0.07069188356399536 + ], + "credit_map": [ + 0.017090557143092155, + 0.018598487600684166, + 0.021044109016656876, + 0.03612349554896355, + 0.009870287962257862, + 0.01616312935948372, + 0.016912586987018585, + 0.01985851489007473, + 0.006572543643414974, + 0.008077646605670452, + 0.006356381345540285, + 0.009537894278764725, + 0.007354615721851587, + 0.005570314824581146, + 0.009552597999572754, + 0.006348284427076578, + 0.0061986083164811134, + 0.010027065873146057, + 0.012918300926685333, + 0.009277760051190853, + 0.007389754056930542, + 0.004073415417224169, + 0.008037296123802662, + 0.004687686916440725, + 0.004130619112402201, + 0.004216424189507961, + 0.004159788601100445, + 0.0047136046923696995, + 0.005518904887139797, + 0.0031661444809287786, + 0.0032859239727258682, + 0.0021297847852110863, + 0.011024299077689648, + 0.005435673054307699, + 0.008617670275270939, + 0.004758263938128948, + 0.005311558488756418, + 0.005738247185945511, + 0.004926121328026056, + 0.007732790429145098, + 0.012231624685227871, + 0.012466629967093468, + 0.008381261490285397, + 0.01120832934975624, + 0.0119086354970932, + 0.015903372317552567, + 0.010730139911174774, + 0.009502149187028408, + 0.018380772322416306, + 0.020359352231025696, + 0.017720382660627365, + 0.010112423449754715, + 0.01554164383560419, + 0.012748418375849724, + 0.01539684645831585, + 0.010557789355516434, + 0.024511298164725304, + 0.03338669613003731, + 0.025034908205270767, + 0.020410317927598953, + 0.08636907488107681, + 0.041326407343149185, + 0.029121525585651398, + 0.023472756147384644, + 0.006212960463017225, + 0.009289674460887909, + 0.0063413348980247974, + 0.017521386966109276, + 0.006711097434163094, + 0.0077287862077355385, + 0.010091564618051052, + 0.006121400278061628, + 0.006816170644015074, + 0.009929720312356949, + 0.005506919231265783, + 0.013107341714203358, + 0.009479283355176449, + 0.010958756320178509, + 0.008043043315410614, + 0.006850647274404764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.36287620663642883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609816402196884 + ], + "opposite_logprob_tokens": [ + -4.1410980224609375 + ], + "logprob_reward": -0.01609816402196884, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609816402196884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14530281722545624, + 0.09005295485258102, + 0.0668485015630722, + 0.13047009706497192, + 0.12167750298976898, + 0.09333755075931549, + 0.03299727290868759, + 0.22316423058509827, + 0.05666179955005646, + 0.039487309753894806 + ], + "credit_map": [ + 0.015938518568873405, + 0.0172344371676445, + 0.021510113030672073, + 0.020730504766106606, + 0.010843638330698013, + 0.026590071618556976, + 0.016885217279195786, + 0.015570329502224922, + 0.009670769795775414, + 0.012550944462418556, + 0.010756737552583218, + 0.014207807369530201, + 0.01041320338845253, + 0.007666683290153742, + 0.013483839109539986, + 0.011302966624498367, + 0.006472137290984392, + 0.00891792867332697, + 0.014733022078871727, + 0.00950256735086441, + 0.008277506567537785, + 0.004726526793092489, + 0.009091225452721119, + 0.0051275803707540035, + 0.01354382187128067, + 0.01802978478372097, + 0.017350727692246437, + 0.015145715326070786, + 0.031287048012018204, + 0.015448885969817638, + 0.011258934624493122, + 0.008405179716646671, + 0.037290167063474655, + 0.009747030213475227, + 0.03193516656756401, + 0.005763813387602568, + 0.008982903324067593, + 0.0070493342354893684, + 0.008016969077289104, + 0.012892127968370914, + 0.018899550661444664, + 0.013001023791730404, + 0.011752030812203884, + 0.009684523567557335, + 0.010249687358736992, + 0.013201257213950157, + 0.008900033310055733, + 0.007649444509297609, + 0.006187812425196171, + 0.006370456889271736, + 0.004154367838054895, + 0.0027900610584765673, + 0.0040900628082454205, + 0.0033618404995650053, + 0.0034419065341353416, + 0.0026007636915892363, + 0.015487042255699635, + 0.02060665376484394, + 0.014290299266576767, + 0.0161452479660511, + 0.07081545889377594, + 0.03996007516980171, + 0.022686120122671127, + 0.02317332848906517, + 0.005222519394010305, + 0.009676062501966953, + 0.003964010626077652, + 0.011460508219897747, + 0.006387288216501474, + 0.006225567311048508, + 0.007004303392022848, + 0.006721536163240671, + 0.004558489192277193, + 0.00679585849866271, + 0.004280773922801018, + 0.008310673758387566, + 0.0037646114360541105, + 0.003852282417938113, + 0.00449123652651906, + 0.003433383069932461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.01609816402196884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346659287810326 + ], + "opposite_logprob_tokens": [ + -3.6513466835021973 + ], + "logprob_reward": -0.026346659287810326, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346659287810326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23771697282791138, + 0.19197027385234833, + 0.15804287791252136, + 0.08960670977830887, + 0.08594925701618195, + 0.05563879758119583, + 0.020842740312218666, + 0.07156252861022949, + 0.04318031296133995, + 0.04548952728509903 + ], + "credit_map": [ + 0.035053499042987823, + 0.03156140819191933, + 0.032043565064668655, + 0.04517485946416855, + 0.013589777983725071, + 0.04382172226905823, + 0.018453938886523247, + 0.0180182047188282, + 0.026177652180194855, + 0.026268111541867256, + 0.03580502048134804, + 0.03191060200333595, + 0.0158405601978302, + 0.017860831692814827, + 0.02233545295894146, + 0.01577201671898365, + 0.01985190063714981, + 0.019605236127972603, + 0.04405459016561508, + 0.03278074041008949, + 0.009400607086718082, + 0.007481650449335575, + 0.011700309813022614, + 0.013167830184102058, + 0.012090873904526234, + 0.010730997659265995, + 0.016549892723560333, + 0.01639508083462715, + 0.009965653531253338, + 0.0068909842520952225, + 0.009881841950118542, + 0.00710138538852334, + 0.016275571659207344, + 0.009760614484548569, + 0.022821003571152687, + 0.00809317734092474, + 0.007578084245324135, + 0.0066175321117043495, + 0.0062753790989518166, + 0.008527891710400581, + 0.010197252966463566, + 0.007118706125766039, + 0.006427672225981951, + 0.005396263208240271, + 0.00608238996937871, + 0.008695527911186218, + 0.005419880151748657, + 0.006301100365817547, + 0.002387657528743148, + 0.0028960902709513903, + 0.0022944121155887842, + 0.0016964614624157548, + 0.004132072441279888, + 0.002517079235985875, + 0.003013591980561614, + 0.001905373064801097, + 0.005692202597856522, + 0.00917384959757328, + 0.004386608023196459, + 0.009921589866280556, + 0.015223933383822441, + 0.009561932645738125, + 0.00917759072035551, + 0.008424818515777588, + 0.00406538276001811, + 0.007833126932382584, + 0.0031746267341077328, + 0.007325927726924419, + 0.004940412938594818, + 0.005346676334738731, + 0.0070738461799919605, + 0.0034203133545815945, + 0.004830044228583574, + 0.007539475802332163, + 0.005582133308053017, + 0.008286966010928154, + 0.004896698519587517, + 0.006491849198937416, + 0.004558814689517021, + 0.003303540637716651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026346659287810326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031931623816490173 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031931623816490173, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031931623816490173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28816866874694824, + 0.14123964309692383, + 0.040178533643484116, + 0.0495123453438282, + 0.06963137537240982, + 0.10826445370912552, + 0.0489707887172699, + 0.09243295341730118, + 0.062094174325466156, + 0.09950706362724304 + ], + "credit_map": [ + 0.031071290373802185, + 0.029753291979432106, + 0.037827860563993454, + 0.05602793022990227, + 0.019072357565164566, + 0.047697290778160095, + 0.027422945946455002, + 0.03929571807384491, + 0.027902726083993912, + 0.013272983953356743, + 0.03406029939651489, + 0.015920748934149742, + 0.01577436737716198, + 0.008841116912662983, + 0.017976466566324234, + 0.007490941323339939, + 0.003684345865622163, + 0.004961492959409952, + 0.007248350419104099, + 0.007449179422110319, + 0.003391799284145236, + 0.0029212830122560263, + 0.004196721129119396, + 0.006325362250208855, + 0.005889417603611946, + 0.006514549721032381, + 0.01007553655654192, + 0.007979776710271835, + 0.005206518340855837, + 0.005416195839643478, + 0.00480047520250082, + 0.003629873041063547, + 0.011746584437787533, + 0.012932387180626392, + 0.013063802383840084, + 0.006968367379158735, + 0.006372867617756128, + 0.004728756379336119, + 0.005945408716797829, + 0.007873200811445713, + 0.020388532429933548, + 0.016321558505296707, + 0.011433747597038746, + 0.0104129733517766, + 0.011908003129065037, + 0.01760380156338215, + 0.008574798703193665, + 0.011621033772826195, + 0.0065619926899671555, + 0.01023048534989357, + 0.0040090433321893215, + 0.0041399188339710236, + 0.006621827837079763, + 0.0060889325104653835, + 0.006734635680913925, + 0.00458395155146718, + 0.006804784294217825, + 0.016551049426198006, + 0.007134352345019579, + 0.013996331952512264, + 0.0145060233771801, + 0.009413298219442368, + 0.012577933259308338, + 0.011449175886809826, + 0.006066554691642523, + 0.011359630152583122, + 0.005409265868365765, + 0.012516294606029987, + 0.007259951904416084, + 0.006276176776736975, + 0.007341627497226, + 0.005864676088094711, + 0.013568826951086521, + 0.014197762124240398, + 0.014314563944935799, + 0.017391715198755264, + 0.00964797381311655, + 0.012727106921374798, + 0.010702133178710938, + 0.006956983357667923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0031931623816490173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman performing a handstand yoga pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a form-fitting tracksuit with a vibrant red top and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a minimalist room with large floor-to-ceiling windows and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and smoothly tilt upward to reveal the subject's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's face visible at the end with her eyes closed, a faint smile, and beads of sweat on her forehead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532645836472511 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532645836472511, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532645836472511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10683409124612808, + 0.07840533554553986, + 0.054658349603414536, + 0.02004580944776535, + 0.08227337151765823, + 0.06932640075683594, + 0.08414492011070251, + 0.11264250427484512, + 0.26517465710639954, + 0.12649448215961456 + ], + "credit_map": [ + 0.018018599599599838, + 0.0118692172691226, + 0.012453454546630383, + 0.01668008603155613, + 0.007531910203397274, + 0.02000189572572708, + 0.012482876889407635, + 0.007796051446348429, + 0.009643584489822388, + 0.007761156652122736, + 0.01528011541813612, + 0.010561826638877392, + 0.01042838953435421, + 0.005965387914329767, + 0.012165510095655918, + 0.006599371321499348, + 0.0057702865451574326, + 0.009410791099071503, + 0.011464053764939308, + 0.009637038223445415, + 0.004591555334627628, + 0.002645061584189534, + 0.005539453588426113, + 0.00560011574998498, + 0.0026208513882011175, + 0.0029828851111233234, + 0.003813383402302861, + 0.002336296485736966, + 0.002729244064539671, + 0.002102822298184037, + 0.0019060184713453054, + 0.0015543093904852867, + 0.013664530590176582, + 0.007917052134871483, + 0.01997537352144718, + 0.008904913440346718, + 0.007378743030130863, + 0.007050022482872009, + 0.007496837060898542, + 0.009885908104479313, + 0.008332313969731331, + 0.01092162262648344, + 0.00672931969165802, + 0.004768935963511467, + 0.00863347016274929, + 0.015987809747457504, + 0.007216613739728928, + 0.006736319046467543, + 0.010568887926638126, + 0.017371006309986115, + 0.005126299802213907, + 0.006783280521631241, + 0.012012737803161144, + 0.011040057055652142, + 0.011227361857891083, + 0.010015292093157768, + 0.011585501953959465, + 0.017282715067267418, + 0.007597239222377539, + 0.015331856906414032, + 0.02651064842939377, + 0.014209904707968235, + 0.011064655147492886, + 0.009059994481503963, + 0.019463304430246353, + 0.039099764078855515, + 0.025363033637404442, + 0.0526309460401535, + 0.03554452955722809, + 0.033374298363924026, + 0.041833437979221344, + 0.0178653784096241, + 0.016484130173921585, + 0.01577870547771454, + 0.013752801343798637, + 0.022617219015955925, + 0.01527155376970768, + 0.019597874954342842, + 0.01314583420753479, + 0.00984637625515461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.012532645836472511 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_tensor([9382, 9382])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_tensor([9382, 9382])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3fd5bd14319e341b6a63c24a0390135f391cc0d3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_tensor([9382, 9382])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion tracking shot, shot upward from a low angle, captures this breathtaking moment with a strikingly cinematic composition. On a futuristic asphalt street, lashed by dense rain and reflecting off the vibrant neon lights, Sheriff Black Cat, in his police uniform, is at the apex of his leap from his motorcycle. His jet-black fur gleams wetly under the shimmering cyan and magenta billboards, and raindrops sling off his body in shimmering arcs. He wears a tight-fitting navy blue uniform, the gold shield-shaped badge on his chest reflecting a sharp highlight, and the stiff collar of his white shirt remains crisp in the rain. His body is fully extended in the air, his muscles clearly defined. One front paw thrusts forward with thunderous force, pointed at a target off-screen. His eyes are focused and sharp, their yellow pupils like two bright lights in the darkness. Behind him, the futuristic police motorcycle remains tilted, its red and blue lights spinning silently, casting intermittent halos of light on the wet road. The entire scene is filled with razor-sharp detail, from the distinct hair on his beard to the texture of the fabric of his uniform, while a subtle layer of film grain covers the image, creating an atmosphere of tension, movement, and classic action cinema.", + "seed": 28302, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28302_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03404340762645006, + "vlm_logprob": -0.03404340762645006, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03404340762645006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04983214661478996 + ], + "opposite_logprob_tokens": [ + -3.0498321056365967 + ], + "logprob_reward": -0.04983214661478996, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04983214661478996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12580366432666779, + 0.03795574977993965, + 0.05871296301484108, + 0.04617445915937424, + 0.05462424084544182, + 0.1124754473567009, + 0.06859596818685532, + 0.3280239999294281, + 0.033490411937236786, + 0.13414311408996582 + ], + "credit_map": [ + 0.01560320146381855, + 0.014726669527590275, + 0.012862777337431908, + 0.024200523272156715, + 0.02270599454641342, + 0.01123378798365593, + 0.01397504098713398, + 0.010495667345821857, + 0.006171417888253927, + 0.004057818092405796, + 0.006171247456222773, + 0.005667403340339661, + 0.004221457988023758, + 0.004178537055850029, + 0.00492113409563899, + 0.00256673083640635, + 0.013594729825854301, + 0.005948045291006565, + 0.0074325199238955975, + 0.007518455386161804, + 0.005819264799356461, + 0.007184639573097229, + 0.006964036263525486, + 0.0042512728832662106, + 0.004592747427523136, + 0.0056620254181325436, + 0.004924342036247253, + 0.004953387193381786, + 0.011075280606746674, + 0.005655848421156406, + 0.005018502473831177, + 0.004292324185371399, + 0.005838894750922918, + 0.005792555399239063, + 0.0071555133908987045, + 0.006970720365643501, + 0.00960935465991497, + 0.004759816452860832, + 0.006343742832541466, + 0.00815364345908165, + 0.015168754383921623, + 0.009035327471792698, + 0.03246621415019035, + 0.010342301800847054, + 0.014227627776563168, + 0.005909922067075968, + 0.017836084589362144, + 0.007489215582609177, + 0.005095391534268856, + 0.0054592834785580635, + 0.013608266599476337, + 0.011509530246257782, + 0.008876007981598377, + 0.006634309887886047, + 0.00854821503162384, + 0.008864965289831161, + 0.028148476034402847, + 0.015854623168706894, + 0.12795394659042358, + 0.07390264421701431, + 0.013106999918818474, + 0.0075605567544698715, + 0.021870872005820274, + 0.0396258719265461, + 0.00280137755908072, + 0.002516192616894841, + 0.002679112134501338, + 0.006486546713858843, + 0.004144619684666395, + 0.0028207767754793167, + 0.004606917500495911, + 0.007434867322444916, + 0.007841097190976143, + 0.008306993171572685, + 0.03837093338370323, + 0.04672810062766075, + 0.01155016664415598, + 0.004791508428752422, + 0.009516088292002678, + 0.007038240786641836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.04983214661478996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620647557079792 + ], + "opposite_logprob_tokens": [ + -2.8120646476745605 + ], + "logprob_reward": -0.0620647557079792, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620647557079792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3889998197555542, + 0.04623247683048248, + 0.1226923018693924, + 0.026201650500297546, + 0.0694567933678627, + 0.07373370230197906, + 0.07742388546466827, + 0.035660382360219955, + 0.07451213151216507, + 0.08508683741092682 + ], + "credit_map": [ + 0.05324375256896019, + 0.053344953805208206, + 0.04004799574613571, + 0.06273920834064484, + 0.05767136439681053, + 0.038019269704818726, + 0.048357028514146805, + 0.035576242953538895, + 0.006887533236294985, + 0.003622297430410981, + 0.005887690465897322, + 0.009530947543680668, + 0.0035196617245674133, + 0.0034787876065820456, + 0.007931799627840519, + 0.005373756401240826, + 0.019645441323518753, + 0.012037832289934158, + 0.0170266292989254, + 0.015737222507596016, + 0.013515360653400421, + 0.018957896158099174, + 0.013455004431307316, + 0.012316920794546604, + 0.004610947333276272, + 0.0028915624134242535, + 0.00344858318567276, + 0.0023806190583854914, + 0.0031871877145022154, + 0.0032763401977717876, + 0.003624474164098501, + 0.002781935967504978, + 0.008943083696067333, + 0.006046704947948456, + 0.008197233080863953, + 0.00893381517380476, + 0.015787625685334206, + 0.006054835394024849, + 0.007790402974933386, + 0.0077030956745147705, + 0.010844585485756397, + 0.006465916521847248, + 0.02196139469742775, + 0.006837902590632439, + 0.006720505189150572, + 0.0031150782015174627, + 0.01291652861982584, + 0.0048717958852648735, + 0.010177260264754295, + 0.009790908545255661, + 0.00883746612817049, + 0.00964451301842928, + 0.007643691264092922, + 0.010815093293786049, + 0.009959857910871506, + 0.010555092245340347, + 0.005475242156535387, + 0.0053751785308122635, + 0.003010628279298544, + 0.0062002986669540405, + 0.0036774342879652977, + 0.002270692028105259, + 0.0035013104788959026, + 0.0061495983973145485, + 0.008586261421442032, + 0.005079158581793308, + 0.0052818115800619125, + 0.0168566033244133, + 0.008208747953176498, + 0.005262932274490595, + 0.010963782668113708, + 0.014272836036980152, + 0.013576824218034744, + 0.005830005276948214, + 0.027607804164290428, + 0.010807405225932598, + 0.00760286208242178, + 0.0068632569164037704, + 0.007001698017120361, + 0.005796984303742647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0620647557079792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013375869020819664 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013375869020819664, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013375869020819664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21857216954231262, + 0.08614612370729446, + 0.13215389847755432, + 0.02892572432756424, + 0.13286545872688293, + 0.09168484807014465, + 0.09758402407169342, + 0.02295273169875145, + 0.07771139591932297, + 0.11140371114015579 + ], + "credit_map": [ + 0.028099028393626213, + 0.03381511569023132, + 0.020995454862713814, + 0.03823251649737358, + 0.03633681684732437, + 0.015760792419314384, + 0.026898274198174477, + 0.018434172496199608, + 0.01055849064141512, + 0.009030391462147236, + 0.01149587519466877, + 0.011957259848713875, + 0.01140926405787468, + 0.009690369479358196, + 0.013641306199133396, + 0.008363167755305767, + 0.02182784676551819, + 0.015525955706834793, + 0.022499721497297287, + 0.012714889831840992, + 0.01217742171138525, + 0.017313219606876373, + 0.01620400696992874, + 0.013890828937292099, + 0.004197055473923683, + 0.002892043674364686, + 0.003797091543674469, + 0.0028463334310799837, + 0.004193723201751709, + 0.004049685318022966, + 0.0046422951854765415, + 0.0023074978962540627, + 0.01918262057006359, + 0.013996955007314682, + 0.019947316497564316, + 0.019396575167775154, + 0.023577094078063965, + 0.009719113819301128, + 0.013085583224892616, + 0.013960201293230057, + 0.009467207826673985, + 0.006986180320382118, + 0.01625935360789299, + 0.026638755574822426, + 0.00948372296988964, + 0.004325783811509609, + 0.013002313673496246, + 0.005521534476429224, + 0.0074464925564825535, + 0.010070106945931911, + 0.020439092069864273, + 0.011271527037024498, + 0.009578951634466648, + 0.015280075371265411, + 0.015497338026762009, + 0.008000444620847702, + 0.002765261335298419, + 0.0027904908638447523, + 0.0019540840294212103, + 0.003769183298572898, + 0.004012754186987877, + 0.002058928832411766, + 0.002240725327283144, + 0.0033613049890846014, + 0.009328696876764297, + 0.005289201159030199, + 0.005501942243427038, + 0.023345598950982094, + 0.009300332516431808, + 0.00624029291793704, + 0.008733538910746574, + 0.009971784427762032, + 0.02163531817495823, + 0.009364625439047813, + 0.021532388404011726, + 0.022880785167217255, + 0.010103830136358738, + 0.004995130933821201, + 0.012734591960906982, + 0.008157029747962952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0013375869020819664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369936719536781 + ], + "opposite_logprob_tokens": [ + -3.4086992740631104 + ], + "logprob_reward": -0.03369936719536781, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369936719536781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22491823136806488, + 0.16937455534934998, + 0.11979344487190247, + 0.03531322628259659, + 0.03774715214967728, + 0.03850061446428299, + 0.06147018447518349, + 0.12191805988550186, + 0.08313318341970444, + 0.10783132910728455 + ], + "credit_map": [ + 0.0334162637591362, + 0.038308095186948776, + 0.02124694734811783, + 0.028579160571098328, + 0.030995404347777367, + 0.02056061290204525, + 0.02958674356341362, + 0.022224988788366318, + 0.03216386213898659, + 0.012443711049854755, + 0.03499113395810127, + 0.024695713073015213, + 0.016042077913880348, + 0.011160529218614101, + 0.027165831997990608, + 0.010711713694036007, + 0.030238676816225052, + 0.013316430151462555, + 0.017928346991539, + 0.015403782948851585, + 0.00981411524116993, + 0.011144415475428104, + 0.016019809991121292, + 0.0059278723783791065, + 0.005073179490864277, + 0.004747974220663309, + 0.005061807110905647, + 0.005240701604634523, + 0.0036354847252368927, + 0.004143653437495232, + 0.0041398582980036736, + 0.003270566463470459, + 0.007497583050280809, + 0.003081619506701827, + 0.006078898441046476, + 0.0072485171258449554, + 0.004474222660064697, + 0.002322265412658453, + 0.0034575851168483496, + 0.0035864587407559156, + 0.004217325709760189, + 0.005192347336560488, + 0.006491077598184347, + 0.008344723843038082, + 0.004412428941577673, + 0.002532579703256488, + 0.004344993270933628, + 0.0029651401564478874, + 0.005672657396644354, + 0.005855164956301451, + 0.014718998223543167, + 0.004305411130189896, + 0.0062546092085540295, + 0.006535402964800596, + 0.011485031805932522, + 0.006642910186201334, + 0.01067121047526598, + 0.007723862770944834, + 0.029538700357079506, + 0.020289938896894455, + 0.00527803273871541, + 0.006776426453143358, + 0.009893866255879402, + 0.031746022403240204, + 0.010709498077630997, + 0.005100603681057692, + 0.005628571379929781, + 0.024947350844740868, + 0.00983310304582119, + 0.008407285436987877, + 0.009290862828493118, + 0.009215903468430042, + 0.01027916930615902, + 0.006669108755886555, + 0.020305955782532692, + 0.034903548657894135, + 0.009843403473496437, + 0.007625179830938578, + 0.010975638404488564, + 0.007229329086840153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.03369936719536781 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283181712031364 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283181712031364, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283181712031364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12852011620998383, + 0.05570646747946739, + 0.062152184545993805, + 0.05358239263296127, + 0.05488298833370209, + 0.09549077600240707, + 0.06148037314414978, + 0.39007195830345154, + 0.030012959614396095, + 0.06809979677200317 + ], + "credit_map": [ + 0.02161233127117157, + 0.020713621750473976, + 0.01080398540943861, + 0.017208946868777275, + 0.01944730244576931, + 0.013086678460240364, + 0.016139104962348938, + 0.00950813852250576, + 0.009242487139999866, + 0.004598165862262249, + 0.010530482977628708, + 0.007795009762048721, + 0.007033949717879295, + 0.004323310684412718, + 0.007832037284970284, + 0.004351020324975252, + 0.013477494940161705, + 0.007713066413998604, + 0.009247700683772564, + 0.005978969391435385, + 0.006586616393178701, + 0.00631966907531023, + 0.0084921196103096, + 0.004336543381214142, + 0.008632364682853222, + 0.00536473048850894, + 0.006910845171660185, + 0.005477076396346092, + 0.007568393833935261, + 0.007859863340854645, + 0.0052742725238204, + 0.006494841072708368, + 0.007545690983533859, + 0.005626539699733257, + 0.007773407734930515, + 0.0057732886634767056, + 0.007712431252002716, + 0.005563178565353155, + 0.008046427741646767, + 0.006842021830379963, + 0.010606899857521057, + 0.006882412359118462, + 0.010511708445847034, + 0.02947242558002472, + 0.008541707880795002, + 0.005740037653595209, + 0.017962269484996796, + 0.005773309618234634, + 0.0050943829119205475, + 0.006682320032268763, + 0.007772224023938179, + 0.006384782958775759, + 0.006214586552232504, + 0.009440877474844456, + 0.010929103940725327, + 0.00896209292113781, + 0.03458282724022865, + 0.021057836711406708, + 0.13301746547222137, + 0.12756967544555664, + 0.0143744433298707, + 0.006745412480086088, + 0.036789555102586746, + 0.01593472808599472, + 0.0027003195136785507, + 0.0029524737037718296, + 0.003138151718303561, + 0.004758818540722132, + 0.003220601938664913, + 0.0030250851996243, + 0.005642544012516737, + 0.004574964288622141, + 0.007453016005456448, + 0.004658565390855074, + 0.011337456293404102, + 0.023504624143242836, + 0.006899048574268818, + 0.0041924468241631985, + 0.006410857196897268, + 0.0036437774542719126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.023283181712031364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion tracking shot, shot upward from a low angle, captures this breathtaking moment with a strikingly cinematic composition. On a futuristic asphalt street, lashed by dense rain and reflecting off the vibrant neon lights, Sheriff Black Cat, in his police uniform, is at the apex of his leap from his motorcycle. His jet-black fur gleams wetly under the shimmering cyan and magenta billboards, and raindrops sling off his body in shimmering arcs. He wears a tight-fitting navy blue uniform, the gold shield-shaped badge on his chest reflecting a sharp highlight, and the stiff collar of his white shirt remains crisp in the rain. His body is fully extended in the air, his muscles clearly defined. One front paw thrusts forward with thunderous force, pointed at a target off-screen. His eyes are focused and sharp, their yellow pupils like two bright lights in the darkness. Behind him, the futuristic police motorcycle remains tilted, its red and blue lights spinning silently, casting intermittent halos of light on the wet road. The entire scene is filled with razor-sharp detail, from the distinct hair on his beard to the texture of the fabric of his uniform, while a subtle layer of film grain covers the image, creating an atmosphere of tension, movement, and classic action cinema.", + "seed": 28303, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28303_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04579690967220813, + "vlm_logprob": -0.04579690967220813, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.04579690967220813, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1273500621318817 + ], + "opposite_logprob_tokens": [ + -2.127350091934204 + ], + "logprob_reward": -0.1273500621318817, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1273500621318817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17771518230438232, + 0.021577445790171623, + 0.1277865618467331, + 0.3878755271434784, + 0.020572498440742493, + 0.050129104405641556, + 0.13872411847114563, + 0.024275965988636017, + 0.015733225271105766, + 0.03561042994260788 + ], + "credit_map": [ + 0.021123068407177925, + 0.020158667117357254, + 0.021905185654759407, + 0.026626871898770332, + 0.017326580360531807, + 0.017950307577848434, + 0.026635970920324326, + 0.025988519191741943, + 0.0024782216642051935, + 0.0026739235036075115, + 0.002319308929145336, + 0.002565183211117983, + 0.0039328099228441715, + 0.002950687427073717, + 0.0020212307572364807, + 0.0026360778138041496, + 0.012810922227799892, + 0.020024649798870087, + 0.015111410990357399, + 0.020889654755592346, + 0.018389027565717697, + 0.01563330367207527, + 0.014887329190969467, + 0.01004026085138321, + 0.016791604459285736, + 0.04473521560430527, + 0.022566484287381172, + 0.14411115646362305, + 0.08499843627214432, + 0.025899240747094154, + 0.0181424580514431, + 0.03063090331852436, + 0.0014278346206992865, + 0.0017074343049898744, + 0.0025673212949186563, + 0.005182985216379166, + 0.0016435043653473258, + 0.002305371919646859, + 0.003206187393516302, + 0.0025318567641079426, + 0.004355656448751688, + 0.008619704283773899, + 0.005946371704339981, + 0.005912142340093851, + 0.00471503334119916, + 0.005775280296802521, + 0.004989056382328272, + 0.009815855883061886, + 0.013134443201124668, + 0.045122258365154266, + 0.03732254356145859, + 0.009591002017259598, + 0.006787148304283619, + 0.011271831579506397, + 0.008045936934649944, + 0.007448954973369837, + 0.0019273612415418029, + 0.0029912847094237804, + 0.002642291598021984, + 0.003011985681951046, + 0.0028910371474921703, + 0.004343799315392971, + 0.0035337794106453657, + 0.002934426302090287, + 0.001736255595460534, + 0.0016496427124366164, + 0.002538719680160284, + 0.0019170920131728053, + 0.0016664251452311873, + 0.001874924753792584, + 0.0023054806515574455, + 0.0020446833223104477, + 0.0036081657744944096, + 0.0036397860385477543, + 0.00446778628975153, + 0.00477351900190115, + 0.0038632366340607405, + 0.00464373966678977, + 0.005788716021925211, + 0.004825477954000235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.1273500621318817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004360109567642 + ], + "opposite_logprob_tokens": [ + -2.6950435638427734 + ], + "logprob_reward": -0.07004360109567642, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004360109567642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163126602768898, + 0.025978999212384224, + 0.09002453833818436, + 0.15938834846019745, + 0.11131107062101364, + 0.055808357894420624, + 0.27976274490356445, + 0.0422513522207737, + 0.035807881504297256, + 0.036540135741233826 + ], + "credit_map": [ + 0.021771755069494247, + 0.018453998491168022, + 0.023904139176011086, + 0.022456778213381767, + 0.015026474371552467, + 0.013886588625609875, + 0.025107430294156075, + 0.022519439458847046, + 0.003972527105361223, + 0.0025818529538810253, + 0.0035728798247873783, + 0.004076363053172827, + 0.0028028436936438084, + 0.003037923015654087, + 0.002413610927760601, + 0.0035209981724619865, + 0.009715674445033073, + 0.01412502583116293, + 0.010288115590810776, + 0.016096480190753937, + 0.012793087400496006, + 0.011058480478823185, + 0.00945372600108385, + 0.0064939530566334724, + 0.011994521133601665, + 0.023519854992628098, + 0.018241094425320625, + 0.02310451678931713, + 0.026120072230696678, + 0.01888917200267315, + 0.01484504621475935, + 0.0226740762591362, + 0.008445718325674534, + 0.010252125561237335, + 0.015034657903015614, + 0.014841894619166851, + 0.008511664345860481, + 0.023557066917419434, + 0.015780530869960785, + 0.014887413941323757, + 0.005731365643441677, + 0.007053025532513857, + 0.009010379202663898, + 0.00856709759682417, + 0.007557854987680912, + 0.006929569877684116, + 0.005554132163524628, + 0.005404936149716377, + 0.03203541785478592, + 0.10625292360782623, + 0.08156733959913254, + 0.013697796501219273, + 0.012632389552891254, + 0.010781299322843552, + 0.013161100447177887, + 0.009634464979171753, + 0.008932179771363735, + 0.008609680458903313, + 0.005353942047804594, + 0.004312536679208279, + 0.0036758100613951683, + 0.0032616991084069014, + 0.003581661032512784, + 0.004523843992501497, + 0.004962460603564978, + 0.004871661309152842, + 0.005657605826854706, + 0.00398147851228714, + 0.003417885396629572, + 0.00467622559517622, + 0.004114375449717045, + 0.004126188810914755, + 0.0039036828093230724, + 0.004067742731422186, + 0.004250143188983202, + 0.004735942929983139, + 0.00488057266920805, + 0.006213916931301355, + 0.004649811889976263, + 0.003838320728391409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.07004360109567642 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335086883045733 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335086883045733, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335086883045733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18057863414287567, + 0.07523083686828613, + 0.047175053507089615, + 0.3732762336730957, + 0.05663711205124855, + 0.03535695746541023, + 0.08978043496608734, + 0.034581102430820465, + 0.027974706143140793, + 0.07940896600484848 + ], + "credit_map": [ + 0.0219511017203331, + 0.020635295659303665, + 0.026325223967432976, + 0.0255748238414526, + 0.01820957101881504, + 0.016057509928941727, + 0.027051763609051704, + 0.024773335084319115, + 0.009709850884974003, + 0.0051343608647584915, + 0.009279024787247181, + 0.010202188976109028, + 0.014131300151348114, + 0.011845550499856472, + 0.00815789308398962, + 0.006770670413970947, + 0.005443249829113483, + 0.00744246831163764, + 0.006000573746860027, + 0.007148697506636381, + 0.006851112004369497, + 0.0064230626448988914, + 0.003858179785311222, + 0.00400770828127861, + 0.017932916060090065, + 0.05149823799729347, + 0.021495508030056953, + 0.1328098475933075, + 0.07162955403327942, + 0.023459656164050102, + 0.01867031119763851, + 0.03578023612499237, + 0.004616053309291601, + 0.005242673214524984, + 0.007326183374971151, + 0.008374280296266079, + 0.0038347640074789524, + 0.012409969232976437, + 0.007983001880347729, + 0.006850185338407755, + 0.005232597701251507, + 0.005271653179079294, + 0.005157751496881247, + 0.005002534948289394, + 0.003120958339422941, + 0.004135067109018564, + 0.0033302984666079283, + 0.0041060978546738625, + 0.008610252290964127, + 0.03149092197418213, + 0.027870897203683853, + 0.005182859022170305, + 0.004308175761252642, + 0.003971712663769722, + 0.004693370778113604, + 0.0036522403825074434, + 0.007493687327951193, + 0.008094326592981815, + 0.004327451344579458, + 0.0029116233345121145, + 0.003228110494092107, + 0.003978620748966932, + 0.0025138838682323694, + 0.002033397788181901, + 0.0037982838694006205, + 0.004354675766080618, + 0.004862726666033268, + 0.0028651533648371696, + 0.0036472314968705177, + 0.002840786473825574, + 0.0023105733562260866, + 0.0032952737528830767, + 0.01055601704865694, + 0.0104151526466012, + 0.01651652529835701, + 0.008338799700140953, + 0.008603292517364025, + 0.010178367607295513, + 0.007625699043273926, + 0.0071751126088202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.001335086883045733 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606578752398491 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.01606578752398491, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606578752398491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09612105041742325, + 0.0259792972356081, + 0.046704716980457306, + 0.4496403634548187, + 0.1130758747458458, + 0.04206426069140434, + 0.11605875194072723, + 0.05845479294657707, + 0.014910323545336723, + 0.036990586668252945 + ], + "credit_map": [ + 0.01641770824790001, + 0.01205157395452261, + 0.015087985433638096, + 0.012606917880475521, + 0.008946079760789871, + 0.006677529774606228, + 0.011897893622517586, + 0.012435362674295902, + 0.00280013564042747, + 0.0022074526641517878, + 0.002784798387438059, + 0.0028809993527829647, + 0.0034257571678608656, + 0.0033133914694190025, + 0.0040153381414711475, + 0.004551423713564873, + 0.005622842349112034, + 0.0066666752099990845, + 0.0058213635347783566, + 0.007627479732036591, + 0.006680059712380171, + 0.007218129467219114, + 0.003523975843563676, + 0.0035441950894892216, + 0.019696062430739403, + 0.05483231320977211, + 0.02938155084848404, + 0.16130989789962769, + 0.09210427105426788, + 0.02816259115934372, + 0.021889962255954742, + 0.042263712733983994, + 0.008169916458427906, + 0.01037509087473154, + 0.01549486257135868, + 0.016338838264346123, + 0.008096395060420036, + 0.022060545161366463, + 0.016074420884251595, + 0.016465799883008003, + 0.006836907472461462, + 0.007753393147140741, + 0.006895150989294052, + 0.004239960107952356, + 0.004209220875054598, + 0.004527329932898283, + 0.0031090753618627787, + 0.004493224900215864, + 0.015483230352401733, + 0.03803987801074982, + 0.030700359493494034, + 0.0064926352351903915, + 0.008039992302656174, + 0.005746763199567795, + 0.005665136501193047, + 0.005890753120183945, + 0.013970805332064629, + 0.012051546946167946, + 0.007910347543656826, + 0.004397244658321142, + 0.004934351425617933, + 0.006386647000908852, + 0.004105098079890013, + 0.004698752425611019, + 0.001839259872213006, + 0.0016668324824422598, + 0.0017741923220455647, + 0.0014751917915418744, + 0.0016470763366669416, + 0.002361530438065529, + 0.0017494471976533532, + 0.002396792871877551, + 0.0032429031562060118, + 0.00375228188931942, + 0.003942278213799, + 0.004711982794106007, + 0.0021178945899009705, + 0.0077465595677495, + 0.006222781725227833, + 0.005253905896097422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.01606578752398491 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190010726451874 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190010726451874, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190010726451874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05812407284975052, + 0.037214718759059906, + 0.03218702971935272, + 0.18738031387329102, + 0.06149487942457199, + 0.08565062284469604, + 0.40423479676246643, + 0.045531049370765686, + 0.03615643084049225, + 0.05202607437968254 + ], + "credit_map": [ + 0.007288144435733557, + 0.007499292027205229, + 0.008482360281050205, + 0.007300710771232843, + 0.007778440602123737, + 0.006805694196373224, + 0.00720404926687479, + 0.005765383597463369, + 0.00442684069275856, + 0.0038182351272553205, + 0.005513215437531471, + 0.004872528836131096, + 0.00440209312364459, + 0.005657200235873461, + 0.0046611656434834, + 0.0038634410593658686, + 0.0034889469388872385, + 0.004249798599630594, + 0.004382086452096701, + 0.004652224946767092, + 0.004760464187711477, + 0.00429249182343483, + 0.003132150275632739, + 0.003228869754821062, + 0.016800526529550552, + 0.0322263166308403, + 0.018318112939596176, + 0.020063066855072975, + 0.03246867656707764, + 0.01968558505177498, + 0.018268967047333717, + 0.02954907715320587, + 0.00421918136999011, + 0.0085225161164999, + 0.007494739722460508, + 0.010225625708699226, + 0.005859099794179201, + 0.00758480466902256, + 0.007865259423851967, + 0.009723653085529804, + 0.016325820237398148, + 0.012707923538982868, + 0.01467464491724968, + 0.009843812324106693, + 0.008424917235970497, + 0.00803370587527752, + 0.007133818697184324, + 0.008505980484187603, + 0.04286627471446991, + 0.15432405471801758, + 0.125220388174057, + 0.02082069218158722, + 0.0175323486328125, + 0.013111941516399384, + 0.018532339483499527, + 0.011826731264591217, + 0.00692307623103261, + 0.006978996563702822, + 0.00851285271346569, + 0.00465738819912076, + 0.003341519273817539, + 0.006106261163949966, + 0.0034381647128611803, + 0.005572788883000612, + 0.00402466207742691, + 0.005461052060127258, + 0.006004126276820898, + 0.0032492063473910093, + 0.006856531370431185, + 0.00379124004393816, + 0.0032914767507463694, + 0.0034781377762556076, + 0.006628841161727905, + 0.006930401548743248, + 0.012080206535756588, + 0.004903071094304323, + 0.004649674985557795, + 0.00750426109880209, + 0.005365950986742973, + 0.003963673021644354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.014190010726451874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion tracking shot, shot upward from a low angle, captures this breathtaking moment with a strikingly cinematic composition. On a futuristic asphalt street, lashed by dense rain and reflecting off the vibrant neon lights, Sheriff Black Cat, in his police uniform, is at the apex of his leap from his motorcycle. His jet-black fur gleams wetly under the shimmering cyan and magenta billboards, and raindrops sling off his body in shimmering arcs. He wears a tight-fitting navy blue uniform, the gold shield-shaped badge on his chest reflecting a sharp highlight, and the stiff collar of his white shirt remains crisp in the rain. His body is fully extended in the air, his muscles clearly defined. One front paw thrusts forward with thunderous force, pointed at a target off-screen. His eyes are focused and sharp, their yellow pupils like two bright lights in the darkness. Behind him, the futuristic police motorcycle remains tilted, its red and blue lights spinning silently, casting intermittent halos of light on the wet road. The entire scene is filled with razor-sharp detail, from the distinct hair on his beard to the texture of the fabric of his uniform, while a subtle layer of film grain covers the image, creating an atmosphere of tension, movement, and classic action cinema.", + "seed": 28304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28304_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0391720799729228, + "vlm_logprob": -0.0391720799729228, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.0391720799729228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11372540891170502 + ], + "opposite_logprob_tokens": [ + -2.238725423812866 + ], + "logprob_reward": -0.11372540891170502, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11372540891170502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10753128677606583, + 0.051869891583919525, + 0.15494683384895325, + 0.07530909031629562, + 0.12121220678091049, + 0.03676372021436691, + 0.2251489907503128, + 0.09008953720331192, + 0.10312943160533905, + 0.03399898484349251 + ], + "credit_map": [ + 0.015254649333655834, + 0.015516980551183224, + 0.019689036533236504, + 0.011377638205885887, + 0.011925376951694489, + 0.009219981729984283, + 0.012160176411271095, + 0.012387446127831936, + 0.006956629455089569, + 0.004748650826513767, + 0.007914302870631218, + 0.007181593682616949, + 0.004547185264527798, + 0.005070315673947334, + 0.008269927464425564, + 0.007181287743151188, + 0.01767609640955925, + 0.0268083605915308, + 0.016831284388899803, + 0.013506288640201092, + 0.017683321610093117, + 0.02607557363808155, + 0.013660978525876999, + 0.022704945877194405, + 0.015071207657456398, + 0.011828340590000153, + 0.01558583416044712, + 0.008233328349888325, + 0.007338386029005051, + 0.007180154323577881, + 0.00491176126524806, + 0.005160082131624222, + 0.01631048321723938, + 0.022816162556409836, + 0.024772539734840393, + 0.011982637457549572, + 0.010213762521743774, + 0.017906971275806427, + 0.00962778925895691, + 0.007581877056509256, + 0.004652183037251234, + 0.004317490383982658, + 0.00630256487056613, + 0.005680411122739315, + 0.003128773532807827, + 0.0025328395422548056, + 0.004837133921682835, + 0.005312327761203051, + 0.02541830763220787, + 0.04124275967478752, + 0.034182947129011154, + 0.0425860658288002, + 0.017126401886343956, + 0.016145948320627213, + 0.021505147218704224, + 0.026941442862153053, + 0.011532244272530079, + 0.0106219332665205, + 0.011794223450124264, + 0.01744403876364231, + 0.008204328827559948, + 0.007532051298767328, + 0.012825191020965576, + 0.010135534219443798, + 0.015129734762012959, + 0.007300070486962795, + 0.044647905975580215, + 0.014454078860580921, + 0.003795253112912178, + 0.0031727987807244062, + 0.00939674861729145, + 0.0052328468300402164, + 0.003571840701624751, + 0.004377740900963545, + 0.012554741464555264, + 0.004824791569262743, + 0.001813932554796338, + 0.0015725779812783003, + 0.0035062930546700954, + 0.001777068478986621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.11372540891170502 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486757829785347 + ], + "opposite_logprob_tokens": [ + -3.048675775527954 + ], + "logprob_reward": -0.0486757829785347, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486757829785347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14052528142929077, + 0.09821737557649612, + 0.1441168487071991, + 0.27209022641181946, + 0.0803825706243515, + 0.059924282133579254, + 0.06951256096363068, + 0.021131351590156555, + 0.06734806299209595, + 0.046751417219638824 + ], + "credit_map": [ + 0.022752003744244576, + 0.013954073190689087, + 0.036413904279470444, + 0.01826895959675312, + 0.011487139388918877, + 0.009232188574969769, + 0.01060106884688139, + 0.01781594194471836, + 0.012769976630806923, + 0.011233130469918251, + 0.01815561018884182, + 0.010689979419112206, + 0.01138343382626772, + 0.009246277622878551, + 0.013199915178120136, + 0.011539052240550518, + 0.019020000472664833, + 0.02092694491147995, + 0.02077599987387657, + 0.015805717557668686, + 0.01500711776316166, + 0.01980705000460148, + 0.01883097179234028, + 0.013943055644631386, + 0.1008557602763176, + 0.015172894112765789, + 0.10583456605672836, + 0.015239514410495758, + 0.007359018549323082, + 0.010752229019999504, + 0.008015713654458523, + 0.008860504254698753, + 0.012872201390564442, + 0.010535540990531445, + 0.01878720335662365, + 0.008281547576189041, + 0.005267447326332331, + 0.013401414267718792, + 0.0053742630407214165, + 0.005862954072654247, + 0.0070608798414468765, + 0.00769390631467104, + 0.010622399859130383, + 0.008237971924245358, + 0.005846500862389803, + 0.00326556246727705, + 0.007636321242898703, + 0.009560740552842617, + 0.011076490394771099, + 0.010026396252214909, + 0.008205287158489227, + 0.015546574257314205, + 0.005620023235678673, + 0.007058793678879738, + 0.006436177529394627, + 0.005542818456888199, + 0.0024363526608794928, + 0.0035088821314275265, + 0.0030710839200764894, + 0.0031423321925103664, + 0.0016489571426063776, + 0.001984167378395796, + 0.0030082829762250185, + 0.002331292489543557, + 0.009832783602178097, + 0.005125638097524643, + 0.026809973642230034, + 0.010258782655000687, + 0.0024023237638175488, + 0.0030738175846636295, + 0.006065795198082924, + 0.003778951009735465, + 0.004210331477224827, + 0.0058752866461873055, + 0.01607142761349678, + 0.006482973694801331, + 0.003991077654063702, + 0.0021432156208902597, + 0.005281669553369284, + 0.0026954328641295433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0486757829785347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23078086972236633, + 0.15040403604507446, + 0.07151831686496735, + 0.21032597124576569, + 0.07194337248802185, + 0.06762765347957611, + 0.03932727500796318, + 0.023479873314499855, + 0.06882087886333466, + 0.06577177345752716 + ], + "credit_map": [ + 0.026736164465546608, + 0.026069119572639465, + 0.03890146315097809, + 0.02705451101064682, + 0.026903940364718437, + 0.021019306033849716, + 0.029673844575881958, + 0.03442252799868584, + 0.019979998469352722, + 0.01904256083071232, + 0.023683441802859306, + 0.021209828555583954, + 0.015264004468917847, + 0.01431626919656992, + 0.021830622106790543, + 0.015077301301062107, + 0.009524241089820862, + 0.00909558404237032, + 0.009454919956624508, + 0.01525198481976986, + 0.005383366718888283, + 0.005179839674383402, + 0.011153259314596653, + 0.0064751217141747475, + 0.07268483191728592, + 0.009246465750038624, + 0.08119237422943115, + 0.00925562996417284, + 0.005017575342208147, + 0.017675526440143585, + 0.007717347703874111, + 0.0075362143106758595, + 0.01098704431205988, + 0.006885809823870659, + 0.014436069875955582, + 0.005598907358944416, + 0.005832155700773001, + 0.01583711802959442, + 0.005163745954632759, + 0.007202522363513708, + 0.008281567133963108, + 0.007234704680740833, + 0.010448649525642395, + 0.007680879905819893, + 0.005758907180279493, + 0.006236491724848747, + 0.014578742906451225, + 0.0074077085591852665, + 0.005285407416522503, + 0.004159430507570505, + 0.0045362841337919235, + 0.0050166151486337185, + 0.0037077635060995817, + 0.009470845572650433, + 0.0025388156063854694, + 0.004612110089510679, + 0.004105567466467619, + 0.0026971425395458937, + 0.003062047529965639, + 0.002223390154540539, + 0.0017665198538452387, + 0.005134979262948036, + 0.0019491229904815555, + 0.0025411027017980814, + 0.007792623247951269, + 0.006220415234565735, + 0.01528297457844019, + 0.0070643643848598, + 0.006006633397191763, + 0.014069647528231144, + 0.006329774856567383, + 0.006054447963833809, + 0.007238478399813175, + 0.006489951629191637, + 0.025428561493754387, + 0.00921617727726698, + 0.0036431814078241587, + 0.0030276526231318712, + 0.0073457094840705395, + 0.0033820613753050566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194242656230927 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194242656230927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194242656230927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11683604121208191, + 0.07289845496416092, + 0.1359889805316925, + 0.17143505811691284, + 0.057317107915878296, + 0.0379897877573967, + 0.11223116517066956, + 0.04148435592651367, + 0.2159755975008011, + 0.037843458354473114 + ], + "credit_map": [ + 0.014479177072644234, + 0.016175633296370506, + 0.020417092368006706, + 0.014617751352488995, + 0.013983977027237415, + 0.011571226641535759, + 0.014168844558298588, + 0.011422340758144855, + 0.01090426929295063, + 0.009063130244612694, + 0.012915940023958683, + 0.009868201799690723, + 0.007083685137331486, + 0.006623927969485521, + 0.010678953491151333, + 0.005760343745350838, + 0.015807412564754486, + 0.021445928141474724, + 0.01745852641761303, + 0.011457941494882107, + 0.016961660236120224, + 0.018601227551698685, + 0.014293995685875416, + 0.019962284713983536, + 0.062021590769290924, + 0.011529151350259781, + 0.06326332688331604, + 0.010770440101623535, + 0.005120323039591312, + 0.006362246815115213, + 0.006654948927462101, + 0.005713026504963636, + 0.007334682159125805, + 0.009529715403914452, + 0.011583367362618446, + 0.006882625166326761, + 0.004008270800113678, + 0.008104768581688404, + 0.005238919984549284, + 0.004634758457541466, + 0.004729963839054108, + 0.003999658394604921, + 0.005429496988654137, + 0.003833338851109147, + 0.0037220839876681566, + 0.00565994530916214, + 0.006166661158204079, + 0.0044486382976174355, + 0.01297515444457531, + 0.02473382093012333, + 0.01662164181470871, + 0.014216040261089802, + 0.01230716984719038, + 0.01179117988795042, + 0.009529203176498413, + 0.01005695853382349, + 0.006481009069830179, + 0.005796522833406925, + 0.007496431469917297, + 0.004734185058623552, + 0.0031554207671433687, + 0.00401827460154891, + 0.004659166559576988, + 0.005143344402313232, + 0.032498039305210114, + 0.01513729989528656, + 0.09918935596942902, + 0.030196594074368477, + 0.004969149827957153, + 0.005389458499848843, + 0.020238127559423447, + 0.008357555605471134, + 0.0036975322291254997, + 0.005403056275099516, + 0.007208762690424919, + 0.004202425014227629, + 0.0035842403303831816, + 0.005369486752897501, + 0.004764096345752478, + 0.0036138570867478848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.014194242656230927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069188714027405 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069188714027405, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069188714027405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08853361010551453, + 0.0850164070725441, + 0.04772014543414116, + 0.13006016612052917, + 0.1616363674402237, + 0.08652603626251221, + 0.0882350355386734, + 0.025457415729761124, + 0.260834664106369, + 0.02598019503057003 + ], + "credit_map": [ + 0.017076922580599785, + 0.007769228424876928, + 0.01909794472157955, + 0.013802742585539818, + 0.005328938364982605, + 0.006033778190612793, + 0.010579661466181278, + 0.00884439516812563, + 0.01097796019166708, + 0.007726694922894239, + 0.013278337195515633, + 0.007681038696318865, + 0.006251861806958914, + 0.012369093485176563, + 0.009958677925169468, + 0.016772735863924026, + 0.0077932970598340034, + 0.007697266526520252, + 0.006134646013379097, + 0.006428850814700127, + 0.004997392650693655, + 0.0038408630061894655, + 0.006078669801354408, + 0.004749158397316933, + 0.03494792804121971, + 0.013211929239332676, + 0.03556549921631813, + 0.009108904749155045, + 0.009285525418817997, + 0.011720473878085613, + 0.008013267070055008, + 0.008206617087125778, + 0.024671724066138268, + 0.022086704149842262, + 0.04060922563076019, + 0.01766466163098812, + 0.00879007950425148, + 0.022711971774697304, + 0.014719261787831783, + 0.010382718406617641, + 0.010641224682331085, + 0.011151842772960663, + 0.012124475091695786, + 0.01100916787981987, + 0.009231018833816051, + 0.009949848055839539, + 0.010876202024519444, + 0.011542250402271748, + 0.012181667611002922, + 0.013109715655446053, + 0.008917445316910744, + 0.02234615944325924, + 0.0061494167894124985, + 0.01115479227155447, + 0.0067363400012254715, + 0.007639497518539429, + 0.0035660755820572376, + 0.0037707474548369646, + 0.0035038040950894356, + 0.003165401751175523, + 0.00230263639241457, + 0.0029356200248003006, + 0.0028326804749667645, + 0.0033804483246058226, + 0.03900202363729477, + 0.01760789379477501, + 0.11885187774896622, + 0.035791076719760895, + 0.006233477499336004, + 0.008324182592332363, + 0.024922989308834076, + 0.010101119056344032, + 0.002136463997885585, + 0.0032077678479254246, + 0.009354665875434875, + 0.003377884393557906, + 0.0018869036575779319, + 0.001615297980606556, + 0.0028756405226886272, + 0.001525570172816515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.016069188714027405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion tracking shot, shot upward from a low angle, captures this breathtaking moment with a strikingly cinematic composition. On a futuristic asphalt street, lashed by dense rain and reflecting off the vibrant neon lights, Sheriff Black Cat, in his police uniform, is at the apex of his leap from his motorcycle. His jet-black fur gleams wetly under the shimmering cyan and magenta billboards, and raindrops sling off his body in shimmering arcs. He wears a tight-fitting navy blue uniform, the gold shield-shaped badge on his chest reflecting a sharp highlight, and the stiff collar of his white shirt remains crisp in the rain. His body is fully extended in the air, his muscles clearly defined. One front paw thrusts forward with thunderous force, pointed at a target off-screen. His eyes are focused and sharp, their yellow pupils like two bright lights in the darkness. Behind him, the futuristic police motorcycle remains tilted, its red and blue lights spinning silently, casting intermittent halos of light on the wet road. The entire scene is filled with razor-sharp detail, from the distinct hair on his beard to the texture of the fabric of his uniform, while a subtle layer of film grain covers the image, creating an atmosphere of tension, movement, and classic action cinema.", + "seed": 28305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_7/index_9382_seed_28305_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.032161621702834965, + "vlm_logprob": -0.032161621702834965, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.032161621702834965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055756233632564545 + ], + "opposite_logprob_tokens": [ + -2.9307563304901123 + ], + "logprob_reward": -0.055756233632564545, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055756233632564545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11282021552324295, + 0.06888201832771301, + 0.1608337014913559, + 0.05267736315727234, + 0.0727539211511612, + 0.02767946384847164, + 0.09741344302892685, + 0.18165303766727448, + 0.19066935777664185, + 0.034617453813552856 + ], + "credit_map": [ + 0.010095124132931232, + 0.009444303810596466, + 0.012557144276797771, + 0.009747656062245369, + 0.010257954709231853, + 0.011653770692646503, + 0.023919550701975822, + 0.02514471672475338, + 0.005833502858877182, + 0.003626521211117506, + 0.018966833129525185, + 0.0052078161388635635, + 0.0033588209189474583, + 0.003926265984773636, + 0.005693785846233368, + 0.022268472239375114, + 0.012381711043417454, + 0.01584649831056595, + 0.02036917768418789, + 0.02170747146010399, + 0.02329530380666256, + 0.018295513466000557, + 0.02364097349345684, + 0.02529706247150898, + 0.007439519744366407, + 0.00462741032242775, + 0.005023476202040911, + 0.005558098666369915, + 0.010371170938014984, + 0.007191348355263472, + 0.006628214847296476, + 0.005838125944137573, + 0.005812433548271656, + 0.010522241704165936, + 0.008356550708413124, + 0.007949240505695343, + 0.006728771608322859, + 0.015446707606315613, + 0.009325838647782803, + 0.00861214566975832, + 0.0043495530262589455, + 0.002888283459469676, + 0.00369049608707428, + 0.005195403005927801, + 0.0041573322378098965, + 0.003354133339598775, + 0.0020143850706517696, + 0.0020298792514950037, + 0.016949158161878586, + 0.02144278958439827, + 0.009195643477141857, + 0.018031589686870575, + 0.008745498023927212, + 0.006415352690964937, + 0.007440920453518629, + 0.009192494675517082, + 0.041839588433504105, + 0.019855916500091553, + 0.015780620276927948, + 0.02033333294093609, + 0.007224157452583313, + 0.010626052506268024, + 0.01837020367383957, + 0.047623176127672195, + 0.01969696208834648, + 0.015521252527832985, + 0.026845190674066544, + 0.026580076664686203, + 0.010568737052381039, + 0.013996903784573078, + 0.017447104677557945, + 0.060013141483068466, + 0.0044983504340052605, + 0.0036132230889052153, + 0.004335700534284115, + 0.0028088868129998446, + 0.0029245836194604635, + 0.006764175835996866, + 0.006345613393932581, + 0.003326923120766878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.055756233632564545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868214204907417 + ], + "opposite_logprob_tokens": [ + -3.04868221282959 + ], + "logprob_reward": -0.04868214204907417, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868214204907417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14622776210308075, + 0.21238219738006592, + 0.1021006628870964, + 0.034921206533908844, + 0.04907439276576042, + 0.03759068623185158, + 0.103595070540905, + 0.17868764698505402, + 0.08585786819458008, + 0.04956252500414848 + ], + "credit_map": [ + 0.008542121388018131, + 0.006474788300693035, + 0.00860206875950098, + 0.007551492191851139, + 0.008912392891943455, + 0.010266612283885479, + 0.04865250363945961, + 0.04722578451037407, + 0.016209250316023827, + 0.013236712664365768, + 0.05523949861526489, + 0.017339767888188362, + 0.014377495273947716, + 0.010271095670759678, + 0.021662592887878418, + 0.06404579430818558, + 0.010036190040409565, + 0.01428611297160387, + 0.01083142589777708, + 0.013840294443070889, + 0.017757190391421318, + 0.010922796092927456, + 0.011111066676676273, + 0.013315588235855103, + 0.004395828116685152, + 0.004172285553067923, + 0.0037883613258600235, + 0.0037454322446137667, + 0.006685200612992048, + 0.004529679659754038, + 0.004442061763256788, + 0.0031623595859855413, + 0.006089228671044111, + 0.010818042792379856, + 0.005906061269342899, + 0.005455882754176855, + 0.006510200910270214, + 0.005290554370731115, + 0.005255650728940964, + 0.003748771036043763, + 0.003749035531654954, + 0.003667775308713317, + 0.006387932226061821, + 0.013002203777432442, + 0.0032981715630739927, + 0.002813776256516576, + 0.00259763584472239, + 0.002074155956506729, + 0.015236233361065388, + 0.025084273889660835, + 0.014330049976706505, + 0.018289443105459213, + 0.007825089618563652, + 0.0068153454922139645, + 0.006218559108674526, + 0.009796073660254478, + 0.04327455908060074, + 0.018293917179107666, + 0.012980068102478981, + 0.018708443269133568, + 0.00838549342006445, + 0.009717020206153393, + 0.017413916066288948, + 0.049914225935935974, + 0.007665917742997408, + 0.01112392544746399, + 0.01756913587450981, + 0.00978062767535448, + 0.004135193768888712, + 0.0051498450338840485, + 0.006121027749031782, + 0.02431219071149826, + 0.007205022964626551, + 0.005374780856072903, + 0.0051139299757778645, + 0.004764133598655462, + 0.0037594579625874758, + 0.009664461016654968, + 0.009181602858006954, + 0.004499136470258236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.04868214204907417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021991373505443335 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.0021991373505443335, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021991373505443335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27942559123039246, + 0.29577842354774475, + 0.0537942536175251, + 0.016413936391472816, + 0.023422129452228546, + 0.03198831155896187, + 0.06024656072258949, + 0.08104453235864639, + 0.0610380657017231, + 0.09684824198484421 + ], + "credit_map": [ + 0.012926154769957066, + 0.009645710699260235, + 0.014678881503641605, + 0.009147454053163528, + 0.016150567680597305, + 0.0118514159694314, + 0.10344211012125015, + 0.10158327221870422, + 0.024821126833558083, + 0.011726882308721542, + 0.09517345577478409, + 0.014560121111571789, + 0.0096811568364501, + 0.011869044043123722, + 0.016556911170482635, + 0.11138972640037537, + 0.007934379391372204, + 0.0072615547105669975, + 0.0062503148801624775, + 0.008344881236553192, + 0.006907639559358358, + 0.0050504752434790134, + 0.006075206212699413, + 0.005969800986349583, + 0.0023452364839613438, + 0.0016415331047028303, + 0.002139830496162176, + 0.0022606311831623316, + 0.0024299074430018663, + 0.0018796033691614866, + 0.0022524185478687286, + 0.001464775181375444, + 0.003319650888442993, + 0.003220276441425085, + 0.0028181441593915224, + 0.00341354263946414, + 0.003482441883534193, + 0.0024940914008766413, + 0.0026982766576111317, + 0.0019757067784667015, + 0.003057178808376193, + 0.003204868407920003, + 0.005349141079932451, + 0.010315325111150742, + 0.0035738074220716953, + 0.0024704337120056152, + 0.002164491917937994, + 0.0018530661473050714, + 0.009403294883668423, + 0.013979689218103886, + 0.008676563389599323, + 0.007685940712690353, + 0.004653421696275473, + 0.005601930897682905, + 0.004845898132771254, + 0.005399818066507578, + 0.015134830959141254, + 0.007515668869018555, + 0.007965650409460068, + 0.00724835554137826, + 0.005066580139100552, + 0.011219050735235214, + 0.008905943483114243, + 0.017988456413149834, + 0.005586969666182995, + 0.009410565719008446, + 0.011865847744047642, + 0.006984046660363674, + 0.003816270735114813, + 0.0037272206973284483, + 0.004355217330157757, + 0.015291932038962841, + 0.013065267354249954, + 0.010172074660658836, + 0.009895793162286282, + 0.007879932411015034, + 0.014424911700189114, + 0.01447364967316389, + 0.017220230773091316, + 0.00971638597548008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0021991373505443335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810246288776398 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810246288776398, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810246288776398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2210162729024887, + 0.030256161466240883, + 0.059393610805273056, + 0.06871723383665085, + 0.0370839461684227, + 0.04109096899628639, + 0.19513598084449768, + 0.02407905086874962, + 0.22919870913028717, + 0.0940280556678772 + ], + "credit_map": [ + 0.012693686410784721, + 0.007355932146310806, + 0.011967800557613373, + 0.009439279325306416, + 0.012058740481734276, + 0.011140442453324795, + 0.07908237725496292, + 0.0772780105471611, + 0.0030297827906906605, + 0.003093771869316697, + 0.0038388725370168686, + 0.0038629816845059395, + 0.005263619590550661, + 0.0026449928991496563, + 0.004504989366978407, + 0.0040171509608626366, + 0.0060440474189817905, + 0.006764921359717846, + 0.007395819760859013, + 0.01094791665673256, + 0.006713957991451025, + 0.006120673380792141, + 0.00834942702203989, + 0.0070568532682955265, + 0.009747683070600033, + 0.008378643542528152, + 0.008420130237936974, + 0.007333928253501654, + 0.0070038107223808765, + 0.00902138277888298, + 0.011802351102232933, + 0.007009310182183981, + 0.005926608573645353, + 0.007062658201903105, + 0.0033502988517284393, + 0.004133411217480898, + 0.004159078933298588, + 0.004438301548361778, + 0.004060015082359314, + 0.0039535751566290855, + 0.004940279293805361, + 0.0037154227029532194, + 0.006544611416757107, + 0.01129081565886736, + 0.005090013146400452, + 0.0036588944494724274, + 0.0035553090274333954, + 0.0022956193424761295, + 0.03689941391348839, + 0.03704476356506348, + 0.0394228920340538, + 0.03051774762570858, + 0.011387248523533344, + 0.011025073938071728, + 0.010465958155691624, + 0.018372882157564163, + 0.00267564645037055, + 0.0028616059571504593, + 0.0033926854375749826, + 0.0048898970708251, + 0.002015058882534504, + 0.0025505570229142904, + 0.0027420511469244957, + 0.0029515507631003857, + 0.04085591435432434, + 0.02375044859945774, + 0.027016423642635345, + 0.0411895252764225, + 0.008420883677899837, + 0.010891799814999104, + 0.0188845694065094, + 0.058189138770103455, + 0.013575769029557705, + 0.01867174543440342, + 0.00877611618489027, + 0.01035303995013237, + 0.006140951532870531, + 0.011333832517266273, + 0.017818890511989594, + 0.0073577119037508965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.03810246288776398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black cat wearing a navy blue police uniform with a gold shield badge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat captured in mid-air at the apex of a leap from a motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic asphalt street with dense rain and vibrant neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a slow-motion effect with visible film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606813259422779 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606813259422779, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606813259422779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11403501033782959, + 0.27511343359947205, + 0.09894180297851562, + 0.03921669349074364, + 0.06876151263713837, + 0.09501789510250092, + 0.0835038498044014, + 0.07103607803583145, + 0.09167436510324478, + 0.06269930303096771 + ], + "credit_map": [ + 0.012636363506317139, + 0.010079702362418175, + 0.011963723227381706, + 0.010870568454265594, + 0.009534010663628578, + 0.01088740210980177, + 0.024984164163470268, + 0.02307906746864319, + 0.022188294678926468, + 0.012650717981159687, + 0.08527801930904388, + 0.01594547927379608, + 0.009698518551886082, + 0.010884811170399189, + 0.01684669777750969, + 0.10162089765071869, + 0.013596435077488422, + 0.01288516540080309, + 0.013432888314127922, + 0.01212344691157341, + 0.011608010157942772, + 0.007839533500373363, + 0.013981278985738754, + 0.013475049287080765, + 0.006996883545070887, + 0.004223974887281656, + 0.0037117924075573683, + 0.0032163038849830627, + 0.007488970644772053, + 0.0041075036861002445, + 0.0059667425230145454, + 0.003504521446302533, + 0.00835566595196724, + 0.011699584312736988, + 0.008228983730077744, + 0.007201612461358309, + 0.008020726032555103, + 0.011087420396506786, + 0.006415967829525471, + 0.00775155471637845, + 0.010785838589072227, + 0.009762599132955074, + 0.017584405839443207, + 0.028668511658906937, + 0.009725365787744522, + 0.006570178084075451, + 0.006525969598442316, + 0.005395024549216032, + 0.008578626438975334, + 0.019402630627155304, + 0.016169078648090363, + 0.01123118121176958, + 0.005227033980190754, + 0.007356686517596245, + 0.0060448539443314075, + 0.009493759833276272, + 0.015186426229774952, + 0.007235812488943338, + 0.008021971210837364, + 0.006570108700543642, + 0.0030174441635608673, + 0.004705640021711588, + 0.007553375791758299, + 0.01874529756605625, + 0.01676143892109394, + 0.009042903780937195, + 0.01767716556787491, + 0.015605092979967594, + 0.00409168004989624, + 0.004830573685467243, + 0.007378305308520794, + 0.01628721132874489, + 0.009813442826271057, + 0.005323507357388735, + 0.007549541536718607, + 0.005404084455221891, + 0.0067958589643239975, + 0.010386046022176743, + 0.011228367686271667, + 0.006198452785611153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01606813259422779 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_tensor([2556, 2556])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_tensor([2556, 2556])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7316be32ff8fdc8a4f14fb8ec197f03ce6b72a2c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_tensor([2556, 2556])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An elegant, sweeping, high-angle drone shot gracefully tracks a vibrant metallic red Corvette C8 Stingray, its signature mid-engine silhouette defined by sharp aerodynamic lines and large side air intakes, as it effortlessly cruises along a winding coastal highway. The entire scene is bathed in the warm, saturated light of golden hour, with the low-hanging sun casting long, dramatic shadows stretching across the smooth asphalt. The intense golden glow gleams off the car's immaculate, highly polished paint and multi-spoke alloy wheels. On one side of the road, rugged, sun-bathed cliffs rise dramatically, their earthy hues gleaming in the afternoon sun. On the other side, a vast expanse of deep blue ocean shimmers endlessly toward the horizon, gentle white waves lapping against the rocky coastline below. The Corvette navigates the road's graceful S-curves with smooth, powerful precision, its motion creating a breathtaking aura of freedom and high-performance elegance.", + "seed": 11508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_2556_seed_11508_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029417386557906867, + "vlm_logprob": -0.029417386557906867, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.029417386557906867, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038812875747680664 + ], + "opposite_logprob_tokens": [ + -4.413812637329102 + ], + "logprob_reward": -0.038812875747680664, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.038812875747680664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25970885157585144, + 0.18678651750087738, + 0.09606572985649109, + 0.051168810576200485, + 0.1454750895500183, + 0.03443518653512001, + 0.04507696256041527, + 0.09154338389635086, + 0.046029042452573776, + 0.04371041804552078 + ], + "credit_map": [ + 0.018509667366743088, + 0.015810508280992508, + 0.02886587753891945, + 0.03879198431968689, + 0.03675505146384239, + 0.03698102757334709, + 0.03393455222249031, + 0.05006017908453941, + 0.013835426419973373, + 0.011170487850904465, + 0.03837514668703079, + 0.02482534386217594, + 0.020390285179018974, + 0.038528796285390854, + 0.01714443974196911, + 0.022516585886478424, + 0.005261972080916166, + 0.0067070601508021355, + 0.008618551306426525, + 0.01525581069290638, + 0.02741457335650921, + 0.012926020659506321, + 0.01158373337239027, + 0.008297998458147049, + 0.004855537787079811, + 0.004681940656155348, + 0.008222116157412529, + 0.007373195607215166, + 0.0071270279586315155, + 0.005872812122106552, + 0.008368929848074913, + 0.004667246248573065, + 0.008850408717989922, + 0.02821514382958412, + 0.012322583235800266, + 0.012034771032631397, + 0.04279327765107155, + 0.01578441821038723, + 0.015300778672099113, + 0.010173692367970943, + 0.003955742344260216, + 0.005159029271453619, + 0.0041163042187690735, + 0.0035555707290768623, + 0.004722238052636385, + 0.0037015583366155624, + 0.004317460115998983, + 0.0049072797410190105, + 0.006683946121484041, + 0.006072231102734804, + 0.0061530740931630135, + 0.00450483150780201, + 0.005591158755123615, + 0.007007836829870939, + 0.003833683440461755, + 0.005230199079960585, + 0.01057346910238266, + 0.009998597204685211, + 0.011713125742971897, + 0.010316668078303337, + 0.010322979651391506, + 0.012981352396309376, + 0.014901950024068356, + 0.01073523610830307, + 0.004279058892279863, + 0.004903030581772327, + 0.007415277883410454, + 0.006093532778322697, + 0.005286897532641888, + 0.00711857108399272, + 0.005127962213009596, + 0.0058047049678862095, + 0.004269868601113558, + 0.0038789797108620405, + 0.007947910577058792, + 0.0052420105785131454, + 0.0054158782586455345, + 0.007322354707866907, + 0.005366560071706772, + 0.004266852047294378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.038812875747680664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662858977913857 + ], + "opposite_logprob_tokens": [ + -4.758662700653076 + ], + "logprob_reward": -0.008662858977913857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662858977913857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.200955331325531, + 0.172251895070076, + 0.09287094324827194, + 0.07836127281188965, + 0.043603766709566116, + 0.02613619901239872, + 0.026811793446540833, + 0.19577467441558838, + 0.07662826776504517, + 0.08660587668418884 + ], + "credit_map": [ + 0.01962423324584961, + 0.013842599466443062, + 0.01993607170879841, + 0.0273672416806221, + 0.022357473149895668, + 0.031207798048853874, + 0.024588679894804955, + 0.042031239718198776, + 0.010122478939592838, + 0.007587556727230549, + 0.03737155348062515, + 0.02608657255768776, + 0.023603806272149086, + 0.034028612077236176, + 0.016926907002925873, + 0.016524408012628555, + 0.005485649220645428, + 0.006344311870634556, + 0.010318215936422348, + 0.011276938021183014, + 0.027837669476866722, + 0.01325622946023941, + 0.012326211668550968, + 0.006025717593729496, + 0.005426284857094288, + 0.0055291843600571156, + 0.00809738039970398, + 0.00425503496080637, + 0.02906220033764839, + 0.01436157152056694, + 0.007247085217386484, + 0.004382532089948654, + 0.0043619656935334206, + 0.00788155384361744, + 0.004648292437195778, + 0.006381392944604158, + 0.004398546181619167, + 0.0050423769280314445, + 0.0056810323148965836, + 0.005208609160035849, + 0.003463577479124069, + 0.0039209648966789246, + 0.0031769960187375546, + 0.0027110434602946043, + 0.002680235542356968, + 0.0026367150712758303, + 0.003106585470959544, + 0.004440081771463156, + 0.003447573632001877, + 0.004435820505023003, + 0.0038518814835697412, + 0.002905742498114705, + 0.0030044475570321083, + 0.003483615117147565, + 0.0025207626167684793, + 0.0031619491055607796, + 0.024306440725922585, + 0.027479492127895355, + 0.032518550753593445, + 0.017632562667131424, + 0.019042961299419403, + 0.019435809925198555, + 0.02670842409133911, + 0.028650440275669098, + 0.005626104772090912, + 0.00964356679469347, + 0.01012091338634491, + 0.008227295242249966, + 0.017862938344478607, + 0.0079263960942626, + 0.007156828418374062, + 0.01006422657519579, + 0.007995061576366425, + 0.01409947220236063, + 0.012247043661773205, + 0.014226619154214859, + 0.007527367677539587, + 0.007470646407455206, + 0.014655960723757744, + 0.008383707143366337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.008662858977913857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29033854603767395, + 0.08859483152627945, + 0.19659370183944702, + 0.044236861169338226, + 0.09131523966789246, + 0.0688779354095459, + 0.06360378116369247, + 0.06116046383976936, + 0.052534423768520355, + 0.04274425283074379 + ], + "credit_map": [ + 0.018175560981035233, + 0.01951332576572895, + 0.041125401854515076, + 0.05953102558851242, + 0.021918373182415962, + 0.05326259136199951, + 0.030461881309747696, + 0.04635037109255791, + 0.006757257506251335, + 0.0071290601044893265, + 0.015990808606147766, + 0.012475759722292423, + 0.014967773109674454, + 0.013745828531682491, + 0.008739628829061985, + 0.00878871325403452, + 0.012114938348531723, + 0.0115080326795578, + 0.020404472947120667, + 0.023288292810320854, + 0.0656202882528305, + 0.028576670214533806, + 0.018812505528330803, + 0.016268493607640266, + 0.004141001496464014, + 0.003866678336635232, + 0.004106381442397833, + 0.003216404002159834, + 0.012522926554083824, + 0.007812939584255219, + 0.004076170735061169, + 0.004494357388466597, + 0.00471076462417841, + 0.0248267725110054, + 0.0065894718281924725, + 0.007759497966617346, + 0.023875562474131584, + 0.010710968635976315, + 0.006903307978063822, + 0.0059388973750174046, + 0.011263535358011723, + 0.01086666714400053, + 0.00856771320104599, + 0.0064819552935659885, + 0.008229700848460197, + 0.007553321309387684, + 0.006600632797926664, + 0.00931441318243742, + 0.009399163536727428, + 0.009901453740894794, + 0.008681672625243664, + 0.004286319483071566, + 0.007436162326484919, + 0.009044879116117954, + 0.005420365836471319, + 0.009433765895664692, + 0.011074593290686607, + 0.010656923055648804, + 0.00916300993412733, + 0.005352626089006662, + 0.005233367905020714, + 0.005198087077587843, + 0.007894093170762062, + 0.006587762385606766, + 0.004977749660611153, + 0.005737924948334694, + 0.006178719457238913, + 0.0038122436963021755, + 0.008481660857796669, + 0.012202596291899681, + 0.004785989411175251, + 0.006357539910823107, + 0.003399741603061557, + 0.00458567263558507, + 0.0063642458990216255, + 0.004061071202158928, + 0.00602423632517457, + 0.008095545694231987, + 0.004527243319898844, + 0.0056864917278289795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637557547539473 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637557547539473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637557547539473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14632245898246765, + 0.086698979139328, + 0.14033864438533783, + 0.04277827963232994, + 0.10469017922878265, + 0.05619152635335922, + 0.17299622297286987, + 0.10245823115110397, + 0.06131381541490555, + 0.08621175587177277 + ], + "credit_map": [ + 0.010106944479048252, + 0.009666265919804573, + 0.02000376209616661, + 0.024134017527103424, + 0.012620450928807259, + 0.02910546585917473, + 0.014699295163154602, + 0.02598625235259533, + 0.006634254939854145, + 0.007949239574372768, + 0.015379166230559349, + 0.008305953815579414, + 0.015502982772886753, + 0.013213654048740864, + 0.009045461192727089, + 0.01066826656460762, + 0.008972647599875927, + 0.009121161885559559, + 0.014978778548538685, + 0.016419945284724236, + 0.047116395086050034, + 0.022264810279011726, + 0.012762803584337234, + 0.008702084422111511, + 0.004578119609504938, + 0.003997156396508217, + 0.005189426243305206, + 0.003703744150698185, + 0.009684931486845016, + 0.006673560477793217, + 0.005444440990686417, + 0.00350689678452909, + 0.005024921149015427, + 0.03126172348856926, + 0.0053250896744430065, + 0.006331492681056261, + 0.03284924104809761, + 0.01070480328053236, + 0.007322366815060377, + 0.005870526656508446, + 0.008565198630094528, + 0.010160316713154316, + 0.005657646805047989, + 0.005569216329604387, + 0.006029510870575905, + 0.007021781522780657, + 0.006020094268023968, + 0.00716775469481945, + 0.025445712730288506, + 0.023230530321598053, + 0.022664833813905716, + 0.017703983932733536, + 0.024323387071490288, + 0.025483611971139908, + 0.017195304855704308, + 0.016948826611042023, + 0.012218636460602283, + 0.013312085531651974, + 0.018625443801283836, + 0.00938135851174593, + 0.00940397847443819, + 0.014599655754864216, + 0.012348431162536144, + 0.012568631209433079, + 0.007367194164544344, + 0.007267035078257322, + 0.006452250760048628, + 0.004402916878461838, + 0.01227555237710476, + 0.007781318388879299, + 0.007368631195276976, + 0.00839891005307436, + 0.008606321178376675, + 0.011638183146715164, + 0.016054127365350723, + 0.007950151339173317, + 0.00967944785952568, + 0.01140591874718666, + 0.011135933920741081, + 0.009741664864122868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004637557547539473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901961147785187 + ], + "opposite_logprob_tokens": [ + -2.464019536972046 + ], + "logprob_reward": -0.08901961147785187, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901961147785187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.137764573097229, + 0.3115087151527405, + 0.08099150657653809, + 0.09356097877025604, + 0.08402502536773682, + 0.055793121457099915, + 0.051355138421058655, + 0.0816977471113205, + 0.06244650110602379, + 0.04085671156644821 + ], + "credit_map": [ + 0.008372845128178596, + 0.008109909482300282, + 0.029091620817780495, + 0.019651856273412704, + 0.011199315078556538, + 0.026738043874502182, + 0.011957705952227116, + 0.02264326624572277, + 0.020771214738488197, + 0.023625211790204048, + 0.046984363347291946, + 0.04670628160238266, + 0.04271556809544563, + 0.07175831496715546, + 0.02742769569158554, + 0.031520068645477295, + 0.0054096500389277935, + 0.006778383161872625, + 0.011666857637465, + 0.012305868789553642, + 0.018955035135149956, + 0.010902788490056992, + 0.0080256387591362, + 0.006947281304746866, + 0.0061217304319143295, + 0.008380504325032234, + 0.007848158478736877, + 0.006078827194869518, + 0.03264731541275978, + 0.016992583870887756, + 0.009453058242797852, + 0.006038799416273832, + 0.006998545955866575, + 0.015119286254048347, + 0.010140668600797653, + 0.00948723591864109, + 0.012594969011843204, + 0.01245960034430027, + 0.008591635152697563, + 0.00863308273255825, + 0.004849448800086975, + 0.017503537237644196, + 0.006393219344317913, + 0.004884779918938875, + 0.005395686719566584, + 0.005283636972308159, + 0.005076758563518524, + 0.006406049709767103, + 0.007714252918958664, + 0.005639362148940563, + 0.006010850891470909, + 0.004944093059748411, + 0.005252297967672348, + 0.008680492639541626, + 0.005098819732666016, + 0.00801496859639883, + 0.009900064207613468, + 0.011148912832140923, + 0.014098520390689373, + 0.00833336915820837, + 0.007400388363748789, + 0.010699393227696419, + 0.008496946655213833, + 0.011620153672993183, + 0.005150849465280771, + 0.006315301172435284, + 0.00691957725211978, + 0.008400852791965008, + 0.009382518008351326, + 0.006560204550623894, + 0.011969078332185745, + 0.007748117204755545, + 0.00565628707408905, + 0.008005625568330288, + 0.004486285615712404, + 0.005902099423110485, + 0.004442629404366016, + 0.003892584703862667, + 0.00463121710345149, + 0.0038399852346628904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.08901961147785187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An elegant, sweeping, high-angle drone shot gracefully tracks a vibrant metallic red Corvette C8 Stingray, its signature mid-engine silhouette defined by sharp aerodynamic lines and large side air intakes, as it effortlessly cruises along a winding coastal highway. The entire scene is bathed in the warm, saturated light of golden hour, with the low-hanging sun casting long, dramatic shadows stretching across the smooth asphalt. The intense golden glow gleams off the car's immaculate, highly polished paint and multi-spoke alloy wheels. On one side of the road, rugged, sun-bathed cliffs rise dramatically, their earthy hues gleaming in the afternoon sun. On the other side, a vast expanse of deep blue ocean shimmers endlessly toward the horizon, gentle white waves lapping against the rocky coastline below. The Corvette navigates the road's graceful S-curves with smooth, powerful precision, its motion creating a breathtaking aura of freedom and high-performance elegance.", + "seed": 11509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_2556_seed_11509_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019284386187791824, + "vlm_logprob": -0.019284386187791824, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.019284386187791824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015725649893283844 + ], + "opposite_logprob_tokens": [ + -5.390725612640381 + ], + "logprob_reward": -0.015725649893283844, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.015725649893283844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07469575107097626, + 0.3537319004535675, + 0.019463015720248222, + 0.028719458729028702, + 0.2649146318435669, + 0.04026402160525322, + 0.035997942090034485, + 0.04509640857577324, + 0.09984153509140015, + 0.03727540001273155 + ], + "credit_map": [ + 0.008725863881409168, + 0.00777793163433671, + 0.005990686360746622, + 0.0121763925999403, + 0.012075861915946007, + 0.010059967637062073, + 0.009539698250591755, + 0.008349340409040451, + 0.044408515095710754, + 0.03817044943571091, + 0.03161893039941788, + 0.07521332055330276, + 0.029301859438419342, + 0.01671287603676319, + 0.08023729175329208, + 0.03806865215301514, + 0.0027655402664095163, + 0.0033459067344665527, + 0.0013326184125617146, + 0.0027100383304059505, + 0.0024157625157386065, + 0.0019491075072437525, + 0.0028083010111004114, + 0.0021357389632612467, + 0.0035585169680416584, + 0.002793149556964636, + 0.004606036469340324, + 0.003816620446741581, + 0.00358657562173903, + 0.0025837868452072144, + 0.004693471826612949, + 0.00308129726909101, + 0.03995742276310921, + 0.019624434411525726, + 0.08834085613489151, + 0.025498388335108757, + 0.011607430875301361, + 0.025727665051817894, + 0.026306821033358574, + 0.027851590886712074, + 0.005107483826577663, + 0.007249459624290466, + 0.006937902886420488, + 0.0040903775952756405, + 0.005567054729908705, + 0.003375947242602706, + 0.004484591074287891, + 0.003451199736446142, + 0.0037621345836669207, + 0.007425092626363039, + 0.0037384331226348877, + 0.0070517826825380325, + 0.0030505117028951645, + 0.0032286427449434996, + 0.004616150166839361, + 0.003125188872218132, + 0.0073310378938913345, + 0.004896069411188364, + 0.004323466215282679, + 0.008428158238530159, + 0.0038068084977567196, + 0.00738519337028265, + 0.005852764938026667, + 0.00307290768250823, + 0.023952584713697433, + 0.00781323667615652, + 0.0069381422363221645, + 0.0070320554077625275, + 0.01110355369746685, + 0.033016353845596313, + 0.0037046775687485933, + 0.006280919536948204, + 0.005905744154006243, + 0.004859464708715677, + 0.004151924513280392, + 0.004499171394854784, + 0.0039453147910535336, + 0.0053215851075947285, + 0.00347177404910326, + 0.005120417103171349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.015725649893283844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960902199149132 + ], + "opposite_logprob_tokens": [ + -5.130960941314697 + ], + "logprob_reward": -0.005960902199149132, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960902199149132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11196018010377884, + 0.6386401057243347, + 0.018903713673353195, + 0.026921911165118217, + 0.07187177985906601, + 0.014387799426913261, + 0.02741936966776848, + 0.03157269209623337, + 0.029591254889965057, + 0.0287312138825655 + ], + "credit_map": [ + 0.013775293715298176, + 0.013131599873304367, + 0.016924580559134483, + 0.013581742532551289, + 0.014170431531965733, + 0.016596589237451553, + 0.012623514980077744, + 0.011156431399285793, + 0.08004824072122574, + 0.06728582829236984, + 0.05869549885392189, + 0.13440312445163727, + 0.05245252326130867, + 0.030814271420240402, + 0.1459733247756958, + 0.0689672976732254, + 0.0029010954312980175, + 0.0038270498625934124, + 0.0012378101237118244, + 0.0020607097540050745, + 0.0024325335398316383, + 0.0021842466667294502, + 0.0024426088202744722, + 0.001817658543586731, + 0.003237086348235607, + 0.003465137677267194, + 0.0022560907527804375, + 0.003295858157798648, + 0.005055143032222986, + 0.0028297051321715117, + 0.004239197354763746, + 0.002543692709878087, + 0.013296800665557384, + 0.006539443973451853, + 0.016342893242836, + 0.006495542824268341, + 0.004742088727653027, + 0.013068326748907566, + 0.0059892782010138035, + 0.005397401750087738, + 0.0012384946458041668, + 0.002395854564383626, + 0.0021127969957888126, + 0.0015941128367558122, + 0.0018437894759699702, + 0.0012979377061128616, + 0.0024365659337490797, + 0.001468246104195714, + 0.002275946084409952, + 0.006393739487975836, + 0.002396808471530676, + 0.005341507960110903, + 0.0026739847380667925, + 0.0023870980367064476, + 0.003052565734833479, + 0.0028977200854569674, + 0.0031968618277460337, + 0.003887461731210351, + 0.004834813065826893, + 0.004626951180398464, + 0.002934229327365756, + 0.0033088482450693846, + 0.0040590655989944935, + 0.004724461119621992, + 0.003888867562636733, + 0.0056065102107822895, + 0.00373608130030334, + 0.004230756312608719, + 0.002111823530867696, + 0.00336357019841671, + 0.0036357564385980368, + 0.00301789166405797, + 0.002696614246815443, + 0.002215675311163068, + 0.00430133705958724, + 0.005315139889717102, + 0.00347815896384418, + 0.004158214665949345, + 0.0031107354443520308, + 0.0034553399309515953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.005960902199149132 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281434550881386 + ], + "opposite_logprob_tokens": [ + -3.7732813358306885 + ], + "logprob_reward": -0.023281434550881386, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281434550881386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08763454109430313, + 0.41912442445755005, + 0.08789380639791489, + 0.026437852531671524, + 0.15913397073745728, + 0.015026100911200047, + 0.11918915808200836, + 0.022241687402129173, + 0.030760815367102623, + 0.03255753964185715 + ], + "credit_map": [ + 0.010745865292847157, + 0.00935538299381733, + 0.014960899017751217, + 0.010067131370306015, + 0.011431586928665638, + 0.011703756637871265, + 0.009943629615008831, + 0.009426293894648552, + 0.05306481942534447, + 0.044410813599824905, + 0.03735595941543579, + 0.08793894946575165, + 0.034550443291664124, + 0.020666969940066338, + 0.09597206860780716, + 0.04516443982720375, + 0.009678516536951065, + 0.014133164659142494, + 0.0059535508044064045, + 0.006652216427028179, + 0.02078230492770672, + 0.010947393253445625, + 0.011411771178245544, + 0.008334900252521038, + 0.002937320852652192, + 0.0033828015439212322, + 0.0029686070047318935, + 0.002617573132738471, + 0.0025625661946833134, + 0.0056409514509141445, + 0.0034857133869081736, + 0.0028423196636140347, + 0.018821116536855698, + 0.009788171388208866, + 0.052303846925497055, + 0.01728418655693531, + 0.006569360848516226, + 0.013369083404541016, + 0.022018376737833023, + 0.018979843705892563, + 0.0019923371728509665, + 0.0020481771789491177, + 0.002920152386650443, + 0.0018641336355358362, + 0.0013901394559070468, + 0.0014597619883716106, + 0.002049841219559312, + 0.0013015599688515067, + 0.013653657399117947, + 0.029022889211773872, + 0.010628062300384045, + 0.013540455140173435, + 0.01260611042380333, + 0.01624789834022522, + 0.009840969927608967, + 0.013649115338921547, + 0.0020231802482157946, + 0.0015815369552001357, + 0.002904471242800355, + 0.005984076764434576, + 0.0016030940460041165, + 0.0021593591663986444, + 0.003935738001018763, + 0.0020502314437180758, + 0.0062386575154960155, + 0.007361170370131731, + 0.002163499593734741, + 0.002478941110894084, + 0.0027761508245021105, + 0.0048696426674723625, + 0.0022002775222063065, + 0.0026724794879555702, + 0.006480330601334572, + 0.00608464702963829, + 0.0037416494451463223, + 0.003383491886779666, + 0.00317671918310225, + 0.0037830921355634928, + 0.002426857827231288, + 0.003480752930045128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.023281434550881386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028168782591819763 + ], + "opposite_logprob_tokens": [ + -5.877816677093506 + ], + "logprob_reward": -0.0028168782591819763, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028168782591819763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09813740849494934, + 0.30087026953697205, + 0.09841171652078629, + 0.04573792591691017, + 0.15935446321964264, + 0.11832466721534729, + 0.03303583711385727, + 0.032296840101480484, + 0.06638914346694946, + 0.0474417470395565 + ], + "credit_map": [ + 0.0122258011251688, + 0.011859237216413021, + 0.010485108941793442, + 0.01185964047908783, + 0.016187040135264397, + 0.013215779326856136, + 0.011230375617742538, + 0.011074423789978027, + 0.037574682384729385, + 0.03139014169573784, + 0.02812497690320015, + 0.06343968957662582, + 0.02433670312166214, + 0.015363233163952827, + 0.06780152767896652, + 0.03283930942416191, + 0.011411353945732117, + 0.015781842172145844, + 0.006850556470453739, + 0.007136504165828228, + 0.023031160235404968, + 0.013025391846895218, + 0.012120871804654598, + 0.00905403308570385, + 0.005689044017344713, + 0.0050965952686965466, + 0.0036277063190937042, + 0.004179909359663725, + 0.005530361086130142, + 0.009830293245613575, + 0.006653430871665478, + 0.00513058016076684, + 0.02082609385251999, + 0.010202805511653423, + 0.04573620855808258, + 0.01515820063650608, + 0.008711561560630798, + 0.03187822923064232, + 0.014561077579855919, + 0.012280289083719254, + 0.014671881683170795, + 0.020175781100988388, + 0.021988827735185623, + 0.012627575546503067, + 0.014356589876115322, + 0.009496580809354782, + 0.014227010309696198, + 0.01078041736036539, + 0.00392042426392436, + 0.00695616751909256, + 0.00248920195735991, + 0.00627694558352232, + 0.0029381602071225643, + 0.0036703268997371197, + 0.0036698738113045692, + 0.0031147394329309464, + 0.004718421027064323, + 0.0028910303954035044, + 0.004277646541595459, + 0.006738127209246159, + 0.003263040678575635, + 0.003483604174107313, + 0.0042428686283528805, + 0.0026821016799658537, + 0.014023631811141968, + 0.0053873504512012005, + 0.005650849547237158, + 0.0067108264192938805, + 0.007050464395433664, + 0.020002974197268486, + 0.0033378556836396456, + 0.004225190728902817, + 0.00654201814904809, + 0.007649480830878019, + 0.005125297233462334, + 0.005471990909427404, + 0.005901235621422529, + 0.006607454270124435, + 0.0041486890986561775, + 0.00599557813256979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.0028168782591819763 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048637066036462784 + ], + "opposite_logprob_tokens": [ + -3.0486371517181396 + ], + "logprob_reward": -0.048637066036462784, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048637066036462784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05624784156680107, + 0.7177457809448242, + 0.02143326960504055, + 0.018786320462822914, + 0.0909908190369606, + 0.020871594548225403, + 0.02039923332631588, + 0.015143483877182007, + 0.02549625001847744, + 0.012885372154414654 + ], + "credit_map": [ + 0.007081953808665276, + 0.006885539274662733, + 0.0064164684154093266, + 0.0076916879042983055, + 0.0066469814628362656, + 0.008661879226565361, + 0.006590241566300392, + 0.006273091305047274, + 0.09092883765697479, + 0.07667611539363861, + 0.06469624489545822, + 0.15076391398906708, + 0.0589110292494297, + 0.034348975867033005, + 0.16408801078796387, + 0.07733264565467834, + 0.0027854335494339466, + 0.003727744799107313, + 0.0016650879988446832, + 0.0023880007211118937, + 0.003653779160231352, + 0.002051105722784996, + 0.0031019672751426697, + 0.0020601488649845123, + 0.0026405039243400097, + 0.0018575565190985799, + 0.0019491424318403006, + 0.0020060900133103132, + 0.0017450008308514953, + 0.0034322920255362988, + 0.003143204376101494, + 0.002012530341744423, + 0.009233490563929081, + 0.004873265046626329, + 0.031246645376086235, + 0.010236509144306183, + 0.004247384611517191, + 0.016725316643714905, + 0.008405113592743874, + 0.006023100111633539, + 0.003391327103599906, + 0.002069039037451148, + 0.005508396774530411, + 0.0024449375923722982, + 0.0018727983115240932, + 0.0012034968240186572, + 0.002966305473819375, + 0.0014152944786474109, + 0.0018074499676004052, + 0.004361949395388365, + 0.0016043586656451225, + 0.0037755309604108334, + 0.0022524890955537558, + 0.002028395887464285, + 0.0022342693991959095, + 0.0023347889073193073, + 0.0025412377435714006, + 0.0015899965073913336, + 0.00156346894800663, + 0.002142119687050581, + 0.0017124285222962499, + 0.0018165330402553082, + 0.0018149178940802813, + 0.001962781883776188, + 0.004987749271094799, + 0.002002219669520855, + 0.0024600017350167036, + 0.0027214677538722754, + 0.002565604168921709, + 0.006793390028178692, + 0.0018253200687468052, + 0.0021404996514320374, + 0.001349598285742104, + 0.001437741331756115, + 0.0018861137796193361, + 0.0018801498226821423, + 0.0013257426908239722, + 0.0017595065291970968, + 0.001627083751372993, + 0.001619436196051538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.048637066036462784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An elegant, sweeping, high-angle drone shot gracefully tracks a vibrant metallic red Corvette C8 Stingray, its signature mid-engine silhouette defined by sharp aerodynamic lines and large side air intakes, as it effortlessly cruises along a winding coastal highway. The entire scene is bathed in the warm, saturated light of golden hour, with the low-hanging sun casting long, dramatic shadows stretching across the smooth asphalt. The intense golden glow gleams off the car's immaculate, highly polished paint and multi-spoke alloy wheels. On one side of the road, rugged, sun-bathed cliffs rise dramatically, their earthy hues gleaming in the afternoon sun. On the other side, a vast expanse of deep blue ocean shimmers endlessly toward the horizon, gentle white waves lapping against the rocky coastline below. The Corvette navigates the road's graceful S-curves with smooth, powerful precision, its motion creating a breathtaking aura of freedom and high-performance elegance.", + "seed": 11510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_2556_seed_11510_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02130461223423481, + "vlm_logprob": -0.02130461223423481, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02130461223423481, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025859925895929337 + ], + "opposite_logprob_tokens": [ + -4.150859832763672 + ], + "logprob_reward": -0.025859925895929337, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.025859925895929337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12716315686702728, + 0.056528326123952866, + 0.0392351895570755, + 0.10001999139785767, + 0.12691770493984222, + 0.09407517313957214, + 0.19630414247512817, + 0.053598225116729736, + 0.13758070766925812, + 0.0685773491859436 + ], + "credit_map": [ + 0.011137705296278, + 0.009429887868463993, + 0.02360628731548786, + 0.020246630534529686, + 0.015939388424158096, + 0.01228076871484518, + 0.011680683121085167, + 0.022841811180114746, + 0.005713069345802069, + 0.005123221315443516, + 0.012127939611673355, + 0.009864330291748047, + 0.006923253182321787, + 0.004436981398612261, + 0.0068316347897052765, + 0.005507900845259428, + 0.003361767390742898, + 0.0038031362928450108, + 0.007439704611897469, + 0.005371827632188797, + 0.0049094557762146, + 0.006663715001195669, + 0.0037343758158385754, + 0.0039512114599347115, + 0.005129670258611441, + 0.01328196469694376, + 0.009140132926404476, + 0.014451862312853336, + 0.025867290794849396, + 0.012212572619318962, + 0.01369224302470684, + 0.00624425895512104, + 0.007352869957685471, + 0.0053494637832045555, + 0.03813120350241661, + 0.012848006561398506, + 0.008529186248779297, + 0.03716424107551575, + 0.00798528827726841, + 0.009557456709444523, + 0.005779679398983717, + 0.00508197071030736, + 0.02966107241809368, + 0.008104515261948109, + 0.007305721286684275, + 0.021441617980599403, + 0.011049671098589897, + 0.005650932900607586, + 0.02027299255132675, + 0.013684102334082127, + 0.0359865203499794, + 0.03350355848670006, + 0.017952071502804756, + 0.020520785823464394, + 0.01828480325639248, + 0.03609932214021683, + 0.005919505842030048, + 0.00556194968521595, + 0.00644364720210433, + 0.006300944369286299, + 0.00651546660810709, + 0.007358662784099579, + 0.008624881505966187, + 0.006873172242194414, + 0.012393975630402565, + 0.01512296125292778, + 0.019518254324793816, + 0.021516501903533936, + 0.019412189722061157, + 0.01938747614622116, + 0.017177782952785492, + 0.013051571324467659, + 0.005658109672367573, + 0.005316601600497961, + 0.007772480137646198, + 0.011263472959399223, + 0.012070300057530403, + 0.00911009032279253, + 0.0089932382106781, + 0.0083930604159832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.025859925895929337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602276548743248 + ], + "opposite_logprob_tokens": [ + -3.895602226257324 + ], + "logprob_reward": -0.020602276548743248, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602276548743248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21930257976055145, + 0.06367995589971542, + 0.04480645805597305, + 0.023627925664186478, + 0.2717495858669281, + 0.1404770165681839, + 0.09217928349971771, + 0.033987123519182205, + 0.06793826073408127, + 0.04225181043148041 + ], + "credit_map": [ + 0.0168538186699152, + 0.01356261596083641, + 0.025519315153360367, + 0.057154033333063126, + 0.014455973170697689, + 0.055627670139074326, + 0.017012152820825577, + 0.01911700703203678, + 0.004341338761150837, + 0.00534789776429534, + 0.01213088259100914, + 0.010199359618127346, + 0.006957656238228083, + 0.01029004156589508, + 0.007490211632102728, + 0.0069225686602294445, + 0.0036335785407572985, + 0.004483739845454693, + 0.007786377798765898, + 0.007963722571730614, + 0.006863395683467388, + 0.004848892334848642, + 0.005717905703932047, + 0.0035088439472019672, + 0.0014644560869783163, + 0.0032661387231200933, + 0.002571037271991372, + 0.004099757876247168, + 0.0031377917621284723, + 0.0028758798725903034, + 0.004012640565633774, + 0.0022002249024808407, + 0.015117019414901733, + 0.012110857293009758, + 0.09213615953922272, + 0.02910204976797104, + 0.015140717849135399, + 0.07696924358606339, + 0.014000196009874344, + 0.017173340544104576, + 0.006767716724425554, + 0.005052311811596155, + 0.04688790440559387, + 0.011156044900417328, + 0.00930021982640028, + 0.03738132864236832, + 0.01536373421549797, + 0.008567746728658676, + 0.012416726909577847, + 0.007386591751128435, + 0.018361208960413933, + 0.0090901218354702, + 0.011029958724975586, + 0.013428808189928532, + 0.009543084539473057, + 0.010922782123088837, + 0.003524088067933917, + 0.002764415228739381, + 0.004420917015522718, + 0.004498542286455631, + 0.0043524629436433315, + 0.004532419610768557, + 0.005211928393691778, + 0.004682349506765604, + 0.0059746624901890755, + 0.007046133279800415, + 0.009897526353597641, + 0.00834373664110899, + 0.009227766655385494, + 0.007088548503816128, + 0.011028492823243141, + 0.009331393986940384, + 0.003062986070290208, + 0.0032087499275803566, + 0.008486359380185604, + 0.006896919570863247, + 0.006168000400066376, + 0.004473463166505098, + 0.0054231383837759495, + 0.004532193299382925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.020602276548743248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198238030076027 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198238030076027, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198238030076027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5374810695648193, + 0.04306478798389435, + 0.02836364321410656, + 0.021405905485153198, + 0.052214909344911575, + 0.0222949106246233, + 0.061286237090826035, + 0.0314156748354435, + 0.17060139775276184, + 0.03187147155404091 + ], + "credit_map": [ + 0.030893752351403236, + 0.025766219943761826, + 0.05667810142040253, + 0.16847816109657288, + 0.025317376479506493, + 0.1533714234828949, + 0.037773918360471725, + 0.03920214995741844, + 0.004332520999014378, + 0.004663657862693071, + 0.007128993049263954, + 0.008054887875914574, + 0.00528109772130847, + 0.004982598591595888, + 0.0042147268541157246, + 0.004406304564327002, + 0.004019510000944138, + 0.0025344770401716232, + 0.005803764797747135, + 0.003320151474326849, + 0.003205664688721299, + 0.0036680905614048243, + 0.00262687006033957, + 0.003185112727805972, + 0.0017041824758052826, + 0.002491994760930538, + 0.002743219956755638, + 0.003918771166354418, + 0.003653845516964793, + 0.002447444247081876, + 0.0027454805094748735, + 0.0017009665025398135, + 0.0048556262627244, + 0.003491793991997838, + 0.01136784814298153, + 0.005348674021661282, + 0.004634632728993893, + 0.012914218939840794, + 0.0049761878326535225, + 0.004625930450856686, + 0.003127774689346552, + 0.0012840500567108393, + 0.0037612321320921183, + 0.0031912748236209154, + 0.001532921683974564, + 0.003483828390017152, + 0.0032594932708889246, + 0.0026543359272181988, + 0.009378906339406967, + 0.005858284421265125, + 0.008758549578487873, + 0.006659944541752338, + 0.00976900290697813, + 0.006030924618244171, + 0.006180004682391882, + 0.008650621399283409, + 0.0037427942734211683, + 0.0029926879797130823, + 0.004114597104489803, + 0.0038781336043030024, + 0.0036419336684048176, + 0.003491937881335616, + 0.004395797848701477, + 0.00515779061242938, + 0.01226761844009161, + 0.012925850227475166, + 0.021313047036528587, + 0.015065322630107403, + 0.04469882324337959, + 0.028984298929572105, + 0.0200559850782156, + 0.015290462411940098, + 0.00223046843893826, + 0.0028519416227936745, + 0.0037296602968126535, + 0.003995557781308889, + 0.006238527595996857, + 0.00565260648727417, + 0.003600150579586625, + 0.00357256131246686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.014198238030076027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066960990428925 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066960990428925, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066960990428925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1592692881822586, + 0.06296917796134949, + 0.09389396011829376, + 0.03924711421132088, + 0.20746636390686035, + 0.1380554884672165, + 0.04929909482598305, + 0.12845714390277863, + 0.08901823312044144, + 0.03232416883111 + ], + "credit_map": [ + 0.011190003715455532, + 0.008503618650138378, + 0.019539758563041687, + 0.03537966310977936, + 0.010969256982207298, + 0.035447701811790466, + 0.013714920729398727, + 0.02452436089515686, + 0.006646865978837013, + 0.004167630337178707, + 0.009160250425338745, + 0.010690771043300629, + 0.0062827761285007, + 0.0123876528814435, + 0.006300510372966528, + 0.007332716137170792, + 0.008409877307713032, + 0.006555565632879734, + 0.01988762430846691, + 0.014066725969314575, + 0.012141696177423, + 0.011144551448523998, + 0.009917519986629486, + 0.011770389974117279, + 0.0054741124622523785, + 0.0035250878427177668, + 0.005425688810646534, + 0.006113728508353233, + 0.004800281021744013, + 0.004995738621801138, + 0.004372822120785713, + 0.00453965226188302, + 0.010741416364908218, + 0.009350338019430637, + 0.06898434460163116, + 0.02059388905763626, + 0.012448832392692566, + 0.060587674379348755, + 0.011016903445124626, + 0.013742955401539803, + 0.018348706886172295, + 0.008841088972985744, + 0.0287565179169178, + 0.015927119180560112, + 0.012123079039156437, + 0.023396814242005348, + 0.01555889192968607, + 0.015103260055184364, + 0.003071361454203725, + 0.0037611296866089106, + 0.007840120233595371, + 0.004808644764125347, + 0.011480201035737991, + 0.008400036953389645, + 0.005133120343089104, + 0.004804477095603943, + 0.015908632427453995, + 0.013055885210633278, + 0.015258884988725185, + 0.016277968883514404, + 0.013848813250660896, + 0.018767286092042923, + 0.01498423982411623, + 0.02035542018711567, + 0.008632085286080837, + 0.006960604339838028, + 0.01049010455608368, + 0.007881895639002323, + 0.022351913154125214, + 0.015624213963747025, + 0.009750723838806152, + 0.007326687686145306, + 0.0021204580552875996, + 0.002980307675898075, + 0.002909076167270541, + 0.00397079112008214, + 0.005905890371650457, + 0.005822158418595791, + 0.004037851933389902, + 0.004577632062137127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.016066960990428925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795659705996513 + ], + "opposite_logprob_tokens": [ + -3.5297956466674805 + ], + "logprob_reward": -0.029795659705996513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795659705996513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30485934019088745, + 0.07108833640813828, + 0.03789139539003372, + 0.04318128898739815, + 0.039807260036468506, + 0.1808830201625824, + 0.12569227814674377, + 0.04550405219197273, + 0.10944557189941406, + 0.041647396981716156 + ], + "credit_map": [ + 0.020039713010191917, + 0.015400510281324387, + 0.030998704954981804, + 0.08441530168056488, + 0.019390448927879333, + 0.08434119820594788, + 0.023335399106144905, + 0.026938093826174736, + 0.0045326161198318005, + 0.004716688301414251, + 0.011829507537186146, + 0.011238007806241512, + 0.006692796945571899, + 0.01759832724928856, + 0.0070500122383236885, + 0.007430385332554579, + 0.0026952556800097227, + 0.0030579722952097654, + 0.00871602725237608, + 0.006905870046466589, + 0.005362797528505325, + 0.002710820874199271, + 0.004779880866408348, + 0.0036627757363021374, + 0.002507734578102827, + 0.005160290747880936, + 0.004692854825407267, + 0.007943318225443363, + 0.009539240971207619, + 0.004793810658156872, + 0.0055675501935184, + 0.002976491814479232, + 0.003236922901123762, + 0.004457317758351564, + 0.005916673690080643, + 0.00638003321364522, + 0.005274849012494087, + 0.0048455302603542805, + 0.0069907414726912975, + 0.002705194288864732, + 0.008765327744185925, + 0.004861986730247736, + 0.06213550642132759, + 0.014514011330902576, + 0.011265857145190239, + 0.047746237367391586, + 0.020283499732613564, + 0.011310625821352005, + 0.0077512688003480434, + 0.009274166077375412, + 0.01871911622583866, + 0.018083512783050537, + 0.02677583508193493, + 0.019468870013952255, + 0.01180296577513218, + 0.013816572725772858, + 0.0037491843104362488, + 0.004039200488477945, + 0.005747090093791485, + 0.00645601749420166, + 0.007287651766091585, + 0.006711867172271013, + 0.00626105023548007, + 0.005251995287835598, + 0.007034209556877613, + 0.009172366000711918, + 0.01468036137521267, + 0.01292975153774023, + 0.019561877474188805, + 0.017211392521858215, + 0.0156673826277256, + 0.01318824477493763, + 0.0028604785911738873, + 0.0035403131041675806, + 0.0074358247220516205, + 0.006070673931390047, + 0.0064651803113520145, + 0.004235952161252499, + 0.005597841460257769, + 0.005441140849143267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.029795659705996513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An elegant, sweeping, high-angle drone shot gracefully tracks a vibrant metallic red Corvette C8 Stingray, its signature mid-engine silhouette defined by sharp aerodynamic lines and large side air intakes, as it effortlessly cruises along a winding coastal highway. The entire scene is bathed in the warm, saturated light of golden hour, with the low-hanging sun casting long, dramatic shadows stretching across the smooth asphalt. The intense golden glow gleams off the car's immaculate, highly polished paint and multi-spoke alloy wheels. On one side of the road, rugged, sun-bathed cliffs rise dramatically, their earthy hues gleaming in the afternoon sun. On the other side, a vast expanse of deep blue ocean shimmers endlessly toward the horizon, gentle white waves lapping against the rocky coastline below. The Corvette navigates the road's graceful S-curves with smooth, powerful precision, its motion creating a breathtaking aura of freedom and high-performance elegance.", + "seed": 11511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_8/index_2556_seed_11511_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022703537531197072, + "vlm_logprob": -0.022703537531197072, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.022703537531197072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02396811544895172 + ], + "opposite_logprob_tokens": [ + -4.27396821975708 + ], + "logprob_reward": -0.02396811544895172, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02396811544895172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1367858201265335, + 0.07804601639509201, + 0.04550044611096382, + 0.12069039791822433, + 0.08427097648382187, + 0.04711851105093956, + 0.05426955595612526, + 0.13745035231113434, + 0.19128447771072388, + 0.1045834869146347 + ], + "credit_map": [ + 0.022209806367754936, + 0.01563234254717827, + 0.026363054290413857, + 0.017238594591617584, + 0.00963319931179285, + 0.008924768306314945, + 0.008756018243730068, + 0.028028031811118126, + 0.00786617398262024, + 0.007634483277797699, + 0.01581747457385063, + 0.012839430011808872, + 0.007058258168399334, + 0.0065077729523181915, + 0.009205400943756104, + 0.011117011308670044, + 0.0032824473455548286, + 0.0023271727841347456, + 0.00656397920101881, + 0.007791178300976753, + 0.005866880062967539, + 0.007171063218265772, + 0.006204706151038408, + 0.00629301555454731, + 0.0077336858958005905, + 0.004196584224700928, + 0.04054750129580498, + 0.009013347327709198, + 0.006652554962784052, + 0.031219953671097755, + 0.009416451677680016, + 0.01191030815243721, + 0.003929793369024992, + 0.007246409077197313, + 0.014353277161717415, + 0.010835262946784496, + 0.016202783212065697, + 0.011292407289147377, + 0.013354186899960041, + 0.007056848146021366, + 0.004035626538097858, + 0.00880279392004013, + 0.005345778074115515, + 0.006068470422178507, + 0.00969651248306036, + 0.0034634640906006098, + 0.005420027766376734, + 0.004285831935703754, + 0.004456175025552511, + 0.007857903838157654, + 0.005060083232820034, + 0.0066059608943760395, + 0.009001518599689007, + 0.004717840813100338, + 0.010509634390473366, + 0.006060435436666012, + 0.008584365248680115, + 0.004269177559763193, + 0.006314999423921108, + 0.005762821063399315, + 0.009668867103755474, + 0.03732158616185188, + 0.01949635148048401, + 0.046032171696424484, + 0.008879469707608223, + 0.007112764287739992, + 0.008692894130945206, + 0.014056927524507046, + 0.014605354517698288, + 0.06528989970684052, + 0.05570661649107933, + 0.016940545290708542, + 0.006633359007537365, + 0.007062390446662903, + 0.008173724636435509, + 0.012027840130031109, + 0.015835419297218323, + 0.007288950029760599, + 0.02002389170229435, + 0.027537906542420387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02396811544895172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964101292192936 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964101292192936, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964101292192936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09685607999563217, + 0.09177367389202118, + 0.04425371065735817, + 0.4050443172454834, + 0.06836909055709839, + 0.03466744348406792, + 0.06062214449048042, + 0.021872628480196, + 0.11451355367898941, + 0.062027353793382645 + ], + "credit_map": [ + 0.008227233774960041, + 0.012960651889443398, + 0.01861722394824028, + 0.012788841500878334, + 0.012462662532925606, + 0.007590531371533871, + 0.007231625262647867, + 0.01697731390595436, + 0.010754711925983429, + 0.008530108258128166, + 0.017821695655584335, + 0.011060931719839573, + 0.007562330458313227, + 0.005903964396566153, + 0.01070066262036562, + 0.019439272582530975, + 0.004986390471458435, + 0.0036504825111478567, + 0.007180023938417435, + 0.004929648712277412, + 0.00488108117133379, + 0.006665117107331753, + 0.00618597911670804, + 0.005774992518126965, + 0.0233779214322567, + 0.01828087493777275, + 0.13357196748256683, + 0.03222040832042694, + 0.03360573202371597, + 0.0806412398815155, + 0.02779337577521801, + 0.05555284023284912, + 0.0033742922823876143, + 0.005870725028216839, + 0.013714585453271866, + 0.006978793069720268, + 0.013662456534802914, + 0.0116485096514225, + 0.007542245555669069, + 0.00557748693972826, + 0.0024661060888320208, + 0.003322555683553219, + 0.004440405406057835, + 0.0031455953139811754, + 0.00707379262894392, + 0.005058587063103914, + 0.004052282311022282, + 0.005108123179525137, + 0.0047529879957437515, + 0.012821466661989689, + 0.005475964862853289, + 0.004927137866616249, + 0.01440398022532463, + 0.004815888591110706, + 0.007774594239890575, + 0.00565012963488698, + 0.0013358937576413155, + 0.0029528921004384756, + 0.002212103921920061, + 0.0033674510195851326, + 0.003990595694631338, + 0.002167593454942107, + 0.0029302772600203753, + 0.002915821736678481, + 0.004699822515249252, + 0.004246026277542114, + 0.005785084795206785, + 0.008118411526083946, + 0.008599120192229748, + 0.02277926541864872, + 0.037561509758234024, + 0.022724324837327003, + 0.005966922268271446, + 0.007126528304070234, + 0.005924561060965061, + 0.007464287336915731, + 0.010278590954840183, + 0.00771610951051116, + 0.008730691857635975, + 0.00881966669112444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.005964101292192936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634631097316742 + ], + "opposite_logprob_tokens": [ + -3.651346206665039 + ], + "logprob_reward": -0.02634631097316742, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634631097316742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05729629844427109, + 0.08611155301332474, + 0.13273528218269348, + 0.2556816339492798, + 0.12624560296535492, + 0.03688375651836395, + 0.01912854053080082, + 0.0760503038764, + 0.07744751125574112, + 0.13241952657699585 + ], + "credit_map": [ + 0.009830772876739502, + 0.006376529578119516, + 0.007800368592143059, + 0.007536973338574171, + 0.0060067991726100445, + 0.005073009990155697, + 0.006745476741343737, + 0.007926367223262787, + 0.006257574539631605, + 0.006914911326020956, + 0.018245993182063103, + 0.017132535576820374, + 0.009938612580299377, + 0.008084727451205254, + 0.011450204066932201, + 0.008086996152997017, + 0.017429780215024948, + 0.009893964976072311, + 0.024688711389899254, + 0.016068413853645325, + 0.017619099467992783, + 0.01888267509639263, + 0.015875376760959625, + 0.012277249246835709, + 0.012912114150822163, + 0.007672553416341543, + 0.09780915081501007, + 0.012868804857134819, + 0.00801186915487051, + 0.08595769852399826, + 0.008638083934783936, + 0.0218113474547863, + 0.012587846256792545, + 0.020032191649079323, + 0.01834113895893097, + 0.012724128551781178, + 0.01730705238878727, + 0.01947840116918087, + 0.013364026322960854, + 0.012410828843712807, + 0.0020446565467864275, + 0.0030434574000537395, + 0.00466657429933548, + 0.0033800029195845127, + 0.010492960922420025, + 0.004607850685715675, + 0.004342973232269287, + 0.0043052807450294495, + 0.0017128274776041508, + 0.004836721811443567, + 0.002060499507933855, + 0.0020188831258565187, + 0.0031148274429142475, + 0.0013040074845775962, + 0.002313910750672221, + 0.00176686339545995, + 0.004112031310796738, + 0.003938448149710894, + 0.003540162229910493, + 0.0051446412689983845, + 0.006543006747961044, + 0.018882883712649345, + 0.010917154140770435, + 0.022971972823143005, + 0.004015337210148573, + 0.0030140141025185585, + 0.0035211555659770966, + 0.00585866067558527, + 0.005896236281841993, + 0.02803734317421913, + 0.017976583912968636, + 0.009128178469836712, + 0.007538802921772003, + 0.010463605634868145, + 0.00780488271266222, + 0.007450842764228582, + 0.0171393733471632, + 0.02121718041598797, + 0.024284709244966507, + 0.036520130932331085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.02634631097316742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190834015607834 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190834015607834, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190834015607834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1277407705783844, + 0.09747201204299927, + 0.18512733280658722, + 0.27911242842674255, + 0.05711836740374565, + 0.02334793657064438, + 0.016930369660258293, + 0.09284024685621262, + 0.052269067615270615, + 0.06804147362709045 + ], + "credit_map": [ + 0.021464217454195023, + 0.013085691258311272, + 0.021593524143099785, + 0.01958366669714451, + 0.010911800898611546, + 0.009887142106890678, + 0.009798669256269932, + 0.02141605131328106, + 0.01060019712895155, + 0.007839508354663849, + 0.021286889910697937, + 0.018251702189445496, + 0.010301541537046432, + 0.005737685598433018, + 0.011732815764844418, + 0.011721671558916569, + 0.020767370238900185, + 0.018465857952833176, + 0.029120586812496185, + 0.02294020541012287, + 0.025349026545882225, + 0.024538390338420868, + 0.02661707065999508, + 0.017328836023807526, + 0.01659761182963848, + 0.011246905662119389, + 0.08904893696308136, + 0.015422679483890533, + 0.017828330397605896, + 0.09225206822156906, + 0.013001526705920696, + 0.02371436357498169, + 0.004222494550049305, + 0.007972730323672295, + 0.009451192803680897, + 0.004416611045598984, + 0.012595956213772297, + 0.009234171360731125, + 0.0047495705075562, + 0.004475635942071676, + 0.002037612022832036, + 0.003867741907015443, + 0.0035936743952333927, + 0.00278649409301579, + 0.003755066078156233, + 0.0018006681930273771, + 0.003033182118088007, + 0.002473497064784169, + 0.0017928988672792912, + 0.00301525741815567, + 0.0017001911764964461, + 0.001728466828353703, + 0.0023907627910375595, + 0.001527972985059023, + 0.0024047621991485357, + 0.0023700580932199955, + 0.004996306262910366, + 0.003455297788605094, + 0.0025672330521047115, + 0.0044040982611477375, + 0.006203619297593832, + 0.025479832664132118, + 0.013410738669335842, + 0.032323118299245834, + 0.002840409753844142, + 0.002368456218391657, + 0.0020066292490810156, + 0.0028364118188619614, + 0.0038331502582877874, + 0.01694491133093834, + 0.006810794584453106, + 0.014628306031227112, + 0.004097587428987026, + 0.005461378488689661, + 0.0035679107531905174, + 0.0043866937048733234, + 0.007453257218003273, + 0.015154204331338406, + 0.010075959376990795, + 0.01784447766840458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.014190834015607834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a metallic red Corvette C8 Stingray with a mid-engine silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving along a winding coastal highway with smooth, powerful motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by the warm, saturated light of golden hour with long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature rugged sun-bathed cliffs on one side and a deep blue ocean on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a high-angle drone perspective that tracks the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043048325926065445 + ], + "opposite_logprob_tokens": [ + -3.16804838180542 + ], + "logprob_reward": -0.043048325926065445, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043048325926065445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03848252817988396, + 0.037289757281541824, + 0.0385965034365654, + 0.5478346943855286, + 0.04013240709900856, + 0.03475023806095123, + 0.015433198772370815, + 0.09068135917186737, + 0.11374076455831528, + 0.04305851459503174 + ], + "credit_map": [ + 0.004812374245375395, + 0.005355606321245432, + 0.007162246387451887, + 0.004107002168893814, + 0.0032934851478785276, + 0.0031211753375828266, + 0.0028610185254365206, + 0.0077696191146969795, + 0.0035835150629281998, + 0.002975698094815016, + 0.007642786018550396, + 0.005284413229674101, + 0.003393088933080435, + 0.003274047514423728, + 0.004060559906065464, + 0.007075647823512554, + 0.0029113460332155228, + 0.0033303319942206144, + 0.00551726296544075, + 0.005431196186691523, + 0.007163988891988993, + 0.006683950312435627, + 0.0038130381144583225, + 0.0037453908007591963, + 0.02644522860646248, + 0.015125240199267864, + 0.2228529006242752, + 0.02824372984468937, + 0.013274146243929863, + 0.17167474329471588, + 0.015786312520503998, + 0.05443241074681282, + 0.0032857104670256376, + 0.002900714287534356, + 0.006633277051150799, + 0.004688679706305265, + 0.0059841712936758995, + 0.004985498730093241, + 0.00726604089140892, + 0.004388314206153154, + 0.003201783401891589, + 0.006001298315823078, + 0.004277267958968878, + 0.003609401872381568, + 0.006414539180696011, + 0.002403407357633114, + 0.00421448890119791, + 0.004628049209713936, + 0.0018989742966368794, + 0.001381470588967204, + 0.0021325345151126385, + 0.002088806126266718, + 0.0016773870447650552, + 0.0013383530313149095, + 0.0026121477130800486, + 0.0023035251069813967, + 0.005014684051275253, + 0.0029395725578069687, + 0.0023270866367965937, + 0.0041394964791834354, + 0.00597303407266736, + 0.02549532800912857, + 0.012970011681318283, + 0.03182215243577957, + 0.004977790173143148, + 0.0041191657073795795, + 0.006174742244184017, + 0.008702073246240616, + 0.008526131510734558, + 0.03502199798822403, + 0.033851128071546555, + 0.012367739342153072, + 0.0028778824489563704, + 0.002915232442319393, + 0.0015021066647022963, + 0.002148654544726014, + 0.003803602419793606, + 0.005092231556773186, + 0.010009640827775002, + 0.014709161594510078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.043048325926065445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_tensor([9491, 9491])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_tensor([9491, 9491])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9503bd7eaf718c4231b40c0c454c1fa3a0069808 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_tensor([9491, 9491])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, high-angle drone shot tracks smoothly alongside a lone rider clad in full, matte-black futuristic motocross armor, its articulated composite panels gleaming subtly in the sunlight. The rider's posture is focused and low on a powerful, angular dirt bike, a matching matte black and gunmetal gray machine with exposed, heavy-duty suspension. As they relentlessly traverse the landscape, the bike's oversized rear tire aggressively kicks up a thick, billowing cloud of ochre dust that hangs suspended in the air. The camera follows, capturing their swift and decisive navigation across the rugged terrain: a vast, sun-scorched wasteland stretching to a jagged mesa on the distant horizon. The entire scene is bathed in the warm, harsh light of the late afternoon sun, casting long, dramatic shadows and brightly backlit, swirling clouds of dust, creating a golden haze against the deep blue sky. The warm tones, predominantly earthy reds and browns, contrast sharply with the rider's dark, anonymous silhouette, evoking a powerful atmosphere of solitary, high-speed determination on this epic, unfolding journey.", + "seed": 24986, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_9491_seed_24986_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08025979667436331, + "vlm_logprob": -0.08025979667436331, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.08025979667436331, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18550992012023926 + ], + "opposite_logprob_tokens": [ + -1.8105099201202393 + ], + "logprob_reward": -0.18550992012023926, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18550992012023926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0942024290561676, + 0.04510214179754257, + 0.027475543320178986, + 0.08779360353946686, + 0.0662994384765625, + 0.22387449443340302, + 0.20399276912212372, + 0.11335348337888718, + 0.04845643788576126, + 0.08944965898990631 + ], + "credit_map": [ + 0.006408278364688158, + 0.008667897433042526, + 0.013234983198344707, + 0.017583170905709267, + 0.009188421070575714, + 0.013099041767418385, + 0.013237850740551949, + 0.012782789766788483, + 0.002218185691162944, + 0.0035698269493877888, + 0.007246172521263361, + 0.0053863427601754665, + 0.003755672834813595, + 0.007064059842377901, + 0.0074235741049051285, + 0.008438311517238617, + 0.002729971893131733, + 0.004108946770429611, + 0.00418452825397253, + 0.0037850039079785347, + 0.0023547131568193436, + 0.003785226959735155, + 0.0037580744829028845, + 0.0027690797578543425, + 0.007557975593954325, + 0.006392245180904865, + 0.01599343866109848, + 0.00981680303812027, + 0.01173515897244215, + 0.01835496537387371, + 0.008742929436266422, + 0.009200090542435646, + 0.005753668956458569, + 0.004401186015456915, + 0.010391909629106522, + 0.006541321985423565, + 0.007025144062936306, + 0.01062134187668562, + 0.009558646939694881, + 0.012006229721009731, + 0.009300545789301395, + 0.0061144898645579815, + 0.010280292481184006, + 0.008592583239078522, + 0.08663860708475113, + 0.08829537034034729, + 0.007376326713711023, + 0.007276294287294149, + 0.020553555339574814, + 0.026144837960600853, + 0.015706414356827736, + 0.02393535152077675, + 0.050217319279909134, + 0.03257058560848236, + 0.018258769065141678, + 0.016605960205197334, + 0.008216159418225288, + 0.012989349663257599, + 0.015282235108315945, + 0.026438463479280472, + 0.010964171029627323, + 0.010932201519608498, + 0.015568464063107967, + 0.0129624605178833, + 0.004024907946586609, + 0.005160134751349688, + 0.006380162667483091, + 0.0077631911262869835, + 0.008479351177811623, + 0.005418960005044937, + 0.005793184041976929, + 0.005436548963189125, + 0.007511375471949577, + 0.0050646234303712845, + 0.007708920165896416, + 0.00740527780726552, + 0.027291756123304367, + 0.023715270683169365, + 0.005214782897382975, + 0.005537663120776415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.18550992012023926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641253061592579 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641253061592579, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641253061592579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16796043515205383, + 0.0914110466837883, + 0.04397638142108917, + 0.09183329343795776, + 0.04281420633196831, + 0.3108865022659302, + 0.05464242771267891, + 0.06601110100746155, + 0.08562882989645004, + 0.04483575373888016 + ], + "credit_map": [ + 0.010780161246657372, + 0.018766717985272408, + 0.025652192533016205, + 0.024828432127833366, + 0.015988348051905632, + 0.024714099243283272, + 0.022442422807216644, + 0.024788055568933487, + 0.007990892045199871, + 0.008811526000499725, + 0.010752109810709953, + 0.011461274698376656, + 0.01102654729038477, + 0.011964681558310986, + 0.014569946564733982, + 0.014834071509540081, + 0.0038797776214778423, + 0.006858263164758682, + 0.007455157116055489, + 0.005499567370861769, + 0.003798454999923706, + 0.005408854223787785, + 0.006379670929163694, + 0.0046966345980763435, + 0.0081289391964674, + 0.007693124003708363, + 0.016119640320539474, + 0.008562960661947727, + 0.012758045457303524, + 0.02142101153731346, + 0.00794676598161459, + 0.009202804416418076, + 0.003960827831178904, + 0.0026235918048769236, + 0.007147268857806921, + 0.0038968948647379875, + 0.004914259072393179, + 0.007678908295929432, + 0.005051426123827696, + 0.0075410278514027596, + 0.012103808112442493, + 0.006477500312030315, + 0.00957520306110382, + 0.008097795769572258, + 0.1261957585811615, + 0.12999841570854187, + 0.00905215460807085, + 0.009385871700942516, + 0.004765875171869993, + 0.006724529433995485, + 0.006316387094557285, + 0.006110229063779116, + 0.00948651134967804, + 0.007697190158069134, + 0.005018273834139109, + 0.008523432537913322, + 0.004709410481154919, + 0.0057327947579324245, + 0.007094628177583218, + 0.014180931262671947, + 0.008957834914326668, + 0.007287605665624142, + 0.007735657971352339, + 0.01031224150210619, + 0.007496797479689121, + 0.010926253162324429, + 0.010282471776008606, + 0.01350387278944254, + 0.01525264885276556, + 0.009872650727629662, + 0.009773125872015953, + 0.008521010167896748, + 0.003744716290384531, + 0.002474797423928976, + 0.0037721244152635336, + 0.004127694759517908, + 0.013138647191226482, + 0.011586727574467659, + 0.0033147032372653484, + 0.0026763437781482935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.007641253061592579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019461041083559394 + ], + "opposite_logprob_tokens": [ + -6.251945972442627 + ], + "logprob_reward": -0.0019461041083559394, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019461041083559394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22611762583255768, + 0.0396495945751667, + 0.03319394215941429, + 0.06123695895075798, + 0.06816557794809341, + 0.38806989789009094, + 0.055549174547195435, + 0.031914014369249344, + 0.02505185827612877, + 0.07105135917663574 + ], + "credit_map": [ + 0.010220816358923912, + 0.011733608320355415, + 0.04322366043925285, + 0.015711769461631775, + 0.020355386659502983, + 0.03355924412608147, + 0.03735519200563431, + 0.05395795404911041, + 0.0024182144552469254, + 0.003018298652023077, + 0.004974732641130686, + 0.0033305156975984573, + 0.005852420348674059, + 0.0077676610089838505, + 0.005344417877495289, + 0.006943336222320795, + 0.002619626000523567, + 0.003097031032666564, + 0.004890862852334976, + 0.0037259033415466547, + 0.004997605457901955, + 0.005372330546379089, + 0.004707359243184328, + 0.0037832215894013643, + 0.0048311143182218075, + 0.005005254875868559, + 0.009546219371259212, + 0.005195156205445528, + 0.007637473288923502, + 0.012588337995111942, + 0.00667565455660224, + 0.00975775346159935, + 0.006993637885898352, + 0.0035540482494980097, + 0.010487821884453297, + 0.005388295743614435, + 0.010380026884377003, + 0.016847915947437286, + 0.006077978294342756, + 0.008435847237706184, + 0.013279715552926064, + 0.00913249421864748, + 0.012095054611563683, + 0.015710990875959396, + 0.15217554569244385, + 0.15678270161151886, + 0.015322321094572544, + 0.013571033254265785, + 0.004678521305322647, + 0.0051267207600176334, + 0.007004283834248781, + 0.006071534473448992, + 0.010084105655550957, + 0.00831271056085825, + 0.00509867025539279, + 0.009172623977065086, + 0.0026392508298158646, + 0.003641747171059251, + 0.0030554148834198713, + 0.0047232406213879585, + 0.005265960469841957, + 0.004042292013764381, + 0.0034080049954354763, + 0.005138104781508446, + 0.002002227585762739, + 0.0023993931245058775, + 0.0019314682576805353, + 0.003367307595908642, + 0.007402324583381414, + 0.004213117528706789, + 0.0020334513392299414, + 0.0017025676788762212, + 0.0043569086119532585, + 0.002175799570977688, + 0.004647674970328808, + 0.002944502281025052, + 0.02658289298415184, + 0.02343675307929516, + 0.0032726626377552748, + 0.00363416550680995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0019461041083559394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633504942059517 + ], + "opposite_logprob_tokens": [ + -3.6513350009918213 + ], + "logprob_reward": -0.02633504942059517, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633504942059517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1062757596373558, + 0.25526943802833557, + 0.04321643337607384, + 0.10550349205732346, + 0.035906948149204254, + 0.1763521432876587, + 0.04211421310901642, + 0.04776717722415924, + 0.027473054826259613, + 0.16012132167816162 + ], + "credit_map": [ + 0.005750179756432772, + 0.007242395076900721, + 0.017969293519854546, + 0.008483603596687317, + 0.009558670222759247, + 0.017083458602428436, + 0.021997788920998573, + 0.018190380185842514, + 0.01819758675992489, + 0.020236101001501083, + 0.024894287809729576, + 0.012024291791021824, + 0.04022781550884247, + 0.0449288934469223, + 0.03470734879374504, + 0.06005312502384186, + 0.0030765479896217585, + 0.005258023273199797, + 0.008847976103425026, + 0.004283415153622627, + 0.00410252483561635, + 0.0051952446810901165, + 0.006041564047336578, + 0.006411141715943813, + 0.00850630085915327, + 0.007923711091279984, + 0.017793577164411545, + 0.00722276559099555, + 0.016107384115457535, + 0.029418107122182846, + 0.008288304321467876, + 0.010243344120681286, + 0.0033274840097874403, + 0.0020189599599689245, + 0.0057861157692968845, + 0.0028267488814890385, + 0.004223214462399483, + 0.007364911027252674, + 0.003975701984018087, + 0.006383815314620733, + 0.009152084589004517, + 0.004380709491670132, + 0.008245191536843777, + 0.004947964567691088, + 0.06779599189758301, + 0.06998851150274277, + 0.005273187533020973, + 0.0065685114823281765, + 0.0037067399825900793, + 0.004055891185998917, + 0.004435044247657061, + 0.004996181465685368, + 0.009489224292337894, + 0.006531869526952505, + 0.0035662197042256594, + 0.005333046428859234, + 0.005096713546663523, + 0.004938764031976461, + 0.00595741905272007, + 0.004406880587339401, + 0.008930860087275505, + 0.006624776870012283, + 0.004866050090640783, + 0.006945719011127949, + 0.0021126028150320053, + 0.002659441437572241, + 0.0025115276221185923, + 0.0037946519441902637, + 0.007314582820981741, + 0.004248383454978466, + 0.002488597994670272, + 0.002343268133699894, + 0.009203101508319378, + 0.004101371392607689, + 0.010137781500816345, + 0.004961212631314993, + 0.06170957535505295, + 0.055692512542009354, + 0.006490515545010567, + 0.007825261913239956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.02633504942059517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986665666103363 + ], + "opposite_logprob_tokens": [ + -1.8048666715621948 + ], + "logprob_reward": -0.17986665666103363, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986665666103363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031062781810760498, + 0.036687642335891724, + 0.027192607522010803, + 0.07373819500207901, + 0.06430011987686157, + 0.5877805948257446, + 0.11732436716556549, + 0.012627382762730122, + 0.010705074295401573, + 0.038581233471632004 + ], + "credit_map": [ + 0.002236324129626155, + 0.00217594881542027, + 0.0048540313728153706, + 0.0033651995472609997, + 0.0028589414432644844, + 0.005504133645445108, + 0.005567967426031828, + 0.004500235430896282, + 0.0030106701888144016, + 0.0033620537724345922, + 0.004864167887717485, + 0.004402425140142441, + 0.0033062894362956285, + 0.006031371187418699, + 0.0057130856439471245, + 0.005997577216476202, + 0.002289731753990054, + 0.003289641812443733, + 0.004329031333327293, + 0.0033957718405872583, + 0.0028536284808069468, + 0.0036407485604286194, + 0.003855803981423378, + 0.003538250457495451, + 0.006167543586343527, + 0.005470234900712967, + 0.012917869724333286, + 0.007419859059154987, + 0.010203643701970577, + 0.016454750671982765, + 0.007453951518982649, + 0.007650342769920826, + 0.0054271770641207695, + 0.00402887724339962, + 0.009029502049088478, + 0.005817592144012451, + 0.007429257035255432, + 0.01307880599051714, + 0.008895428851246834, + 0.010593478567898273, + 0.02431274764239788, + 0.01308067049831152, + 0.027603361755609512, + 0.01812702789902687, + 0.2299700379371643, + 0.23634129762649536, + 0.01891310326755047, + 0.01943233422935009, + 0.010447059758007526, + 0.013028732500970364, + 0.00811921339482069, + 0.019161788746714592, + 0.02444973960518837, + 0.017794933170080185, + 0.012429659254848957, + 0.01189324539154768, + 0.0010193580528721213, + 0.001218520337715745, + 0.0019099210621789098, + 0.002347006695345044, + 0.001311805215664208, + 0.0012680236250162125, + 0.0017647941131144762, + 0.0017879537772387266, + 0.0010206144070252776, + 0.0015245034592226148, + 0.001217551063746214, + 0.0016238895477727056, + 0.0015770216705277562, + 0.0012530499370768666, + 0.0012735590571537614, + 0.001214885269291699, + 0.00146663305349648, + 0.0011387558188289404, + 0.0018052684608846903, + 0.0014007827267050743, + 0.01544345822185278, + 0.014681419357657433, + 0.0013473286526277661, + 0.001297584269195795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.17986665666103363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle drone shot tracks smoothly alongside a lone rider clad in full, matte-black futuristic motocross armor, its articulated composite panels gleaming subtly in the sunlight. The rider's posture is focused and low on a powerful, angular dirt bike, a matching matte black and gunmetal gray machine with exposed, heavy-duty suspension. As they relentlessly traverse the landscape, the bike's oversized rear tire aggressively kicks up a thick, billowing cloud of ochre dust that hangs suspended in the air. The camera follows, capturing their swift and decisive navigation across the rugged terrain: a vast, sun-scorched wasteland stretching to a jagged mesa on the distant horizon. The entire scene is bathed in the warm, harsh light of the late afternoon sun, casting long, dramatic shadows and brightly backlit, swirling clouds of dust, creating a golden haze against the deep blue sky. The warm tones, predominantly earthy reds and browns, contrast sharply with the rider's dark, anonymous silhouette, evoking a powerful atmosphere of solitary, high-speed determination on this epic, unfolding journey.", + "seed": 24987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_9491_seed_24987_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.043670110253151505, + "vlm_logprob": -0.043670110253151505, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.043670110253151505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02609931491315365 + ], + "opposite_logprob_tokens": [ + -3.77609920501709 + ], + "logprob_reward": -0.02609931491315365, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02609931491315365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17502592504024506, + 0.11896820366382599, + 0.07298466563224792, + 0.12126551568508148, + 0.1708202064037323, + 0.14901615679264069, + 0.056330498307943344, + 0.06508199870586395, + 0.02060617506504059, + 0.049900613725185394 + ], + "credit_map": [ + 0.017228716984391212, + 0.01800278201699257, + 0.02763681858778, + 0.02485029399394989, + 0.015099925920367241, + 0.017920417711138725, + 0.034281015396118164, + 0.0200059674680233, + 0.012376588769257069, + 0.019908780232071877, + 0.017136240378022194, + 0.013826203532516956, + 0.01585289277136326, + 0.009924277663230896, + 0.016904307529330254, + 0.01303892582654953, + 0.006748097017407417, + 0.01096431165933609, + 0.010453733615577221, + 0.009130747988820076, + 0.007808044087141752, + 0.005091908387839794, + 0.011533919721841812, + 0.011253911070525646, + 0.01673777960240841, + 0.013222677633166313, + 0.011996614746749401, + 0.014381440356373787, + 0.013188151642680168, + 0.011932790279388428, + 0.01430666632950306, + 0.02549940161406994, + 0.02679605968296528, + 0.015293829143047333, + 0.021973924711346626, + 0.010895654559135437, + 0.013060344383120537, + 0.04914626479148865, + 0.0095664132386446, + 0.024087710306048393, + 0.011502588167786598, + 0.012403134256601334, + 0.01810421794652939, + 0.020001262426376343, + 0.02153162658214569, + 0.01807328499853611, + 0.025956962257623672, + 0.021443087607622147, + 0.0049670846201479435, + 0.0060016387142241, + 0.008909459225833416, + 0.00695951608940959, + 0.010752405039966106, + 0.005831682123243809, + 0.006990982685238123, + 0.005917734000831842, + 0.00857858918607235, + 0.010033460333943367, + 0.0045609730295836926, + 0.006468372885137796, + 0.012549931183457375, + 0.006746902596205473, + 0.008487497456371784, + 0.007656275760382414, + 0.002080754144117236, + 0.002270370488986373, + 0.002341445768252015, + 0.0026672673411667347, + 0.003171969437971711, + 0.00232391688041389, + 0.002597369719296694, + 0.0031530815176665783, + 0.006052190903574228, + 0.006563789676874876, + 0.003985459916293621, + 0.005835612304508686, + 0.007712711580097675, + 0.007836997509002686, + 0.006335519719868898, + 0.005578337237238884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.02609931491315365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031987475231289864 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031987475231289864, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031987475231289864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06604022532701492, + 0.05774915590882301, + 0.03334754332900047, + 0.03372042626142502, + 0.5818086266517639, + 0.09666183590888977, + 0.03344893828034401, + 0.048605404794216156, + 0.028164973482489586, + 0.020452899858355522 + ], + "credit_map": [ + 0.007931312546133995, + 0.007614884525537491, + 0.010334036312997341, + 0.008357303217053413, + 0.007589899003505707, + 0.00818120688199997, + 0.008192279376089573, + 0.007839305326342583, + 0.006742753088474274, + 0.013006157241761684, + 0.005972094833850861, + 0.005957695189863443, + 0.007854578085243702, + 0.004700675141066313, + 0.007718358188867569, + 0.005796842277050018, + 0.0034584440290927887, + 0.005353446118533611, + 0.008595535531640053, + 0.0031793529633432627, + 0.0029367124661803246, + 0.002861723070964217, + 0.0032142263371497393, + 0.003748103976249695, + 0.004639317747205496, + 0.004370618611574173, + 0.004958880599588156, + 0.004646265413612127, + 0.003386562457308173, + 0.002671087859198451, + 0.003570704720914364, + 0.005476988386362791, + 0.09424297511577606, + 0.07263489067554474, + 0.09665212780237198, + 0.03226124122738838, + 0.06208516284823418, + 0.09104181081056595, + 0.02304873988032341, + 0.10984165966510773, + 0.006427457090467215, + 0.007861620746552944, + 0.011198936961591244, + 0.016456570476293564, + 0.018673546612262726, + 0.007567331660538912, + 0.019537340849637985, + 0.008939028717577457, + 0.002792007988318801, + 0.0032158605754375458, + 0.005798994563519955, + 0.005512508563697338, + 0.005198145750910044, + 0.003918492700904608, + 0.0038327891379594803, + 0.0031801399309188128, + 0.005822837818413973, + 0.007483292371034622, + 0.0035624001175165176, + 0.005618254654109478, + 0.009347215294837952, + 0.005990085192024708, + 0.005942335352301598, + 0.004838981200009584, + 0.007833989337086678, + 0.008969912305474281, + 0.0015002082800492644, + 0.0016304838936775923, + 0.0028026504442095757, + 0.002197287045419216, + 0.0014654527185484767, + 0.0017649894580245018, + 0.0024149701930582523, + 0.00276273675262928, + 0.0016549021238461137, + 0.0020128474570810795, + 0.0019870176911354065, + 0.004775294102728367, + 0.0023405691608786583, + 0.0025045631919056177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0031987475231289864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005623904871754348 + ], + "opposite_logprob_tokens": [ + -7.5005621910095215 + ], + "logprob_reward": -0.0005623904871754348, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005623904871754348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09580958634614944, + 0.06415778398513794, + 0.05112868547439575, + 0.05807606875896454, + 0.4520100951194763, + 0.028733309358358383, + 0.03739292919635773, + 0.0157844927161932, + 0.09514978528022766, + 0.10175720602273941 + ], + "credit_map": [ + 0.010452043265104294, + 0.010763355530798435, + 0.012400766834616661, + 0.009896942414343357, + 0.015696857124567032, + 0.015137375332415104, + 0.013003583997488022, + 0.008458662778139114, + 0.005908594466745853, + 0.012676442973315716, + 0.011315921321511269, + 0.00697516230866313, + 0.00765971839427948, + 0.005421532783657312, + 0.006760801654309034, + 0.0074396091513335705, + 0.006345236208289862, + 0.005063665099442005, + 0.015710584819316864, + 0.005888457875698805, + 0.004130299668759108, + 0.004124594386667013, + 0.004661307670176029, + 0.005204540677368641, + 0.005211486481130123, + 0.0044602504931390285, + 0.006117863580584526, + 0.005877729505300522, + 0.0046862089075148106, + 0.013510249555110931, + 0.004664547275751829, + 0.01354773249477148, + 0.07866784930229187, + 0.06709635257720947, + 0.07175982743501663, + 0.0277097187936306, + 0.05611976236104965, + 0.05107678845524788, + 0.019551357254385948, + 0.08002842217683792, + 0.002191603649407625, + 0.002633925061672926, + 0.003192055970430374, + 0.005484302062541246, + 0.003366092685610056, + 0.004099185112863779, + 0.0037504162173718214, + 0.004015726502984762, + 0.003968958742916584, + 0.0039022588171064854, + 0.006212266627699137, + 0.00642424589022994, + 0.004043992608785629, + 0.005318023730069399, + 0.003959130961447954, + 0.0035640534479171038, + 0.001815769704990089, + 0.002113945782184601, + 0.0014977416722103953, + 0.00251441216096282, + 0.0020199762657284737, + 0.002171536907553673, + 0.0018734647892415524, + 0.0017776458989828825, + 0.03027394972741604, + 0.03568591922521591, + 0.004804878029972315, + 0.004707238636910915, + 0.0076905363239347935, + 0.003958146553486586, + 0.003956341650336981, + 0.004072781652212143, + 0.013854901306331158, + 0.01461348868906498, + 0.00580095499753952, + 0.007245061453431845, + 0.012582911178469658, + 0.027370454743504524, + 0.010463585145771503, + 0.009825841523706913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.0005623904871754348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630713447928429 + ], + "opposite_logprob_tokens": [ + -4.758630752563477 + ], + "logprob_reward": -0.008630713447928429, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630713447928429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08885272592306137, + 0.05647788196802139, + 0.10262124240398407, + 0.06752019375562668, + 0.445570707321167, + 0.031370777636766434, + 0.09516238421201706, + 0.0325765386223793, + 0.04045255854725838, + 0.03939497470855713 + ], + "credit_map": [ + 0.007458111271262169, + 0.011245512403547764, + 0.010483729653060436, + 0.011024629697203636, + 0.009077673777937889, + 0.011480344459414482, + 0.017528051510453224, + 0.010554688051342964, + 0.00727826077491045, + 0.012484485283493996, + 0.00508912606164813, + 0.00581078976392746, + 0.008347593247890472, + 0.004326974041759968, + 0.007717978674918413, + 0.005422681104391813, + 0.013245818205177784, + 0.016549909487366676, + 0.019850898534059525, + 0.011800315231084824, + 0.009749680757522583, + 0.008554973639547825, + 0.009054524824023247, + 0.013815122656524181, + 0.007309015840291977, + 0.006682096514850855, + 0.007194484118372202, + 0.006974213756620884, + 0.005328647792339325, + 0.013655176386237144, + 0.005071653984487057, + 0.01530491467565298, + 0.07558896392583847, + 0.06306681782007217, + 0.07420256733894348, + 0.0276778731495142, + 0.048919547349214554, + 0.059965550899505615, + 0.019781358540058136, + 0.07636809349060059, + 0.003157747210934758, + 0.0035520452074706554, + 0.0031162698287516832, + 0.004761329386383295, + 0.004081111401319504, + 0.003231676062569022, + 0.0054417806677520275, + 0.004028822295367718, + 0.008928374387323856, + 0.00982456374913454, + 0.01992536149919033, + 0.015822652727365494, + 0.009063465520739555, + 0.013650425709784031, + 0.008716553449630737, + 0.00923099648207426, + 0.005660302937030792, + 0.006939226295799017, + 0.0029047131538391113, + 0.0031784174498170614, + 0.004788346588611603, + 0.003216344164684415, + 0.003092156257480383, + 0.0027970336377620697, + 0.013438197784125805, + 0.015395901165902615, + 0.0024307710118591785, + 0.0028752877842634916, + 0.0023114304058253765, + 0.0014334233710542321, + 0.0014031180180609226, + 0.0011644348269328475, + 0.00528247095644474, + 0.005133291240781546, + 0.003707664320245385, + 0.003782979678362608, + 0.003983622882515192, + 0.009773164987564087, + 0.003737029153853655, + 0.00399475684389472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.008630713447928429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985938489437103 + ], + "opposite_logprob_tokens": [ + -1.8048593997955322 + ], + "logprob_reward": -0.17985938489437103, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985938489437103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09574083238840103, + 0.038066476583480835, + 0.03728281706571579, + 0.04438291862607002, + 0.4641200602054596, + 0.05287591367959976, + 0.11379289627075195, + 0.09147126972675323, + 0.032819826155900955, + 0.02944699488580227 + ], + "credit_map": [ + 0.00717273261398077, + 0.008404741063714027, + 0.01324047427624464, + 0.018054665997624397, + 0.009325031191110611, + 0.010050562210381031, + 0.01890137605369091, + 0.010591253638267517, + 0.004267814103513956, + 0.007670633029192686, + 0.006086408160626888, + 0.004214419983327389, + 0.004519559442996979, + 0.0029254253022372723, + 0.004328867886215448, + 0.004053348209708929, + 0.004156712908297777, + 0.004209699109196663, + 0.011624001897871494, + 0.003714548423886299, + 0.003031330183148384, + 0.003423385787755251, + 0.00318036787211895, + 0.003942773211747408, + 0.0035162989515811205, + 0.0043695904314517975, + 0.005692045204341412, + 0.005493570119142532, + 0.003935020882636309, + 0.0073995208367705345, + 0.0041054533794522285, + 0.009871420450508595, + 0.07093613594770432, + 0.04996049404144287, + 0.06578530371189117, + 0.025700565427541733, + 0.043703049421310425, + 0.10610611736774445, + 0.01917034387588501, + 0.08275806158781052, + 0.0058102356269955635, + 0.006379872560501099, + 0.0055897715501487255, + 0.0077844238840043545, + 0.005228950642049313, + 0.006119646597653627, + 0.008993290364742279, + 0.006969718728214502, + 0.011930934153497219, + 0.012051786296069622, + 0.01943800412118435, + 0.01826864667236805, + 0.012326067313551903, + 0.016609037294983864, + 0.011565291322767735, + 0.01160312257707119, + 0.013006729073822498, + 0.016865096986293793, + 0.007475145626813173, + 0.0071874987334012985, + 0.014909693971276283, + 0.008996067568659782, + 0.01267784833908081, + 0.010353188030421734, + 0.010715155862271786, + 0.012214398942887783, + 0.0012266666162759066, + 0.002313041128218174, + 0.0027991626411676407, + 0.0013104879762977362, + 0.0011034180643036962, + 0.001137496787123382, + 0.004077798221260309, + 0.004314000252634287, + 0.0031406658235937357, + 0.004093735944479704, + 0.003939042799174786, + 0.004206925630569458, + 0.003361496375873685, + 0.0023133286740630865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.17985938489437103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle drone shot tracks smoothly alongside a lone rider clad in full, matte-black futuristic motocross armor, its articulated composite panels gleaming subtly in the sunlight. The rider's posture is focused and low on a powerful, angular dirt bike, a matching matte black and gunmetal gray machine with exposed, heavy-duty suspension. As they relentlessly traverse the landscape, the bike's oversized rear tire aggressively kicks up a thick, billowing cloud of ochre dust that hangs suspended in the air. The camera follows, capturing their swift and decisive navigation across the rugged terrain: a vast, sun-scorched wasteland stretching to a jagged mesa on the distant horizon. The entire scene is bathed in the warm, harsh light of the late afternoon sun, casting long, dramatic shadows and brightly backlit, swirling clouds of dust, creating a golden haze against the deep blue sky. The warm tones, predominantly earthy reds and browns, contrast sharply with the rider's dark, anonymous silhouette, evoking a powerful atmosphere of solitary, high-speed determination on this epic, unfolding journey.", + "seed": 24988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_9491_seed_24988_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044811687991023065, + "vlm_logprob": -0.044811687991023065, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.044811687991023065, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06450437754392624 + ], + "opposite_logprob_tokens": [ + -2.814504384994507 + ], + "logprob_reward": -0.06450437754392624, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06450437754392624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11957842111587524, + 0.11368647962808609, + 0.10374628007411957, + 0.12712466716766357, + 0.0837169885635376, + 0.10539183020591736, + 0.07484754920005798, + 0.046330526471138, + 0.11066804826259613, + 0.11490924656391144 + ], + "credit_map": [ + 0.01140683051198721, + 0.0081694470718503, + 0.0202675499022007, + 0.022109083831310272, + 0.013173097744584084, + 0.013746063224971294, + 0.015883639454841614, + 0.014822707511484623, + 0.013141060248017311, + 0.008631507866084576, + 0.01615321822464466, + 0.019231287762522697, + 0.011641810648143291, + 0.009767147712409496, + 0.019578810781240463, + 0.015541638247668743, + 0.017130108550190926, + 0.027247831225395203, + 0.0071051944978535175, + 0.007901336997747421, + 0.007129843812435865, + 0.006114698015153408, + 0.010535819455981255, + 0.020581450313329697, + 0.011503010056912899, + 0.017509793862700462, + 0.012821917422115803, + 0.018303804099559784, + 0.007780648302286863, + 0.030857916921377182, + 0.01164179015904665, + 0.016705788671970367, + 0.012281417846679688, + 0.010463232174515724, + 0.011764167807996273, + 0.010500369593501091, + 0.011289406567811966, + 0.009233135730028152, + 0.010191367007791996, + 0.007993891835212708, + 0.010290456004440784, + 0.00965111330151558, + 0.01508916076272726, + 0.018566913902759552, + 0.011095375753939152, + 0.012589801102876663, + 0.01603079028427601, + 0.012078208848834038, + 0.008228612132370472, + 0.011724665760993958, + 0.011505179107189178, + 0.008845090866088867, + 0.006561122834682465, + 0.011029105633497238, + 0.008258420042693615, + 0.008695350959897041, + 0.0060091763734817505, + 0.004936599172651768, + 0.0057082707062363625, + 0.006189859472215176, + 0.004299734719097614, + 0.005258745513856411, + 0.005588166415691376, + 0.00833997130393982, + 0.02163388580083847, + 0.008571996353566647, + 0.027790740132331848, + 0.010552605614066124, + 0.008801686577498913, + 0.012323519214987755, + 0.008855313993990421, + 0.01213829405605793, + 0.014626078307628632, + 0.009699860587716103, + 0.017102282494306564, + 0.019555656239390373, + 0.009655124507844448, + 0.010281474329531193, + 0.01875283010303974, + 0.015235942788422108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.06450437754392624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102861508727074 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.004102861508727074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102861508727074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11120053380727768, + 0.08215061575174332, + 0.1352105736732483, + 0.10518594831228256, + 0.09695413708686829, + 0.0803522914648056, + 0.0771779790520668, + 0.11043247580528259, + 0.13772422075271606, + 0.06361117959022522 + ], + "credit_map": [ + 0.011906549334526062, + 0.011049226857721806, + 0.01694308966398239, + 0.018637381494045258, + 0.012179437093436718, + 0.01478093583136797, + 0.013200344517827034, + 0.012503569014370441, + 0.009207337163388729, + 0.005783779080957174, + 0.016544533893465996, + 0.016135942190885544, + 0.006608990021049976, + 0.006649306043982506, + 0.010539257898926735, + 0.010681466199457645, + 0.02828145958483219, + 0.04452558234333992, + 0.005384344141930342, + 0.008696306496858597, + 0.006374616175889969, + 0.007329327519983053, + 0.006033958401530981, + 0.028584977611899376, + 0.008253468200564384, + 0.01542186550796032, + 0.009702412411570549, + 0.008330837823450565, + 0.01929454505443573, + 0.019750332459807396, + 0.015000680461525917, + 0.009431803598999977, + 0.016718665137887, + 0.012795373797416687, + 0.014018652960658073, + 0.011330202221870422, + 0.011661113239824772, + 0.00991086382418871, + 0.011035047471523285, + 0.00948422122746706, + 0.010598537512123585, + 0.00930478610098362, + 0.00964258424937725, + 0.012419509701430798, + 0.00591254560276866, + 0.013621683232486248, + 0.009576283395290375, + 0.00927636120468378, + 0.01044156588613987, + 0.01113378256559372, + 0.011179673485457897, + 0.009797636419534683, + 0.004954914562404156, + 0.013528885319828987, + 0.007548391819000244, + 0.00859313365072012, + 0.010515617206692696, + 0.015852084383368492, + 0.013290303759276867, + 0.019165808334946632, + 0.017081791535019875, + 0.011554146185517311, + 0.01163836196064949, + 0.011334359645843506, + 0.028673943132162094, + 0.008497404865920544, + 0.03757430985569954, + 0.013559558428823948, + 0.00861450843513012, + 0.013717355206608772, + 0.011279013939201832, + 0.01580812595784664, + 0.010398388840258121, + 0.006468727719038725, + 0.010322343558073044, + 0.008850268088281155, + 0.004963423125445843, + 0.007806055713444948, + 0.008740011602640152, + 0.00606196466833353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.004102861508727074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738320458680391 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738320458680391, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738320458680391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061155665665864944, + 0.03772455081343651, + 0.0723905861377716, + 0.12559859454631805, + 0.09066581726074219, + 0.06121304631233215, + 0.1391013115644455, + 0.2379610389471054, + 0.14252299070358276, + 0.03166640177369118 + ], + "credit_map": [ + 0.007159616332501173, + 0.006195529364049435, + 0.009546615183353424, + 0.009823868051171303, + 0.007373438682407141, + 0.007021232042461634, + 0.0073946574702858925, + 0.0066407062113285065, + 0.0037761968560516834, + 0.002944891806691885, + 0.007145997602492571, + 0.00888162013143301, + 0.0026585450395941734, + 0.0031424995977431536, + 0.0038628545589745045, + 0.005311948712915182, + 0.014944470487535, + 0.02294604480266571, + 0.003974145278334618, + 0.004838188644498587, + 0.0036960241850465536, + 0.0040791514329612255, + 0.0037428801879286766, + 0.014169680885970592, + 0.010739127174019814, + 0.015659520402550697, + 0.011664209887385368, + 0.01297016441822052, + 0.02330164611339569, + 0.022098358720541, + 0.017748035490512848, + 0.011417527683079243, + 0.006657148711383343, + 0.010091685689985752, + 0.006912032142281532, + 0.015247921459376812, + 0.012063167989253998, + 0.012120719067752361, + 0.013063063845038414, + 0.014510082080960274, + 0.009769000113010406, + 0.007828585803508759, + 0.009159956127405167, + 0.009295407682657242, + 0.004287456162273884, + 0.0068542431108653545, + 0.006849192548543215, + 0.007169210817664862, + 0.015981465578079224, + 0.017602313309907913, + 0.02967764437198639, + 0.017046039924025536, + 0.010227879509329796, + 0.01664263941347599, + 0.013798067346215248, + 0.018125249072909355, + 0.024446146562695503, + 0.01998535543680191, + 0.04552756994962692, + 0.02805953659117222, + 0.017083480954170227, + 0.02863231673836708, + 0.024587010964751244, + 0.04963962733745575, + 0.0234097708016634, + 0.01153147779405117, + 0.033694472163915634, + 0.01525058038532734, + 0.006921359803527594, + 0.011824086308479309, + 0.01221859734505415, + 0.02767263539135456, + 0.004558421205729246, + 0.005825326312333345, + 0.004423300735652447, + 0.003929378930479288, + 0.003173904726281762, + 0.002955390838906169, + 0.003281589364632964, + 0.003519090823829174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.006738320458680391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059431265108287334 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059431265108287334, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059431265108287334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041115131229162216, + 0.11661656200885773, + 0.45990675687789917, + 0.14973261952400208, + 0.04095456749200821, + 0.03477530926465988, + 0.027039028704166412, + 0.042388901114463806, + 0.06671226769685745, + 0.020758822560310364 + ], + "credit_map": [ + 0.004684643354266882, + 0.003794199787080288, + 0.006542534567415714, + 0.008486240170896053, + 0.004038291517645121, + 0.00412339810281992, + 0.005125122610479593, + 0.004320700652897358, + 0.01282031275331974, + 0.009174910373985767, + 0.026135293766856194, + 0.026195688173174858, + 0.0054353149607777596, + 0.009721259586513042, + 0.010801639407873154, + 0.016332155093550682, + 0.09310977905988693, + 0.14789678156375885, + 0.024044401943683624, + 0.030871329829096794, + 0.022021442651748657, + 0.02509860321879387, + 0.017651282250881195, + 0.09921317547559738, + 0.012826981022953987, + 0.02090093307197094, + 0.014512590132653713, + 0.019262518733739853, + 0.024024508893489838, + 0.025122417137026787, + 0.019601808860898018, + 0.013480881229043007, + 0.0042041633278131485, + 0.005054524168372154, + 0.004577042534947395, + 0.0075101363472640514, + 0.002290035132318735, + 0.006628637667745352, + 0.003301889169961214, + 0.0073881433345377445, + 0.005212169606238604, + 0.0036024763248860836, + 0.00407362449914217, + 0.006233673542737961, + 0.0031927202362567186, + 0.0037529475521296263, + 0.004619887564331293, + 0.004087815992534161, + 0.0032318199519068003, + 0.003742993576452136, + 0.004208570811897516, + 0.0032019075006246567, + 0.001806897111237049, + 0.004726260434836149, + 0.002761437091976404, + 0.0033591436222195625, + 0.004341281019151211, + 0.005130100529640913, + 0.006181178614497185, + 0.005935209337621927, + 0.0047860583290457726, + 0.004103364888578653, + 0.004352733958512545, + 0.007558979094028473, + 0.01625850424170494, + 0.004200084134936333, + 0.01911475695669651, + 0.005314646288752556, + 0.004056415054947138, + 0.007112099789083004, + 0.00406271917745471, + 0.006593046244233847, + 0.003770421724766493, + 0.002006822731345892, + 0.004472588188946247, + 0.0022831931710243225, + 0.00132363336160779, + 0.0029475244227796793, + 0.002177434042096138, + 0.0017772061983123422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.0059431265108287334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276975393295288 + ], + "opposite_logprob_tokens": [ + -2.0177698135375977 + ], + "logprob_reward": -0.14276975393295288, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14276975393295288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08519618958234787, + 0.05262121930718422, + 0.11893393844366074, + 0.2210044413805008, + 0.1150740385055542, + 0.10932499170303345, + 0.06563040614128113, + 0.12564167380332947, + 0.0714881494641304, + 0.035084955394268036 + ], + "credit_map": [ + 0.007566808257251978, + 0.006666491739451885, + 0.01622469164431095, + 0.014409673400223255, + 0.010279049165546894, + 0.009423943236470222, + 0.011038580909371376, + 0.009586947970092297, + 0.004536970052868128, + 0.0036621266044676304, + 0.007102209609001875, + 0.008407010696828365, + 0.006474869791418314, + 0.007280604913830757, + 0.008535472676157951, + 0.0066219535656273365, + 0.023299744352698326, + 0.037320658564567566, + 0.006401787977665663, + 0.008281175047159195, + 0.006080711726099253, + 0.0074129956774413586, + 0.005994449369609356, + 0.024142418056726456, + 0.019886502996087074, + 0.029847003519535065, + 0.01919330470263958, + 0.029309596866369247, + 0.031005624681711197, + 0.04625155031681061, + 0.023304009810090065, + 0.022206850349903107, + 0.016384867951273918, + 0.015309778973460197, + 0.013265512883663177, + 0.017763957381248474, + 0.009641459211707115, + 0.016576731577515602, + 0.009551219642162323, + 0.016580507159233093, + 0.015385436825454235, + 0.014665741473436356, + 0.012382208369672298, + 0.01788102649152279, + 0.009002982638776302, + 0.016386155039072037, + 0.011496026068925858, + 0.0121254101395607, + 0.008675863035023212, + 0.008122008293867111, + 0.010418565943837166, + 0.007941538468003273, + 0.00585544528439641, + 0.012251392938196659, + 0.005837828852236271, + 0.0065277693793177605, + 0.013636612333357334, + 0.015925107523798943, + 0.019665315747261047, + 0.01902913860976696, + 0.013214551843702793, + 0.01342849899083376, + 0.013087722472846508, + 0.017654726281762123, + 0.01682877168059349, + 0.005722720175981522, + 0.01636595092713833, + 0.004726556595414877, + 0.004642069805413485, + 0.00911670457571745, + 0.004090277943760157, + 0.009995107538998127, + 0.005966621451079845, + 0.004220635164529085, + 0.008021250367164612, + 0.004161520395427942, + 0.0022979979403316975, + 0.004254264757037163, + 0.003257234813645482, + 0.002905428409576416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.14276975393295288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle drone shot tracks smoothly alongside a lone rider clad in full, matte-black futuristic motocross armor, its articulated composite panels gleaming subtly in the sunlight. The rider's posture is focused and low on a powerful, angular dirt bike, a matching matte black and gunmetal gray machine with exposed, heavy-duty suspension. As they relentlessly traverse the landscape, the bike's oversized rear tire aggressively kicks up a thick, billowing cloud of ochre dust that hangs suspended in the air. The camera follows, capturing their swift and decisive navigation across the rugged terrain: a vast, sun-scorched wasteland stretching to a jagged mesa on the distant horizon. The entire scene is bathed in the warm, harsh light of the late afternoon sun, casting long, dramatic shadows and brightly backlit, swirling clouds of dust, creating a golden haze against the deep blue sky. The warm tones, predominantly earthy reds and browns, contrast sharply with the rider's dark, anonymous silhouette, evoking a powerful atmosphere of solitary, high-speed determination on this epic, unfolding journey.", + "seed": 24989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000065/group_9/index_9491_seed_24989_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03892466882243752, + "vlm_logprob": -0.03892466882243752, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03892466882243752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10304668545722961 + ], + "opposite_logprob_tokens": [ + -2.3530466556549072 + ], + "logprob_reward": -0.10304668545722961, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10304668545722961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046113885939121246, + 0.05486376956105232, + 0.04262254759669304, + 0.108572818338871, + 0.10321278125047684, + 0.024746594950556755, + 0.05175890028476715, + 0.5182704329490662, + 0.03382089361548424, + 0.016017349436879158 + ], + "credit_map": [ + 0.004474600777029991, + 0.004840533249080181, + 0.004250745289027691, + 0.006071234587579966, + 0.006567221600562334, + 0.006908438168466091, + 0.005875078961253166, + 0.00712603610008955, + 0.008367670699954033, + 0.0039321864023804665, + 0.011778082698583603, + 0.01155170425772667, + 0.005389625206589699, + 0.0045915101654827595, + 0.00522690499201417, + 0.004026089329272509, + 0.003622292308136821, + 0.003408970544114709, + 0.007672849576920271, + 0.006953198928385973, + 0.005730018950998783, + 0.005455279257148504, + 0.0043962388299405575, + 0.005383704323321581, + 0.017290007323026657, + 0.015929516404867172, + 0.009617196395993233, + 0.00889106560498476, + 0.022492630407214165, + 0.016929594799876213, + 0.009143145754933357, + 0.008279671892523766, + 0.011114193126559258, + 0.011453184299170971, + 0.013420100323855877, + 0.015147469937801361, + 0.01446832437068224, + 0.01070729922503233, + 0.01504063606262207, + 0.011861585080623627, + 0.002436941023916006, + 0.0028048024978488684, + 0.003241761354729533, + 0.0038320759776979685, + 0.0029798150062561035, + 0.0024652848951518536, + 0.004696452524513006, + 0.002289464697241783, + 0.0056090266443789005, + 0.00296161905862391, + 0.0078398697078228, + 0.006982357706874609, + 0.003630078863352537, + 0.012998616322875023, + 0.006765059195458889, + 0.00497227581217885, + 0.060330770909786224, + 0.03203679621219635, + 0.13982662558555603, + 0.044013623148202896, + 0.01831272430717945, + 0.18139305710792542, + 0.017317861318588257, + 0.02503903955221176, + 0.003930786624550819, + 0.005713317077606916, + 0.004643749911338091, + 0.00500983651727438, + 0.002389335073530674, + 0.006465091835707426, + 0.003184611676260829, + 0.0024841688573360443, + 0.0028583176899701357, + 0.0028194300830364227, + 0.0018435056554153562, + 0.0016052948776632547, + 0.0019387826323509216, + 0.001956677995622158, + 0.001352670369669795, + 0.0016426722286269069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.10304668545722961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260316655039787 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260316655039787, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260316655039787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04361019656062126, + 0.22888857126235962, + 0.04798854514956474, + 0.05460949242115021, + 0.08602125197649002, + 0.029799306765198708, + 0.06260837614536285, + 0.3968561887741089, + 0.030637873336672783, + 0.018980205059051514 + ], + "credit_map": [ + 0.004952896386384964, + 0.0038513154722750187, + 0.005480150692164898, + 0.007648106198757887, + 0.004432556685060263, + 0.0042413449846208096, + 0.006552825216203928, + 0.006451006513088942, + 0.03914592042565346, + 0.016171574592590332, + 0.05695260316133499, + 0.04974924027919769, + 0.010118882171809673, + 0.020216288045048714, + 0.01820201799273491, + 0.01833205670118332, + 0.005971135105937719, + 0.004240876995027065, + 0.007989292964339256, + 0.008249109610915184, + 0.005851780064404011, + 0.005071260966360569, + 0.004710924345999956, + 0.005904172547161579, + 0.009156647138297558, + 0.00717246113345027, + 0.004838941153138876, + 0.004417089279741049, + 0.009104528464376926, + 0.008008860051631927, + 0.005548419896513224, + 0.006362550426274538, + 0.008983106352388859, + 0.010004733689129353, + 0.01122736744582653, + 0.013514278456568718, + 0.011866229586303234, + 0.008724399842321873, + 0.012511670589447021, + 0.009189477190375328, + 0.00352246081456542, + 0.00236335676163435, + 0.005553151480853558, + 0.0036323319654911757, + 0.002907866844907403, + 0.003134253202006221, + 0.0032141185365617275, + 0.0054717715829610825, + 0.007118024863302708, + 0.00395123241469264, + 0.010079825296998024, + 0.006025555543601513, + 0.005229826085269451, + 0.01899554952979088, + 0.0061677368357777596, + 0.005040636286139488, + 0.04608564451336861, + 0.024699946865439415, + 0.10918552428483963, + 0.0340365394949913, + 0.012531941756606102, + 0.13920758664608002, + 0.011981700547039509, + 0.019127320498228073, + 0.004145273007452488, + 0.004049201961606741, + 0.004408641252666712, + 0.004401204641908407, + 0.0024247507099062204, + 0.005020133685320616, + 0.003190189367160201, + 0.0029984836000949144, + 0.0031093705911189318, + 0.003017927287146449, + 0.0025845218915492296, + 0.002099419943988323, + 0.0018303940305486321, + 0.0020809213165193796, + 0.0016086052637547255, + 0.00264904648065567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.005260316655039787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107639167457819 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.01107639167457819, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107639167457819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.112720787525177, + 0.19299134612083435, + 0.12648889422416687, + 0.11768757551908493, + 0.051678020507097244, + 0.11893326044082642, + 0.05147041752934456, + 0.1478443145751953, + 0.042479708790779114, + 0.03770570084452629 + ], + "credit_map": [ + 0.010181190446019173, + 0.011990033090114594, + 0.008820599876344204, + 0.011543654836714268, + 0.007365179713815451, + 0.03066791035234928, + 0.009056861512362957, + 0.023095348849892616, + 0.03149424493312836, + 0.013651888817548752, + 0.04979003220796585, + 0.03952552005648613, + 0.00811108760535717, + 0.019224893301725388, + 0.015101718716323376, + 0.01609194278717041, + 0.018355179578065872, + 0.01213026326149702, + 0.014963720925152302, + 0.016265541315078735, + 0.011759601533412933, + 0.0193537175655365, + 0.01336289756000042, + 0.020297961309552193, + 0.011575676500797272, + 0.01101148221641779, + 0.016088368371129036, + 0.015150804072618484, + 0.01623096503317356, + 0.013907909393310547, + 0.012225590646266937, + 0.021496761590242386, + 0.0055129509419202805, + 0.005536938086152077, + 0.0066682444885373116, + 0.008921678178012371, + 0.008438513614237309, + 0.005325750447809696, + 0.005987861193716526, + 0.005286078434437513, + 0.014570878818631172, + 0.009535790421068668, + 0.020155612379312515, + 0.014542181976139545, + 0.011903096921741962, + 0.012510882690548897, + 0.012607567943632603, + 0.023107243701815605, + 0.004373587667942047, + 0.0034210265148431063, + 0.0074692703783512115, + 0.00964015070348978, + 0.005608420819044113, + 0.0061789993196725845, + 0.010060545057058334, + 0.004718413110822439, + 0.01732688769698143, + 0.0091708330437541, + 0.04008853808045387, + 0.012716936878859997, + 0.00474574975669384, + 0.051333457231521606, + 0.0045110615901649, + 0.007950841449201107, + 0.005087976343929768, + 0.006937783677130938, + 0.006415721960365772, + 0.007116713095456362, + 0.003393147373571992, + 0.00492151640355587, + 0.004432125482708216, + 0.004174721427261829, + 0.0029132328927516937, + 0.0037086359225213528, + 0.005833683535456657, + 0.004401265177875757, + 0.0031176062766462564, + 0.006168085150420666, + 0.0036766850389540195, + 0.007886499166488647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.01107639167457819 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247035529464483 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.005247035529464483, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247035529464483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027498403564095497, + 0.13603554666042328, + 0.05895741283893585, + 0.03322143852710724, + 0.03373868018388748, + 0.024351827800273895, + 0.03781412914395332, + 0.5830307006835938, + 0.036537863314151764, + 0.028814038261771202 + ], + "credit_map": [ + 0.0030261704232543707, + 0.003128337673842907, + 0.0046935854479670525, + 0.004169073887169361, + 0.0027939477004110813, + 0.0035801860503852367, + 0.003203471191227436, + 0.00290363235399127, + 0.023360177874565125, + 0.010004144161939621, + 0.03400333970785141, + 0.02879420667886734, + 0.005768167786300182, + 0.012470097281038761, + 0.010514694266021252, + 0.011120705865323544, + 0.006497129797935486, + 0.005554720759391785, + 0.011479164473712444, + 0.009385909885168076, + 0.006216698791831732, + 0.006948056630790234, + 0.0055269235745072365, + 0.007348799612373114, + 0.004750391934067011, + 0.004584743175655603, + 0.00310460920445621, + 0.002957515651360154, + 0.005746561102569103, + 0.004867124371230602, + 0.0031034580897539854, + 0.004107031971216202, + 0.004167604725807905, + 0.0031426872592419386, + 0.005634305067360401, + 0.004065996035933495, + 0.004964163526892662, + 0.0038552486803382635, + 0.003903365461155772, + 0.0040053073316812515, + 0.003547122236341238, + 0.002037854166701436, + 0.005117088556289673, + 0.00267158099450171, + 0.002778652822598815, + 0.0030177996959537268, + 0.0021325338166207075, + 0.0030491945799440145, + 0.005140949506312609, + 0.0023527105804532766, + 0.006440230645239353, + 0.0035679121501743793, + 0.0023908603470772505, + 0.011946476064622402, + 0.0026039020158350468, + 0.0033710855059325695, + 0.0674586072564125, + 0.035984836518764496, + 0.16215960681438446, + 0.05049305409193039, + 0.0170386154204607, + 0.20463810861110687, + 0.017030347138643265, + 0.028227495029568672, + 0.003928775433450937, + 0.006148237735033035, + 0.005981616210192442, + 0.0039050783962011337, + 0.002148375380784273, + 0.008238904178142548, + 0.0027493208181113005, + 0.0034375502727925777, + 0.004793287254869938, + 0.00589330680668354, + 0.0036849549505859613, + 0.002602085005491972, + 0.002564222551882267, + 0.0045280070044100285, + 0.0016479872865602374, + 0.003100185189396143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.005247035529464483 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone rider wearing full matte-black futuristic motocross armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a powerful, angular dirt bike with exposed heavy-duty suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike kick up a thick, billowing cloud of ochre dust while moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast, sun-scorched wasteland with a jagged mesa on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a sweeping high-angle drone view that tracks smoothly alongside the rider?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999291479587555 + ], + "opposite_logprob_tokens": [ + -2.694993019104004 + ], + "logprob_reward": -0.06999291479587555, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999291479587555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056020088493824005, + 0.1306396722793579, + 0.06447505205869675, + 0.07152910530567169, + 0.03125372901558876, + 0.2169078141450882, + 0.036757078021764755, + 0.26979535818099976, + 0.09483180195093155, + 0.027790255844593048 + ], + "credit_map": [ + 0.003821566002443433, + 0.006031028460711241, + 0.007169757038354874, + 0.00609293719753623, + 0.00536376005038619, + 0.012385936453938484, + 0.006656823679804802, + 0.008498274721205235, + 0.022781936451792717, + 0.00983068160712719, + 0.03117423504590988, + 0.02714064158499241, + 0.006408947054296732, + 0.01236292440444231, + 0.010756755247712135, + 0.010183549486100674, + 0.007777954917401075, + 0.005818240810185671, + 0.010181707330048084, + 0.012221611104905605, + 0.006085499655455351, + 0.0067064701579511166, + 0.006570490077137947, + 0.009113078005611897, + 0.010155614465475082, + 0.00628221919760108, + 0.01089633721858263, + 0.009099937975406647, + 0.006751624867320061, + 0.007772641256451607, + 0.008637937717139721, + 0.011932793073356152, + 0.003410050878301263, + 0.0027767277788370848, + 0.00564656313508749, + 0.0044469148851931095, + 0.0038196085952222347, + 0.002740453230217099, + 0.003592157969251275, + 0.004821254871785641, + 0.026538275182247162, + 0.017139378935098648, + 0.040969278663396835, + 0.02258402295410633, + 0.023752296343445778, + 0.024876292794942856, + 0.018783019855618477, + 0.04226526618003845, + 0.005087564699351788, + 0.003060065908357501, + 0.00495533412322402, + 0.004882656037807465, + 0.004078507889062166, + 0.005835713818669319, + 0.004869766533374786, + 0.003987468779087067, + 0.030682504177093506, + 0.01726919412612915, + 0.0733809843659401, + 0.023587938398122787, + 0.009849070571362972, + 0.09208374470472336, + 0.010295466519892216, + 0.01264646090567112, + 0.010008458979427814, + 0.0076510231010615826, + 0.014809994027018547, + 0.01845344342291355, + 0.009886357001960278, + 0.010593404062092304, + 0.013270111754536629, + 0.010159004479646683, + 0.0031778933480381966, + 0.004071772564202547, + 0.00303567829541862, + 0.002562611596658826, + 0.0030291930306702852, + 0.004701581783592701, + 0.0034567846450954676, + 0.003754740348085761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.06999291479587555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_tensor([4541, 4541])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_tensor([4541, 4541])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..51b591879e153a1481e10a96f608ee67ade11ece --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_tensor([4541, 4541])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium-sized shot captures a man in his 30s with neatly cropped brown hair standing in a well-lit, organized workshop. Dressed in a navy short-sleeved polo shirt and khaki pants, he exudes a professional demeanor. The background consists of a clean concrete floor and a slightly out-of-focus workbench with neatly arranged tools. The man tightly grasps a vibrant red cylindrical fire extinguisher with both hands, its black nozzle and pressure gauge clearly visible. As he speaks directly to the camera with a serious and clear expression, he begins to vigorously shake the extinguisher up and down, a movement that conveys a sense of urgent preparedness. He then pauses, holding the extinguisher steady with his right hand and gesturing vigorously with his left hand, pointing his index finger at the nozzle to emphasize a key point in his explanation. Finally, he nods, gently and steadily bending his knees, and deliberately placing the extinguisher upright on the floor, his movements precise and measured. The entire scene is bathed in bright, cool fluorescent light from an overhead light source, creating clear reflections on the extinguisher's smooth surface and establishing an atmosphere of focused instruction.", + "seed": 13920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_4541_seed_13920_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032762344600632785, + "vlm_logprob": -0.032762344600632785, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.032762344600632785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02088659256696701 + ], + "opposite_logprob_tokens": [ + -4.520886421203613 + ], + "logprob_reward": -0.02088659256696701, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02088659256696701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2097054123878479, + 0.08585792779922485, + 0.04298204556107521, + 0.06140950322151184, + 0.0497133806347847, + 0.038534168154001236, + 0.08283455669879913, + 0.08681608736515045, + 0.15490685403347015, + 0.18724006414413452 + ], + "credit_map": [ + 0.024689946323633194, + 0.01739339902997017, + 0.047851402312517166, + 0.023069892078638077, + 0.024294842034578323, + 0.03140305355191231, + 0.01980525255203247, + 0.021197626367211342, + 0.009573463350534439, + 0.006776844151318073, + 0.010229218751192093, + 0.007990275509655476, + 0.023556040599942207, + 0.009192398749291897, + 0.010218093171715736, + 0.008321591652929783, + 0.005922286305576563, + 0.006689767353236675, + 0.005916297901421785, + 0.0037297550588846207, + 0.005484850145876408, + 0.006211022846400738, + 0.0049657090567052364, + 0.004062360152602196, + 0.009473937563598156, + 0.0053807529620826244, + 0.01034894771873951, + 0.006439664866775274, + 0.006970967631787062, + 0.010343015193939209, + 0.004817450884729624, + 0.0076347654685378075, + 0.0075349765829741955, + 0.005260659847408533, + 0.005624324083328247, + 0.004057367332279682, + 0.0043174633756279945, + 0.0058998712338507175, + 0.009427659213542938, + 0.007591060362756252, + 0.0063377185724675655, + 0.0039772349409759045, + 0.006537488661706448, + 0.003246834734454751, + 0.004148618783801794, + 0.006409151013940573, + 0.0040266672149300575, + 0.003850453533232212, + 0.01375550590455532, + 0.005377712659537792, + 0.01850009523332119, + 0.006140637677162886, + 0.007329095155000687, + 0.010918736457824707, + 0.01190795935690403, + 0.008904816582798958, + 0.01502213068306446, + 0.008840038441121578, + 0.016082394868135452, + 0.0077945152297616005, + 0.008202589116990566, + 0.009619159623980522, + 0.009932393208146095, + 0.011322869919240475, + 0.015827583149075508, + 0.01905827596783638, + 0.02537607215344906, + 0.01434691809117794, + 0.018415894359350204, + 0.019787514582276344, + 0.014977654442191124, + 0.027116941288113594, + 0.01881897635757923, + 0.0297220591455698, + 0.02298727072775364, + 0.03021775186061859, + 0.01692023128271103, + 0.016597529873251915, + 0.0315018892288208, + 0.02047433890402317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02088659256696701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609886810183525 + ], + "opposite_logprob_tokens": [ + -4.141098976135254 + ], + "logprob_reward": -0.01609886810183525, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609886810183525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12545886635780334, + 0.08644858747720718, + 0.0741744190454483, + 0.05161483958363533, + 0.03391261771321297, + 0.03617384657263756, + 0.05932839214801788, + 0.11118081957101822, + 0.2232777327299118, + 0.19842985272407532 + ], + "credit_map": [ + 0.01983882673084736, + 0.0165447685867548, + 0.019268224015831947, + 0.013719498179852962, + 0.010822839103639126, + 0.013199843466281891, + 0.018810611218214035, + 0.013254269026219845, + 0.011374915950000286, + 0.008210002444684505, + 0.01385835837572813, + 0.008553928695619106, + 0.016010645776987076, + 0.01406150870025158, + 0.0066429851576685905, + 0.007736254017800093, + 0.012893615290522575, + 0.008308034390211105, + 0.01394141186028719, + 0.00505835609510541, + 0.00823927577584982, + 0.012294911779463291, + 0.005345683079212904, + 0.008093140088021755, + 0.006517757661640644, + 0.0052599310874938965, + 0.011024558916687965, + 0.003290698165073991, + 0.007204764988273382, + 0.009781011380255222, + 0.0034426026977598667, + 0.005093521438539028, + 0.004202115815132856, + 0.0034613381139934063, + 0.005241789389401674, + 0.0035389778204262257, + 0.002703000558540225, + 0.005350870545953512, + 0.004620707128196955, + 0.004793821834027767, + 0.005237115081399679, + 0.003083211602643132, + 0.0046360064297914505, + 0.0024345223791897297, + 0.003647606587037444, + 0.004228696227073669, + 0.008143189363181591, + 0.004763501696288586, + 0.009828910231590271, + 0.005811075679957867, + 0.011406801640987396, + 0.0048983776941895485, + 0.00540109584107995, + 0.007258387748152018, + 0.008037303574383259, + 0.006686445791274309, + 0.01953883469104767, + 0.013000084087252617, + 0.016354886814951897, + 0.011113123036921024, + 0.012084366753697395, + 0.012391434982419014, + 0.012341049499809742, + 0.014357046224176884, + 0.03261794149875641, + 0.0244424007833004, + 0.023378197103738785, + 0.03468039631843567, + 0.020693140104413033, + 0.03553953394293785, + 0.019213462248444557, + 0.03271270915865898, + 0.03358127176761627, + 0.027413351461291313, + 0.030589072033762932, + 0.018600033596158028, + 0.01704671047627926, + 0.023517580702900887, + 0.015735289081931114, + 0.03194655478000641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.01609886810183525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032076595816761255 + ], + "opposite_logprob_tokens": [ + -5.753207683563232 + ], + "logprob_reward": -0.0032076595816761255, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032076595816761255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15502041578292847, + 0.09662491083145142, + 0.05426359176635742, + 0.040058668702840805, + 0.033516719937324524, + 0.04872816428542137, + 0.02295856736600399, + 0.03644454479217529, + 0.0897187814116478, + 0.42266562581062317 + ], + "credit_map": [ + 0.027503840625286102, + 0.015445823781192303, + 0.032086778432130814, + 0.009284738451242447, + 0.013821151107549667, + 0.02020704746246338, + 0.018910203129053116, + 0.017760852351784706, + 0.010996940545737743, + 0.00785552617162466, + 0.010289820842444897, + 0.008636611513793468, + 0.0272614024579525, + 0.010561110451817513, + 0.011371693573892117, + 0.009651809930801392, + 0.006595808546990156, + 0.006508637685328722, + 0.005428689066320658, + 0.007065598387271166, + 0.0056882454082369804, + 0.0071058026514947414, + 0.009971329011023045, + 0.005899481009691954, + 0.00681881420314312, + 0.0030970056541264057, + 0.00599867757409811, + 0.003578934585675597, + 0.005113718565553427, + 0.006486319005489349, + 0.004731266293674707, + 0.004233929794281721, + 0.0061735487543046474, + 0.004515550099313259, + 0.005016586743295193, + 0.003147135255858302, + 0.0035055261105298996, + 0.005354692693799734, + 0.002609687391668558, + 0.003193994052708149, + 0.009177920408546925, + 0.003747698850929737, + 0.006705032661557198, + 0.0038564656861126423, + 0.004065041895955801, + 0.007377000525593758, + 0.008158030919730663, + 0.005640976596623659, + 0.004130115732550621, + 0.0021288085263222456, + 0.004230836872011423, + 0.0019079677294939756, + 0.0022567985579371452, + 0.003093474544584751, + 0.002730666659772396, + 0.0024798973463475704, + 0.005839258898049593, + 0.0037295026704669, + 0.004959246143698692, + 0.004108037333935499, + 0.004446413367986679, + 0.004026075825095177, + 0.004723594523966312, + 0.004612411372363567, + 0.009031257592141628, + 0.008828694000840187, + 0.011440315283834934, + 0.013270813971757889, + 0.013614204712212086, + 0.013257353566586971, + 0.00879365298897028, + 0.01148249488323927, + 0.06791436672210693, + 0.047920916229486465, + 0.06868550181388855, + 0.03609419986605644, + 0.03746524453163147, + 0.07880908250808716, + 0.023783091455698013, + 0.06199318915605545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0032076595816761255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003241240978241 + ], + "opposite_logprob_tokens": [ + -2.3503241539001465 + ], + "logprob_reward": -0.1003241240978241, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003241240978241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16082201898097992, + 0.25614282488822937, + 0.05081162974238396, + 0.04553842172026634, + 0.05037403106689453, + 0.025085413828492165, + 0.03223434463143349, + 0.04293690621852875, + 0.10263429582118988, + 0.23342014849185944 + ], + "credit_map": [ + 0.024238869547843933, + 0.011837583966553211, + 0.019660012796521187, + 0.019773758947849274, + 0.012999070808291435, + 0.02508585900068283, + 0.024991432204842567, + 0.022235432639718056, + 0.029105324298143387, + 0.02289065718650818, + 0.02249385230243206, + 0.023645637556910515, + 0.0693620815873146, + 0.03071906790137291, + 0.02775920182466507, + 0.0301669891923666, + 0.00813745241612196, + 0.005261802114546299, + 0.008013462647795677, + 0.006796622183173895, + 0.00391050148755312, + 0.0061249504797160625, + 0.007119457703083754, + 0.005447378847748041, + 0.005938540678471327, + 0.004076594486832619, + 0.006824520882219076, + 0.0033797798678278923, + 0.004693166818469763, + 0.009400754235684872, + 0.0055798194371163845, + 0.005645245313644409, + 0.007699641399085522, + 0.008261162787675858, + 0.006717517506331205, + 0.007863117381930351, + 0.005849109962582588, + 0.007493910845369101, + 0.0028968651313334703, + 0.00359270628541708, + 0.00425367895513773, + 0.0031564850360155106, + 0.0034408930223435163, + 0.0031101428903639317, + 0.003213948803022504, + 0.0035463208332657814, + 0.00237487256526947, + 0.001989071723073721, + 0.004777516704052687, + 0.002907942049205303, + 0.005632318556308746, + 0.005108916666358709, + 0.0031612724997103214, + 0.0031791352666914463, + 0.0039621684700250626, + 0.003505073254927993, + 0.0057635861448943615, + 0.005531659349799156, + 0.004986214451491833, + 0.005023840349167585, + 0.003854649607092142, + 0.004590904340147972, + 0.005076962523162365, + 0.00810908991843462, + 0.010250231251120567, + 0.01272039394825697, + 0.013455281034111977, + 0.011507941409945488, + 0.014540194533765316, + 0.013946622610092163, + 0.010737580247223377, + 0.015476043336093426, + 0.029173003509640694, + 0.018199127167463303, + 0.046494510024785995, + 0.02503982186317444, + 0.02016546204686165, + 0.02833603136241436, + 0.025539156049489975, + 0.04047303646802902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.1003241240978241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329447865486145 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.02329447865486145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329447865486145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18023334443569183, + 0.04904171824455261, + 0.06303194165229797, + 0.06249526888132095, + 0.042614907026290894, + 0.030518734827637672, + 0.13458001613616943, + 0.14322711527347565, + 0.14662803709506989, + 0.14762884378433228 + ], + "credit_map": [ + 0.021493272855877876, + 0.01852376200258732, + 0.03212564066052437, + 0.01668163575232029, + 0.020839715376496315, + 0.03905322030186653, + 0.013570069335401058, + 0.01794603280723095, + 0.005348706617951393, + 0.004251466132700443, + 0.00780753418803215, + 0.003883163444697857, + 0.010946894064545631, + 0.0068028587847948074, + 0.0040929545648396015, + 0.005908142775297165, + 0.011399395763874054, + 0.0056771766394376755, + 0.013363627716898918, + 0.005241920705884695, + 0.006423022598028183, + 0.008164594881236553, + 0.005276082083582878, + 0.0074861240573227406, + 0.007813469506800175, + 0.0045760213397443295, + 0.01365253422409296, + 0.004058349411934614, + 0.006137598771601915, + 0.015093705616891384, + 0.006038607098162174, + 0.005124983377754688, + 0.006659423001110554, + 0.0041984799318015575, + 0.006346292793750763, + 0.0032028467394411564, + 0.004063178785145283, + 0.007203658577054739, + 0.00444916682317853, + 0.006491860840469599, + 0.004723251331597567, + 0.002886691130697727, + 0.003933368716388941, + 0.003038733033463359, + 0.0025965666864067316, + 0.003762150416150689, + 0.005739770829677582, + 0.003838203614577651, + 0.021075492724776268, + 0.008349394425749779, + 0.03501712158322334, + 0.009782280772924423, + 0.009150446392595768, + 0.017382612451910973, + 0.014688555151224136, + 0.01913411170244217, + 0.019036870449781418, + 0.013483530841767788, + 0.029523305594921112, + 0.0127558633685112, + 0.011368174105882645, + 0.01627807877957821, + 0.01972155272960663, + 0.021059738472104073, + 0.01921370066702366, + 0.013397566974163055, + 0.028765203431248665, + 0.013103273697197437, + 0.01101813092827797, + 0.036285921931266785, + 0.009163515642285347, + 0.01568073220551014, + 0.025539441034197807, + 0.01944814994931221, + 0.018039532005786896, + 0.013316850177943707, + 0.011916589923202991, + 0.02429181896150112, + 0.014371197670698166, + 0.020705265924334526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02329447865486145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium-sized shot captures a man in his 30s with neatly cropped brown hair standing in a well-lit, organized workshop. Dressed in a navy short-sleeved polo shirt and khaki pants, he exudes a professional demeanor. The background consists of a clean concrete floor and a slightly out-of-focus workbench with neatly arranged tools. The man tightly grasps a vibrant red cylindrical fire extinguisher with both hands, its black nozzle and pressure gauge clearly visible. As he speaks directly to the camera with a serious and clear expression, he begins to vigorously shake the extinguisher up and down, a movement that conveys a sense of urgent preparedness. He then pauses, holding the extinguisher steady with his right hand and gesturing vigorously with his left hand, pointing his index finger at the nozzle to emphasize a key point in his explanation. Finally, he nods, gently and steadily bending his knees, and deliberately placing the extinguisher upright on the floor, his movements precise and measured. The entire scene is bathed in bright, cool fluorescent light from an overhead light source, creating clear reflections on the extinguisher's smooth surface and establishing an atmosphere of focused instruction.", + "seed": 13921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_4541_seed_13921_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03631749302148819, + "vlm_logprob": -0.03631749302148819, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03631749302148819, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029242299497127533 + ], + "opposite_logprob_tokens": [ + -3.9042422771453857 + ], + "logprob_reward": -0.029242299497127533, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029242299497127533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033916812390089035, + 0.012573471292853355, + 0.02277219481766224, + 0.047186028212308884, + 0.6339883208274841, + 0.10896684229373932, + 0.036307983100414276, + 0.02774900197982788, + 0.014599448069930077, + 0.06193986535072327 + ], + "credit_map": [ + 0.00578691903501749, + 0.003913376480340958, + 0.002945770975202322, + 0.002034761244431138, + 0.0029590597841888666, + 0.0076532866805791855, + 0.002466802019625902, + 0.0061568389646708965, + 0.0017720033647492528, + 0.0017563805449754, + 0.0012430889764800668, + 0.000904685992281884, + 0.0012043422320857644, + 0.002405531005933881, + 0.0016935967141762376, + 0.0015938442666083574, + 0.002888291608542204, + 0.005054679233580828, + 0.0018266686238348484, + 0.002121726982295513, + 0.003566431114450097, + 0.003577723866328597, + 0.0020722467452287674, + 0.001664427574723959, + 0.009578547440469265, + 0.010621276684105396, + 0.005875352770090103, + 0.005134562496095896, + 0.004334411583840847, + 0.00574503093957901, + 0.003224731422960758, + 0.0026721202302724123, + 0.15941597521305084, + 0.022155411541461945, + 0.2755555808544159, + 0.0314403660595417, + 0.023778360337018967, + 0.017308823764324188, + 0.0710059106349945, + 0.03332793340086937, + 0.0077003249898552895, + 0.008426460437476635, + 0.034765973687171936, + 0.0036345110274851322, + 0.035354163497686386, + 0.004776539746671915, + 0.004338223487138748, + 0.009970645420253277, + 0.0038514994084835052, + 0.003224093234166503, + 0.00370596069842577, + 0.0040002986788749695, + 0.004138915799558163, + 0.005725101567804813, + 0.004222201183438301, + 0.007439917419105768, + 0.003268861910328269, + 0.0036568674258887768, + 0.0032598101533949375, + 0.0037601825315505266, + 0.0035990849137306213, + 0.0014159317361190915, + 0.005316197406500578, + 0.0034720676485449076, + 0.0018625204684212804, + 0.0020563986618071795, + 0.0016870849067345262, + 0.0012714380864053965, + 0.001919581787660718, + 0.0021086372435092926, + 0.001244052778929472, + 0.0024497348349541426, + 0.008065854199230671, + 0.010685077868402004, + 0.006572393700480461, + 0.009993284940719604, + 0.0066146706230938435, + 0.008575210347771645, + 0.006823655217885971, + 0.004609720781445503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.029242299497127533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802891872823238 + ], + "opposite_logprob_tokens": [ + -4.63480281829834 + ], + "logprob_reward": -0.009802891872823238, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802891872823238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0467192567884922, + 0.06176956743001938, + 0.04843176528811455, + 0.04153425991535187, + 0.5416605472564697, + 0.04170829802751541, + 0.08410603553056717, + 0.06620009988546371, + 0.01497607957571745, + 0.052894044667482376 + ], + "credit_map": [ + 0.0060552977956831455, + 0.008154934272170067, + 0.010498803108930588, + 0.005399174522608519, + 0.003522559069097042, + 0.005760632921010256, + 0.003126953961327672, + 0.004200908355414867, + 0.008847297169268131, + 0.009930280968546867, + 0.008709890767931938, + 0.0035593812353909016, + 0.00559586938470602, + 0.016164390370249748, + 0.004122122656553984, + 0.004840339533984661, + 0.011218591593205929, + 0.016359949484467506, + 0.00348531361669302, + 0.002438361756503582, + 0.0031166968401521444, + 0.007261502556502819, + 0.0024057282134890556, + 0.0021456237882375717, + 0.0064342389814555645, + 0.008915466256439686, + 0.004308748058974743, + 0.006036458536982536, + 0.0036109716165810823, + 0.0063108718022704124, + 0.0033690372947603464, + 0.0025484738871455193, + 0.131712406873703, + 0.018626444041728973, + 0.23613783717155457, + 0.027883736416697502, + 0.020341748371720314, + 0.016121961176395416, + 0.06058184802532196, + 0.030254565179347992, + 0.004343047738075256, + 0.0028217346407473087, + 0.013539141975343227, + 0.0028491495177149773, + 0.008107973262667656, + 0.0021733336616307497, + 0.0033004917204380035, + 0.004573431797325611, + 0.008308998309075832, + 0.008316249586641788, + 0.007884401828050613, + 0.010400312021374702, + 0.009700817987322807, + 0.014219931326806545, + 0.010147768072783947, + 0.01512756198644638, + 0.006885282229632139, + 0.00896396767348051, + 0.007987011224031448, + 0.00938411708921194, + 0.008406245149672031, + 0.0033944027964025736, + 0.013881749473512173, + 0.007297327741980553, + 0.0016227200394496322, + 0.0023772187996655703, + 0.0016580981900915504, + 0.0016901730559766293, + 0.0022373239044100046, + 0.0013670118059962988, + 0.0018701201770454645, + 0.0021534147672355175, + 0.0066326335072517395, + 0.009063939563930035, + 0.005592224653810263, + 0.007838808000087738, + 0.006043982692062855, + 0.008133026771247387, + 0.0045164222829043865, + 0.005073014181107283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.009802891872823238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110340960323811 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110340960323811, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110340960323811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05582752823829651, + 0.04600398615002632, + 0.026410626247525215, + 0.028540214523673058, + 0.26686519384384155, + 0.19694861769676208, + 0.10712471604347229, + 0.08975140005350113, + 0.10057629644870758, + 0.0819515660405159 + ], + "credit_map": [ + 0.007922651246190071, + 0.010126041248440742, + 0.010308452881872654, + 0.003806409426033497, + 0.0050506312400102615, + 0.007999292574822903, + 0.004259509965777397, + 0.006354530807584524, + 0.0052278367802500725, + 0.0053285956382751465, + 0.005846055224537849, + 0.002856940496712923, + 0.005524095613509417, + 0.005722328554838896, + 0.010035277344286442, + 0.005462846253067255, + 0.00393154751509428, + 0.0064214435406029224, + 0.002021410269662738, + 0.0017840511864051223, + 0.0025304316077381372, + 0.005252220667898655, + 0.0021747590508311987, + 0.0022947557736188173, + 0.003288649022579193, + 0.00721865426748991, + 0.0027862321585416794, + 0.004098382778465748, + 0.003220592625439167, + 0.00451138848438859, + 0.0021381268743425608, + 0.0012781813275068998, + 0.06223224848508835, + 0.009550688788294792, + 0.11476747691631317, + 0.015709303319454193, + 0.011583421379327774, + 0.009203480556607246, + 0.029240524396300316, + 0.014577997848391533, + 0.014645189978182316, + 0.014087098650634289, + 0.06693906337022781, + 0.007026881445199251, + 0.06901077181100845, + 0.007780138403177261, + 0.00681894738227129, + 0.010640502907335758, + 0.012674711644649506, + 0.010508913546800613, + 0.009361214935779572, + 0.014864902943372726, + 0.011709306389093399, + 0.01752452924847603, + 0.01431454811245203, + 0.016166560351848602, + 0.00973505713045597, + 0.01140273455530405, + 0.011590955778956413, + 0.01342443935573101, + 0.011148883029818535, + 0.004431485664099455, + 0.018589777871966362, + 0.009428045712411404, + 0.009938360191881657, + 0.013722503557801247, + 0.01058267056941986, + 0.010084970854222775, + 0.01738230139017105, + 0.011905667372047901, + 0.0140272481366992, + 0.0129325520247221, + 0.007215599063783884, + 0.011648469604551792, + 0.008031963370740414, + 0.0130302170291543, + 0.009250507690012455, + 0.0063791717402637005, + 0.02015233226120472, + 0.006243289913982153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.004110340960323811 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003231555223465 + ], + "opposite_logprob_tokens": [ + -2.35032320022583 + ], + "logprob_reward": -0.1003231555223465, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003231555223465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04333996772766113, + 0.0904957503080368, + 0.12030506134033203, + 0.06656202673912048, + 0.33555805683135986, + 0.05917394906282425, + 0.10052425414323807, + 0.05748514458537102, + 0.0581558495759964, + 0.06839992105960846 + ], + "credit_map": [ + 0.00537949288263917, + 0.006731660105288029, + 0.004332526586949825, + 0.005440277513116598, + 0.006305241025984287, + 0.00570962717756629, + 0.005591968074440956, + 0.0038491738960146904, + 0.010647608898580074, + 0.010824267752468586, + 0.005478211212903261, + 0.00611191987991333, + 0.009957539848983288, + 0.014684721827507019, + 0.0180959515273571, + 0.014695528894662857, + 0.02480141445994377, + 0.031189177185297012, + 0.006703975144773722, + 0.012976918369531631, + 0.011244609020650387, + 0.012820946052670479, + 0.014017668552696705, + 0.006550351157784462, + 0.009063840843737125, + 0.012966232374310493, + 0.00661948136985302, + 0.011263956315815449, + 0.007760976441204548, + 0.008021567948162556, + 0.0055131614208221436, + 0.005352802574634552, + 0.08244334161281586, + 0.012715676799416542, + 0.14053861796855927, + 0.01695958338677883, + 0.01927475817501545, + 0.009438471868634224, + 0.03652752935886383, + 0.017660055309534073, + 0.004770574159920216, + 0.006099025718867779, + 0.014830321073532104, + 0.0037512467242777348, + 0.01741926558315754, + 0.003949263598769903, + 0.004209547769278288, + 0.004144703503698111, + 0.009766294620931149, + 0.01414974220097065, + 0.008647584356367588, + 0.01402912475168705, + 0.013428916223347187, + 0.0126948282122612, + 0.011780885048210621, + 0.016026876866817474, + 0.005142473615705967, + 0.007607793435454369, + 0.004121473059058189, + 0.013247865252196789, + 0.00759071484208107, + 0.0034186290577054024, + 0.010024978779256344, + 0.006331218406558037, + 0.00600411556661129, + 0.00823941919952631, + 0.00533165130764246, + 0.006849453318864107, + 0.008635226637125015, + 0.008277278393507004, + 0.00885789655148983, + 0.005960807669907808, + 0.006967524066567421, + 0.009088464081287384, + 0.007727756164968014, + 0.011769639328122139, + 0.007180237676948309, + 0.007893667556345463, + 0.009979713708162308, + 0.0077929203398525715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.1003231555223465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810877725481987 + ], + "opposite_logprob_tokens": [ + -3.2881088256835938 + ], + "logprob_reward": -0.03810877725481987, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810877725481987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0908140018582344, + 0.03599310666322708, + 0.04610570892691612, + 0.14580492675304413, + 0.2103794366121292, + 0.19583342969417572, + 0.07142596691846848, + 0.14716006815433502, + 0.013586612418293953, + 0.042896755039691925 + ], + "credit_map": [ + 0.012944746762514114, + 0.013416089117527008, + 0.00905646849423647, + 0.009112527593970299, + 0.007929976098239422, + 0.020629240199923515, + 0.0069772545248270035, + 0.010747701860964298, + 0.004521522205322981, + 0.004809272941201925, + 0.004212255589663982, + 0.002361846389248967, + 0.003667996497824788, + 0.008104912005364895, + 0.005297055933624506, + 0.0030182451009750366, + 0.007493549957871437, + 0.012203756719827652, + 0.004684151615947485, + 0.0027777799405157566, + 0.004866392817348242, + 0.007695756386965513, + 0.003814059542492032, + 0.002570263110101223, + 0.024199487641453743, + 0.0327778197824955, + 0.01940138451755047, + 0.02028084173798561, + 0.010116181336343288, + 0.01999903842806816, + 0.011869311332702637, + 0.007160861045122147, + 0.0483267642557621, + 0.007744687609374523, + 0.08969693630933762, + 0.013647409155964851, + 0.009034859016537666, + 0.007548948284238577, + 0.02297232858836651, + 0.011407498270273209, + 0.016043514013290405, + 0.021316908299922943, + 0.048573870211839676, + 0.007427498232573271, + 0.07059239596128464, + 0.00893208384513855, + 0.008594639599323273, + 0.014352531172335148, + 0.007174128666520119, + 0.007446840405464172, + 0.008590778335928917, + 0.008120878599584103, + 0.008626184426248074, + 0.011080889031291008, + 0.008935388177633286, + 0.011450878344476223, + 0.01480949018150568, + 0.019285043701529503, + 0.01672203838825226, + 0.020391328260302544, + 0.01873587816953659, + 0.007407120428979397, + 0.031202109530568123, + 0.018607070669531822, + 0.0017377277836203575, + 0.0017557363025844097, + 0.0017294707940891385, + 0.0012367964955046773, + 0.0019717917311936617, + 0.0012724233092740178, + 0.001707523362711072, + 0.0021751434542238712, + 0.005634440574795008, + 0.005369257647544146, + 0.004767692182213068, + 0.007329246029257774, + 0.005001612473279238, + 0.005811498500406742, + 0.00519559858366847, + 0.003787406487390399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.03810877725481987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium-sized shot captures a man in his 30s with neatly cropped brown hair standing in a well-lit, organized workshop. Dressed in a navy short-sleeved polo shirt and khaki pants, he exudes a professional demeanor. The background consists of a clean concrete floor and a slightly out-of-focus workbench with neatly arranged tools. The man tightly grasps a vibrant red cylindrical fire extinguisher with both hands, its black nozzle and pressure gauge clearly visible. As he speaks directly to the camera with a serious and clear expression, he begins to vigorously shake the extinguisher up and down, a movement that conveys a sense of urgent preparedness. He then pauses, holding the extinguisher steady with his right hand and gesturing vigorously with his left hand, pointing his index finger at the nozzle to emphasize a key point in his explanation. Finally, he nods, gently and steadily bending his knees, and deliberately placing the extinguisher upright on the floor, his movements precise and measured. The entire scene is bathed in bright, cool fluorescent light from an overhead light source, creating clear reflections on the extinguisher's smooth surface and establishing an atmosphere of focused instruction.", + "seed": 13922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_4541_seed_13922_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.036814744817093015, + "vlm_logprob": -0.036814744817093015, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.036814744817093015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029233615845441818 + ], + "opposite_logprob_tokens": [ + -3.904233694076538 + ], + "logprob_reward": -0.029233615845441818, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029233615845441818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17921428382396698, + 0.06608710438013077, + 0.07328648120164871, + 0.06433407962322235, + 0.08968963474035263, + 0.050800543278455734, + 0.04503331705927849, + 0.1488981395959854, + 0.07583919912576675, + 0.20681726932525635 + ], + "credit_map": [ + 0.02756350114941597, + 0.01949622668325901, + 0.01980244182050228, + 0.01842823252081871, + 0.018070098012685776, + 0.027422292158007622, + 0.01521204598248005, + 0.033219438046216965, + 0.005711941048502922, + 0.004578730091452599, + 0.006463208701461554, + 0.012501298449933529, + 0.0074285357259213924, + 0.008854818530380726, + 0.010694900527596474, + 0.009853672236204147, + 0.011988772079348564, + 0.004724967293441296, + 0.0083309942856431, + 0.005632111337035894, + 0.00804968737065792, + 0.0144208874553442, + 0.00847496185451746, + 0.011664099991321564, + 0.007646610494703054, + 0.007212192751467228, + 0.008243395015597343, + 0.011417059227824211, + 0.006076278630644083, + 0.007674591615796089, + 0.006243383511900902, + 0.00982057023793459, + 0.022617487236857414, + 0.006791521795094013, + 0.012442759238183498, + 0.006517360918223858, + 0.006565576884895563, + 0.021241899579763412, + 0.005300549324601889, + 0.00821248535066843, + 0.007876244373619556, + 0.004945980850607157, + 0.005712303798645735, + 0.004250695463269949, + 0.008491686545312405, + 0.005065061617642641, + 0.007888700813055038, + 0.006569867022335529, + 0.007245111279189587, + 0.0033027969766408205, + 0.005695068743079901, + 0.003462818218395114, + 0.005740343127399683, + 0.005760473664849997, + 0.006936109159141779, + 0.006890593096613884, + 0.02098517119884491, + 0.029634149745106697, + 0.014196022413671017, + 0.010816019959747791, + 0.0254978034645319, + 0.018838293850421906, + 0.013669071719050407, + 0.015261608175933361, + 0.008370930328965187, + 0.013745496049523354, + 0.00882271770387888, + 0.012076415121555328, + 0.010055909864604473, + 0.006978108547627926, + 0.007992631755769253, + 0.007796987891197205, + 0.023041775450110435, + 0.029413925483822823, + 0.0187834445387125, + 0.022805463522672653, + 0.0244988314807415, + 0.02321665920317173, + 0.031195254996418953, + 0.0338619090616703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.029233615845441818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016094528138637543 + ], + "opposite_logprob_tokens": [ + -4.14109468460083 + ], + "logprob_reward": -0.016094528138637543, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016094528138637543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14818693697452545, + 0.03860326483845711, + 0.0708310455083847, + 0.07740038633346558, + 0.1058630719780922, + 0.022878995165228844, + 0.10105346143245697, + 0.21104532480239868, + 0.12243017554283142, + 0.1017073318362236 + ], + "credit_map": [ + 0.01968899369239807, + 0.018883116543293, + 0.017045874148607254, + 0.016341930255293846, + 0.011743858456611633, + 0.022088827565312386, + 0.01492366660386324, + 0.027470681816339493, + 0.005171310156583786, + 0.00457215029746294, + 0.0075494637712836266, + 0.0035117825027555227, + 0.0043016583658754826, + 0.005344397388398647, + 0.002394207287579775, + 0.005758300889283419, + 0.016643455252051353, + 0.0067496532574296, + 0.009226268157362938, + 0.003808538196608424, + 0.005749636795371771, + 0.016780808568000793, + 0.004404971841722727, + 0.0074677239172160625, + 0.013764124363660812, + 0.008091535419225693, + 0.009745231829583645, + 0.010578437708318233, + 0.007210632786154747, + 0.012767741456627846, + 0.006547815632075071, + 0.008694877848029137, + 0.025488678365945816, + 0.00787530466914177, + 0.014228726737201214, + 0.009744924493134022, + 0.007037789560854435, + 0.02501659095287323, + 0.006561155430972576, + 0.009909912012517452, + 0.003988925833255053, + 0.0022332787048071623, + 0.0032784754876047373, + 0.0026647227350622416, + 0.002373970113694668, + 0.004436143208295107, + 0.0015256961341947317, + 0.0023777871392667294, + 0.020555393770337105, + 0.010833846405148506, + 0.015607014298439026, + 0.00972910039126873, + 0.00797648262232542, + 0.015374261885881424, + 0.009863642044365406, + 0.011113734915852547, + 0.030792949721217155, + 0.05404412001371384, + 0.02530050091445446, + 0.013509197160601616, + 0.021030081436038017, + 0.035758957266807556, + 0.01124534010887146, + 0.019364189356565475, + 0.013228104449808598, + 0.023230956867337227, + 0.013504799455404282, + 0.01649410091340542, + 0.01353510469198227, + 0.01464575994759798, + 0.014720086008310318, + 0.013071281835436821, + 0.013628639280796051, + 0.011873183771967888, + 0.015052673406898975, + 0.011862607672810555, + 0.010694132186472416, + 0.01075754314661026, + 0.014515184797346592, + 0.013323376886546612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.016094528138637543 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025027638766914606 + ], + "opposite_logprob_tokens": [ + -6.002502918243408 + ], + "logprob_reward": -0.0025027638766914606, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025027638766914606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1878095120191574, + 0.08005000650882721, + 0.06734868139028549, + 0.11327461153268814, + 0.06173473596572876, + 0.03665368631482124, + 0.05076755955815315, + 0.18563051521778107, + 0.11739157140254974, + 0.09933912009000778 + ], + "credit_map": [ + 0.025003746151924133, + 0.013285817578434944, + 0.023544006049633026, + 0.022765502333641052, + 0.02114548534154892, + 0.0270242840051651, + 0.024396488443017006, + 0.03064417466521263, + 0.007933209650218487, + 0.005059769377112389, + 0.008472954854369164, + 0.013257404789328575, + 0.00872267596423626, + 0.009781881235539913, + 0.01452608872205019, + 0.012296025641262531, + 0.008419536985456944, + 0.008557799272239208, + 0.007892630994319916, + 0.009146017022430897, + 0.007885686121881008, + 0.00845640990883112, + 0.008809315040707588, + 0.00818129163235426, + 0.012748154811561108, + 0.011107951402664185, + 0.013443490490317345, + 0.022695280611515045, + 0.013408186845481396, + 0.010516833513975143, + 0.016855400055646896, + 0.01249930914491415, + 0.00980342272669077, + 0.00673965597525239, + 0.007746310904622078, + 0.008906426839530468, + 0.007419488858431578, + 0.008764146827161312, + 0.006474174093455076, + 0.005881108343601227, + 0.004687931388616562, + 0.0035727634094655514, + 0.004945293068885803, + 0.005751197692006826, + 0.003508396213874221, + 0.003305765800178051, + 0.006086032371968031, + 0.004796304274350405, + 0.00817677192389965, + 0.004107884131371975, + 0.006677658762782812, + 0.0038262135349214077, + 0.0053153447806835175, + 0.0065317703410983086, + 0.008699529804289341, + 0.007432387676090002, + 0.01820344850420952, + 0.024529239162802696, + 0.016889430582523346, + 0.015561195090413094, + 0.02487458847463131, + 0.02851771004498005, + 0.020975619554519653, + 0.0360792800784111, + 0.01365173514932394, + 0.02776501141488552, + 0.008744558319449425, + 0.010555630549788475, + 0.017542622983455658, + 0.013937264680862427, + 0.009032453410327435, + 0.016162296757102013, + 0.010381302796304226, + 0.01094738021492958, + 0.009932360611855984, + 0.014799337834119797, + 0.01192446518689394, + 0.010724326595664024, + 0.015471380203962326, + 0.015158567577600479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0025027638766914606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295311152935028 + ], + "opposite_logprob_tokens": [ + -2.2379531860351562 + ], + "logprob_reward": -0.11295311152935028, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295311152935028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11479218304157257, + 0.03732365369796753, + 0.056349314749240875, + 0.08950432389974594, + 0.07799571752548218, + 0.06547178328037262, + 0.06458576023578644, + 0.24359893798828125, + 0.1693677008152008, + 0.0810106098651886 + ], + "credit_map": [ + 0.01899280585348606, + 0.010397208854556084, + 0.013169960118830204, + 0.01389179565012455, + 0.012997090816497803, + 0.015741491690278053, + 0.018251560628414154, + 0.011350266635417938, + 0.0046634613536298275, + 0.0036577004939317703, + 0.005117164924740791, + 0.0044435374438762665, + 0.005097993649542332, + 0.004762580618262291, + 0.00566290644928813, + 0.003918307833373547, + 0.009669279679656029, + 0.005707727745175362, + 0.007292189169675112, + 0.006947833579033613, + 0.006560932844877243, + 0.009148248471319675, + 0.005530507769435644, + 0.005492593627423048, + 0.015570662915706635, + 0.008627106435596943, + 0.011471256613731384, + 0.00815984234213829, + 0.007441118825227022, + 0.01720317266881466, + 0.008523652330040932, + 0.012507511302828789, + 0.017136864364147186, + 0.006316852755844593, + 0.009792155586183071, + 0.010327368043363094, + 0.005975967738777399, + 0.014339915476739407, + 0.006248990073800087, + 0.007857597433030605, + 0.01455154363065958, + 0.006085690576583147, + 0.010792993009090424, + 0.006340357940644026, + 0.007326565217226744, + 0.010005693882703781, + 0.0037838490679860115, + 0.006585089955478907, + 0.011048462241888046, + 0.008918949402868748, + 0.008186036720871925, + 0.006671454757452011, + 0.006369762122631073, + 0.007727212738245726, + 0.008374903351068497, + 0.00728897750377655, + 0.028958963230252266, + 0.04644547402858734, + 0.026395542547106743, + 0.016881484538316727, + 0.020584234967827797, + 0.04590027779340744, + 0.014742042869329453, + 0.043690916150808334, + 0.01854511722922325, + 0.03481527045369148, + 0.02155335247516632, + 0.019128775224089622, + 0.024485714733600616, + 0.018030980601906776, + 0.019008973613381386, + 0.013799508102238178, + 0.009092175401747227, + 0.00753073301166296, + 0.009227828122675419, + 0.00875830464065075, + 0.007984696887433529, + 0.009700106456875801, + 0.010456600226461887, + 0.018260164186358452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.11295311152935028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328970469534397 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328970469534397, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328970469534397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09079302847385406, + 0.048868559300899506, + 0.04829905182123184, + 0.06718975305557251, + 0.04743463546037674, + 0.046330466866493225, + 0.06255292147397995, + 0.25613531470298767, + 0.16204975545406342, + 0.17034649848937988 + ], + "credit_map": [ + 0.013527528382837772, + 0.00854461919516325, + 0.00957180093973875, + 0.009570777416229248, + 0.012027647346258163, + 0.011654999107122421, + 0.013995721004903316, + 0.011899929493665695, + 0.0045444900169968605, + 0.003551220055669546, + 0.006895828060805798, + 0.007815125398337841, + 0.006087669637054205, + 0.0065351747907698154, + 0.007452747784554958, + 0.005986304022371769, + 0.00498352013528347, + 0.004635094199329615, + 0.004529042635113001, + 0.004544041585177183, + 0.008106700144708157, + 0.006848332472145557, + 0.00639771856367588, + 0.00825460348278284, + 0.010734977200627327, + 0.008489456959068775, + 0.00847690086811781, + 0.007582513149827719, + 0.00702908169478178, + 0.011463087052106857, + 0.006425526924431324, + 0.006988211069256067, + 0.007927069440484047, + 0.0056220958940684795, + 0.005456310231238604, + 0.005408633500337601, + 0.004840570967644453, + 0.007425670977681875, + 0.004742200952023268, + 0.006012082099914551, + 0.010437888093292713, + 0.0035776954609900713, + 0.008013253100216389, + 0.004405026789754629, + 0.004116271622478962, + 0.006987414322793484, + 0.003909483551979065, + 0.004883434157818556, + 0.01072310097515583, + 0.007995674386620522, + 0.009474942460656166, + 0.007437128107994795, + 0.006328405812382698, + 0.007750705815851688, + 0.0055015189573168755, + 0.007341444492340088, + 0.034352220594882965, + 0.060652825981378555, + 0.028541402891278267, + 0.01664753258228302, + 0.0229375958442688, + 0.04982399567961693, + 0.015018918551504612, + 0.028160836547613144, + 0.020010488107800484, + 0.04388046637177467, + 0.014698644168674946, + 0.014446448534727097, + 0.018887287005782127, + 0.020511934533715248, + 0.014367305673658848, + 0.015247191302478313, + 0.030939936637878418, + 0.015581476502120495, + 0.032970674335956573, + 0.017790045589208603, + 0.01431130152195692, + 0.014930792152881622, + 0.017904097214341164, + 0.025918172672390938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.02328970469534397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium-sized shot captures a man in his 30s with neatly cropped brown hair standing in a well-lit, organized workshop. Dressed in a navy short-sleeved polo shirt and khaki pants, he exudes a professional demeanor. The background consists of a clean concrete floor and a slightly out-of-focus workbench with neatly arranged tools. The man tightly grasps a vibrant red cylindrical fire extinguisher with both hands, its black nozzle and pressure gauge clearly visible. As he speaks directly to the camera with a serious and clear expression, he begins to vigorously shake the extinguisher up and down, a movement that conveys a sense of urgent preparedness. He then pauses, holding the extinguisher steady with his right hand and gesturing vigorously with his left hand, pointing his index finger at the nozzle to emphasize a key point in his explanation. Finally, he nods, gently and steadily bending his knees, and deliberately placing the extinguisher upright on the floor, his movements precise and measured. The entire scene is bathed in bright, cool fluorescent light from an overhead light source, creating clear reflections on the extinguisher's smooth surface and establishing an atmosphere of focused instruction.", + "seed": 13923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_1/index_4541_seed_13923_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030003429111093282, + "vlm_logprob": -0.030003429111093282, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.030003429111093282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029230257496237755 + ], + "opposite_logprob_tokens": [ + -3.9042303562164307 + ], + "logprob_reward": -0.029230257496237755, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029230257496237755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05591326206922531, + 0.01919100433588028, + 0.03247329220175743, + 0.011219245381653309, + 0.46626755595207214, + 0.047164641320705414, + 0.19785311818122864, + 0.07366485893726349, + 0.05938355624675751, + 0.03686951845884323 + ], + "credit_map": [ + 0.005378759931772947, + 0.006687588524073362, + 0.004822302144020796, + 0.005762971471995115, + 0.015485034324228764, + 0.006632041651755571, + 0.004943670704960823, + 0.006200889125466347, + 0.00205675489269197, + 0.0019989421125501394, + 0.002580232685431838, + 0.001983508002012968, + 0.004629592411220074, + 0.0019901515915989876, + 0.002086122753098607, + 0.0018656965112313628, + 0.00253790570423007, + 0.009534904733300209, + 0.0020705151837319136, + 0.008953932672739029, + 0.002399314660578966, + 0.003456461476162076, + 0.0021309067960828543, + 0.0013893485302105546, + 0.0011521129636093974, + 0.0010579328518360853, + 0.0013703657314181328, + 0.0012316616484895349, + 0.0019990773871541023, + 0.0012332410551607609, + 0.0017488745506852865, + 0.0014259772142395377, + 0.029209254309535027, + 0.027585482224822044, + 0.15606525540351868, + 0.15734721720218658, + 0.010814816690981388, + 0.028365764766931534, + 0.04069908708333969, + 0.016180632635951042, + 0.010624313727021217, + 0.014072132296860218, + 0.004228748846799135, + 0.002709076041355729, + 0.003314812434837222, + 0.006085634231567383, + 0.0023833513259887695, + 0.003746566828340292, + 0.02697773464024067, + 0.05942138284444809, + 0.014783110469579697, + 0.02056298777461052, + 0.01764870434999466, + 0.039986249059438705, + 0.008447583764791489, + 0.010025338269770145, + 0.011562950909137726, + 0.0072929272428154945, + 0.008382671512663364, + 0.011896197684109211, + 0.00601737666875124, + 0.009955509565770626, + 0.008242319338023663, + 0.010314897634088993, + 0.015186267904937267, + 0.004823520313948393, + 0.013706506229937077, + 0.003958128392696381, + 0.004363986197859049, + 0.006571022793650627, + 0.0030254325829446316, + 0.007748681586235762, + 0.006681288126856089, + 0.0031420262530446053, + 0.005327529273927212, + 0.002572022844105959, + 0.0030075344257056713, + 0.008060730993747711, + 0.002410554327070713, + 0.005667827557772398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.029230257496237755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964812356978655 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964812356978655, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964812356978655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10565708577632904, + 0.04930371791124344, + 0.15103723108768463, + 0.019539417698979378, + 0.25033441185951233, + 0.027495212852954865, + 0.036950044333934784, + 0.025640884414315224, + 0.20904886722564697, + 0.12499307841062546 + ], + "credit_map": [ + 0.00985674187541008, + 0.01401565596461296, + 0.01809968613088131, + 0.01706750877201557, + 0.014914501458406448, + 0.016569834202528, + 0.0055872551165521145, + 0.009545915760099888, + 0.00601767934858799, + 0.005832511931657791, + 0.009330490604043007, + 0.005862726364284754, + 0.0075460514053702354, + 0.0051518091931939125, + 0.005509935785084963, + 0.0040525151416659355, + 0.009657245129346848, + 0.05112861469388008, + 0.00978037342429161, + 0.04391380771994591, + 0.006946731358766556, + 0.016468705609440804, + 0.008966743014752865, + 0.004175030626356602, + 0.001851739245466888, + 0.0036522657610476017, + 0.003538705175742507, + 0.00337782921269536, + 0.0022370119113475084, + 0.0021614686120301485, + 0.0017082003178074956, + 0.0010122000239789486, + 0.016760950908064842, + 0.014559400267899036, + 0.08271872252225876, + 0.083613321185112, + 0.006901801563799381, + 0.015744304284453392, + 0.02131141908466816, + 0.008724491111934185, + 0.005517087876796722, + 0.004241830203682184, + 0.002775758970528841, + 0.002704764250665903, + 0.0028115822933614254, + 0.0047863805666565895, + 0.0029015783220529556, + 0.001756233163177967, + 0.0068351831287145615, + 0.008502631448209286, + 0.0028774274978786707, + 0.0028288268949836493, + 0.0036267379764467478, + 0.007366874720901251, + 0.0024884238373488188, + 0.0024239416234195232, + 0.0036769951693713665, + 0.0031891833059489727, + 0.0026277443394064903, + 0.004372011870145798, + 0.002690422348678112, + 0.0029393970035016537, + 0.003583548590540886, + 0.0025615831837058067, + 0.05778682231903076, + 0.01797245442867279, + 0.03467218205332756, + 0.01138406340032816, + 0.01840122975409031, + 0.031379662454128265, + 0.011344239115715027, + 0.02610824443399906, + 0.016438079997897148, + 0.008348517119884491, + 0.023267144337296486, + 0.012250924482941628, + 0.010266469791531563, + 0.0258360356092453, + 0.009470658376812935, + 0.019115259870886803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.005964812356978655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110340960323811 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110340960323811, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110340960323811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08248404413461685, + 0.025579271838068962, + 0.24760012328624725, + 0.06265662610530853, + 0.11334466189146042, + 0.041916802525520325, + 0.057536132633686066, + 0.052284397184848785, + 0.18870475888252258, + 0.1278931051492691 + ], + "credit_map": [ + 0.008656945079565048, + 0.009175499901175499, + 0.012790313921868801, + 0.007855192758142948, + 0.018044084310531616, + 0.009657998569309711, + 0.00763773825019598, + 0.008666273206472397, + 0.0025288718752563, + 0.002530595287680626, + 0.0042921146377921104, + 0.003781983396038413, + 0.003103717928752303, + 0.0024617298040539026, + 0.0029925797134637833, + 0.0038876777980476618, + 0.015922201797366142, + 0.0823315903544426, + 0.014365742914378643, + 0.07393057644367218, + 0.011599548161029816, + 0.026660846546292305, + 0.01585187204182148, + 0.006937745958566666, + 0.0047845724038779736, + 0.010548816993832588, + 0.00584019860252738, + 0.007854439318180084, + 0.009035787545144558, + 0.008100745268166065, + 0.008432778529822826, + 0.008059285581111908, + 0.008268711157143116, + 0.006195371504873037, + 0.032502833753824234, + 0.038704294711351395, + 0.003910773899406195, + 0.008232290856540203, + 0.0114445174112916, + 0.004085872787982225, + 0.006861924659460783, + 0.008647777140140533, + 0.0039809513837099075, + 0.0039403545670211315, + 0.004413470160216093, + 0.004840452689677477, + 0.0051911515183746815, + 0.004040719475597143, + 0.008196922019124031, + 0.016044318675994873, + 0.004132312256842852, + 0.006197638809680939, + 0.005282132420688868, + 0.01135976891964674, + 0.0032238943967968225, + 0.003099147230386734, + 0.005618199706077576, + 0.006772094406187534, + 0.005217135883867741, + 0.0077792261727154255, + 0.006019300315529108, + 0.006290791556239128, + 0.008860515430569649, + 0.005727136041969061, + 0.03729735687375069, + 0.02191442809998989, + 0.025411445647478104, + 0.01337144523859024, + 0.022569410502910614, + 0.03246679529547691, + 0.014461314305663109, + 0.021212557330727577, + 0.024092676118016243, + 0.01092529110610485, + 0.0142589146271348, + 0.010044735856354237, + 0.012295100837945938, + 0.022811608389019966, + 0.011148463003337383, + 0.022316325455904007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.004110340960323811 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205366179347038 + ], + "opposite_logprob_tokens": [ + -2.812053680419922 + ], + "logprob_reward": -0.06205366179347038, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205366179347038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05652794986963272, + 0.014996218495070934, + 0.04688732326030731, + 0.01611674576997757, + 0.3315277397632599, + 0.028271974995732307, + 0.115081787109375, + 0.079270139336586, + 0.26567041873931885, + 0.0456496924161911 + ], + "credit_map": [ + 0.004466920625418425, + 0.00582102220505476, + 0.0069311847910285, + 0.005711954552680254, + 0.010235895402729511, + 0.005727706477046013, + 0.007541239261627197, + 0.010092020966112614, + 0.0018643736839294434, + 0.002080711768940091, + 0.0017556976526975632, + 0.001731311553157866, + 0.0018514522816985846, + 0.001969411037862301, + 0.0016447360394522548, + 0.0020985230803489685, + 0.003091522492468357, + 0.015142697840929031, + 0.0029703383333981037, + 0.012629354372620583, + 0.002970397472381592, + 0.005109117366373539, + 0.0028962544165551662, + 0.0020776470191776752, + 0.0012823013821616769, + 0.0015365405706688762, + 0.001958776731044054, + 0.0021128873340785503, + 0.0031873479019850492, + 0.001554810325615108, + 0.0023442676756531, + 0.002139812568202615, + 0.020887991413474083, + 0.019559774547815323, + 0.11044472455978394, + 0.11166414618492126, + 0.008146420121192932, + 0.019860193133354187, + 0.028898565098643303, + 0.01206593681126833, + 0.004183283541351557, + 0.003344665514305234, + 0.0030022989958524704, + 0.0030559669248759747, + 0.004162251017987728, + 0.0029151758644729853, + 0.003754127537831664, + 0.0038542060647159815, + 0.014270161278545856, + 0.03069785051047802, + 0.008512614294886589, + 0.010255804285407066, + 0.012046429328620434, + 0.0270431786775589, + 0.005138130392879248, + 0.007117621134966612, + 0.007665333803743124, + 0.014507710933685303, + 0.007966283708810806, + 0.01139779482036829, + 0.008218459784984589, + 0.008006933145225048, + 0.013633052818477154, + 0.007874572649598122, + 0.04092860966920853, + 0.028199004009366035, + 0.037990521639585495, + 0.01913086697459221, + 0.038435086607933044, + 0.04862130805850029, + 0.0189933180809021, + 0.033371686935424805, + 0.006733120884746313, + 0.004597194492816925, + 0.004402310587465763, + 0.005131816025823355, + 0.004146374296396971, + 0.008630067110061646, + 0.006387866567820311, + 0.005620942916721106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.06205366179347038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s with neatly cropped brown hair wearing a navy short-sleeved polo shirt and khaki pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a well-lit, organized workshop with a clean concrete floor and a slightly out-of-focus workbench with tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a vibrant red cylindrical fire extinguisher with a black nozzle and pressure gauge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man shake the extinguisher up and down vigorously before placing it upright on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, medium-sized shot with bright, cool fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865807294845581 + ], + "opposite_logprob_tokens": [ + -3.0486581325531006 + ], + "logprob_reward": -0.04865807294845581, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865807294845581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057469308376312256, + 0.034162480384111404, + 0.07251686602830887, + 0.017175493761897087, + 0.16751398146152496, + 0.027449505403637886, + 0.017686471343040466, + 0.03504123166203499, + 0.5305467247962952, + 0.04043792933225632 + ], + "credit_map": [ + 0.004309463314712048, + 0.007988186553120613, + 0.004899052903056145, + 0.006994377821683884, + 0.01884453371167183, + 0.0077062975615262985, + 0.002671359572559595, + 0.004056037403643131, + 0.003602463286370039, + 0.0038255411200225353, + 0.005022770259529352, + 0.004457446746528149, + 0.005227037705481052, + 0.00333976186811924, + 0.005824797786772251, + 0.002862662775442004, + 0.004522888921201229, + 0.022678282111883163, + 0.004270115867257118, + 0.020317204296588898, + 0.00448747631162405, + 0.00746934674680233, + 0.005599863827228546, + 0.003171689109876752, + 0.0016629479359835386, + 0.004028098192065954, + 0.001931288163177669, + 0.0026742585469037294, + 0.001823527505621314, + 0.0020525106228888035, + 0.0021067094057798386, + 0.000896153156645596, + 0.010767159052193165, + 0.009201831184327602, + 0.054960839450359344, + 0.0571005716919899, + 0.0040315892547369, + 0.010506371967494488, + 0.015186531469225883, + 0.0057590920478105545, + 0.005320698954164982, + 0.007336669135838747, + 0.0025954770389944315, + 0.0016196920769289136, + 0.0030280936043709517, + 0.0043972269631922245, + 0.001961392816156149, + 0.0011902565602213144, + 0.0032838592305779457, + 0.0030159293673932552, + 0.0013232730561867356, + 0.0018612050916999578, + 0.002482220297679305, + 0.0026033406611531973, + 0.0013453643769025803, + 0.0017712799599394202, + 0.0038751461543142796, + 0.003840136341750622, + 0.003739154664799571, + 0.006474133115261793, + 0.003632918233051896, + 0.003669080790132284, + 0.005801855586469173, + 0.004008808173239231, + 0.14698122441768646, + 0.04277759790420532, + 0.11633934825658798, + 0.02962579019367695, + 0.03330910578370094, + 0.06795898824930191, + 0.021725529804825783, + 0.07182915508747101, + 0.005278788041323423, + 0.0038913257885724306, + 0.004676048643887043, + 0.00296152476221323, + 0.003959267865866423, + 0.01013820432126522, + 0.0025838795118033886, + 0.0069488901644945145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.04865807294845581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_tensor([7087, 7087])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_tensor([7087, 7087])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4aba287b57fcca1451acb6fc6778c2731aff306c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_tensor([7087, 7087])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle tracking shot sweeps alongside an origami figure, capturing it as it negotiates a high-speed turn through a metropolis made entirely of paper. The camera dynamically follows the racer, maintaining sharp focus as the background blurs with motion. Folded from a single sheet of vibrant, matte red paper, the figure mimics the aggressive stance of a professional motorcycle racer, its body tilted at an extreme angle, almost parallel to the paper ground. Its form is a masterpiece of intricate folding, with sharp creases defining its helmeted head, angular shoulders, and hunched limbs for maximum aerodynamics. The vehicle is a large, intricately folded wheel made of smooth black paper, its structure suggestive of spokes and a tread pattern formed by clever geometric pleats. The entire scene is rendered in ultra-sharp 4K, revealing the crisp paper texture and subtle fiber grain of the material. As the wheel rounds a corner on the gray paper street, the dynamic movement is so intense that the towering, soft-colored origami skyscrapers in the background transform into vertical ribbons of color, creating a powerful sense of speed and a thrilling, focused atmosphere. Bright, direct lighting illuminates the scene, casting sharp, clean shadows and highlighting every fold and edge, making the surreal yet tangible world pop with crystal clarity.", + "seed": 19899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_7087_seed_19899_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08278514612466097, + "vlm_logprob": -0.08278514612466097, + "vlm_yes_no_margin": 3.324999988079071, + "vlm_token_credit": -0.08278514612466097, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036547161638736725 + ], + "opposite_logprob_tokens": [ + -3.4115471839904785 + ], + "logprob_reward": -0.036547161638736725, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036547161638736725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30376100540161133, + 0.06096377968788147, + 0.028065595775842667, + 0.0637531578540802, + 0.09782476723194122, + 0.22217664122581482, + 0.04758186265826225, + 0.029244007542729378, + 0.06978041678667068, + 0.07684879750013351 + ], + "credit_map": [ + 0.030926203355193138, + 0.030513642355799675, + 0.04719872027635574, + 0.037639833986759186, + 0.03366393223404884, + 0.05457281693816185, + 0.04641377553343773, + 0.02283204346895218, + 0.00782669149339199, + 0.008109508082270622, + 0.007307383697479963, + 0.007800845894962549, + 0.004776281304657459, + 0.006563075352460146, + 0.011031892150640488, + 0.007548096589744091, + 0.0036171143874526024, + 0.003053882857784629, + 0.003779167076572776, + 0.002773083746433258, + 0.0025663613341748714, + 0.005322880577296019, + 0.0042204540222883224, + 0.0027326513081789017, + 0.00754967238754034, + 0.008847388438880444, + 0.00971574429422617, + 0.008652162738144398, + 0.010672132484614849, + 0.0062362817116081715, + 0.006141388788819313, + 0.005938380025327206, + 0.01201982144266367, + 0.014244668185710907, + 0.01297806203365326, + 0.01122479047626257, + 0.01508937869220972, + 0.009823421947658062, + 0.010190206579864025, + 0.012254404835402966, + 0.028869062662124634, + 0.03072355128824711, + 0.02446606568992138, + 0.02518358640372753, + 0.025805236771702766, + 0.0242548156529665, + 0.03350413590669632, + 0.029370158910751343, + 0.0034531543496996164, + 0.013064616359770298, + 0.0029078200459480286, + 0.003117011161521077, + 0.005200880579650402, + 0.002874711761251092, + 0.005009773653000593, + 0.011953892186284065, + 0.00421938207000494, + 0.002474145730957389, + 0.003945521079003811, + 0.002984407125040889, + 0.0026652079541236162, + 0.005478458013385534, + 0.00431010453030467, + 0.0031667801085859537, + 0.008125470019876957, + 0.012251515872776508, + 0.009174635633826256, + 0.01073822658509016, + 0.008709313347935677, + 0.00569053553044796, + 0.009623137302696705, + 0.005467568524181843, + 0.009267247281968594, + 0.007542482577264309, + 0.011212059296667576, + 0.00584095623344183, + 0.015099724754691124, + 0.010769020766019821, + 0.009274901822209358, + 0.007842388935387135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.036547161638736725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790733456611633 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790733456611633, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790733456611633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2732473611831665, + 0.05380217358469963, + 0.024643005803227425, + 0.10132522135972977, + 0.05502378195524216, + 0.3620447516441345, + 0.021214095875620842, + 0.02554112672805786, + 0.05684072896838188, + 0.026317855343222618 + ], + "credit_map": [ + 0.035616252571344376, + 0.033950041979551315, + 0.05817434936761856, + 0.02724631503224373, + 0.02050304226577282, + 0.05206792801618576, + 0.02792203798890114, + 0.01776735670864582, + 0.005038324743509293, + 0.007841764949262142, + 0.006917024031281471, + 0.0075768958777189255, + 0.004760509356856346, + 0.0056529478169977665, + 0.01059779617935419, + 0.0054169087670743465, + 0.0036201849579811096, + 0.0023170262575149536, + 0.0030957551207393408, + 0.002564354334026575, + 0.0018077190034091473, + 0.003919357433915138, + 0.00411733565852046, + 0.0032012693118304014, + 0.011959963478147984, + 0.01609322614967823, + 0.013852749951183796, + 0.013985414057970047, + 0.019168173894286156, + 0.009203783236443996, + 0.009289435110986233, + 0.007772475481033325, + 0.008321212604641914, + 0.006988379638642073, + 0.007552231661975384, + 0.006029452197253704, + 0.007229906041175127, + 0.005154912825673819, + 0.007138627581298351, + 0.006609051954001188, + 0.037013594061136246, + 0.04452041909098625, + 0.039328403770923615, + 0.04433203116059303, + 0.04574320837855339, + 0.038569044321775436, + 0.06021881103515625, + 0.05231921374797821, + 0.0017950809560716152, + 0.0021031550131738186, + 0.0027000519912689924, + 0.0026014556642621756, + 0.002655668882653117, + 0.0036963187158107758, + 0.002914915792644024, + 0.0027474474627524614, + 0.003385921474546194, + 0.002624531975015998, + 0.0037859200965613127, + 0.0023831725120544434, + 0.002296938095241785, + 0.0039433762431144714, + 0.004408496432006359, + 0.0027127692010253668, + 0.006930210627615452, + 0.008593943901360035, + 0.006225925870239735, + 0.008487140759825706, + 0.00974238757044077, + 0.00521397078409791, + 0.007160948123782873, + 0.004486200399696827, + 0.0024734840262681246, + 0.0024035705719143152, + 0.0031008042860776186, + 0.002346082590520382, + 0.006599812302738428, + 0.0033135050907731056, + 0.0035131219774484634, + 0.0025674737989902496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.009790733456611633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081461794674397 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081461794674397, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081461794674397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5829340219497681, + 0.09013867378234863, + 0.015950053930282593, + 0.054661478847265244, + 0.023049725219607353, + 0.17224064469337463, + 0.01200647559016943, + 0.017471062019467354, + 0.01622319221496582, + 0.015324696898460388 + ], + "credit_map": [ + 0.10911758989095688, + 0.06804385036230087, + 0.108448825776577, + 0.03940321505069733, + 0.07457566261291504, + 0.09976829588413239, + 0.046363603323698044, + 0.03721294552087784, + 0.009385897777974606, + 0.014730318449437618, + 0.008457419462502003, + 0.012404165230691433, + 0.009191189892590046, + 0.009468755684792995, + 0.01746968924999237, + 0.009031233377754688, + 0.0035799650941044092, + 0.0015186761738732457, + 0.002283854875713587, + 0.0014637255808338523, + 0.00140211614780128, + 0.0013132077874615788, + 0.0024165159557014704, + 0.0019719903357326984, + 0.005964171141386032, + 0.007717698812484741, + 0.005900371354073286, + 0.007472868077456951, + 0.011281300336122513, + 0.006688231602311134, + 0.0048986077308654785, + 0.004738220479339361, + 0.0028432905673980713, + 0.0031165045220404863, + 0.003120201174169779, + 0.002434932626783848, + 0.0031338108237832785, + 0.002817607717588544, + 0.002872616518288851, + 0.002710757777094841, + 0.02028779126703739, + 0.022991415113210678, + 0.01807406172156334, + 0.018820159137248993, + 0.0207293052226305, + 0.019576754420995712, + 0.027507079765200615, + 0.024254068732261658, + 0.0008350945427082479, + 0.002956132870167494, + 0.0006711377645842731, + 0.0010280159767717123, + 0.0013070531422272325, + 0.0009161242051050067, + 0.0014380031498149037, + 0.0028549134731292725, + 0.0023393104784190655, + 0.0015337920049205422, + 0.002437916351482272, + 0.0017542594578117132, + 0.0017057409277185798, + 0.0024685978423804045, + 0.003104399424046278, + 0.002127045067027211, + 0.0018398623215034604, + 0.001777510973624885, + 0.0022324391175061464, + 0.0021909417118877172, + 0.001835850183852017, + 0.0029999196995049715, + 0.0016368330689147115, + 0.001709834556095302, + 0.0016859311144798994, + 0.0014994367957115173, + 0.002430273685604334, + 0.001358024775981903, + 0.0030105046462267637, + 0.0019112786976620555, + 0.0018785626161843538, + 0.0015506836352869868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011081461794674397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345850229263306 + ], + "opposite_logprob_tokens": [ + -1.3134584426879883 + ], + "logprob_reward": -0.31345850229263306, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31345850229263306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4407692849636078, + 0.0428345650434494, + 0.07147612422704697, + 0.04827476292848587, + 0.026718508452177048, + 0.14325103163719177, + 0.02780940942466259, + 0.09261105209589005, + 0.08360788971185684, + 0.02264733612537384 + ], + "credit_map": [ + 0.0801796168088913, + 0.05812884867191315, + 0.07772719115018845, + 0.03138415887951851, + 0.05974045395851135, + 0.06930452585220337, + 0.036981336772441864, + 0.027323191985487938, + 0.005731215700507164, + 0.0057367137633264065, + 0.004765140824019909, + 0.0048872437328100204, + 0.0035621756687760353, + 0.004188768565654755, + 0.00885752122849226, + 0.005105795804411173, + 0.0103166988119483, + 0.006584246642887592, + 0.008736897259950638, + 0.006918848492205143, + 0.0069538913667202, + 0.013822623528540134, + 0.009211075492203236, + 0.008931850083172321, + 0.005848044995218515, + 0.0073208375833928585, + 0.00609069736674428, + 0.006941569037735462, + 0.00903463363647461, + 0.004736820701509714, + 0.0041924165561795235, + 0.004109748173505068, + 0.002733555855229497, + 0.005430532153695822, + 0.0030072529334574938, + 0.0034124706871807575, + 0.0033780434168875217, + 0.0018230268033221364, + 0.003247531596571207, + 0.00368609931319952, + 0.014722464606165886, + 0.018110549077391624, + 0.014574805274605751, + 0.01635558158159256, + 0.01760585978627205, + 0.01678602397441864, + 0.023312892764806747, + 0.021782882511615753, + 0.0023605357855558395, + 0.0042099421843886375, + 0.0034250847529619932, + 0.0020765920635312796, + 0.003528050146996975, + 0.004428000655025244, + 0.004484997596591711, + 0.003296209964901209, + 0.011813553981482983, + 0.008779427036643028, + 0.01575588807463646, + 0.009319344535470009, + 0.00958780013024807, + 0.014908856712281704, + 0.012742291204631329, + 0.00970390159636736, + 0.009735357016324997, + 0.007693008985370398, + 0.011818774044513702, + 0.010799487121403217, + 0.009293085895478725, + 0.015997393056750298, + 0.009129400365054607, + 0.009141388349235058, + 0.002876604674383998, + 0.0023146013263612986, + 0.002825925126671791, + 0.0021023268345743418, + 0.003842882812023163, + 0.0035166009329259396, + 0.0026451742742210627, + 0.0025232217740267515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.31345850229263306 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304787144064903 + ], + "opposite_logprob_tokens": [ + -3.1680479049682617 + ], + "logprob_reward": -0.04304787144064903, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304787144064903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2268521785736084, + 0.053468625992536545, + 0.04875395447015762, + 0.10394913703203201, + 0.041857268661260605, + 0.29819783568382263, + 0.08251301944255829, + 0.04479895159602165, + 0.0684298649430275, + 0.031179223209619522 + ], + "credit_map": [ + 0.018008986487984657, + 0.023445794358849525, + 0.060453254729509354, + 0.017782015725970268, + 0.01854712888598442, + 0.06457864493131638, + 0.014738430269062519, + 0.009297922253608704, + 0.009643105790019035, + 0.009669795632362366, + 0.005927106365561485, + 0.0060763247311115265, + 0.004120221361517906, + 0.0058517009019851685, + 0.007494950667023659, + 0.004685416352003813, + 0.007607569452375174, + 0.0055230227299034595, + 0.005816389806568623, + 0.00615414185449481, + 0.004562365356832743, + 0.006462912540882826, + 0.006874358747154474, + 0.005753188859671354, + 0.01025085337460041, + 0.015317709185183048, + 0.01324116438627243, + 0.015166040509939194, + 0.02260197140276432, + 0.010458489879965782, + 0.00895779300481081, + 0.007955110631883144, + 0.005228023510426283, + 0.005124819464981556, + 0.0068339514546096325, + 0.005140106193721294, + 0.005671344697475433, + 0.004578975960612297, + 0.004747690167278051, + 0.0045323544181883335, + 0.03480095788836479, + 0.043866828083992004, + 0.027409333735704422, + 0.03369225189089775, + 0.035811323672533035, + 0.03483923152089119, + 0.046628013253211975, + 0.04114988446235657, + 0.0063268886879086494, + 0.020973175764083862, + 0.006007019896060228, + 0.005767674185335636, + 0.009870581328868866, + 0.004274049773812294, + 0.00922752358019352, + 0.020066097378730774, + 0.0075618801638484, + 0.00414417777210474, + 0.005173626355826855, + 0.004301963374018669, + 0.003954221494495869, + 0.007467512972652912, + 0.006616557948291302, + 0.00557900732383132, + 0.007393998093903065, + 0.006657698191702366, + 0.008091285824775696, + 0.00895195547491312, + 0.008006449788808823, + 0.014656320214271545, + 0.007047956343740225, + 0.007624201010912657, + 0.0029522739350795746, + 0.0028283370193094015, + 0.0039001908153295517, + 0.0024918955750763416, + 0.004854964092373848, + 0.004556143190711737, + 0.0034121416974812746, + 0.006183274555951357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.04304787144064903 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle tracking shot sweeps alongside an origami figure, capturing it as it negotiates a high-speed turn through a metropolis made entirely of paper. The camera dynamically follows the racer, maintaining sharp focus as the background blurs with motion. Folded from a single sheet of vibrant, matte red paper, the figure mimics the aggressive stance of a professional motorcycle racer, its body tilted at an extreme angle, almost parallel to the paper ground. Its form is a masterpiece of intricate folding, with sharp creases defining its helmeted head, angular shoulders, and hunched limbs for maximum aerodynamics. The vehicle is a large, intricately folded wheel made of smooth black paper, its structure suggestive of spokes and a tread pattern formed by clever geometric pleats. The entire scene is rendered in ultra-sharp 4K, revealing the crisp paper texture and subtle fiber grain of the material. As the wheel rounds a corner on the gray paper street, the dynamic movement is so intense that the towering, soft-colored origami skyscrapers in the background transform into vertical ribbons of color, creating a powerful sense of speed and a thrilling, focused atmosphere. Bright, direct lighting illuminates the scene, casting sharp, clean shadows and highlighting every fold and edge, making the surreal yet tangible world pop with crystal clarity.", + "seed": 19900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_7087_seed_19900_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01491705235093832, + "vlm_logprob": -0.01491705235093832, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01491705235093832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025794638320803642 + ], + "opposite_logprob_tokens": [ + -3.775794744491577 + ], + "logprob_reward": -0.025794638320803642, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025794638320803642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11205382645130157, + 0.04818012937903404, + 0.08339889347553253, + 0.20690201222896576, + 0.026774579659104347, + 0.18591059744358063, + 0.2146269828081131, + 0.04382960870862007, + 0.04469608888030052, + 0.03362734988331795 + ], + "credit_map": [ + 0.010576503351330757, + 0.012748529203236103, + 0.01336163841187954, + 0.009873220697045326, + 0.01691349595785141, + 0.01807226426899433, + 0.010309398174285889, + 0.020198775455355644, + 0.006987837143242359, + 0.005276200827211142, + 0.005379441659897566, + 0.006651361007243395, + 0.004236586391925812, + 0.0071386126801371574, + 0.005211174488067627, + 0.007298917043954134, + 0.009103002026677132, + 0.0043915314599871635, + 0.014908251352608204, + 0.005820207763463259, + 0.003983207978308201, + 0.026620663702487946, + 0.004623084794729948, + 0.013948941603302956, + 0.022517027333378792, + 0.008487045764923096, + 0.06368158757686615, + 0.008857044391334057, + 0.005342343356460333, + 0.01158882211893797, + 0.00814518891274929, + 0.07828294485807419, + 0.00448561692610383, + 0.0025734691880643368, + 0.003266308456659317, + 0.003359454683959484, + 0.002815490821376443, + 0.00537198968231678, + 0.0022352358791977167, + 0.002667014952749014, + 0.05844692140817642, + 0.009342112578451633, + 0.05880739167332649, + 0.0281680840998888, + 0.008690687827765942, + 0.005389545578509569, + 0.008584127761423588, + 0.008481723256409168, + 0.01795930787920952, + 0.018891558051109314, + 0.019920464605093002, + 0.04983365535736084, + 0.022818049415946007, + 0.02715812437236309, + 0.027443399652838707, + 0.03060241974890232, + 0.006629276089370251, + 0.003508166875690222, + 0.011419960297644138, + 0.005141888279467821, + 0.004997945390641689, + 0.004394491668790579, + 0.004190916661173105, + 0.003546962281689048, + 0.009618102572858334, + 0.004045173060148954, + 0.009493883699178696, + 0.005600404925644398, + 0.0035863746888935566, + 0.003938407637178898, + 0.0033728277776390314, + 0.005040912423282862, + 0.007106445729732513, + 0.002124690916389227, + 0.008988169953227043, + 0.004856197163462639, + 0.0025874741841107607, + 0.002885090187191963, + 0.0021646302193403244, + 0.0029146536253392696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.025794638320803642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953673273324966 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953673273324966, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953673273324966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12929868698120117, + 0.04361920803785324, + 0.07352149486541748, + 0.2640780210494995, + 0.11878019571304321, + 0.0741490051150322, + 0.09339780360460281, + 0.04395328834652901, + 0.08775115013122559, + 0.07145120203495026 + ], + "credit_map": [ + 0.03083517588675022, + 0.011680424213409424, + 0.02349701151251793, + 0.0072020478546619415, + 0.006858284585177898, + 0.023506702855229378, + 0.0064024124294519424, + 0.01931663043797016, + 0.0040183584205806255, + 0.006127944216132164, + 0.004967434331774712, + 0.010237356647849083, + 0.004676563199609518, + 0.003797756740823388, + 0.005550834350287914, + 0.004242961760610342, + 0.0077965096570551395, + 0.00504247797653079, + 0.012418961152434349, + 0.00548793887719512, + 0.004768203478306532, + 0.023130420595407486, + 0.004098695237189531, + 0.010778287425637245, + 0.027826746925711632, + 0.012265177443623543, + 0.07405681908130646, + 0.018323170021176338, + 0.009760621935129166, + 0.015806538984179497, + 0.016555169597268105, + 0.08948377519845963, + 0.017625734210014343, + 0.011600964702665806, + 0.01597978174686432, + 0.01592845283448696, + 0.01576988212764263, + 0.011968103237450123, + 0.012897919863462448, + 0.017009355127811432, + 0.01274195872247219, + 0.00543354544788599, + 0.012346399948000908, + 0.02236257493495941, + 0.006322972476482391, + 0.004111253656446934, + 0.005367413628846407, + 0.005462882108986378, + 0.008698557503521442, + 0.006413354072719812, + 0.013016892597079277, + 0.022841228172183037, + 0.009129240177571774, + 0.007404847536236048, + 0.012073937803506851, + 0.013819749467074871, + 0.0073259915225207806, + 0.003966719843447208, + 0.013074594549834728, + 0.00444800266996026, + 0.004654288291931152, + 0.0034404501784592867, + 0.0033364698756486177, + 0.0037067702505737543, + 0.03125269338488579, + 0.0037925669457763433, + 0.031773850321769714, + 0.005764933302998543, + 0.0038497152272611856, + 0.0037757500540465117, + 0.003302083583548665, + 0.004239561501890421, + 0.021229702979326248, + 0.004309795796871185, + 0.023803865537047386, + 0.007608301937580109, + 0.003550917375832796, + 0.0032445730175822973, + 0.003204967826604843, + 0.004499076399952173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.005953673273324966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11948899924755096, + 0.08107303082942963, + 0.16686490178108215, + 0.0878416895866394, + 0.03337092325091362, + 0.2161724716424942, + 0.08224352449178696, + 0.07973416894674301, + 0.1023079976439476, + 0.030902285128831863 + ], + "credit_map": [ + 0.018957901746034622, + 0.014050337485969067, + 0.01577477902173996, + 0.007761777378618717, + 0.01294170692563057, + 0.026560114696621895, + 0.009313163347542286, + 0.014129220508038998, + 0.010489686392247677, + 0.008067204616963863, + 0.009244636632502079, + 0.013446701690554619, + 0.005512491334229708, + 0.013050776906311512, + 0.00889038760215044, + 0.012371141463518143, + 0.017849024385213852, + 0.013058889657258987, + 0.024779148399829865, + 0.020093703642487526, + 0.013945085927844048, + 0.03964540734887123, + 0.011869596317410469, + 0.02562406286597252, + 0.008362715132534504, + 0.005093386862426996, + 0.02346762642264366, + 0.006798312999308109, + 0.004146823659539223, + 0.006328382063657045, + 0.004303692374378443, + 0.02934074215590954, + 0.0062206159345805645, + 0.004150315187871456, + 0.004672378301620483, + 0.0037504236679524183, + 0.0031915418803691864, + 0.004514068365097046, + 0.0035307228099554777, + 0.003340858733281493, + 0.04336594417691231, + 0.014322774484753609, + 0.041239168494939804, + 0.04586329683661461, + 0.019346339628100395, + 0.010994646698236465, + 0.023415112867951393, + 0.01762518286705017, + 0.007559021003544331, + 0.0064322249963879585, + 0.012410787865519524, + 0.019851112738251686, + 0.006698427256196737, + 0.006195778958499432, + 0.010011766105890274, + 0.013084404170513153, + 0.011709440499544144, + 0.004724097438156605, + 0.025141369551420212, + 0.010288191959261894, + 0.011632133275270462, + 0.006534289568662643, + 0.004045143723487854, + 0.005659506190568209, + 0.02436593547463417, + 0.006524396128952503, + 0.026630613952875137, + 0.007715173065662384, + 0.019924728199839592, + 0.006997177377343178, + 0.004638015292584896, + 0.0055119553580880165, + 0.008591977879405022, + 0.0025542655494064093, + 0.009621265344321728, + 0.0025727851316332817, + 0.0021957855205982924, + 0.0017047388246282935, + 0.0013731771614402533, + 0.002288288436830044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.00978341419249773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280037567019463 + ], + "opposite_logprob_tokens": [ + -3.773280143737793 + ], + "logprob_reward": -0.023280037567019463, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280037567019463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26320114731788635, + 0.028688188642263412, + 0.09485965967178345, + 0.1449650526046753, + 0.049229998141527176, + 0.20764398574829102, + 0.07707270234823227, + 0.04210596904158592, + 0.04750306159257889, + 0.04473023861646652 + ], + "credit_map": [ + 0.04693690687417984, + 0.026215536519885063, + 0.04637407884001732, + 0.01698465645313263, + 0.02697817049920559, + 0.047157544642686844, + 0.014157245866954327, + 0.03839700296521187, + 0.003423627931624651, + 0.003586437785997987, + 0.0036645252257585526, + 0.003374354913830757, + 0.002683453494682908, + 0.005853359587490559, + 0.0021971911191940308, + 0.003905234858393669, + 0.013370458967983723, + 0.008319806307554245, + 0.017080847173929214, + 0.007154108956456184, + 0.005062656011432409, + 0.025522714480757713, + 0.005529963411390781, + 0.012819097377359867, + 0.021317295730113983, + 0.006635936442762613, + 0.04209813103079796, + 0.006347835063934326, + 0.004253120627254248, + 0.0078103914856910706, + 0.005978006403893232, + 0.05052432045340538, + 0.010233432054519653, + 0.006253179628401995, + 0.007065408863127232, + 0.004524658899754286, + 0.0047128223814070225, + 0.006457189563661814, + 0.004590766504406929, + 0.005392536520957947, + 0.07820498943328857, + 0.009141416288912296, + 0.07979273796081543, + 0.016728220507502556, + 0.006037162616848946, + 0.0038335397839546204, + 0.005856336560100317, + 0.008049565367400646, + 0.011986834928393364, + 0.006729929242283106, + 0.017749745398759842, + 0.008296124637126923, + 0.006363603286445141, + 0.004056035075336695, + 0.006440845783799887, + 0.015449581667780876, + 0.006946427747607231, + 0.003689766861498356, + 0.010264693759381771, + 0.0043337284587323666, + 0.0045204758644104, + 0.004141787998378277, + 0.003361399983987212, + 0.0048476834781467915, + 0.01420584786683321, + 0.0025419716257601976, + 0.01476406492292881, + 0.00465362099930644, + 0.004332916811108589, + 0.0026119656395167112, + 0.0023641998413950205, + 0.0020284715574234724, + 0.011052492074668407, + 0.0027342962566763163, + 0.013652816414833069, + 0.005460549611598253, + 0.0026528420858085155, + 0.002582237357273698, + 0.002838020445778966, + 0.003756979713216424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.023280037567019463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097734984010458 + ], + "opposite_logprob_tokens": [ + -4.6347737312316895 + ], + "logprob_reward": -0.0097734984010458, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097734984010458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08269508183002472, + 0.03942989557981491, + 0.0997280403971672, + 0.28687918186187744, + 0.0813140943646431, + 0.04385081306099892, + 0.061988335102796555, + 0.05717096105217934, + 0.0981263592839241, + 0.1488173007965088 + ], + "credit_map": [ + 0.011533626355230808, + 0.012577655725181103, + 0.009527957066893578, + 0.0069755082949995995, + 0.012595188803970814, + 0.014263942837715149, + 0.007552533410489559, + 0.0076686604879796505, + 0.005854678340256214, + 0.0032059645745903254, + 0.005079173948615789, + 0.004647243767976761, + 0.0031182128004729748, + 0.006824469193816185, + 0.0032382397912442684, + 0.007461910601705313, + 0.011417737230658531, + 0.007591423112899065, + 0.01976688951253891, + 0.00725548155605793, + 0.004581885412335396, + 0.03222659230232239, + 0.0038927393034100533, + 0.012995285913348198, + 0.04670098423957825, + 0.011721045710146427, + 0.08017484843730927, + 0.014495954848825932, + 0.008474129252135754, + 0.01655086688697338, + 0.012042992748320103, + 0.09671832621097565, + 0.014055455103516579, + 0.010248275473713875, + 0.012434553354978561, + 0.00872589461505413, + 0.006642390973865986, + 0.008696497417986393, + 0.00705043924972415, + 0.013460577465593815, + 0.008689812384545803, + 0.0037790522910654545, + 0.008141430094838142, + 0.007390079088509083, + 0.004872533492743969, + 0.003498824778944254, + 0.004313534125685692, + 0.003165540285408497, + 0.0117696113884449, + 0.005828613881021738, + 0.013269063085317612, + 0.006991621572524309, + 0.0038270149379968643, + 0.004740694537758827, + 0.005090842489153147, + 0.010470868088304996, + 0.007158072665333748, + 0.005800532642751932, + 0.0071874517016112804, + 0.007726668380200863, + 0.012782510370016098, + 0.006800646428018808, + 0.004793238826096058, + 0.0049218288622796535, + 0.025216083973646164, + 0.007544186897575855, + 0.026792239397764206, + 0.013132303953170776, + 0.005161446053534746, + 0.006439610850065947, + 0.005449654534459114, + 0.008390817791223526, + 0.04581453651189804, + 0.011764354072511196, + 0.040796082466840744, + 0.011716692708432674, + 0.008187570609152317, + 0.008852570317685604, + 0.008780652657151222, + 0.012904823757708073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.0097734984010458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle tracking shot sweeps alongside an origami figure, capturing it as it negotiates a high-speed turn through a metropolis made entirely of paper. The camera dynamically follows the racer, maintaining sharp focus as the background blurs with motion. Folded from a single sheet of vibrant, matte red paper, the figure mimics the aggressive stance of a professional motorcycle racer, its body tilted at an extreme angle, almost parallel to the paper ground. Its form is a masterpiece of intricate folding, with sharp creases defining its helmeted head, angular shoulders, and hunched limbs for maximum aerodynamics. The vehicle is a large, intricately folded wheel made of smooth black paper, its structure suggestive of spokes and a tread pattern formed by clever geometric pleats. The entire scene is rendered in ultra-sharp 4K, revealing the crisp paper texture and subtle fiber grain of the material. As the wheel rounds a corner on the gray paper street, the dynamic movement is so intense that the towering, soft-colored origami skyscrapers in the background transform into vertical ribbons of color, creating a powerful sense of speed and a thrilling, focused atmosphere. Bright, direct lighting illuminates the scene, casting sharp, clean shadows and highlighting every fold and edge, making the surreal yet tangible world pop with crystal clarity.", + "seed": 19901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_7087_seed_19901_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019066299218684434, + "vlm_logprob": -0.019066299218684434, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.019066299218684434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04058618098497391 + ], + "opposite_logprob_tokens": [ + -3.290586233139038 + ], + "logprob_reward": -0.04058618098497391, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04058618098497391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14052055776119232, + 0.0654606893658638, + 0.058980122208595276, + 0.17041459679603577, + 0.062393203377723694, + 0.08773943036794662, + 0.2129228264093399, + 0.08386345207691193, + 0.01881750486791134, + 0.09888768196105957 + ], + "credit_map": [ + 0.011101139709353447, + 0.022203506901860237, + 0.021198401227593422, + 0.026030538603663445, + 0.015613858588039875, + 0.012094482779502869, + 0.01791694574058056, + 0.014361652545630932, + 0.005297134630382061, + 0.0031790975481271744, + 0.011252718977630138, + 0.005102174822241068, + 0.004551639314740896, + 0.012193749658763409, + 0.0062060365453362465, + 0.01767812855541706, + 0.011247930116951466, + 0.006108627188950777, + 0.005282441154122353, + 0.00828371662646532, + 0.0064748250879347324, + 0.005152693018317223, + 0.006211463827639818, + 0.010218414478003979, + 0.028789479285478592, + 0.02156762033700943, + 0.03342040255665779, + 0.014495494775474072, + 0.01856568641960621, + 0.0174618698656559, + 0.013046086765825748, + 0.02306791953742504, + 0.006425793282687664, + 0.005771205760538578, + 0.008782762102782726, + 0.00969192199409008, + 0.006624637171626091, + 0.00949780736118555, + 0.005228780675679445, + 0.010370288975536823, + 0.009050028398633003, + 0.011544731445610523, + 0.009284603409469128, + 0.010310374200344086, + 0.010799567215144634, + 0.008590295910835266, + 0.012622155249118805, + 0.015537668950855732, + 0.02948797680437565, + 0.006591601762920618, + 0.09348701685667038, + 0.011263283900916576, + 0.009104326367378235, + 0.007553829345852137, + 0.012375456281006336, + 0.04305930808186531, + 0.015342692844569683, + 0.009123723953962326, + 0.019825726747512817, + 0.007345461752265692, + 0.005811708047986031, + 0.0059279161505401134, + 0.009435911662876606, + 0.011050297878682613, + 0.002796122571453452, + 0.0018708015559241176, + 0.002087836153805256, + 0.002582357032224536, + 0.0014880085363984108, + 0.002063998021185398, + 0.0019146003760397434, + 0.004013777710497379, + 0.01686573214828968, + 0.006380635313689709, + 0.03167537599802017, + 0.007654153276234865, + 0.004153142683207989, + 0.009607014246284962, + 0.008138810284435749, + 0.01441280823200941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.04058618098497391 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041007245890796185 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.0041007245890796185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041007245890796185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1898350566625595, + 0.09178285300731659, + 0.08372415602207184, + 0.1893845498561859, + 0.07102400809526443, + 0.0988742932677269, + 0.054166607558727264, + 0.07769809663295746, + 0.05762172490358353, + 0.08588862419128418 + ], + "credit_map": [ + 0.016097374260425568, + 0.03358229622244835, + 0.027748331427574158, + 0.026348812505602837, + 0.022161338478326797, + 0.018434198573231697, + 0.022990858182311058, + 0.022471873089671135, + 0.011838158592581749, + 0.005962664261460304, + 0.014880468137562275, + 0.010825518518686295, + 0.007297567091882229, + 0.01775312051177025, + 0.010506179183721542, + 0.012719188816845417, + 0.0208881963044405, + 0.009846460074186325, + 0.007344263605773449, + 0.009971639141440392, + 0.00788366049528122, + 0.006318188738077879, + 0.007511841598898172, + 0.013959914445877075, + 0.03204698488116264, + 0.027129720896482468, + 0.034491732716560364, + 0.01551795657724142, + 0.02077045850455761, + 0.01836598664522171, + 0.014550632797181606, + 0.02651108056306839, + 0.007567828521132469, + 0.005904390476644039, + 0.011528734117746353, + 0.011707484722137451, + 0.0075675430707633495, + 0.010053756646811962, + 0.005443602800369263, + 0.011250671930611134, + 0.020978756248950958, + 0.00901380367577076, + 0.014309357851743698, + 0.009116841480135918, + 0.021526498720049858, + 0.006570895202457905, + 0.008931552059948444, + 0.008426600135862827, + 0.011787663213908672, + 0.0031328322365880013, + 0.016024595126509666, + 0.004652468953281641, + 0.004355079494416714, + 0.0022821491584181786, + 0.004513363353908062, + 0.007418453227728605, + 0.015112381428480148, + 0.00814304780215025, + 0.018351281061768532, + 0.006270545534789562, + 0.005190014839172363, + 0.0056778788566589355, + 0.008488251827657223, + 0.01046469621360302, + 0.011967490427196026, + 0.007124565076082945, + 0.005224233493208885, + 0.007346353959292173, + 0.006525268312543631, + 0.0072685470804572105, + 0.005567015148699284, + 0.006598255131393671, + 0.01253210473805666, + 0.00806751661002636, + 0.02114715240895748, + 0.010732854716479778, + 0.007847454398870468, + 0.005575140472501516, + 0.009129196405410767, + 0.01085721142590046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0041007245890796185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10196813941001892, + 0.06551209092140198, + 0.16544374823570251, + 0.16406826674938202, + 0.12162426859140396, + 0.04904962331056595, + 0.1591060906648636, + 0.04808517545461655, + 0.027814950793981552, + 0.09732760488986969 + ], + "credit_map": [ + 0.00845666229724884, + 0.02007836289703846, + 0.016209257766604424, + 0.017734810709953308, + 0.007508579641580582, + 0.012506491504609585, + 0.009246278554201126, + 0.010227708145976067, + 0.011775046586990356, + 0.004428272135555744, + 0.01161468680948019, + 0.007982979528605938, + 0.005484029185026884, + 0.009134837426245213, + 0.007690442260354757, + 0.0074018086306750774, + 0.01784244179725647, + 0.009708348661661148, + 0.017158640548586845, + 0.019227057695388794, + 0.010785573162138462, + 0.016921302303671837, + 0.014827627688646317, + 0.058972761034965515, + 0.030308451503515244, + 0.023131199181079865, + 0.02655516192317009, + 0.012956587597727776, + 0.014340151101350784, + 0.01820889301598072, + 0.011738868430256844, + 0.026828965172171593, + 0.01666172966361046, + 0.010614418424665928, + 0.016223391517996788, + 0.019064923748373985, + 0.012616516090929508, + 0.015962321311235428, + 0.012072761543095112, + 0.018408220261335373, + 0.007854343391954899, + 0.004258653149008751, + 0.007589028216898441, + 0.00534812593832612, + 0.005996872205287218, + 0.005383496172726154, + 0.006492895539849997, + 0.006126211490482092, + 0.031084921211004257, + 0.006327113136649132, + 0.07080690562725067, + 0.008724857121706009, + 0.007325826212763786, + 0.005202620290219784, + 0.0078284228220582, + 0.021805448457598686, + 0.008659095503389835, + 0.004988718777894974, + 0.011280491016805172, + 0.004528497811406851, + 0.0033605662174522877, + 0.003510843263939023, + 0.005628778133541346, + 0.0061281900852918625, + 0.00505294743925333, + 0.003370852442458272, + 0.003393234685063362, + 0.004411712288856506, + 0.002271869918331504, + 0.0024414658546447754, + 0.002817588159814477, + 0.004055282101035118, + 0.013768323697149754, + 0.008189844898879528, + 0.03109101578593254, + 0.00723537802696228, + 0.004839461762458086, + 0.006613244768232107, + 0.008879459463059902, + 0.016710884869098663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.00978058110922575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979704923927784 + ], + "opposite_logprob_tokens": [ + -3.529797077178955 + ], + "logprob_reward": -0.02979704923927784, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979704923927784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20972052216529846, + 0.08512666821479797, + 0.20019102096557617, + 0.04044090583920479, + 0.04438167065382004, + 0.028166165575385094, + 0.0732031837105751, + 0.030764982104301453, + 0.03723648563027382, + 0.25076842308044434 + ], + "credit_map": [ + 0.019736070185899734, + 0.03889594227075577, + 0.029113374650478363, + 0.03306926414370537, + 0.022557280957698822, + 0.025681471452116966, + 0.0178411565721035, + 0.022825950756669044, + 0.01461803913116455, + 0.0056556095369160175, + 0.014889450743794441, + 0.007524101063609123, + 0.0059570493176579475, + 0.01574154570698738, + 0.006506863981485367, + 0.014234010130167007, + 0.0586669035255909, + 0.025046145543456078, + 0.01816294714808464, + 0.02174926921725273, + 0.01840231567621231, + 0.015467009507119656, + 0.019103331491351128, + 0.02359309047460556, + 0.00579623831436038, + 0.004247567150741816, + 0.0068498398177325726, + 0.005490968003869057, + 0.0038260763976722956, + 0.0035824147053062916, + 0.0035714146215468645, + 0.007076389156281948, + 0.004949941299855709, + 0.0033920679707080126, + 0.00883176177740097, + 0.0053322832100093365, + 0.004849704913794994, + 0.00754927285015583, + 0.0030389162711799145, + 0.0064377207309007645, + 0.0056842658668756485, + 0.0026745956856757402, + 0.0063744354993104935, + 0.0032561274711042643, + 0.0030433693900704384, + 0.0023167221806943417, + 0.0020790458656847477, + 0.0027376036159694195, + 0.017476920038461685, + 0.0029700868763029575, + 0.03156503662467003, + 0.004910982679575682, + 0.003434313228353858, + 0.002293397206813097, + 0.0038659111596643925, + 0.006686532404273748, + 0.005835718009620905, + 0.0032329887617379427, + 0.004613177850842476, + 0.0025219714734703302, + 0.0026231359224766493, + 0.003603146178647876, + 0.0025700319092720747, + 0.0057648117654025555, + 0.007626062724739313, + 0.00517255300655961, + 0.0038613490760326385, + 0.003787220921367407, + 0.004011225886642933, + 0.004630721174180508, + 0.002899761078879237, + 0.005247590132057667, + 0.039711493998765945, + 0.016421495005488396, + 0.07799766212701797, + 0.018932849168777466, + 0.012194487266242504, + 0.022172290831804276, + 0.021764032542705536, + 0.04157409816980362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.02979704923927784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066960170865059 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066960170865059, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066960170865059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2601988613605499, + 0.04407036304473877, + 0.04138275608420372, + 0.04492376372218132, + 0.03157159313559532, + 0.06844528019428253, + 0.10649371892213821, + 0.10705319046974182, + 0.17865784466266632, + 0.11720259487628937 + ], + "credit_map": [ + 0.022465668618679047, + 0.037820879369974136, + 0.04258815571665764, + 0.04168560728430748, + 0.01772407442331314, + 0.03465903177857399, + 0.021574081853032112, + 0.041681356728076935, + 0.008614161051809788, + 0.0032384691294282675, + 0.006160857155919075, + 0.004903287161141634, + 0.003807872999459505, + 0.006037796847522259, + 0.005331835243850946, + 0.005976085085421801, + 0.009760250337421894, + 0.004551637452095747, + 0.004144389182329178, + 0.005481455475091934, + 0.002660804195329547, + 0.0029877738561481237, + 0.0033097518607974052, + 0.008486691862344742, + 0.009097044356167316, + 0.0077828760258853436, + 0.004789391532540321, + 0.005412066821008921, + 0.004405315034091473, + 0.0054738870821893215, + 0.002693185582756996, + 0.005269997753202915, + 0.003277173964306712, + 0.002415660535916686, + 0.006585079710930586, + 0.0034261976834386587, + 0.003824270097538829, + 0.004948954097926617, + 0.0028632539324462414, + 0.004231004044413567, + 0.016445400193333626, + 0.005046749487519264, + 0.00601645652204752, + 0.004827129188925028, + 0.020552700385451317, + 0.0049396478570997715, + 0.005117617081850767, + 0.005499587394297123, + 0.013247909024357796, + 0.0045130131766200066, + 0.0258951373398304, + 0.005513802636414766, + 0.0046315984800457954, + 0.0060981702990829945, + 0.01050290185958147, + 0.03609118238091469, + 0.01941126212477684, + 0.013977810740470886, + 0.02077944576740265, + 0.008966079913079739, + 0.007853938266634941, + 0.008993372321128845, + 0.011468001641333103, + 0.015603278763592243, + 0.0357549674808979, + 0.01393892616033554, + 0.023854287341237068, + 0.042066045105457306, + 0.011735312640666962, + 0.014022339135408401, + 0.017506347969174385, + 0.019779620692133904, + 0.019421780481934547, + 0.008080675266683102, + 0.03850659355521202, + 0.009708459489047527, + 0.005609983578324318, + 0.00776717159897089, + 0.010747615247964859, + 0.017360316589474678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011066960170865059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle tracking shot sweeps alongside an origami figure, capturing it as it negotiates a high-speed turn through a metropolis made entirely of paper. The camera dynamically follows the racer, maintaining sharp focus as the background blurs with motion. Folded from a single sheet of vibrant, matte red paper, the figure mimics the aggressive stance of a professional motorcycle racer, its body tilted at an extreme angle, almost parallel to the paper ground. Its form is a masterpiece of intricate folding, with sharp creases defining its helmeted head, angular shoulders, and hunched limbs for maximum aerodynamics. The vehicle is a large, intricately folded wheel made of smooth black paper, its structure suggestive of spokes and a tread pattern formed by clever geometric pleats. The entire scene is rendered in ultra-sharp 4K, revealing the crisp paper texture and subtle fiber grain of the material. As the wheel rounds a corner on the gray paper street, the dynamic movement is so intense that the towering, soft-colored origami skyscrapers in the background transform into vertical ribbons of color, creating a powerful sense of speed and a thrilling, focused atmosphere. Bright, direct lighting illuminates the scene, casting sharp, clean shadows and highlighting every fold and edge, making the surreal yet tangible world pop with crystal clarity.", + "seed": 19902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_11/index_7087_seed_19902_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13169516287744046, + "vlm_logprob": -0.13169516287744046, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.13169516287744046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021394770592451096 + ], + "opposite_logprob_tokens": [ + -4.021394729614258 + ], + "logprob_reward": -0.021394770592451096, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021394770592451096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16901840269565582, + 0.031418848782777786, + 0.034956369549036026, + 0.09700961410999298, + 0.04588674381375313, + 0.24738572537899017, + 0.11704393476247787, + 0.03970479220151901, + 0.17311716079711914, + 0.04445846006274223 + ], + "credit_map": [ + 0.02322019822895527, + 0.02514735795557499, + 0.02097116969525814, + 0.019424917176365852, + 0.020883359014987946, + 0.021056484431028366, + 0.022058136761188507, + 0.016256775707006454, + 0.004770372062921524, + 0.0038745731581002474, + 0.0030469573102891445, + 0.003291425993666053, + 0.00452720420435071, + 0.003691965015605092, + 0.004652649164199829, + 0.0035636997781693935, + 0.004523338750004768, + 0.003799168858677149, + 0.0033079704735428095, + 0.004945139400660992, + 0.00376502750441432, + 0.004124210216104984, + 0.005374650936573744, + 0.005116861313581467, + 0.0048564737662673, + 0.005354596301913261, + 0.004837692715227604, + 0.005764467641711235, + 0.004940079990774393, + 0.03271791711449623, + 0.007292157504707575, + 0.031246226280927658, + 0.0029927208088338375, + 0.004132212605327368, + 0.0027694280724972486, + 0.003625920508056879, + 0.004056508187204599, + 0.01035214401781559, + 0.008216851390898228, + 0.009740956127643585, + 0.03067159652709961, + 0.022738855332136154, + 0.016967549920082092, + 0.02397305704653263, + 0.028038250282406807, + 0.046789225190877914, + 0.044570308178663254, + 0.03363687917590141, + 0.009730174206197262, + 0.00896328967064619, + 0.007225331850349903, + 0.007145348004996777, + 0.010178145952522755, + 0.04325732961297035, + 0.010984756052494049, + 0.019559552893042564, + 0.0042479271069169044, + 0.004588830750435591, + 0.004184073302894831, + 0.003507128218188882, + 0.0055045923218131065, + 0.006406990811228752, + 0.0067642671056091785, + 0.004500982351601124, + 0.025466514751315117, + 0.013376709073781967, + 0.01739896833896637, + 0.017619991675019264, + 0.01729634217917919, + 0.02224038727581501, + 0.0305160004645586, + 0.029202254489064217, + 0.005242452025413513, + 0.004383858293294907, + 0.004769863560795784, + 0.004731997847557068, + 0.006631848867982626, + 0.004531787242740393, + 0.007591992616653442, + 0.006574660073965788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.021394770592451096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190698698163033 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190698698163033, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190698698163033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19590352475643158, + 0.039671141654253006, + 0.08800092339515686, + 0.10949072241783142, + 0.07358365505933762, + 0.14628975093364716, + 0.09856175631284714, + 0.04837258160114288, + 0.1547497808933258, + 0.04537609592080116 + ], + "credit_map": [ + 0.032612163573503494, + 0.02781907096505165, + 0.019750718027353287, + 0.020883630961179733, + 0.021879876032471657, + 0.0286247618496418, + 0.022438278421759605, + 0.0218950342386961, + 0.0061842151917517185, + 0.004150772001594305, + 0.00427566422149539, + 0.004059601575136185, + 0.005346691235899925, + 0.005144291091710329, + 0.00536221731454134, + 0.005147687159478664, + 0.013608160428702831, + 0.009623429737985134, + 0.006451488472521305, + 0.013311989605426788, + 0.0067816562950611115, + 0.014147152192890644, + 0.010403052903711796, + 0.013673996552824974, + 0.006950202863663435, + 0.008087068796157837, + 0.006402433384209871, + 0.00904504768550396, + 0.008612977340817451, + 0.02960710972547531, + 0.010363942012190819, + 0.030421949923038483, + 0.007032334804534912, + 0.007972297258675098, + 0.005422165151685476, + 0.007916098460555077, + 0.00820807833224535, + 0.01269535906612873, + 0.014669608324766159, + 0.009667710401117802, + 0.013049024157226086, + 0.009838271886110306, + 0.007190507836639881, + 0.011526834219694138, + 0.011202354915440083, + 0.03364144265651703, + 0.025302115827798843, + 0.03453919664025307, + 0.007982267998158932, + 0.00712653249502182, + 0.0064490074291825294, + 0.006624406669288874, + 0.007594482507556677, + 0.03394121676683426, + 0.009825273416936398, + 0.019018568098545074, + 0.004739102441817522, + 0.005691492930054665, + 0.004127624444663525, + 0.003947325516492128, + 0.007104692980647087, + 0.00873357430100441, + 0.007873349823057652, + 0.006155423820018768, + 0.021153371781110764, + 0.012018331326544285, + 0.013282103464007378, + 0.016702642664313316, + 0.012850594706833363, + 0.021143926307559013, + 0.0253386702388525, + 0.03226013854146004, + 0.00525600416585803, + 0.00563765037804842, + 0.005066271405667067, + 0.00472762156277895, + 0.007190707139670849, + 0.005525499582290649, + 0.005998455453664064, + 0.0059738908894360065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.018190698698163033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818835735321045 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818835735321045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818835735321045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15157034993171692, + 0.03061344474554062, + 0.1594703644514084, + 0.1015145406126976, + 0.09815184772014618, + 0.11648361384868622, + 0.1164432168006897, + 0.05238516628742218, + 0.11671962589025497, + 0.05664779990911484 + ], + "credit_map": [ + 0.012879200279712677, + 0.017725041136145592, + 0.015936629846692085, + 0.013740481808781624, + 0.015819357708096504, + 0.026520170271396637, + 0.03135045990347862, + 0.017599018290638924, + 0.0036877314560115337, + 0.004104677587747574, + 0.003612060332670808, + 0.0031972231809049845, + 0.0043107387609779835, + 0.0037716750521212816, + 0.0047160061076283455, + 0.0032133315689861774, + 0.013816755264997482, + 0.013909022323787212, + 0.009625723585486412, + 0.01173563301563263, + 0.013521031476557255, + 0.04155763238668442, + 0.017877109348773956, + 0.03742746263742447, + 0.008265354670584202, + 0.01143520139157772, + 0.007096708752214909, + 0.010264651849865913, + 0.01513458788394928, + 0.019076332449913025, + 0.013174205087125301, + 0.017067493870854378, + 0.008056934922933578, + 0.007713709957897663, + 0.005718361586332321, + 0.009805116802453995, + 0.010192531161010265, + 0.018740477040410042, + 0.019676340743899345, + 0.018248379230499268, + 0.007545880042016506, + 0.009309535846114159, + 0.0058900401927530766, + 0.007967312820255756, + 0.00813045259565115, + 0.038282111287117004, + 0.009660443291068077, + 0.029697837308049202, + 0.009023575112223625, + 0.008442354388535023, + 0.006963014602661133, + 0.006547995377331972, + 0.00949089415371418, + 0.044978342950344086, + 0.011937212198972702, + 0.019059833139181137, + 0.0047350372187793255, + 0.006135087925940752, + 0.00414629140868783, + 0.004414803814142942, + 0.007421461865305901, + 0.010419354774057865, + 0.008681446313858032, + 0.006431685760617256, + 0.010725216008722782, + 0.010454234667122364, + 0.012181800790131092, + 0.01044343039393425, + 0.013460558839142323, + 0.016605202108621597, + 0.017906155437231064, + 0.024943027645349503, + 0.005397805478423834, + 0.005074977409094572, + 0.005184303503483534, + 0.005872142035514116, + 0.009374884888529778, + 0.006451919209212065, + 0.010255714878439903, + 0.009036049246788025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.01818835735321045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486538827419281 + ], + "opposite_logprob_tokens": [ + -1.2236539125442505 + ], + "logprob_reward": -0.3486538827419281, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486538827419281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2842651307582855, + 0.05990736186504364, + 0.0869305208325386, + 0.18329709768295288, + 0.13748851418495178, + 0.05718415975570679, + 0.061720505356788635, + 0.03444631025195122, + 0.05870813876390457, + 0.03605235368013382 + ], + "credit_map": [ + 0.03531656041741371, + 0.03324544429779053, + 0.044039398431777954, + 0.021841716021299362, + 0.03738360106945038, + 0.03340652585029602, + 0.04532228782773018, + 0.0337095782160759, + 0.007228217087686062, + 0.008293751627206802, + 0.004212405066937208, + 0.0064959158189594746, + 0.006946893408894539, + 0.009045826271176338, + 0.008634735830128193, + 0.009049610234797001, + 0.009804793633520603, + 0.007041413336992264, + 0.004071092698723078, + 0.012936114333570004, + 0.005469920579344034, + 0.020792579278349876, + 0.00974737573415041, + 0.017067231237888336, + 0.012317397631704807, + 0.011914796195924282, + 0.010966882109642029, + 0.016265034675598145, + 0.012632643803954124, + 0.04985671117901802, + 0.02028188295662403, + 0.04906173422932625, + 0.014754547737538815, + 0.011280415579676628, + 0.008546779863536358, + 0.013750055804848671, + 0.013666149228811264, + 0.024317031726241112, + 0.029125753790140152, + 0.022047771140933037, + 0.008913322351872921, + 0.004415188450366259, + 0.0036422377452254295, + 0.00527146365493536, + 0.006945300847291946, + 0.007203352637588978, + 0.009984240867197514, + 0.010809049010276794, + 0.0042215860448777676, + 0.004401650745421648, + 0.0036844927817583084, + 0.004146554507315159, + 0.003974062390625477, + 0.01913350448012352, + 0.0072308750823140144, + 0.014927779324352741, + 0.003922790288925171, + 0.003302093595266342, + 0.003728275652974844, + 0.002643054584041238, + 0.005445646587759256, + 0.005012436769902706, + 0.006339397747069597, + 0.004052616190165281, + 0.007574614603072405, + 0.006024225149303675, + 0.0061206393875181675, + 0.008833198808133602, + 0.006853052880614996, + 0.005811353679746389, + 0.008573430590331554, + 0.008917626924812794, + 0.0030257017351686954, + 0.0033548520877957344, + 0.003112907288596034, + 0.0036032681819051504, + 0.005677425302565098, + 0.005430375225841999, + 0.005564217455685139, + 0.006283603608608246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.3486538827419281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an origami figure folded from a single sheet of vibrant matte red paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the origami figure mimicking the stance of a professional motorcycle racer with a helmeted head and hunched limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle a large wheel made of smooth black paper with geometric pleats suggesting spokes and tread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a metropolis made entirely of paper with gray streets and towering soft-colored skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dramatic low-angle tracking shot that follows the racer while the background blurs with motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520481050014496 + ], + "opposite_logprob_tokens": [ + -1.502048134803772 + ], + "logprob_reward": -0.2520481050014496, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520481050014496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16120421886444092, + 0.04605640098452568, + 0.045203518122434616, + 0.07290139049291611, + 0.04408080503344536, + 0.20798872411251068, + 0.22252032160758972, + 0.0686977356672287, + 0.0524311289191246, + 0.07891581207513809 + ], + "credit_map": [ + 0.01865016296505928, + 0.01747942715883255, + 0.023986192420125008, + 0.013276268728077412, + 0.02206043154001236, + 0.015951460227370262, + 0.03203892707824707, + 0.01776133105158806, + 0.004404922481626272, + 0.0041291965171694756, + 0.00356748653575778, + 0.005909607745707035, + 0.005816168151795864, + 0.007646794430911541, + 0.007565800100564957, + 0.007016421295702457, + 0.003649976337328553, + 0.0027210249099880457, + 0.0027298058848828077, + 0.00482290331274271, + 0.0037335618399083614, + 0.013039914891123772, + 0.004513878375291824, + 0.009992449544370174, + 0.007415376603603363, + 0.008085198700428009, + 0.0068174018524587154, + 0.008342023938894272, + 0.011130329221487045, + 0.008679484017193317, + 0.01150325033813715, + 0.010928318835794926, + 0.0033076603431254625, + 0.004799596033990383, + 0.0020989319309592247, + 0.003386079566553235, + 0.004312180448323488, + 0.011165609583258629, + 0.00583786191418767, + 0.00917288102209568, + 0.021202348172664642, + 0.013495372608304024, + 0.009958538226783276, + 0.016789628192782402, + 0.01724374294281006, + 0.043543968349695206, + 0.03921137750148773, + 0.04654374718666077, + 0.018394818529486656, + 0.016955388709902763, + 0.012834705412387848, + 0.011890900321304798, + 0.019500205293297768, + 0.08377670496702194, + 0.021577728912234306, + 0.03758985549211502, + 0.007512071635574102, + 0.0071981679648160934, + 0.005617013666778803, + 0.007781308609992266, + 0.0074658580124378204, + 0.01050330139696598, + 0.013721903786063194, + 0.008898104541003704, + 0.007220201659947634, + 0.00435557309538126, + 0.006869658827781677, + 0.005921689327806234, + 0.003568005980923772, + 0.007530082017183304, + 0.00772234657779336, + 0.009243562817573547, + 0.007653506007045507, + 0.008363040164113045, + 0.007929093204438686, + 0.007268993183970451, + 0.010771376080811024, + 0.00902662891894579, + 0.01289792638272047, + 0.015005244873464108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.2520481050014496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_tensor([779, 779])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_tensor([779, 779])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a48ab67cefce444bb58759fc8a94b5c5bbe1ea9c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_tensor([779, 779])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a low angle, the camera remains still, focused on the surface of a thick, viscous, dark brown mud. In the dim, slanting light of dusk, this murky swamp takes on an ominous, oily texture. Bubbles of varying sizes slowly and powerfully rise from the depths of the mud, forming smooth, taut bubbles. Their surfaces tremble subtly from the internal pressure, finally bursting with a silent, tense \"poof,\" releasing turbid gas and tiny splashes of mud. The entire surface of the mud thus appears to be in a constant, irregular state of peristalsis, as if a vast, unknown creature is breathing or stirring heavily beneath. In the background, several dead tree roots covered in dark green moss twist and protrude from the mire, adding to the dead silence. The image is dominated by cold tones, with the light creating fleeting highlights on the slippery mud bubbles, creating an unsettling sense of suspense and a primal, latent atmosphere of danger.", + "seed": 16676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_779_seed_16676_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3408416770398617, + "vlm_logprob": -0.3408416770398617, + "vlm_yes_no_margin": 2.174999976158142, + "vlm_token_credit": -0.3408416770398617, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.154790997505188 + ], + "opposite_logprob_tokens": [ + -2.0297908782958984 + ], + "logprob_reward": -0.154790997505188, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.154790997505188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13937005400657654, + 0.028384383767843246, + 0.034857261925935745, + 0.047608375549316406, + 0.28255191445350647, + 0.09464596211910248, + 0.060580503195524216, + 0.10643564164638519, + 0.0976446345448494, + 0.10792131721973419 + ], + "credit_map": [ + 0.022585973143577576, + 0.025039829313755035, + 0.011545299552381039, + 0.022649703547358513, + 0.009683718904852867, + 0.02433403953909874, + 0.011862930841743946, + 0.0116685526445508, + 0.0033295471221208572, + 0.004443855490535498, + 0.003679779591038823, + 0.0038730765227228403, + 0.002766238758340478, + 0.004583894740790129, + 0.0032959473319351673, + 0.0024120425805449486, + 0.003814839757978916, + 0.00349319726228714, + 0.004601429682224989, + 0.004430183209478855, + 0.0027582908514887094, + 0.004934861324727535, + 0.005912677384912968, + 0.004911783616989851, + 0.006997112650424242, + 0.0044172233901917934, + 0.005807055626064539, + 0.006762554869055748, + 0.004688595421612263, + 0.007086718454957008, + 0.00461586145684123, + 0.007233253214508295, + 0.05629535764455795, + 0.022290077060461044, + 0.03737632930278778, + 0.026049254462122917, + 0.022366337478160858, + 0.07177266478538513, + 0.02198575623333454, + 0.0244161207228899, + 0.017327509820461273, + 0.01416661124676466, + 0.01152532547712326, + 0.008361239917576313, + 0.01220723520964384, + 0.010555027984082699, + 0.008682560175657272, + 0.01182045228779316, + 0.00847171526402235, + 0.007138432003557682, + 0.004559863358736038, + 0.006317718420177698, + 0.010713550262153149, + 0.008298209868371487, + 0.005726728122681379, + 0.009354286827147007, + 0.021079154685139656, + 0.01057122740894556, + 0.010492416098713875, + 0.011455608531832695, + 0.01138296164572239, + 0.02318151853978634, + 0.009976252913475037, + 0.008296499028801918, + 0.0143289715051651, + 0.01414670143276453, + 0.008187217637896538, + 0.014800451695919037, + 0.010075339116156101, + 0.017751526087522507, + 0.008082717657089233, + 0.010271708481013775, + 0.0176008902490139, + 0.010911999270319939, + 0.012648514471948147, + 0.008044148795306683, + 0.018919141963124275, + 0.020761609077453613, + 0.008309976197779179, + 0.010725035332143307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.154790997505188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539023876190186 + ], + "opposite_logprob_tokens": [ + -0.42890235781669617 + ], + "logprob_reward": -1.0539023876190186, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539023876190186, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05032182112336159, + 0.062797412276268, + 0.04828866571187973, + 0.039416417479515076, + 0.07526881247758865, + 0.1144043356180191, + 0.042526017874479294, + 0.07901130616664886, + 0.36959409713745117, + 0.1183711364865303 + ], + "credit_map": [ + 0.011843766085803509, + 0.006005030125379562, + 0.004775012843310833, + 0.008101806044578552, + 0.0030373085755854845, + 0.009110127575695515, + 0.0032841074280440807, + 0.004164663143455982, + 0.01201268844306469, + 0.005913181230425835, + 0.006541365757584572, + 0.007586719002574682, + 0.00457358593121171, + 0.012395678088068962, + 0.007933385670185089, + 0.005840811878442764, + 0.005343626253306866, + 0.0057681831531226635, + 0.0069017778150737286, + 0.00750675518065691, + 0.0042439657263457775, + 0.005308924242854118, + 0.006719626951962709, + 0.00649580592289567, + 0.004058169201016426, + 0.004357985686510801, + 0.0053575350902974606, + 0.007414496038109064, + 0.0029111988842487335, + 0.004765601828694344, + 0.005118356551975012, + 0.005433073733001947, + 0.011605544947087765, + 0.004669892601668835, + 0.013524278067052364, + 0.0075215245597064495, + 0.008605772629380226, + 0.017330685630440712, + 0.004004712216556072, + 0.008006397634744644, + 0.015973135828971863, + 0.012560628354549408, + 0.012767382897436619, + 0.011492057703435421, + 0.019007105380296707, + 0.01726287603378296, + 0.009986121207475662, + 0.015355034731328487, + 0.006060325540602207, + 0.004804844968020916, + 0.0053857360035181046, + 0.005649050232023001, + 0.004883374087512493, + 0.005418068263679743, + 0.0038688774220645428, + 0.006455743685364723, + 0.020961657166481018, + 0.005788646172732115, + 0.005382302217185497, + 0.005759562365710735, + 0.005483794957399368, + 0.022826533764600754, + 0.006831227336078882, + 0.005977578926831484, + 0.05389592424035072, + 0.041546471416950226, + 0.02851935289800167, + 0.05067729949951172, + 0.03199942409992218, + 0.07639378309249878, + 0.03449811413884163, + 0.05206369236111641, + 0.02524987794458866, + 0.011876340955495834, + 0.008297552354633808, + 0.0073257735930383205, + 0.022998766973614693, + 0.027594100683927536, + 0.007616368588060141, + 0.007412348873913288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -1.0539023876190186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650331757962704 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.008650331757962704, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650331757962704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11536344885826111, + 0.06516920775175095, + 0.06984272599220276, + 0.025215445086359978, + 0.04510151594877243, + 0.13380548357963562, + 0.05174598470330238, + 0.05841968208551407, + 0.1858297437429428, + 0.24950675666332245 + ], + "credit_map": [ + 0.02810225635766983, + 0.009006187319755554, + 0.009785732254385948, + 0.027805635705590248, + 0.008164852857589722, + 0.01653343439102173, + 0.009123224765062332, + 0.006842123344540596, + 0.010504718869924545, + 0.006960672792047262, + 0.0074662817642092705, + 0.007632013875991106, + 0.005988392513245344, + 0.014710020273923874, + 0.006113353651016951, + 0.005793747957795858, + 0.008254824206233025, + 0.0075539471581578255, + 0.009674370288848877, + 0.012715838849544525, + 0.006088345777243376, + 0.009989539161324501, + 0.007205633912235498, + 0.008360220119357109, + 0.0035482081584632397, + 0.002313886536285281, + 0.003230435773730278, + 0.004885237198323011, + 0.0021656136959791183, + 0.003459382336586714, + 0.003089764853939414, + 0.002522916067391634, + 0.010713723488152027, + 0.005118379835039377, + 0.003972045611590147, + 0.005124709568917751, + 0.003398233326151967, + 0.010254938155412674, + 0.0032516266219317913, + 0.003267859108746052, + 0.019950902089476585, + 0.015655068680644035, + 0.01344611868262291, + 0.013125774450600147, + 0.025834448635578156, + 0.015586997382342815, + 0.009157862514257431, + 0.021048307418823242, + 0.0133975213393569, + 0.004123563878238201, + 0.0037340063136070967, + 0.003938069101423025, + 0.006080847233533859, + 0.01369002927094698, + 0.0028953661676496267, + 0.0038865834940224886, + 0.007469628006219864, + 0.008920956403017044, + 0.007375550456345081, + 0.007197422441095114, + 0.006538201589137316, + 0.008000396192073822, + 0.007490918505936861, + 0.005426611285656691, + 0.021614357829093933, + 0.023337719962000847, + 0.016772115603089333, + 0.025890951976180077, + 0.016998527571558952, + 0.03234303370118141, + 0.01929561235010624, + 0.029577430337667465, + 0.04252327233552933, + 0.023213038221001625, + 0.03287472948431969, + 0.021332452073693275, + 0.036410339176654816, + 0.049818940460681915, + 0.01842801459133625, + 0.024905985221266747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.008650331757962704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538296170532703 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538296170532703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538296170532703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13721400499343872, + 0.09046556800603867, + 0.053111232817173004, + 0.0466424822807312, + 0.2800656259059906, + 0.057922713458538055, + 0.03245498612523079, + 0.05752869322896004, + 0.18650442361831665, + 0.05809022858738899 + ], + "credit_map": [ + 0.03439773619174957, + 0.011603692546486855, + 0.01240503042936325, + 0.029747921973466873, + 0.010961147956550121, + 0.02102597989141941, + 0.008278261870145798, + 0.008794247172772884, + 0.009916291572153568, + 0.010639469139277935, + 0.009531567804515362, + 0.010223086923360825, + 0.012557758949697018, + 0.01607949659228325, + 0.014323791489005089, + 0.007194100879132748, + 0.00646754028275609, + 0.004627191927284002, + 0.006793814245611429, + 0.00633857911452651, + 0.00430953549221158, + 0.011411583051085472, + 0.007908081635832787, + 0.0052549089305102825, + 0.006588772404938936, + 0.006030723452568054, + 0.004765844903886318, + 0.005918995011597872, + 0.0055490280501544476, + 0.007896922528743744, + 0.004051451105624437, + 0.005840746685862541, + 0.05110230669379234, + 0.012317607179284096, + 0.06412450969219208, + 0.014481477439403534, + 0.02417953871190548, + 0.08916551619768143, + 0.010323218069970608, + 0.014371464028954506, + 0.009616999886929989, + 0.006689570844173431, + 0.0062353117391467094, + 0.005462036933749914, + 0.010354316793382168, + 0.007620032411068678, + 0.004563660826534033, + 0.0073807816952466965, + 0.00685231015086174, + 0.00404178025200963, + 0.0024746109265834093, + 0.0033691697753965855, + 0.0029890413861721754, + 0.00686471676453948, + 0.0023825792595744133, + 0.00348077854141593, + 0.011834646575152874, + 0.005708345212042332, + 0.005099669564515352, + 0.005204349756240845, + 0.006460839416831732, + 0.012601143680512905, + 0.005503347609192133, + 0.005116351414471865, + 0.020972073078155518, + 0.028326289728283882, + 0.01770734041929245, + 0.029693732038140297, + 0.021370509639382362, + 0.02400965616106987, + 0.01902690902352333, + 0.02539791725575924, + 0.011583848856389523, + 0.005319440271705389, + 0.007452795282006264, + 0.005660738330334425, + 0.005317134782671928, + 0.012294581159949303, + 0.004417524207383394, + 0.006044165231287479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.012538296170532703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47432637214660645 + ], + "opposite_logprob_tokens": [ + -0.9743263721466064 + ], + "logprob_reward": -0.47432637214660645, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47432637214660645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2799973487854004, + 0.11066266149282455, + 0.05015186592936516, + 0.03335268050432205, + 0.06615982204675674, + 0.1260187029838562, + 0.05211303383111954, + 0.05860535055398941, + 0.1319621205329895, + 0.09097641706466675 + ], + "credit_map": [ + 0.07677516341209412, + 0.023560984060168266, + 0.02332908846437931, + 0.06910782307386398, + 0.017300007864832878, + 0.0333700031042099, + 0.021182922646403313, + 0.01537135150283575, + 0.013877232559025288, + 0.02166452631354332, + 0.010290602222084999, + 0.012485071085393429, + 0.011081630364060402, + 0.016584159806370735, + 0.014399650506675243, + 0.010279793292284012, + 0.006024185102432966, + 0.005016318988054991, + 0.005793800111860037, + 0.008247051388025284, + 0.005218397825956345, + 0.0063869706355035305, + 0.007576623000204563, + 0.005888518877327442, + 0.005346124526113272, + 0.0027618976309895515, + 0.0044248150661587715, + 0.00306005054153502, + 0.003977582324296236, + 0.007184478919953108, + 0.002868965268135071, + 0.003728764597326517, + 0.009876614436507225, + 0.006233776453882456, + 0.014777063392102718, + 0.00527170579880476, + 0.009892593137919903, + 0.008188372477889061, + 0.0035832722205668688, + 0.008336428552865982, + 0.018691062927246094, + 0.018107468262314796, + 0.01378070842474699, + 0.010665527544915676, + 0.02474209852516651, + 0.014152543619275093, + 0.01131890807300806, + 0.014560393057763577, + 0.009414095431566238, + 0.005119714420288801, + 0.005746927112340927, + 0.004869580268859863, + 0.005805802531540394, + 0.00796020869165659, + 0.0040087103843688965, + 0.009187993593513966, + 0.018466182053089142, + 0.003175581805408001, + 0.0033154271077364683, + 0.0037603743840008974, + 0.004050418734550476, + 0.019885925576090813, + 0.0026820257771760225, + 0.003269414883106947, + 0.01554910745471716, + 0.02374306507408619, + 0.012859568931162357, + 0.02266867831349373, + 0.013938763178884983, + 0.015517893247306347, + 0.011663817800581455, + 0.016021214425563812, + 0.014189299196004868, + 0.009115047752857208, + 0.008749125525355339, + 0.02014869824051857, + 0.011057361029088497, + 0.011363265104591846, + 0.007805454544723034, + 0.008548173122107983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.47432637214660645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera remains still, focused on the surface of a thick, viscous, dark brown mud. In the dim, slanting light of dusk, this murky swamp takes on an ominous, oily texture. Bubbles of varying sizes slowly and powerfully rise from the depths of the mud, forming smooth, taut bubbles. Their surfaces tremble subtly from the internal pressure, finally bursting with a silent, tense \"poof,\" releasing turbid gas and tiny splashes of mud. The entire surface of the mud thus appears to be in a constant, irregular state of peristalsis, as if a vast, unknown creature is breathing or stirring heavily beneath. In the background, several dead tree roots covered in dark green moss twist and protrude from the mire, adding to the dead silence. The image is dominated by cold tones, with the light creating fleeting highlights on the slippery mud bubbles, creating an unsettling sense of suspense and a primal, latent atmosphere of danger.", + "seed": 16677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_779_seed_16677_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1759633630514145, + "vlm_logprob": -0.1759633630514145, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.1759633630514145, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023965438827872276 + ], + "opposite_logprob_tokens": [ + -4.273965358734131 + ], + "logprob_reward": -0.023965438827872276, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.023965438827872276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13044261932373047, + 0.061849310994148254, + 0.07563406229019165, + 0.10384528338909149, + 0.17906740307807922, + 0.031396009027957916, + 0.07084152847528458, + 0.06981417536735535, + 0.11412873864173889, + 0.16298086941242218 + ], + "credit_map": [ + 0.011713597923517227, + 0.019984301179647446, + 0.01663276180624962, + 0.01663285493850708, + 0.013877361081540585, + 0.01362068671733141, + 0.017359204590320587, + 0.020621847361326218, + 0.0063521419651806355, + 0.01003413088619709, + 0.007012429181486368, + 0.00924866832792759, + 0.005961303599178791, + 0.00833617802709341, + 0.004706881940364838, + 0.010197577998042107, + 0.0075585562735795975, + 0.01034503523260355, + 0.008431089110672474, + 0.00910935364663601, + 0.010666696354746819, + 0.008116135373711586, + 0.01061105728149414, + 0.010796141810715199, + 0.008038945496082306, + 0.008554683066904545, + 0.01313218753784895, + 0.01700683683156967, + 0.014306674711406231, + 0.012994756922125816, + 0.012129370123147964, + 0.017681829631328583, + 0.024213770404458046, + 0.017832046374678612, + 0.024000246077775955, + 0.020095309242606163, + 0.014782696962356567, + 0.02001873403787613, + 0.020476803183555603, + 0.03764781355857849, + 0.0024097191635519266, + 0.004011605866253376, + 0.002830838318914175, + 0.003935299813747406, + 0.0050241840071976185, + 0.0036615165881812572, + 0.004484476521611214, + 0.005038370378315449, + 0.005031159613281488, + 0.0055469125509262085, + 0.009411558508872986, + 0.011801597662270069, + 0.009000599384307861, + 0.008063677698373795, + 0.009925861842930317, + 0.012060163542628288, + 0.0075645786710083485, + 0.008687064982950687, + 0.005879458039999008, + 0.008719167672097683, + 0.011763479560613632, + 0.01031124871224165, + 0.005545251537114382, + 0.011343921534717083, + 0.010542759671807289, + 0.009253698401153088, + 0.010288028046488762, + 0.016321973875164986, + 0.011633340269327164, + 0.018625354394316673, + 0.013431062921881676, + 0.02403252199292183, + 0.01635718159377575, + 0.01421167142689228, + 0.017594562843441963, + 0.03244069963693619, + 0.02609436772763729, + 0.01885899342596531, + 0.02217259630560875, + 0.015250791795551777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.023965438827872276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761830806732178 + ], + "opposite_logprob_tokens": [ + -0.8261830806732178 + ], + "logprob_reward": -0.5761830806732178, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761830806732178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1118038147687912, + 0.07217800617218018, + 0.1236715093255043, + 0.23417378962039948, + 0.07768528908491135, + 0.07798118889331818, + 0.06337682157754898, + 0.06081056967377663, + 0.038491155952215195, + 0.1398278772830963 + ], + "credit_map": [ + 0.006941125262528658, + 0.01118488423526287, + 0.014582999050617218, + 0.01704169064760208, + 0.019485022872686386, + 0.01245200727134943, + 0.014590495266020298, + 0.015525591559708118, + 0.007366649340838194, + 0.011449056677520275, + 0.006008717231452465, + 0.007911233231425285, + 0.008233454078435898, + 0.009514115750789642, + 0.008457188494503498, + 0.013237596489489079, + 0.0064368476159870625, + 0.005739334039390087, + 0.01064365729689598, + 0.010668790899217129, + 0.006245394703000784, + 0.02982024848461151, + 0.011347596533596516, + 0.04276963695883751, + 0.024770352989435196, + 0.02515251748263836, + 0.022637566551566124, + 0.04581933468580246, + 0.030617324635386467, + 0.026501165702939034, + 0.03431380167603493, + 0.024361729621887207, + 0.007996220141649246, + 0.006390747614204884, + 0.010000835172832012, + 0.009200495667755604, + 0.006078124977648258, + 0.008881372399628162, + 0.012975923717021942, + 0.01616157405078411, + 0.005835841875523329, + 0.007651990745216608, + 0.010940920561552048, + 0.00886240229010582, + 0.014267844147980213, + 0.012047753669321537, + 0.009319443255662918, + 0.009054991416633129, + 0.00769417779520154, + 0.006250171456485987, + 0.008633365854620934, + 0.00976929347962141, + 0.009266403503715992, + 0.007856390438973904, + 0.005717897322028875, + 0.008189120329916477, + 0.006138959899544716, + 0.006878682412207127, + 0.006245202384889126, + 0.009262757375836372, + 0.008623228408396244, + 0.007406106684356928, + 0.007272099144756794, + 0.008983537554740906, + 0.004533691331744194, + 0.005230196751654148, + 0.005103825591504574, + 0.004039133433252573, + 0.00460463622584939, + 0.0055674114264547825, + 0.004986620042473078, + 0.004425642546266317, + 0.01410039234906435, + 0.0243578739464283, + 0.013380316086113453, + 0.02144145406782627, + 0.018082493916153908, + 0.018046259880065918, + 0.01718714088201523, + 0.013231941498816013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.5761830806732178 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071418300271034 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071418300271034, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071418300271034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13870128989219666, + 0.05468998849391937, + 0.24156898260116577, + 0.23317840695381165, + 0.05722406506538391, + 0.05959280580282211, + 0.09178649634122849, + 0.04533359035849571, + 0.040252991020679474, + 0.037671394646167755 + ], + "credit_map": [ + 0.014348439872264862, + 0.01667362079024315, + 0.017942793667316437, + 0.02162924036383629, + 0.022036848589777946, + 0.014029244892299175, + 0.015238973312079906, + 0.016802126541733742, + 0.004413004964590073, + 0.005381720140576363, + 0.008987528271973133, + 0.011049489490687847, + 0.005791792180389166, + 0.005661363713443279, + 0.006648023147135973, + 0.006757070310413837, + 0.015786083415150642, + 0.027378816157579422, + 0.030085554346442223, + 0.041308991611003876, + 0.033573996275663376, + 0.027925429865717888, + 0.03386877104640007, + 0.03164133056998253, + 0.020766733214259148, + 0.019583310931921005, + 0.03009386733174324, + 0.029526308178901672, + 0.04003872722387314, + 0.028277592733502388, + 0.027410779148340225, + 0.03748110309243202, + 0.006057922262698412, + 0.005737341474741697, + 0.0065050008706748486, + 0.007805297616869211, + 0.007426989730447531, + 0.006284803152084351, + 0.007560849189758301, + 0.009845864959061146, + 0.005541788414120674, + 0.008329655043780804, + 0.0062897903844714165, + 0.007381632924079895, + 0.01002091821283102, + 0.007949519902467728, + 0.007359167095273733, + 0.00672033429145813, + 0.008523104712367058, + 0.01009932067245245, + 0.00944611057639122, + 0.012011872604489326, + 0.014043244533240795, + 0.012175781652331352, + 0.011880194768309593, + 0.013606872409582138, + 0.004409571643918753, + 0.00475451210513711, + 0.005514062941074371, + 0.0060021099634468555, + 0.007297747302800417, + 0.005479085724800825, + 0.005212405230849981, + 0.006664098706096411, + 0.005288505926728249, + 0.0036370158195495605, + 0.00404027383774519, + 0.005286103580147028, + 0.00442166393622756, + 0.0060957265086472034, + 0.004063056781888008, + 0.007420643698424101, + 0.003362429328262806, + 0.0035759767051786184, + 0.003342381678521633, + 0.005253826268017292, + 0.006797676905989647, + 0.0055702608078718185, + 0.003875188762322068, + 0.0058936551213264465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.016071418300271034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042236328125 + ], + "opposite_logprob_tokens": [ + -2.812042236328125 + ], + "logprob_reward": -0.062042236328125, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042236328125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17537975311279297, + 0.078499436378479, + 0.0801498219370842, + 0.17549222707748413, + 0.1263038069009781, + 0.06138811260461807, + 0.11522377282381058, + 0.07387038320302963, + 0.05435571447014809, + 0.05933701619505882 + ], + "credit_map": [ + 0.017407940700650215, + 0.0221870094537735, + 0.021071720868349075, + 0.02759724110364914, + 0.024822711944580078, + 0.018570849671959877, + 0.02374402992427349, + 0.019978225231170654, + 0.008525645360350609, + 0.010466846637427807, + 0.013376213610172272, + 0.011332781985402107, + 0.00926265399903059, + 0.010077152401208878, + 0.008083864115178585, + 0.007374268490821123, + 0.006499274633824825, + 0.010711709037423134, + 0.007813812233507633, + 0.01165437139570713, + 0.013230392709374428, + 0.010202604345977306, + 0.009489796124398708, + 0.010547852143645287, + 0.014892768114805222, + 0.01526719145476818, + 0.018704349175095558, + 0.03487226739525795, + 0.025276023894548416, + 0.022005708888173103, + 0.026668217033147812, + 0.017805667594075203, + 0.014865224249660969, + 0.012690902687609196, + 0.016144907101988792, + 0.015174767933785915, + 0.014038275927305222, + 0.01543018501251936, + 0.015336967073380947, + 0.022622572258114815, + 0.0058482009917497635, + 0.006848567631095648, + 0.007223314605653286, + 0.007930616848170757, + 0.010187218897044659, + 0.008712160401046276, + 0.008049670606851578, + 0.006588355638086796, + 0.017339326441287994, + 0.013450785540044308, + 0.01246084924787283, + 0.012469053268432617, + 0.015627538785338402, + 0.014859136193990707, + 0.011007756926119328, + 0.01800931803882122, + 0.010691070929169655, + 0.01263067964464426, + 0.006564108654856682, + 0.007628949359059334, + 0.01326274499297142, + 0.008851001970469952, + 0.005928391125053167, + 0.008313427679240704, + 0.007681502029299736, + 0.007707857061177492, + 0.006147235631942749, + 0.007684675976634026, + 0.0062994277104735374, + 0.0071694026701152325, + 0.004456479102373123, + 0.00720912916585803, + 0.00627681240439415, + 0.005906271282583475, + 0.006794350687414408, + 0.009180494584143162, + 0.011126705445349216, + 0.007474509533494711, + 0.0053985691629350185, + 0.007179296109825373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.062042236328125 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155464112758636 + ], + "opposite_logprob_tokens": [ + -1.7015546560287476 + ], + "logprob_reward": -0.20155464112758636, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155464112758636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10753439366817474, + 0.17640668153762817, + 0.12137512862682343, + 0.12364611029624939, + 0.07781223952770233, + 0.061380572617053986, + 0.05664317309856415, + 0.05727004259824753, + 0.0682210847735405, + 0.14971058070659637 + ], + "credit_map": [ + 0.010593765415251255, + 0.013298995792865753, + 0.013810213655233383, + 0.016775483265519142, + 0.015020429156720638, + 0.012681787833571434, + 0.013503135181963444, + 0.011850571259856224, + 0.01208520494401455, + 0.0224104356020689, + 0.02030160091817379, + 0.03783594071865082, + 0.01727212592959404, + 0.014875048771500587, + 0.027972407639026642, + 0.0236539077013731, + 0.008369863964617252, + 0.010948512703180313, + 0.011670779436826706, + 0.013718736357986927, + 0.01020418293774128, + 0.022477714344859123, + 0.010600228793919086, + 0.03338509425520897, + 0.01265290193259716, + 0.011729069985449314, + 0.012677918188273907, + 0.020405659452080727, + 0.02264363132417202, + 0.016246221959590912, + 0.014244971796870232, + 0.013045716099441051, + 0.006853101775050163, + 0.007358573842793703, + 0.010076379403471947, + 0.012262384407222271, + 0.010915156453847885, + 0.009585107676684856, + 0.009527471847832203, + 0.011234053410589695, + 0.0071495999582111835, + 0.006907433271408081, + 0.00787642877548933, + 0.0048968056216835976, + 0.008862738497555256, + 0.00952063500881195, + 0.00557425431907177, + 0.010592671111226082, + 0.004848779644817114, + 0.005508167203515768, + 0.005487557966262102, + 0.009157565422356129, + 0.010134682059288025, + 0.006291404832154512, + 0.005635119508951902, + 0.00957989040762186, + 0.004526734352111816, + 0.005050982814282179, + 0.004927473608404398, + 0.005613342858850956, + 0.011859020218253136, + 0.008838187903165817, + 0.0047659194096922874, + 0.011688372120261192, + 0.0062553188763558865, + 0.0050154197961091995, + 0.0054916576482355595, + 0.008349821902811527, + 0.006157059222459793, + 0.012326042167842388, + 0.007633068133145571, + 0.016992688179016113, + 0.010971707291901112, + 0.013748141005635262, + 0.01410601194947958, + 0.02903578244149685, + 0.021073220297694206, + 0.023140940815210342, + 0.011044754646718502, + 0.02659001015126705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.20155464112758636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera remains still, focused on the surface of a thick, viscous, dark brown mud. In the dim, slanting light of dusk, this murky swamp takes on an ominous, oily texture. Bubbles of varying sizes slowly and powerfully rise from the depths of the mud, forming smooth, taut bubbles. Their surfaces tremble subtly from the internal pressure, finally bursting with a silent, tense \"poof,\" releasing turbid gas and tiny splashes of mud. The entire surface of the mud thus appears to be in a constant, irregular state of peristalsis, as if a vast, unknown creature is breathing or stirring heavily beneath. In the background, several dead tree roots covered in dark green moss twist and protrude from the mire, adding to the dead silence. The image is dominated by cold tones, with the light creating fleeting highlights on the slippery mud bubbles, creating an unsettling sense of suspense and a primal, latent atmosphere of danger.", + "seed": 16678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_779_seed_16678_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16524986494332553, + "vlm_logprob": -0.16524986494332553, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.16524986494332553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05061575770378113 + ], + "opposite_logprob_tokens": [ + -3.1756157875061035 + ], + "logprob_reward": -0.05061575770378113, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05061575770378113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3993472456932068, + 0.18121285736560822, + 0.07768568396568298, + 0.03558076173067093, + 0.06043308600783348, + 0.08797598630189896, + 0.05152576044201851, + 0.023929767310619354, + 0.02301584556698799, + 0.05929301679134369 + ], + "credit_map": [ + 0.05107782036066055, + 0.0432465635240078, + 0.06354013830423355, + 0.04204768314957619, + 0.037303581833839417, + 0.04657738283276558, + 0.06812120229005814, + 0.04743286594748497, + 0.01993429660797119, + 0.019249549135565758, + 0.02060004323720932, + 0.029898617416620255, + 0.021487602964043617, + 0.02216695435345173, + 0.024398840963840485, + 0.02347695454955101, + 0.012834841385483742, + 0.00940645206719637, + 0.009005685336887836, + 0.010922788642346859, + 0.009697233326733112, + 0.008545300923287868, + 0.008895945735275745, + 0.008377439342439175, + 0.005595796275883913, + 0.003859691321849823, + 0.0038677023258060217, + 0.0036796825006604195, + 0.00434880843386054, + 0.004140103701502085, + 0.005390924867242575, + 0.004698055796325207, + 0.007946640253067017, + 0.006226836703717709, + 0.00892279390245676, + 0.007729987148195505, + 0.00709916464984417, + 0.0065258704125881195, + 0.007943297736346722, + 0.008038494735956192, + 0.012275942601263523, + 0.015026677399873734, + 0.00895319040864706, + 0.007419578731060028, + 0.010785463266074657, + 0.010341024026274681, + 0.00975325983017683, + 0.013420847244560719, + 0.00858885608613491, + 0.0062169418670237064, + 0.0070810867473483086, + 0.0047093541361391544, + 0.005192614160478115, + 0.005380658898502588, + 0.006264805793762207, + 0.008091440424323082, + 0.004045163281261921, + 0.0027092676609754562, + 0.003464084817096591, + 0.0018301352392882109, + 0.002527860691770911, + 0.0026791987475007772, + 0.0032691906671971083, + 0.003404865274205804, + 0.0026765193324536085, + 0.003262328216806054, + 0.0028980316128581762, + 0.002328620059415698, + 0.0031065684743225574, + 0.00237034703604877, + 0.004173025023192167, + 0.002200403483584523, + 0.00774080865085125, + 0.007396215107291937, + 0.006341647356748581, + 0.007083694916218519, + 0.01092994213104248, + 0.005810529924929142, + 0.008790083229541779, + 0.005200093146413565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.05061575770378113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233660936355591 + ], + "opposite_logprob_tokens": [ + -0.8983660936355591 + ], + "logprob_reward": -0.5233660936355591, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233660936355591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08495866507291794, + 0.2798835039138794, + 0.045354753732681274, + 0.04906434193253517, + 0.09975555539131165, + 0.04090890660881996, + 0.04448246210813522, + 0.016631215810775757, + 0.06063118949532509, + 0.27832943201065063 + ], + "credit_map": [ + 0.010264058597385883, + 0.008152189664542675, + 0.012252170592546463, + 0.009983286261558533, + 0.00916249118745327, + 0.009835653938353062, + 0.013633961789309978, + 0.011674847453832626, + 0.021603506058454514, + 0.024098847061395645, + 0.027475079521536827, + 0.07046902924776077, + 0.0207205880433321, + 0.03455573692917824, + 0.0436411127448082, + 0.03731958568096161, + 0.004900659900158644, + 0.00433695875108242, + 0.0056339772418141365, + 0.00932666938751936, + 0.004471474792808294, + 0.005329628009349108, + 0.005152524448931217, + 0.006202857941389084, + 0.006042348686605692, + 0.004869659896939993, + 0.005132319871336222, + 0.008094672113656998, + 0.00599235575646162, + 0.005468579474836588, + 0.006983146071434021, + 0.00648125633597374, + 0.012847832404077053, + 0.011488051153719425, + 0.011457106098532677, + 0.013573609292507172, + 0.009278289042413235, + 0.011009970679879189, + 0.013826778158545494, + 0.016273917630314827, + 0.005223672837018967, + 0.005913906265050173, + 0.004125220701098442, + 0.004653646610677242, + 0.004717396106570959, + 0.004440814256668091, + 0.006942640524357557, + 0.004891606047749519, + 0.006376639474183321, + 0.006568576209247112, + 0.005071880295872688, + 0.0036697029136121273, + 0.004317319951951504, + 0.005343826487660408, + 0.005453205201774836, + 0.007681309711188078, + 0.0020907241851091385, + 0.0025051601696759462, + 0.0019949811976403, + 0.0015453064115718007, + 0.0020631321240216494, + 0.0020731117110699415, + 0.001821424113586545, + 0.0025373748503625393, + 0.007185902912169695, + 0.009352383203804493, + 0.003978596068918705, + 0.006346249021589756, + 0.013359861448407173, + 0.006635553203523159, + 0.00895755086094141, + 0.004815084859728813, + 0.031416360288858414, + 0.03288818150758743, + 0.029807107523083687, + 0.037172831594944, + 0.05091957002878189, + 0.027994774281978607, + 0.04232838749885559, + 0.025802195072174072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.5233660936355591 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632498629391193 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632498629391193, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632498629391193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18844729661941528, + 0.20454056560993195, + 0.07029875367879868, + 0.10574111342430115, + 0.08776294440031052, + 0.03918432071805, + 0.04370852932333946, + 0.017353195697069168, + 0.15837039053440094, + 0.08459291607141495 + ], + "credit_map": [ + 0.02826784923672676, + 0.019797982648015022, + 0.02680388279259205, + 0.024350235238671303, + 0.016297221183776855, + 0.0215822272002697, + 0.02080274373292923, + 0.030545148998498917, + 0.020303817465901375, + 0.01860821805894375, + 0.018556972965598106, + 0.038511719554662704, + 0.02201949432492256, + 0.020381182432174683, + 0.029514769092202187, + 0.03664438799023628, + 0.010399763472378254, + 0.005718397907912731, + 0.007662042509764433, + 0.01126672700047493, + 0.0067754462361335754, + 0.007963410578668118, + 0.007393872365355492, + 0.013119086623191833, + 0.011470519937574863, + 0.007958807051181793, + 0.01154132280498743, + 0.016751371324062347, + 0.01027512364089489, + 0.011378886178135872, + 0.01652509719133377, + 0.01983998715877533, + 0.010282744653522968, + 0.009202579036355019, + 0.009424206800758839, + 0.01197564322501421, + 0.012405161745846272, + 0.010586493648588657, + 0.011808616109192371, + 0.012077500112354755, + 0.005220755934715271, + 0.005286857485771179, + 0.004530167207121849, + 0.0035726677160710096, + 0.00474221957847476, + 0.004123648628592491, + 0.0044031403958797455, + 0.0073048630729317665, + 0.005717921536415815, + 0.005102133844047785, + 0.00579554820433259, + 0.0032831053249537945, + 0.004866304341703653, + 0.004803972318768501, + 0.008723869919776917, + 0.00541567150503397, + 0.002046296140179038, + 0.0017040047096088529, + 0.002657291479408741, + 0.0022403241600841284, + 0.0020576005335897207, + 0.0019055759767070413, + 0.0027822654228657484, + 0.0019598372746258974, + 0.016865288838744164, + 0.016719616949558258, + 0.01042471919208765, + 0.016105007380247116, + 0.026682451367378235, + 0.01868426986038685, + 0.036958422511816025, + 0.015930620953440666, + 0.009329534135758877, + 0.009893840178847313, + 0.01023381669074297, + 0.013131989166140556, + 0.014035506173968315, + 0.008526683785021305, + 0.011628526262938976, + 0.007813015952706337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.007632498629391193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306123033165932 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306123033165932, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306123033165932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19612768292427063, + 0.31293803453445435, + 0.10612194240093231, + 0.08814134448766708, + 0.09425397962331772, + 0.030894139781594276, + 0.04783667251467705, + 0.02780446782708168, + 0.03402424976229668, + 0.06185755506157875 + ], + "credit_map": [ + 0.027891889214515686, + 0.030525319278240204, + 0.020062001422047615, + 0.02775304764509201, + 0.019984588027000427, + 0.021115433424711227, + 0.020945534110069275, + 0.027849851176142693, + 0.02618112787604332, + 0.026539400219917297, + 0.032291315495967865, + 0.07270228862762451, + 0.02607002854347229, + 0.03843550384044647, + 0.053090620785951614, + 0.037627749145030975, + 0.014239220879971981, + 0.01129642128944397, + 0.013023052364587784, + 0.017488040030002594, + 0.011205628514289856, + 0.014027189463376999, + 0.010905243456363678, + 0.013937139883637428, + 0.010410844348371029, + 0.007297472096979618, + 0.009247485548257828, + 0.012302142567932606, + 0.007686777971684933, + 0.009997675195336342, + 0.01306651346385479, + 0.018132422119379044, + 0.011597773991525173, + 0.009213126264512539, + 0.011225624941289425, + 0.012481140904128551, + 0.011477119289338589, + 0.011225375346839428, + 0.011088849976658821, + 0.015944968909025192, + 0.005076889414340258, + 0.0035588934551924467, + 0.0034089682158082724, + 0.0037324281875044107, + 0.0037965301889926195, + 0.0031653232872486115, + 0.004111292771995068, + 0.004043810069561005, + 0.007282159756869078, + 0.005534410942345858, + 0.005672438070178032, + 0.0051228818483650684, + 0.006991097237914801, + 0.00613380316644907, + 0.006709856912493706, + 0.004390018060803413, + 0.003901383373886347, + 0.003982330206781626, + 0.0038098765071481466, + 0.0026356929447501898, + 0.002374535659328103, + 0.0030706345569342375, + 0.004153281915932894, + 0.003876729402691126, + 0.004248775541782379, + 0.0036310802679508924, + 0.0030704718083143234, + 0.0036344260443001986, + 0.006144476588815451, + 0.003856504801660776, + 0.006154654547572136, + 0.003283857833594084, + 0.007593944203108549, + 0.006394760217517614, + 0.00665787048637867, + 0.007297619245946407, + 0.01074258703738451, + 0.006387254223227501, + 0.009512713178992271, + 0.007270800415426493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04306123033165932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157374441623688 + ], + "opposite_logprob_tokens": [ + -1.7015737295150757 + ], + "logprob_reward": -0.20157374441623688, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157374441623688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10556752979755402, + 0.11905623227357864, + 0.0586259588599205, + 0.08275821805000305, + 0.2954016625881195, + 0.073427215218544, + 0.08788607269525528, + 0.08589838445186615, + 0.06449810415506363, + 0.026880614459514618 + ], + "credit_map": [ + 0.013521777465939522, + 0.01183015014976263, + 0.01704902946949005, + 0.010948240756988525, + 0.009898305870592594, + 0.012749554589390755, + 0.01210942491889, + 0.01746104285120964, + 0.010821089148521423, + 0.009313254617154598, + 0.013803176581859589, + 0.02537130005657673, + 0.008881769143044949, + 0.014690625481307507, + 0.016785115003585815, + 0.01938990317285061, + 0.0063962857238948345, + 0.0059504881501197815, + 0.00578127009794116, + 0.012157531455159187, + 0.00791129283607006, + 0.006580376531928778, + 0.006486913654953241, + 0.00736180180683732, + 0.010745582170784473, + 0.006697512231767178, + 0.011344694532454014, + 0.013483843766152859, + 0.007848522625863552, + 0.008444330655038357, + 0.010946926660835743, + 0.013246813789010048, + 0.04368888586759567, + 0.035345885902643204, + 0.030008409172296524, + 0.035140473395586014, + 0.04096057638525963, + 0.03486085310578346, + 0.032579366117715836, + 0.04281722754240036, + 0.010642197914421558, + 0.007136365864425898, + 0.00923237856477499, + 0.007428234908729792, + 0.009931288659572601, + 0.00840845424681902, + 0.009315697476267815, + 0.011332601308822632, + 0.010290933772921562, + 0.009903455153107643, + 0.011527168564498425, + 0.008711769245564938, + 0.009087725542485714, + 0.009676759131252766, + 0.016394468024373055, + 0.012293784879148006, + 0.010599709115922451, + 0.011515060439705849, + 0.01274588517844677, + 0.009264068678021431, + 0.007937728427350521, + 0.009148241952061653, + 0.015128116123378277, + 0.009559571743011475, + 0.006106757093220949, + 0.007388245780020952, + 0.0055213067680597305, + 0.007173451129347086, + 0.009408054873347282, + 0.007425103802233934, + 0.014490553177893162, + 0.006984624546021223, + 0.003672810038551688, + 0.0024255395401269197, + 0.0028851677197963, + 0.003544280305504799, + 0.00328321591950953, + 0.0030954135581851006, + 0.00499582989141345, + 0.0029783567879348993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.20157374441623688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera remains still, focused on the surface of a thick, viscous, dark brown mud. In the dim, slanting light of dusk, this murky swamp takes on an ominous, oily texture. Bubbles of varying sizes slowly and powerfully rise from the depths of the mud, forming smooth, taut bubbles. Their surfaces tremble subtly from the internal pressure, finally bursting with a silent, tense \"poof,\" releasing turbid gas and tiny splashes of mud. The entire surface of the mud thus appears to be in a constant, irregular state of peristalsis, as if a vast, unknown creature is breathing or stirring heavily beneath. In the background, several dead tree roots covered in dark green moss twist and protrude from the mire, adding to the dead silence. The image is dominated by cold tones, with the light creating fleeting highlights on the slippery mud bubbles, creating an unsettling sense of suspense and a primal, latent atmosphere of danger.", + "seed": 16679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_15/index_779_seed_16679_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14852189607918262, + "vlm_logprob": -0.14852189607918262, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.14852189607918262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08092805743217468 + ], + "opposite_logprob_tokens": [ + -2.705928087234497 + ], + "logprob_reward": -0.08092805743217468, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08092805743217468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21947410702705383, + 0.06476911902427673, + 0.024971144273877144, + 0.03773513808846474, + 0.055213235318660736, + 0.04583030566573143, + 0.35217195749282837, + 0.11458196491003036, + 0.06294509768486023, + 0.022307919338345528 + ], + "credit_map": [ + 0.030071843415498734, + 0.02397722564637661, + 0.01815319061279297, + 0.026692846789956093, + 0.026766128838062286, + 0.029858486726880074, + 0.03336355462670326, + 0.030590815469622612, + 0.00917394645512104, + 0.0064386408776044846, + 0.005521310959011316, + 0.007846859283745289, + 0.006355742458254099, + 0.010788388550281525, + 0.008008630014955997, + 0.010635602287948132, + 0.0028259861283004284, + 0.0020693906117230654, + 0.0019064090447500348, + 0.003072444349527359, + 0.0025137655902653933, + 0.005028586834669113, + 0.003897264366969466, + 0.0036572981625795364, + 0.006853777915239334, + 0.004161984194070101, + 0.0037202786188572645, + 0.005689868237823248, + 0.003087878692895174, + 0.005171331577003002, + 0.0048504965379834175, + 0.004199523478746414, + 0.007954045198857784, + 0.0036983333993703127, + 0.003329409519210458, + 0.0037329092156141996, + 0.00548009667545557, + 0.01381534431129694, + 0.005320369731634855, + 0.011882725171744823, + 0.008319370448589325, + 0.004460856784135103, + 0.00374995032325387, + 0.0024769320152699947, + 0.005736758001148701, + 0.00832378026098013, + 0.003496803343296051, + 0.009265856817364693, + 0.020427165552973747, + 0.01550808921456337, + 0.0087443171069026, + 0.01081534568220377, + 0.019482485949993134, + 0.12779736518859863, + 0.014746345579624176, + 0.13465085625648499, + 0.008837574161589146, + 0.005003102123737335, + 0.0036783521063625813, + 0.004744782578200102, + 0.005771150812506676, + 0.039442028850317, + 0.005106764379888773, + 0.04199821501970291, + 0.003766448935493827, + 0.0016480698250234127, + 0.0016853435663506389, + 0.0017263784538954496, + 0.002201676834374666, + 0.023530038073658943, + 0.0031351130455732346, + 0.025252029299736023, + 0.0026580558624118567, + 0.0015705750556662679, + 0.0014030839083716273, + 0.0020586696919053793, + 0.0013788103824481368, + 0.00593571225181222, + 0.002472300548106432, + 0.004830712452530861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.08092805743217468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347811222076416 + ], + "opposite_logprob_tokens": [ + -1.3134781122207642 + ], + "logprob_reward": -0.31347811222076416, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31347811222076416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04454602301120758, + 0.041582223027944565, + 0.09453721344470978, + 0.03189590945839882, + 0.12435873597860336, + 0.19351260364055634, + 0.3693643808364868, + 0.02601901814341545, + 0.033200956881046295, + 0.040982939302921295 + ], + "credit_map": [ + 0.004245120566338301, + 0.005384079180657864, + 0.0029653655365109444, + 0.005004166159778833, + 0.007146595977246761, + 0.005019664764404297, + 0.006806461606174707, + 0.007974566891789436, + 0.007124548312276602, + 0.0034529846161603928, + 0.0034538554027676582, + 0.004151421133428812, + 0.005584249272942543, + 0.005435138009488583, + 0.007059028837829828, + 0.0053210011683404446, + 0.013554089702665806, + 0.007646332029253244, + 0.010789230465888977, + 0.01109086163341999, + 0.009415417909622192, + 0.010469154454767704, + 0.019276704639196396, + 0.01229542214423418, + 0.005704555194824934, + 0.00322909839451313, + 0.002624043496325612, + 0.004157403018325567, + 0.0037485980428755283, + 0.004357855301350355, + 0.0032516783103346825, + 0.0048226783983409405, + 0.013805189169943333, + 0.008406953886151314, + 0.008338323794305325, + 0.011629688553512096, + 0.01149031799286604, + 0.028603194281458855, + 0.015281875617802143, + 0.026803195476531982, + 0.011872912757098675, + 0.010175596922636032, + 0.006269761826843023, + 0.00621476536616683, + 0.013730351813137531, + 0.06841077655553818, + 0.008816633373498917, + 0.06802179664373398, + 0.03374519571661949, + 0.029743952676653862, + 0.017488112673163414, + 0.01580938696861267, + 0.03624451905488968, + 0.10269828885793686, + 0.01917395554482937, + 0.11446095257997513, + 0.0018204995431005955, + 0.001300950301811099, + 0.0013777880230918527, + 0.0019971714355051517, + 0.0017219658475369215, + 0.00781636219471693, + 0.002181737218052149, + 0.0078025455586612225, + 0.0021685862448066473, + 0.0012698948848992586, + 0.0012705570552498102, + 0.0009918275754898787, + 0.0019743740558624268, + 0.011645661666989326, + 0.00183842983096838, + 0.012041624635457993, + 0.007061045616865158, + 0.002918707439675927, + 0.003921865951269865, + 0.00466197868809104, + 0.0029210455249994993, + 0.0068506659008562565, + 0.003596374299377203, + 0.009051257744431496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.31347811222076416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201764017343521 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201764017343521, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201764017343521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15853995084762573, + 0.03420793265104294, + 0.08159390836954117, + 0.02002687193453312, + 0.09680777043104172, + 0.14925765991210938, + 0.21442042291164398, + 0.15543393790721893, + 0.04394039884209633, + 0.045771125704050064 + ], + "credit_map": [ + 0.01808488555252552, + 0.019744353368878365, + 0.01921544037759304, + 0.02397596463561058, + 0.017403479665517807, + 0.018066076561808586, + 0.025295360013842583, + 0.0167543925344944, + 0.004777179565280676, + 0.0035842389333993196, + 0.0031990492716431618, + 0.004109441302716732, + 0.0044177439995110035, + 0.004398788325488567, + 0.005031703505665064, + 0.004689788445830345, + 0.010849234648048878, + 0.0072559453547000885, + 0.008948768489062786, + 0.009815354831516743, + 0.007544481195509434, + 0.008657936938107014, + 0.018032317981123924, + 0.010489868931472301, + 0.0033145342022180557, + 0.0017470496241003275, + 0.0015956319402903318, + 0.00299138855189085, + 0.002205345081165433, + 0.0028829302173107862, + 0.0019950265996158123, + 0.0032949631568044424, + 0.013654942624270916, + 0.008295612409710884, + 0.011532106436789036, + 0.01453336887061596, + 0.009643551893532276, + 0.011738097295165062, + 0.015082540921866894, + 0.012327547185122967, + 0.01592407375574112, + 0.00947996973991394, + 0.007827069610357285, + 0.00527834240347147, + 0.012055907398462296, + 0.04191311448812485, + 0.009058719500899315, + 0.04772046580910683, + 0.01420212909579277, + 0.015746956691145897, + 0.006944471038877964, + 0.0077336374670267105, + 0.018924647942185402, + 0.0686875656247139, + 0.009281350299715996, + 0.07289966195821762, + 0.012130656279623508, + 0.008975641801953316, + 0.005696834530681372, + 0.005273227114230394, + 0.007837917655706406, + 0.0506829135119915, + 0.00756401102989912, + 0.057272739708423615, + 0.0019764434546232224, + 0.0011207975912839174, + 0.0010519024217501283, + 0.0011459836969152093, + 0.0016747048357501626, + 0.01679813489317894, + 0.002325209090486169, + 0.017847223207354546, + 0.00959806703031063, + 0.0038744318298995495, + 0.0054145753383636475, + 0.006359088700264692, + 0.0033650712575763464, + 0.005610150285065174, + 0.004671808332204819, + 0.006877933628857136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.014201764017343521 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578688010573387 + ], + "opposite_logprob_tokens": [ + -3.895578622817993 + ], + "logprob_reward": -0.020578688010573387, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578688010573387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07244209200143814, + 0.05723877623677254, + 0.050480201840400696, + 0.05016438663005829, + 0.40362656116485596, + 0.04674872010946274, + 0.20894847810268402, + 0.04084287956357002, + 0.04426465928554535, + 0.02524317242205143 + ], + "credit_map": [ + 0.008714110590517521, + 0.007844914682209492, + 0.005669957026839256, + 0.007740371394902468, + 0.011145763099193573, + 0.008613284677267075, + 0.010358482599258423, + 0.012355209328234196, + 0.010131016373634338, + 0.006708945147693157, + 0.004484161734580994, + 0.004925013054162264, + 0.008318890817463398, + 0.007228608708828688, + 0.0069420491345226765, + 0.008500095456838608, + 0.008429711684584618, + 0.00403059646487236, + 0.005363265983760357, + 0.006454111076891422, + 0.005529956892132759, + 0.005702569615095854, + 0.008433558978140354, + 0.006536436267197132, + 0.010145355947315693, + 0.00516850408166647, + 0.0035084846895188093, + 0.00604967912659049, + 0.007744248956441879, + 0.006420023273676634, + 0.0048611098900437355, + 0.0062669855542480946, + 0.029218779876828194, + 0.024490615352988243, + 0.014942260459065437, + 0.014919348992407322, + 0.04498985409736633, + 0.12459775060415268, + 0.026022084057331085, + 0.12444590777158737, + 0.007995344698429108, + 0.0038397363387048244, + 0.003560842014849186, + 0.0025758109986782074, + 0.004702945705503225, + 0.010462590493261814, + 0.004052295349538326, + 0.009559160098433495, + 0.009833911433815956, + 0.004734487272799015, + 0.002678951481357217, + 0.006105764303356409, + 0.006832964718341827, + 0.08469711989164352, + 0.007440023124217987, + 0.08662527799606323, + 0.004206196404993534, + 0.0034435552079230547, + 0.0032248797360807657, + 0.0020227485802024603, + 0.0075479065999388695, + 0.007828214205801487, + 0.003931405954062939, + 0.008637976832687855, + 0.0027862798888236284, + 0.0011645050253719091, + 0.0013652831548824906, + 0.0016187438741326332, + 0.0016137544298544526, + 0.0158267579972744, + 0.0022518234327435493, + 0.01763751171529293, + 0.0038392245769500732, + 0.001988442149013281, + 0.0025119963102042675, + 0.0030952426604926586, + 0.001994840335100889, + 0.005488410592079163, + 0.0025269396137446165, + 0.003798078279942274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.020578688010573387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are bubbles of varying sizes rising from the dark brown mud surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dead tree roots covered in dark green moss visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold tones with slanting dusk light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134228587150574 + ], + "opposite_logprob_tokens": [ + -1.3134229183197021 + ], + "logprob_reward": -0.3134228587150574, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134228587150574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12422388792037964, + 0.037282392382621765, + 0.02921556681394577, + 0.05707412213087082, + 0.03683534264564514, + 0.1396147757768631, + 0.43387940526008606, + 0.07158496230840683, + 0.03652474656701088, + 0.033764831721782684 + ], + "credit_map": [ + 0.015054658055305481, + 0.009159396402537823, + 0.014484452083706856, + 0.01963411271572113, + 0.008982268162071705, + 0.01812007650732994, + 0.024514004588127136, + 0.014274923130869865, + 0.006734875496476889, + 0.005246582441031933, + 0.002066447166725993, + 0.00395301403477788, + 0.0030347539577633142, + 0.0068069323897361755, + 0.004653695039451122, + 0.004786092322319746, + 0.004789181984961033, + 0.002573417965322733, + 0.002305231522768736, + 0.002993369707837701, + 0.0022669844329357147, + 0.005240476690232754, + 0.005131315439939499, + 0.003915589768439531, + 0.009075173176825047, + 0.0052595725283026695, + 0.00446704588830471, + 0.007941300980746746, + 0.006283960770815611, + 0.006146595813333988, + 0.00943132396787405, + 0.008469153195619583, + 0.006121648475527763, + 0.00358340865932405, + 0.0029132310301065445, + 0.0025518652983009815, + 0.0051468717865645885, + 0.005858028773218393, + 0.004648394882678986, + 0.0060118939727544785, + 0.008168811909854412, + 0.005789079703390598, + 0.003641376970335841, + 0.004185374826192856, + 0.007099652662873268, + 0.05234183743596077, + 0.005549502559006214, + 0.052839141339063644, + 0.03976400941610336, + 0.03866429626941681, + 0.020202601328492165, + 0.016588591039180756, + 0.04956149682402611, + 0.11566770076751709, + 0.025609806180000305, + 0.12782087922096252, + 0.010833155363798141, + 0.0067333076149225235, + 0.00627012737095356, + 0.008163430728018284, + 0.005000620614737272, + 0.015810122713446617, + 0.00758442934602499, + 0.011189768090844154, + 0.0034073484130203724, + 0.0014225638005882502, + 0.001876345369964838, + 0.001504944870248437, + 0.0014969953335821629, + 0.011382224038243294, + 0.0024156684521585703, + 0.013018655590713024, + 0.003671130631119013, + 0.0019388622604310513, + 0.001988843083381653, + 0.0015150747494772077, + 0.001941671478562057, + 0.008885737508535385, + 0.002280597807839513, + 0.0115429125726223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.3134228587150574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_tensor([8701, 8701])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_tensor([8701, 8701])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8875954c3a266ccc0e840338a4d9ca27347ccae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_tensor([8701, 8701])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot captures an endless stream of glossy photographic prints as they traverse the frame vertically, tumbling past the camera in a continuous, fluid motion. Each print, with a high-gloss finish and gently curled edges, twirls and flutters, revealing fleeting, out-of-focus glimpses of its contents—vibrant, saturated landscapes, intimate black-and-white portraits, and candid moments of joy. As this waterfall of memories tumbles downward, a powerful off-camera light source strikes the lens, causing a bright, film-like anamorphic lens flare to bloom and dominate the image. This flare appears as a warm, golden-white core with long, horizontal streaks of electric blue light that stretch from edge to edge, shimmering dynamically across the smooth surface of the photograph. The interplay between the chaotic dance of the falling prints and the stylized, dreamlike light creates a deeply nostalgic and visually stunning atmosphere, evoking potent memories against a dark, blurred backdrop.", + "seed": 16829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_8701_seed_16829_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19618657771497966, + "vlm_logprob": -0.19618657771497966, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.19618657771497966, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.903260350227356 + ], + "opposite_logprob_tokens": [ + -0.528260350227356 + ], + "logprob_reward": -0.903260350227356, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.903260350227356, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1221376359462738, + 0.19625821709632874, + 0.05683939903974533, + 0.2661072313785553, + 0.058663539588451385, + 0.07561405748128891, + 0.09070803225040436, + 0.05111357197165489, + 0.02405346930027008, + 0.05850491672754288 + ], + "credit_map": [ + 0.008939807303249836, + 0.037269480526447296, + 0.01090378500521183, + 0.014621556736528873, + 0.011217484250664711, + 0.008930479176342487, + 0.012405430898070335, + 0.01784961298108101, + 0.008249050006270409, + 0.01556954812258482, + 0.014131233096122742, + 0.015220426954329014, + 0.056436073035001755, + 0.021151306107640266, + 0.012937570922076702, + 0.052563004195690155, + 0.004927904345095158, + 0.007293053902685642, + 0.0056232973001897335, + 0.007005596999078989, + 0.009966042824089527, + 0.0049284109845757484, + 0.009770170785486698, + 0.007324919570237398, + 0.028065530583262444, + 0.0430566631257534, + 0.014877486042678356, + 0.052425310015678406, + 0.049942631274461746, + 0.03068593703210354, + 0.03018469363451004, + 0.01686897873878479, + 0.00455825449898839, + 0.00764194130897522, + 0.006852425634860992, + 0.007342050317674875, + 0.008228994905948639, + 0.00843721441924572, + 0.009425911121070385, + 0.006176747381687164, + 0.0043723550625145435, + 0.008976138196885586, + 0.007974972017109394, + 0.00912390649318695, + 0.012421241030097008, + 0.01096766721457243, + 0.011025817133486271, + 0.010751958936452866, + 0.00760293286293745, + 0.011991266161203384, + 0.015134346671402454, + 0.012751370668411255, + 0.011794324032962322, + 0.012922445312142372, + 0.00943464133888483, + 0.009076699614524841, + 0.009048611856997013, + 0.010693601332604885, + 0.007702638395130634, + 0.005689800716936588, + 0.004550389014184475, + 0.00476499367505312, + 0.0032774880528450012, + 0.005386047996580601, + 0.004643369000405073, + 0.003925468772649765, + 0.0026572225615382195, + 0.001956423744559288, + 0.003093782579526305, + 0.0030528486240655184, + 0.0016977600753307343, + 0.0030265948735177517, + 0.005807696375995874, + 0.011049973778426647, + 0.00645084073767066, + 0.0072125960141420364, + 0.008320847526192665, + 0.007047573104500771, + 0.004625597968697548, + 0.007989793084561825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.903260350227356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638358250260353 + ], + "opposite_logprob_tokens": [ + -3.651383638381958 + ], + "logprob_reward": -0.02638358250260353, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638358250260353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13523298501968384, + 0.10422199219465256, + 0.05969554930925369, + 0.2877223789691925, + 0.0833754613995552, + 0.10492614656686783, + 0.09585435688495636, + 0.03091946244239807, + 0.05534063279628754, + 0.04271097481250763 + ], + "credit_map": [ + 0.02332664653658867, + 0.012267244048416615, + 0.0169580839574337, + 0.011380194686353207, + 0.011667081154882908, + 0.021454181522130966, + 0.01364787295460701, + 0.024531660601496696, + 0.0059324768371880054, + 0.011765893548727036, + 0.011187237687408924, + 0.011377090588212013, + 0.02190772071480751, + 0.010847369208931923, + 0.010108096525073051, + 0.021096112206578255, + 0.0075597274117171764, + 0.008943972177803516, + 0.005259021185338497, + 0.008496708236634731, + 0.0092821279540658, + 0.006805213633924723, + 0.0067901816219091415, + 0.006558598019182682, + 0.027942677959799767, + 0.04645282030105591, + 0.017547689378261566, + 0.059277985244989395, + 0.05574175715446472, + 0.02935172989964485, + 0.03205035999417305, + 0.019357381388545036, + 0.006555333733558655, + 0.007158797699958086, + 0.0076803783886134624, + 0.01961977407336235, + 0.005676943343132734, + 0.01168187614530325, + 0.018385007977485657, + 0.006617351435124874, + 0.008923493325710297, + 0.013116471469402313, + 0.012167368084192276, + 0.012147484347224236, + 0.018638256937265396, + 0.01592918112874031, + 0.01209237053990364, + 0.011911524459719658, + 0.008396160788834095, + 0.010385139845311642, + 0.01563822478055954, + 0.015065017156302929, + 0.013181143440306187, + 0.012486723251640797, + 0.01160783227533102, + 0.0090941172093153, + 0.0042407456785440445, + 0.004661956802010536, + 0.0038682345766574144, + 0.004395258147269487, + 0.0034306233283132315, + 0.0034647351130843163, + 0.0030952950473874807, + 0.0037626142147928476, + 0.01121170911937952, + 0.007720721885561943, + 0.005177265498787165, + 0.0037992580328136683, + 0.005932557862251997, + 0.007426365278661251, + 0.00412233080714941, + 0.0099504254758358, + 0.0049765040166676044, + 0.009256177581846714, + 0.004140607081353664, + 0.005265820771455765, + 0.0055271899327635765, + 0.005244802683591843, + 0.0037627327255904675, + 0.004537138622254133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.02638358250260353 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797108359634876 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.009797108359634876, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797108359634876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1871621459722519, + 0.06591443717479706, + 0.04314383864402771, + 0.26708197593688965, + 0.0848807767033577, + 0.09501568973064423, + 0.06671486049890518, + 0.04317602142691612, + 0.037804048508405685, + 0.10910618305206299 + ], + "credit_map": [ + 0.017705203965306282, + 0.045523639768362045, + 0.015602839179337025, + 0.01808314025402069, + 0.01564297080039978, + 0.01732797548174858, + 0.02012598142027855, + 0.03715040534734726, + 0.004109750501811504, + 0.00641763536259532, + 0.007187946233898401, + 0.007998711429536343, + 0.010438058525323868, + 0.01006259024143219, + 0.0062937685288488865, + 0.013405981473624706, + 0.0029003953095525503, + 0.005148947238922119, + 0.004711374640464783, + 0.00661048898473382, + 0.006941620260477066, + 0.0038899697829037905, + 0.008566663600504398, + 0.00437437929213047, + 0.026960372924804688, + 0.03743163123726845, + 0.01744087040424347, + 0.06290571391582489, + 0.047957733273506165, + 0.03184626251459122, + 0.026532234624028206, + 0.016007183119654655, + 0.00732691865414381, + 0.006021028850227594, + 0.008677024394273758, + 0.021183263510465622, + 0.006290436256676912, + 0.011974603869020939, + 0.014799118041992188, + 0.008608399890363216, + 0.006707482505589724, + 0.010176043957471848, + 0.013391140848398209, + 0.010665260255336761, + 0.016917821019887924, + 0.01262167189270258, + 0.013427086174488068, + 0.01110918540507555, + 0.005506615620106459, + 0.009335009381175041, + 0.014522097073495388, + 0.009798172861337662, + 0.006745151709765196, + 0.006037828978151083, + 0.006932650692760944, + 0.007837341167032719, + 0.00653294799849391, + 0.009217873215675354, + 0.005883058998733759, + 0.005722232162952423, + 0.004117335192859173, + 0.004201361909508705, + 0.003084162250161171, + 0.004417048767209053, + 0.006618273910135031, + 0.005205496679991484, + 0.00532505614683032, + 0.002704156329855323, + 0.0047714184038341045, + 0.005149729084223509, + 0.002572783036157489, + 0.005457138177007437, + 0.010180657729506493, + 0.01905064284801483, + 0.013246465474367142, + 0.013738034293055534, + 0.0153729859739542, + 0.012316801585257053, + 0.009234040975570679, + 0.015966562554240227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.009797108359634876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196552991867065 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196552991867065, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196552991867065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2508988082408905, + 0.06924773007631302, + 0.05937650799751282, + 0.11848904192447662, + 0.04505448788404465, + 0.09360934048891068, + 0.09279979020357132, + 0.10594561696052551, + 0.07153052091598511, + 0.09304822981357574 + ], + "credit_map": [ + 0.03286295756697655, + 0.06838487833738327, + 0.021069612354040146, + 0.025055192410945892, + 0.02613985911011696, + 0.02874397300183773, + 0.0271526500582695, + 0.021489640697836876, + 0.0035609917249530554, + 0.004752765875309706, + 0.007262664381414652, + 0.00853621494024992, + 0.01406113151460886, + 0.008980280719697475, + 0.004463235381990671, + 0.017630431801080704, + 0.007369159255176783, + 0.00549674779176712, + 0.008055138401687145, + 0.01045287773013115, + 0.004984910134226084, + 0.004405319225043058, + 0.012024495750665665, + 0.006587852258235216, + 0.010593189857900143, + 0.016744976863265038, + 0.006298589985817671, + 0.028666434809565544, + 0.020825952291488647, + 0.014036094769835472, + 0.013425198383629322, + 0.007898584008216858, + 0.003470629919320345, + 0.00506111653521657, + 0.005107079166918993, + 0.006752047222107649, + 0.005701178684830666, + 0.006533963605761528, + 0.006789195351302624, + 0.0056392718106508255, + 0.005467729177325964, + 0.008500974625349045, + 0.010158184915781021, + 0.009233296848833561, + 0.01870003528892994, + 0.015903141349554062, + 0.012120927684009075, + 0.013525037094950676, + 0.009144742041826248, + 0.011649048887193203, + 0.013261837884783745, + 0.014220459386706352, + 0.012894676998257637, + 0.011264191009104252, + 0.00914717186242342, + 0.011217648163437843, + 0.013225707225501537, + 0.019535519182682037, + 0.012109555304050446, + 0.012219212017953396, + 0.012157412245869637, + 0.011662938632071018, + 0.00936629343777895, + 0.015668967738747597, + 0.009963716380298138, + 0.007496012840420008, + 0.011171232908964157, + 0.006022625602781773, + 0.007473752833902836, + 0.010111276060342789, + 0.006536873988807201, + 0.012755023315548897, + 0.008927407674491405, + 0.013449531979858875, + 0.012355044484138489, + 0.010522929020226002, + 0.012969870120286942, + 0.010377643629908562, + 0.00867857038974762, + 0.015767211094498634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.018196552991867065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023295294493436813 + ], + "opposite_logprob_tokens": [ + -3.7732954025268555 + ], + "logprob_reward": -0.023295294493436813, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295294493436813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16935791075229645, + 0.23436905443668365, + 0.044348228722810745, + 0.19644343852996826, + 0.05161182954907417, + 0.08967700600624084, + 0.07251457124948502, + 0.04749755188822746, + 0.02520490251481533, + 0.06897547841072083 + ], + "credit_map": [ + 0.023771779611706734, + 0.033209800720214844, + 0.018339022994041443, + 0.0165716540068388, + 0.01592390239238739, + 0.0153910918161273, + 0.02237599715590477, + 0.023774679750204086, + 0.013813582248985767, + 0.016419246792793274, + 0.011753850616514683, + 0.016614655032753944, + 0.05331762135028839, + 0.03638606518507004, + 0.017516234889626503, + 0.068547822535038, + 0.0052401283755898476, + 0.005871303379535675, + 0.005235876888036728, + 0.0061669680289924145, + 0.005005642306059599, + 0.004793517291545868, + 0.007288692519068718, + 0.004746103659272194, + 0.021600520238280296, + 0.02835051529109478, + 0.012332825921475887, + 0.04467310756444931, + 0.033488769084215164, + 0.02254592254757881, + 0.020582368597388268, + 0.012869437225162983, + 0.0037292768247425556, + 0.004989203996956348, + 0.005173644050955772, + 0.007884969934821129, + 0.006659252103418112, + 0.007744515314698219, + 0.008448239415884018, + 0.006982734892517328, + 0.00624507712200284, + 0.01006392017006874, + 0.007965871132910252, + 0.009142829105257988, + 0.021138859912753105, + 0.013718913309276104, + 0.00953611359000206, + 0.011865430511534214, + 0.005260730627924204, + 0.008336416445672512, + 0.012529593892395496, + 0.011358516290783882, + 0.008673661388456821, + 0.008013911545276642, + 0.008734707720577717, + 0.009607037529349327, + 0.00698052765801549, + 0.010263264179229736, + 0.006438072770833969, + 0.005309967789798975, + 0.004848825745284557, + 0.005699848290532827, + 0.0029262450989335775, + 0.005030806642025709, + 0.002607365371659398, + 0.0031361368019133806, + 0.0039033873472362757, + 0.002406190149486065, + 0.002759761642664671, + 0.003524064552038908, + 0.003233096795156598, + 0.0036349042784422636, + 0.006859068293124437, + 0.012881495989859104, + 0.008811328560113907, + 0.008851282298564911, + 0.010620107874274254, + 0.00793219543993473, + 0.005674806889146566, + 0.007345204707235098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.023295294493436813 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures an endless stream of glossy photographic prints as they traverse the frame vertically, tumbling past the camera in a continuous, fluid motion. Each print, with a high-gloss finish and gently curled edges, twirls and flutters, revealing fleeting, out-of-focus glimpses of its contents—vibrant, saturated landscapes, intimate black-and-white portraits, and candid moments of joy. As this waterfall of memories tumbles downward, a powerful off-camera light source strikes the lens, causing a bright, film-like anamorphic lens flare to bloom and dominate the image. This flare appears as a warm, golden-white core with long, horizontal streaks of electric blue light that stretch from edge to edge, shimmering dynamically across the smooth surface of the photograph. The interplay between the chaotic dance of the falling prints and the stylized, dreamlike light creates a deeply nostalgic and visually stunning atmosphere, evoking potent memories against a dark, blurred backdrop.", + "seed": 16830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_8701_seed_16830_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15664433157071472, + "vlm_logprob": -0.15664433157071472, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.15664433157071472, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6999443769454956 + ], + "opposite_logprob_tokens": [ + -0.6999443769454956 + ], + "logprob_reward": -0.6999443769454956, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6999443769454956, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16328488290309906, + 0.10093828290700912, + 0.06833714246749878, + 0.12752656638622284, + 0.05099848657846451, + 0.08706533163785934, + 0.10945591330528259, + 0.15545159578323364, + 0.06511765718460083, + 0.07182417064905167 + ], + "credit_map": [ + 0.020150113850831985, + 0.020919345319271088, + 0.015343118458986282, + 0.021913614124059677, + 0.016909636557102203, + 0.01827647164463997, + 0.02588932029902935, + 0.023883266374468803, + 0.012153217568993568, + 0.014275752939283848, + 0.016165772452950478, + 0.013366896659135818, + 0.009121255949139595, + 0.012241605669260025, + 0.01197813544422388, + 0.011635645292699337, + 0.008143041282892227, + 0.005815517157316208, + 0.01303973887115717, + 0.011837727390229702, + 0.00786650087684393, + 0.007476080674678087, + 0.007203286979347467, + 0.006955246906727552, + 0.009640338830649853, + 0.024601684883236885, + 0.009567836299538612, + 0.015437005087733269, + 0.015037336386740208, + 0.02229677513241768, + 0.009330054745078087, + 0.021615540608763695, + 0.006632263772189617, + 0.004119782242923975, + 0.005587135907262564, + 0.004475988447666168, + 0.007159715052694082, + 0.008057861588895321, + 0.00798665452748537, + 0.006979086436331272, + 0.008745111525058746, + 0.011227203533053398, + 0.010535812005400658, + 0.011393530294299126, + 0.008425187319517136, + 0.010902793146669865, + 0.012846921570599079, + 0.012988769449293613, + 0.010191625915467739, + 0.012616753578186035, + 0.00891223456710577, + 0.014911306090652943, + 0.015724921599030495, + 0.014618808403611183, + 0.01853475160896778, + 0.013945515267550945, + 0.00718813668936491, + 0.014573166146874428, + 0.00530881155282259, + 0.006620524451136589, + 0.04990076273679733, + 0.0422692745923996, + 0.015142177231609821, + 0.014448733069002628, + 0.010341278277337551, + 0.006544274743646383, + 0.010141616687178612, + 0.009683213196694851, + 0.00728588504716754, + 0.006835682317614555, + 0.007710185367614031, + 0.006575517822057009, + 0.01527674775570631, + 0.0083279674872756, + 0.006505593657493591, + 0.007526441477239132, + 0.00829903595149517, + 0.007868641056120396, + 0.008989713154733181, + 0.00903003104031086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.6999443769454956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814033046364784 + ], + "opposite_logprob_tokens": [ + -3.288140296936035 + ], + "logprob_reward": -0.03814033046364784, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814033046364784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10487290471792221, + 0.05315428972244263, + 0.08562043309211731, + 0.1152535006403923, + 0.18290968239307404, + 0.06025324761867523, + 0.07389345020055771, + 0.1679709106683731, + 0.09240737557411194, + 0.06366419792175293 + ], + "credit_map": [ + 0.010574117302894592, + 0.012900102883577347, + 0.007799206301569939, + 0.01239797379821539, + 0.01684235967695713, + 0.011577095836400986, + 0.017873788252472878, + 0.014908261597156525, + 0.005784396547824144, + 0.008903490379452705, + 0.006799199618399143, + 0.004938928410410881, + 0.00621159840375185, + 0.006611704360693693, + 0.0059872702695429325, + 0.007917700335383415, + 0.012227692641317844, + 0.006491069681942463, + 0.012172523885965347, + 0.018152398988604546, + 0.008940710686147213, + 0.00914828572422266, + 0.008950642310082912, + 0.009537111036479473, + 0.009635180234909058, + 0.016419656574726105, + 0.008596545085310936, + 0.013713833875954151, + 0.009207493625581264, + 0.023135708644986153, + 0.01248839870095253, + 0.022056682035326958, + 0.025862645357847214, + 0.021069686859846115, + 0.015404951758682728, + 0.02103148214519024, + 0.022260015830397606, + 0.027004845440387726, + 0.028021398931741714, + 0.022254662588238716, + 0.007270829286426306, + 0.007632138207554817, + 0.0060621402226388454, + 0.006115972064435482, + 0.008123737759888172, + 0.007571517489850521, + 0.008067109622061253, + 0.00940980575978756, + 0.009463025256991386, + 0.011672730557620525, + 0.007573002949357033, + 0.009108462370932102, + 0.00929584912955761, + 0.008829557336866856, + 0.008474085479974747, + 0.0094767389819026, + 0.008168697357177734, + 0.015378202311694622, + 0.007451454643160105, + 0.007714204490184784, + 0.05280567705631256, + 0.0442267507314682, + 0.015117790549993515, + 0.017108144238591194, + 0.015527920797467232, + 0.010406030341982841, + 0.01521873939782381, + 0.015005217865109444, + 0.008833115920424461, + 0.00944778323173523, + 0.010539217852056026, + 0.007429353427141905, + 0.009626317769289017, + 0.008249102160334587, + 0.009828325361013412, + 0.006114146672189236, + 0.007867848500609398, + 0.007030846551060677, + 0.005620777606964111, + 0.009326831437647343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.03814033046364784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643973920494318 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.007643973920494318, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643973920494318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11153785139322281, + 0.045469727367162704, + 0.05039934813976288, + 0.2696595788002014, + 0.13556165993213654, + 0.06352011114358902, + 0.11189183592796326, + 0.10104744136333466, + 0.02368992380797863, + 0.08722249418497086 + ], + "credit_map": [ + 0.015599764883518219, + 0.014922084286808968, + 0.008357567712664604, + 0.013964843936264515, + 0.016473716124892235, + 0.011738421395421028, + 0.017756545916199684, + 0.01272490806877613, + 0.004539484158158302, + 0.006285755429416895, + 0.00545112369582057, + 0.005923776421695948, + 0.005720385815948248, + 0.005711642559617758, + 0.006758233066648245, + 0.005079326685518026, + 0.006835512351244688, + 0.003971642814576626, + 0.008313324302434921, + 0.009426996111869812, + 0.0048471312038600445, + 0.005918676033616066, + 0.005794860888272524, + 0.005291208159178495, + 0.01678430661559105, + 0.056968796998262405, + 0.018771370872855186, + 0.04499908536672592, + 0.027305416762828827, + 0.040555451065301895, + 0.024257071316242218, + 0.04001808166503906, + 0.017636943608522415, + 0.011056260205805302, + 0.01163106132298708, + 0.01627138815820217, + 0.014928789809346199, + 0.025300169363617897, + 0.020363163203001022, + 0.018373876810073853, + 0.007644796743988991, + 0.00946461409330368, + 0.009585036896169186, + 0.007621037308126688, + 0.007429536897689104, + 0.007530112750828266, + 0.0069295987486839294, + 0.007315377704799175, + 0.00955058354884386, + 0.012612627819180489, + 0.010384500958025455, + 0.012250608764588833, + 0.015151659958064556, + 0.016729407012462616, + 0.016110656782984734, + 0.019101791083812714, + 0.003783061169087887, + 0.009174303151667118, + 0.00429965415969491, + 0.004201096948236227, + 0.033014070242643356, + 0.0263072457164526, + 0.011132410727441311, + 0.009135596454143524, + 0.0019703989382833242, + 0.003546937368810177, + 0.0029648668132722378, + 0.0038865217939019203, + 0.003639262868091464, + 0.002134861657395959, + 0.003306421684101224, + 0.002240650588646531, + 0.014257299713790417, + 0.010674957185983658, + 0.011949896812438965, + 0.009169980883598328, + 0.010219935327768326, + 0.00885334238409996, + 0.010202447883784771, + 0.011894629336893559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.007643973920494318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201059006154537 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201059006154537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201059006154537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11784318834543228, + 0.05960811302065849, + 0.06296300143003464, + 0.130977064371109, + 0.059515468776226044, + 0.18762080371379852, + 0.1683914214372635, + 0.06569412350654602, + 0.09368806332349777, + 0.05369880050420761 + ], + "credit_map": [ + 0.01404197234660387, + 0.016931846737861633, + 0.009969713166356087, + 0.015666155144572258, + 0.018784714862704277, + 0.01171316858381033, + 0.01753642037510872, + 0.013199190609157085, + 0.005742942448705435, + 0.008116538636386395, + 0.006049534305930138, + 0.005330361425876617, + 0.008156392723321915, + 0.008464973419904709, + 0.00951152853667736, + 0.008235848508775234, + 0.00780492601916194, + 0.0047972360625863075, + 0.011816018261015415, + 0.009959105402231216, + 0.0073349857702851295, + 0.0067950086668133736, + 0.007757109589874744, + 0.006698616314679384, + 0.010665645822882652, + 0.01794997788965702, + 0.012609164230525494, + 0.021818935871124268, + 0.014495705254375935, + 0.02178875356912613, + 0.012406319379806519, + 0.019242551177740097, + 0.006768761668354273, + 0.005566167179495096, + 0.006905397865921259, + 0.005888856947422028, + 0.006826985161751509, + 0.010301269590854645, + 0.009314322844147682, + 0.007943712174892426, + 0.027520935982465744, + 0.022587893530726433, + 0.02146069146692753, + 0.02292027696967125, + 0.021423503756523132, + 0.016087697818875313, + 0.0252845399081707, + 0.030335264280438423, + 0.01428942196071148, + 0.015186554752290249, + 0.013753407634794712, + 0.02044951356947422, + 0.02609233558177948, + 0.025392666459083557, + 0.027849819511175156, + 0.025377700105309486, + 0.006148329935967922, + 0.006101553328335285, + 0.005608221050351858, + 0.005587248597294092, + 0.016270706430077553, + 0.008346850983798504, + 0.010719822719693184, + 0.006911393254995346, + 0.009493761695921421, + 0.019696570932865143, + 0.006834571715444326, + 0.013922488316893578, + 0.011694532819092274, + 0.009356175549328327, + 0.013371003791689873, + 0.009318950586020947, + 0.010067579336464405, + 0.006865185219794512, + 0.00585970189422369, + 0.004468816798180342, + 0.0053522465750575066, + 0.006383712403476238, + 0.006816618610173464, + 0.007884942926466465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.014201059006154537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291917517781258 + ], + "opposite_logprob_tokens": [ + -3.773291826248169 + ], + "logprob_reward": -0.023291917517781258, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291917517781258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26649871468544006, + 0.03388075530529022, + 0.04219949617981911, + 0.1882060468196869, + 0.050270579755306244, + 0.07834616303443909, + 0.09690821170806885, + 0.12158811092376709, + 0.08710899204015732, + 0.03499290719628334 + ], + "credit_map": [ + 0.03344075381755829, + 0.027597231790423393, + 0.029436953365802765, + 0.03860296308994293, + 0.034019675105810165, + 0.027891743928194046, + 0.04043915495276451, + 0.03507024049758911, + 0.004225845914334059, + 0.006485181860625744, + 0.0038621029816567898, + 0.003449804149568081, + 0.005165209993720055, + 0.0037649297155439854, + 0.0027991498354822397, + 0.004128532018512487, + 0.004729556385427713, + 0.003371570957824588, + 0.0057119084522128105, + 0.005327762104570866, + 0.006081419065594673, + 0.004594577942043543, + 0.006710285320878029, + 0.005672418978065252, + 0.012597245164215565, + 0.04109535366296768, + 0.00953226163983345, + 0.019216462969779968, + 0.021629266440868378, + 0.0340254120528698, + 0.016023771837353706, + 0.034086283296346664, + 0.005969190504401922, + 0.004462364129722118, + 0.005737924482673407, + 0.006148000247776508, + 0.005525548942387104, + 0.005911689717322588, + 0.010484238155186176, + 0.006031625904142857, + 0.012574615888297558, + 0.00894530862569809, + 0.009055175818502903, + 0.007816188037395477, + 0.010819079354405403, + 0.008013366721570492, + 0.009469102136790752, + 0.011653336696326733, + 0.008960398845374584, + 0.010428965091705322, + 0.008174424059689045, + 0.012334713712334633, + 0.016104644164443016, + 0.014323138631880283, + 0.015334579162299633, + 0.011247352696955204, + 0.005515668541193008, + 0.010519743897020817, + 0.00510626845061779, + 0.007346886210143566, + 0.03931059688329697, + 0.02341560646891594, + 0.01937563344836235, + 0.010997719131410122, + 0.007967302575707436, + 0.02044844813644886, + 0.004653856158256531, + 0.012546820566058159, + 0.01033836416900158, + 0.008450194261968136, + 0.011682270094752312, + 0.011021746322512627, + 0.0037367672193795443, + 0.004104364197701216, + 0.0034165075048804283, + 0.003343892050907016, + 0.005131259560585022, + 0.003863163758069277, + 0.005822427105158567, + 0.005574529524892569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.023291917517781258 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures an endless stream of glossy photographic prints as they traverse the frame vertically, tumbling past the camera in a continuous, fluid motion. Each print, with a high-gloss finish and gently curled edges, twirls and flutters, revealing fleeting, out-of-focus glimpses of its contents—vibrant, saturated landscapes, intimate black-and-white portraits, and candid moments of joy. As this waterfall of memories tumbles downward, a powerful off-camera light source strikes the lens, causing a bright, film-like anamorphic lens flare to bloom and dominate the image. This flare appears as a warm, golden-white core with long, horizontal streaks of electric blue light that stretch from edge to edge, shimmering dynamically across the smooth surface of the photograph. The interplay between the chaotic dance of the falling prints and the stylized, dreamlike light creates a deeply nostalgic and visually stunning atmosphere, evoking potent memories against a dark, blurred backdrop.", + "seed": 16831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_8701_seed_16831_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.7319559950381518, + "vlm_logprob": -0.7319559950381518, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.7319559950381518, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.5324347019195557 + ], + "opposite_logprob_tokens": [ + -0.03243475407361984 + ], + "logprob_reward": -3.5324347019195557, + "yes_no_margin_reward": -3.5, + "token_credit_reward": -3.5324347019195557, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11266140639781952, + 0.05118364840745926, + 0.05461058020591736, + 0.04483446478843689, + 0.09369270503520966, + 0.11586582660675049, + 0.037012729793787, + 0.11299875378608704, + 0.06768426299095154, + 0.30945566296577454 + ], + "credit_map": [ + 0.011690364219248295, + 0.012117912992835045, + 0.012424929067492485, + 0.01885857619345188, + 0.018943745642900467, + 0.017599044367671013, + 0.012853465043008327, + 0.008173366077244282, + 0.006190089508891106, + 0.0040053511038422585, + 0.007603250909596682, + 0.009228578768670559, + 0.008756355382502079, + 0.005602088291198015, + 0.00539309810847044, + 0.0044048381969332695, + 0.004401242360472679, + 0.006328865885734558, + 0.006338014733046293, + 0.00589905958622694, + 0.007314285263419151, + 0.009313893504440784, + 0.007730513345450163, + 0.007284708321094513, + 0.004057940095663071, + 0.007332547567784786, + 0.004659189842641354, + 0.00680804206058383, + 0.00539845647290349, + 0.005860513541847467, + 0.006498679984360933, + 0.0042190938256680965, + 0.013471881859004498, + 0.010181794874370098, + 0.006244011223316193, + 0.015995554625988007, + 0.012242168188095093, + 0.01408145297318697, + 0.007821246981620789, + 0.013654599897563457, + 0.017600499093532562, + 0.010561649687588215, + 0.016974380239844322, + 0.01901012472808361, + 0.013030626811087132, + 0.015299713239073753, + 0.011230899021029472, + 0.012157943099737167, + 0.004628675989806652, + 0.004025762435048819, + 0.005193364340811968, + 0.005713548976927996, + 0.004117252305150032, + 0.004649294074624777, + 0.004737596493214369, + 0.003947234246879816, + 0.009874986484646797, + 0.008241401053965092, + 0.03240817040205002, + 0.01267209928482771, + 0.006153704598546028, + 0.005586749874055386, + 0.030675530433654785, + 0.0073860990814864635, + 0.010972825810313225, + 0.008484742604196072, + 0.007399152498692274, + 0.00864232238382101, + 0.009315058588981628, + 0.005979340057820082, + 0.009677705354988575, + 0.007213116157799959, + 0.022289009764790535, + 0.02352426014840603, + 0.025188902392983437, + 0.07882669568061829, + 0.018674951046705246, + 0.026415428146719933, + 0.03368906304240227, + 0.0808473601937294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -3.5324347019195557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04326719790697098 + ], + "opposite_logprob_tokens": [ + -3.168267250061035 + ], + "logprob_reward": -0.04326719790697098, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04326719790697098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10624506324529648, + 0.0684681385755539, + 0.06323429942131042, + 0.03338362276554108, + 0.059564266353845596, + 0.23388387262821198, + 0.03659241646528244, + 0.15140368044376373, + 0.17078545689582825, + 0.07643917202949524 + ], + "credit_map": [ + 0.011028106324374676, + 0.013821546919643879, + 0.012461825273931026, + 0.01775493659079075, + 0.018353549763560295, + 0.012681186199188232, + 0.013233496807515621, + 0.0069104149006307125, + 0.009067190811038017, + 0.006772232707589865, + 0.007257194258272648, + 0.011774474754929543, + 0.008433439768850803, + 0.0072379764169454575, + 0.0068743424490094185, + 0.011051289737224579, + 0.006028329487890005, + 0.006995754782110453, + 0.00850600004196167, + 0.006648840848356485, + 0.007115205749869347, + 0.011168067343533039, + 0.0076036532409489155, + 0.009168445132672787, + 0.002838252577930689, + 0.004673427436500788, + 0.0037188748829066753, + 0.005956385284662247, + 0.004542557056993246, + 0.004298912361264229, + 0.004728243220597506, + 0.0026269685477018356, + 0.008976515382528305, + 0.007335363887250423, + 0.004377693869173527, + 0.00832295697182417, + 0.006363260094076395, + 0.008198724128305912, + 0.007130964193493128, + 0.008858787827193737, + 0.036494653671979904, + 0.023426122963428497, + 0.03501833975315094, + 0.03821403533220291, + 0.029060564935207367, + 0.02722688764333725, + 0.02340736612677574, + 0.021035898476839066, + 0.0034581287764012814, + 0.005117390770465136, + 0.003977715037763119, + 0.005147228017449379, + 0.005604558624327183, + 0.005336965899914503, + 0.0042663756757974625, + 0.0036840559914708138, + 0.012132849544286728, + 0.012662743218243122, + 0.048654038459062576, + 0.012229514308273792, + 0.008379597216844559, + 0.005240822210907936, + 0.04220840334892273, + 0.009895715862512589, + 0.03073214180767536, + 0.017816446721553802, + 0.037347760051488876, + 0.022277576848864555, + 0.015537266619503498, + 0.013724284246563911, + 0.01882951147854328, + 0.014520463533699512, + 0.009339109063148499, + 0.008392008021473885, + 0.012074385769665241, + 0.008304806426167488, + 0.007495254278182983, + 0.011954440735280514, + 0.00900112371891737, + 0.009878053329885006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.04326719790697098 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016107195988297462 + ], + "opposite_logprob_tokens": [ + -4.141107082366943 + ], + "logprob_reward": -0.016107195988297462, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016107195988297462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13105779886245728, + 0.05949947237968445, + 0.09079749137163162, + 0.04039091244339943, + 0.07359398901462555, + 0.10670411586761475, + 0.04856616631150246, + 0.05664122477173805, + 0.2803223729133606, + 0.11242642998695374 + ], + "credit_map": [ + 0.011541083455085754, + 0.015712415799498558, + 0.014359332621097565, + 0.020925423130393028, + 0.02718491479754448, + 0.014795532450079918, + 0.0158492811024189, + 0.010689808987081051, + 0.008344386704266071, + 0.005336620844900608, + 0.008913572877645493, + 0.009517079219222069, + 0.009101740084588528, + 0.006632672157138586, + 0.006027800031006336, + 0.005625598598271608, + 0.007066949736326933, + 0.017187412828207016, + 0.010491786524653435, + 0.01918691210448742, + 0.009660429321229458, + 0.010241985321044922, + 0.009076220914721489, + 0.007885799743235111, + 0.003278515301644802, + 0.007461430039256811, + 0.004180108197033405, + 0.005876599345356226, + 0.005717353895306587, + 0.0050009265542030334, + 0.005177418235689402, + 0.003698560642078519, + 0.007990945130586624, + 0.008225110359489918, + 0.006315264385193586, + 0.01557149738073349, + 0.00893483217805624, + 0.009906958788633347, + 0.007679078262299299, + 0.008970308117568493, + 0.016475848853588104, + 0.009433100000023842, + 0.016394704580307007, + 0.01698983833193779, + 0.011619948782026768, + 0.014268049038946629, + 0.010483738034963608, + 0.011038897559046745, + 0.0052039786241948605, + 0.011293930932879448, + 0.006935355253517628, + 0.004589924588799477, + 0.005365100223571062, + 0.0040817479602992535, + 0.004863658919930458, + 0.006232470739632845, + 0.006686836015433073, + 0.00601347116753459, + 0.011069676838815212, + 0.0065164281986653805, + 0.006571069359779358, + 0.00416257418692112, + 0.011153181083500385, + 0.004467989318072796, + 0.05109917372465134, + 0.03046783246099949, + 0.06176029145717621, + 0.03630886599421501, + 0.02521338313817978, + 0.022176088765263557, + 0.031532905995845795, + 0.021763823926448822, + 0.008974806405603886, + 0.008822453208267689, + 0.010213731788098812, + 0.02744077704846859, + 0.006177336443215609, + 0.010175909847021103, + 0.012014910578727722, + 0.028606506064534187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.016107195988297462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381472185254097 + ], + "opposite_logprob_tokens": [ + -3.288147211074829 + ], + "logprob_reward": -0.0381472185254097, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381472185254097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06378594040870667, + 0.02667882852256298, + 0.048224300146102905, + 0.01762339472770691, + 0.033430591225624084, + 0.06634820252656937, + 0.05834019556641579, + 0.2922339141368866, + 0.11824686825275421, + 0.27508774399757385 + ], + "credit_map": [ + 0.005659003276377916, + 0.00838248711079359, + 0.006903559900820255, + 0.010783462785184383, + 0.011380044743418694, + 0.00722202192991972, + 0.007077845744788647, + 0.006377514451742172, + 0.002994090784341097, + 0.002232972765341401, + 0.003072121413424611, + 0.004043590277433395, + 0.003650824073702097, + 0.0030820996034890413, + 0.004162249155342579, + 0.0034408802166581154, + 0.003021024167537689, + 0.011036441661417484, + 0.003961396869271994, + 0.008915639482438564, + 0.007465943694114685, + 0.005262572318315506, + 0.004058731719851494, + 0.004502551630139351, + 0.0018837573006749153, + 0.002580336295068264, + 0.0015707432758063078, + 0.0018504589097574353, + 0.00295807677321136, + 0.002087592612951994, + 0.0028830820228904486, + 0.0018093474209308624, + 0.004063601139932871, + 0.0037579152267426252, + 0.0019408006919547915, + 0.007202500943094492, + 0.0037108082324266434, + 0.004511237144470215, + 0.003584371181204915, + 0.004659353289753199, + 0.009477894753217697, + 0.006428576074540615, + 0.009472724050283432, + 0.011425732634961605, + 0.007450096774846315, + 0.008736714720726013, + 0.006423959042876959, + 0.006932499352842569, + 0.004552183207124472, + 0.0075792246498167515, + 0.006079085171222687, + 0.013086393475532532, + 0.0060049318708479404, + 0.005180001258850098, + 0.009382469579577446, + 0.006475904956459999, + 0.024331482127308846, + 0.0240581464022398, + 0.10400034487247467, + 0.016388194635510445, + 0.0111478790640831, + 0.00785672478377819, + 0.08846161514520645, + 0.015989528968930244, + 0.014378576539456844, + 0.013251050375401974, + 0.01536468230187893, + 0.018131909891963005, + 0.009252007119357586, + 0.00949669536203146, + 0.028332963585853577, + 0.01003897562623024, + 0.01589811220765114, + 0.02245951071381569, + 0.01935397833585739, + 0.08888885378837585, + 0.013186000287532806, + 0.014955302700400352, + 0.027707375586032867, + 0.07263863831758499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0381472185254097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823660850524902 + ], + "opposite_logprob_tokens": [ + -3.5298237800598145 + ], + "logprob_reward": -0.029823660850524902, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823660850524902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10312388837337494, + 0.02397114410996437, + 0.04193869233131409, + 0.03079569712281227, + 0.04185916483402252, + 0.0345437154173851, + 0.08159253001213074, + 0.09249919652938843, + 0.15584661066532135, + 0.3938293755054474 + ], + "credit_map": [ + 0.009222953580319881, + 0.014087093062698841, + 0.010661534033715725, + 0.016406364738941193, + 0.019081344828009605, + 0.012147354893386364, + 0.011603934690356255, + 0.009913310408592224, + 0.002248877426609397, + 0.002763909287750721, + 0.0026218320708721876, + 0.0037766098976135254, + 0.002899444429203868, + 0.0026398268528282642, + 0.0036147700157016516, + 0.003405872266739607, + 0.0024092921521514654, + 0.00784626230597496, + 0.004125318955630064, + 0.008074981160461903, + 0.00575783709064126, + 0.0037832388188689947, + 0.004514953121542931, + 0.00542680686339736, + 0.0032918048091232777, + 0.004257954657077789, + 0.0023332873824983835, + 0.0034972697030752897, + 0.004546779673546553, + 0.004131430294364691, + 0.006066483911126852, + 0.002670684829354286, + 0.004807247314602137, + 0.004438112955540419, + 0.0031129501294344664, + 0.005331738386303186, + 0.005093677900731564, + 0.006124441046267748, + 0.006584170274436474, + 0.00636682566255331, + 0.003150174394249916, + 0.0038093666080385447, + 0.004193326458334923, + 0.004785032942891121, + 0.004496573004871607, + 0.005196100566536188, + 0.0036910013295710087, + 0.005222138948738575, + 0.005933318752795458, + 0.015525443479418755, + 0.008335841819643974, + 0.01020111981779337, + 0.011280236765742302, + 0.008099311031401157, + 0.010002408176660538, + 0.012214850634336472, + 0.008152092806994915, + 0.007128457073122263, + 0.03139307722449303, + 0.005424355622380972, + 0.0038067223504185677, + 0.003230880945920944, + 0.02805667743086815, + 0.005306925158947706, + 0.016240853816270828, + 0.0187209602445364, + 0.020941074937582016, + 0.02266860008239746, + 0.015912076458334923, + 0.013683599419891834, + 0.03345547616481781, + 0.014223982580006123, + 0.02449311688542366, + 0.028695544227957726, + 0.027534985914826393, + 0.11115237325429916, + 0.022569872438907623, + 0.03119954653084278, + 0.03878745436668396, + 0.10939648002386093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.029823660850524902 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures an endless stream of glossy photographic prints as they traverse the frame vertically, tumbling past the camera in a continuous, fluid motion. Each print, with a high-gloss finish and gently curled edges, twirls and flutters, revealing fleeting, out-of-focus glimpses of its contents—vibrant, saturated landscapes, intimate black-and-white portraits, and candid moments of joy. As this waterfall of memories tumbles downward, a powerful off-camera light source strikes the lens, causing a bright, film-like anamorphic lens flare to bloom and dominate the image. This flare appears as a warm, golden-white core with long, horizontal streaks of electric blue light that stretch from edge to edge, shimmering dynamically across the smooth surface of the photograph. The interplay between the chaotic dance of the falling prints and the stylized, dreamlike light creates a deeply nostalgic and visually stunning atmosphere, evoking potent memories against a dark, blurred backdrop.", + "seed": 16832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000068/group_6/index_8701_seed_16832_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22745333444327115, + "vlm_logprob": -0.22745333444327115, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.22745333444327115, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8287236094474792 + ], + "opposite_logprob_tokens": [ + -0.5787236094474792 + ], + "logprob_reward": -0.8287236094474792, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8287236094474792, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14403055608272552, + 0.09188354760408401, + 0.1434682309627533, + 0.07669682055711746, + 0.1413401961326599, + 0.03480718657374382, + 0.09671600908041, + 0.0527087077498436, + 0.043832533061504364, + 0.1745162457227707 + ], + "credit_map": [ + 0.012802347540855408, + 0.018659420311450958, + 0.017854338511824608, + 0.017887713387608528, + 0.03418949991464615, + 0.009857799857854843, + 0.01982545666396618, + 0.012953978031873703, + 0.006810742896050215, + 0.025188617408275604, + 0.00760317500680685, + 0.0082792853936553, + 0.00758703239262104, + 0.008276018314063549, + 0.00785774551331997, + 0.020280923694372177, + 0.007875816896557808, + 0.017781438305974007, + 0.015498513355851173, + 0.0090941172093153, + 0.037064164876937866, + 0.01216836180537939, + 0.016000792384147644, + 0.027985038235783577, + 0.006596629042178392, + 0.02606336399912834, + 0.003491026349365711, + 0.006714377552270889, + 0.004617112223058939, + 0.0066494811326265335, + 0.007714544422924519, + 0.014850287698209286, + 0.009482048451900482, + 0.013014303520321846, + 0.012992508709430695, + 0.037392668426036835, + 0.008881276473402977, + 0.017283916473388672, + 0.02005561627447605, + 0.022237854078412056, + 0.004400908946990967, + 0.0032516908831894398, + 0.002762238262221217, + 0.002568074269220233, + 0.0034716951195150614, + 0.00456186942756176, + 0.007602350786328316, + 0.006188362371176481, + 0.007634339854121208, + 0.023482367396354675, + 0.006561986170709133, + 0.009160961955785751, + 0.012528463266789913, + 0.007076529320329428, + 0.017994750291109085, + 0.012276612222194672, + 0.005729244090616703, + 0.006656074430793524, + 0.006071879994124174, + 0.005118818953633308, + 0.004839595407247543, + 0.005740150343626738, + 0.007990716025233269, + 0.010562225244939327, + 0.00438481941819191, + 0.005975150968879461, + 0.005250457674264908, + 0.007746380753815174, + 0.004057817626744509, + 0.0049650659784674644, + 0.005193323828279972, + 0.006259516347199678, + 0.02908279001712799, + 0.021081386134028435, + 0.018283184617757797, + 0.011765220202505589, + 0.009681974537670612, + 0.01973593421280384, + 0.018637044355273247, + 0.04624871164560318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.8287236094474792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332976832985878 + ], + "opposite_logprob_tokens": [ + -3.773329734802246 + ], + "logprob_reward": -0.02332976832985878, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332976832985878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10311251133680344, + 0.16479472815990448, + 0.1324075162410736, + 0.12246894836425781, + 0.07793423533439636, + 0.024150831624865532, + 0.09665226936340332, + 0.07152630388736725, + 0.09139923751354218, + 0.11555340886116028 + ], + "credit_map": [ + 0.011385052464902401, + 0.012707296758890152, + 0.010465900413691998, + 0.012798023410141468, + 0.02284586802124977, + 0.00752346683293581, + 0.014504489488899708, + 0.010882420465350151, + 0.011649597436189651, + 0.056252337992191315, + 0.009285153821110725, + 0.008052119053900242, + 0.011822628788650036, + 0.011993841268122196, + 0.011127067729830742, + 0.04461197927594185, + 0.007751148659735918, + 0.021051058545708656, + 0.01301746629178524, + 0.007651418447494507, + 0.03423363342881203, + 0.009938166476786137, + 0.01306013111025095, + 0.02570449374616146, + 0.009624749422073364, + 0.04511769488453865, + 0.0054924990981817245, + 0.008917052298784256, + 0.006211832165718079, + 0.00958564504981041, + 0.011962223798036575, + 0.025557253509759903, + 0.005756719503551722, + 0.006654858589172363, + 0.007517707068473101, + 0.019994055852293968, + 0.005561243277043104, + 0.009234732016921043, + 0.011688041500747204, + 0.011526877991855145, + 0.002503086579963565, + 0.0023512395564466715, + 0.001645875396206975, + 0.002128446474671364, + 0.002515894128009677, + 0.0030608773231506348, + 0.005633578170090914, + 0.004311834461987019, + 0.005487869028002024, + 0.015519872307777405, + 0.006490402389317751, + 0.011533301323652267, + 0.00942962896078825, + 0.006805919576436281, + 0.02321542426943779, + 0.018169855698943138, + 0.005010605789721012, + 0.008424133062362671, + 0.008331445045769215, + 0.009226840920746326, + 0.007310305722057819, + 0.007132658734917641, + 0.014281264506280422, + 0.011809052899479866, + 0.01097122672945261, + 0.009417334571480751, + 0.01752888597548008, + 0.01487086433917284, + 0.01048369612544775, + 0.009229044429957867, + 0.00904296338558197, + 0.009855222888290882, + 0.01936456747353077, + 0.016079258173704147, + 0.012027673423290253, + 0.008407128043472767, + 0.00775644788518548, + 0.012237370014190674, + 0.011816970072686672, + 0.027863988652825356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.02332976832985878 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801830165088177 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.009801830165088177, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801830165088177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1254550963640213, + 0.07276763767004013, + 0.13334278762340546, + 0.035989176481962204, + 0.10253981500864029, + 0.05559945106506348, + 0.11123347282409668, + 0.09804040938615799, + 0.08146189153194427, + 0.18357031047344208 + ], + "credit_map": [ + 0.012204146943986416, + 0.015798848122358322, + 0.013631898909807205, + 0.015076669864356518, + 0.02892439439892769, + 0.008290687575936317, + 0.018110128119587898, + 0.013418322429060936, + 0.004656991921365261, + 0.01951100118458271, + 0.004571356810629368, + 0.005780094303190708, + 0.007950631901621819, + 0.006012831814587116, + 0.006564158480614424, + 0.017720570787787437, + 0.008471783250570297, + 0.023889675736427307, + 0.012357998639345169, + 0.008425907231867313, + 0.03541991114616394, + 0.008332185447216034, + 0.01386735588312149, + 0.022577974945306778, + 0.003695365274325013, + 0.003933875821530819, + 0.004441510885953903, + 0.005258901510387659, + 0.0050777774304151535, + 0.0055426969192922115, + 0.004525037948042154, + 0.0035140106920152903, + 0.007140778936445713, + 0.008478857576847076, + 0.009599094279110432, + 0.02751360461115837, + 0.007425266318023205, + 0.012170304544270039, + 0.01493863482028246, + 0.015273275785148144, + 0.004336547572165728, + 0.006435580085963011, + 0.003501898841932416, + 0.006096659228205681, + 0.007804481778293848, + 0.007348406128585339, + 0.009532331489026546, + 0.01054354477673769, + 0.010275173000991344, + 0.01703036017715931, + 0.007917041890323162, + 0.013329323381185532, + 0.010968747548758984, + 0.007268374785780907, + 0.02454313263297081, + 0.019901316612958908, + 0.009221840649843216, + 0.010794045403599739, + 0.012758311815559864, + 0.011615765281021595, + 0.01108178123831749, + 0.009491640143096447, + 0.01603676751255989, + 0.017040252685546875, + 0.008158039301633835, + 0.01104775071144104, + 0.016970042139291763, + 0.013779752887785435, + 0.008340101689100266, + 0.008453154005110264, + 0.007833652198314667, + 0.0068794009275734425, + 0.02747766301035881, + 0.023332223296165466, + 0.020832858979701996, + 0.01573200151324272, + 0.011893649585545063, + 0.020451363176107407, + 0.019338632002472878, + 0.04451191797852516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.009801830165088177 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329925447702408 + ], + "opposite_logprob_tokens": [ + -3.773299217224121 + ], + "logprob_reward": -0.02329925447702408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329925447702408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03222160041332245, + 0.098825603723526, + 0.06035982817411423, + 0.04597713053226471, + 0.042765114456415176, + 0.014984901994466782, + 0.12148896604776382, + 0.06531381607055664, + 0.22412557899951935, + 0.29393747448921204 + ], + "credit_map": [ + 0.005097682122141123, + 0.00301003223285079, + 0.00409912271425128, + 0.004425411578267813, + 0.0056176939979195595, + 0.0025493965949863195, + 0.004110110457986593, + 0.0033121490851044655, + 0.006656178273260593, + 0.03411988168954849, + 0.00490109296515584, + 0.005029289983212948, + 0.007445193361490965, + 0.006739998236298561, + 0.007154883351176977, + 0.026779090985655785, + 0.0037066263612359762, + 0.006792312953621149, + 0.006512410473078489, + 0.00335711264051497, + 0.01652728021144867, + 0.00428816257044673, + 0.007073191460222006, + 0.012102733366191387, + 0.0037478951271623373, + 0.01561637781560421, + 0.0024069510400295258, + 0.003702603979036212, + 0.0029980302788317204, + 0.004156510345637798, + 0.0041758171282708645, + 0.009172946214675903, + 0.0029243272729218006, + 0.0040525333024561405, + 0.004344552755355835, + 0.010649562813341618, + 0.003133174264803529, + 0.005586546845734119, + 0.006306718103587627, + 0.0057676988653838634, + 0.0014559292467311025, + 0.0018103447509929538, + 0.0017683871556073427, + 0.001345262979157269, + 0.0018935934640467167, + 0.0026332198176532984, + 0.0018414977239444852, + 0.0022366668563336134, + 0.006575976498425007, + 0.009727176278829575, + 0.010272298008203506, + 0.015151843428611755, + 0.007188612129539251, + 0.009310652501881123, + 0.03736793249845505, + 0.025894470512866974, + 0.006843026727437973, + 0.007035321090370417, + 0.008106132969260216, + 0.008179917931556702, + 0.006679993588477373, + 0.006968826521188021, + 0.010841685347259045, + 0.010658914223313332, + 0.01980694942176342, + 0.03029860556125641, + 0.032447025179862976, + 0.026932094246149063, + 0.028970107436180115, + 0.021600870415568352, + 0.02184591256082058, + 0.04222399741411209, + 0.04836626723408699, + 0.03937701880931854, + 0.030875617638230324, + 0.020187709480524063, + 0.01636282168328762, + 0.03328549116849899, + 0.03113793209195137, + 0.07434462010860443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.02329925447702408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are glossy photographic prints moving vertically in a tumbling, fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the prints display a high-gloss finish with gently curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright anamorphic lens flare with a warm golden-white core and horizontal electric blue streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred behind the falling prints?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521122097969055 + ], + "opposite_logprob_tokens": [ + -1.5021121501922607 + ], + "logprob_reward": -0.2521122097969055, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521122097969055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02040702849626541, + 0.05833775922656059, + 0.03690559044480324, + 0.04065185412764549, + 0.02870657667517662, + 0.01903962343931198, + 0.05175889655947685, + 0.13688744604587555, + 0.040727648884058, + 0.5665775537490845 + ], + "credit_map": [ + 0.005193517077714205, + 0.002745784120634198, + 0.002032727934420109, + 0.0028843365143984556, + 0.0027129126247018576, + 0.0011104571167379618, + 0.001972029684111476, + 0.001755263190716505, + 0.004311060532927513, + 0.01823807880282402, + 0.003571047680452466, + 0.0035015116445720196, + 0.004114873707294464, + 0.004344561602920294, + 0.004955054260790348, + 0.015301575884222984, + 0.0026795563753694296, + 0.005676743574440479, + 0.003468900453299284, + 0.00181768205948174, + 0.010232595726847649, + 0.0021349708549678326, + 0.0038654019590467215, + 0.007029741536825895, + 0.0027292435988783836, + 0.010522868484258652, + 0.0026387013494968414, + 0.002937600715085864, + 0.003573748981580138, + 0.007580981124192476, + 0.003896030830219388, + 0.00677267974242568, + 0.0025393739342689514, + 0.0031101712957024574, + 0.0023493387270718813, + 0.007759415544569492, + 0.0022198264487087727, + 0.0032599351834505796, + 0.0029855056200176477, + 0.00448300875723362, + 0.001960666850209236, + 0.001519816112704575, + 0.0014002303360030055, + 0.001813254551962018, + 0.0017916600918397307, + 0.002384619088843465, + 0.004827361088246107, + 0.003342015901580453, + 0.0035855029709637165, + 0.009810416027903557, + 0.0051532587967813015, + 0.005619802977889776, + 0.005398482084274292, + 0.004701781552284956, + 0.009316547773778439, + 0.00817310530692339, + 0.015144003555178642, + 0.015671085566282272, + 0.013395501300692558, + 0.01490107923746109, + 0.013112439773976803, + 0.01933278515934944, + 0.019590137526392937, + 0.025740409269928932, + 0.00369994156062603, + 0.006265871226787567, + 0.004590251948684454, + 0.004806949757039547, + 0.0038970992900431156, + 0.003970632795244455, + 0.004909609444439411, + 0.008587293326854706, + 0.08090565353631973, + 0.06453222781419754, + 0.06861498206853867, + 0.04541162773966789, + 0.02958887815475464, + 0.06308777630329132, + 0.06586641818284988, + 0.1485699713230133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.2521122097969055 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_tensor([7339, 7339])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_tensor([7339, 7339])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b33df336cee699d592e027ecf4b57c58e6296ddb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_tensor([7339, 7339])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, static, extreme close-up offers a comprehensive view of a pristine, pale white ceramic bowl, its rim filled with thick, glossy cream. The cream's surface is spotless, a milky white that reflects the soft, directed morning light from an unseen window, taking on a pearlescent sheen. The bowl itself has a simple, matte appearance, contrasting sharply with its glazed interior. The bowl rests on a dark, rustic wooden tabletop, its deep veins providing a textural, earthy foundation. From the top of the frame, the tip of a slender, white nose descends with deliberate slowness, stopping just millimeters above the liquid, its shadow temporarily darkening the pure white. As one gently and deeply inhales the scent, the viscous cream surface produces only the faintest, almost imperceptible tremor—a single, slow-motion ripple that momentarily disrupts the perfect reflection before the entire scene returns to absolute stillness, creating an atmosphere of tranquility, intimacy, and profound sensory focus, focused on a moment of quiet indulgence.", + "seed": 10900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_7339_seed_10900_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07252880446612835, + "vlm_logprob": -0.07252880446612835, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.07252880446612835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08091805130243301 + ], + "opposite_logprob_tokens": [ + -3.705918073654175 + ], + "logprob_reward": -0.08091805130243301, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.08091805130243301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4411965310573578, + 0.1788637787103653, + 0.06502175331115723, + 0.04272407665848732, + 0.02078132890164852, + 0.034899190068244934, + 0.07825692743062973, + 0.05950193852186203, + 0.06006540358066559, + 0.018689030781388283 + ], + "credit_map": [ + 0.035872697830200195, + 0.14070925116539001, + 0.020268024876713753, + 0.10380911827087402, + 0.07607374340295792, + 0.01893218420445919, + 0.029520666226744652, + 0.01601083017885685, + 0.022518599405884743, + 0.024906693026423454, + 0.036180172115564346, + 0.018194440752267838, + 0.026295894756913185, + 0.014855616725981236, + 0.016923069953918457, + 0.018989285454154015, + 0.00850656908005476, + 0.010600117035210133, + 0.009102951735258102, + 0.007644381374120712, + 0.011134745553135872, + 0.003652349580079317, + 0.009450739249587059, + 0.004929899238049984, + 0.005334926303476095, + 0.005031125154346228, + 0.009081580676138401, + 0.005615562200546265, + 0.004068807698786259, + 0.003397561376914382, + 0.005741744767874479, + 0.004452764522284269, + 0.00258077634498477, + 0.0029801810160279274, + 0.00412810267880559, + 0.002259621163830161, + 0.0026889964938163757, + 0.0017493723426014185, + 0.002570562995970249, + 0.0018237170297652483, + 0.004687144886702299, + 0.004747047554701567, + 0.00688712764531374, + 0.004094053991138935, + 0.004189672879874706, + 0.0022710368502885103, + 0.0047677503898739815, + 0.0032553572673350573, + 0.010291688144207, + 0.014086969196796417, + 0.00929963868111372, + 0.010627701878547668, + 0.008991249836981297, + 0.005799008533358574, + 0.013153553009033203, + 0.006007120478898287, + 0.0061587742529809475, + 0.008993147872388363, + 0.006550310179591179, + 0.007906469516456127, + 0.007817908190190792, + 0.005930309649556875, + 0.011337975971400738, + 0.004807048477232456, + 0.007215091027319431, + 0.008229720406234264, + 0.01342572271823883, + 0.008852590806782246, + 0.0067848688922822475, + 0.0027279886417090893, + 0.0071336086839437485, + 0.005695811007171869, + 0.003221417311578989, + 0.0031446930952370167, + 0.0028833402320742607, + 0.0017606940818950534, + 0.0019092997536063194, + 0.0019268800970166922, + 0.0019893969874829054, + 0.001853307825513184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.08091805130243301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011120719835162163 + ], + "opposite_logprob_tokens": [ + -4.511120796203613 + ], + "logprob_reward": -0.011120719835162163, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011120719835162163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36802545189857483, + 0.18569831550121307, + 0.09520703554153442, + 0.04974566400051117, + 0.05210137367248535, + 0.046258337795734406, + 0.05987511947751045, + 0.04064887762069702, + 0.052637211978435516, + 0.04980265721678734 + ], + "credit_map": [ + 0.038653820753097534, + 0.0722213089466095, + 0.040987372398376465, + 0.053722746670246124, + 0.06915770471096039, + 0.024058952927589417, + 0.04995299503207207, + 0.019270537421107292, + 0.025135977193713188, + 0.028430677950382233, + 0.0300411619246006, + 0.0189899280667305, + 0.02081380970776081, + 0.019281288608908653, + 0.02053501270711422, + 0.02247045561671257, + 0.01319065410643816, + 0.012435066513717175, + 0.01980111002922058, + 0.01114747580140829, + 0.01002244371920824, + 0.005252755247056484, + 0.013906093314290047, + 0.00945143774151802, + 0.007307351101189852, + 0.007341138087213039, + 0.01012498140335083, + 0.006187539082020521, + 0.004752001259475946, + 0.003854138310998678, + 0.006034005433320999, + 0.0041445097886025906, + 0.009033702313899994, + 0.005910568404942751, + 0.011742498725652695, + 0.0054274494759738445, + 0.004449051339179277, + 0.0033798988442867994, + 0.005496697500348091, + 0.006661504507064819, + 0.007805336732417345, + 0.00543231749907136, + 0.010862557217478752, + 0.00550081068649888, + 0.004479917231947184, + 0.0035711380187422037, + 0.0035959023516625166, + 0.005010358057916164, + 0.011277496814727783, + 0.0075969938188791275, + 0.015521838329732418, + 0.006333638913929462, + 0.0043208906427025795, + 0.004289624281227589, + 0.003925676923245192, + 0.0066089569590985775, + 0.00545903854072094, + 0.004659928847104311, + 0.008948029018938541, + 0.005236020311713219, + 0.0034002375323325396, + 0.0033415774814784527, + 0.005375445354729891, + 0.004228596109896898, + 0.006500833667814732, + 0.006485239136964083, + 0.009008744731545448, + 0.009780457243323326, + 0.004762057214975357, + 0.0033204255159944296, + 0.008494662120938301, + 0.004284796305000782, + 0.006619823630899191, + 0.0059570325538516045, + 0.00878479890525341, + 0.007113473955541849, + 0.005186787340790033, + 0.004339487757533789, + 0.005859662778675556, + 0.005941590294241905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011120719835162163 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020622245967388153 + ], + "opposite_logprob_tokens": [ + -3.8956222534179688 + ], + "logprob_reward": -0.020622245967388153, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020622245967388153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6367343664169312, + 0.12649068236351013, + 0.05949240177869797, + 0.06603850424289703, + 0.028664249926805496, + 0.020656073465943336, + 0.014579880982637405, + 0.012721152976155281, + 0.015116596594452858, + 0.019506094977259636 + ], + "credit_map": [ + 0.044791966676712036, + 0.1931341290473938, + 0.03567656874656677, + 0.15958784520626068, + 0.0962684378027916, + 0.02530846744775772, + 0.05823260173201561, + 0.02373441867530346, + 0.012139391154050827, + 0.017922868952155113, + 0.012949667870998383, + 0.021183263510465622, + 0.012960731983184814, + 0.008101123385131359, + 0.028640875592827797, + 0.012592769227921963, + 0.007411022204905748, + 0.007615525741130114, + 0.010377934202551842, + 0.01028859056532383, + 0.006730330642312765, + 0.004040253348648548, + 0.007711462676525116, + 0.005317286122590303, + 0.008780856616795063, + 0.00880245491862297, + 0.010048056952655315, + 0.011596208438277245, + 0.006714646238833666, + 0.005041940603405237, + 0.009192995727062225, + 0.005861350800842047, + 0.00439657736569643, + 0.003294740105047822, + 0.00561551982536912, + 0.003752944292500615, + 0.0024145834613591433, + 0.0021934008691459894, + 0.003964880481362343, + 0.003031607251614332, + 0.002717119175940752, + 0.0031803695019334555, + 0.0029976365622133017, + 0.0032497129868716, + 0.0022233566269278526, + 0.0012471145018935204, + 0.0030392655171453953, + 0.0020015002228319645, + 0.0017532347701489925, + 0.002272333949804306, + 0.0019499670015648007, + 0.002366524888202548, + 0.0015499051660299301, + 0.0010653563076630235, + 0.002373355207964778, + 0.0012492046225816011, + 0.001572102541103959, + 0.00136292795650661, + 0.0024277183692902327, + 0.0017148703336715698, + 0.0010237634414806962, + 0.0013296152465045452, + 0.0019420742755755782, + 0.0013480822090059519, + 0.0020071102771908045, + 0.0016976770712062716, + 0.0030419081449508667, + 0.0015769912861287594, + 0.0015898948768153787, + 0.001554780756123364, + 0.0021779912058264017, + 0.0014702433254569769, + 0.003571336390450597, + 0.0028872715774923563, + 0.004120250698179007, + 0.0017193469684571028, + 0.0017591442447155714, + 0.0012617908651009202, + 0.0015085466438904405, + 0.0026784101501107216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.020622245967388153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004993408918381 + ], + "opposite_logprob_tokens": [ + -2.695050001144409 + ], + "logprob_reward": -0.07004993408918381, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004993408918381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5700809359550476, + 0.10574464499950409, + 0.030222173780202866, + 0.040070950984954834, + 0.05327499285340309, + 0.07237156480550766, + 0.05196778476238251, + 0.02379092015326023, + 0.029313327744603157, + 0.02316269837319851 + ], + "credit_map": [ + 0.04387721046805382, + 0.16597887873649597, + 0.05123884230852127, + 0.11190169304609299, + 0.10466542840003967, + 0.02677684836089611, + 0.0433199368417263, + 0.022322138771414757, + 0.017025571316480637, + 0.016301695257425308, + 0.02246592938899994, + 0.00967657845467329, + 0.011601860634982586, + 0.008598609827458858, + 0.009361968375742435, + 0.010712439194321632, + 0.003833117661997676, + 0.0041851201094686985, + 0.004328995943069458, + 0.004084884189069271, + 0.0028646860737353563, + 0.003231858368963003, + 0.004822915885597467, + 0.0028705967124551535, + 0.006579494569450617, + 0.005112074315547943, + 0.008480716496706009, + 0.004486303310841322, + 0.004046265035867691, + 0.0026106114964932203, + 0.004307185765355825, + 0.004448302555829287, + 0.008466470055282116, + 0.007297864183783531, + 0.011979630216956139, + 0.0054617649875581264, + 0.006574885919690132, + 0.0034722192212939262, + 0.004614526871591806, + 0.005407634191215038, + 0.011417013593018055, + 0.01028793677687645, + 0.013510070741176605, + 0.008581160567700863, + 0.006348108872771263, + 0.005318629089742899, + 0.008498289622366428, + 0.00841036718338728, + 0.0064019993878901005, + 0.006533701904118061, + 0.00629132054746151, + 0.008564210496842861, + 0.005437423475086689, + 0.0045190416276454926, + 0.008792692795395851, + 0.005427395459264517, + 0.0038068366702646017, + 0.003270351095125079, + 0.004941035993397236, + 0.0028989375568926334, + 0.0019720594864338636, + 0.002026352332904935, + 0.00228518177755177, + 0.0025901671033352613, + 0.005386005155742168, + 0.004064683802425861, + 0.006900164298713207, + 0.0029783190693706274, + 0.0021503486204892397, + 0.0020967356394976377, + 0.002562318230047822, + 0.0031747561879456043, + 0.003621321637183428, + 0.0034412986133247614, + 0.00408704811707139, + 0.0028887782245874405, + 0.002154112560674548, + 0.0015326881548389792, + 0.0027112902607768774, + 0.0027261620853096247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07004993408918381 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799330711364746 + ], + "opposite_logprob_tokens": [ + -1.8049330711364746 + ], + "logprob_reward": -0.1799330711364746, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799330711364746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5419790148735046, + 0.2556089460849762, + 0.02808259427547455, + 0.021162167191505432, + 0.04757362976670265, + 0.02081381343305111, + 0.019764216616749763, + 0.025357719510793686, + 0.027945538982748985, + 0.01171236764639616 + ], + "credit_map": [ + 0.03990240395069122, + 0.15003253519535065, + 0.03871507570147514, + 0.14796781539916992, + 0.07290715724229813, + 0.030076662078499794, + 0.0424889400601387, + 0.019888384267687798, + 0.03895806148648262, + 0.03880341351032257, + 0.05855362117290497, + 0.01946181245148182, + 0.033271390944719315, + 0.019532324746251106, + 0.018256165087223053, + 0.028772152960300446, + 0.00446246238425374, + 0.003748514922335744, + 0.005739601328969002, + 0.003964447882026434, + 0.002614131895825267, + 0.0016501299105584621, + 0.0031715051736682653, + 0.0027318010106682777, + 0.003438464365899563, + 0.0024192079436033964, + 0.004852725192904472, + 0.0024874869268387556, + 0.0022316633258014917, + 0.0013374665286391973, + 0.002289342926815152, + 0.0021058088168501854, + 0.007312482222914696, + 0.006357589270919561, + 0.010709241963922977, + 0.004932926967740059, + 0.006276105064898729, + 0.0031294890213757753, + 0.004248019307851791, + 0.0046077752485871315, + 0.003075281623750925, + 0.0030509342905133963, + 0.0035526473075151443, + 0.002384777646511793, + 0.0019280255073681474, + 0.001702275825664401, + 0.00286648771725595, + 0.0022533833980560303, + 0.002541180467233062, + 0.002882556989789009, + 0.0021936148405075073, + 0.003562128869816661, + 0.002170151798054576, + 0.0013198006199672818, + 0.003688693977892399, + 0.0014060877729207277, + 0.004172150045633316, + 0.0029041809029877186, + 0.005697781220078468, + 0.003727648174390197, + 0.0019525846000760794, + 0.0018315281486138701, + 0.0022175379563122988, + 0.002854307647794485, + 0.003315465524792671, + 0.004119950346648693, + 0.0041125258430838585, + 0.004510195925831795, + 0.003210179740563035, + 0.0016669691540300846, + 0.004866255447268486, + 0.0021439949050545692, + 0.0015446633333340287, + 0.001366900047287345, + 0.0021844038274139166, + 0.0021247833501547575, + 0.0012001704890280962, + 0.0009298464865423739, + 0.0011326463427394629, + 0.001228954060934484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.1799330711364746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static, extreme close-up offers a comprehensive view of a pristine, pale white ceramic bowl, its rim filled with thick, glossy cream. The cream's surface is spotless, a milky white that reflects the soft, directed morning light from an unseen window, taking on a pearlescent sheen. The bowl itself has a simple, matte appearance, contrasting sharply with its glazed interior. The bowl rests on a dark, rustic wooden tabletop, its deep veins providing a textural, earthy foundation. From the top of the frame, the tip of a slender, white nose descends with deliberate slowness, stopping just millimeters above the liquid, its shadow temporarily darkening the pure white. As one gently and deeply inhales the scent, the viscous cream surface produces only the faintest, almost imperceptible tremor—a single, slow-motion ripple that momentarily disrupts the perfect reflection before the entire scene returns to absolute stillness, creating an atmosphere of tranquility, intimacy, and profound sensory focus, focused on a moment of quiet indulgence.", + "seed": 10901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_7339_seed_10901_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12495251819491386, + "vlm_logprob": -0.12495251819491386, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.12495251819491386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054259780794382095 + ], + "opposite_logprob_tokens": [ + -3.929259777069092 + ], + "logprob_reward": -0.054259780794382095, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.054259780794382095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32669797539711, + 0.18248584866523743, + 0.11089432239532471, + 0.07169459015130997, + 0.0359744094312191, + 0.04585173353552818, + 0.0844273492693901, + 0.03635203838348389, + 0.07274702191352844, + 0.0328747034072876 + ], + "credit_map": [ + 0.028805436566472054, + 0.03079579584300518, + 0.019556302577257156, + 0.045733433216810226, + 0.053555723279714584, + 0.10618966817855835, + 0.020777074620127678, + 0.021284542977809906, + 0.018071305006742477, + 0.021006623283028603, + 0.031633052974939346, + 0.017359480261802673, + 0.03637262433767319, + 0.02105148322880268, + 0.022758541628718376, + 0.014232754707336426, + 0.009451950900256634, + 0.012270787730813026, + 0.021576112136244774, + 0.015164943411946297, + 0.012713488191366196, + 0.01148607861250639, + 0.019349591806530952, + 0.008881372399628162, + 0.007806890178471804, + 0.005904499441385269, + 0.012167421169579029, + 0.008287741802632809, + 0.009198883548378944, + 0.010134374722838402, + 0.009269173257052898, + 0.008925605565309525, + 0.004932962823659182, + 0.002820003079250455, + 0.0053125908598303795, + 0.004677127115428448, + 0.0036450133193284273, + 0.005311078857630491, + 0.005520701874047518, + 0.00375493336468935, + 0.005080593749880791, + 0.006836751010268927, + 0.009406703524291515, + 0.005091560538858175, + 0.007221258245408535, + 0.004512932617217302, + 0.004623261746019125, + 0.0030786737333983183, + 0.007871131412684917, + 0.00895602721720934, + 0.015252328477799892, + 0.01094709150493145, + 0.010079928673803806, + 0.011732125654816628, + 0.011715095490217209, + 0.00787361990660429, + 0.0036098577547818422, + 0.0032182903960347176, + 0.00847706664353609, + 0.004136030096560717, + 0.003987943287938833, + 0.00454986235126853, + 0.004080279730260372, + 0.004292710218578577, + 0.006911853328347206, + 0.00718571525067091, + 0.01266239583492279, + 0.009410157799720764, + 0.00964784063398838, + 0.01091220136731863, + 0.008896022103726864, + 0.007120836526155472, + 0.00427594967186451, + 0.002857442479580641, + 0.005999130196869373, + 0.002690673340111971, + 0.00480066891759634, + 0.005703653674572706, + 0.0026084082201123238, + 0.003938780166208744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.054259780794382095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029863810166716576 + ], + "opposite_logprob_tokens": [ + -3.5298638343811035 + ], + "logprob_reward": -0.029863810166716576, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029863810166716576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2728007435798645, + 0.29496312141418457, + 0.16340135037899017, + 0.05697717145085335, + 0.03927480801939964, + 0.045334141701459885, + 0.051189109683036804, + 0.02936263382434845, + 0.021418584510684013, + 0.02527836337685585 + ], + "credit_map": [ + 0.02449411153793335, + 0.03360234200954437, + 0.02837310917675495, + 0.03261208534240723, + 0.043009597808122635, + 0.054254643619060516, + 0.03206338360905647, + 0.02439146488904953, + 0.02980530820786953, + 0.034356627613306046, + 0.046997033059597015, + 0.03325531631708145, + 0.03481881320476532, + 0.036376502364873886, + 0.050991788506507874, + 0.0283617302775383, + 0.015233085490763187, + 0.013428330421447754, + 0.02999621257185936, + 0.01856192573904991, + 0.02137697860598564, + 0.02124929055571556, + 0.024407582357525826, + 0.01914794184267521, + 0.007455234881490469, + 0.009252366609871387, + 0.008265611715614796, + 0.006007197313010693, + 0.007625105325132608, + 0.006619241554290056, + 0.0066248569637537, + 0.005127554293721914, + 0.005024708807468414, + 0.00569212157279253, + 0.005024667363613844, + 0.004431293345987797, + 0.006342761218547821, + 0.004820681642740965, + 0.003733778139576316, + 0.004204795695841312, + 0.004878686275333166, + 0.00534416688606143, + 0.009361516684293747, + 0.005093729123473167, + 0.005674054846167564, + 0.0058781118132174015, + 0.004583178088068962, + 0.00452069565653801, + 0.006795777007937431, + 0.005502916872501373, + 0.00662960484623909, + 0.007105763535946608, + 0.007257693447172642, + 0.006580245681107044, + 0.007297868374735117, + 0.004019239451736212, + 0.003206737106665969, + 0.0031316003296524286, + 0.004426243714988232, + 0.002888521645218134, + 0.004050020594149828, + 0.005243490915745497, + 0.0025114913005381823, + 0.0039045270532369614, + 0.002476769033819437, + 0.0021808648016303778, + 0.004079821519553661, + 0.00207054172642529, + 0.002430156571790576, + 0.0027720644138753414, + 0.002432145643979311, + 0.0029762217309325933, + 0.0024144407361745834, + 0.0022681921254843473, + 0.0055236355401575565, + 0.00239786459133029, + 0.003440924920141697, + 0.004156507086008787, + 0.002407949883490801, + 0.002668847795575857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.029863810166716576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02639089711010456 + ], + "opposite_logprob_tokens": [ + -3.651390790939331 + ], + "logprob_reward": -0.02639089711010456, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639089711010456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33979299664497375, + 0.3056977391242981, + 0.16050581634044647, + 0.05478910729289055, + 0.029323462396860123, + 0.028733136132359505, + 0.02798587456345558, + 0.013314027339220047, + 0.01843436062335968, + 0.021423449739813805 + ], + "credit_map": [ + 0.029228748753666878, + 0.04267065227031708, + 0.031002622097730637, + 0.03962850943207741, + 0.06377113610506058, + 0.05961478874087334, + 0.04565262421965599, + 0.028223954141139984, + 0.028640812262892723, + 0.024239834398031235, + 0.04085468873381615, + 0.035183385014534, + 0.047187961637973785, + 0.037327609956264496, + 0.06255394965410233, + 0.02970951981842518, + 0.016499776393175125, + 0.01818588376045227, + 0.017347443848848343, + 0.020479027181863785, + 0.025694726034998894, + 0.01649862714111805, + 0.030618377029895782, + 0.01518197637051344, + 0.00663692457601428, + 0.00814003124833107, + 0.004536696709692478, + 0.0052336412481963634, + 0.011125095188617706, + 0.006258565466850996, + 0.007643758784979582, + 0.00521440152078867, + 0.0026182897854596376, + 0.002609415678307414, + 0.004939152393490076, + 0.0037292763590812683, + 0.0028362050652503967, + 0.0046386756002902985, + 0.004367801360785961, + 0.003584647784009576, + 0.002747531048953533, + 0.003105162875726819, + 0.004665894899517298, + 0.0032952569890767336, + 0.0034804069437086582, + 0.004021492321044207, + 0.0043072146363556385, + 0.0031101785134524107, + 0.0034127847757190466, + 0.0034425356425344944, + 0.004763002507388592, + 0.003023667959496379, + 0.003026171587407589, + 0.003924504388123751, + 0.0034415344707667828, + 0.002951674163341522, + 0.001376309897750616, + 0.0013734942767769098, + 0.0029148911125957966, + 0.0015043889870867133, + 0.0015120975440368056, + 0.001388768432661891, + 0.0018263333477079868, + 0.0014177454868331552, + 0.0018747762078419328, + 0.0015489573124796152, + 0.002014861674979329, + 0.002916518831625581, + 0.002061624778434634, + 0.0025044013746082783, + 0.003576492890715599, + 0.0019367298809811473, + 0.002378634177148342, + 0.0017938950331881642, + 0.004188519902527332, + 0.002124306745827198, + 0.0027318948414176702, + 0.0033737861085683107, + 0.0021331394091248512, + 0.0026992755010724068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02639089711010456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709876894950867 + ], + "opposite_logprob_tokens": [ + -2.127098798751831 + ], + "logprob_reward": -0.12709876894950867, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709876894950867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5119585990905762, + 0.1615712195634842, + 0.07386205345392227, + 0.03540601581335068, + 0.07033205777406693, + 0.024960020557045937, + 0.039665933698415756, + 0.025194112211465836, + 0.03902501240372658, + 0.018024995923042297 + ], + "credit_map": [ + 0.05201917141675949, + 0.08436700701713562, + 0.03325318172574043, + 0.03899345546960831, + 0.14763274788856506, + 0.06184135004878044, + 0.05292646586894989, + 0.04092523455619812, + 0.013981636613607407, + 0.014523159712553024, + 0.024135682731866837, + 0.016541533172130585, + 0.026003075763583183, + 0.024557875469326973, + 0.018317708745598793, + 0.02351054921746254, + 0.006445513106882572, + 0.007178834639489651, + 0.008333735167980194, + 0.008767517283558846, + 0.010437349788844585, + 0.013314490206539631, + 0.012639966793358326, + 0.00674464413896203, + 0.00452641025185585, + 0.0038549236487597227, + 0.0029857412446290255, + 0.004499790258705616, + 0.004466285463422537, + 0.005767870228737593, + 0.005573512054979801, + 0.0037314819637686014, + 0.007958690635859966, + 0.007890533655881882, + 0.010769397020339966, + 0.00746395718306303, + 0.012023057788610458, + 0.010407622903585434, + 0.0069079529494047165, + 0.006910843774676323, + 0.0027247797697782516, + 0.0036306476686149836, + 0.0035783371422439814, + 0.003560737706720829, + 0.003205907763913274, + 0.003200407838448882, + 0.00296009611338377, + 0.0020991056226193905, + 0.004509539809077978, + 0.003580312943086028, + 0.007417640648782253, + 0.004643869120627642, + 0.003997103311121464, + 0.005702633410692215, + 0.005555947311222553, + 0.004258889239281416, + 0.0020590373314917088, + 0.0035060965456068516, + 0.005671376828104258, + 0.002712006215006113, + 0.003338926937431097, + 0.0030893709044903517, + 0.002514702035114169, + 0.002302596578374505, + 0.0032403916120529175, + 0.003944621421396732, + 0.00611071428284049, + 0.004524921532720327, + 0.005288969725370407, + 0.0058235889300704, + 0.0058090537786483765, + 0.004282746929675341, + 0.0015650327550247312, + 0.001965639181435108, + 0.002428637584671378, + 0.002312357071787119, + 0.002592226956039667, + 0.002962198108434677, + 0.0024760449305176735, + 0.0017228578217327595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.12709876894950867 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871493339538574 + ], + "opposite_logprob_tokens": [ + -1.1371493339538574 + ], + "logprob_reward": -0.3871493339538574, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871493339538574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40040305256843567, + 0.18316563963890076, + 0.033192675560712814, + 0.06675627827644348, + 0.15209585428237915, + 0.08445882052183151, + 0.015460491180419922, + 0.017552247270941734, + 0.029197033494710922, + 0.01771789975464344 + ], + "credit_map": [ + 0.02076583541929722, + 0.038122184574604034, + 0.016139617189764977, + 0.0676272064447403, + 0.07601694762706757, + 0.12433315813541412, + 0.037772197276353836, + 0.01962590590119362, + 0.017267733812332153, + 0.01903422735631466, + 0.03779321908950806, + 0.018926599994301796, + 0.03016442060470581, + 0.021836576983332634, + 0.01779056526720524, + 0.020352303981781006, + 0.0033083679154515266, + 0.0031773250084370375, + 0.00467702466994524, + 0.0041821859776973724, + 0.00430008489638567, + 0.005144932772964239, + 0.004329496528953314, + 0.004073259886354208, + 0.007355580106377602, + 0.005972667131572962, + 0.009536444209516048, + 0.007552933879196644, + 0.010561932809650898, + 0.011607589200139046, + 0.007595100440084934, + 0.00657403189688921, + 0.01715468429028988, + 0.012307889759540558, + 0.02465338446199894, + 0.015071974135935307, + 0.023368965834379196, + 0.02702113427221775, + 0.014586003497242928, + 0.01793181151151657, + 0.010954495519399643, + 0.009117084555327892, + 0.01119800191372633, + 0.00885750725865364, + 0.013718029484152794, + 0.01290829572826624, + 0.008835497312247753, + 0.008869906887412071, + 0.0017659295117482543, + 0.0016805087216198444, + 0.0028013442642986774, + 0.0018083592876791954, + 0.0014926862204447389, + 0.0019918966572731733, + 0.0019166190177202225, + 0.0020031477324664593, + 0.0014870878076180816, + 0.0012552746338769794, + 0.0035962588153779507, + 0.00260612228885293, + 0.0015815615188330412, + 0.0020256415009498596, + 0.003192376112565398, + 0.0018079244764521718, + 0.0030477172695100307, + 0.0030808201991021633, + 0.0025823204778134823, + 0.0037732829805463552, + 0.0035474172327667475, + 0.00504206633195281, + 0.0046979160979390144, + 0.003425493137910962, + 0.0018960675224661827, + 0.0018787519074976444, + 0.0027907444164156914, + 0.0022398773580789566, + 0.0020165005698800087, + 0.0024337461218237877, + 0.0023430476430803537, + 0.002119164215400815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.3871493339538574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static, extreme close-up offers a comprehensive view of a pristine, pale white ceramic bowl, its rim filled with thick, glossy cream. The cream's surface is spotless, a milky white that reflects the soft, directed morning light from an unseen window, taking on a pearlescent sheen. The bowl itself has a simple, matte appearance, contrasting sharply with its glazed interior. The bowl rests on a dark, rustic wooden tabletop, its deep veins providing a textural, earthy foundation. From the top of the frame, the tip of a slender, white nose descends with deliberate slowness, stopping just millimeters above the liquid, its shadow temporarily darkening the pure white. As one gently and deeply inhales the scent, the viscous cream surface produces only the faintest, almost imperceptible tremor—a single, slow-motion ripple that momentarily disrupts the perfect reflection before the entire scene returns to absolute stillness, creating an atmosphere of tranquility, intimacy, and profound sensory focus, focused on a moment of quiet indulgence.", + "seed": 10902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_7339_seed_10902_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13919915929436683, + "vlm_logprob": -0.13919915929436683, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.13919915929436683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04687969759106636 + ], + "opposite_logprob_tokens": [ + -3.921879768371582 + ], + "logprob_reward": -0.04687969759106636, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04687969759106636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7159626483917236, + 0.035141684114933014, + 0.038092393428087234, + 0.03626639023423195, + 0.03606489300727844, + 0.036161426454782486, + 0.040034011006355286, + 0.02102590911090374, + 0.03092864900827408, + 0.01032197568565607 + ], + "credit_map": [ + 0.13631944358348846, + 0.28140774369239807, + 0.02727789804339409, + 0.027992237359285355, + 0.1043710857629776, + 0.04123246297240257, + 0.026854567229747772, + 0.07050717622041702, + 0.004822437185794115, + 0.005235485732555389, + 0.003992461133748293, + 0.001987128285691142, + 0.005333881359547377, + 0.006995301228016615, + 0.0034164683893322945, + 0.0033585287164896727, + 0.003562387079000473, + 0.0049658166244626045, + 0.0036209195386618376, + 0.0030223785433918238, + 0.012188354507088661, + 0.004867583978921175, + 0.0035265404731035233, + 0.0023384145461022854, + 0.0035059747751802206, + 0.004977730102837086, + 0.0031803748570382595, + 0.00369636039249599, + 0.006908060982823372, + 0.0036031855270266533, + 0.005376765970140696, + 0.005017940886318684, + 0.0036245814990252256, + 0.006093678995966911, + 0.003537857672199607, + 0.0036737655755132437, + 0.005404330790042877, + 0.00416222820058465, + 0.005164255853742361, + 0.004404198843985796, + 0.004207782447338104, + 0.006651477888226509, + 0.003072052961215377, + 0.003165025031194091, + 0.008367929607629776, + 0.004476823378354311, + 0.0029042160604149103, + 0.003316119546070695, + 0.0038114076014608145, + 0.0055756778456270695, + 0.0032465390395373106, + 0.0037332982756197453, + 0.00859193317592144, + 0.004388966131955385, + 0.005850652698427439, + 0.0048355404287576675, + 0.0019840258173644543, + 0.003139758249744773, + 0.002651108894497156, + 0.0016854553250595927, + 0.0034113596193492413, + 0.002716248156502843, + 0.0027359144296497107, + 0.00270204059779644, + 0.003976576495915651, + 0.00395731208845973, + 0.0035471925511956215, + 0.0029961224645376205, + 0.003325815312564373, + 0.004905145149677992, + 0.003508344991132617, + 0.004712143447250128, + 0.0010093566961586475, + 0.0012728305300697684, + 0.0013570805313065648, + 0.0010838625021278858, + 0.0013478567125275731, + 0.0013903439976274967, + 0.001546311192214489, + 0.001314335037022829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.04687969759106636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02986866980791092 + ], + "opposite_logprob_tokens": [ + -3.5298686027526855 + ], + "logprob_reward": -0.02986866980791092, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02986866980791092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2808723449707031, + 0.2750501334667206, + 0.09840124100446701, + 0.07279448956251144, + 0.12289764732122421, + 0.04057576507329941, + 0.03222832828760147, + 0.03447837010025978, + 0.015821630135178566, + 0.026880035176873207 + ], + "credit_map": [ + 0.04371485114097595, + 0.08697543293237686, + 0.014473192393779755, + 0.010776910930871964, + 0.04278986155986786, + 0.04822879657149315, + 0.017962615936994553, + 0.015950681641697884, + 0.026583625003695488, + 0.04064912721514702, + 0.023075954988598824, + 0.012402772903442383, + 0.037839096039533615, + 0.09263837337493896, + 0.023191353306174278, + 0.01866983063519001, + 0.01007221732288599, + 0.013942200690507889, + 0.008824853226542473, + 0.008987093344330788, + 0.028648795560002327, + 0.010644187219440937, + 0.012018559500575066, + 0.00526332575827837, + 0.00931550282984972, + 0.015130183659493923, + 0.008237716741859913, + 0.005304205697029829, + 0.010100391693413258, + 0.007684354204684496, + 0.007838723249733448, + 0.00918341614305973, + 0.020730797201395035, + 0.02413748949766159, + 0.017820322886109352, + 0.009536297991871834, + 0.015182914212346077, + 0.012832973152399063, + 0.0082158874720335, + 0.01444096490740776, + 0.004946381784975529, + 0.005911956541240215, + 0.0038984722923487425, + 0.003750220872461796, + 0.0072445496916770935, + 0.005915905814617872, + 0.004192415624856949, + 0.0047158668749034405, + 0.0033981390297412872, + 0.0037271291948854923, + 0.00592381414026022, + 0.0030277136247605085, + 0.004788586404174566, + 0.004322210792452097, + 0.0035488642752170563, + 0.0034918703604489565, + 0.0038776625879108906, + 0.004612200893461704, + 0.00473641837015748, + 0.0032563181594014168, + 0.004790347069501877, + 0.0055633471347391605, + 0.004167147912085056, + 0.003474928205832839, + 0.0018394900253042579, + 0.002415137132629752, + 0.002818033332005143, + 0.001486517139710486, + 0.0017423862591385841, + 0.001971092540770769, + 0.0014911845792084932, + 0.002057787962257862, + 0.002813054248690605, + 0.0035050944425165653, + 0.00321978353895247, + 0.0026307750958949327, + 0.0035276238340884447, + 0.003754207631573081, + 0.003974235616624355, + 0.0034552596043795347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02986866980791092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984853833913803 + ], + "opposite_logprob_tokens": [ + -3.529848575592041 + ], + "logprob_reward": -0.02984853833913803, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984853833913803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41523876786231995, + 0.25464189052581787, + 0.04555482417345047, + 0.04358173906803131, + 0.11138498038053513, + 0.06337913870811462, + 0.02809046022593975, + 0.013281630352139473, + 0.009400137700140476, + 0.015446508303284645 + ], + "credit_map": [ + 0.08456647396087646, + 0.1500389724969864, + 0.02421296015381813, + 0.016121618449687958, + 0.05354858934879303, + 0.03546018898487091, + 0.03200482204556465, + 0.019285114482045174, + 0.022221127524971962, + 0.036359865218400955, + 0.0220660250633955, + 0.011152959428727627, + 0.0365631990134716, + 0.08469290286302567, + 0.023776734247803688, + 0.01780906319618225, + 0.006779236253350973, + 0.008268258534371853, + 0.003805224783718586, + 0.003324144519865513, + 0.008583246730268002, + 0.005526602268218994, + 0.0044518280774354935, + 0.004816282074898481, + 0.005364915821701288, + 0.007721954956650734, + 0.004271320067346096, + 0.0034443072509020567, + 0.010277783498167992, + 0.004805429372936487, + 0.00401184381917119, + 0.003684181021526456, + 0.013862024061381817, + 0.020039085298776627, + 0.011506547220051289, + 0.009997650049626827, + 0.015749627724289894, + 0.011834999546408653, + 0.01508833933621645, + 0.013306694105267525, + 0.007780124898999929, + 0.011835929937660694, + 0.007467916235327721, + 0.005030556116253138, + 0.009754722006618977, + 0.009136416018009186, + 0.005997326225042343, + 0.006376140285283327, + 0.0027028436306864023, + 0.0035511190071702003, + 0.003907267469912767, + 0.002212353516370058, + 0.005389961879700422, + 0.003391440724954009, + 0.003979851026087999, + 0.0029556199442595243, + 0.0016254071379080415, + 0.0020527136512100697, + 0.0013981630327180028, + 0.0009766715811565518, + 0.002374177798628807, + 0.0025680027902126312, + 0.0010823342017829418, + 0.0012041585287079215, + 0.0010763569734990597, + 0.0011501895496621728, + 0.00148397299926728, + 0.0010394274722784758, + 0.0009387479512952268, + 0.001023519434966147, + 0.0012782145058736205, + 0.0014097081730142236, + 0.0015114715788513422, + 0.002864752896130085, + 0.002195423236116767, + 0.0013450449332594872, + 0.0021359764505177736, + 0.0018251556903123856, + 0.002083249157294631, + 0.0014854317996650934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02984853833913803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042059659957886 + ], + "opposite_logprob_tokens": [ + -1.9104206562042236 + ], + "logprob_reward": -0.16042059659957886, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042059659957886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4547828435897827, + 0.15415863692760468, + 0.07967457175254822, + 0.06688061356544495, + 0.057362720370292664, + 0.03704345226287842, + 0.0845256969332695, + 0.02654394321143627, + 0.01731400191783905, + 0.021713485941290855 + ], + "credit_map": [ + 0.06470630317926407, + 0.1548868864774704, + 0.020836925134062767, + 0.020317403599619865, + 0.07158125936985016, + 0.029258666560053825, + 0.023708052933216095, + 0.06948735564947128, + 0.016469581052660942, + 0.020269179716706276, + 0.012390971183776855, + 0.008240317925810814, + 0.02511253021657467, + 0.042472075670957565, + 0.016971902921795845, + 0.012232082895934582, + 0.009530261158943176, + 0.013833515346050262, + 0.005853882525116205, + 0.004957735072821379, + 0.02591732330620289, + 0.008692444302141666, + 0.005234663374722004, + 0.005654748994857073, + 0.006372049916535616, + 0.009768099524080753, + 0.0061164782382547855, + 0.004947308450937271, + 0.02075940929353237, + 0.007731775753200054, + 0.0060553462244570255, + 0.005130148958414793, + 0.005864403210580349, + 0.006513127125799656, + 0.0056252856738865376, + 0.006885404698550701, + 0.009638902731239796, + 0.006500157527625561, + 0.009336952120065689, + 0.006998489610850811, + 0.0038618813268840313, + 0.0046703931875526905, + 0.004046274349093437, + 0.0034059216268360615, + 0.008664615452289581, + 0.005751464050263166, + 0.003250706009566784, + 0.003392196726053953, + 0.007863856852054596, + 0.011124344542622566, + 0.0058208913542330265, + 0.009435108862817287, + 0.015876926481723785, + 0.010297676548361778, + 0.013820206746459007, + 0.010286685079336166, + 0.0021874569356441498, + 0.002948506036773324, + 0.0025163055397570133, + 0.0029299966990947723, + 0.004853765480220318, + 0.003198569407686591, + 0.004580455366522074, + 0.003328890772536397, + 0.0014730006223544478, + 0.002076136413961649, + 0.0016868335660547018, + 0.0019655178766697645, + 0.00270009390078485, + 0.0021261211950331926, + 0.003002055687829852, + 0.002284244168549776, + 0.0021251798607409, + 0.002915819874033332, + 0.003056858666241169, + 0.002181262243539095, + 0.003148299176245928, + 0.0027245772071182728, + 0.0035543525591492653, + 0.00200713612139225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.16042059659957886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42897829413414 + ], + "opposite_logprob_tokens": [ + -1.0539783239364624 + ], + "logprob_reward": -0.42897829413414, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42897829413414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29357537627220154, + 0.18239672482013702, + 0.12561537325382233, + 0.06291651725769043, + 0.057293638586997986, + 0.08518558740615845, + 0.15840847790241241, + 0.010002868250012398, + 0.014970839954912663, + 0.00963452085852623 + ], + "credit_map": [ + 0.035302530974149704, + 0.08441158384084702, + 0.015909826382994652, + 0.014655214734375477, + 0.0450434684753418, + 0.03485613688826561, + 0.022349180653691292, + 0.041047461330890656, + 0.016510166227817535, + 0.022180434316396713, + 0.018521888181567192, + 0.012300943024456501, + 0.04515638202428818, + 0.03774779289960861, + 0.01816627010703087, + 0.011812851764261723, + 0.013986365869641304, + 0.021420573815703392, + 0.010106518864631653, + 0.009693234227597713, + 0.03005928546190262, + 0.016353504732251167, + 0.01582854799926281, + 0.008167336694896221, + 0.008638444356620312, + 0.009329264983534813, + 0.005239165388047695, + 0.004626275505870581, + 0.014950363896787167, + 0.006697889417409897, + 0.006643385626375675, + 0.00679172296077013, + 0.006454649847000837, + 0.008185883052647114, + 0.007305345498025417, + 0.005123756360262632, + 0.011098279617726803, + 0.008726948872208595, + 0.00459049828350544, + 0.005808279383927584, + 0.00974191352725029, + 0.01119051594287157, + 0.009920505806803703, + 0.007994034327566624, + 0.017182964831590652, + 0.014571169391274452, + 0.006896990817040205, + 0.007687493693083525, + 0.014015397988259792, + 0.02093811333179474, + 0.009374682791531086, + 0.01611175760626793, + 0.035683684051036835, + 0.01638580672442913, + 0.026247411966323853, + 0.019651617854833603, + 0.000975974602624774, + 0.0014532540226355195, + 0.0018176758894696832, + 0.0008503610151819885, + 0.0011692048283293843, + 0.0012134419521316886, + 0.0015505433548241854, + 0.0009724125848151743, + 0.001633747131563723, + 0.0017590813804417849, + 0.001709891832433641, + 0.0015007874462753534, + 0.0024560296442359686, + 0.0019383929902687669, + 0.0022197277285158634, + 0.001753180637024343, + 0.0011799207422882318, + 0.0013475916348397732, + 0.0015106042847037315, + 0.0008030485478229821, + 0.0012539938325062394, + 0.0012999351602047682, + 0.001194806769490242, + 0.0010446208762004972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.42897829413414 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static, extreme close-up offers a comprehensive view of a pristine, pale white ceramic bowl, its rim filled with thick, glossy cream. The cream's surface is spotless, a milky white that reflects the soft, directed morning light from an unseen window, taking on a pearlescent sheen. The bowl itself has a simple, matte appearance, contrasting sharply with its glazed interior. The bowl rests on a dark, rustic wooden tabletop, its deep veins providing a textural, earthy foundation. From the top of the frame, the tip of a slender, white nose descends with deliberate slowness, stopping just millimeters above the liquid, its shadow temporarily darkening the pure white. As one gently and deeply inhales the scent, the viscous cream surface produces only the faintest, almost imperceptible tremor—a single, slow-motion ripple that momentarily disrupts the perfect reflection before the entire scene returns to absolute stillness, creating an atmosphere of tranquility, intimacy, and profound sensory focus, focused on a moment of quiet indulgence.", + "seed": 10903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_0/index_7339_seed_10903_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12100811786949635, + "vlm_logprob": -0.12100811786949635, + "vlm_yes_no_margin": 2.924999976158142, + "vlm_token_credit": -0.12100811786949635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05412416532635689 + ], + "opposite_logprob_tokens": [ + -3.929124116897583 + ], + "logprob_reward": -0.05412416532635689, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05412416532635689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1076800525188446, + 0.2635449171066284, + 0.058464664965867996, + 0.032593246549367905, + 0.01927431859076023, + 0.21307557821273804, + 0.06101662293076515, + 0.048224546015262604, + 0.14476703107357025, + 0.05135896056890488 + ], + "credit_map": [ + 0.011558684520423412, + 0.010083233006298542, + 0.016353249549865723, + 0.01246931217610836, + 0.01118514034897089, + 0.00973370298743248, + 0.019028140231966972, + 0.0172685906291008, + 0.028068773448467255, + 0.10270620882511139, + 0.009963853284716606, + 0.01037947740405798, + 0.03579701855778694, + 0.017122870311141014, + 0.014221717603504658, + 0.04528500884771347, + 0.004592792596668005, + 0.007847790606319904, + 0.007781671825796366, + 0.006162138655781746, + 0.01222176756709814, + 0.0077524189837276936, + 0.007193904835730791, + 0.004912178497761488, + 0.003343038260936737, + 0.003202376887202263, + 0.005299457814544439, + 0.003996413666754961, + 0.0053876228630542755, + 0.0046782707795500755, + 0.00352649437263608, + 0.003159572137519717, + 0.001370732788927853, + 0.003404814749956131, + 0.002818615408614278, + 0.0022758925333619118, + 0.0037359357811510563, + 0.002305473666638136, + 0.0016847086371853948, + 0.0016781463054940104, + 0.015848131850361824, + 0.08575496077537537, + 0.007346848491579294, + 0.008815830573439598, + 0.04368466138839722, + 0.011882235296070576, + 0.007380971685051918, + 0.032361939549446106, + 0.0033539917785674334, + 0.025842862203717232, + 0.0023609879426658154, + 0.0019324113382026553, + 0.019939254969358444, + 0.002353444928303361, + 0.0019952652510255575, + 0.003238398116081953, + 0.0030461247079074383, + 0.01911298744380474, + 0.0026574323419481516, + 0.0023809962440282106, + 0.011478693224489689, + 0.002272471319884062, + 0.0018398264655843377, + 0.0054360139183700085, + 0.0075040594674646854, + 0.041209105402231216, + 0.005925777833908796, + 0.004174128174781799, + 0.05502373352646828, + 0.007998520508408546, + 0.007071061059832573, + 0.01586063764989376, + 0.003518124110996723, + 0.013085375539958477, + 0.0028615654446184635, + 0.002939016092568636, + 0.017123151570558548, + 0.003403429640457034, + 0.003596680471673608, + 0.004831614904105663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.05412416532635689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020632172003388405 + ], + "opposite_logprob_tokens": [ + -3.895632266998291 + ], + "logprob_reward": -0.020632172003388405, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020632172003388405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08448679745197296, + 0.05573868751525879, + 0.059150174260139465, + 0.02966219373047352, + 0.03379464149475098, + 0.21313364803791046, + 0.019691303372383118, + 0.23876693844795227, + 0.23749792575836182, + 0.0280776284635067 + ], + "credit_map": [ + 0.006855476647615433, + 0.01676602102816105, + 0.011230241507291794, + 0.008811297826468945, + 0.013462294824421406, + 0.007111233659088612, + 0.010009068064391613, + 0.010241159237921238, + 0.004995203111320734, + 0.00941177736967802, + 0.006886565592139959, + 0.006640520878136158, + 0.012237956747412682, + 0.004479085095226765, + 0.005295099690556526, + 0.005792477633804083, + 0.004765965510159731, + 0.0073317610658705235, + 0.00834871456027031, + 0.00684911897405982, + 0.012874537147581577, + 0.006240226328372955, + 0.006529158912599087, + 0.006210689898580313, + 0.0024616585578769445, + 0.0036616690922528505, + 0.0031097084283828735, + 0.004431011620908976, + 0.0052721076644957066, + 0.004036948084831238, + 0.0038565013092011213, + 0.0028325868770480156, + 0.003218875266611576, + 0.0048779998905956745, + 0.003727974835783243, + 0.0033175330609083176, + 0.008914567530155182, + 0.004112487193197012, + 0.0030730823054909706, + 0.0025521216448396444, + 0.013049459084868431, + 0.07586658000946045, + 0.008086154237389565, + 0.005934083368629217, + 0.08084306120872498, + 0.010930991731584072, + 0.005570055451244116, + 0.012853274121880531, + 0.0017402670346200466, + 0.004593048710376024, + 0.001869727624580264, + 0.0019590076990425587, + 0.004850499797612429, + 0.0018825582228600979, + 0.0015189293771982193, + 0.0012772645568475127, + 0.015591554343700409, + 0.09951343387365341, + 0.013005572371184826, + 0.007792712189257145, + 0.061532072722911835, + 0.010627208277583122, + 0.006657428108155727, + 0.02404695749282837, + 0.012756881304085255, + 0.06816252321004868, + 0.008655146695673466, + 0.005526110529899597, + 0.09356722980737686, + 0.012862770818173885, + 0.008553133346140385, + 0.027414139360189438, + 0.0032606415916234255, + 0.004608056508004665, + 0.0035256966948509216, + 0.003778978483751416, + 0.0037459132727235556, + 0.0020078876987099648, + 0.004696741700172424, + 0.0024537132121622562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.020632172003388405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609886810183525 + ], + "opposite_logprob_tokens": [ + -4.141098976135254 + ], + "logprob_reward": -0.01609886810183525, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609886810183525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16309207677841187, + 0.07055608928203583, + 0.10156761854887009, + 0.08184304088354111, + 0.022219479084014893, + 0.08713238686323166, + 0.032437827438116074, + 0.2506594657897949, + 0.17199181020259857, + 0.018500179052352905 + ], + "credit_map": [ + 0.015435444191098213, + 0.03289758786559105, + 0.011884186416864395, + 0.01759415492415428, + 0.028546683490276337, + 0.013749940320849419, + 0.025273462757468224, + 0.017710614949464798, + 0.005672065541148186, + 0.01451815664768219, + 0.007955617271363735, + 0.00784735381603241, + 0.012202015146613121, + 0.007263540755957365, + 0.0073167430236935616, + 0.0077805910259485245, + 0.006832752376794815, + 0.01627301797270775, + 0.009696313180029392, + 0.00885795708745718, + 0.03030555509030819, + 0.010761071927845478, + 0.010011011734604836, + 0.008829936385154724, + 0.006955786142498255, + 0.01056830957531929, + 0.0073623862117528915, + 0.011291496455669403, + 0.019383110105991364, + 0.009336533024907112, + 0.009712651371955872, + 0.0072327712550759315, + 0.0019856183789670467, + 0.004337902646511793, + 0.00229088612832129, + 0.002173200948163867, + 0.005492447875440121, + 0.002551317447796464, + 0.0017118991818279028, + 0.0016762049635872245, + 0.009230171330273151, + 0.012866731733083725, + 0.012305661104619503, + 0.012511633336544037, + 0.01201058179140091, + 0.008391955867409706, + 0.009821695275604725, + 0.0099939638748765, + 0.0029504906851798296, + 0.007122897543013096, + 0.0027600876055657864, + 0.0038139531388878822, + 0.006432921160012484, + 0.002926743123680353, + 0.003172254655510187, + 0.0032584776636213064, + 0.015631141141057014, + 0.10808948427438736, + 0.008166537620127201, + 0.007593514397740364, + 0.0685414969921112, + 0.010302506387233734, + 0.006855465471744537, + 0.025479314848780632, + 0.009149203076958656, + 0.05129506066441536, + 0.006107898894697428, + 0.004969557281583548, + 0.06782317906618118, + 0.009015437215566635, + 0.006275019142776728, + 0.01735646091401577, + 0.0015970751410350204, + 0.004249799530953169, + 0.000989318359643221, + 0.001105321105569601, + 0.00582880200818181, + 0.001179965678602457, + 0.0016923316288739443, + 0.0018575650174170732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.01609886810183525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270836442708969 + ], + "opposite_logprob_tokens": [ + -2.1270835399627686 + ], + "logprob_reward": -0.1270836442708969, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270836442708969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10969682037830353, + 0.06431589275598526, + 0.0907951146364212, + 0.018575366586446762, + 0.01335906982421875, + 0.05735296010971069, + 0.05248470604419708, + 0.3239229917526245, + 0.26045024394989014, + 0.009046892635524273 + ], + "credit_map": [ + 0.010932493954896927, + 0.013527940027415752, + 0.008704950101673603, + 0.009963947348296642, + 0.025174425914883614, + 0.008812544867396355, + 0.017644518986344337, + 0.014935990795493126, + 0.00640120217576623, + 0.014924533665180206, + 0.004650472197681665, + 0.004923244472593069, + 0.012136436067521572, + 0.00720018008723855, + 0.006272138562053442, + 0.007807687856256962, + 0.006621683947741985, + 0.010805070400238037, + 0.0144775016233325, + 0.010693193413317204, + 0.02097361534833908, + 0.008836827240884304, + 0.008686352521181107, + 0.009700869210064411, + 0.0014481504913419485, + 0.0033340954687446356, + 0.0017457656795158982, + 0.0027002000715583563, + 0.003530655987560749, + 0.001804571831598878, + 0.0023187289480119944, + 0.0016931983409449458, + 0.0013308326015248895, + 0.0014485487481579185, + 0.0019876009318977594, + 0.0020053584594279528, + 0.0022079581394791603, + 0.001534431939944625, + 0.001652750070206821, + 0.00119158870074898, + 0.003721440676599741, + 0.023868883028626442, + 0.0019204936688765883, + 0.0024719582870602608, + 0.016350790858268738, + 0.0024163404013961554, + 0.001703666872344911, + 0.004899386782199144, + 0.0037868234794586897, + 0.012301255017518997, + 0.008177561685442924, + 0.005281195044517517, + 0.012051504105329514, + 0.0027361768297851086, + 0.004494684282690287, + 0.0036555123515427113, + 0.020863603800535202, + 0.13745549321174622, + 0.01358609925955534, + 0.010080201551318169, + 0.08636701852083206, + 0.01368328183889389, + 0.009347021579742432, + 0.032540254294872284, + 0.013650098815560341, + 0.07786674052476883, + 0.010670438408851624, + 0.007626035716384649, + 0.10270542651414871, + 0.013976728543639183, + 0.008324731141328812, + 0.0256300400942564, + 0.000808012904599309, + 0.0020692991092801094, + 0.000777563254814595, + 0.0007745102629996836, + 0.0024741217494010925, + 0.0006937812431715429, + 0.0007984139956533909, + 0.000651189882773906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.1270836442708969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle, static, extreme close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine, pale white ceramic bowl with a matte exterior and glazed interior visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bowl filled with thick, glossy, milky white cream that has a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender white nose descend from the top of the frame and stop just millimeters above the cream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cream surface produce only a single, slow-motion ripple before returning to absolute stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871017396450043 + ], + "opposite_logprob_tokens": [ + -1.1371017694473267 + ], + "logprob_reward": -0.3871017396450043, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871017396450043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1743873506784439, + 0.21054551005363464, + 0.053919240832328796, + 0.023975791409611702, + 0.0408843494951725, + 0.2982988655567169, + 0.016176050528883934, + 0.1230655238032341, + 0.04783419519662857, + 0.010913077741861343 + ], + "credit_map": [ + 0.016933510079979897, + 0.02847779355943203, + 0.015135053545236588, + 0.015915386378765106, + 0.043879859149456024, + 0.018148193135857582, + 0.02034711092710495, + 0.015550446696579456, + 0.022041188552975655, + 0.07810168713331223, + 0.009171754121780396, + 0.010579122230410576, + 0.030812477692961693, + 0.013320661149919033, + 0.011813344433903694, + 0.03470528498291969, + 0.004233960062265396, + 0.007329236716032028, + 0.00594715541228652, + 0.004785304423421621, + 0.014732817187905312, + 0.0054552326910197735, + 0.006543847266584635, + 0.004891687072813511, + 0.0029162063729017973, + 0.0026591981295496225, + 0.003256128402426839, + 0.004037928301841021, + 0.0035501294769346714, + 0.0025223391130566597, + 0.002777507295832038, + 0.0022563538514077663, + 0.0036445464938879013, + 0.00596819631755352, + 0.006255027372390032, + 0.0042279805056750774, + 0.00949003268033266, + 0.0035934632178395987, + 0.004340853076428175, + 0.0033642516937106848, + 0.014791719615459442, + 0.1290084719657898, + 0.008628019131720066, + 0.006841687019914389, + 0.11579561233520508, + 0.01119754184037447, + 0.0060756392776966095, + 0.005960206966847181, + 0.0015025801258161664, + 0.0026211158838123083, + 0.0017034192569553852, + 0.0025735623203217983, + 0.0033118294086307287, + 0.0012530023232102394, + 0.0019911681301891804, + 0.0012193730799481273, + 0.007935401052236557, + 0.050799835473299026, + 0.007170277182012796, + 0.004245083313435316, + 0.03155694156885147, + 0.005262200720608234, + 0.0037615522742271423, + 0.012334226630628109, + 0.002898595994338393, + 0.013738472945988178, + 0.0023852046579122543, + 0.002290965523570776, + 0.0175881776958704, + 0.002703289268538356, + 0.0024130698293447495, + 0.0038164211437106133, + 0.0009606844396330416, + 0.002118184231221676, + 0.0013197659282013774, + 0.0011129900813102722, + 0.0026818106416612864, + 0.0010323114693164825, + 0.0009144042269326746, + 0.0007729270146228373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.3871017396450043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_tensor([6454, 6454])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_tensor([6454, 6454])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa81267ca9e37035bfbe1ba8f35874f1553c8a01 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_tensor([6454, 6454])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, extreme close-up captures an empty, unbranded aluminum soda can resting on a rough, dark gray asphalt surface. The midday sun glints off its silvery, cylindrical form, revealing faint marks of wear and an open pull tab. From the top of the frame, the massive black rubber tread of a heavy vehicle's tire slowly and deliberately rolls into view, its deep grooves filled with dust. As the tire makes contact, immense downward pressure is instantly applied, and the can squirms violently, its smooth sides wrinkling and folding inward with a sharp, metallic creak. The camera remains steady as the weight relentlessly compresses the structure, forcing the aluminum to scream and twist as it is systematically squeezed, flattening into a contorted shape against the hard ground. The tire rolls completely past, revealing the can's final form: a wrinkled, flattened, and shattered disc of metal, its original shape obliterated, now a mere silvery fragment pressed against the dark, richly textured asphalt, casting a small, sharp shadow in the bright sunlight.", + "seed": 30640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_6454_seed_30640_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19190712026320397, + "vlm_logprob": -0.19190712026320397, + "vlm_yes_no_margin": 3.199999952316284, + "vlm_token_credit": -0.19190712026320397, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2371402531862259 + ], + "opposite_logprob_tokens": [ + -1.6121402978897095 + ], + "logprob_reward": -0.2371402531862259, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2371402531862259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15008197724819183, + 0.11586695909500122, + 0.03191628307104111, + 0.04191679134964943, + 0.03438341245055199, + 0.0795455351471901, + 0.01649087853729725, + 0.05762537196278572, + 0.21274833381175995, + 0.25942447781562805 + ], + "credit_map": [ + 0.021087054163217545, + 0.018060708418488503, + 0.01979896053671837, + 0.019287511706352234, + 0.012136354111135006, + 0.018365047872066498, + 0.027339156717061996, + 0.014007184654474258, + 0.017391573637723923, + 0.01445973664522171, + 0.010978027246892452, + 0.014435060322284698, + 0.010337349958717823, + 0.011238640174269676, + 0.027382662519812584, + 0.009643904864788055, + 0.007486519869416952, + 0.002567221410572529, + 0.0029759444296360016, + 0.002807430922985077, + 0.002926414832472801, + 0.0026850560680031776, + 0.004599745851010084, + 0.005867949221283197, + 0.004901317413896322, + 0.005131164565682411, + 0.0045800721272826195, + 0.00456191785633564, + 0.004413172602653503, + 0.004198803566396236, + 0.0036768114659935236, + 0.010453528724610806, + 0.008701457642018795, + 0.005169948097318411, + 0.004114855546504259, + 0.0033060945570468903, + 0.003844715654850006, + 0.00307366787455976, + 0.0027096918784081936, + 0.003462983760982752, + 0.0074754455126821995, + 0.016780542209744453, + 0.017357423901557922, + 0.006172527093440294, + 0.008127243258059025, + 0.008975920267403126, + 0.007248528767377138, + 0.00740790506824851, + 0.0013984584948047996, + 0.0022721050772815943, + 0.002569336909800768, + 0.00164445408154279, + 0.001746745896525681, + 0.0021870683412998915, + 0.002290647476911545, + 0.0023820633068680763, + 0.006612658500671387, + 0.006935404147952795, + 0.010198939591646194, + 0.0048103053122758865, + 0.006113529670983553, + 0.008133789524435997, + 0.004461098462343216, + 0.010359646752476692, + 0.027370642870664597, + 0.017451992258429527, + 0.04337269812822342, + 0.030474163591861725, + 0.015819832682609558, + 0.029893508180975914, + 0.022008446976542473, + 0.026357043534517288, + 0.02319801226258278, + 0.029492031782865524, + 0.048615746200084686, + 0.02737485244870186, + 0.03802892565727234, + 0.02675836905837059, + 0.03500417619943619, + 0.03095235303044319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.2371402531862259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640897762030363 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007640897762030363, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640897762030363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048037972301244736, + 0.09208427369594574, + 0.023096660152077675, + 0.03451571613550186, + 0.023140281438827515, + 0.04035027325153351, + 0.04048129543662071, + 0.06051267683506012, + 0.07480394095182419, + 0.5629770159721375 + ], + "credit_map": [ + 0.006715708412230015, + 0.004566436167806387, + 0.007655912544578314, + 0.0042749191634356976, + 0.003259846707805991, + 0.009904882870614529, + 0.005262075923383236, + 0.006398186553269625, + 0.018507778644561768, + 0.007174098864197731, + 0.012513458728790283, + 0.008920560590922832, + 0.010341782122850418, + 0.013224845752120018, + 0.008877390995621681, + 0.012524357996881008, + 0.0034289383329451084, + 0.0027603849302977324, + 0.0033390377648174763, + 0.0032799167092889547, + 0.002977637806907296, + 0.0021399848628789186, + 0.0030974752735346556, + 0.0020732830744236708, + 0.009452038444578648, + 0.0038865457754582167, + 0.002641481114551425, + 0.0024962371680885553, + 0.005377943627536297, + 0.002209922531619668, + 0.00153478910215199, + 0.006916755344718695, + 0.006299639120697975, + 0.0033622130285948515, + 0.0028516657184809446, + 0.001917560352012515, + 0.0018808824243023992, + 0.002821501810103655, + 0.0015438908012583852, + 0.002462926786392927, + 0.004918224178254604, + 0.0067489962093532085, + 0.00716417608782649, + 0.004086794797331095, + 0.004576570354402065, + 0.005817456636577845, + 0.002963755512610078, + 0.004074296448379755, + 0.003702020039781928, + 0.0073768701404333115, + 0.005582055542618036, + 0.004009226802736521, + 0.004288914147764444, + 0.003968643490225077, + 0.005839632824063301, + 0.005713929422199726, + 0.005996500141918659, + 0.010825165547430515, + 0.007828467525541782, + 0.0066044144332408905, + 0.008409800007939339, + 0.009229076094925404, + 0.004996816627681255, + 0.006622431334108114, + 0.009654752910137177, + 0.005330453161150217, + 0.014783737249672413, + 0.009709765203297138, + 0.004576580598950386, + 0.009695829823613167, + 0.007881876081228256, + 0.01317093800753355, + 0.05242149159312248, + 0.06319060176610947, + 0.11502614617347717, + 0.05974360927939415, + 0.08001455664634705, + 0.05727018415927887, + 0.07087115943431854, + 0.06443919241428375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.007640897762030363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190113827586174 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018190113827586174, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018190113827586174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02910708449780941, + 0.03725987672805786, + 0.04137749597430229, + 0.11765970289707184, + 0.049174997955560684, + 0.16674606502056122, + 0.22574877738952637, + 0.32321876287460327, + 0.0071731992065906525, + 0.0025340926367789507 + ], + "credit_map": [ + 0.004378713201731443, + 0.0027901227585971355, + 0.004314148798584938, + 0.004215449094772339, + 0.0019137674244120717, + 0.0053398790769279, + 0.003118641208857298, + 0.0030363621190190315, + 0.005794727709144354, + 0.0032090451568365097, + 0.00472201406955719, + 0.004682514350861311, + 0.004561135545372963, + 0.004527036566287279, + 0.005642073694616556, + 0.0041213259100914, + 0.007674434222280979, + 0.005088907200843096, + 0.004879110027104616, + 0.005203490145504475, + 0.005504282657057047, + 0.004688239190727472, + 0.004223044496029615, + 0.004115984309464693, + 0.02925269864499569, + 0.013704291544854641, + 0.010966906324028969, + 0.010488774627447128, + 0.012925639748573303, + 0.007731238845735788, + 0.005424438044428825, + 0.027165710926055908, + 0.013735398650169373, + 0.00782476831227541, + 0.007032633293420076, + 0.003614111104980111, + 0.004944006912410259, + 0.00477852625772357, + 0.0033679045736789703, + 0.003877646056935191, + 0.019078634679317474, + 0.027932418510317802, + 0.04605462774634361, + 0.014513172209262848, + 0.013374640606343746, + 0.017620258033275604, + 0.012762675061821938, + 0.015409626998007298, + 0.01961636357009411, + 0.04462341591715813, + 0.03182365745306015, + 0.020450206473469734, + 0.02340896613895893, + 0.021796803921461105, + 0.03081481158733368, + 0.033214520663022995, + 0.030217424035072327, + 0.057122934609651566, + 0.04498477652668953, + 0.034330811351537704, + 0.043570730835199356, + 0.04926711693406105, + 0.029057029634714127, + 0.03466792404651642, + 0.0007580856326967478, + 0.0006560571491718292, + 0.0012581817572936416, + 0.0011350448476150632, + 0.000519916822668165, + 0.0011208511423319578, + 0.0009674842585809529, + 0.0007575771887786686, + 0.0002536518150009215, + 0.0002121574361808598, + 0.0007452518912032247, + 0.00026860504294745624, + 0.00025030074175447226, + 0.00028704115538857877, + 0.00020269946253392845, + 0.0003143849316984415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.018190113827586174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031939942855387926 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.0031939942855387926, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031939942855387926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10589402168989182, + 0.07241240888834, + 0.0938555970788002, + 0.09105291217565536, + 0.12264968454837799, + 0.0557563342154026, + 0.08360982686281204, + 0.11557845771312714, + 0.09439829736948013, + 0.16479246318340302 + ], + "credit_map": [ + 0.01574064791202545, + 0.01073592808097601, + 0.01213053334504366, + 0.01277703046798706, + 0.0088084377348423, + 0.025225209072232246, + 0.010996940545737743, + 0.009479292668402195, + 0.009673490189015865, + 0.007789051625877619, + 0.008647991344332695, + 0.008260182105004787, + 0.009440441615879536, + 0.010292531922459602, + 0.011610628105700016, + 0.0066980901174247265, + 0.011943548917770386, + 0.009105219505727291, + 0.008229849860072136, + 0.014075640588998795, + 0.014295482076704502, + 0.009057451970875263, + 0.01960909552872181, + 0.007539312355220318, + 0.03108433075249195, + 0.008033515885472298, + 0.00665217824280262, + 0.005418213550001383, + 0.02185550518333912, + 0.004773393739014864, + 0.004201298579573631, + 0.009034475311636925, + 0.03489796817302704, + 0.01676689088344574, + 0.01580190099775791, + 0.009207300841808319, + 0.011094645597040653, + 0.01478557102382183, + 0.008687133900821209, + 0.011408266611397266, + 0.006975208874791861, + 0.008194751106202602, + 0.013390576466917992, + 0.005175343248993158, + 0.004789181984961033, + 0.005461737979203463, + 0.005635612178593874, + 0.006133923772722483, + 0.007391506340354681, + 0.01994171366095543, + 0.015321348793804646, + 0.006920807063579559, + 0.007746923249214888, + 0.006961224600672722, + 0.008959336206316948, + 0.010366959497332573, + 0.010752803646028042, + 0.02094912715256214, + 0.01573370210826397, + 0.012317994609475136, + 0.016696244478225708, + 0.017917674034833908, + 0.009389174170792103, + 0.011821743100881577, + 0.01162207592278719, + 0.00743777584284544, + 0.01403635460883379, + 0.0154165243729949, + 0.00707081938162446, + 0.014769782312214375, + 0.013581043109297752, + 0.010463927872478962, + 0.015579833649098873, + 0.01783687062561512, + 0.03432871773838997, + 0.01749589294195175, + 0.02292245253920555, + 0.016817057505249977, + 0.020584311336278915, + 0.019227318465709686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.0031939942855387926 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933703422546387 + ], + "opposite_logprob_tokens": [ + -0.6933703422546387 + ], + "logprob_reward": -0.6933703422546387, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933703422546387, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.055948738008737564, + 0.05328386649489403, + 0.04990136995911598, + 0.08788321912288666, + 0.1874743402004242, + 0.1157282143831253, + 0.02162587083876133, + 0.03480934351682663, + 0.12793029844760895, + 0.2654147148132324 + ], + "credit_map": [ + 0.009167138487100601, + 0.005248001776635647, + 0.0056012640707194805, + 0.006043692119419575, + 0.005622474476695061, + 0.013048610650002956, + 0.005372247193008661, + 0.005845310166478157, + 0.008134029805660248, + 0.004810994490981102, + 0.008298732340335846, + 0.0064455000683665276, + 0.006669326685369015, + 0.0075093223713338375, + 0.005733260419219732, + 0.005682696588337421, + 0.011910936795175076, + 0.005276914220303297, + 0.006168383173644543, + 0.0030562072061002254, + 0.005136242602020502, + 0.005793705116957426, + 0.0036908306647092104, + 0.008868145756423473, + 0.015307721681892872, + 0.01346750557422638, + 0.008187534287571907, + 0.013372926041483879, + 0.010847320780158043, + 0.008998097851872444, + 0.010565551929175854, + 0.007136569358408451, + 0.05343935266137123, + 0.025612328201532364, + 0.02551770769059658, + 0.0146713238209486, + 0.014967378228902817, + 0.021402355283498764, + 0.01400624867528677, + 0.017857642844319344, + 0.017141543328762054, + 0.014851763844490051, + 0.0131145678460598, + 0.011415194720029831, + 0.011701335199177265, + 0.02488078735768795, + 0.008836762979626656, + 0.013786262832581997, + 0.0016192573821172118, + 0.004953847266733646, + 0.003871927736327052, + 0.0017389209242537618, + 0.0024180226027965546, + 0.0024733569007366896, + 0.0023561466950923204, + 0.002194392029196024, + 0.005193246062844992, + 0.0036204892676323652, + 0.004683859646320343, + 0.0031269388273358345, + 0.003762921318411827, + 0.004665594082325697, + 0.0033451004419475794, + 0.006411194335669279, + 0.01602649688720703, + 0.012278190813958645, + 0.028319189324975014, + 0.01450780313462019, + 0.00921835470944643, + 0.01866042986512184, + 0.00928915198892355, + 0.019630679860711098, + 0.027927331626415253, + 0.02122991345822811, + 0.06883541494607925, + 0.020864544436335564, + 0.033346664160490036, + 0.039724286645650864, + 0.025130849331617355, + 0.028355687856674194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.6933703422546387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, high-angle, extreme close-up captures an empty, unbranded aluminum soda can resting on a rough, dark gray asphalt surface. The midday sun glints off its silvery, cylindrical form, revealing faint marks of wear and an open pull tab. From the top of the frame, the massive black rubber tread of a heavy vehicle's tire slowly and deliberately rolls into view, its deep grooves filled with dust. As the tire makes contact, immense downward pressure is instantly applied, and the can squirms violently, its smooth sides wrinkling and folding inward with a sharp, metallic creak. The camera remains steady as the weight relentlessly compresses the structure, forcing the aluminum to scream and twist as it is systematically squeezed, flattening into a contorted shape against the hard ground. The tire rolls completely past, revealing the can's final form: a wrinkled, flattened, and shattered disc of metal, its original shape obliterated, now a mere silvery fragment pressed against the dark, richly textured asphalt, casting a small, sharp shadow in the bright sunlight.", + "seed": 30641, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_6454_seed_30641_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2034510691417381, + "vlm_logprob": -0.2034510691417381, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.2034510691417381, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02373485080897808 + ], + "opposite_logprob_tokens": [ + -4.148735046386719 + ], + "logprob_reward": -0.02373485080897808, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02373485080897808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23144792020320892, + 0.07982231676578522, + 0.07788834720849991, + 0.20492865145206451, + 0.034686747938394547, + 0.07972020655870438, + 0.05417429655790329, + 0.14940057694911957, + 0.04246916249394417, + 0.045461781322956085 + ], + "credit_map": [ + 0.029420889914035797, + 0.030667265877127647, + 0.029277002438902855, + 0.04355272278189659, + 0.01723996177315712, + 0.027505632489919662, + 0.020634520798921585, + 0.03314993903040886, + 0.012654837220907211, + 0.008219987154006958, + 0.011294377036392689, + 0.01525916438549757, + 0.006776335649192333, + 0.007886751554906368, + 0.008045420050621033, + 0.009685449302196503, + 0.007767477072775364, + 0.010276333428919315, + 0.010764931328594685, + 0.008966140449047089, + 0.012232561595737934, + 0.00833356473594904, + 0.009322524070739746, + 0.01022481918334961, + 0.019862333312630653, + 0.018911141902208328, + 0.0206125620752573, + 0.023803023621439934, + 0.021846460178494453, + 0.022648923099040985, + 0.02633048966526985, + 0.05091370642185211, + 0.0017428669380024076, + 0.00306614744476974, + 0.0023438262287527323, + 0.010237145237624645, + 0.007086573634296656, + 0.002691177884116769, + 0.003665882395580411, + 0.0038531250320374966, + 0.011168702505528927, + 0.013800718821585178, + 0.009485458955168724, + 0.009100006893277168, + 0.01078774780035019, + 0.0116953793913126, + 0.008140474557876587, + 0.0055417148396372795, + 0.005533879157155752, + 0.009033342823386192, + 0.007589146960526705, + 0.00497493427246809, + 0.009244020096957684, + 0.006362643092870712, + 0.004535758402198553, + 0.00690056849271059, + 0.012715895660221577, + 0.044402461498975754, + 0.018888548016548157, + 0.007636625319719315, + 0.037417154759168625, + 0.010027917101979256, + 0.00961190927773714, + 0.008700070902705193, + 0.006006069481372833, + 0.00596644077450037, + 0.004213735461235046, + 0.004621536936610937, + 0.006524885073304176, + 0.006397897377610207, + 0.004710185807198286, + 0.004028411116451025, + 0.003648921847343445, + 0.0057822479866445065, + 0.005515242461115122, + 0.006677465513348579, + 0.007862021215260029, + 0.0047135259956121445, + 0.005526388995349407, + 0.00573596591129899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.02373485080897808 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652104064822197 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.008652104064822197, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652104064822197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1452282965183258, + 0.04664567485451698, + 0.06392551958560944, + 0.19430245459079742, + 0.2717876732349396, + 0.06527398526668549, + 0.06195329502224922, + 0.08043540269136429, + 0.028822580352425575, + 0.04162511229515076 + ], + "credit_map": [ + 0.0324745699763298, + 0.01782255247235298, + 0.01989704929292202, + 0.02492700330913067, + 0.009029004722833633, + 0.011128386482596397, + 0.010980413295328617, + 0.01896931789815426, + 0.005292590241879225, + 0.0066540902480483055, + 0.006326072849333286, + 0.007813426665961742, + 0.0036446081940084696, + 0.005370176397264004, + 0.005266249645501375, + 0.0062784659676253796, + 0.008571230806410313, + 0.007662896998226643, + 0.00938121322542429, + 0.011345229111611843, + 0.006567748263478279, + 0.005464029032737017, + 0.0055589391849935055, + 0.009374228306114674, + 0.0158708356320858, + 0.012404261156916618, + 0.01625462807714939, + 0.027344344183802605, + 0.013557719066739082, + 0.017594166100025177, + 0.03383002057671547, + 0.05744647979736328, + 0.010164221748709679, + 0.01554363127797842, + 0.009821251034736633, + 0.08961942791938782, + 0.09613976627588272, + 0.021921394392848015, + 0.01769392006099224, + 0.01088405679911375, + 0.009675317443907261, + 0.012900423258543015, + 0.009589772671461105, + 0.005016731098294258, + 0.007688683923333883, + 0.009649972431361675, + 0.0056566339917480946, + 0.005096453242003918, + 0.006051513832062483, + 0.012368443422019482, + 0.0076191299594938755, + 0.0038133196067065, + 0.010140227153897285, + 0.008668815717101097, + 0.0035827304236590862, + 0.009709116071462631, + 0.009295879863202572, + 0.022164681926369667, + 0.0062483507208526134, + 0.005391353275626898, + 0.01894753985106945, + 0.008071579970419407, + 0.004869068041443825, + 0.00544694671407342, + 0.003941751550883055, + 0.004827486351132393, + 0.002957028802484274, + 0.003890895051881671, + 0.0038586861919611692, + 0.0033872078638523817, + 0.002887974726036191, + 0.003071550512686372, + 0.0033032631035894156, + 0.005357012152671814, + 0.004647816997021437, + 0.004853262100368738, + 0.00864064134657383, + 0.006513467989861965, + 0.004215124528855085, + 0.00409452011808753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008652104064822197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644185960292816 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644185960292816, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644185960292816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054222796112298965, + 0.02920352853834629, + 0.034001294523477554, + 0.24071691930294037, + 0.19312727451324463, + 0.12828850746154785, + 0.047507334500551224, + 0.24012775719165802, + 0.021858196705579758, + 0.010946317575871944 + ], + "credit_map": [ + 0.0077324784360826015, + 0.007118826266378164, + 0.007490141317248344, + 0.010512921959161758, + 0.003255232237279415, + 0.003933200147002935, + 0.007304695900529623, + 0.006875309161841869, + 0.004138170275837183, + 0.0038709943182766438, + 0.003970890771597624, + 0.005345134995877743, + 0.0027697323821485043, + 0.002796715823933482, + 0.003446592018008232, + 0.0028653028421103954, + 0.003350586164742708, + 0.004537091124802828, + 0.004458896350115538, + 0.005927129182964563, + 0.00389684340916574, + 0.003937417175620794, + 0.0040521579794585705, + 0.0038411777932196856, + 0.016501471400260925, + 0.01222066581249237, + 0.01479351706802845, + 0.034861970692873, + 0.011376602575182915, + 0.019302722066640854, + 0.03826317563652992, + 0.09339682012796402, + 0.007436077576130629, + 0.008575479499995708, + 0.009844666346907616, + 0.0707283616065979, + 0.057952579110860825, + 0.014466860331594944, + 0.01190310437232256, + 0.012220178730785847, + 0.01065520104020834, + 0.019082948565483093, + 0.01676323264837265, + 0.013466820120811462, + 0.02013307437300682, + 0.01974550262093544, + 0.01439161691814661, + 0.014050140976905823, + 0.005540510173887014, + 0.00760142644867301, + 0.00621851347386837, + 0.0037409556098282337, + 0.007194377947598696, + 0.007020141463726759, + 0.00426649022847414, + 0.0059249261394143105, + 0.011972939595580101, + 0.09191790968179703, + 0.02643701620399952, + 0.007051179185509682, + 0.0757058709859848, + 0.009719165973365307, + 0.010773245245218277, + 0.006550466641783714, + 0.002828003605827689, + 0.003350471379235387, + 0.0021004651207476854, + 0.002424937905743718, + 0.003038995200768113, + 0.003159653628244996, + 0.0020301330368965864, + 0.00292553985491395, + 0.0010326006449759007, + 0.001999289495870471, + 0.001389472046867013, + 0.0014042808907106519, + 0.001635044813156128, + 0.0014788171974942088, + 0.0011750427074730396, + 0.000831771525554359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.008644185960292816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019418209558352828 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019418209558352828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019418209558352828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07879536598920822, + 0.059723105281591415, + 0.24653688073158264, + 0.18468032777309418, + 0.0768638327717781, + 0.06752273440361023, + 0.08048901706933975, + 0.06656930595636368, + 0.04803832247853279, + 0.090781070291996 + ], + "credit_map": [ + 0.01490264106541872, + 0.009083331562578678, + 0.010473390109837055, + 0.01339022908359766, + 0.005919397342950106, + 0.005249978508800268, + 0.009277204982936382, + 0.010499191470444202, + 0.011845534667372704, + 0.005429637152701616, + 0.007444129791110754, + 0.011142675764858723, + 0.0045143598690629005, + 0.005832910072058439, + 0.007850706577301025, + 0.005663149058818817, + 0.023616578429937363, + 0.02825269289314747, + 0.027451539412140846, + 0.043997202068567276, + 0.04143776744604111, + 0.023474588990211487, + 0.026881247758865356, + 0.03142527490854263, + 0.012616838328540325, + 0.012934726662933826, + 0.01657046377658844, + 0.02716692164540291, + 0.012641838751733303, + 0.019446276128292084, + 0.03213455528020859, + 0.051168713718652725, + 0.0034980480559170246, + 0.00447005545720458, + 0.004515249282121658, + 0.02034907601773739, + 0.021022576838731766, + 0.008331299759447575, + 0.009238156490027905, + 0.005439369473606348, + 0.005972668528556824, + 0.007938350550830364, + 0.00761687196791172, + 0.01018806267529726, + 0.009146532975137234, + 0.009970557875931263, + 0.008715247735381126, + 0.00797443836927414, + 0.007857910357415676, + 0.01648884452879429, + 0.008218608796596527, + 0.0056748660281300545, + 0.014224747195839882, + 0.013485847972333431, + 0.006149346474558115, + 0.00838884711265564, + 0.006483270321041346, + 0.018509503453969955, + 0.007266269065439701, + 0.005336957052350044, + 0.015279563143849373, + 0.004117659293115139, + 0.004457851406186819, + 0.005118239670991898, + 0.005852936767041683, + 0.006348901428282261, + 0.005393665749579668, + 0.0064005679450929165, + 0.006957503966987133, + 0.005897806026041508, + 0.005367391277104616, + 0.005819548852741718, + 0.006075629033148289, + 0.012756265699863434, + 0.011943794786930084, + 0.01377043779939413, + 0.017726508900523186, + 0.010166715830564499, + 0.01015431247651577, + 0.00818740762770176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.0019418209558352828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742823839187622 + ], + "opposite_logprob_tokens": [ + -0.4742823541164398 + ], + "logprob_reward": -0.9742823839187622, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742823839187622, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0584186352789402, + 0.06745108217000961, + 0.16079562902450562, + 0.07593602687120438, + 0.07888871431350708, + 0.041393015533685684, + 0.03784222528338432, + 0.3847584128379822, + 0.024697476997971535, + 0.06981880217790604 + ], + "credit_map": [ + 0.010727937333285809, + 0.007463804446160793, + 0.008167536929249763, + 0.01005710382014513, + 0.004033653065562248, + 0.004990710411220789, + 0.004572527017444372, + 0.008405356667935848, + 0.010164059698581696, + 0.009993324987590313, + 0.009869573637843132, + 0.011632881127297878, + 0.006656830199062824, + 0.006802328862249851, + 0.007383398711681366, + 0.004948682617396116, + 0.0158724132925272, + 0.017030222341418266, + 0.02051982283592224, + 0.02613672986626625, + 0.029740700498223305, + 0.015729526057839394, + 0.015210984274744987, + 0.020555226132273674, + 0.008229195140302181, + 0.006124970503151417, + 0.006745876744389534, + 0.010501177050173283, + 0.005563892889767885, + 0.00605497881770134, + 0.009971641935408115, + 0.022744297981262207, + 0.0033566548954695463, + 0.0059091467410326, + 0.004970476496964693, + 0.02426019124686718, + 0.019818240776658058, + 0.007579883094877005, + 0.007736811414361, + 0.005257311277091503, + 0.005705450661480427, + 0.005111482460051775, + 0.003903967095538974, + 0.004952150862663984, + 0.00553229171782732, + 0.005214752163738012, + 0.0045549385249614716, + 0.006417982745915651, + 0.0040620299987494946, + 0.007980834692716599, + 0.002601999556645751, + 0.002402396872639656, + 0.006654899101704359, + 0.005524231120944023, + 0.0032691466622054577, + 0.005346686579287052, + 0.028505543246865273, + 0.13958397507667542, + 0.044444676488637924, + 0.011219749227166176, + 0.11492870002985, + 0.014335156418383121, + 0.01934128627181053, + 0.012399289757013321, + 0.002612380078062415, + 0.0030603474006056786, + 0.0024466312024742365, + 0.003026894060894847, + 0.0038793538697063923, + 0.0038423235528171062, + 0.0030507901683449745, + 0.0027787566650658846, + 0.009916190057992935, + 0.00878337025642395, + 0.009019646793603897, + 0.012110504321753979, + 0.0078116110526025295, + 0.005773127544671297, + 0.007979877293109894, + 0.008424470201134682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.9742823839187622 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, high-angle, extreme close-up captures an empty, unbranded aluminum soda can resting on a rough, dark gray asphalt surface. The midday sun glints off its silvery, cylindrical form, revealing faint marks of wear and an open pull tab. From the top of the frame, the massive black rubber tread of a heavy vehicle's tire slowly and deliberately rolls into view, its deep grooves filled with dust. As the tire makes contact, immense downward pressure is instantly applied, and the can squirms violently, its smooth sides wrinkling and folding inward with a sharp, metallic creak. The camera remains steady as the weight relentlessly compresses the structure, forcing the aluminum to scream and twist as it is systematically squeezed, flattening into a contorted shape against the hard ground. The tire rolls completely past, revealing the can's final form: a wrinkled, flattened, and shattered disc of metal, its original shape obliterated, now a mere silvery fragment pressed against the dark, richly textured asphalt, casting a small, sharp shadow in the bright sunlight.", + "seed": 30642, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_6454_seed_30642_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1560152147896588, + "vlm_logprob": -0.1560152147896588, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.1560152147896588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32012197375297546 + ], + "opposite_logprob_tokens": [ + -1.3201220035552979 + ], + "logprob_reward": -0.32012197375297546, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32012197375297546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11352363228797913, + 0.329059898853302, + 0.057522714138031006, + 0.07453341037034988, + 0.03342103585600853, + 0.03929703310132027, + 0.049601972103118896, + 0.06712768971920013, + 0.10083829611539841, + 0.13507428765296936 + ], + "credit_map": [ + 0.014385996386408806, + 0.015399064868688583, + 0.01524158101528883, + 0.01290010754019022, + 0.012723750434815884, + 0.01011042483150959, + 0.018388286232948303, + 0.014374420046806335, + 0.041519321501255035, + 0.07648026198148727, + 0.1172567829489708, + 0.02954653836786747, + 0.019678274169564247, + 0.018673023208975792, + 0.014848470687866211, + 0.011057235300540924, + 0.008301297202706337, + 0.008480827324092388, + 0.0069551775231957436, + 0.0065119583159685135, + 0.006102106533944607, + 0.009554893709719181, + 0.004658290185034275, + 0.00695816008374095, + 0.010061575099825859, + 0.011072017252445221, + 0.009558950550854206, + 0.009531889110803604, + 0.014619804918766022, + 0.006163809914141893, + 0.0068298084661364555, + 0.006695553194731474, + 0.0032271489035338163, + 0.004845521878451109, + 0.0031918720342218876, + 0.003247353481128812, + 0.006814838852733374, + 0.004809515085071325, + 0.0037120827473700047, + 0.003572705201804638, + 0.005310372915118933, + 0.003952053375542164, + 0.007099778391420841, + 0.003867378458380699, + 0.004601598251610994, + 0.003981621935963631, + 0.004353570286184549, + 0.006130660884082317, + 0.0073871007189154625, + 0.007277744822204113, + 0.009726217947900295, + 0.00622494937852025, + 0.006152817979454994, + 0.004920843057334423, + 0.003247033804655075, + 0.004665259271860123, + 0.011002644896507263, + 0.009905495680868626, + 0.008616750128567219, + 0.007631656248122454, + 0.00841179396957159, + 0.008948960341513157, + 0.00564937898889184, + 0.006961015518754721, + 0.014422427862882614, + 0.014813126064836979, + 0.010395103134214878, + 0.011649017222225666, + 0.020596496760845184, + 0.011012427508831024, + 0.01044539175927639, + 0.007504303008317947, + 0.022958602756261826, + 0.021479111164808273, + 0.015342501923441887, + 0.008909231051802635, + 0.017484022304415703, + 0.027320463210344315, + 0.009045790880918503, + 0.012534567154943943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.32012197375297546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554189190268517 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012554189190268517, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554189190268517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19857770204544067, + 0.21331171691417694, + 0.08503890782594681, + 0.022653304040431976, + 0.030267609283328056, + 0.03778528422117233, + 0.11013814061880112, + 0.03864928334951401, + 0.07179053127765656, + 0.1917874813079834 + ], + "credit_map": [ + 0.026904642581939697, + 0.02901338040828705, + 0.026739001274108887, + 0.023244457319378853, + 0.02826518565416336, + 0.017650658264756203, + 0.02548365853726864, + 0.021276723593473434, + 0.02581978589296341, + 0.027694035321474075, + 0.024092556908726692, + 0.028675934299826622, + 0.02364363707602024, + 0.016969583928585052, + 0.04370783269405365, + 0.02270834892988205, + 0.012870335951447487, + 0.012658759020268917, + 0.009465363807976246, + 0.010339726693928242, + 0.009883434511721134, + 0.010319072753190994, + 0.0074584889225661755, + 0.012043722905218601, + 0.003512441413477063, + 0.003461439860984683, + 0.0022233070340007544, + 0.0021754319313913584, + 0.004929955583065748, + 0.0022494480945169926, + 0.0019944573286920786, + 0.0021068218629807234, + 0.004103274550288916, + 0.004075895994901657, + 0.0022343595046550035, + 0.0027060064021497965, + 0.006980754900723696, + 0.0036602404434233904, + 0.0036575302947312593, + 0.0028495481237769127, + 0.005109432153403759, + 0.003197798505425453, + 0.007754836697131395, + 0.0041066547855734825, + 0.0052789356559515, + 0.004134548362344503, + 0.004244891926646233, + 0.003958184737712145, + 0.019137324765324593, + 0.016250312328338623, + 0.030049139633774757, + 0.013476403430104256, + 0.007538496050983667, + 0.007922508753836155, + 0.005736800841987133, + 0.01002715714275837, + 0.006014701910316944, + 0.005031167063862085, + 0.00553507125005126, + 0.004759076051414013, + 0.005337214563041925, + 0.003720876993611455, + 0.0034952033311128616, + 0.004755974281579256, + 0.009710812009871006, + 0.01018540933728218, + 0.007810244336724281, + 0.010522241704165936, + 0.014541074633598328, + 0.007499178405851126, + 0.006016280502080917, + 0.005505293142050505, + 0.06377538293600082, + 0.05566989257931709, + 0.018037624657154083, + 0.014565943740308285, + 0.01036370825022459, + 0.01129971444606781, + 0.008560063317418098, + 0.009515147656202316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.012554189190268517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545948848128319 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545948848128319, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545948848128319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3156721591949463, + 0.1302458941936493, + 0.13130611181259155, + 0.053451068699359894, + 0.024755161255598068, + 0.07512548565864563, + 0.03153176233172417, + 0.027407033368945122, + 0.13631339371204376, + 0.0741918534040451 + ], + "credit_map": [ + 0.03392302244901657, + 0.04897362366318703, + 0.04418617859482765, + 0.03902939707040787, + 0.043452661484479904, + 0.028785182163119316, + 0.039135631173849106, + 0.03818647935986519, + 0.023724686354398727, + 0.024981629103422165, + 0.022046979516744614, + 0.01176268421113491, + 0.009078187867999077, + 0.018855702131986618, + 0.009989889338612556, + 0.009806147776544094, + 0.02155127376317978, + 0.01733580231666565, + 0.014272268861532211, + 0.016193877905607224, + 0.013826279900968075, + 0.02063172683119774, + 0.009756959043443203, + 0.01773795112967491, + 0.008674516342580318, + 0.010194055736064911, + 0.00557590601965785, + 0.005235010292381048, + 0.0089143430814147, + 0.005517180077731609, + 0.004576180130243301, + 0.004763883538544178, + 0.0034336233511567116, + 0.0036545880138874054, + 0.002658034674823284, + 0.0021630411501973867, + 0.005277771502733231, + 0.002520474372431636, + 0.0028324744198471308, + 0.002215156564489007, + 0.009089439176023006, + 0.008648972027003765, + 0.012261267751455307, + 0.006772659253329039, + 0.010339060798287392, + 0.009069160558283329, + 0.0085750138387084, + 0.010369924828410149, + 0.004405585117638111, + 0.006282276473939419, + 0.005787197034806013, + 0.0041282339952886105, + 0.003083533374592662, + 0.002558537060394883, + 0.0017603366868570447, + 0.003526062937453389, + 0.004036077298223972, + 0.0029930814635008574, + 0.00397794833406806, + 0.0029965678695589304, + 0.004271695390343666, + 0.002731131389737129, + 0.0029848599806427956, + 0.0034156746696680784, + 0.022820433601737022, + 0.02358374372124672, + 0.0129745714366436, + 0.014673206023871899, + 0.028047407045960426, + 0.01454805675894022, + 0.010757356882095337, + 0.008908630348742008, + 0.017758971080183983, + 0.014421803876757622, + 0.007724075112491846, + 0.008984834887087345, + 0.006732147186994553, + 0.007713397033512592, + 0.005486030597239733, + 0.005370600614696741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.012545948848128319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17441734671592712, + 0.08971042931079865, + 0.24994555115699768, + 0.06000952795147896, + 0.04606732726097107, + 0.08902699500322342, + 0.03875109180808067, + 0.06019585207104683, + 0.11298679560422897, + 0.07888907194137573 + ], + "credit_map": [ + 0.019700851291418076, + 0.026135273277759552, + 0.02286512963473797, + 0.02318538911640644, + 0.018520258367061615, + 0.021014545112848282, + 0.015407150611281395, + 0.027588756754994392, + 0.014148857444524765, + 0.018066462129354477, + 0.010382265783846378, + 0.008420112542808056, + 0.008819595910608768, + 0.011990317143499851, + 0.009724849835038185, + 0.008157971315085888, + 0.03040662594139576, + 0.034985560923814774, + 0.02393503673374653, + 0.023520098999142647, + 0.030923493206501007, + 0.04479096457362175, + 0.023528151214122772, + 0.037855617702007294, + 0.008869970217347145, + 0.009421875700354576, + 0.007963410578668118, + 0.006680741440504789, + 0.010460052639245987, + 0.005046809557825327, + 0.006243697367608547, + 0.005322964861989021, + 0.005519286263734102, + 0.005691196769475937, + 0.004285405855625868, + 0.005022370722144842, + 0.0075834039598703384, + 0.006203057710081339, + 0.005304434802383184, + 0.006458168383687735, + 0.011849982663989067, + 0.008854840882122517, + 0.014168087393045425, + 0.01001618430018425, + 0.013351544737815857, + 0.010717042721807957, + 0.009078106842935085, + 0.010991203598678112, + 0.006470778025686741, + 0.005841465666890144, + 0.007593600079417229, + 0.0038223862648010254, + 0.004585764370858669, + 0.0038985006976872683, + 0.0023417090997099876, + 0.004196887835860252, + 0.01102873682975769, + 0.007618661038577557, + 0.006646940018981695, + 0.007181521970778704, + 0.00782082136720419, + 0.008247299119830132, + 0.005443226080387831, + 0.006208641454577446, + 0.014065652154386044, + 0.01788513734936714, + 0.00833433959633112, + 0.011947821825742722, + 0.028546256944537163, + 0.014457550831139088, + 0.009374813176691532, + 0.008375225588679314, + 0.014611450955271721, + 0.015386833809316158, + 0.011558080092072487, + 0.008343375287950039, + 0.007339805364608765, + 0.010635806247591972, + 0.004664641339331865, + 0.006349081639200449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890313267707825 + ], + "opposite_logprob_tokens": [ + -1.0539031028747559 + ], + "logprob_reward": -0.42890313267707825, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890313267707825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2402210235595703, + 0.2606510519981384, + 0.07546965032815933, + 0.08095130324363708, + 0.07176917046308517, + 0.03527950868010521, + 0.03582588955760002, + 0.07487480342388153, + 0.08685087412595749, + 0.038106705993413925 + ], + "credit_map": [ + 0.03567478060722351, + 0.03576422482728958, + 0.028940515592694283, + 0.02300349995493889, + 0.03302447870373726, + 0.025807736441493034, + 0.02820548415184021, + 0.029800277203321457, + 0.046794407069683075, + 0.08506524562835693, + 0.050974830985069275, + 0.010981380939483643, + 0.01725146360695362, + 0.02296162210404873, + 0.012457510456442833, + 0.01416458748281002, + 0.009590938687324524, + 0.010541809722781181, + 0.010020971298217773, + 0.00885800737887621, + 0.009877013973891735, + 0.010959895327687263, + 0.005553889088332653, + 0.010067122988402843, + 0.011528450064361095, + 0.013422392308712006, + 0.00764913996681571, + 0.006673740688711405, + 0.019168980419635773, + 0.007029784377664328, + 0.009216202422976494, + 0.006262610666453838, + 0.008211687207221985, + 0.010350394994020462, + 0.00795054528862238, + 0.006780157797038555, + 0.014616499654948711, + 0.011041345074772835, + 0.006808861158788204, + 0.006009677425026894, + 0.005922865122556686, + 0.004006272181868553, + 0.004751896485686302, + 0.003831068053841591, + 0.006935955956578255, + 0.0044035823084414005, + 0.0025257470551878214, + 0.0029021159280091524, + 0.005749008152633905, + 0.0057108039036393166, + 0.007489573210477829, + 0.004409781191498041, + 0.0022994577884674072, + 0.00367160071618855, + 0.0025521942880004644, + 0.003943467978388071, + 0.013812657445669174, + 0.009669249877333641, + 0.012986225076019764, + 0.011039717122912407, + 0.006995207164436579, + 0.007157548796385527, + 0.006870124954730272, + 0.006344062741845846, + 0.013370397500693798, + 0.013245692476630211, + 0.01153342891484499, + 0.013929239474236965, + 0.012903321534395218, + 0.008596986532211304, + 0.006302534136921167, + 0.0069692619144916534, + 0.007704766001552343, + 0.007641688920557499, + 0.002875291043892503, + 0.005126219242811203, + 0.004553398583084345, + 0.0033696084283292294, + 0.003300559939816594, + 0.0035351726692169905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.42890313267707825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up captures an empty, unbranded aluminum soda can resting on a rough, dark gray asphalt surface. The midday sun glints off its silvery, cylindrical form, revealing faint marks of wear and an open pull tab. From the top of the frame, the massive black rubber tread of a heavy vehicle's tire slowly and deliberately rolls into view, its deep grooves filled with dust. As the tire makes contact, immense downward pressure is instantly applied, and the can squirms violently, its smooth sides wrinkling and folding inward with a sharp, metallic creak. The camera remains steady as the weight relentlessly compresses the structure, forcing the aluminum to scream and twist as it is systematically squeezed, flattening into a contorted shape against the hard ground. The tire rolls completely past, revealing the can's final form: a wrinkled, flattened, and shattered disc of metal, its original shape obliterated, now a mere silvery fragment pressed against the dark, richly textured asphalt, casting a small, sharp shadow in the bright sunlight.", + "seed": 30643, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_1/index_6454_seed_30643_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14730587536469103, + "vlm_logprob": -0.14730587536469103, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.14730587536469103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23681068420410156 + ], + "opposite_logprob_tokens": [ + -1.6118106842041016 + ], + "logprob_reward": -0.23681068420410156, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23681068420410156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09240178018808365, + 0.032104622572660446, + 0.09202397614717484, + 0.03767368942499161, + 0.2676314115524292, + 0.03977971524000168, + 0.32083067297935486, + 0.06360036879777908, + 0.04040422663092613, + 0.013549570925533772 + ], + "credit_map": [ + 0.012102792039513588, + 0.007389286067336798, + 0.014777328819036484, + 0.009783327579498291, + 0.009201984852552414, + 0.010979099199175835, + 0.01746237650513649, + 0.01070559024810791, + 0.002822935814037919, + 0.0028911512345075607, + 0.007341715972870588, + 0.00410482008010149, + 0.0020708783995360136, + 0.005750172305852175, + 0.00413935212418437, + 0.0029835933819413185, + 0.008472402580082417, + 0.006553880870342255, + 0.010301580652594566, + 0.005983161274343729, + 0.029213786125183105, + 0.005507039371877909, + 0.007604462094604969, + 0.01838766410946846, + 0.005691491533070803, + 0.0035948671866208315, + 0.002057489473372698, + 0.004861638881266117, + 0.0032559148967266083, + 0.00314216292463243, + 0.008347627706825733, + 0.006722495891153812, + 0.04143725708127022, + 0.022495616227388382, + 0.03952119126915932, + 0.029816925525665283, + 0.021232249215245247, + 0.02671680971980095, + 0.043409623205661774, + 0.04300174117088318, + 0.002446831436827779, + 0.0030184832867234945, + 0.0055465311743319035, + 0.005697040818631649, + 0.007177427411079407, + 0.006869640666991472, + 0.006175360642373562, + 0.00284840096719563, + 0.027490882202982903, + 0.04945078119635582, + 0.06852667778730392, + 0.01804567687213421, + 0.0686730295419693, + 0.02812158316373825, + 0.02373170293867588, + 0.03679033741354942, + 0.010979103855788708, + 0.0035654441453516483, + 0.007946433499455452, + 0.006014769896864891, + 0.00634789140895009, + 0.011134299449622631, + 0.008992310613393784, + 0.008620118722319603, + 0.00479520158842206, + 0.004963407292962074, + 0.005370140541344881, + 0.004476455971598625, + 0.005927919875830412, + 0.003850600915029645, + 0.005461920518428087, + 0.005558576434850693, + 0.001028700964525342, + 0.0011869955342262983, + 0.0020395927131175995, + 0.001969963777810335, + 0.00202743848785758, + 0.0015729960286989808, + 0.0015854280209168792, + 0.0021384556312114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.23681068420410156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362489465624094 + ], + "opposite_logprob_tokens": [ + -5.62862491607666 + ], + "logprob_reward": -0.00362489465624094, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362489465624094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06106525659561157, + 0.044919028878211975, + 0.15515686571598053, + 0.10931136459112167, + 0.1154463067650795, + 0.1810745894908905, + 0.06509272754192352, + 0.1310575008392334, + 0.035243622958660126, + 0.10163278877735138 + ], + "credit_map": [ + 0.008345534093677998, + 0.005531797651201487, + 0.008865690790116787, + 0.008357422426342964, + 0.006162672303617001, + 0.006495427805930376, + 0.010610361583530903, + 0.006696346681565046, + 0.0066217160783708096, + 0.0051345182582736015, + 0.0048460871912539005, + 0.005830197595059872, + 0.0035170698538422585, + 0.006592609453946352, + 0.00527891656383872, + 0.007097912952303886, + 0.010935061611235142, + 0.007954833097755909, + 0.014006220735609531, + 0.00809710193425417, + 0.05458863452076912, + 0.010346478782594204, + 0.013016024604439735, + 0.0362125001847744, + 0.012600831687450409, + 0.00701978849247098, + 0.014845077879726887, + 0.008735045790672302, + 0.008478300645947456, + 0.009619679301977158, + 0.022844407707452774, + 0.025168225169181824, + 0.0076863705180585384, + 0.0049916584976017475, + 0.00904002133756876, + 0.012326134368777275, + 0.010189764201641083, + 0.007240291219204664, + 0.03875426575541496, + 0.025217793881893158, + 0.01166690792888403, + 0.014566699042916298, + 0.020895224064588547, + 0.02719137817621231, + 0.03812101110816002, + 0.03369775414466858, + 0.024591973051428795, + 0.010343623347580433, + 0.006199914962053299, + 0.009641082026064396, + 0.017463073134422302, + 0.003954109735786915, + 0.011196933686733246, + 0.004122626036405563, + 0.005316153168678284, + 0.007198828738182783, + 0.031090691685676575, + 0.007773717865347862, + 0.009314892813563347, + 0.00649002380669117, + 0.01164962351322174, + 0.04477434232831001, + 0.00902477651834488, + 0.010939422063529491, + 0.003490944392979145, + 0.00340523524209857, + 0.005463141016662121, + 0.003455640748143196, + 0.005827413406223059, + 0.003932470455765724, + 0.004536767955869436, + 0.005132006015628576, + 0.011192826554179192, + 0.007695731241255999, + 0.021897681057453156, + 0.007539013866335154, + 0.024467477574944496, + 0.01479120459407568, + 0.006178936921060085, + 0.00786991510540247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.00362489465624094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189528957009315 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189528957009315, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189528957009315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04032761603593826, + 0.027091791853308678, + 0.19198806583881378, + 0.04183816909790039, + 0.13408543169498444, + 0.4976770281791687, + 0.03573158383369446, + 0.00992550328373909, + 0.013881026767194271, + 0.007453791331499815 + ], + "credit_map": [ + 0.004146050661802292, + 0.003562227124348283, + 0.005567007698118687, + 0.0054586175829172134, + 0.0036937007680535316, + 0.004158032592386007, + 0.008551966398954391, + 0.005190013907849789, + 0.0027097004931420088, + 0.002042641630396247, + 0.004268336575478315, + 0.0030891401693224907, + 0.0015952729154378176, + 0.0048911995254457, + 0.004279129672795534, + 0.004216372966766357, + 0.016485316678881645, + 0.011729075573384762, + 0.02660302259027958, + 0.010752391070127487, + 0.06205682083964348, + 0.013634791597723961, + 0.014234068803489208, + 0.03649257868528366, + 0.005818005185574293, + 0.004492684733122587, + 0.006651518400758505, + 0.003988575655966997, + 0.0031469855457544327, + 0.003438152838498354, + 0.006665132474154234, + 0.007637114729732275, + 0.016105428338050842, + 0.00692006666213274, + 0.011108312755823135, + 0.009467175230383873, + 0.012777674943208694, + 0.01046422403305769, + 0.0368955135345459, + 0.030347032472491264, + 0.03286433592438698, + 0.041861649602651596, + 0.05598119646310806, + 0.07071332633495331, + 0.10759004950523376, + 0.08941922336816788, + 0.06709979474544525, + 0.032147444784641266, + 0.003271628636866808, + 0.005457892548292875, + 0.0051574124954640865, + 0.0023937609512358904, + 0.00981213804334402, + 0.0036266872193664312, + 0.0026563096325844526, + 0.0033557522110641003, + 0.0015979636227712035, + 0.0006962972693145275, + 0.0009896359406411648, + 0.0007893359288573265, + 0.0013741225702688098, + 0.0027416450902819633, + 0.0008972608484327793, + 0.0008392430027015507, + 0.0013119328068569303, + 0.0012519690208137035, + 0.0020015467889606953, + 0.0015623884974047542, + 0.0019136619521304965, + 0.0017240096349269152, + 0.0021107622887939215, + 0.0020047558937221766, + 0.0009481446468271315, + 0.0007551844464614987, + 0.0008191994274966419, + 0.000678171927575022, + 0.001535492716357112, + 0.001005337806418538, + 0.0008324632653966546, + 0.000879797269590199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.018189528957009315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036165802739560604 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036165802739560604, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036165802739560604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047074876725673676, + 0.04477421194314957, + 0.17761830985546112, + 0.027646899223327637, + 0.14554846286773682, + 0.1958620548248291, + 0.17060628533363342, + 0.034975986927747726, + 0.07787510007619858, + 0.07801782339811325 + ], + "credit_map": [ + 0.006056039594113827, + 0.0048227449879050255, + 0.005176350008696318, + 0.0077634588815271854, + 0.00375081948004663, + 0.0035646012984216213, + 0.009450720623135567, + 0.0064901430159807205, + 0.004295643884688616, + 0.005584043450653553, + 0.0035632774233818054, + 0.0060068280436098576, + 0.0039085447788238525, + 0.007486569695174694, + 0.0057832165621221066, + 0.008146088570356369, + 0.013996890746057034, + 0.010054283775389194, + 0.01627209596335888, + 0.01133935246616602, + 0.05450214818120003, + 0.013741366565227509, + 0.02144804410636425, + 0.03626413643360138, + 0.004968911875039339, + 0.002829496981576085, + 0.003725439077243209, + 0.003272575791925192, + 0.002033262513577938, + 0.0018146013608202338, + 0.004183069337159395, + 0.00481954263523221, + 0.013565591536462307, + 0.013323536142706871, + 0.022745396941900253, + 0.019593074917793274, + 0.013169558718800545, + 0.01757829450070858, + 0.027066871523857117, + 0.018506143242120743, + 0.012796676717698574, + 0.01595781184732914, + 0.021618349477648735, + 0.028172967955470085, + 0.0426851250231266, + 0.03369973227381706, + 0.028116490691900253, + 0.012814904563128948, + 0.014943143352866173, + 0.023586764931678772, + 0.03101411834359169, + 0.011078312061727047, + 0.04266924411058426, + 0.013212282210588455, + 0.015363839454948902, + 0.01873857155442238, + 0.0062300655990839005, + 0.0023602796718478203, + 0.004196864552795887, + 0.003330441191792488, + 0.004020648077130318, + 0.006371478084474802, + 0.0046460372395813465, + 0.00382017414085567, + 0.006276633124798536, + 0.007221860811114311, + 0.009041011333465576, + 0.009214984253048897, + 0.009983551688492298, + 0.008584115654230118, + 0.015713253989815712, + 0.011839685030281544, + 0.006571121048182249, + 0.00717483414337039, + 0.009182358160614967, + 0.006479295901954174, + 0.019163325428962708, + 0.010777263902127743, + 0.011266179382801056, + 0.007403443101793528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.0036165802739560604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an empty, unbranded aluminum soda can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the can resting on a rough, dark gray asphalt surface with visible dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive black rubber tire with deep grooves roll into the frame from the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can violently squirm, wrinkle, and flatten as the tire applies downward pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and high-angle throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742876887321472 + ], + "opposite_logprob_tokens": [ + -0.9742876887321472 + ], + "logprob_reward": -0.4742876887321472, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742876887321472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035050567239522934, + 0.04075099527835846, + 0.20100995898246765, + 0.0669710636138916, + 0.3270450234413147, + 0.13957014679908752, + 0.07347818464040756, + 0.06481795012950897, + 0.022214388474822044, + 0.02909175679087639 + ], + "credit_map": [ + 0.004039794206619263, + 0.00369002646766603, + 0.005553298629820347, + 0.005708698183298111, + 0.002807329874485731, + 0.0034324501175433397, + 0.005707848351448774, + 0.004111120942980051, + 0.005251685623079538, + 0.003253303002566099, + 0.00906758476048708, + 0.00502051692456007, + 0.00274061830714345, + 0.007102408912032843, + 0.004239413421601057, + 0.004075463395565748, + 0.017924634739756584, + 0.014235316775739193, + 0.018645573407411575, + 0.012546945363283157, + 0.059666428714990616, + 0.01542870793491602, + 0.022399943321943283, + 0.04016239568591118, + 0.008050798438489437, + 0.004510568454861641, + 0.00648321770131588, + 0.0057167187333106995, + 0.005847593769431114, + 0.007111057639122009, + 0.014714546501636505, + 0.014536567963659763, + 0.023780841380357742, + 0.03430837020277977, + 0.04031269997358322, + 0.03673036769032478, + 0.023920563980937004, + 0.03280811756849289, + 0.06644203513860703, + 0.06874202191829681, + 0.00888922531157732, + 0.011167548596858978, + 0.015768030658364296, + 0.020831849426031113, + 0.026868004351854324, + 0.025218147784471512, + 0.02040657214820385, + 0.010420766659080982, + 0.0058636912144720554, + 0.01186368241906166, + 0.019698945805430412, + 0.006871565245091915, + 0.007028812076896429, + 0.007704581134021282, + 0.004322874825447798, + 0.010124028660356998, + 0.0151599096134305, + 0.003934985492378473, + 0.005865261424332857, + 0.0033832842018455267, + 0.005672375205904245, + 0.02154831774532795, + 0.0032416856847703457, + 0.0060121286660432816, + 0.0022209982853382826, + 0.002349246060475707, + 0.0026124382857233286, + 0.0027664126828312874, + 0.003370999824255705, + 0.0024174926802515984, + 0.0031025554053485394, + 0.0033742443192750216, + 0.0033786138519644737, + 0.0033629911486059427, + 0.00397562887519598, + 0.002953103743493557, + 0.005689698737114668, + 0.003942061681300402, + 0.0033835433423519135, + 0.002406115410849452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.4742876887321472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_tensor([3534, 3534])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_tensor([3534, 3534])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6a8b6ab74faaf80ec91c8259a2e2e0f7f6d564b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_tensor([3534, 3534])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, slightly low-angle, medium tracking shot follows a motocross rider as he descends to the right, completing a breathtaking stunt. Clad in a vibrant red and white jersey with bold, abstract graphic elements, black protective pants, and heavy-duty boots, the rider is a figure of profound concentration. His full-face helmet, adorned with a sharp black and white pattern, obscures his expression, but his determined stance and grip on the mud-stained, bright yellow handlebars of his dirt bike convey absolute control. The camera steadily tracks his trajectory as he prepares to descend the steep slope of a large, compact dirt ramp. As the bike's rear tire makes contact, it sends a plume of loose, black soil into the air. The bike's suspension compresses violently under the impact, and the rider, knees bent and body leaning forward, deftly absorbs the shock, guiding the machine steadily and deliberately down the ramp. In the background, the cavernous, packed arena is a swirling vortex of movement and color, with thousands of spectators roaring in the glare of the intense stadium floodlights, which are out of focus and cast sharp, dramatic highlights and deep shadows, creating a high-energy, exhilarating, triumphant atmosphere.", + "seed": 28107, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_3534_seed_28107_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042199179716408255, + "vlm_logprob": -0.042199179716408255, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.042199179716408255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009232801385223866 + ], + "opposite_logprob_tokens": [ + -5.009232997894287 + ], + "logprob_reward": -0.009232801385223866, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009232801385223866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19758199155330658, + 0.0873970314860344, + 0.14459706842899323, + 0.07510971277952194, + 0.07102745026350021, + 0.058354511857032776, + 0.07989417761564255, + 0.0887966901063919, + 0.0836079865694046, + 0.11363334953784943 + ], + "credit_map": [ + 0.016972672194242477, + 0.017580056563019753, + 0.014408612623810768, + 0.011172221042215824, + 0.014377336017787457, + 0.009217468090355396, + 0.06418834626674652, + 0.04966528341174126, + 0.009040636010468006, + 0.011021639220416546, + 0.010037953965365887, + 0.009888147003948689, + 0.009636195376515388, + 0.014507033862173557, + 0.011614836752414703, + 0.01165059395134449, + 0.015043194405734539, + 0.020112359896302223, + 0.02755475603044033, + 0.01454857923090458, + 0.012566342949867249, + 0.01545469369739294, + 0.021546822041273117, + 0.01777034066617489, + 0.008252434432506561, + 0.011544045992195606, + 0.00808377843350172, + 0.017150210216641426, + 0.006726662162691355, + 0.005401955917477608, + 0.010569493286311626, + 0.0073811374604702, + 0.009427840821444988, + 0.008912008255720139, + 0.009700184687972069, + 0.008900478482246399, + 0.007476693484932184, + 0.010396352969110012, + 0.008868968114256859, + 0.007344924379140139, + 0.007622326258569956, + 0.0054292231798172, + 0.01025032252073288, + 0.004385489504784346, + 0.005803669337183237, + 0.013304105959832668, + 0.004668053705245256, + 0.006891328375786543, + 0.012673482298851013, + 0.007655081804841757, + 0.009798911400139332, + 0.00834859162569046, + 0.013785960152745247, + 0.010226006619632244, + 0.00797615759074688, + 0.009429988451302052, + 0.01343668345361948, + 0.008504935540258884, + 0.013221091590821743, + 0.010250833816826344, + 0.012480597011744976, + 0.013659242540597916, + 0.00714063597843051, + 0.01010267622768879, + 0.008209233172237873, + 0.01478111743927002, + 0.012454848736524582, + 0.011827141977846622, + 0.011092379689216614, + 0.007581579964607954, + 0.0076868291944265366, + 0.009974857792258263, + 0.014037242159247398, + 0.010790148749947548, + 0.017538700252771378, + 0.012556671164929867, + 0.011031849309802055, + 0.019343115389347076, + 0.014849253930151463, + 0.013486377894878387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.009232801385223866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329086884856224 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.02329086884856224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329086884856224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17437759041786194, + 0.060782626271247864, + 0.15014195442199707, + 0.0597689226269722, + 0.14859139919281006, + 0.0620807521045208, + 0.03488670289516449, + 0.047482699155807495, + 0.0629463717341423, + 0.19894097745418549 + ], + "credit_map": [ + 0.03433060273528099, + 0.020333293825387955, + 0.018032928928732872, + 0.013037727214396, + 0.017679082229733467, + 0.013139946386218071, + 0.03148917853832245, + 0.02633482776582241, + 0.006943151820451021, + 0.01191968284547329, + 0.006283971015363932, + 0.01049701776355505, + 0.0063009620644152164, + 0.006230308674275875, + 0.00688124867156148, + 0.005726282484829426, + 0.024903537705540657, + 0.01168337743729353, + 0.031210478395223618, + 0.010075800120830536, + 0.010343621484935284, + 0.010027812793850899, + 0.01926676742732525, + 0.032630570232868195, + 0.008098036982119083, + 0.005832416005432606, + 0.007535136770457029, + 0.005945936776697636, + 0.006196431815624237, + 0.007969207130372524, + 0.01111565250903368, + 0.007076104171574116, + 0.020521940663456917, + 0.022200927138328552, + 0.016139380633831024, + 0.016058681532740593, + 0.018508490175008774, + 0.023379791527986526, + 0.015570234507322311, + 0.01621195487678051, + 0.008026137948036194, + 0.0048662275075912476, + 0.01005939207971096, + 0.0067565180361270905, + 0.005105352029204369, + 0.01017020270228386, + 0.008073546923696995, + 0.009023375809192657, + 0.005024665966629982, + 0.0033468599431216717, + 0.0058311717584729195, + 0.004044908564537764, + 0.0034051022958010435, + 0.004186010919511318, + 0.004684643819928169, + 0.004363338928669691, + 0.006438564043492079, + 0.00494996877387166, + 0.006671491079032421, + 0.004806925542652607, + 0.0067895338870584965, + 0.0070194341242313385, + 0.003943684510886669, + 0.0068630981259047985, + 0.007579468190670013, + 0.012206584215164185, + 0.006645900197327137, + 0.00979013741016388, + 0.008119741454720497, + 0.004907228052616119, + 0.006356440018862486, + 0.007340872194617987, + 0.024523939937353134, + 0.015942515805363655, + 0.03529316559433937, + 0.021564045920968056, + 0.012945916503667831, + 0.04158705845475197, + 0.021843325346708298, + 0.02524099498987198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.02329086884856224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895011454820633 + ], + "opposite_logprob_tokens": [ + -2.5789501667022705 + ], + "logprob_reward": -0.07895011454820633, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895011454820633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47835299372673035, + 0.09892399609088898, + 0.10707785189151764, + 0.04713399335741997, + 0.04300115630030632, + 0.05766545236110687, + 0.04406129568815231, + 0.030774546787142754, + 0.03584645316004753, + 0.05716225132346153 + ], + "credit_map": [ + 0.05560929700732231, + 0.020372839644551277, + 0.025162432342767715, + 0.016719121485948563, + 0.025705212727189064, + 0.018231524154543877, + 0.1797049343585968, + 0.1368476152420044, + 0.011666896753013134, + 0.009065473452210426, + 0.012039571069180965, + 0.011061408556997776, + 0.012712939642369747, + 0.020467234775424004, + 0.0108757009729743, + 0.011034769006073475, + 0.019579924643039703, + 0.005710546392947435, + 0.014194396324455738, + 0.00882206205278635, + 0.008675815537571907, + 0.008029024116694927, + 0.026430495083332062, + 0.01563558727502823, + 0.0058036320842802525, + 0.005319374147802591, + 0.006160920485854149, + 0.005846637301146984, + 0.0051984828896820545, + 0.006736261770129204, + 0.007110327482223511, + 0.004958357661962509, + 0.006153288763016462, + 0.005019973032176495, + 0.006096628028899431, + 0.00443748664110899, + 0.003743353532627225, + 0.0067034270614385605, + 0.0056899916380643845, + 0.005157007370144129, + 0.007525332272052765, + 0.0038126464933156967, + 0.010191207751631737, + 0.005117869935929775, + 0.005750098265707493, + 0.012707364745438099, + 0.006218592636287212, + 0.00634233932942152, + 0.005803433712571859, + 0.004012513440102339, + 0.007174652069807053, + 0.006301408167928457, + 0.00531201995909214, + 0.004569757264107466, + 0.005183318629860878, + 0.005704190116375685, + 0.0034626731649041176, + 0.00317999254912138, + 0.004913959186524153, + 0.0036290925927460194, + 0.004985253792256117, + 0.004594995640218258, + 0.0028886660002171993, + 0.0031199119985103607, + 0.004652809351682663, + 0.0044673169031739235, + 0.005122159142047167, + 0.004641354084014893, + 0.004335075616836548, + 0.003285252721980214, + 0.004946500062942505, + 0.004395982716232538, + 0.011585853062570095, + 0.004738634917885065, + 0.007932666689157486, + 0.007519397418946028, + 0.0034482975024729967, + 0.00962509773671627, + 0.006172407418489456, + 0.006139897275716066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07895011454820633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895650714635849 + ], + "opposite_logprob_tokens": [ + -2.5789566040039062 + ], + "logprob_reward": -0.07895650714635849, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895650714635849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13578881323337555, + 0.07507634162902832, + 0.4309508800506592, + 0.0294553991407156, + 0.15104927122592926, + 0.06270740181207657, + 0.0275979395955801, + 0.03454508259892464, + 0.029560137540102005, + 0.023268697783350945 + ], + "credit_map": [ + 0.017546646296977997, + 0.011023626662790775, + 0.01913028210401535, + 0.013250702060759068, + 0.014094725251197815, + 0.01456014346331358, + 0.02529827132821083, + 0.02088441513478756, + 0.008087017573416233, + 0.00767799885943532, + 0.009652075357735157, + 0.00980659294873476, + 0.007601440418511629, + 0.011408207938075066, + 0.009081339463591576, + 0.011761668138206005, + 0.08787180483341217, + 0.012315163388848305, + 0.05843563750386238, + 0.022477230057120323, + 0.03073568269610405, + 0.0375647246837616, + 0.09715370833873749, + 0.08439694344997406, + 0.0037603494711220264, + 0.003174844430759549, + 0.0036623571068048477, + 0.0031156204640865326, + 0.003579073352739215, + 0.0028793702367693186, + 0.00551232323050499, + 0.003771463641896844, + 0.02128651551902294, + 0.017302213236689568, + 0.020820006728172302, + 0.027911925688385963, + 0.014114799909293652, + 0.01622161455452442, + 0.019136693328619003, + 0.014255501329898834, + 0.00843700673431158, + 0.005175188649445772, + 0.01355497632175684, + 0.007023239508271217, + 0.0043526384979486465, + 0.007294503506273031, + 0.006272920873016119, + 0.010596927255392075, + 0.002687229309231043, + 0.004294776823371649, + 0.0038022694643586874, + 0.003042819444090128, + 0.00359652959741652, + 0.0031170325819402933, + 0.00309420726262033, + 0.0039630732499063015, + 0.003662727540358901, + 0.002715543145313859, + 0.004980392288416624, + 0.006828244309872389, + 0.002513023791834712, + 0.0038141754921525717, + 0.003228332381695509, + 0.006802643649280071, + 0.0027192039415240288, + 0.005651434883475304, + 0.004053553566336632, + 0.004239700268954039, + 0.0034191031008958817, + 0.002772885374724865, + 0.0034718483220785856, + 0.0032324090134352446, + 0.002462737960740924, + 0.0030691891442984343, + 0.005031453911215067, + 0.0023939074017107487, + 0.0027145897038280964, + 0.0025615734048187733, + 0.00242570205591619, + 0.002609544899314642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.07895650714635849 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020565606653690338 + ], + "opposite_logprob_tokens": [ + -3.8955655097961426 + ], + "logprob_reward": -0.020565606653690338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020565606653690338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6090296506881714, + 0.09402064234018326, + 0.039868514984846115, + 0.03083937056362629, + 0.04083207622170448, + 0.07633478939533234, + 0.032352279871702194, + 0.026294583454728127, + 0.020359305664896965, + 0.030068757012486458 + ], + "credit_map": [ + 0.08564665168523788, + 0.03647685423493385, + 0.059615056961774826, + 0.03890678286552429, + 0.04603487625718117, + 0.050664037466049194, + 0.16618981957435608, + 0.12549558281898499, + 0.01405704952776432, + 0.008865793235599995, + 0.00960144866257906, + 0.012100175954401493, + 0.010399146005511284, + 0.012038393877446651, + 0.015137733891606331, + 0.01182090025395155, + 0.007662551011890173, + 0.00236413418315351, + 0.005277227610349655, + 0.003542887046933174, + 0.0033263596706092358, + 0.0032833837904036045, + 0.008651514537632465, + 0.005760455038398504, + 0.0038751638494431973, + 0.003135260194540024, + 0.005046000704169273, + 0.002866556402295828, + 0.0033666256349533796, + 0.004868505988270044, + 0.0045307110995054245, + 0.003150549251586199, + 0.005154108628630638, + 0.0052226996049284935, + 0.00661060307174921, + 0.005101562477648258, + 0.003893873654305935, + 0.005895655136555433, + 0.0043053277768194675, + 0.0046482449397444725, + 0.010674496181309223, + 0.00486252224072814, + 0.016506826505064964, + 0.007509589195251465, + 0.006739993579685688, + 0.010515199974179268, + 0.008526942692697048, + 0.010999218560755253, + 0.00504344142973423, + 0.003031040308997035, + 0.004921958781778812, + 0.0035219709388911724, + 0.004051369149237871, + 0.0037975884042680264, + 0.003956562373787165, + 0.0040283468551933765, + 0.005458081606775522, + 0.0031434125266969204, + 0.004195828456431627, + 0.0025748410262167454, + 0.0017890093149617314, + 0.003633162472397089, + 0.0019496137974783778, + 0.003550633555278182, + 0.00279587646946311, + 0.003440768225118518, + 0.002577574923634529, + 0.0020725445356220007, + 0.002396731171756983, + 0.0021270595025271177, + 0.0018452837830409408, + 0.003103465773165226, + 0.0038346017245203257, + 0.002850349759683013, + 0.0047691622748970985, + 0.004058776888996363, + 0.0024760288652032614, + 0.004577680956572294, + 0.003429643576964736, + 0.0040725162252783775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.020565606653690338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, slightly low-angle, medium tracking shot follows a motocross rider as he descends to the right, completing a breathtaking stunt. Clad in a vibrant red and white jersey with bold, abstract graphic elements, black protective pants, and heavy-duty boots, the rider is a figure of profound concentration. His full-face helmet, adorned with a sharp black and white pattern, obscures his expression, but his determined stance and grip on the mud-stained, bright yellow handlebars of his dirt bike convey absolute control. The camera steadily tracks his trajectory as he prepares to descend the steep slope of a large, compact dirt ramp. As the bike's rear tire makes contact, it sends a plume of loose, black soil into the air. The bike's suspension compresses violently under the impact, and the rider, knees bent and body leaning forward, deftly absorbs the shock, guiding the machine steadily and deliberately down the ramp. In the background, the cavernous, packed arena is a swirling vortex of movement and color, with thousands of spectators roaring in the glare of the intense stadium floodlights, which are out of focus and cast sharp, dramatic highlights and deep shadows, creating a high-energy, exhilarating, triumphant atmosphere.", + "seed": 28108, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_3534_seed_28108_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04805014319717884, + "vlm_logprob": -0.04805014319717884, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.04805014319717884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012269143015146255 + ], + "opposite_logprob_tokens": [ + -4.637269020080566 + ], + "logprob_reward": -0.012269143015146255, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012269143015146255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11513567715883255, + 0.04090212285518646, + 0.04196520522236824, + 0.3274822235107422, + 0.03622686490416527, + 0.034160785377025604, + 0.2838967740535736, + 0.05068456754088402, + 0.03289408981800079, + 0.03665168210864067 + ], + "credit_map": [ + 0.02320619858801365, + 0.020676737651228905, + 0.008842702023684978, + 0.005642122123390436, + 0.006776487920433283, + 0.02972026728093624, + 0.009950131177902222, + 0.010321035981178284, + 0.004210882820188999, + 0.006457137875258923, + 0.0037720841355621815, + 0.003646416589617729, + 0.004347693175077438, + 0.0076299626380205154, + 0.006454129237681627, + 0.0043838173151016235, + 0.004257620312273502, + 0.009349764324724674, + 0.007770158350467682, + 0.002603508299216628, + 0.004925455432385206, + 0.006767767947167158, + 0.0028145064134150743, + 0.0034764264710247517, + 0.021913854405283928, + 0.11949600279331207, + 0.050178349018096924, + 0.009880454279482365, + 0.013453884981572628, + 0.08893639594316483, + 0.011109720915555954, + 0.01251356303691864, + 0.008268444798886776, + 0.009062683209776878, + 0.0022349520586431026, + 0.003961944952607155, + 0.0038849206175655127, + 0.005881207529455423, + 0.0018709264695644379, + 0.001061787479557097, + 0.004877292085438967, + 0.007976140826940536, + 0.0028849185910075903, + 0.0028565728571265936, + 0.003395245410501957, + 0.0063650477677583694, + 0.0026446299161762, + 0.0031609409488737583, + 0.10773488879203796, + 0.02119900844991207, + 0.09230810403823853, + 0.01261480525135994, + 0.00919718574732542, + 0.022934840992093086, + 0.008783006109297276, + 0.00912491325289011, + 0.006544939242303371, + 0.007198130711913109, + 0.0055397809483110905, + 0.0028919223695993423, + 0.011910106986761093, + 0.009815542958676815, + 0.0032992148771882057, + 0.003484937362372875, + 0.0023288382217288017, + 0.004984891973435879, + 0.002933090552687645, + 0.0036799204535782337, + 0.006712689064443111, + 0.004118624608963728, + 0.004282928537577391, + 0.003853109898045659, + 0.004025331698358059, + 0.005059924442321062, + 0.0025821730960160494, + 0.007083419244736433, + 0.003917313646525145, + 0.003067492740228772, + 0.004675596486777067, + 0.00624043820425868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.012269143015146255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289354518055916 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289354518055916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289354518055916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10929711908102036, + 0.06064433604478836, + 0.041909899562597275, + 0.2410423755645752, + 0.0543411523103714, + 0.04425624758005142, + 0.16638155281543732, + 0.1494838297367096, + 0.032029490917921066, + 0.1006140187382698 + ], + "credit_map": [ + 0.025425683706998825, + 0.017458166927099228, + 0.0071428436785936356, + 0.004685414023697376, + 0.008346411399543285, + 0.028528308495879173, + 0.00767116341739893, + 0.010039123706519604, + 0.007136572618037462, + 0.009628823027014732, + 0.0036175965797156096, + 0.0040959808975458145, + 0.012292354367673397, + 0.008924677036702633, + 0.009804625064134598, + 0.005143702495843172, + 0.006332716904580593, + 0.008481297641992569, + 0.003285788930952549, + 0.0021545619238168, + 0.004372939467430115, + 0.008553839288651943, + 0.004982742480933666, + 0.003746010595932603, + 0.017449690029025078, + 0.0812968909740448, + 0.047784361988306046, + 0.009127859026193619, + 0.010296108201146126, + 0.05307150259613991, + 0.010735240764915943, + 0.01128068845719099, + 0.0103662284091115, + 0.012327807955443859, + 0.004507013596594334, + 0.005504683591425419, + 0.0047942535020411015, + 0.008184473030269146, + 0.005222474690526724, + 0.0034342168364673853, + 0.009819475933909416, + 0.009269644506275654, + 0.004379402846097946, + 0.0025251400656998158, + 0.00360141322016716, + 0.00884824339300394, + 0.002957048825919628, + 0.0028558748308569193, + 0.052183378487825394, + 0.012701369822025299, + 0.049289003014564514, + 0.011460901238024235, + 0.008201459422707558, + 0.012308553792536259, + 0.010905245319008827, + 0.009331616573035717, + 0.01733187586069107, + 0.03073686920106411, + 0.01730630360543728, + 0.009528109803795815, + 0.022522490471601486, + 0.033773571252822876, + 0.00791909359395504, + 0.010365510359406471, + 0.003094037529081106, + 0.0045900847762823105, + 0.002958964090794325, + 0.00493509229272604, + 0.00423640850931406, + 0.004632893018424511, + 0.003571130335330963, + 0.004010877106338739, + 0.01086539775133133, + 0.007093977648764849, + 0.013477416709065437, + 0.02701684832572937, + 0.00582431023940444, + 0.009014325216412544, + 0.012654254212975502, + 0.01466747559607029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.023289354518055916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1128908172249794 + ], + "opposite_logprob_tokens": [ + -2.2378907203674316 + ], + "logprob_reward": -0.1128908172249794, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1128908172249794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08729548752307892, + 0.033370815217494965, + 0.09251730889081955, + 0.4431610405445099, + 0.04128766059875488, + 0.0424799844622612, + 0.15317997336387634, + 0.04875591769814491, + 0.03246115520596504, + 0.025490645319223404 + ], + "credit_map": [ + 0.013936526142060757, + 0.019368590787053108, + 0.00564585393294692, + 0.004258106462657452, + 0.005953759886324406, + 0.024094555526971817, + 0.007236572913825512, + 0.006801525596529245, + 0.005267045926302671, + 0.005065653473138809, + 0.0024673070292919874, + 0.0025600146036595106, + 0.003276468487456441, + 0.0062249344773590565, + 0.005176630802452564, + 0.0033327583223581314, + 0.015479306690394878, + 0.02023339457809925, + 0.01517394557595253, + 0.005325882695615292, + 0.007396139670163393, + 0.013526195660233498, + 0.008124380372464657, + 0.007258064579218626, + 0.03516824170947075, + 0.12981896102428436, + 0.11680128425359726, + 0.02288748137652874, + 0.022013084962964058, + 0.06450437009334564, + 0.02454209513962269, + 0.027425521984696388, + 0.005617651157081127, + 0.0060269637033343315, + 0.007401777431368828, + 0.0066515374928712845, + 0.004519075620919466, + 0.005103026982396841, + 0.0031345118768513203, + 0.002833114704117179, + 0.00852584931999445, + 0.00948280468583107, + 0.005229567177593708, + 0.0026194583624601364, + 0.0037870926316827536, + 0.006851481273770332, + 0.0030267280526459217, + 0.0029570029582828283, + 0.03518140688538551, + 0.020416175946593285, + 0.027492035180330276, + 0.01429795753210783, + 0.011052354238927364, + 0.02254297211766243, + 0.009659585542976856, + 0.012537474744021893, + 0.008567475713789463, + 0.00904233567416668, + 0.012469964101910591, + 0.0029232942033559084, + 0.004364679101854563, + 0.006163799203932285, + 0.0031936801970005035, + 0.0020306874066591263, + 0.005266688764095306, + 0.004666772671043873, + 0.005417598411440849, + 0.003704654984176159, + 0.003392521757632494, + 0.005349515005946159, + 0.002402895363047719, + 0.0022605115082114935, + 0.003714613616466522, + 0.0029178494587540627, + 0.0029838746413588524, + 0.004912225063890219, + 0.0021638229954987764, + 0.003173474920913577, + 0.003103364259004593, + 0.0025214191991835833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.1128908172249794 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202240288257599 + ], + "opposite_logprob_tokens": [ + -2.8120224475860596 + ], + "logprob_reward": -0.06202240288257599, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202240288257599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1031726598739624, + 0.09604820609092712, + 0.06855230778455734, + 0.1443992406129837, + 0.07273639738559723, + 0.0544067919254303, + 0.13593490421772003, + 0.22149334847927094, + 0.023781726136803627, + 0.07947442680597305 + ], + "credit_map": [ + 0.018735412508249283, + 0.013562621548771858, + 0.01287255622446537, + 0.0077177491039037704, + 0.008526310324668884, + 0.019359128549695015, + 0.010010228492319584, + 0.012388653121888638, + 0.018727092072367668, + 0.016440050676465034, + 0.010712060146033764, + 0.009035338647663593, + 0.006619881838560104, + 0.015576755627989769, + 0.008248372934758663, + 0.010688653215765953, + 0.006569558288902044, + 0.012942173518240452, + 0.008090944029390812, + 0.004902231507003307, + 0.008302567526698112, + 0.01156034879386425, + 0.008192215114831924, + 0.007992266677320004, + 0.01514881756156683, + 0.03938334807753563, + 0.012359009124338627, + 0.009784000925719738, + 0.010488797910511494, + 0.03316904604434967, + 0.015046153217554092, + 0.009020071476697922, + 0.01126627903431654, + 0.019380265846848488, + 0.004242211580276489, + 0.010868500918149948, + 0.007631428074091673, + 0.012869531288743019, + 0.00402031047269702, + 0.0024578701704740524, + 0.013007495552301407, + 0.01095082238316536, + 0.00539704505354166, + 0.0033961189910769463, + 0.003799691330641508, + 0.009768359363079071, + 0.00409331452101469, + 0.003993940073996782, + 0.04423815757036209, + 0.013537378050386906, + 0.03666561841964722, + 0.0091709541156888, + 0.007978495210409164, + 0.012037304230034351, + 0.00548922922462225, + 0.006817764602601528, + 0.03220854699611664, + 0.041569992899894714, + 0.029036832973361015, + 0.012965467758476734, + 0.017556626349687576, + 0.05077039822936058, + 0.0188777893781662, + 0.018507694825530052, + 0.0032811975106596947, + 0.0023820321075618267, + 0.0028995475731790066, + 0.0023209410719573498, + 0.0035061126109212637, + 0.004986248444765806, + 0.00227330788038671, + 0.0021323389373719692, + 0.00921660102903843, + 0.007188719697296619, + 0.009286885149776936, + 0.017331097275018692, + 0.0069008152931928635, + 0.008032243698835373, + 0.010229532606899738, + 0.011288526467978954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.06202240288257599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029778998345136642 + ], + "opposite_logprob_tokens": [ + -3.5297789573669434 + ], + "logprob_reward": -0.029778998345136642, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029778998345136642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3163785934448242, + 0.14254024624824524, + 0.05377891659736633, + 0.20093342661857605, + 0.046271972358226776, + 0.05870847776532173, + 0.07293958961963654, + 0.05940878763794899, + 0.02456830069422722, + 0.02447163686156273 + ], + "credit_map": [ + 0.07134626060724258, + 0.069099061191082, + 0.023376649245619774, + 0.014755131676793098, + 0.014366029761731625, + 0.08024487644433975, + 0.017084060236811638, + 0.026106584817171097, + 0.02372499369084835, + 0.02870919741690159, + 0.010128304362297058, + 0.011682121083140373, + 0.009724658913910389, + 0.03552338853478432, + 0.009162805043160915, + 0.013884807005524635, + 0.008288118988275528, + 0.011626477353274822, + 0.009631827473640442, + 0.003735224949195981, + 0.0032404030207544565, + 0.008313317783176899, + 0.0034785603638738394, + 0.005464993417263031, + 0.014535268768668175, + 0.07081960886716843, + 0.0332561656832695, + 0.006836053915321827, + 0.008473413996398449, + 0.05015698820352554, + 0.007900002412497997, + 0.008955937810242176, + 0.0167708583176136, + 0.008365762419998646, + 0.00349799869582057, + 0.0032128659076988697, + 0.0021214974112808704, + 0.008826752193272114, + 0.0020975943189114332, + 0.0013786490308120847, + 0.014482279308140278, + 0.011067585088312626, + 0.006033245008438826, + 0.004490107763558626, + 0.0034102676436305046, + 0.01186655554920435, + 0.0035661146976053715, + 0.003792333183810115, + 0.014729870483279228, + 0.009827123954892159, + 0.012191723100841045, + 0.006881914101541042, + 0.006531756836920977, + 0.010790695436298847, + 0.005815623793751001, + 0.006170890759676695, + 0.010961675085127354, + 0.009209143929183483, + 0.013143139891326427, + 0.003740146290510893, + 0.00554180471226573, + 0.009786288253962994, + 0.0038893106393516064, + 0.003137284889817238, + 0.0030151803512126207, + 0.0024218810722231865, + 0.005363895557820797, + 0.0027047244366258383, + 0.0028980434872210026, + 0.002742716111242771, + 0.0025862071197479963, + 0.002835653256624937, + 0.0032403261866420507, + 0.002465370809659362, + 0.002800209680572152, + 0.005562617909163237, + 0.0021562608890235424, + 0.0026544227730482817, + 0.0026910919696092606, + 0.002901338739320636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.029778998345136642 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, slightly low-angle, medium tracking shot follows a motocross rider as he descends to the right, completing a breathtaking stunt. Clad in a vibrant red and white jersey with bold, abstract graphic elements, black protective pants, and heavy-duty boots, the rider is a figure of profound concentration. His full-face helmet, adorned with a sharp black and white pattern, obscures his expression, but his determined stance and grip on the mud-stained, bright yellow handlebars of his dirt bike convey absolute control. The camera steadily tracks his trajectory as he prepares to descend the steep slope of a large, compact dirt ramp. As the bike's rear tire makes contact, it sends a plume of loose, black soil into the air. The bike's suspension compresses violently under the impact, and the rider, knees bent and body leaning forward, deftly absorbs the shock, guiding the machine steadily and deliberately down the ramp. In the background, the cavernous, packed arena is a swirling vortex of movement and color, with thousands of spectators roaring in the glare of the intense stadium floodlights, which are out of focus and cast sharp, dramatic highlights and deep shadows, creating a high-energy, exhilarating, triumphant atmosphere.", + "seed": 28109, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_3534_seed_28109_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044269345700740814, + "vlm_logprob": -0.044269345700740814, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.044269345700740814, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01740304008126259 + ], + "opposite_logprob_tokens": [ + -4.1424031257629395 + ], + "logprob_reward": -0.01740304008126259, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01740304008126259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055876296013593674, + 0.14005789160728455, + 0.0640488713979721, + 0.08639019727706909, + 0.050066057592630386, + 0.4284311830997467, + 0.06773627549409866, + 0.0568586066365242, + 0.029809381812810898, + 0.02072519063949585 + ], + "credit_map": [ + 0.005067174788564444, + 0.00828308705240488, + 0.007584609556943178, + 0.008336980827152729, + 0.005427790340036154, + 0.008394981734454632, + 0.006136548239737749, + 0.00664512487128377, + 0.007246518507599831, + 0.013574455864727497, + 0.008547470904886723, + 0.008378838188946247, + 0.012440488673746586, + 0.04015257582068443, + 0.011819512583315372, + 0.03789803013205528, + 0.008463658392429352, + 0.00999827217310667, + 0.011213567107915878, + 0.0060516465455293655, + 0.004580779001116753, + 0.00980625580996275, + 0.0056402855552732944, + 0.008294404484331608, + 0.010981877334415913, + 0.013027453795075417, + 0.019319530576467514, + 0.005280692130327225, + 0.010251273401081562, + 0.008654255419969559, + 0.007545342203229666, + 0.011329776607453823, + 0.008683626540005207, + 0.005911004729568958, + 0.005469799507409334, + 0.0031198316719383, + 0.005361766088753939, + 0.012436057440936565, + 0.003980493173003197, + 0.005103473551571369, + 0.023750966414809227, + 0.015182864852249622, + 0.16534553468227386, + 0.02654748596251011, + 0.013033121824264526, + 0.12773846089839935, + 0.021679479628801346, + 0.03515327721834183, + 0.008651891723275185, + 0.008486022241413593, + 0.0064399223774671555, + 0.007701046299189329, + 0.011551826260983944, + 0.010511317290365696, + 0.004499141592532396, + 0.009895111434161663, + 0.010398110374808311, + 0.006889476906508207, + 0.006627440452575684, + 0.005850524641573429, + 0.010389120317995548, + 0.007264420390129089, + 0.004275129642337561, + 0.005164382513612509, + 0.0036784526892006397, + 0.0032966812141239643, + 0.005278305616229773, + 0.004262115340679884, + 0.0031976960599422455, + 0.0037067777011543512, + 0.002303627086803317, + 0.004085725173354149, + 0.002924900036305189, + 0.0030720916111022234, + 0.0027091260999441147, + 0.002405740786343813, + 0.0031385046895593405, + 0.0023772302083671093, + 0.0016372704412788153, + 0.0024603279307484627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.01740304008126259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865330085158348 + ], + "opposite_logprob_tokens": [ + -3.0486533641815186 + ], + "logprob_reward": -0.04865330085158348, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865330085158348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08684699237346649, + 0.05290835723280907, + 0.11309874802827835, + 0.10846888273954391, + 0.07231328636407852, + 0.1863897293806076, + 0.20251692831516266, + 0.09098059684038162, + 0.05864064767956734, + 0.02783583477139473 + ], + "credit_map": [ + 0.009014290757477283, + 0.010316052474081516, + 0.011788524687290192, + 0.011571952141821384, + 0.009364858269691467, + 0.010753195732831955, + 0.010369861498475075, + 0.013668256811797619, + 0.004278259817510843, + 0.010361963883042336, + 0.004165997728705406, + 0.0044643511064350605, + 0.011220928281545639, + 0.005957517307251692, + 0.006427435670047998, + 0.006031904835253954, + 0.013645960949361324, + 0.013155743479728699, + 0.020036544650793076, + 0.027843432500958443, + 0.008468580432236195, + 0.005857840646058321, + 0.017405493184924126, + 0.006685154512524605, + 0.011796843260526657, + 0.011954755522310734, + 0.018479397520422935, + 0.008905410766601562, + 0.020259592682123184, + 0.009567187167704105, + 0.013370838016271591, + 0.014134865254163742, + 0.011217305436730385, + 0.008793228305876255, + 0.009308510459959507, + 0.007088031619787216, + 0.0067128343507647514, + 0.01254531741142273, + 0.009576678276062012, + 0.007071378640830517, + 0.011083451099693775, + 0.011624389328062534, + 0.05978446826338768, + 0.021667204797267914, + 0.009082849137485027, + 0.043019603937864304, + 0.012810109183192253, + 0.01731763780117035, + 0.02766506001353264, + 0.026253722608089447, + 0.013518569990992546, + 0.03107605315744877, + 0.030245885252952576, + 0.028258448466658592, + 0.015727709978818893, + 0.0297714751213789, + 0.020978927612304688, + 0.011687085963785648, + 0.011350642889738083, + 0.010959816165268421, + 0.007699275389313698, + 0.01350419782102108, + 0.008339055813848972, + 0.0064615970477461815, + 0.008705627173185349, + 0.009782965295016766, + 0.0079796826466918, + 0.006304107140749693, + 0.007518903352320194, + 0.007603910285979509, + 0.0040856595151126385, + 0.0066597918048501015, + 0.004727599211037159, + 0.0036930926144123077, + 0.002963543636724353, + 0.0035232051741331816, + 0.0034333469811826944, + 0.002940129255875945, + 0.002627528505399823, + 0.0039273896254599094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.04865330085158348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201982870697975 + ], + "opposite_logprob_tokens": [ + -2.8120198249816895 + ], + "logprob_reward": -0.06201982870697975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201982870697975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0842851921916008, + 0.09377022087574005, + 0.06267768889665604, + 0.07364625483751297, + 0.06090116873383522, + 0.3034987449645996, + 0.1298912614583969, + 0.0943811759352684, + 0.05523516982793808, + 0.041713062673807144 + ], + "credit_map": [ + 0.006631891708821058, + 0.00902988389134407, + 0.0111624076962471, + 0.01029888354241848, + 0.008286023512482643, + 0.011051575653254986, + 0.013924267143011093, + 0.013900269754230976, + 0.008818727917969227, + 0.01064742635935545, + 0.012091752141714096, + 0.009176859632134438, + 0.019414039328694344, + 0.014155236072838306, + 0.00806821696460247, + 0.011397967115044594, + 0.009475139901041985, + 0.01038267556577921, + 0.010705636814236641, + 0.007007749285548925, + 0.005214368924498558, + 0.006918240804225206, + 0.005326936952769756, + 0.00764694157987833, + 0.009648221544921398, + 0.009568742476403713, + 0.01499849371612072, + 0.006270131096243858, + 0.008658445440232754, + 0.007763702422380447, + 0.008396212942898273, + 0.008342310786247253, + 0.009547365829348564, + 0.004195884335786104, + 0.007665029726922512, + 0.005151759833097458, + 0.006056781392544508, + 0.016048483550548553, + 0.006116596516221762, + 0.006119273602962494, + 0.041976165026426315, + 0.02956560254096985, + 0.08585909754037857, + 0.029235446825623512, + 0.01624886877834797, + 0.04772099107503891, + 0.0213338490575552, + 0.031558748334646225, + 0.016706692054867744, + 0.017228009179234505, + 0.015084256418049335, + 0.01701454073190689, + 0.010241832584142685, + 0.02104571834206581, + 0.008011525496840477, + 0.02455868571996689, + 0.022926097735762596, + 0.010547616519033909, + 0.013669201172888279, + 0.008776129223406315, + 0.008266901597380638, + 0.01597241498529911, + 0.0066813016310334206, + 0.00754151726141572, + 0.008684543892741203, + 0.008614711463451385, + 0.00924076046794653, + 0.006146565545350313, + 0.005518313962966204, + 0.00742139108479023, + 0.0029127250891178846, + 0.006696164608001709, + 0.00496348412707448, + 0.004803283605724573, + 0.006323712877929211, + 0.004397442564368248, + 0.004633297212421894, + 0.004581410903483629, + 0.003817415563389659, + 0.008193018846213818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.06201982870697975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999925523996353 + ], + "opposite_logprob_tokens": [ + -2.6949992179870605 + ], + "logprob_reward": -0.06999925523996353, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999925523996353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08756273984909058, + 0.26822689175605774, + 0.11316148936748505, + 0.05241720378398895, + 0.0416799820959568, + 0.1274288147687912, + 0.10253839939832687, + 0.05923330783843994, + 0.11347632110118866, + 0.03427477180957794 + ], + "credit_map": [ + 0.00810197088867426, + 0.010266496799886227, + 0.014370734803378582, + 0.012400315143167973, + 0.008068208582699299, + 0.010393116623163223, + 0.011244234628975391, + 0.012717672623693943, + 0.013087401166558266, + 0.025260692462325096, + 0.01658363826572895, + 0.014849666506052017, + 0.021794090047478676, + 0.07108236104249954, + 0.02547551691532135, + 0.08009354770183563, + 0.01938467286527157, + 0.01760006882250309, + 0.01733754202723503, + 0.015236922539770603, + 0.007173283025622368, + 0.012466078624129295, + 0.010238292627036572, + 0.013724641874432564, + 0.005487448535859585, + 0.0070350016467273235, + 0.009918561205267906, + 0.0048630498349666595, + 0.006629172246903181, + 0.005213250871747732, + 0.0049462756142020226, + 0.00832444615662098, + 0.006145286373794079, + 0.0036777330096811056, + 0.006145230028778315, + 0.0032309521920979023, + 0.003278199350461364, + 0.009048543870449066, + 0.005371495615690947, + 0.00478254584595561, + 0.011290003545582294, + 0.01382053829729557, + 0.033590786159038544, + 0.014544958248734474, + 0.009064467623829842, + 0.01860012114048004, + 0.011140411719679832, + 0.01537752989679575, + 0.020112110301852226, + 0.013964584097266197, + 0.01424914225935936, + 0.007936663925647736, + 0.006107210647314787, + 0.019615257158875465, + 0.005798226222395897, + 0.014755220152437687, + 0.01442564744502306, + 0.006964060477912426, + 0.011073569767177105, + 0.00365641457028687, + 0.004100863356143236, + 0.00991157628595829, + 0.003759742947295308, + 0.0053414348512887955, + 0.013500364497303963, + 0.012146336026489735, + 0.02397383563220501, + 0.014605560339987278, + 0.0057924366556108, + 0.015723485499620438, + 0.006410927977412939, + 0.021323392167687416, + 0.004218120593577623, + 0.0050646960735321045, + 0.0035999638494104147, + 0.005270928610116243, + 0.0031416115816682577, + 0.0042047505266964436, + 0.0031329505145549774, + 0.005641752853989601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.06999925523996353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327130362391472 + ], + "opposite_logprob_tokens": [ + -3.773271322250366 + ], + "logprob_reward": -0.02327130362391472, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327130362391472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1544346809387207, + 0.1265321671962738, + 0.06351320445537567, + 0.2652718126773834, + 0.055461812764406204, + 0.11088137328624725, + 0.13937197625637054, + 0.030055521056056023, + 0.018513724207878113, + 0.03596370294690132 + ], + "credit_map": [ + 0.014628013595938683, + 0.015772953629493713, + 0.020647253841161728, + 0.015865761786699295, + 0.01875472255051136, + 0.01806832291185856, + 0.023294484242796898, + 0.02740318328142166, + 0.009030050598084927, + 0.01567268557846546, + 0.011733246967196465, + 0.010350942611694336, + 0.013928334228694439, + 0.027652382850646973, + 0.010332822799682617, + 0.027831707149744034, + 0.0100587522611022, + 0.009444822557270527, + 0.007794227916747332, + 0.011124929413199425, + 0.005005932413041592, + 0.006289889570325613, + 0.007512771990150213, + 0.006281889043748379, + 0.030057506635785103, + 0.032004233449697495, + 0.07575157284736633, + 0.021141713485121727, + 0.0405307337641716, + 0.02745928429067135, + 0.013324668630957603, + 0.025002099573612213, + 0.009785524569451809, + 0.0029856893233954906, + 0.006453195121139288, + 0.002976944437250495, + 0.007017367519438267, + 0.016169985756278038, + 0.003820764133706689, + 0.006252345163375139, + 0.012411619536578655, + 0.015733890235424042, + 0.012469629757106304, + 0.014707553200423717, + 0.01019645668566227, + 0.015500937588512897, + 0.014163719490170479, + 0.015697579830884933, + 0.020185677334666252, + 0.015576304867863655, + 0.0155409574508667, + 0.020818015560507774, + 0.00800318643450737, + 0.022982442751526833, + 0.00862008798867464, + 0.02764531411230564, + 0.006819121539592743, + 0.0027950790245085955, + 0.0047496953047811985, + 0.004979935009032488, + 0.0020850165747106075, + 0.0044088708236813545, + 0.002133631380274892, + 0.002084174659103155, + 0.003341139294207096, + 0.0027146681677550077, + 0.0029117802623659372, + 0.002065685112029314, + 0.0016018395544961095, + 0.0022791833616793156, + 0.0014987081522122025, + 0.0021007226314395666, + 0.005766488611698151, + 0.0046692825853824615, + 0.006524053402245045, + 0.004222062882035971, + 0.002208058023825288, + 0.004484097007662058, + 0.0022894071880728006, + 0.005800254177302122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.02327130362391472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, slightly low-angle, medium tracking shot follows a motocross rider as he descends to the right, completing a breathtaking stunt. Clad in a vibrant red and white jersey with bold, abstract graphic elements, black protective pants, and heavy-duty boots, the rider is a figure of profound concentration. His full-face helmet, adorned with a sharp black and white pattern, obscures his expression, but his determined stance and grip on the mud-stained, bright yellow handlebars of his dirt bike convey absolute control. The camera steadily tracks his trajectory as he prepares to descend the steep slope of a large, compact dirt ramp. As the bike's rear tire makes contact, it sends a plume of loose, black soil into the air. The bike's suspension compresses violently under the impact, and the rider, knees bent and body leaning forward, deftly absorbs the shock, guiding the machine steadily and deliberately down the ramp. In the background, the cavernous, packed arena is a swirling vortex of movement and color, with thousands of spectators roaring in the glare of the intense stadium floodlights, which are out of focus and cast sharp, dramatic highlights and deep shadows, creating a high-energy, exhilarating, triumphant atmosphere.", + "seed": 28110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_10/index_3534_seed_28110_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06668058224022388, + "vlm_logprob": -0.06668058224022388, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.06668058224022388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016135700047016144 + ], + "opposite_logprob_tokens": [ + -4.2661356925964355 + ], + "logprob_reward": -0.016135700047016144, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016135700047016144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054821424186229706, + 0.03519277647137642, + 0.019050762057304382, + 0.1936243176460266, + 0.017392579466104507, + 0.41891226172447205, + 0.11760681867599487, + 0.031866274774074554, + 0.08802350610494614, + 0.02350926585495472 + ], + "credit_map": [ + 0.007369234226644039, + 0.008518123999238014, + 0.007565282750874758, + 0.005255465861409903, + 0.006738182157278061, + 0.0071013253182172775, + 0.00471379654482007, + 0.007560012396425009, + 0.006579335778951645, + 0.007474248297512531, + 0.003961560316383839, + 0.003104252740740776, + 0.0030420394614338875, + 0.0056684291921556, + 0.002827260410413146, + 0.002535649109631777, + 0.002927279332652688, + 0.0038319185841828585, + 0.002328620757907629, + 0.002327504102140665, + 0.0015340985264629126, + 0.0025276807136833668, + 0.002423891332000494, + 0.0011497671948745847, + 0.04252953827381134, + 0.012853079475462437, + 0.036134008318185806, + 0.008127822540700436, + 0.007608907762914896, + 0.051091037690639496, + 0.008584639057517052, + 0.026695286855101585, + 0.002784458687528968, + 0.002829537959769368, + 0.0015090678352862597, + 0.0016007327940315008, + 0.001940040965564549, + 0.001892401254735887, + 0.0025964207015931606, + 0.0022399183362722397, + 0.17509548366069794, + 0.17353498935699463, + 0.008318934589624405, + 0.011142886243760586, + 0.013703524135053158, + 0.018381541594862938, + 0.009333403781056404, + 0.009401519782841206, + 0.04539690166711807, + 0.045109253376722336, + 0.005842592567205429, + 0.0037835624534636736, + 0.004564202390611172, + 0.004215166438370943, + 0.0053005642257630825, + 0.0033945704344660044, + 0.005943930707871914, + 0.003100588219240308, + 0.004575352184474468, + 0.0035623598378151655, + 0.003055890556424856, + 0.005213333293795586, + 0.0030259345658123493, + 0.0033888863399624825, + 0.01485035102814436, + 0.011387799866497517, + 0.023104434832930565, + 0.009722886607050896, + 0.006044487003237009, + 0.01236656866967678, + 0.005333393812179565, + 0.005213584750890732, + 0.002236381871625781, + 0.00267842342145741, + 0.0037697881925851107, + 0.004792127758264542, + 0.0021274005994200706, + 0.0035683035384863615, + 0.0022589012514799833, + 0.00207794108428061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.016135700047016144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347355917096138 + ], + "opposite_logprob_tokens": [ + -3.6513473987579346 + ], + "logprob_reward": -0.026347355917096138, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347355917096138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09900789707899094, + 0.05650908499956131, + 0.026353875175118446, + 0.2731204032897949, + 0.03568736091256142, + 0.1726439893245697, + 0.12480399757623672, + 0.09947085380554199, + 0.075846366584301, + 0.03655615821480751 + ], + "credit_map": [ + 0.021058334037661552, + 0.02032649517059326, + 0.008925138972699642, + 0.007051900960505009, + 0.009397552348673344, + 0.015552854165434837, + 0.007008762564510107, + 0.009686874225735664, + 0.010922585614025593, + 0.011512435972690582, + 0.005048913881182671, + 0.005373611114919186, + 0.005403450224548578, + 0.0085183996707201, + 0.005994210485368967, + 0.003735482692718506, + 0.004248145967721939, + 0.005505433306097984, + 0.0037730224430561066, + 0.0028491225093603134, + 0.0020127727184444666, + 0.0028798619750887156, + 0.001980197848752141, + 0.00310532096773386, + 0.05544450506567955, + 0.023036494851112366, + 0.03904491662979126, + 0.01681356504559517, + 0.019237976521253586, + 0.08798792958259583, + 0.01827329397201538, + 0.013281731866300106, + 0.004852022510021925, + 0.005730125121772289, + 0.0032384400255978107, + 0.0030677304603159428, + 0.004162054508924484, + 0.005037476774305105, + 0.004948558751493692, + 0.004650958813726902, + 0.06764058023691177, + 0.06348375231027603, + 0.004226522985845804, + 0.007490768563002348, + 0.008640037849545479, + 0.00894874706864357, + 0.0069770063273608685, + 0.005236589815467596, + 0.044491373002529144, + 0.04852055758237839, + 0.005040473770350218, + 0.007648791652172804, + 0.005340262781828642, + 0.004168978426605463, + 0.005420307628810406, + 0.004173266235738993, + 0.02151157334446907, + 0.009653833694756031, + 0.014606510289013386, + 0.011706854216754436, + 0.006324979942291975, + 0.018307825550436974, + 0.007699945941567421, + 0.009659340605139732, + 0.013720209710299969, + 0.014485161751508713, + 0.007037628907710314, + 0.010998720303177834, + 0.006250204984098673, + 0.014036373235285282, + 0.004399018362164497, + 0.004919059574604034, + 0.005340749863535166, + 0.004902909975498915, + 0.00654245913028717, + 0.00551200658082962, + 0.002933270065113902, + 0.005205169785767794, + 0.002691644011065364, + 0.0034279506653547287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.026347355917096138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900544047355652 + ], + "opposite_logprob_tokens": [ + -2.464005470275879 + ], + "logprob_reward": -0.08900544047355652, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900544047355652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051948171108961105, + 0.028998874127864838, + 0.021884186193346977, + 0.052503008395433426, + 0.010031095705926418, + 0.6885043382644653, + 0.08173758536577225, + 0.014696603640913963, + 0.032616619020700455, + 0.017079532146453857 + ], + "credit_map": [ + 0.008904216811060905, + 0.008425656706094742, + 0.005741259083151817, + 0.004768907558172941, + 0.004661700688302517, + 0.007584364619106054, + 0.006376306526362896, + 0.005485759116709232, + 0.004916008096188307, + 0.004837303888052702, + 0.0036211302503943443, + 0.002607821486890316, + 0.0034535678569227457, + 0.0037414487451314926, + 0.003332077758386731, + 0.0024895172100514174, + 0.0022479433100670576, + 0.0028578066267073154, + 0.0027279320638626814, + 0.0024937139824032784, + 0.004151289816945791, + 0.002254748484119773, + 0.0029729364905506372, + 0.002177816117182374, + 0.005097961984574795, + 0.0033811074681580067, + 0.011099141091108322, + 0.004061595071107149, + 0.008092532865703106, + 0.006101647857576609, + 0.006530804093927145, + 0.00813821516931057, + 0.001568054431118071, + 0.0013708298793062568, + 0.0012974214041605592, + 0.001250231871381402, + 0.0013384370831772685, + 0.0013438158202916384, + 0.0009202127694152296, + 0.0009420925052836537, + 0.28430041670799255, + 0.293066143989563, + 0.014454745687544346, + 0.018061866983771324, + 0.020843250676989555, + 0.03373105823993683, + 0.011436434462666512, + 0.01261038426309824, + 0.026190800592303276, + 0.026538662612438202, + 0.005796434357762337, + 0.008768342435359955, + 0.0036930344067513943, + 0.002776171313598752, + 0.004156285896897316, + 0.0038178565446287394, + 0.0024841506965458393, + 0.002010223688557744, + 0.0019942843355238438, + 0.001910412684082985, + 0.0012439701240509748, + 0.0024444900918751955, + 0.001089352066628635, + 0.0015197191387414932, + 0.003775859484449029, + 0.005483766086399555, + 0.004252956714481115, + 0.007252342067658901, + 0.002884737914428115, + 0.004108077380806208, + 0.002332942094653845, + 0.002525935648009181, + 0.0022246455773711205, + 0.0023374406155198812, + 0.0033125607296824455, + 0.002273328136652708, + 0.0015004262095317245, + 0.002222184557467699, + 0.0014847595011815429, + 0.0017241875175386667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.08900544047355652 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901241421699524 + ], + "opposite_logprob_tokens": [ + -2.464012384414673 + ], + "logprob_reward": -0.08901241421699524, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901241421699524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10767790675163269, + 0.025681979954242706, + 0.016363492235541344, + 0.16551943123340607, + 0.03172939643263817, + 0.4572894275188446, + 0.13171201944351196, + 0.021664148196578026, + 0.01866811141371727, + 0.023694060742855072 + ], + "credit_map": [ + 0.024143626913428307, + 0.022226333618164062, + 0.009454644285142422, + 0.008189721964299679, + 0.01078188419342041, + 0.017459198832511902, + 0.005939373280853033, + 0.009483123198151588, + 0.007070387713611126, + 0.004870830103754997, + 0.0024544941261410713, + 0.0018547277431935072, + 0.0016229349421337247, + 0.0041730329394340515, + 0.002039265586063266, + 0.001596307149156928, + 0.0020128136966377497, + 0.0025178082287311554, + 0.002047635382041335, + 0.002263968111947179, + 0.0024622133933007717, + 0.001893990091048181, + 0.0017582589061930776, + 0.0014068045420572162, + 0.03055517002940178, + 0.01133920531719923, + 0.03670132905244827, + 0.008619306609034538, + 0.005973664578050375, + 0.028159065172076225, + 0.01021843682974577, + 0.03395324572920799, + 0.005125782452523708, + 0.005754516925662756, + 0.0031519108451902866, + 0.0033782029058784246, + 0.0031057021114975214, + 0.0038841012865304947, + 0.0036843316629529, + 0.0036448463797569275, + 0.19335788488388062, + 0.19421806931495667, + 0.006915905047208071, + 0.010470434091985226, + 0.0139059042558074, + 0.021040787920355797, + 0.007794050965458155, + 0.00958639569580555, + 0.040845535695552826, + 0.04858449473977089, + 0.009119944646954536, + 0.012277782894670963, + 0.006357965525239706, + 0.004172522109001875, + 0.005799042992293835, + 0.004554723855108023, + 0.0038449796847999096, + 0.0031785035971552134, + 0.0028395578265190125, + 0.0029969820752739906, + 0.0014204959152266383, + 0.0034983528312295675, + 0.0017011164454743266, + 0.0021841602865606546, + 0.0027593441773205996, + 0.0026934652123600245, + 0.002473310334607959, + 0.0031030664686113596, + 0.0012940310407429934, + 0.0034410248044878244, + 0.0012775772484019399, + 0.0016262914286926389, + 0.0026142976712435484, + 0.0028260070830583572, + 0.004695019219070673, + 0.0043802266009151936, + 0.0015859910054132342, + 0.0033444485161453485, + 0.0019012095872312784, + 0.0023468625731766224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.08901241421699524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motocross rider wearing a red and white jersey with black pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider on a dirt bike with bright yellow handlebars and a black and white patterned helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider descend a steep dirt ramp while the camera tracks from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible plume of black soil rising from the bike's rear tire upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a packed arena with out-of-focus stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290200054645538 + ], + "opposite_logprob_tokens": [ + -2.2379019260406494 + ], + "logprob_reward": -0.11290200054645538, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290200054645538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15473483502864838, + 0.03242980316281319, + 0.02412566728889942, + 0.25805771350860596, + 0.038587454706430435, + 0.2174840122461319, + 0.22920119762420654, + 0.022611085325479507, + 0.006559961009770632, + 0.016208281740546227 + ], + "credit_map": [ + 0.033895641565322876, + 0.030478160828351974, + 0.01812455803155899, + 0.010888393968343735, + 0.009267366491258144, + 0.02559741772711277, + 0.010674060322344303, + 0.01580924354493618, + 0.005528424400836229, + 0.0062981098890304565, + 0.004226022865623236, + 0.0037678941152989864, + 0.002770226914435625, + 0.005068744998425245, + 0.002512268954887986, + 0.002258113119751215, + 0.0026798471808433533, + 0.003371031256392598, + 0.003669396508485079, + 0.0032276480924338102, + 0.0029923776164650917, + 0.0029924430418759584, + 0.0029108065646141768, + 0.002282114466652274, + 0.06187846139073372, + 0.01914491504430771, + 0.043660592287778854, + 0.01120904739946127, + 0.009169349446892738, + 0.07651989907026291, + 0.01081171166151762, + 0.025663724169135094, + 0.005907590035349131, + 0.006076561287045479, + 0.004291585646569729, + 0.004489134531468153, + 0.0037917857989668846, + 0.005057717673480511, + 0.0045256358571350574, + 0.004447446670383215, + 0.09046299755573273, + 0.08779436349868774, + 0.0036875116638839245, + 0.005712205544114113, + 0.007249151822179556, + 0.009687626734375954, + 0.005141682457178831, + 0.007748476695269346, + 0.07753598690032959, + 0.08859098702669144, + 0.01216195896267891, + 0.01790504716336727, + 0.009802930057048798, + 0.00689713703468442, + 0.009446166455745697, + 0.006860989611595869, + 0.004357580095529556, + 0.002323446562513709, + 0.003738226369023323, + 0.0029327464289963245, + 0.0013996448833495378, + 0.003947876859456301, + 0.0017741824267432094, + 0.0021373818162828684, + 0.0009035500697791576, + 0.0009464137838222086, + 0.0014083547284826636, + 0.0009664163808338344, + 0.0005246733780950308, + 0.0008953534415923059, + 0.00044070262811146677, + 0.00047449718113057315, + 0.0022982109803706408, + 0.001871328568086028, + 0.0029282483737915754, + 0.0034206360578536987, + 0.0007832630071789026, + 0.002590036252513528, + 0.00107204623054713, + 0.0012445119209587574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.11290200054645538 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_tensor([6236, 6236])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_tensor([6236, 6236])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..02307db7215cc57d2987e7f54b6c99963c2cc2a9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_tensor([6236, 6236])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot smoothly follows a pristine, cherry-red classic convertible as it gracefully navigates a sun-drenched town street corner. The car's flawless paintwork possesses a deep, liquid sheen, while its extensive, mirror-polished chrome bumpers, grille, and side moldings brilliantly reflect the distorted, warm-toned images of the historic red-brick facades lining the street. With its cream-colored leather interior and large, slim-rimmed steering wheel, the car glides effortlessly over the cobblestone road. The scene is bathed in the bright, warm glow of midday sunlight, casting crisp, well-defined shadows and creating dazzling, mirrored highlights on the car's metallic surfaces. In the background, quaint storefronts with large glass windows and striped canvas awnings create a nostalgic, idyllic atmosphere, capturing a tranquil and timeless moment of classic automotive beauty.", + "seed": 14250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_6236_seed_14250_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08442132538184524, + "vlm_logprob": -0.08442132538184524, + "vlm_yes_no_margin": 3.549999952316284, + "vlm_token_credit": -0.08442132538184524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0056104338727891445 + ], + "opposite_logprob_tokens": [ + -5.505610466003418 + ], + "logprob_reward": -0.0056104338727891445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0056104338727891445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06341306120157242, + 0.06343258917331696, + 0.10203190892934799, + 0.17124100029468536, + 0.060769207775592804, + 0.04107046127319336, + 0.19130820035934448, + 0.18485331535339355, + 0.08331663906574249, + 0.03856368735432625 + ], + "credit_map": [ + 0.007650358136743307, + 0.008022001013159752, + 0.008585037663578987, + 0.009301678277552128, + 0.00718445610255003, + 0.006348863244056702, + 0.005839050747454166, + 0.010481610894203186, + 0.006457848474383354, + 0.005958891473710537, + 0.00818860623985529, + 0.007154813967645168, + 0.010638964362442493, + 0.010977492667734623, + 0.007288431748747826, + 0.00676753232255578, + 0.008676457218825817, + 0.004042759537696838, + 0.0277696680277586, + 0.008111544884741306, + 0.007031023968011141, + 0.03532259911298752, + 0.0046338425017893314, + 0.006444002967327833, + 0.03220094367861748, + 0.015124034136533737, + 0.023302102461457253, + 0.020459486171603203, + 0.01861848123371601, + 0.020923607051372528, + 0.023812083527445793, + 0.01680024154484272, + 0.0062188152223825455, + 0.007723846472799778, + 0.007455247454345226, + 0.009525317698717117, + 0.007353859953582287, + 0.006987107452005148, + 0.007448544725775719, + 0.008056461811065674, + 0.004705829545855522, + 0.004501692485064268, + 0.005521195475012064, + 0.004665971267968416, + 0.005670326296240091, + 0.00484032928943634, + 0.005117152817547321, + 0.006047962699085474, + 0.008197023533284664, + 0.004834465682506561, + 0.005370824132114649, + 0.07006829977035522, + 0.007063407450914383, + 0.005954815074801445, + 0.07958272844552994, + 0.010236630216240883, + 0.011417684145271778, + 0.01829461008310318, + 0.028047295287251472, + 0.037880271673202515, + 0.01784353144466877, + 0.01669391617178917, + 0.028305787593126297, + 0.026370197534561157, + 0.007327566854655743, + 0.00781443901360035, + 0.016292843967676163, + 0.012544927187263966, + 0.005751217249780893, + 0.005101202055811882, + 0.01206125132739544, + 0.016423190012574196, + 0.00687935808673501, + 0.0044006649404764175, + 0.005480946507304907, + 0.005660746246576309, + 0.003411155892536044, + 0.0034163151867687702, + 0.00511521939188242, + 0.004199279006570578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.0056104338727891445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023323478177189827 + ], + "opposite_logprob_tokens": [ + -3.7733235359191895 + ], + "logprob_reward": -0.023323478177189827, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023323478177189827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039662010967731476, + 0.03580418974161148, + 0.056832872331142426, + 0.12543822824954987, + 0.07932258397340775, + 0.03531498834490776, + 0.3770942687988281, + 0.14768041670322418, + 0.06437148153781891, + 0.03847899287939072 + ], + "credit_map": [ + 0.004302672576159239, + 0.002762420801445842, + 0.0064724585972726345, + 0.005139108747243881, + 0.005541315767914057, + 0.006009470205754042, + 0.004127743188291788, + 0.005306817125529051, + 0.0056465864181518555, + 0.0034120583441108465, + 0.0037999681662768126, + 0.003894524183124304, + 0.006407765205949545, + 0.006171508226543665, + 0.0029287857469171286, + 0.003542987396940589, + 0.004446924198418856, + 0.004284236114472151, + 0.014527874998748302, + 0.005759215448051691, + 0.005955751519650221, + 0.00975953321903944, + 0.004272205289453268, + 0.00782713107764721, + 0.022889522835612297, + 0.009411617182195187, + 0.01670096255838871, + 0.014907400123775005, + 0.01619507558643818, + 0.01635476015508175, + 0.016254642978310585, + 0.01272424217313528, + 0.010367125272750854, + 0.009168867953121662, + 0.009415701031684875, + 0.008316807448863983, + 0.00906752422451973, + 0.010249930433928967, + 0.01084055658429861, + 0.011896071024239063, + 0.004841392394155264, + 0.003833327442407608, + 0.003956970293074846, + 0.003360043279826641, + 0.005757291801273823, + 0.005037947557866573, + 0.0036283826921135187, + 0.004899629857391119, + 0.01604096218943596, + 0.011643735691905022, + 0.011592917144298553, + 0.1332087218761444, + 0.018111273646354675, + 0.012717999517917633, + 0.14979329705238342, + 0.02398534305393696, + 0.009301861748099327, + 0.014255528338253498, + 0.024993441998958588, + 0.030465533956885338, + 0.013016016222536564, + 0.013453913852572441, + 0.022878969088196754, + 0.01931515336036682, + 0.007575268857181072, + 0.007053849752992392, + 0.010439776815474033, + 0.009593505412340164, + 0.004008994437754154, + 0.005008595529943705, + 0.008109556511044502, + 0.012581936083734035, + 0.0046774884685873985, + 0.00476272776722908, + 0.006629306823015213, + 0.0057126861065626144, + 0.0038436632603406906, + 0.005012715235352516, + 0.004083503037691116, + 0.003756899619475007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.023323478177189827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34864118695259094 + ], + "opposite_logprob_tokens": [ + -1.2236411571502686 + ], + "logprob_reward": -0.34864118695259094, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34864118695259094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07855246216058731, + 0.07186412811279297, + 0.16831690073013306, + 0.11838728934526443, + 0.11970090866088867, + 0.034485649317502975, + 0.10730406641960144, + 0.21359768509864807, + 0.05357929319143295, + 0.034211646765470505 + ], + "credit_map": [ + 0.00849646981805563, + 0.006461076904088259, + 0.014534550718963146, + 0.009962485171854496, + 0.010053224861621857, + 0.007243048865348101, + 0.00947105884552002, + 0.012330536730587482, + 0.007505915593355894, + 0.00860380195081234, + 0.01235772855579853, + 0.007426633033901453, + 0.010040533728897572, + 0.008326876908540726, + 0.00950820092111826, + 0.008094429038465023, + 0.011722428724169731, + 0.005944021046161652, + 0.020429017022252083, + 0.010276537388563156, + 0.015330293215811253, + 0.054421138018369675, + 0.008192107081413269, + 0.042001351714134216, + 0.02103528566658497, + 0.010204719379544258, + 0.014524245634675026, + 0.012897390872240067, + 0.0143794696778059, + 0.015172434970736504, + 0.01737515814602375, + 0.012798566371202469, + 0.016792312264442444, + 0.015041155740618706, + 0.013103212229907513, + 0.019897401332855225, + 0.01615160144865513, + 0.013163307681679726, + 0.011763456277549267, + 0.013788454234600067, + 0.0048067886382341385, + 0.003954964689910412, + 0.0034852458629757166, + 0.004140205681324005, + 0.003640406299382448, + 0.00657548476010561, + 0.004209207836538553, + 0.0036733404267579317, + 0.006517329718917608, + 0.0077200885862112045, + 0.009191521443426609, + 0.024480437859892845, + 0.011571146547794342, + 0.011207285337150097, + 0.02549389749765396, + 0.01112234778702259, + 0.010716337710618973, + 0.013988863676786423, + 0.03426765650510788, + 0.051559656858444214, + 0.013292001560330391, + 0.012574905529618263, + 0.036018699407577515, + 0.041179537773132324, + 0.004840224515646696, + 0.005555405281484127, + 0.010303632356226444, + 0.008114428259432316, + 0.0031800440046936274, + 0.0031331065110862255, + 0.005623403936624527, + 0.012829041108489037, + 0.0037028116639703512, + 0.0046610659919679165, + 0.005737339612096548, + 0.006373530253767967, + 0.002958423225209117, + 0.0033164434134960175, + 0.003965089097619057, + 0.0034969395492225885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.34864118695259094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06562616676092148, + 0.02845214307308197, + 0.043728772550821304, + 0.0670584887266159, + 0.06540009379386902, + 0.023143786936998367, + 0.2800308167934418, + 0.20150355994701385, + 0.1794569492340088, + 0.04559924080967903 + ], + "credit_map": [ + 0.007784826681017876, + 0.005865417420864105, + 0.01347787119448185, + 0.006079452577978373, + 0.008906643837690353, + 0.009557940997183323, + 0.004729845095425844, + 0.009224165230989456, + 0.0036339322105050087, + 0.002677110955119133, + 0.0052159628830850124, + 0.0032676062546670437, + 0.005104430951178074, + 0.002727794460952282, + 0.0025386193301528692, + 0.003286682302132249, + 0.0030153433326631784, + 0.0012369893956929445, + 0.006877393461763859, + 0.0018473640084266663, + 0.003916747402399778, + 0.017564620822668076, + 0.0014052592450752854, + 0.007865050807595253, + 0.010787943378090858, + 0.005085423588752747, + 0.00876240711659193, + 0.01133185438811779, + 0.006575803738087416, + 0.008302436210215092, + 0.007926232181489468, + 0.008286379277706146, + 0.011510936543345451, + 0.005263921804726124, + 0.008338801562786102, + 0.011472636833786964, + 0.0067590209655463696, + 0.007459494285285473, + 0.007203303277492523, + 0.007391971070319414, + 0.004367264453321695, + 0.002215050859376788, + 0.0018397064413875341, + 0.002740137977525592, + 0.001846364582888782, + 0.004210938233882189, + 0.0023652694653719664, + 0.0035590524785220623, + 0.011356914415955544, + 0.008892686106264591, + 0.008404566906392574, + 0.0986618772149086, + 0.015785349532961845, + 0.009305169805884361, + 0.11087371408939362, + 0.016750518232584, + 0.008721349760890007, + 0.010478410869836807, + 0.029975246638059616, + 0.05897034332156181, + 0.014273176901042461, + 0.01154370792210102, + 0.04712025448679924, + 0.020421041175723076, + 0.021560154855251312, + 0.020068466663360596, + 0.027398305013775826, + 0.025263004004955292, + 0.01810224913060665, + 0.02090294472873211, + 0.013725235126912594, + 0.032436564564704895, + 0.004290867131203413, + 0.0038154160138219595, + 0.0073476070538163185, + 0.011185520328581333, + 0.003322889329865575, + 0.004200184717774391, + 0.005783511325716972, + 0.005653242580592632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.01818765513598919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343872770667076 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343872770667076, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343872770667076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06348484009504318, + 0.0436287522315979, + 0.08933532983064651, + 0.024489905685186386, + 0.034748002886772156, + 0.025100423023104668, + 0.44485005736351013, + 0.12352157384157181, + 0.06899227946996689, + 0.08184882998466492 + ], + "credit_map": [ + 0.008494450710713863, + 0.009239079430699348, + 0.009306477382779121, + 0.007159881293773651, + 0.007169486489146948, + 0.005468692630529404, + 0.006520155817270279, + 0.01012661773711443, + 0.006631067954003811, + 0.004988899454474449, + 0.007288256660103798, + 0.006944465916603804, + 0.005342033691704273, + 0.0034870922099798918, + 0.004427565261721611, + 0.004519369453191757, + 0.013688841834664345, + 0.009694745764136314, + 0.01872030459344387, + 0.008750051259994507, + 0.00536977918818593, + 0.01485047023743391, + 0.006687713786959648, + 0.011573415249586105, + 0.0031749873887747526, + 0.0019123406382277608, + 0.003368769306689501, + 0.004865055903792381, + 0.002006575930863619, + 0.002886206842958927, + 0.0035392686259001493, + 0.0027367009315639734, + 0.005418139044195414, + 0.00404090341180563, + 0.004226737190037966, + 0.005546439904719591, + 0.00493971211835742, + 0.0036357303615659475, + 0.0035017377231270075, + 0.0034386080224066973, + 0.0031174886971712112, + 0.002806771080940962, + 0.0024601949844509363, + 0.0030123800970613956, + 0.002797015942633152, + 0.004109895788133144, + 0.002806264441460371, + 0.003990414086729288, + 0.019354209303855896, + 0.010252336040139198, + 0.014779225923120975, + 0.1610768437385559, + 0.020263606682419777, + 0.013246753253042698, + 0.18261152505874634, + 0.023265575990080833, + 0.007380490191280842, + 0.013078195042908192, + 0.015004581771790981, + 0.02452590875327587, + 0.010812767781317234, + 0.011947697028517723, + 0.024019328877329826, + 0.01675260066986084, + 0.007771591190248728, + 0.007754432037472725, + 0.009268464520573616, + 0.012492490001022816, + 0.005569029599428177, + 0.006357655860483646, + 0.008554172702133656, + 0.01122443750500679, + 0.009025883860886097, + 0.010015320032835007, + 0.009620395489037037, + 0.016292503103613853, + 0.008134763687849045, + 0.00841145683079958, + 0.012755478732287884, + 0.0075930263847112656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.026343872770667076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot smoothly follows a pristine, cherry-red classic convertible as it gracefully navigates a sun-drenched town street corner. The car's flawless paintwork possesses a deep, liquid sheen, while its extensive, mirror-polished chrome bumpers, grille, and side moldings brilliantly reflect the distorted, warm-toned images of the historic red-brick facades lining the street. With its cream-colored leather interior and large, slim-rimmed steering wheel, the car glides effortlessly over the cobblestone road. The scene is bathed in the bright, warm glow of midday sunlight, casting crisp, well-defined shadows and creating dazzling, mirrored highlights on the car's metallic surfaces. In the background, quaint storefronts with large glass windows and striped canvas awnings create a nostalgic, idyllic atmosphere, capturing a tranquil and timeless moment of classic automotive beauty.", + "seed": 14251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_6236_seed_14251_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0821492349728942, + "vlm_logprob": -0.0821492349728942, + "vlm_yes_no_margin": 3.575000011920929, + "vlm_token_credit": -0.0821492349728942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009139369241893291 + ], + "opposite_logprob_tokens": [ + -4.884139537811279 + ], + "logprob_reward": -0.009139369241893291, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009139369241893291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08767159283161163, + 0.10919975489377975, + 0.22514232993125916, + 0.05483277514576912, + 0.0909639447927475, + 0.13243858516216278, + 0.08258751779794693, + 0.06669916957616806, + 0.09350255876779556, + 0.05696176365017891 + ], + "credit_map": [ + 0.01325262151658535, + 0.008374463766813278, + 0.009997829794883728, + 0.011481762863695621, + 0.010237984359264374, + 0.009912699460983276, + 0.010659877210855484, + 0.013754351064562798, + 0.009658260270953178, + 0.02684962749481201, + 0.007267461623996496, + 0.02862202376127243, + 0.012262450531125069, + 0.007551288697868586, + 0.009604825638234615, + 0.007383809890598059, + 0.018438247963786125, + 0.008928731083869934, + 0.07291693985462189, + 0.020808301866054535, + 0.014073141850531101, + 0.009337874129414558, + 0.06957904994487762, + 0.011060039512813091, + 0.004665960092097521, + 0.007521787658333778, + 0.010982798412442207, + 0.005861347075551748, + 0.005070408806204796, + 0.006981899030506611, + 0.005636859219521284, + 0.008111708797514439, + 0.005525356158614159, + 0.004600695334374905, + 0.03195828199386597, + 0.01110251434147358, + 0.005573509726673365, + 0.005894612055271864, + 0.006715530063956976, + 0.019593430683016777, + 0.003805150045081973, + 0.007208323106169701, + 0.025515319779515266, + 0.049595676362514496, + 0.0027648601680994034, + 0.0035717652644962072, + 0.018083984032273293, + 0.02189350314438343, + 0.006181920878589153, + 0.004083443898707628, + 0.012387698516249657, + 0.005372715648263693, + 0.0027323763351887465, + 0.00344079895876348, + 0.017159314826130867, + 0.031229248270392418, + 0.004419291391968727, + 0.003334139008074999, + 0.006898846942931414, + 0.010880772024393082, + 0.0016984152607619762, + 0.002332535572350025, + 0.014672096818685532, + 0.02246306650340557, + 0.004802924580872059, + 0.009309642016887665, + 0.014361312612891197, + 0.014321272261440754, + 0.004292196594178677, + 0.005223655141890049, + 0.017303168773651123, + 0.02388838119804859, + 0.00376992579549551, + 0.006690667476505041, + 0.004272544756531715, + 0.007032133638858795, + 0.004247942008078098, + 0.003813444171100855, + 0.01708856038749218, + 0.01004654448479414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.009139369241893291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060507982969284 + ], + "opposite_logprob_tokens": [ + -3.8956050872802734 + ], + "logprob_reward": -0.02060507982969284, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060507982969284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03055904246866703, + 0.3458481431007385, + 0.19672046601772308, + 0.03201543167233467, + 0.036437418311834335, + 0.1219118982553482, + 0.09157147258520126, + 0.03375757858157158, + 0.03582429140806198, + 0.07535424083471298 + ], + "credit_map": [ + 0.0044655511155724525, + 0.002913236618041992, + 0.0031758886761963367, + 0.0031856887508183718, + 0.003683999413624406, + 0.003944181837141514, + 0.004375480115413666, + 0.004815015010535717, + 0.03586788848042488, + 0.09089043736457825, + 0.027511587366461754, + 0.12815536558628082, + 0.01684534177184105, + 0.012565928511321545, + 0.022789517417550087, + 0.011222068220376968, + 0.015789765864610672, + 0.008028829470276833, + 0.06066667288541794, + 0.01839965023100376, + 0.014469772577285767, + 0.010815219022333622, + 0.05815922096371651, + 0.010391313582658768, + 0.0030660289339721203, + 0.0034177640918642282, + 0.007873172871768475, + 0.004552559927105904, + 0.002761234762147069, + 0.004065557383000851, + 0.00229058088734746, + 0.003988530486822128, + 0.002841805573552847, + 0.004665477201342583, + 0.004571692086756229, + 0.00589206675067544, + 0.004152413923293352, + 0.0038266940973699093, + 0.005694800987839699, + 0.004792468622326851, + 0.005456825252622366, + 0.0055338479578495026, + 0.025326743721961975, + 0.04380058869719505, + 0.0035801103804260492, + 0.00460799178108573, + 0.015595034696161747, + 0.018010765314102173, + 0.004906683228909969, + 0.004492342937737703, + 0.024152450263500214, + 0.008142957463860512, + 0.003583484096452594, + 0.004319474566727877, + 0.014629766345024109, + 0.02734430879354477, + 0.0020201834850013256, + 0.0018076555570587516, + 0.0032597039826214314, + 0.006442668382078409, + 0.0008053300553001463, + 0.001155251869931817, + 0.007632365450263023, + 0.01063442137092352, + 0.0016935437452048063, + 0.003315318375825882, + 0.004104470834136009, + 0.005753089673817158, + 0.001715445308946073, + 0.0018835875671356916, + 0.00708901509642601, + 0.010269819758832455, + 0.00382455182261765, + 0.008966726250946522, + 0.005797264631837606, + 0.009815412573516369, + 0.006876522209495306, + 0.004855635110288858, + 0.022207457572221756, + 0.013010668568313122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.02060507982969284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862250089645386 + ], + "opposite_logprob_tokens": [ + -1.2236225605010986 + ], + "logprob_reward": -0.34862250089645386, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34862250089645386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022234901785850525, + 0.8292813897132874, + 0.022439906373620033, + 0.010860010981559753, + 0.01589345932006836, + 0.03129614144563675, + 0.011627932079136372, + 0.026146650314331055, + 0.024337444454431534, + 0.005882140714675188 + ], + "credit_map": [ + 0.003967083990573883, + 0.0020380262285470963, + 0.001828351290896535, + 0.0025744689628481865, + 0.0032937340438365936, + 0.0027132725808769464, + 0.003189476439729333, + 0.002630488248541951, + 0.08772508800029755, + 0.22808289527893066, + 0.05871475860476494, + 0.3068414628505707, + 0.040547456592321396, + 0.02775437757372856, + 0.05377532169222832, + 0.025840099900960922, + 0.002573368838056922, + 0.00195236224681139, + 0.006962639279663563, + 0.001957019791007042, + 0.0015859308186918497, + 0.0011782967485487461, + 0.004978276323527098, + 0.0012520098825916648, + 0.000690213986672461, + 0.001028814585879445, + 0.002697225194424391, + 0.00198321882635355, + 0.0008134540403261781, + 0.0011329713743180037, + 0.001064163981936872, + 0.0014499496901407838, + 0.0014212254900485277, + 0.002050048438832164, + 0.0030129924416542053, + 0.002186657628044486, + 0.001462051528505981, + 0.0012030245270580053, + 0.0027913893572986126, + 0.0017660699086263776, + 0.002543437760323286, + 0.001484008738771081, + 0.007783956825733185, + 0.008249125443398952, + 0.0016182458493858576, + 0.0017204850446432829, + 0.003966180142015219, + 0.003930701408535242, + 0.0005671371472999454, + 0.0007993352483026683, + 0.0036043201107531786, + 0.0014319613110274076, + 0.0006713152397423983, + 0.0006576892919838428, + 0.0014642010210081935, + 0.002431972185149789, + 0.0016781032318249345, + 0.0011483477428555489, + 0.0035073477774858475, + 0.0035785618238151073, + 0.0006689138244837523, + 0.0008629647199995816, + 0.005559316836297512, + 0.009143098257482052, + 0.0013809233205392957, + 0.0021820368710905313, + 0.002600880805402994, + 0.0038128234446048737, + 0.0013676374219357967, + 0.0011906139552593231, + 0.004535727668553591, + 0.0072668008506298065, + 0.0005358611815609038, + 0.0007908393745310605, + 0.0005373522290028632, + 0.0006722924299538136, + 0.0006240417715162039, + 0.0004230031045153737, + 0.0012290022568777204, + 0.001069748541340232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.34862250089645386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199061319231987 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199061319231987, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199061319231987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0812661275267601, + 0.3793219327926636, + 0.13829490542411804, + 0.049538612365722656, + 0.07731263339519501, + 0.054145004600286484, + 0.11774559319019318, + 0.02783529832959175, + 0.05608763545751572, + 0.018452292308211327 + ], + "credit_map": [ + 0.017946390435099602, + 0.008340923115611076, + 0.007212549913674593, + 0.009669533930718899, + 0.012139689177274704, + 0.007374078966677189, + 0.01000524964183569, + 0.00857771560549736, + 0.03987371549010277, + 0.10297276824712753, + 0.027784140780568123, + 0.14045734703540802, + 0.018639184534549713, + 0.01272053737193346, + 0.024866443127393723, + 0.012007798999547958, + 0.017233282327651978, + 0.02588554099202156, + 0.018236026167869568, + 0.02458946593105793, + 0.014632323756814003, + 0.011122707277536392, + 0.015968510881066322, + 0.01062705460935831, + 0.005604492966085672, + 0.005471247714012861, + 0.010519924573600292, + 0.0064196218736469746, + 0.004450553096830845, + 0.005429717246443033, + 0.004745339974761009, + 0.0068977149203419685, + 0.005418128799647093, + 0.007380561903119087, + 0.011886082589626312, + 0.00763482553884387, + 0.008140194229781628, + 0.009232761338353157, + 0.011678962968289852, + 0.015941111370921135, + 0.0016645957948639989, + 0.0030077998526394367, + 0.009526277892291546, + 0.01989368349313736, + 0.001976083731278777, + 0.0019410692621022463, + 0.006822085939347744, + 0.009313405491411686, + 0.0053391484543681145, + 0.005795771721750498, + 0.03666135296225548, + 0.012615268118679523, + 0.004449206404387951, + 0.00622241385281086, + 0.01754978857934475, + 0.029112640768289566, + 0.0023058131337165833, + 0.00160268135368824, + 0.003018316114321351, + 0.0035182444844394922, + 0.0011254233540967107, + 0.0012991437688469887, + 0.0059004537761211395, + 0.009065221063792706, + 0.0029613617807626724, + 0.005076841916888952, + 0.006320004817098379, + 0.008349996991455555, + 0.002715720096603036, + 0.003522659884765744, + 0.009948939085006714, + 0.017192110419273376, + 0.0018364504212513566, + 0.0026678708381950855, + 0.0012648057891055942, + 0.0019128310959786177, + 0.0025628837756812572, + 0.0012284929398447275, + 0.004231028258800507, + 0.002747930120676756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.014199061319231987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180163577198982 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180163577198982, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180163577198982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05525612458586693, + 0.35217681527137756, + 0.1351516842842102, + 0.06739652156829834, + 0.07689699530601501, + 0.04384106397628784, + 0.10135940462350845, + 0.07393466681241989, + 0.06564419716596603, + 0.028342535719275475 + ], + "credit_map": [ + 0.004529320169240236, + 0.004239354748278856, + 0.004665339831262827, + 0.007432423532009125, + 0.006733035668730736, + 0.006768274586647749, + 0.008786534890532494, + 0.012101837433874607, + 0.03621713072061539, + 0.09308331459760666, + 0.026853397488594055, + 0.13168129324913025, + 0.0173769723623991, + 0.011451955884695053, + 0.023464417085051537, + 0.012048332020640373, + 0.011892310343682766, + 0.005959867034107447, + 0.04251177981495857, + 0.010674693621695042, + 0.008237021043896675, + 0.006607248913496733, + 0.042347535490989685, + 0.006921224761754274, + 0.005303612444549799, + 0.005971543490886688, + 0.01781592331826687, + 0.012346512638032436, + 0.005107510834932327, + 0.007788950111716986, + 0.005910026375204325, + 0.007152447942644358, + 0.005960907321423292, + 0.006657782010734081, + 0.012949508614838123, + 0.01471652090549469, + 0.0073545523919165134, + 0.0063864500261843204, + 0.014643462374806404, + 0.008227812126278877, + 0.0033108864445239305, + 0.0040118456818163395, + 0.0081575196236372, + 0.007712905295193195, + 0.00401306850835681, + 0.0022093357983976603, + 0.004973853472620249, + 0.009451651945710182, + 0.00797466840595007, + 0.005088277626782656, + 0.015965264290571213, + 0.0066031236201524734, + 0.0030166474170982838, + 0.0038305826019495726, + 0.02012111432850361, + 0.03875972330570221, + 0.005490443669259548, + 0.004928892012685537, + 0.011781170964241028, + 0.013171649537980556, + 0.0021896669641137123, + 0.003511931747198105, + 0.013495844788849354, + 0.01936507038772106, + 0.003204992273822427, + 0.005736943334341049, + 0.008609420619904995, + 0.010700824670493603, + 0.0027737640775740147, + 0.0036264185328036547, + 0.013042585924267769, + 0.01794925332069397, + 0.002004395704716444, + 0.0033536849077790976, + 0.002295243088155985, + 0.0038497874047607183, + 0.0020948119927197695, + 0.0016120519721880555, + 0.008681802079081535, + 0.0044507598504424095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.018180163577198982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot smoothly follows a pristine, cherry-red classic convertible as it gracefully navigates a sun-drenched town street corner. The car's flawless paintwork possesses a deep, liquid sheen, while its extensive, mirror-polished chrome bumpers, grille, and side moldings brilliantly reflect the distorted, warm-toned images of the historic red-brick facades lining the street. With its cream-colored leather interior and large, slim-rimmed steering wheel, the car glides effortlessly over the cobblestone road. The scene is bathed in the bright, warm glow of midday sunlight, casting crisp, well-defined shadows and creating dazzling, mirrored highlights on the car's metallic surfaces. In the background, quaint storefronts with large glass windows and striped canvas awnings create a nostalgic, idyllic atmosphere, capturing a tranquil and timeless moment of classic automotive beauty.", + "seed": 14252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_6236_seed_14252_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07568713445216417, + "vlm_logprob": -0.07568713445216417, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.07568713445216417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005638527683913708 + ], + "opposite_logprob_tokens": [ + -5.880638599395752 + ], + "logprob_reward": -0.005638527683913708, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.005638527683913708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0925525575876236, + 0.11107668280601501, + 0.042071934789419174, + 0.05042954161763191, + 0.08336083590984344, + 0.16085681319236755, + 0.034514397382736206, + 0.05238493159413338, + 0.3287222981452942, + 0.044029928743839264 + ], + "credit_map": [ + 0.0133402980864048, + 0.011875584721565247, + 0.015992728993296623, + 0.009527960792183876, + 0.012102995067834854, + 0.013421403244137764, + 0.00872048269957304, + 0.007571108173578978, + 0.015173042193055153, + 0.023462830111384392, + 0.012368807569146156, + 0.01668086275458336, + 0.00984484888613224, + 0.016931599006056786, + 0.006777043454349041, + 0.009837659075856209, + 0.007723481394350529, + 0.008711905218660831, + 0.005039579700678587, + 0.006172016263008118, + 0.003555624047294259, + 0.0033578318543732166, + 0.0032298783771693707, + 0.004281620029360056, + 0.004278229083865881, + 0.00519076082855463, + 0.005590661894530058, + 0.004987751599401236, + 0.0047401017509400845, + 0.006693453527987003, + 0.010419927537441254, + 0.008528659120202065, + 0.0044169737957417965, + 0.005804718006402254, + 0.006245824508368969, + 0.005358132999390364, + 0.011756455525755882, + 0.00994524359703064, + 0.019718239083886147, + 0.020115256309509277, + 0.0056039681658148766, + 0.005745720583945513, + 0.00480500515550375, + 0.004582894500344992, + 0.012139298021793365, + 0.009591978043317795, + 0.05937518551945686, + 0.059012770652770996, + 0.002642714185640216, + 0.0036161853931844234, + 0.004841209389269352, + 0.006051459815353155, + 0.002600783482193947, + 0.002088138833642006, + 0.006807032972574234, + 0.005866874009370804, + 0.00416290620341897, + 0.005125467199832201, + 0.007252143695950508, + 0.009922931902110577, + 0.005276413634419441, + 0.0045605250634253025, + 0.009401628747582436, + 0.006682921200990677, + 0.02853178232908249, + 0.1161060556769371, + 0.027087179943919182, + 0.03977485001087189, + 0.03147635981440544, + 0.026902535930275917, + 0.02393706701695919, + 0.03490649536252022, + 0.0052339062094688416, + 0.007276966702193022, + 0.00422545475885272, + 0.00546254450455308, + 0.005787198431789875, + 0.00576819246634841, + 0.005563660990446806, + 0.004712005611509085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.005638527683913708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602861419320107 + ], + "opposite_logprob_tokens": [ + -3.8956029415130615 + ], + "logprob_reward": -0.020602861419320107, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602861419320107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08852796256542206, + 0.2431161254644394, + 0.08209602534770966, + 0.1029951348900795, + 0.08545442670583725, + 0.10847052931785583, + 0.06937409192323685, + 0.10558123141527176, + 0.079248808324337, + 0.035135701298713684 + ], + "credit_map": [ + 0.013043391518294811, + 0.008231263607740402, + 0.019904518499970436, + 0.012437673285603523, + 0.010932416655123234, + 0.008426368236541748, + 0.009447319433093071, + 0.0061050113290548325, + 0.029276076704263687, + 0.06025412306189537, + 0.02754496969282627, + 0.031085431575775146, + 0.021477533504366875, + 0.04025459662079811, + 0.012784883379936218, + 0.02043849602341652, + 0.010934042744338512, + 0.009541193954646587, + 0.013153794221580029, + 0.012743527069687843, + 0.006933613680303097, + 0.007347765378654003, + 0.010323259979486465, + 0.011118825525045395, + 0.007280572783201933, + 0.008912768214941025, + 0.007829422131180763, + 0.009476563893258572, + 0.012906664982438087, + 0.026141073554754257, + 0.01863350346684456, + 0.011814567260444164, + 0.0033582723699510098, + 0.005856129340827465, + 0.0063848611898720264, + 0.00698198052123189, + 0.01457066461443901, + 0.0115499971434474, + 0.018227653577923775, + 0.018524862825870514, + 0.00910486001521349, + 0.006493974011391401, + 0.01068803295493126, + 0.007871568202972412, + 0.013186964206397533, + 0.013083811849355698, + 0.023412132635712624, + 0.024629181250929832, + 0.004894873593002558, + 0.007850768975913525, + 0.00769901555031538, + 0.010770766995847225, + 0.008614296093583107, + 0.004517898429185152, + 0.01401133369654417, + 0.011015134863555431, + 0.00671704625710845, + 0.0077796876430511475, + 0.010765595361590385, + 0.023231051862239838, + 0.010089807212352753, + 0.00599132664501667, + 0.027631253004074097, + 0.013375471346080303, + 0.006998314056545496, + 0.025175675749778748, + 0.006846575532108545, + 0.009643523953855038, + 0.0072533064521849155, + 0.007276283111423254, + 0.0066914004273712635, + 0.009363727644085884, + 0.004574276972562075, + 0.004536251071840525, + 0.0029343608766794205, + 0.003983944654464722, + 0.004964039195328951, + 0.005268269218504429, + 0.00438143964856863, + 0.004493121523410082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.020602861419320107 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343671679496765 + ], + "opposite_logprob_tokens": [ + -1.31343674659729 + ], + "logprob_reward": -0.31343671679496765, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343671679496765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05384398624300957, + 0.11495453864336014, + 0.04022851958870888, + 0.0379057303071022, + 0.1462215781211853, + 0.13364310562610626, + 0.025107642635703087, + 0.025715749710798264, + 0.40684759616851807, + 0.015531592071056366 + ], + "credit_map": [ + 0.009846051223576069, + 0.008395611308515072, + 0.006902405526489019, + 0.0067117540165781975, + 0.006787716876715422, + 0.006524119526147842, + 0.004955868702381849, + 0.0037204609252512455, + 0.013339679688215256, + 0.027702458202838898, + 0.013733526691794395, + 0.014746662229299545, + 0.008838573470711708, + 0.021085575222969055, + 0.006022962741553783, + 0.009485099464654922, + 0.005269800778478384, + 0.0082908496260643, + 0.002900055143982172, + 0.003749252064153552, + 0.006841354072093964, + 0.004088506102561951, + 0.004077823366969824, + 0.005010876804590225, + 0.0029027259442955256, + 0.0031049782410264015, + 0.003969706594944, + 0.0033736727200448513, + 0.004638638347387314, + 0.008637483231723309, + 0.006877545267343521, + 0.004400978796184063, + 0.005789840593934059, + 0.005247696302831173, + 0.005896584130823612, + 0.006916971877217293, + 0.01442506443709135, + 0.01808800734579563, + 0.04329768195748329, + 0.04655972868204117, + 0.005394992884248495, + 0.004662480670958757, + 0.004939671140164137, + 0.005077882669866085, + 0.005874228198081255, + 0.00591737125068903, + 0.050094056874513626, + 0.05168242007493973, + 0.002149915089830756, + 0.0033649916294962168, + 0.0032589146867394447, + 0.00427677808329463, + 0.002021605148911476, + 0.0018279176438227296, + 0.004588301759213209, + 0.003619218012318015, + 0.001843227888457477, + 0.003547549480572343, + 0.00272168405354023, + 0.003945605829358101, + 0.002964975545182824, + 0.0017329988768324256, + 0.0054913898929953575, + 0.003468318609520793, + 0.03367118537425995, + 0.13961713016033173, + 0.035323839634656906, + 0.05032864958047867, + 0.040190961211919785, + 0.03255975991487503, + 0.03893960267305374, + 0.036216456443071365, + 0.0022089211270213127, + 0.002103095641359687, + 0.0017490311292931437, + 0.001402410794980824, + 0.0023455608170479536, + 0.0021547588985413313, + 0.001623929594643414, + 0.0019438841845840216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.31343671679496765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577168092131615 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577168092131615, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577168092131615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10155032575130463, + 0.49555104970932007, + 0.07396253198385239, + 0.038769420236349106, + 0.06458263099193573, + 0.05013725534081459, + 0.0835670456290245, + 0.0257380660623312, + 0.014325786381959915, + 0.05181588977575302 + ], + "credit_map": [ + 0.015720566734671593, + 0.011846452951431274, + 0.01997431553900242, + 0.013080366887152195, + 0.013436724431812763, + 0.012476837262511253, + 0.008820508606731892, + 0.006194559391587973, + 0.05894731730222702, + 0.11923510581254959, + 0.05914730206131935, + 0.06381627917289734, + 0.040433481335639954, + 0.08581513911485672, + 0.026169132441282272, + 0.04198731482028961, + 0.009119102731347084, + 0.010213332250714302, + 0.006919344887137413, + 0.009840419515967369, + 0.010109453462064266, + 0.008463309146463871, + 0.008962806314229965, + 0.01033477671444416, + 0.004063994158059359, + 0.004750777035951614, + 0.002847138326615095, + 0.001985523384064436, + 0.008463580161333084, + 0.007988582365214825, + 0.003926395438611507, + 0.004743433091789484, + 0.0023033295292407274, + 0.0034312340430915356, + 0.0029782636556774378, + 0.0036668821703642607, + 0.007117604836821556, + 0.006949539761990309, + 0.018506892025470734, + 0.019628889858722687, + 0.0022411788813769817, + 0.001550799235701561, + 0.0032602110877633095, + 0.002728260587900877, + 0.0029359397012740374, + 0.0025455825962126255, + 0.017304565757513046, + 0.017570719122886658, + 0.00972678977996111, + 0.015122615732252598, + 0.009267820045351982, + 0.009752820245921612, + 0.007392948493361473, + 0.007565409876406193, + 0.01199432648718357, + 0.012744319625198841, + 0.0016902827192097902, + 0.0019835480488836765, + 0.003435758175328374, + 0.004937971942126751, + 0.0022665609139949083, + 0.002030319534242153, + 0.006491885054856539, + 0.0029017417691648006, + 0.0019038579193875194, + 0.0011753480648621917, + 0.0018891278887167573, + 0.0028997999615967274, + 0.0008137058466672897, + 0.001160097192041576, + 0.002178489463403821, + 0.0023053614422678947, + 0.007093483116477728, + 0.008310345932841301, + 0.002954604336991906, + 0.002891858108341694, + 0.008995924144983292, + 0.01042618416249752, + 0.005403086543083191, + 0.005740410648286343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.020577168092131615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047815486788749695, + 0.26802200078964233, + 0.034103937447071075, + 0.03357115015387535, + 0.014264619909226894, + 0.05847908928990364, + 0.01994701847434044, + 0.06900454312562943, + 0.42096930742263794, + 0.033822838217020035 + ], + "credit_map": [ + 0.006008169613778591, + 0.00360195548273623, + 0.013030501082539558, + 0.0054584466852247715, + 0.005562770180404186, + 0.006406902335584164, + 0.004601410124450922, + 0.003145329188555479, + 0.030565038323402405, + 0.06537918001413345, + 0.03275644779205322, + 0.03311009332537651, + 0.020991817116737366, + 0.04745663329958916, + 0.0148286372423172, + 0.02293415553867817, + 0.0060066417790949345, + 0.0082178246229887, + 0.0036670765839517117, + 0.004851509351283312, + 0.002914471086114645, + 0.0021969452500343323, + 0.001851951819844544, + 0.004397518467158079, + 0.002575516700744629, + 0.0029958724044263363, + 0.003109576879069209, + 0.0031728148460388184, + 0.005229146685451269, + 0.006789397448301315, + 0.005899947602301836, + 0.003798877354711294, + 0.0010926071554422379, + 0.0013458450557664037, + 0.0015767496079206467, + 0.0009723484981805086, + 0.0026980622205883265, + 0.0021034518722444773, + 0.0021776219364255667, + 0.0022979332134127617, + 0.0033185547217726707, + 0.0031810826621949673, + 0.0025496925227344036, + 0.002319260500371456, + 0.008817807771265507, + 0.006055218167603016, + 0.01563243567943573, + 0.016605038195848465, + 0.0018618045141920447, + 0.002653988543897867, + 0.002865527756512165, + 0.00258081522770226, + 0.002634424949064851, + 0.0013351013185456395, + 0.0032105513382703066, + 0.002804805990308523, + 0.0037249845918267965, + 0.005895219277590513, + 0.006458349991589785, + 0.01335871871560812, + 0.007020506076514721, + 0.004243369214236736, + 0.01868712157011032, + 0.009616276249289513, + 0.03243258222937584, + 0.1385623663663864, + 0.0364605151116848, + 0.051719196140766144, + 0.04113639146089554, + 0.03342508152127266, + 0.052311837673187256, + 0.03492133691906929, + 0.004204946104437113, + 0.004476610105484724, + 0.003592119086533785, + 0.004843892529606819, + 0.003394807456061244, + 0.003058383706957102, + 0.004924725275486708, + 0.005327355116605759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.018180398270487785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot smoothly follows a pristine, cherry-red classic convertible as it gracefully navigates a sun-drenched town street corner. The car's flawless paintwork possesses a deep, liquid sheen, while its extensive, mirror-polished chrome bumpers, grille, and side moldings brilliantly reflect the distorted, warm-toned images of the historic red-brick facades lining the street. With its cream-colored leather interior and large, slim-rimmed steering wheel, the car glides effortlessly over the cobblestone road. The scene is bathed in the bright, warm glow of midday sunlight, casting crisp, well-defined shadows and creating dazzling, mirrored highlights on the car's metallic surfaces. In the background, quaint storefronts with large glass windows and striped canvas awnings create a nostalgic, idyllic atmosphere, capturing a tranquil and timeless moment of classic automotive beauty.", + "seed": 14253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_11/index_6236_seed_14253_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08391262758523226, + "vlm_logprob": -0.08391262758523226, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.08391262758523226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00615024846047163 + ], + "opposite_logprob_tokens": [ + -5.381150245666504 + ], + "logprob_reward": -0.00615024846047163, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00615024846047163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12172240018844604, + 0.07200458645820618, + 0.1236150860786438, + 0.029568800702691078, + 0.06976233422756195, + 0.07124926894903183, + 0.0330989696085453, + 0.039062291383743286, + 0.3754299581050873, + 0.06448637694120407 + ], + "credit_map": [ + 0.011989648453891277, + 0.0118684321641922, + 0.020206179469823837, + 0.010295945219695568, + 0.010706113651394844, + 0.035384051501750946, + 0.010600235313177109, + 0.010671778582036495, + 0.006055708508938551, + 0.0053024087101221085, + 0.018604299053549767, + 0.011675252579152584, + 0.006565868388861418, + 0.008136444725096226, + 0.008643475361168385, + 0.007021124474704266, + 0.014310660772025585, + 0.00674151536077261, + 0.023716792464256287, + 0.03654298186302185, + 0.0071074580773711205, + 0.0067482152953743935, + 0.01977633871138096, + 0.008671116083860397, + 0.003912022337317467, + 0.002655231161043048, + 0.0030859734397381544, + 0.0035809590481221676, + 0.0029665089678019285, + 0.0038907763082534075, + 0.005726313684135675, + 0.003751014359295368, + 0.0063847885467112064, + 0.006836935877799988, + 0.00684399763122201, + 0.015056283213198185, + 0.0049587432295084, + 0.00963580422103405, + 0.012035883031785488, + 0.008009890094399452, + 0.0037613105960190296, + 0.005948373582214117, + 0.008769790641963482, + 0.010110907256603241, + 0.010502614080905914, + 0.00803221482783556, + 0.01232055388391018, + 0.011803493835031986, + 0.002452028449624777, + 0.0031325530726462603, + 0.002585147274658084, + 0.007829451933503151, + 0.00286939088255167, + 0.0019339679274708033, + 0.007677271030843258, + 0.0046191527508199215, + 0.002420382807031274, + 0.0026965655852109194, + 0.006922060623764992, + 0.0071219163946807384, + 0.003478333353996277, + 0.0032873114105314016, + 0.006714871618896723, + 0.0064208474941551685, + 0.0062492405995726585, + 0.0051697841845452785, + 0.006715948693454266, + 0.1019236370921135, + 0.010380348190665245, + 0.006256010849028826, + 0.17135371267795563, + 0.06738124787807465, + 0.0037932167761027813, + 0.003884236328303814, + 0.005902469158172607, + 0.02048727683722973, + 0.003963861148804426, + 0.0037282234989106655, + 0.0070940363220870495, + 0.015633054077625275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.00615024846047163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023324759677052498 + ], + "opposite_logprob_tokens": [ + -3.773324728012085 + ], + "logprob_reward": -0.023324759677052498, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023324759677052498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06167091056704521, + 0.09912224113941193, + 0.06357826292514801, + 0.02682727947831154, + 0.042990125715732574, + 0.04142697900533676, + 0.10459867864847183, + 0.017990726977586746, + 0.4782983958721161, + 0.06349639594554901 + ], + "credit_map": [ + 0.005888334475457668, + 0.0037879447918385267, + 0.010882052592933178, + 0.00629077386111021, + 0.006190568208694458, + 0.016564223915338516, + 0.0063002705574035645, + 0.005766748450696468, + 0.009323900565505028, + 0.010756656527519226, + 0.025429829955101013, + 0.013325847685337067, + 0.007559945806860924, + 0.014209965243935585, + 0.00920515414327383, + 0.009310944005846977, + 0.006799202412366867, + 0.004601527005434036, + 0.013309258036315441, + 0.014667218551039696, + 0.004163904581218958, + 0.0038750413805246353, + 0.009799795225262642, + 0.006362317129969597, + 0.003299108473584056, + 0.0037366566248238087, + 0.00281663634814322, + 0.003115519415587187, + 0.003680251073092222, + 0.004468333441764116, + 0.0029090503230690956, + 0.0028017237782478333, + 0.005589499603956938, + 0.006150657776743174, + 0.0047873686999082565, + 0.007205839268863201, + 0.004755810368806124, + 0.004823150113224983, + 0.005868680775165558, + 0.0038091205060482025, + 0.0026319357566535473, + 0.0039070830680429935, + 0.00534403882920742, + 0.007660864386707544, + 0.0031557613983750343, + 0.0037881103344261646, + 0.00824280921369791, + 0.00669637368991971, + 0.005863317288458347, + 0.008258622139692307, + 0.012384421192109585, + 0.019918479025363922, + 0.01068229041993618, + 0.010394182056188583, + 0.019023949280381203, + 0.018073417246341705, + 0.001084645395167172, + 0.0014538101386278868, + 0.0022676941007375717, + 0.003612212371081114, + 0.0018712105229496956, + 0.0015319942031055689, + 0.0028649333398789167, + 0.0033042270224541426, + 0.00792668852955103, + 0.006634247023612261, + 0.009042127057909966, + 0.12467813491821289, + 0.01374413538724184, + 0.008295123465359211, + 0.21711291372776031, + 0.09086500853300095, + 0.003330250270664692, + 0.0030355346389114857, + 0.008834424428641796, + 0.018920963630080223, + 0.0030829389579594135, + 0.0030954780522733927, + 0.008987904526293278, + 0.014208899810910225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.023324759677052498 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486226797103882 + ], + "opposite_logprob_tokens": [ + -1.2236226797103882 + ], + "logprob_reward": -0.3486226797103882, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486226797103882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2690407931804657, + 0.03918089717626572, + 0.17085419595241547, + 0.06015743687748909, + 0.04660340026021004, + 0.06604821234941483, + 0.06856154650449753, + 0.031733445823192596, + 0.21398457884788513, + 0.03383548557758331 + ], + "credit_map": [ + 0.019869990646839142, + 0.019478624686598778, + 0.061020951718091965, + 0.017632272094488144, + 0.023875607177615166, + 0.0940220057964325, + 0.01571427285671234, + 0.017427075654268265, + 0.0036826340947300196, + 0.0049756173975765705, + 0.005406172014772892, + 0.0054720924235880375, + 0.003966003190726042, + 0.006041651125997305, + 0.005799147766083479, + 0.003837576135993004, + 0.021600309759378433, + 0.014310069382190704, + 0.020184936001896858, + 0.056092072278261185, + 0.011607981286942959, + 0.008967451751232147, + 0.024253832176327705, + 0.013837547972798347, + 0.007643380202353001, + 0.005777703132480383, + 0.009142596274614334, + 0.007164692506194115, + 0.008513377048075199, + 0.011946945451200008, + 0.005397569853812456, + 0.004571173805743456, + 0.004898092709481716, + 0.007590517867356539, + 0.005393304396420717, + 0.007875028066337109, + 0.0031180628575384617, + 0.006919919978827238, + 0.00534471683204174, + 0.005463761277496815, + 0.005198608618229628, + 0.008340511471033096, + 0.006938749924302101, + 0.01063120923936367, + 0.00465488201007247, + 0.006278590764850378, + 0.014591400511562824, + 0.009414257481694221, + 0.004508829675614834, + 0.006160809192806482, + 0.00707660149782896, + 0.012705632485449314, + 0.009181453846395016, + 0.00619248952716589, + 0.011295261792838573, + 0.01144046988338232, + 0.002052709925919771, + 0.002587595023214817, + 0.005727611016482115, + 0.005289058201014996, + 0.0032682723831385374, + 0.0036915582604706287, + 0.004916325677186251, + 0.004200316965579987, + 0.004066754598170519, + 0.006768484599888325, + 0.006092814728617668, + 0.08324369788169861, + 0.006245960481464863, + 0.0033027625177055597, + 0.09081540256738663, + 0.013448701240122318, + 0.0032406568061560392, + 0.002610857831314206, + 0.00645155506208539, + 0.00476801535114646, + 0.002741399221122265, + 0.0030836856458336115, + 0.005836693570017815, + 0.005102626513689756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.3486226797103882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03646693751215935, + 0.020106816664338112, + 0.03692442923784256, + 0.01721302606165409, + 0.03905032202601433, + 0.07745502144098282, + 0.03734242543578148, + 0.036061614751815796, + 0.6354238986968994, + 0.06395551562309265 + ], + "credit_map": [ + 0.0035855956375598907, + 0.00592042226344347, + 0.006684770341962576, + 0.0027954624965786934, + 0.0035461997613310814, + 0.008513135835528374, + 0.00248932302929461, + 0.0029320300091058016, + 0.0025090298149734735, + 0.002506040036678314, + 0.002253888640552759, + 0.002670096233487129, + 0.0024985570926219225, + 0.0033540783915668726, + 0.002024526009336114, + 0.002290600212290883, + 0.003930914215743542, + 0.002004980808123946, + 0.00781218009069562, + 0.009318297728896141, + 0.002966357860714197, + 0.0027627120725810528, + 0.004948002751916647, + 0.003180984640493989, + 0.0019521857611835003, + 0.002099375007674098, + 0.00210378784686327, + 0.0019041813211515546, + 0.0016640366520732641, + 0.002806401113048196, + 0.002569871721789241, + 0.0021131851244717836, + 0.005384235177189112, + 0.005468639545142651, + 0.005562816746532917, + 0.006615729071199894, + 0.00337391858920455, + 0.004506805445998907, + 0.004976996220648289, + 0.0031611830927431583, + 0.006374991964548826, + 0.011120329611003399, + 0.00858349446207285, + 0.012366503477096558, + 0.005996723659336567, + 0.010166874155402184, + 0.012694209814071655, + 0.01015189103782177, + 0.002615611767396331, + 0.003101465292274952, + 0.0050101736560463905, + 0.006884592119604349, + 0.0035757285077124834, + 0.003458817722275853, + 0.0066938078962266445, + 0.006002228707075119, + 0.002696065930649638, + 0.0024201979395002127, + 0.003993952181190252, + 0.00501984404399991, + 0.005216649267822504, + 0.006136354524642229, + 0.005927488673478365, + 0.0046510593965649605, + 0.01135857030749321, + 0.009515887126326561, + 0.01183977909386158, + 0.17108692228794098, + 0.017565729096531868, + 0.01056080311536789, + 0.28895747661590576, + 0.11453872919082642, + 0.003917119465768337, + 0.0037150182761251926, + 0.00904939230531454, + 0.013450701721012592, + 0.00473644258454442, + 0.005225751549005508, + 0.008200480602681637, + 0.015660610049962997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.023283764719963074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic convertible car with cherry-red paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car have cream-colored leather interior and large, slim-rimmed steering wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving on a cobblestone road at a town street corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background buildings red-brick facades with striped canvas awnings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle and tracking smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181685358285904 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181685358285904, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181685358285904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16143742203712463, + 0.05357334017753601, + 0.17649102210998535, + 0.07647259533405304, + 0.059972960501909256, + 0.07964245229959488, + 0.058248378336429596, + 0.043915241956710815, + 0.21619930863380432, + 0.0740472823381424 + ], + "credit_map": [ + 0.011796404607594013, + 0.014729726128280163, + 0.03303849324584007, + 0.013421587646007538, + 0.015639927238225937, + 0.050073713064193726, + 0.01195419579744339, + 0.01078338548541069, + 0.004842968191951513, + 0.00667283171787858, + 0.007733295671641827, + 0.007848627865314484, + 0.006543847266584635, + 0.006363933905959129, + 0.008477253839373589, + 0.005090582184493542, + 0.019502468407154083, + 0.01039221789687872, + 0.03010733239352703, + 0.04819445312023163, + 0.012240610085427761, + 0.01056499034166336, + 0.02808198891580105, + 0.01740696094930172, + 0.007785870227962732, + 0.006324263755232096, + 0.010237820446491241, + 0.009693372063338757, + 0.006321801338344812, + 0.008282230235636234, + 0.013054347597062588, + 0.01477289292961359, + 0.0070524755865335464, + 0.009271465241909027, + 0.008250693790614605, + 0.009832033887505531, + 0.005193878896534443, + 0.007571606896817684, + 0.0070257484912872314, + 0.005775061435997486, + 0.005502747837454081, + 0.010671230033040047, + 0.009598619304597378, + 0.012006656266748905, + 0.00879887118935585, + 0.008820720948278904, + 0.012634912505745888, + 0.011608699336647987, + 0.005084269680082798, + 0.005811083596199751, + 0.006992196198552847, + 0.008613732643425465, + 0.005558065604418516, + 0.006199145223945379, + 0.008882464841008186, + 0.01110741775482893, + 0.002385383937507868, + 0.002551319543272257, + 0.007057019509375095, + 0.010612105950713158, + 0.0033848504535853863, + 0.003861259436234832, + 0.008184638805687428, + 0.005878663156181574, + 0.0035583488643169403, + 0.0037371255457401276, + 0.0045442427508533, + 0.06328105181455612, + 0.006150044500827789, + 0.003478090511634946, + 0.09769288450479507, + 0.033757537603378296, + 0.00486449059098959, + 0.004416948184370995, + 0.013461308553814888, + 0.01658746972680092, + 0.004530290141701698, + 0.0038118897937238216, + 0.013103689067065716, + 0.013271195814013481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.018181685358285904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_tensor([3934, 3934])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_tensor([3934, 3934])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3fa877b4fbcb9c63a629c270e54d5f96cc4b6ff8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_tensor([3934, 3934])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-moving close-up zooms in on the broad, rounded armrests of a weathered, old rattan chair. Warm afternoon sunlight slants down, capturing the rough, dry, yellowed texture of the rattan and the details of its flaking varnish. Against this backdrop, marked by the marks of time, a fresh, emerald green sprout stubbornly emerges from a crack in the cane. Its spiraling vines, curling at the top, carefully unfurl in circles at a speed barely perceptible to the naked eye. Sunlight filters through its translucent leaves, casting a soft golden halo across the fine hairs covering them. As the vine slowly unfurls, the fresh green seems to engage in a silent, vibrant rhythm. Against the backdrop of blurred garden greenery and dappled light, the warm tones of the entire image create a poetic atmosphere of profound tranquility, hope, and healing, silently narrating the delicate interplay of decay and rebirth.", + "seed": 20791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_3934_seed_20791_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04314034283161163, + "vlm_logprob": -0.04314034283161163, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04314034283161163, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07400407642126083 + ], + "opposite_logprob_tokens": [ + -2.6990041732788086 + ], + "logprob_reward": -0.07400407642126083, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07400407642126083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17127932608127594, + 0.134047269821167, + 0.11738939583301544, + 0.0691354051232338, + 0.04967731609940529, + 0.07228328287601471, + 0.06836370378732681, + 0.04798886179924011, + 0.07231676578521729, + 0.1975187063217163 + ], + "credit_map": [ + 0.02099682204425335, + 0.017900796607136726, + 0.012913355603814125, + 0.016127722337841988, + 0.037685878574848175, + 0.020612547174096107, + 0.01862877793610096, + 0.026413418352603912, + 0.020969441160559654, + 0.01265824493020773, + 0.018964890390634537, + 0.008609042502939701, + 0.021330352872610092, + 0.01458614319562912, + 0.023696912452578545, + 0.013232233002781868, + 0.015686599537730217, + 0.008814786560833454, + 0.012100795283913612, + 0.016500120982527733, + 0.02759629674255848, + 0.014727464877068996, + 0.01115020364522934, + 0.010813126340508461, + 0.00953941885381937, + 0.008045980706810951, + 0.005620415322482586, + 0.0037119488697499037, + 0.016142848879098892, + 0.006110477261245251, + 0.010157953016459942, + 0.009806366637349129, + 0.008151397109031677, + 0.0056289006024599075, + 0.006632328499108553, + 0.004381870850920677, + 0.005399638786911964, + 0.007774893660098314, + 0.004933365620672703, + 0.006774918641895056, + 0.010219267569482327, + 0.007684736978262663, + 0.008964389562606812, + 0.008233080618083477, + 0.00793625507503748, + 0.008931910619139671, + 0.00890639703720808, + 0.011407246813178062, + 0.008367928676307201, + 0.007815255783498287, + 0.005436967592686415, + 0.005960080306977034, + 0.010952319949865341, + 0.011687599122524261, + 0.011628282256424427, + 0.006515270099043846, + 0.008277837187051773, + 0.0069389427080750465, + 0.003740220097824931, + 0.003906792961061001, + 0.009078855626285076, + 0.004973478615283966, + 0.005640124436467886, + 0.005432612728327513, + 0.012436534278094769, + 0.009571930393576622, + 0.010442323982715607, + 0.005292050540447235, + 0.009896349161863327, + 0.008387352339923382, + 0.008489429019391537, + 0.007800794206559658, + 0.03854460269212723, + 0.03907919302582741, + 0.033162590116262436, + 0.02190803363919258, + 0.013926740735769272, + 0.016604620963335037, + 0.01518887560814619, + 0.01910405606031418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.07400407642126083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330612577497959 + ], + "opposite_logprob_tokens": [ + -3.773306131362915 + ], + "logprob_reward": -0.02330612577497959, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330612577497959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18558244407176971, + 0.09891856461763382, + 0.14880113303661346, + 0.08306550234556198, + 0.09608405083417892, + 0.04229356348514557, + 0.02808307483792305, + 0.022080296650528908, + 0.03917974233627319, + 0.25591158866882324 + ], + "credit_map": [ + 0.023592662066221237, + 0.019782833755016327, + 0.024629196152091026, + 0.015596139244735241, + 0.026344917714595795, + 0.021225932985544205, + 0.024295341223478317, + 0.030115412548184395, + 0.009976077824831009, + 0.011072217486798763, + 0.011708934791386127, + 0.006306662689894438, + 0.016533853486180305, + 0.02042698673903942, + 0.011169486679136753, + 0.011724352836608887, + 0.02203923463821411, + 0.010486427694559097, + 0.01713179051876068, + 0.03328673914074898, + 0.014951453544199467, + 0.01524675264954567, + 0.017701977863907814, + 0.017956756055355072, + 0.011419642716646194, + 0.008094889111816883, + 0.005635524168610573, + 0.005677493289113045, + 0.010977666825056076, + 0.008544591255486012, + 0.017995405942201614, + 0.014720290899276733, + 0.013093702495098114, + 0.011968373320996761, + 0.009153083898127079, + 0.01021785382181406, + 0.009902308695018291, + 0.01887326128780842, + 0.011640598066151142, + 0.011234868317842484, + 0.004835500381886959, + 0.004631449934095144, + 0.005233624950051308, + 0.005886191036552191, + 0.0038370881229639053, + 0.0033722748048603535, + 0.007501480169594288, + 0.006995953619480133, + 0.004041377920657396, + 0.0031781222205609083, + 0.0029717055149376392, + 0.002851753728464246, + 0.0048806932754814625, + 0.0033056798856705427, + 0.0034197999630123377, + 0.0034339413978159428, + 0.003353348234668374, + 0.0023863676469773054, + 0.002131994115188718, + 0.0022560968063771725, + 0.004617752972990274, + 0.0023388590198010206, + 0.002744710771366954, + 0.0022511675488203764, + 0.005400231573730707, + 0.007246289402246475, + 0.0034811634104698896, + 0.0037558148615062237, + 0.005412181839346886, + 0.0029502331744879484, + 0.003881816752254963, + 0.007052013650536537, + 0.059096626937389374, + 0.05881664901971817, + 0.04840288683772087, + 0.02220067009329796, + 0.018183711916208267, + 0.01861121878027916, + 0.015549744479358196, + 0.015050085261464119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.02330612577497959 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013383012264966965 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013383012264966965, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013383012264966965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18589134514331818, + 0.06080123037099838, + 0.06294958293437958, + 0.07518621534109116, + 0.09666294604539871, + 0.12536758184432983, + 0.05050479248166084, + 0.06207254156470299, + 0.1431819200515747, + 0.13738176226615906 + ], + "credit_map": [ + 0.021961607038974762, + 0.018531320616602898, + 0.015052270144224167, + 0.014045050367712975, + 0.044079169631004333, + 0.026977721601724625, + 0.018566923215985298, + 0.02667730674147606, + 0.01025508251041174, + 0.006454435642808676, + 0.00809094961732626, + 0.004726026207208633, + 0.009873601607978344, + 0.007212836295366287, + 0.0060502998530864716, + 0.008138006553053856, + 0.008057629689574242, + 0.006543626543134451, + 0.006835147738456726, + 0.007740332745015621, + 0.014163072220981121, + 0.007098156958818436, + 0.006394971627742052, + 0.006116653326898813, + 0.008030545897781849, + 0.006706644780933857, + 0.005117273423820734, + 0.004718642216175795, + 0.01668742299079895, + 0.006983555853366852, + 0.013880745507776737, + 0.01306139212101698, + 0.023634979501366615, + 0.015466882847249508, + 0.007119405549019575, + 0.00646296888589859, + 0.008109349757432938, + 0.008562332019209862, + 0.00853695161640644, + 0.018770083785057068, + 0.01125167403370142, + 0.008067434653639793, + 0.009397110901772976, + 0.009408515878021717, + 0.01614205911755562, + 0.01126157958060503, + 0.03498174250125885, + 0.024857478216290474, + 0.007119385059922934, + 0.00673317164182663, + 0.0045173210091888905, + 0.004833416547626257, + 0.008516390807926655, + 0.006432401482015848, + 0.006807912141084671, + 0.005544800311326981, + 0.011213287711143494, + 0.009747701697051525, + 0.004554544575512409, + 0.004748832434415817, + 0.013014867901802063, + 0.004900800995528698, + 0.007430801633745432, + 0.006461712997406721, + 0.026771951466798782, + 0.02897294983267784, + 0.01810697838664055, + 0.012635603547096252, + 0.009140417911112309, + 0.010059063322842121, + 0.011352376081049442, + 0.026142599061131477, + 0.02586446702480316, + 0.02816907875239849, + 0.02191115915775299, + 0.012678501196205616, + 0.01040175836533308, + 0.012128590606153011, + 0.01201778743416071, + 0.014210434630513191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0013383012264966965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380990207195282 + ], + "opposite_logprob_tokens": [ + -3.2880990505218506 + ], + "logprob_reward": -0.0380990207195282, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380990207195282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06844677031040192, + 0.04045403003692627, + 0.0455472394824028, + 0.05913067236542702, + 0.03431405499577522, + 0.02828625589609146, + 0.07136912643909454, + 0.06048211455345154, + 0.24609260261058807, + 0.3458770513534546 + ], + "credit_map": [ + 0.010715873911976814, + 0.008043359965085983, + 0.006309292279183865, + 0.00561283016577363, + 0.018368031829595566, + 0.006192927248775959, + 0.005933274514973164, + 0.0072711799293756485, + 0.004356209188699722, + 0.004171113017946482, + 0.00565622141584754, + 0.003816546406596899, + 0.006709813140332699, + 0.004637698642909527, + 0.006389301735907793, + 0.004717125557363033, + 0.006371686235070229, + 0.005433062557131052, + 0.005252803675830364, + 0.0034639856312423944, + 0.00986407045274973, + 0.005603121593594551, + 0.004770016297698021, + 0.004788495134562254, + 0.006084555760025978, + 0.004209564533084631, + 0.0036242350470274687, + 0.0032779520843178034, + 0.009284882806241512, + 0.004052672069519758, + 0.015786925330758095, + 0.012809887528419495, + 0.0041330852545797825, + 0.0033636472653597593, + 0.00290108029730618, + 0.002968923654407263, + 0.00592202739790082, + 0.006397582124918699, + 0.00395608926191926, + 0.00467161787673831, + 0.0030747822020202875, + 0.002776938723400235, + 0.002681533806025982, + 0.0021461546421051025, + 0.005504004657268524, + 0.002841687761247158, + 0.005720968823879957, + 0.0035401852801442146, + 0.009915183298289776, + 0.009060680866241455, + 0.005147166084498167, + 0.006729915738105774, + 0.017717275768518448, + 0.006053530611097813, + 0.010033936239778996, + 0.0067114392295479774, + 0.009243819862604141, + 0.00803244300186634, + 0.005600460804998875, + 0.0040001473389565945, + 0.015788476914167404, + 0.005801931023597717, + 0.007246130146086216, + 0.004768704529851675, + 0.02874838560819626, + 0.031049048528075218, + 0.035617463290691376, + 0.02821587212383747, + 0.018175708130002022, + 0.0201504398137331, + 0.024714110419154167, + 0.05942158028483391, + 0.07258930802345276, + 0.07114414125680923, + 0.06161852553486824, + 0.03132932260632515, + 0.025972170755267143, + 0.026944367215037346, + 0.025218211114406586, + 0.03106103278696537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0380990207195282 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895419001579285 + ], + "opposite_logprob_tokens": [ + -2.5789542198181152 + ], + "logprob_reward": -0.07895419001579285, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895419001579285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056757695972919464, + 0.04735179990530014, + 0.041376613080501556, + 0.02161664515733719, + 0.030921630561351776, + 0.03491911664605141, + 0.033960629254579544, + 0.02145073562860489, + 0.0923537015914917, + 0.6192914247512817 + ], + "credit_map": [ + 0.007798242848366499, + 0.007015572860836983, + 0.005400195252150297, + 0.004316885955631733, + 0.01170248631387949, + 0.0066777849569916725, + 0.006570930127054453, + 0.0072756060399115086, + 0.005175835452973843, + 0.005694158375263214, + 0.006664656102657318, + 0.0037335625384002924, + 0.010434224270284176, + 0.006677545141428709, + 0.004498750902712345, + 0.0044730715453624725, + 0.004569779150187969, + 0.0036597184371203184, + 0.0037825696635991335, + 0.004537272732704878, + 0.008658081293106079, + 0.007876892574131489, + 0.003928061109036207, + 0.004364242311567068, + 0.0027005840092897415, + 0.002348862122744322, + 0.0028838978614658117, + 0.0014560646377503872, + 0.003484902437776327, + 0.0039061622228473425, + 0.0019764446187764406, + 0.0028597277123481035, + 0.006163740064948797, + 0.00458099227398634, + 0.0026097397785633802, + 0.0020632585510611534, + 0.0025513621512800455, + 0.004639597609639168, + 0.0026921741664409637, + 0.0056207687593996525, + 0.003758826991543174, + 0.0035047815181314945, + 0.0036310534924268723, + 0.003376204986125231, + 0.004088340327143669, + 0.0036515556275844574, + 0.008424468338489532, + 0.004483887925744057, + 0.003888311330229044, + 0.0029653545934706926, + 0.0032356041483581066, + 0.004112090449780226, + 0.004087126813828945, + 0.005558617413043976, + 0.005426619201898575, + 0.00468690600246191, + 0.002672291360795498, + 0.0024905637837946415, + 0.002652555936947465, + 0.0029015876352787018, + 0.002343337517231703, + 0.0026780252810567617, + 0.0028029377572238445, + 0.0029094384517520666, + 0.019890863448381424, + 0.01889897510409355, + 0.011288587003946304, + 0.00725597282871604, + 0.007144682575017214, + 0.009898337535560131, + 0.008723879233002663, + 0.009252412244677544, + 0.11857184022665024, + 0.11963007599115372, + 0.10042623430490494, + 0.0633339211344719, + 0.049750201404094696, + 0.054743822664022446, + 0.0497891828417778, + 0.0630461722612381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07895419001579285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-moving close-up zooms in on the broad, rounded armrests of a weathered, old rattan chair. Warm afternoon sunlight slants down, capturing the rough, dry, yellowed texture of the rattan and the details of its flaking varnish. Against this backdrop, marked by the marks of time, a fresh, emerald green sprout stubbornly emerges from a crack in the cane. Its spiraling vines, curling at the top, carefully unfurl in circles at a speed barely perceptible to the naked eye. Sunlight filters through its translucent leaves, casting a soft golden halo across the fine hairs covering them. As the vine slowly unfurls, the fresh green seems to engage in a silent, vibrant rhythm. Against the backdrop of blurred garden greenery and dappled light, the warm tones of the entire image create a poetic atmosphere of profound tranquility, hope, and healing, silently narrating the delicate interplay of decay and rebirth.", + "seed": 20792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_3934_seed_20792_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026218351884745063, + "vlm_logprob": -0.026218351884745063, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.026218351884745063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04216070473194122 + ], + "opposite_logprob_tokens": [ + -3.292160749435425 + ], + "logprob_reward": -0.04216070473194122, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04216070473194122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12909653782844543, + 0.10459254682064056, + 0.11059721559286118, + 0.05361847206950188, + 0.08127260208129883, + 0.07555051892995834, + 0.04166645556688309, + 0.12235474586486816, + 0.11743450164794922, + 0.1638164073228836 + ], + "credit_map": [ + 0.011812659911811352, + 0.011551334522664547, + 0.013062608428299427, + 0.022040676325559616, + 0.017394080758094788, + 0.01650402694940567, + 0.013881804421544075, + 0.022849345579743385, + 0.012178574688732624, + 0.017068875953555107, + 0.012854466214776039, + 0.012500773184001446, + 0.012062568217515945, + 0.01711348257958889, + 0.009582032449543476, + 0.011231777258217335, + 0.006400199607014656, + 0.00651310570538044, + 0.01565677300095558, + 0.007592009846121073, + 0.007460345979779959, + 0.032640233635902405, + 0.014894654974341393, + 0.01943989470601082, + 0.0036447830498218536, + 0.003912948537617922, + 0.007381987292319536, + 0.01166649628430605, + 0.006486779078841209, + 0.004003534093499184, + 0.007328883279114962, + 0.009193064644932747, + 0.006168080493807793, + 0.007598051335662603, + 0.01139088999480009, + 0.011637578718364239, + 0.011608977802097797, + 0.006799266207963228, + 0.012723146006464958, + 0.013346614316105843, + 0.006997515447437763, + 0.006364250089973211, + 0.01763113960623741, + 0.0083014452829957, + 0.006194000598043203, + 0.008593851700425148, + 0.007782299537211657, + 0.013686015270650387, + 0.0031424211338162422, + 0.004882885608822107, + 0.006082691717892885, + 0.0050324429757893085, + 0.004871912766247988, + 0.004953304771333933, + 0.0058807791210711, + 0.006820018403232098, + 0.009800512343645096, + 0.007228131406009197, + 0.03021295741200447, + 0.009221618995070457, + 0.009700593538582325, + 0.0318068191409111, + 0.010447559878230095, + 0.013936554081737995, + 0.01089018490165472, + 0.008401432074606419, + 0.013160165399312973, + 0.009110433980822563, + 0.007749214768409729, + 0.03362337127327919, + 0.011320926249027252, + 0.02317877858877182, + 0.014809371903538704, + 0.016039913520216942, + 0.020363325253129005, + 0.01542005967348814, + 0.014473071321845055, + 0.0389992892742157, + 0.02376718819141388, + 0.019944170489907265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.04216070473194122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597489550709724 + ], + "opposite_logprob_tokens": [ + -3.895597457885742 + ], + "logprob_reward": -0.020597489550709724, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597489550709724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13412560522556305, + 0.09104951471090317, + 0.09445655345916748, + 0.23553727567195892, + 0.07962295413017273, + 0.12871581315994263, + 0.04074607044458389, + 0.0649413987994194, + 0.05508517473936081, + 0.07571954280138016 + ], + "credit_map": [ + 0.012540156953036785, + 0.012806268408894539, + 0.016694195568561554, + 0.024950560182332993, + 0.01889914833009243, + 0.016293836757540703, + 0.01507799793034792, + 0.016863450407981873, + 0.009500681422650814, + 0.014376544393599033, + 0.00895724631845951, + 0.010776174254715443, + 0.01317948941141367, + 0.01180585753172636, + 0.01183168776333332, + 0.010621841065585613, + 0.003968086559325457, + 0.0032288976944983006, + 0.011853464879095554, + 0.00586716877296567, + 0.004454425070434809, + 0.01147608831524849, + 0.01871538534760475, + 0.03489305451512337, + 0.024227503687143326, + 0.011553077958524227, + 0.02622993104159832, + 0.09408886730670929, + 0.01050341222435236, + 0.010739538818597794, + 0.021986186504364014, + 0.03620879352092743, + 0.006439202465116978, + 0.00748950382694602, + 0.013212213292717934, + 0.011719297617673874, + 0.0080491304397583, + 0.007245121989399195, + 0.011331894434988499, + 0.014136605896055698, + 0.009277614764869213, + 0.011014869436621666, + 0.025129010900855064, + 0.013086601160466671, + 0.009568258188664913, + 0.020022641867399216, + 0.01573309674859047, + 0.024883747100830078, + 0.0033475286327302456, + 0.005262534134089947, + 0.005487188696861267, + 0.004633474163711071, + 0.006243305280804634, + 0.0041385190561413765, + 0.006621994078159332, + 0.005011532921344042, + 0.005211520474404097, + 0.003456032369285822, + 0.015413876622915268, + 0.004058533813804388, + 0.004736823961138725, + 0.01699162647128105, + 0.0073459567502141, + 0.007727038115262985, + 0.005284393671900034, + 0.003987000323832035, + 0.005564932245761156, + 0.00399433309212327, + 0.003939324989914894, + 0.01660904288291931, + 0.005950673017650843, + 0.009755481965839863, + 0.008127433247864246, + 0.006053457502275705, + 0.009338789619505405, + 0.006170867010951042, + 0.006740586366504431, + 0.020505284890532494, + 0.008767920546233654, + 0.010015210136771202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.020597489550709724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015151937259361148 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.0015151937259361148, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015151937259361148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09543148428201675, + 0.10007038712501526, + 0.20740818977355957, + 0.11000277101993561, + 0.07270238548517227, + 0.07747740298509598, + 0.07052415609359741, + 0.0709729716181755, + 0.10339757055044174, + 0.09201270341873169 + ], + "credit_map": [ + 0.00994375254958868, + 0.006608248222619295, + 0.011800261214375496, + 0.013941163197159767, + 0.008577841334044933, + 0.011609457433223724, + 0.01232015062123537, + 0.0206306092441082, + 0.009283299557864666, + 0.016150139272212982, + 0.011359211057424545, + 0.011487060226500034, + 0.013200871646404266, + 0.01431115623563528, + 0.01173869613558054, + 0.012539952993392944, + 0.007774236146360636, + 0.008655480109155178, + 0.020115040242671967, + 0.011191311292350292, + 0.010074545629322529, + 0.03837108239531517, + 0.03686235100030899, + 0.07436414062976837, + 0.010359489358961582, + 0.0054022399708628654, + 0.012202277779579163, + 0.03577136620879173, + 0.009142755530774593, + 0.006610178388655186, + 0.01520414650440216, + 0.01531031634658575, + 0.005425043869763613, + 0.005786335561424494, + 0.010097951628267765, + 0.009837775491178036, + 0.009925302118062973, + 0.008553140796720982, + 0.012394331395626068, + 0.010682506486773491, + 0.005824564956128597, + 0.005380657035857439, + 0.013656726107001305, + 0.007373909465968609, + 0.008733518421649933, + 0.015723422169685364, + 0.011778346262872219, + 0.009006260894238949, + 0.004765357822179794, + 0.007302932906895876, + 0.012014167383313179, + 0.006863805465400219, + 0.008284018374979496, + 0.01730058528482914, + 0.008055866695940495, + 0.005937427747994661, + 0.005241000093519688, + 0.004230342339724302, + 0.018226107582449913, + 0.003914894536137581, + 0.005162396002560854, + 0.021645337343215942, + 0.005469413939863443, + 0.00708347512409091, + 0.006959246005862951, + 0.008470669388771057, + 0.009636479429900646, + 0.006909938063472509, + 0.011221611872315407, + 0.028719687834382057, + 0.015585565008223057, + 0.015894372016191483, + 0.008305462077260017, + 0.009249493479728699, + 0.010458751581609249, + 0.009667277336120605, + 0.011359846219420433, + 0.012589504010975361, + 0.021062321960926056, + 0.009320051409304142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.0015151937259361148 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818414404988289 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.01818414404988289, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818414404988289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10473104566335678, + 0.1224370226264, + 0.18800951540470123, + 0.05725216865539551, + 0.15006902813911438, + 0.05825486406683922, + 0.08310367912054062, + 0.08230054378509521, + 0.09138373285531998, + 0.06245839595794678 + ], + "credit_map": [ + 0.007999550551176071, + 0.007891930639743805, + 0.011216488666832447, + 0.016838936135172844, + 0.012461099773645401, + 0.01255486998707056, + 0.017066244035959244, + 0.018701929599046707, + 0.01306906621903181, + 0.02505820244550705, + 0.013817536644637585, + 0.013857688754796982, + 0.014766249805688858, + 0.017877675592899323, + 0.011572814546525478, + 0.012417788617312908, + 0.005465433932840824, + 0.008285585790872574, + 0.009358991868793964, + 0.009932976216077805, + 0.010779371485114098, + 0.04127682372927666, + 0.03828248381614685, + 0.06462785601615906, + 0.004438236355781555, + 0.0031995547469705343, + 0.011593352071940899, + 0.011761910282075405, + 0.004529000259935856, + 0.004618694540113211, + 0.007175465580075979, + 0.009935949929058552, + 0.01205440517514944, + 0.010256139561533928, + 0.02252667210996151, + 0.021437954157590866, + 0.014979715459048748, + 0.019143685698509216, + 0.02842801809310913, + 0.02124243974685669, + 0.004553648643195629, + 0.004620359744876623, + 0.013241336680948734, + 0.006845718715339899, + 0.005067376419901848, + 0.010110592469573021, + 0.00713686877861619, + 0.006678964477032423, + 0.006177273113280535, + 0.0082765007391572, + 0.015422617085278034, + 0.007749812677502632, + 0.009177092462778091, + 0.01831207424402237, + 0.011101949028670788, + 0.006886362098157406, + 0.005651206709444523, + 0.0058537754230201244, + 0.015482455492019653, + 0.006048629991710186, + 0.006988717708736658, + 0.022242000326514244, + 0.010259537026286125, + 0.009774224832654, + 0.008025862276554108, + 0.006389013025909662, + 0.010892268270254135, + 0.010584216564893723, + 0.007826069369912148, + 0.02348465286195278, + 0.010479573160409927, + 0.013702076859772205, + 0.006934792269021273, + 0.004222256597131491, + 0.008966892026364803, + 0.006369893904775381, + 0.007435212377458811, + 0.009087849408388138, + 0.0113228689879179, + 0.00811863038688898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.01818414404988289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048634227365255356 + ], + "opposite_logprob_tokens": [ + -3.0486342906951904 + ], + "logprob_reward": -0.048634227365255356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048634227365255356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20667606592178345, + 0.10744992643594742, + 0.04980956390500069, + 0.08499618619680405, + 0.05906441807746887, + 0.11748554557561874, + 0.1316893845796585, + 0.0766802430152893, + 0.03652534633874893, + 0.12962327897548676 + ], + "credit_map": [ + 0.011956513859331608, + 0.0092341722920537, + 0.019451120868325233, + 0.03920428827404976, + 0.017293540760874748, + 0.015340765938162804, + 0.029783865436911583, + 0.06441181898117065, + 0.010385327972471714, + 0.021144511178135872, + 0.014261565171182156, + 0.013424614444375038, + 0.01018051989376545, + 0.012995620258152485, + 0.01067162211984396, + 0.014386152848601341, + 0.0040772403590381145, + 0.0031852370593696833, + 0.007591376546770334, + 0.005351018626242876, + 0.0031396853737533092, + 0.010063906200230122, + 0.009551722556352615, + 0.006849383935332298, + 0.005893924739211798, + 0.004228147678077221, + 0.008398626931011677, + 0.0289349053055048, + 0.006494954228401184, + 0.007017675321549177, + 0.010819127783179283, + 0.013208831660449505, + 0.004064648877829313, + 0.004736577160656452, + 0.007425105664879084, + 0.005382941570132971, + 0.008473928086459637, + 0.010913239791989326, + 0.009337643161416054, + 0.008730335161089897, + 0.009460145607590675, + 0.008919036947190762, + 0.02178069017827511, + 0.01028496865183115, + 0.012846854515373707, + 0.02390723116695881, + 0.012751169502735138, + 0.017535461112856865, + 0.009192563593387604, + 0.016781801357865334, + 0.01932862401008606, + 0.018273187801241875, + 0.01633065566420555, + 0.011556723155081272, + 0.02016627974808216, + 0.020059553906321526, + 0.0047249277122318745, + 0.00568827148526907, + 0.01378796249628067, + 0.005808663088828325, + 0.0048926579765975475, + 0.019163843244314194, + 0.01011890172958374, + 0.012495024129748344, + 0.003046942874789238, + 0.0036595286801457405, + 0.0035933165345340967, + 0.0035648802295327187, + 0.005152543541043997, + 0.006546392105519772, + 0.005092354491353035, + 0.005869390908628702, + 0.010150153189897537, + 0.014471692033112049, + 0.021842947229743004, + 0.012910020537674427, + 0.011442081071436405, + 0.016466010361909866, + 0.01908125914633274, + 0.0232591200619936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.048634227365255356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-moving close-up zooms in on the broad, rounded armrests of a weathered, old rattan chair. Warm afternoon sunlight slants down, capturing the rough, dry, yellowed texture of the rattan and the details of its flaking varnish. Against this backdrop, marked by the marks of time, a fresh, emerald green sprout stubbornly emerges from a crack in the cane. Its spiraling vines, curling at the top, carefully unfurl in circles at a speed barely perceptible to the naked eye. Sunlight filters through its translucent leaves, casting a soft golden halo across the fine hairs covering them. As the vine slowly unfurls, the fresh green seems to engage in a silent, vibrant rhythm. Against the backdrop of blurred garden greenery and dappled light, the warm tones of the entire image create a poetic atmosphere of profound tranquility, hope, and healing, silently narrating the delicate interplay of decay and rebirth.", + "seed": 20793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_3934_seed_20793_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0397831815527752, + "vlm_logprob": -0.0397831815527752, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.0397831815527752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017382772639393806 + ], + "opposite_logprob_tokens": [ + -4.267382621765137 + ], + "logprob_reward": -0.017382772639393806, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017382772639393806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5258419513702393, + 0.050553008913993835, + 0.03527052327990532, + 0.03417574241757393, + 0.05150843784213066, + 0.07945573329925537, + 0.06634384393692017, + 0.02745787613093853, + 0.043968815356492996, + 0.08542411774396896 + ], + "credit_map": [ + 0.046624042093753815, + 0.06387343257665634, + 0.05298459902405739, + 0.07055608183145523, + 0.10770007967948914, + 0.07120000571012497, + 0.05029226094484329, + 0.06261144578456879, + 0.005164701957255602, + 0.00518882367759943, + 0.006196985021233559, + 0.006971503142267466, + 0.006560586858540773, + 0.008351529017090797, + 0.006617903243750334, + 0.005500977858901024, + 0.003907307051122189, + 0.003399440785869956, + 0.0032994109205901623, + 0.0034877927973866463, + 0.006581373978406191, + 0.004764163866639137, + 0.004391210153698921, + 0.0054398225620388985, + 0.003818424418568611, + 0.004057576414197683, + 0.0036979608703404665, + 0.0037217766512185335, + 0.00585660245269537, + 0.003567732172086835, + 0.004488090053200722, + 0.0049675810150802135, + 0.0047525460831820965, + 0.007230894640088081, + 0.008377454243600368, + 0.004674328491091728, + 0.005399541463702917, + 0.006527048069983721, + 0.007747408002614975, + 0.006799210794270039, + 0.008168155327439308, + 0.014015852473676205, + 0.007176820654422045, + 0.009098769165575504, + 0.006787124089896679, + 0.011545874178409576, + 0.011850456707179546, + 0.01081268209964037, + 0.01161874271929264, + 0.009434575214982033, + 0.007019360084086657, + 0.006453827489167452, + 0.006889518350362778, + 0.011187562718987465, + 0.007397793233394623, + 0.006342459470033646, + 0.002790812635794282, + 0.0029649960342794657, + 0.005339209921658039, + 0.0034588712733238935, + 0.0028063724748790264, + 0.003280245466157794, + 0.0032433229498565197, + 0.003574047703295946, + 0.004581125918775797, + 0.004045578185468912, + 0.006514526903629303, + 0.00937772449105978, + 0.005602874793112278, + 0.0029024723917245865, + 0.006008797790855169, + 0.004935713950544596, + 0.007384662516415119, + 0.007211967371404171, + 0.02322710119187832, + 0.016439709812402725, + 0.00701530696824193, + 0.005129680503159761, + 0.007991550490260124, + 0.011024137027561665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.017382772639393806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637230679392815 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007637230679392815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637230679392815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5394660234451294, + 0.05359940603375435, + 0.030656618997454643, + 0.02900955080986023, + 0.023380527272820473, + 0.050978295505046844, + 0.03117266483604908, + 0.053562313318252563, + 0.11312318593263626, + 0.07505141198635101 + ], + "credit_map": [ + 0.04882962256669998, + 0.06325556337833405, + 0.05692645534873009, + 0.07400993257761002, + 0.1012602150440216, + 0.08208947628736496, + 0.05224797874689102, + 0.060846809297800064, + 0.004345887340605259, + 0.00725441612303257, + 0.006576770916581154, + 0.006087779998779297, + 0.005612288601696491, + 0.008307195268571377, + 0.007606637664139271, + 0.007808431517332792, + 0.0023586228489875793, + 0.002965949010103941, + 0.0035003465600311756, + 0.003398388857021928, + 0.005041802302002907, + 0.004619463812559843, + 0.004761268384754658, + 0.004010776523500681, + 0.0030813729390501976, + 0.002371511422097683, + 0.0038758893497288227, + 0.0029780094046145678, + 0.0030096399132162333, + 0.004765043966472149, + 0.0046662925742566586, + 0.004261791706085205, + 0.002018791390582919, + 0.002475971356034279, + 0.002945856424048543, + 0.0025210233870893717, + 0.0025897566229104996, + 0.0034475005231797695, + 0.0042223683558404446, + 0.0031592599116265774, + 0.0049181729555130005, + 0.01057989988476038, + 0.006211207248270512, + 0.005862030666321516, + 0.0050286827608942986, + 0.0034450003877282143, + 0.008925709873437881, + 0.006007588002830744, + 0.00402394263073802, + 0.004052218981087208, + 0.004380456637591124, + 0.0032412211876362562, + 0.003134483704343438, + 0.0045553529635071754, + 0.003935651388019323, + 0.00384933827444911, + 0.00643136166036129, + 0.005980503745377064, + 0.008664567023515701, + 0.007288474589586258, + 0.004629609175026417, + 0.006662365049123764, + 0.006085741799324751, + 0.007819687016308308, + 0.010050327517092228, + 0.007031425833702087, + 0.020790601149201393, + 0.039335764944553375, + 0.00692874938249588, + 0.00764313293620944, + 0.011193385347723961, + 0.010149799287319183, + 0.008987389504909515, + 0.005354324821382761, + 0.018231645226478577, + 0.0075042336247861385, + 0.006715277209877968, + 0.005201707594096661, + 0.009555262513458729, + 0.013501572422683239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007637230679392815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001043133088387549 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.001043133088387549, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001043133088387549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5814639925956726, + 0.06558392196893692, + 0.05130351334810257, + 0.023286765441298485, + 0.04392298310995102, + 0.039321016520261765, + 0.016467951238155365, + 0.08031847327947617, + 0.051659081131219864, + 0.0466722771525383 + ], + "credit_map": [ + 0.05247991159558296, + 0.07072392851114273, + 0.06277360022068024, + 0.08506496995687485, + 0.10908935964107513, + 0.07753127068281174, + 0.05844348296523094, + 0.06535743921995163, + 0.005327275022864342, + 0.00961386039853096, + 0.007702517788857222, + 0.007957280613481998, + 0.009565279819071293, + 0.009835145436227322, + 0.007634421344846487, + 0.007948142476379871, + 0.004534895997494459, + 0.0069659072905778885, + 0.004392074421048164, + 0.006463515106588602, + 0.007859659381210804, + 0.009518304839730263, + 0.004829492885619402, + 0.006739665288478136, + 0.0029942388646304607, + 0.002595936879515648, + 0.002336250152438879, + 0.002573430771008134, + 0.003620426869019866, + 0.0036853833589702845, + 0.002709881402552128, + 0.0027712169103324413, + 0.0038262761663645506, + 0.0062065282836556435, + 0.004623653832823038, + 0.006188602186739445, + 0.007353891618549824, + 0.004419070202857256, + 0.005732150282710791, + 0.005572808440774679, + 0.0031080031767487526, + 0.0035699130967259407, + 0.00430928124114871, + 0.004513237159699202, + 0.004486772231757641, + 0.006187989376485348, + 0.007126610726118088, + 0.006019208580255508, + 0.0019954813178628683, + 0.0025088442489504814, + 0.0018857801333069801, + 0.0016150762094184756, + 0.00148956838529557, + 0.003916483372449875, + 0.0015690861037001014, + 0.0014876325149089098, + 0.009362432174384594, + 0.005109444260597229, + 0.007552754133939743, + 0.01457877829670906, + 0.004721644334495068, + 0.026818662881851196, + 0.005062890239059925, + 0.007111866492778063, + 0.004342529457062483, + 0.00438136700540781, + 0.00938621535897255, + 0.01284712553024292, + 0.005690310150384903, + 0.003835618495941162, + 0.004415969830006361, + 0.006759942509233952, + 0.004837165120989084, + 0.0034569790586829185, + 0.007450745906680822, + 0.011795446276664734, + 0.004320153966546059, + 0.0027242929209023714, + 0.004727914463728666, + 0.007359577342867851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.001043133088387549 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253676600754261 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253676600754261, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253676600754261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5345757603645325, + 0.09416387975215912, + 0.05717030540108681, + 0.028515903279185295, + 0.03128574788570404, + 0.049525145441293716, + 0.022978248074650764, + 0.07224109023809433, + 0.06642250716686249, + 0.043121375143527985 + ], + "credit_map": [ + 0.04634563624858856, + 0.06277099996805191, + 0.057232100516557693, + 0.08006659895181656, + 0.09310659766197205, + 0.07907754182815552, + 0.05350811034440994, + 0.06246821954846382, + 0.009070053696632385, + 0.01146688312292099, + 0.011665553785860538, + 0.0115659786388278, + 0.012362966313958168, + 0.014210937544703484, + 0.012171262875199318, + 0.011650257743895054, + 0.006219371221959591, + 0.004978332202881575, + 0.005028408952057362, + 0.0060035050846636295, + 0.009572441689670086, + 0.007080803159624338, + 0.006943588610738516, + 0.011343862861394882, + 0.002528984099626541, + 0.003963479306548834, + 0.0035115631762892008, + 0.003938907291740179, + 0.00390704907476902, + 0.0026501554530113935, + 0.0044713797979056835, + 0.0035443883389234543, + 0.0026074398774653673, + 0.004725476261228323, + 0.0037362580187618732, + 0.005068376660346985, + 0.0037480762694031, + 0.0033245179802179337, + 0.0037139682099223137, + 0.004361636936664581, + 0.004789304919540882, + 0.010099384933710098, + 0.006203527562320232, + 0.006620533298701048, + 0.007202947046607733, + 0.002805413445457816, + 0.00648780120536685, + 0.005316236522048712, + 0.00314801256172359, + 0.0027338333893567324, + 0.0025991154834628105, + 0.003293279092758894, + 0.003270405810326338, + 0.002956658136099577, + 0.002965030260384083, + 0.0020119159016758204, + 0.0075213005766272545, + 0.005484101362526417, + 0.007721950765699148, + 0.012568758800625801, + 0.006400818005204201, + 0.019225843250751495, + 0.007187251001596451, + 0.006131072528660297, + 0.007655428722500801, + 0.007932543754577637, + 0.007604458834975958, + 0.00842744018882513, + 0.009709630161523819, + 0.005320265889167786, + 0.0065797544084489346, + 0.013192984275519848, + 0.004060198087245226, + 0.005024257581681013, + 0.005785316228866577, + 0.010414667427539825, + 0.0060665966011583805, + 0.0021453772205859423, + 0.0034935229923576117, + 0.006131441798061132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01253676600754261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031600534915924 + ], + "opposite_logprob_tokens": [ + -1.910315990447998 + ], + "logprob_reward": -0.16031600534915924, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031600534915924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2237883061170578, + 0.11298327147960663, + 0.044639237225055695, + 0.05582598224282265, + 0.05779314041137695, + 0.14364168047904968, + 0.06891268491744995, + 0.03168632462620735, + 0.153274267911911, + 0.10745511949062347 + ], + "credit_map": [ + 0.017266282811760902, + 0.019913170486688614, + 0.022093061357736588, + 0.026827126741409302, + 0.03109860047698021, + 0.0554555207490921, + 0.020077720284461975, + 0.031056838110089302, + 0.008540346287190914, + 0.013613587245345116, + 0.012934491038322449, + 0.013358521275222301, + 0.015072978101670742, + 0.0224543958902359, + 0.012559816241264343, + 0.014449133537709713, + 0.005367457866668701, + 0.005989266559481621, + 0.0041636875830590725, + 0.005851021036505699, + 0.0064855278469622135, + 0.006068553309887648, + 0.005379026755690575, + 0.005334696266800165, + 0.00530786719173193, + 0.005568203050643206, + 0.00743451714515686, + 0.00578624103218317, + 0.0071562351658940315, + 0.008689922280609608, + 0.00899305660277605, + 0.006889940705150366, + 0.003972067963331938, + 0.008808760903775692, + 0.005343499593436718, + 0.01070214994251728, + 0.01006435789167881, + 0.005589979235082865, + 0.007215618621557951, + 0.006096706725656986, + 0.011715972796082497, + 0.021641911938786507, + 0.014288371428847313, + 0.01802627369761467, + 0.012686055153608322, + 0.02060447819530964, + 0.024806523695588112, + 0.019872091710567474, + 0.01057245023548603, + 0.011238249950110912, + 0.0067975218407809734, + 0.007025923579931259, + 0.007111048325896263, + 0.012146388180553913, + 0.00731980474665761, + 0.006701297126710415, + 0.0031524901278316975, + 0.0038813624996691942, + 0.005328305996954441, + 0.004716222640126944, + 0.0035328343510627747, + 0.003471817122772336, + 0.004345833323895931, + 0.0032574571669101715, + 0.02050781063735485, + 0.021112341433763504, + 0.018066031858325005, + 0.020125990733504295, + 0.02932179719209671, + 0.01631881296634674, + 0.011814378201961517, + 0.016007093712687492, + 0.009654794819653034, + 0.007581879384815693, + 0.02102544531226158, + 0.018013250082731247, + 0.009788397699594498, + 0.006474658846855164, + 0.011258509010076523, + 0.023658186197280884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.16031600534915924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-moving close-up zooms in on the broad, rounded armrests of a weathered, old rattan chair. Warm afternoon sunlight slants down, capturing the rough, dry, yellowed texture of the rattan and the details of its flaking varnish. Against this backdrop, marked by the marks of time, a fresh, emerald green sprout stubbornly emerges from a crack in the cane. Its spiraling vines, curling at the top, carefully unfurl in circles at a speed barely perceptible to the naked eye. Sunlight filters through its translucent leaves, casting a soft golden halo across the fine hairs covering them. As the vine slowly unfurls, the fresh green seems to engage in a silent, vibrant rhythm. Against the backdrop of blurred garden greenery and dappled light, the warm tones of the entire image create a poetic atmosphere of profound tranquility, hope, and healing, silently narrating the delicate interplay of decay and rebirth.", + "seed": 20794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_12/index_3934_seed_20794_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010766682727262378, + "vlm_logprob": -0.010766682727262378, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.010766682727262378, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00512962881475687 + ], + "opposite_logprob_tokens": [ + -5.630129814147949 + ], + "logprob_reward": -0.00512962881475687, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00512962881475687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20073159039020538, + 0.28581568598747253, + 0.04677398130297661, + 0.10641718655824661, + 0.07534995675086975, + 0.03912419453263283, + 0.03630012646317482, + 0.0654885545372963, + 0.0647716298699379, + 0.0792270377278328 + ], + "credit_map": [ + 0.041282769292593, + 0.01726544089615345, + 0.022704889997839928, + 0.021415920928120613, + 0.027253685519099236, + 0.024326970800757408, + 0.02092021517455578, + 0.025561703369021416, + 0.039112161844968796, + 0.05347441881895065, + 0.025863396003842354, + 0.037367042154073715, + 0.02720477059483528, + 0.03811364620923996, + 0.03161279857158661, + 0.033067457377910614, + 0.008597156032919884, + 0.005269914399832487, + 0.007727712858468294, + 0.005477787461131811, + 0.003289717948064208, + 0.00459030969068408, + 0.0062668523751199245, + 0.005554527044296265, + 0.02339814230799675, + 0.01916668377816677, + 0.012740721926093102, + 0.007740696892142296, + 0.012055773288011551, + 0.009460396133363247, + 0.01089764665812254, + 0.010957120917737484, + 0.008668004535138607, + 0.007060151547193527, + 0.012227732688188553, + 0.0136464424431324, + 0.0051458426751196384, + 0.010121169500052929, + 0.008976183831691742, + 0.009504429064691067, + 0.005304138641804457, + 0.006067424546927214, + 0.004313817247748375, + 0.005228252615779638, + 0.004880615975707769, + 0.0038507855497300625, + 0.004440563265234232, + 0.00503859668970108, + 0.0051583172753453255, + 0.004286895506083965, + 0.004887064453214407, + 0.003966324031352997, + 0.0033306877594441175, + 0.005082214716821909, + 0.004861753899604082, + 0.004726868588477373, + 0.01010327972471714, + 0.006663820240646601, + 0.010117129422724247, + 0.007738376967608929, + 0.008004664443433285, + 0.005413160193711519, + 0.009083333425223827, + 0.008364789187908173, + 0.009907993488013744, + 0.01210215874016285, + 0.008402064442634583, + 0.006545395590364933, + 0.007079561240971088, + 0.005934756249189377, + 0.005953534971922636, + 0.008846163749694824, + 0.011389460414648056, + 0.008472057059407234, + 0.011777241714298725, + 0.011192990466952324, + 0.008106397464871407, + 0.007986235432326794, + 0.00980416964739561, + 0.01049848087131977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.00512962881475687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103930201381445 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004103930201381445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103930201381445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18452948331832886, + 0.08142577856779099, + 0.09058484435081482, + 0.06052560359239578, + 0.07906342297792435, + 0.12395772337913513, + 0.07440749555826187, + 0.1387246549129486, + 0.05362436920404434, + 0.11315661668777466 + ], + "credit_map": [ + 0.02755814604461193, + 0.020694270730018616, + 0.02141175977885723, + 0.02497866190969944, + 0.02570088393986225, + 0.020489921793341637, + 0.019110767170786858, + 0.024585071951150894, + 0.01036506425589323, + 0.011719880625605583, + 0.007953397929668427, + 0.013623821549117565, + 0.00848409067839384, + 0.008798188529908657, + 0.010426267050206661, + 0.01005506794899702, + 0.009661776944994926, + 0.010165031999349594, + 0.012852675281465054, + 0.011866451241075993, + 0.008022172376513481, + 0.012126256711781025, + 0.014646888710558414, + 0.011243592016398907, + 0.009983438067138195, + 0.010886840522289276, + 0.007468278519809246, + 0.005574102979153395, + 0.00506254006177187, + 0.006492136977612972, + 0.007111440412700176, + 0.007946825586259365, + 0.01509916502982378, + 0.00967593677341938, + 0.01056367065757513, + 0.00782543234527111, + 0.005591271910816431, + 0.009776348248124123, + 0.010338649153709412, + 0.010192944668233395, + 0.015368570573627949, + 0.01709658093750477, + 0.009778446517884731, + 0.01223370898514986, + 0.02792695164680481, + 0.01249304786324501, + 0.013060827739536762, + 0.01599958911538124, + 0.009201504290103912, + 0.009964123368263245, + 0.011383566074073315, + 0.006261092610657215, + 0.009219646453857422, + 0.008581672795116901, + 0.009973357431590557, + 0.009822538122534752, + 0.01969429850578308, + 0.013852069154381752, + 0.01805863343179226, + 0.015212763100862503, + 0.01821763999760151, + 0.014638671651482582, + 0.020541824400424957, + 0.018508754670619965, + 0.004868216812610626, + 0.006791220977902412, + 0.006254445295780897, + 0.005196056328713894, + 0.011170526035130024, + 0.006003940477967262, + 0.006087963469326496, + 0.007252000272274017, + 0.01690356247127056, + 0.01131837721914053, + 0.014796925708651543, + 0.015687603503465652, + 0.013091393746435642, + 0.01221479568630457, + 0.014512130059301853, + 0.014631824567914009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.004103930201381445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021984234917908907 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021984234917908907, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021984234917908907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10341213643550873, + 0.05436303839087486, + 0.027867499738931656, + 0.08371570706367493, + 0.058848313987255096, + 0.27263668179512024, + 0.045118547976017, + 0.2471349835395813, + 0.0672202855348587, + 0.039682768285274506 + ], + "credit_map": [ + 0.012670031748712063, + 0.01004487182945013, + 0.010664022527635098, + 0.01210537925362587, + 0.015763062983751297, + 0.01468717586249113, + 0.014475732110440731, + 0.013001864776015282, + 0.005838178098201752, + 0.009169506840407848, + 0.00710226921364665, + 0.009414401836693287, + 0.0049288286827504635, + 0.004477083217352629, + 0.0067363339476287365, + 0.006696436088532209, + 0.0033088403288275003, + 0.0032519458327442408, + 0.003902692813426256, + 0.004649534355849028, + 0.0020939542446285486, + 0.002959528239443898, + 0.004082127474248409, + 0.003618876915425062, + 0.01798335090279579, + 0.017092499881982803, + 0.008867481723427773, + 0.005782346706837416, + 0.009646743535995483, + 0.0072545683942735195, + 0.008405176922678947, + 0.008683538995683193, + 0.009589689783751965, + 0.006931767333298922, + 0.008664654567837715, + 0.006514152977615595, + 0.0054593621753156185, + 0.006567067001014948, + 0.007643380668014288, + 0.00747823528945446, + 0.02315772883594036, + 0.028321780264377594, + 0.0237453430891037, + 0.028064334765076637, + 0.061245396733284, + 0.03410356864333153, + 0.035077109932899475, + 0.03892141208052635, + 0.005904273595660925, + 0.00699770450592041, + 0.004370182752609253, + 0.003172574331983924, + 0.005214286968111992, + 0.0068190027959644794, + 0.006522342562675476, + 0.006118180695921183, + 0.026882551610469818, + 0.03847597539424896, + 0.015985116362571716, + 0.018002886325120926, + 0.055391646921634674, + 0.02982519008219242, + 0.0320332832634449, + 0.030538348481059074, + 0.0071176644414663315, + 0.00789051502943039, + 0.00446104072034359, + 0.005400385707616806, + 0.017926177009940147, + 0.006820395588874817, + 0.007993623614311218, + 0.009610485285520554, + 0.00498472573235631, + 0.007503810338675976, + 0.0037395551335066557, + 0.0032838552724570036, + 0.0077628884464502335, + 0.0030552728567272425, + 0.0040732817724347115, + 0.005279381759464741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0021984234917908907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606731116771698 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606731116771698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606731116771698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07929907739162445, + 0.07109366357326508, + 0.04513939470052719, + 0.132109135389328, + 0.06158873438835144, + 0.32971686124801636, + 0.08493368327617645, + 0.08359099179506302, + 0.07274941354990005, + 0.03977901488542557 + ], + "credit_map": [ + 0.00890733115375042, + 0.00797442626208067, + 0.010119128040969372, + 0.012056665495038033, + 0.007984253577888012, + 0.008969917893409729, + 0.012127897702157497, + 0.011159457266330719, + 0.011833732016384602, + 0.01294132973998785, + 0.00806033331900835, + 0.012608926743268967, + 0.004430906847119331, + 0.0039681326597929, + 0.008529487997293472, + 0.008720816113054752, + 0.007688590325415134, + 0.004855576902627945, + 0.00561614241451025, + 0.004912573844194412, + 0.0038544093258678913, + 0.005010773427784443, + 0.0078610610216856, + 0.005340267438441515, + 0.030106475576758385, + 0.021290572360157967, + 0.017130207270383835, + 0.00839048158377409, + 0.015642555430531502, + 0.012143367901444435, + 0.012833856977522373, + 0.01457162108272314, + 0.009490825235843658, + 0.009551530703902245, + 0.008524145931005478, + 0.007566988002508879, + 0.005250871181488037, + 0.006064167711883783, + 0.007274018134921789, + 0.00786618608981371, + 0.03394472599029541, + 0.035214029252529144, + 0.02503976970911026, + 0.03135499730706215, + 0.07034710049629211, + 0.04942716658115387, + 0.040010981261730194, + 0.04437808319926262, + 0.00953348446637392, + 0.00995751190930605, + 0.007002588827162981, + 0.006691528484225273, + 0.014014202170073986, + 0.012164444662630558, + 0.013083775527775288, + 0.012486151419579983, + 0.012265977449715137, + 0.009831639006733894, + 0.007698824163526297, + 0.006947434041649103, + 0.014706961810588837, + 0.00804931577295065, + 0.012385809794068336, + 0.011705031618475914, + 0.010195343755185604, + 0.01714327186346054, + 0.0071672494523227215, + 0.006076326593756676, + 0.008060939610004425, + 0.005804646294564009, + 0.007670172024518251, + 0.010631470009684563, + 0.004298180341720581, + 0.005634283181279898, + 0.004693306516855955, + 0.0036042227875441313, + 0.006709670647978783, + 0.004589412827044725, + 0.0047960421070456505, + 0.005453896708786488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.01606731116771698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up view of the armrests of a weathered, old rattan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rattan texture visible as rough, dry, and yellowed with flaking varnish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fresh emerald green sprout emerging from a crack in the cane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sprout's vine slowly unfurl in circles with spiraling vines curling at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow-moving close-up zoom in on the chair armrests?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334119960665703 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.026334119960665703, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334119960665703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13743871450424194, + 0.0724971666932106, + 0.08508898317813873, + 0.14443965256214142, + 0.11551748216152191, + 0.18932554125785828, + 0.12930352985858917, + 0.07488802820444107, + 0.0240721944719553, + 0.027428774163126945 + ], + "credit_map": [ + 0.02859688363969326, + 0.012228170409798622, + 0.015629690140485764, + 0.014781689271330833, + 0.018726695328950882, + 0.014378012157976627, + 0.017089078202843666, + 0.01600847952067852, + 0.00788884423673153, + 0.011589857749640942, + 0.010037098079919815, + 0.011885320767760277, + 0.005589967593550682, + 0.007581944577395916, + 0.008980725891888142, + 0.008943403139710426, + 0.012166657485067844, + 0.011150257661938667, + 0.010229647159576416, + 0.01384783536195755, + 0.007458132691681385, + 0.008675570599734783, + 0.011952598579227924, + 0.00960827711969614, + 0.03139186277985573, + 0.015047082677483559, + 0.01913486048579216, + 0.01090481597930193, + 0.013408314436674118, + 0.02080494537949562, + 0.01596595160663128, + 0.017781810835003853, + 0.023831471800804138, + 0.0151374451816082, + 0.015131779946386814, + 0.01162893045693636, + 0.009343822486698627, + 0.011235382407903671, + 0.0154105918481946, + 0.013798041269183159, + 0.018396399915218353, + 0.019375193864107132, + 0.017598174512386322, + 0.017559193074703217, + 0.03985399380326271, + 0.025303509086370468, + 0.02496972307562828, + 0.026269327849149704, + 0.017736760899424553, + 0.01577049493789673, + 0.010855278931558132, + 0.01065746322274208, + 0.01577056385576725, + 0.022737478837370872, + 0.017659422010183334, + 0.018116049468517303, + 0.007258737459778786, + 0.00762053532525897, + 0.01404204685240984, + 0.00953182578086853, + 0.00866224430501461, + 0.006718046963214874, + 0.01111327763646841, + 0.00994130875915289, + 0.0029612616635859013, + 0.0027946624904870987, + 0.002849672921001911, + 0.00278645777143538, + 0.0034506323281675577, + 0.002183177974075079, + 0.0034282405395060778, + 0.003618086688220501, + 0.0032996363006532192, + 0.0029972451739013195, + 0.0033710310235619545, + 0.002629438880831003, + 0.0039513567462563515, + 0.003320354502648115, + 0.0039420584216713905, + 0.003917650785297155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.026334119960665703 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_tensor([9862, 9862])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_tensor([9862, 9862])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9981d2e46df5ed805de5488a0dc186a9f606c288 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_tensor([9862, 9862])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful, tight close-up captures the viewer's gaze firmly within a chilly, serene dojo. The frame is dominated by a cool, celestial light slanting through a high window. The beam carves a sharp path through the dust-filled air, illuminating the heart of the scene. A mysterious warrior stands there, his face largely obscured by a cold, steel mask carved with intricate ancient patterns, revealing only a pair of eagle-like eyes. He clad in a suit of dark, bladed armor, with sharp metal barbs extending from his shoulders and vambraces. The battle-scarred armor gleams darkly in the dim light. Without warning, he swung his bladed arm with a lightning-fast, gusty whistle, striking a solid wooden training dummy before him with steady, deliberate force. With a dull, explosive thud, the dummy disintegrates, sending countless wood chips and fibers of varying sizes scattering violently in all directions. Several of the largest fragments, in slow motion, streak past the camera, their rough fractured texture clearly visible. At this climax of the unleashing violence, the camera precisely captures a flash of icy coldness in the soldier's eyes through the slit in his mask, his pupils instantly constricting. The entire scene is imbued with a powerful tension, from extreme tranquility to a sudden explosion, conveying a sense of focus, coldness, and overwhelming power.", + "seed": 992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_9862_seed_992_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10202964656054973, + "vlm_logprob": -0.10202964656054973, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.10202964656054973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03392237052321434 + ], + "opposite_logprob_tokens": [ + -3.6589224338531494 + ], + "logprob_reward": -0.03392237052321434, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03392237052321434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1003408282995224, + 0.15022529661655426, + 0.3844599723815918, + 0.04597651585936546, + 0.0217757448554039, + 0.019587168470025063, + 0.04448550567030907, + 0.028195122256875038, + 0.04655838757753372, + 0.158395454287529 + ], + "credit_map": [ + 0.009160308167338371, + 0.011821430176496506, + 0.011363798752427101, + 0.013342363759875298, + 0.013547749258577824, + 0.012758921831846237, + 0.013863544911146164, + 0.014482714235782623, + 0.008035846054553986, + 0.018225373700261116, + 0.016723236069083214, + 0.02868288941681385, + 0.01559881679713726, + 0.023883908987045288, + 0.02441994845867157, + 0.014655264094471931, + 0.018720995634794235, + 0.014557979069650173, + 0.1438979208469391, + 0.14471492171287537, + 0.02612730674445629, + 0.009746660478413105, + 0.016297999769449234, + 0.01039621140807867, + 0.004661172162741423, + 0.006267936434596777, + 0.005668330937623978, + 0.0077979895286262035, + 0.006540936883538961, + 0.004989708308130503, + 0.005218442529439926, + 0.0048319995403289795, + 0.002624099375680089, + 0.002913091331720352, + 0.0032919689547270536, + 0.002523797797039151, + 0.003024499164894223, + 0.0025428158696740866, + 0.002337914891541004, + 0.0025175586342811584, + 0.001846919534727931, + 0.0033378894440829754, + 0.002545598428696394, + 0.002888239687308669, + 0.0027500386349856853, + 0.0023702611215412617, + 0.0018159577157348394, + 0.0020322641357779503, + 0.004865207243710756, + 0.002869831630960107, + 0.008719553239643574, + 0.004567108582705259, + 0.005379033274948597, + 0.005325871519744396, + 0.006855048704892397, + 0.005903853103518486, + 0.002917473902925849, + 0.0025805518962442875, + 0.0032608285546302795, + 0.0031221576500684023, + 0.0035009076818823814, + 0.0037316095549613237, + 0.004920229781419039, + 0.004161363933235407, + 0.0037070983089506626, + 0.0065772198140621185, + 0.006424236111342907, + 0.0070713278837502, + 0.00559839466586709, + 0.00583674618974328, + 0.007219899911433458, + 0.004123467020690441, + 0.04286493360996246, + 0.01424604095518589, + 0.012902645394206047, + 0.0153359891846776, + 0.00801507942378521, + 0.045724887400865555, + 0.009089457802474499, + 0.010216417722404003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.03392237052321434 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758213043212891 + ], + "opposite_logprob_tokens": [ + -5.006758213043213 + ], + "logprob_reward": -0.006758213043212891, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758213043212891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10834848880767822, + 0.06904813647270203, + 0.22718413174152374, + 0.027414314448833466, + 0.09384003281593323, + 0.038148775696754456, + 0.0704842358827591, + 0.03320113942027092, + 0.15900936722755432, + 0.17332135140895844 + ], + "credit_map": [ + 0.007246076595038176, + 0.013739560730755329, + 0.011335550807416439, + 0.015581964515149593, + 0.019677966833114624, + 0.016533304005861282, + 0.013327088207006454, + 0.010906987823545933, + 0.0049717118963599205, + 0.007855119183659554, + 0.006953234784305096, + 0.01271585188806057, + 0.006628220900893211, + 0.01104647945612669, + 0.011849896050989628, + 0.007027622777968645, + 0.010650989599525928, + 0.01005600206553936, + 0.0826297178864479, + 0.08565246313810349, + 0.01794940046966076, + 0.005312327761203051, + 0.009376320987939835, + 0.005556900054216385, + 0.005218382924795151, + 0.003225017338991165, + 0.003195895114913583, + 0.00410118605941534, + 0.0028091557323932648, + 0.0029654712416231632, + 0.0035076530184596777, + 0.0023915553465485573, + 0.007293736096471548, + 0.01154359057545662, + 0.009659741073846817, + 0.012392338365316391, + 0.017500292509794235, + 0.009979586116969585, + 0.012479649856686592, + 0.012991106137633324, + 0.0022353625390678644, + 0.0030038936529308558, + 0.003945382311940193, + 0.004644081927835941, + 0.007961087860167027, + 0.006048329174518585, + 0.005192097742110491, + 0.005118540022522211, + 0.00881757028400898, + 0.005674456246197224, + 0.014388535171747208, + 0.007147905416786671, + 0.009139951318502426, + 0.007748475298285484, + 0.01083635538816452, + 0.006730992347002029, + 0.003380902111530304, + 0.002705928636714816, + 0.004766380880028009, + 0.004482993856072426, + 0.003654451109468937, + 0.00315779447555542, + 0.0054559651762247086, + 0.005596724338829517, + 0.012545217759907246, + 0.0214474368840456, + 0.02262389473617077, + 0.022989453747868538, + 0.020635584369301796, + 0.019347792491316795, + 0.024540109559893608, + 0.014879876747727394, + 0.04491071403026581, + 0.014839882031083107, + 0.018107594922184944, + 0.014108098112046719, + 0.0092692319303751, + 0.05437634512782097, + 0.007907865568995476, + 0.009801643900573254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.006758213043212891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199413359165192 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199413359165192, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199413359165192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12630845606327057, + 0.03707785904407501, + 0.2725566029548645, + 0.037515878677368164, + 0.039180826395750046, + 0.026498299092054367, + 0.034844864159822464, + 0.03708410635590553, + 0.3051728308200836, + 0.08376025408506393 + ], + "credit_map": [ + 0.008636947721242905, + 0.020227305591106415, + 0.014316215179860592, + 0.017982151359319687, + 0.017924897372722626, + 0.01585300639271736, + 0.014776540920138359, + 0.01659139059484005, + 0.00327870761975646, + 0.006797083653509617, + 0.004658000078052282, + 0.008745980449020863, + 0.0025703031569719315, + 0.00535387871786952, + 0.0029405115637928247, + 0.0027333940379321575, + 0.015848882496356964, + 0.011093270033597946, + 0.10196790844202042, + 0.09847257286310196, + 0.018728377297520638, + 0.008018712513148785, + 0.010647577233612537, + 0.00777930673211813, + 0.003318609669804573, + 0.0030613148119300604, + 0.004628545604646206, + 0.006914657074958086, + 0.0038332974072545767, + 0.005561145022511482, + 0.00503741716966033, + 0.0051608942449092865, + 0.0034613218158483505, + 0.003866981714963913, + 0.004038440529257059, + 0.00530704902485013, + 0.007193658500909805, + 0.004387360066175461, + 0.00571387680247426, + 0.005212138406932354, + 0.002263405127450824, + 0.0023748865351080894, + 0.003646040568128228, + 0.004332195036113262, + 0.0047764163464307785, + 0.003048361511901021, + 0.0032717459835112095, + 0.0027852472849190235, + 0.00455855205655098, + 0.0029427488334476948, + 0.009253428317606449, + 0.003730280790477991, + 0.0039315237663686275, + 0.003691828576847911, + 0.004101509228348732, + 0.0026349942199885845, + 0.002907182788476348, + 0.0039169080555438995, + 0.0041690305806696415, + 0.004661723971366882, + 0.006134628318250179, + 0.004566538147628307, + 0.006031008902937174, + 0.004697086755186319, + 0.024266446009278297, + 0.04218291491270065, + 0.04642215371131897, + 0.04310191795229912, + 0.039748404175043106, + 0.03611551225185394, + 0.04604161158204079, + 0.027293875813484192, + 0.023098144680261612, + 0.006838780362159014, + 0.008414439857006073, + 0.00803541112691164, + 0.004537603352218866, + 0.023014528676867485, + 0.004349626135081053, + 0.005471715237945318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.014199413359165192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347937062382698 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347937062382698, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347937062382698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06600981950759888, + 0.043562427163124084, + 0.2448786497116089, + 0.036289580166339874, + 0.03493518754839897, + 0.01726296916604042, + 0.019533934071660042, + 0.035518500953912735, + 0.44911450147628784, + 0.05289449170231819 + ], + "credit_map": [ + 0.00463769631460309, + 0.009035058319568634, + 0.007598512805998325, + 0.008713153190910816, + 0.008216450922191143, + 0.010935540311038494, + 0.007943940348923206, + 0.008929458446800709, + 0.0023613213561475277, + 0.0067943367175757885, + 0.004911113064736128, + 0.011571447364985943, + 0.0036643510684370995, + 0.004827540367841721, + 0.004995183553546667, + 0.00443712854757905, + 0.013158603571355343, + 0.009871737100183964, + 0.0940551608800888, + 0.09244950115680695, + 0.016597090288996696, + 0.004344376269727945, + 0.008844716474413872, + 0.005557442083954811, + 0.0032409094274044037, + 0.0024537849240005016, + 0.004467650316655636, + 0.006024910137057304, + 0.003535280702635646, + 0.0058847288601100445, + 0.0049801296554505825, + 0.0057021817192435265, + 0.003837588941678405, + 0.004962843377143145, + 0.003999225329607725, + 0.0034560803323984146, + 0.006083116866648197, + 0.004732293542474508, + 0.004196294117718935, + 0.00366774364374578, + 0.0012376492377370596, + 0.0015716924099251628, + 0.002256142906844616, + 0.002486177021637559, + 0.0033128950744867325, + 0.0017837801715359092, + 0.002051188610494137, + 0.002563442336395383, + 0.002352614887058735, + 0.001719763851724565, + 0.003559647360816598, + 0.002263751346617937, + 0.002753394888713956, + 0.001986672868952155, + 0.0030571804381906986, + 0.0018409069161862135, + 0.003310097148641944, + 0.003676448483020067, + 0.004470149986445904, + 0.005937930662184954, + 0.003975303377956152, + 0.005011332221329212, + 0.003936228808015585, + 0.00520100723952055, + 0.03603672608733177, + 0.06431026756763458, + 0.059677015990018845, + 0.06342735886573792, + 0.0618189312517643, + 0.05292237922549248, + 0.07070855796337128, + 0.0402132049202919, + 0.0087794354185462, + 0.004863157402724028, + 0.005646809469908476, + 0.005675245076417923, + 0.007408240810036659, + 0.008789513260126114, + 0.006294829770922661, + 0.005437253508716822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.026347937062382698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892029881477356 + ], + "opposite_logprob_tokens": [ + -1.0539202690124512 + ], + "logprob_reward": -0.42892029881477356, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892029881477356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10870131105184555, + 0.05589940398931503, + 0.4144980311393738, + 0.03299736976623535, + 0.10265714675188065, + 0.05453847721219063, + 0.07722163200378418, + 0.03438850864768028, + 0.046478983014822006, + 0.07261910289525986 + ], + "credit_map": [ + 0.009374038316309452, + 0.015025853179395199, + 0.012841222807765007, + 0.01897936873137951, + 0.01211039163172245, + 0.01992572471499443, + 0.010581287555396557, + 0.009863430634140968, + 0.004893495701253414, + 0.008380651473999023, + 0.006716672796756029, + 0.010233847424387932, + 0.005206193774938583, + 0.006374885328114033, + 0.008345762267708778, + 0.005747893825173378, + 0.02028244361281395, + 0.015545930713415146, + 0.15818287432193756, + 0.16003987193107605, + 0.030366532504558563, + 0.006133709568530321, + 0.015905017033219337, + 0.00804166216403246, + 0.00415225001052022, + 0.004641790874302387, + 0.004134554881602526, + 0.0046305167488753796, + 0.004587006289511919, + 0.0038526460994035006, + 0.0031710148323327303, + 0.0038275890983641148, + 0.009024671278893948, + 0.01229821890592575, + 0.010593662969768047, + 0.014008982107043266, + 0.015910694375634193, + 0.011381988413631916, + 0.01418138388544321, + 0.01525755226612091, + 0.004001178778707981, + 0.005184822250157595, + 0.006181598640978336, + 0.006681407336145639, + 0.00898529589176178, + 0.009300381876528263, + 0.007730653043836355, + 0.006473137531429529, + 0.007297309581190348, + 0.006737126968801022, + 0.014281298033893108, + 0.007955516688525677, + 0.010320218279957771, + 0.00980361644178629, + 0.01292875874787569, + 0.007897792384028435, + 0.003033566055819392, + 0.0028952243737876415, + 0.004824473522603512, + 0.0037139039486646652, + 0.004276036284863949, + 0.003420898225158453, + 0.006985520012676716, + 0.005238881800323725, + 0.003429989330470562, + 0.005764184985309839, + 0.004856925457715988, + 0.00599690480157733, + 0.005665323697030544, + 0.007375781890004873, + 0.0072451005689799786, + 0.006144772749394178, + 0.009030099026858807, + 0.008192976005375385, + 0.008877327665686607, + 0.01005580835044384, + 0.0055907117202878, + 0.018055398017168045, + 0.006561548449099064, + 0.00625523179769516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.42892029881477356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, tight close-up captures the viewer's gaze firmly within a chilly, serene dojo. The frame is dominated by a cool, celestial light slanting through a high window. The beam carves a sharp path through the dust-filled air, illuminating the heart of the scene. A mysterious warrior stands there, his face largely obscured by a cold, steel mask carved with intricate ancient patterns, revealing only a pair of eagle-like eyes. He clad in a suit of dark, bladed armor, with sharp metal barbs extending from his shoulders and vambraces. The battle-scarred armor gleams darkly in the dim light. Without warning, he swung his bladed arm with a lightning-fast, gusty whistle, striking a solid wooden training dummy before him with steady, deliberate force. With a dull, explosive thud, the dummy disintegrates, sending countless wood chips and fibers of varying sizes scattering violently in all directions. Several of the largest fragments, in slow motion, streak past the camera, their rough fractured texture clearly visible. At this climax of the unleashing violence, the camera precisely captures a flash of icy coldness in the soldier's eyes through the slit in his mask, his pupils instantly constricting. The entire scene is imbued with a powerful tension, from extreme tranquility to a sudden explosion, conveying a sense of focus, coldness, and overwhelming power.", + "seed": 993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_9862_seed_993_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06177854691632092, + "vlm_logprob": -0.06177854691632092, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.06177854691632092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07751624286174774 + ], + "opposite_logprob_tokens": [ + -2.7025163173675537 + ], + "logprob_reward": -0.07751624286174774, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07751624286174774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2930547297000885, + 0.06688591837882996, + 0.1419156938791275, + 0.12367656826972961, + 0.027205877006053925, + 0.02808254212141037, + 0.2112499177455902, + 0.020812740549445152, + 0.06176001578569412, + 0.025355976074934006 + ], + "credit_map": [ + 0.038633137941360474, + 0.021150680258870125, + 0.02797570452094078, + 0.03443911671638489, + 0.016432998701930046, + 0.04843529313802719, + 0.05504694953560829, + 0.050940852612257004, + 0.007242471911013126, + 0.007776397746056318, + 0.009645404294133186, + 0.009163557551801205, + 0.0067600710317492485, + 0.007587192114442587, + 0.007695065811276436, + 0.011015762574970722, + 0.01758342608809471, + 0.022643782198429108, + 0.01924244500696659, + 0.019985413178801537, + 0.013482029549777508, + 0.021335719153285027, + 0.014944588765501976, + 0.012698283419013023, + 0.012827289290726185, + 0.015357756055891514, + 0.018178259953856468, + 0.013292611576616764, + 0.01840820349752903, + 0.014227231033146381, + 0.015487860888242722, + 0.0158973541110754, + 0.0038613935466855764, + 0.005019066855311394, + 0.0031356019899249077, + 0.0035633172374218702, + 0.0028551435098052025, + 0.0025674288626760244, + 0.0028932008426636457, + 0.003310725325718522, + 0.0028108779806643724, + 0.005275642033666372, + 0.0039045270532369614, + 0.0025320248678326607, + 0.003635386936366558, + 0.003438122570514679, + 0.0030819433741271496, + 0.00340401753783226, + 0.030821988359093666, + 0.0229903943836689, + 0.03413081541657448, + 0.021227631717920303, + 0.016494838520884514, + 0.0342242531478405, + 0.02648824080824852, + 0.024871746078133583, + 0.0026212900411337614, + 0.0019659767858684063, + 0.0030960466247051954, + 0.0021891691721975803, + 0.00214308500289917, + 0.002599267056211829, + 0.002995073329657316, + 0.0032028306741267443, + 0.006878179963678122, + 0.00721924751996994, + 0.007014174945652485, + 0.00807106215506792, + 0.009786218404769897, + 0.006798327900469303, + 0.007084778975695372, + 0.008908024989068508, + 0.00219960929825902, + 0.0023961036931723356, + 0.005523353815078735, + 0.0031241094693541527, + 0.0021496948320418596, + 0.0038080201484262943, + 0.003737901570275426, + 0.0024171844124794006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.07751624286174774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003638435387983918 + ], + "opposite_logprob_tokens": [ + -5.62863826751709 + ], + "logprob_reward": -0.003638435387983918, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003638435387983918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06800679862499237, + 0.029443062841892242, + 0.08554068207740784, + 0.09737803786993027, + 0.26873257756233215, + 0.08770836889743805, + 0.06985289603471756, + 0.14189408719539642, + 0.0898212343454361, + 0.061622269451618195 + ], + "credit_map": [ + 0.008951229974627495, + 0.00575078371912241, + 0.00831751711666584, + 0.009462589398026466, + 0.004249144811183214, + 0.008780810050666332, + 0.011909354478120804, + 0.010585366748273373, + 0.0031113773584365845, + 0.0027655039448291063, + 0.0033660284243524075, + 0.0034449577797204256, + 0.005212259013205767, + 0.004980732686817646, + 0.0032972623594105244, + 0.003264938946813345, + 0.011014004237949848, + 0.005826809909194708, + 0.007388019934296608, + 0.009649721905589104, + 0.0078735938295722, + 0.006374335382133722, + 0.029077814891934395, + 0.0083363838493824, + 0.009080365300178528, + 0.013795115053653717, + 0.021119996905326843, + 0.009915735572576523, + 0.009313995018601418, + 0.012181763537228107, + 0.010455045849084854, + 0.011516017839312553, + 0.0366465225815773, + 0.02969205006957054, + 0.04211219772696495, + 0.056909263134002686, + 0.03338001295924187, + 0.02287326008081436, + 0.027834143489599228, + 0.0192851722240448, + 0.00949521828442812, + 0.011753041297197342, + 0.012051882222294807, + 0.008776378817856312, + 0.008162209764122963, + 0.009885993786156178, + 0.008559169247746468, + 0.019024478271603584, + 0.009640784934163094, + 0.00915489811450243, + 0.011079782620072365, + 0.008087361231446266, + 0.005585040897130966, + 0.009256914258003235, + 0.009093410335481167, + 0.00795469805598259, + 0.016656745225191116, + 0.016127537935972214, + 0.018294259905815125, + 0.013044838793575764, + 0.013759934343397617, + 0.017235910519957542, + 0.019687846302986145, + 0.027087019756436348, + 0.012009973637759686, + 0.014115928672254086, + 0.009838915430009365, + 0.010448839515447617, + 0.012656903825700283, + 0.009499016217887402, + 0.009833619929850101, + 0.011418046429753304, + 0.005757008213549852, + 0.008058912120759487, + 0.008394233882427216, + 0.008263668976724148, + 0.006479482166469097, + 0.00844584871083498, + 0.008430389687418938, + 0.0077927252277731895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.003638435387983918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033700406551361084 + ], + "opposite_logprob_tokens": [ + -3.408700466156006 + ], + "logprob_reward": -0.033700406551361084, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033700406551361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0716828778386116, + 0.06217724084854126, + 0.09883693605661392, + 0.1826675832271576, + 0.16629824042320251, + 0.07072107493877411, + 0.10269026458263397, + 0.16695012152194977, + 0.03654709458351135, + 0.041428498923778534 + ], + "credit_map": [ + 0.00975038018077612, + 0.007210277486592531, + 0.007102290168404579, + 0.008358528837561607, + 0.0038666867185384035, + 0.011043205857276917, + 0.012316915206611156, + 0.012034604325890541, + 0.006110592279583216, + 0.007915306836366653, + 0.005637122318148613, + 0.007971188053488731, + 0.008784539997577667, + 0.009404362179338932, + 0.00779296737164259, + 0.008561166934669018, + 0.01150219701230526, + 0.01467348262667656, + 0.013563457876443863, + 0.01628931239247322, + 0.008423526771366596, + 0.01290810015052557, + 0.012562679126858711, + 0.00891418382525444, + 0.015861475840210915, + 0.032165851444005966, + 0.021615518257021904, + 0.02325611189007759, + 0.024381034076213837, + 0.01957128942012787, + 0.019593695178627968, + 0.026222623884677887, + 0.022193092852830887, + 0.01863701641559601, + 0.025090700015425682, + 0.033166926354169846, + 0.019256213679909706, + 0.013890806585550308, + 0.01917838305234909, + 0.014885115437209606, + 0.00680906418710947, + 0.012062452733516693, + 0.009821130894124508, + 0.007385535631328821, + 0.01036774180829525, + 0.008209834806621075, + 0.007142677903175354, + 0.0089226383715868, + 0.013769318349659443, + 0.011028141714632511, + 0.019969677552580833, + 0.01293147448450327, + 0.00761877978220582, + 0.014277755282819271, + 0.01176818460226059, + 0.011326945386826992, + 0.01885107532143593, + 0.01039742399007082, + 0.019231343641877174, + 0.01821623183786869, + 0.011046783067286015, + 0.022486329078674316, + 0.02992795966565609, + 0.03679298982024193, + 0.0047888364642858505, + 0.0037124615628272295, + 0.005819057580083609, + 0.003843638114631176, + 0.004215824417769909, + 0.004205642733722925, + 0.0047176615335047245, + 0.005243976600468159, + 0.003962084650993347, + 0.004936933051794767, + 0.006115758325904608, + 0.004159941803663969, + 0.004084625281393528, + 0.005245413165539503, + 0.006078093312680721, + 0.0068456558510661125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.033700406551361084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693719655275345 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.033693719655275345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693719655275345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042114466428756714, + 0.023764291778206825, + 0.03972461447119713, + 0.056153811514377594, + 0.6260355114936829, + 0.0413232147693634, + 0.031469400972127914, + 0.06391286849975586, + 0.044080864638090134, + 0.031421009451150894 + ], + "credit_map": [ + 0.006262974347919226, + 0.005025364924222231, + 0.006272096652537584, + 0.004317085724323988, + 0.003540146630257368, + 0.004841088317334652, + 0.005843207705765963, + 0.006012497935444117, + 0.00279498309828341, + 0.0022345606703311205, + 0.003304747398942709, + 0.002824267139658332, + 0.0017686097417026758, + 0.002482681768015027, + 0.0028063920326530933, + 0.00554804690182209, + 0.0049425046890974045, + 0.003407158888876438, + 0.003911009524017572, + 0.004333444871008396, + 0.004217390436679125, + 0.0039152465760707855, + 0.010839935392141342, + 0.004157919902354479, + 0.005396834574639797, + 0.006327225361019373, + 0.013157454319298267, + 0.00642729876562953, + 0.0060191950760781765, + 0.005570923909544945, + 0.006788129918277264, + 0.006466741673648357, + 0.08505644649267197, + 0.06581912934780121, + 0.09534334391355515, + 0.13453035056591034, + 0.07682430744171143, + 0.0574922151863575, + 0.06475802510976791, + 0.04621163383126259, + 0.0043015312403440475, + 0.0059768203645944595, + 0.005302218720316887, + 0.004530992358922958, + 0.0026808343827724457, + 0.004561267793178558, + 0.00421554921194911, + 0.009753992781043053, + 0.004355708137154579, + 0.0033021720591932535, + 0.006553554907441139, + 0.003808512818068266, + 0.0022322158329188824, + 0.0047239623963832855, + 0.003317577298730612, + 0.003175692865625024, + 0.009377093985676765, + 0.004316503182053566, + 0.008731314912438393, + 0.007234050892293453, + 0.004985329229384661, + 0.007636099122464657, + 0.01135930698364973, + 0.010273157618939877, + 0.004307473078370094, + 0.005568617954850197, + 0.006217094603925943, + 0.0052031222730875015, + 0.00698124710470438, + 0.0048254732973873615, + 0.007071200292557478, + 0.003906628116965294, + 0.002760066417977214, + 0.004257373511791229, + 0.004818387795239687, + 0.0033633734565228224, + 0.0036062703002244234, + 0.003895530244335532, + 0.004077432677149773, + 0.004642573185265064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.033693719655275345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603439301252365 + ], + "opposite_logprob_tokens": [ + -1.910343885421753 + ], + "logprob_reward": -0.1603439301252365, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603439301252365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050377629697322845, + 0.024011580273509026, + 0.05033973976969719, + 0.041129257529973984, + 0.13842277228832245, + 0.06754783540964127, + 0.09391860663890839, + 0.23615926504135132, + 0.1570698469877243, + 0.14102347195148468 + ], + "credit_map": [ + 0.00846721138805151, + 0.0024311000015586615, + 0.007234411314129829, + 0.005691887810826302, + 0.0023884291294962168, + 0.00797645840793848, + 0.0064889960922300816, + 0.009699135087430477, + 0.003103991039097309, + 0.0021303053945302963, + 0.0036616004072129726, + 0.0026085954159498215, + 0.002243264578282833, + 0.0036704144440591335, + 0.0032522219698876143, + 0.0033411874901503325, + 0.006049470975995064, + 0.004955066833645105, + 0.0051155309192836285, + 0.006033641751855612, + 0.0067655108869075775, + 0.004455824848264456, + 0.011206250637769699, + 0.005758441053330898, + 0.003581246128305793, + 0.005854400806128979, + 0.008789577521383762, + 0.004447117447853088, + 0.004765721503645182, + 0.004618362057954073, + 0.004500827752053738, + 0.0045720050111413, + 0.019166504964232445, + 0.015678074210882187, + 0.022264130413532257, + 0.028909724205732346, + 0.018438996747136116, + 0.01217613648623228, + 0.013104824349284172, + 0.008684370666742325, + 0.00747072696685791, + 0.009885397739708424, + 0.008821163326501846, + 0.006569146178662777, + 0.005829421803355217, + 0.008906593546271324, + 0.006193188950419426, + 0.013872196897864342, + 0.018913231790065765, + 0.009172763675451279, + 0.016317689791321754, + 0.010457254946231842, + 0.0062758042477071285, + 0.014072624035179615, + 0.008921261876821518, + 0.009787971153855324, + 0.031657300889492035, + 0.02003679797053337, + 0.02710663713514805, + 0.026037946343421936, + 0.019840585067868233, + 0.02484590746462345, + 0.0410795733332634, + 0.04555452615022659, + 0.022222444415092468, + 0.019296744838356972, + 0.02044638991355896, + 0.01957371085882187, + 0.014560648240149021, + 0.016726309433579445, + 0.017532020807266235, + 0.02671157196164131, + 0.01203185971826315, + 0.015682145953178406, + 0.02418043650686741, + 0.017625384032726288, + 0.01607269048690796, + 0.018434705212712288, + 0.017275171354413033, + 0.019721079617738724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.1603439301252365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, tight close-up captures the viewer's gaze firmly within a chilly, serene dojo. The frame is dominated by a cool, celestial light slanting through a high window. The beam carves a sharp path through the dust-filled air, illuminating the heart of the scene. A mysterious warrior stands there, his face largely obscured by a cold, steel mask carved with intricate ancient patterns, revealing only a pair of eagle-like eyes. He clad in a suit of dark, bladed armor, with sharp metal barbs extending from his shoulders and vambraces. The battle-scarred armor gleams darkly in the dim light. Without warning, he swung his bladed arm with a lightning-fast, gusty whistle, striking a solid wooden training dummy before him with steady, deliberate force. With a dull, explosive thud, the dummy disintegrates, sending countless wood chips and fibers of varying sizes scattering violently in all directions. Several of the largest fragments, in slow motion, streak past the camera, their rough fractured texture clearly visible. At this climax of the unleashing violence, the camera precisely captures a flash of icy coldness in the soldier's eyes through the slit in his mask, his pupils instantly constricting. The entire scene is imbued with a powerful tension, from extreme tranquility to a sudden explosion, conveying a sense of focus, coldness, and overwhelming power.", + "seed": 994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_9862_seed_994_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016519937245175243, + "vlm_logprob": -0.016519937245175243, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.016519937245175243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020146971568465233 + ], + "opposite_logprob_tokens": [ + -4.27014684677124 + ], + "logprob_reward": -0.020146971568465233, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020146971568465233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3297913372516632, + 0.06970258802175522, + 0.03340271860361099, + 0.07334747910499573, + 0.1238102912902832, + 0.07523494958877563, + 0.08698152750730515, + 0.027659468352794647, + 0.08667071908712387, + 0.09339892119169235 + ], + "credit_map": [ + 0.03819521889090538, + 0.04750981554389, + 0.020784247666597366, + 0.042132455855607986, + 0.029347945004701614, + 0.04443134739995003, + 0.031380537897348404, + 0.07600978016853333, + 0.008978595025837421, + 0.007933578453958035, + 0.009590566158294678, + 0.0073721399530768394, + 0.010485685430467129, + 0.007775479461997747, + 0.007375573273748159, + 0.010190973058342934, + 0.0037537694443017244, + 0.003523517632856965, + 0.006225363817065954, + 0.006301874294877052, + 0.0036806794814765453, + 0.0035099065862596035, + 0.0032574038486927748, + 0.0031502016354352236, + 0.006964509375393391, + 0.008655432611703873, + 0.018208270892500877, + 0.011006845161318779, + 0.005543902982026339, + 0.009225981310009956, + 0.007920785807073116, + 0.005821746774017811, + 0.009317508898675442, + 0.030530305579304695, + 0.012819620780646801, + 0.037865858525037766, + 0.012113667093217373, + 0.006045960355550051, + 0.009354256093502045, + 0.00576310558244586, + 0.0037030531093478203, + 0.01018235832452774, + 0.005396457388997078, + 0.0057967305183410645, + 0.011146971955895424, + 0.017696721479296684, + 0.011246338486671448, + 0.010066318325698376, + 0.0091191241517663, + 0.010180546902120113, + 0.0093673812225461, + 0.007101329043507576, + 0.013475093990564346, + 0.013206442818045616, + 0.01358453743159771, + 0.010947072878479958, + 0.004640508443117142, + 0.003061345312744379, + 0.0023441521916538477, + 0.002934826537966728, + 0.004519332200288773, + 0.003639178117737174, + 0.002657128730788827, + 0.003862996120005846, + 0.008639008738100529, + 0.006838408298790455, + 0.011443338356912136, + 0.009033836424350739, + 0.011973648332059383, + 0.01325876172631979, + 0.009379521943628788, + 0.016104193404316902, + 0.013227173127233982, + 0.008259975351393223, + 0.01425936073064804, + 0.011289521120488644, + 0.010250605642795563, + 0.012115923687815666, + 0.008609233424067497, + 0.015387126244604588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.020146971568465233 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022134108003228903 + ], + "opposite_logprob_tokens": [ + -6.127213478088379 + ], + "logprob_reward": -0.0022134108003228903, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022134108003228903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07211624085903168, + 0.049155086278915405, + 0.048950571566820145, + 0.10711326450109482, + 0.09703854471445084, + 0.0476653091609478, + 0.1676526516675949, + 0.07534578442573547, + 0.06641319394111633, + 0.2685493528842926 + ], + "credit_map": [ + 0.007408721838146448, + 0.011369270272552967, + 0.005217091646045446, + 0.013842610642313957, + 0.008195134811103344, + 0.010028579272329807, + 0.00959608145058155, + 0.006458749063313007, + 0.005273513495922089, + 0.006818982306867838, + 0.005462221801280975, + 0.00485261483117938, + 0.007264463230967522, + 0.00606674887239933, + 0.006788502912968397, + 0.006628039758652449, + 0.0047552092000842094, + 0.004769986029714346, + 0.007796321529895067, + 0.00875987857580185, + 0.006797751411795616, + 0.005427616648375988, + 0.004990613088011742, + 0.0056531960144639015, + 0.011012114584445953, + 0.012437931261956692, + 0.024660151451826096, + 0.0170510932803154, + 0.009741110727190971, + 0.01296908874064684, + 0.010715305805206299, + 0.008526461198925972, + 0.006285518873482943, + 0.026247018948197365, + 0.005607614293694496, + 0.027632350102066994, + 0.007614298723638058, + 0.009017706848680973, + 0.0074372366070747375, + 0.0071967956610023975, + 0.004905437119305134, + 0.006504922639578581, + 0.005397285334765911, + 0.005614150315523148, + 0.00747894449159503, + 0.007677174638956785, + 0.005205986555665731, + 0.004881410393863916, + 0.01792025752365589, + 0.019701585173606873, + 0.01626230962574482, + 0.013709192164242268, + 0.02715100720524788, + 0.026523077860474586, + 0.025558441877365112, + 0.020826775580644608, + 0.008662737905979156, + 0.011810083873569965, + 0.0071335784159600735, + 0.011436056345701218, + 0.011796554550528526, + 0.009147637523710728, + 0.007458138279616833, + 0.007900995202362537, + 0.007050642743706703, + 0.007067445665597916, + 0.007078459486365318, + 0.008131952956318855, + 0.010125932283699512, + 0.008401591330766678, + 0.008121464401483536, + 0.010435701347887516, + 0.02990352176129818, + 0.0246626827865839, + 0.0381467230618, + 0.03302476182579994, + 0.03200861066579819, + 0.04712827876210213, + 0.03845124691724777, + 0.025223538279533386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.0022134108003228903 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125381201505661, + 0.05909751355648041, + 0.046002645045518875, + 0.1789294183254242, + 0.05470573157072067, + 0.1567929983139038, + 0.23314502835273743, + 0.029662270098924637, + 0.05129265785217285, + 0.06499053537845612 + ], + "credit_map": [ + 0.009996671229600906, + 0.016013551503419876, + 0.01156336534768343, + 0.021838927641510963, + 0.008266352117061615, + 0.019373761489987373, + 0.01426634006202221, + 0.024062223732471466, + 0.005159125197678804, + 0.010114572010934353, + 0.00834273174405098, + 0.006491439417004585, + 0.008260082453489304, + 0.006947393529117107, + 0.006388939451426268, + 0.007393225096166134, + 0.004886009730398655, + 0.004618625622242689, + 0.005303023848682642, + 0.008598947897553444, + 0.005771814379841089, + 0.006316075101494789, + 0.005127069540321827, + 0.005381077062338591, + 0.016755152493715286, + 0.02813362330198288, + 0.033872269093990326, + 0.03646417707204819, + 0.012702210806310177, + 0.02332909032702446, + 0.016857879236340523, + 0.010814993642270565, + 0.00499607902020216, + 0.013211471028625965, + 0.0065360595472157, + 0.014723963104188442, + 0.005070394836366177, + 0.003923057112842798, + 0.0035476794000715017, + 0.002697024028748274, + 0.010231290012598038, + 0.016339872032403946, + 0.014249378815293312, + 0.013385536149144173, + 0.020852649584412575, + 0.036398231983184814, + 0.023744752630591393, + 0.02159127965569496, + 0.0232776440680027, + 0.027527617290616035, + 0.023316962644457817, + 0.01847757399082184, + 0.03826131671667099, + 0.03772196173667908, + 0.03573629632592201, + 0.028825627639889717, + 0.0031645868439227343, + 0.004151728469878435, + 0.0022953564766794443, + 0.003031213767826557, + 0.006044148467481136, + 0.003995808772742748, + 0.0037147216498851776, + 0.00326470285654068, + 0.005821635480970144, + 0.0052357506938278675, + 0.005651036277413368, + 0.004505170974880457, + 0.006676332093775272, + 0.007882832549512386, + 0.004886552691459656, + 0.01063334196805954, + 0.006778297014534473, + 0.004588764626532793, + 0.012210500426590443, + 0.009124794974923134, + 0.00708362041041255, + 0.008957115933299065, + 0.008311803452670574, + 0.007935629226267338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.011075802147388458 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072855442762375 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072855442762375, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072855442762375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12525302171707153, + 0.04627404734492302, + 0.027039317414164543, + 0.21200720965862274, + 0.03579659387469292, + 0.05513278394937515, + 0.40168774127960205, + 0.03184850886464119, + 0.03139333426952362, + 0.03356742858886719 + ], + "credit_map": [ + 0.017860278487205505, + 0.012691210955381393, + 0.01136376615613699, + 0.013532467186450958, + 0.009933126159012318, + 0.014742826111614704, + 0.012918497435748577, + 0.03221086040139198, + 0.005686861928552389, + 0.010435972362756729, + 0.00406138738617301, + 0.004229097627103329, + 0.006114952731877565, + 0.005168742034584284, + 0.0051622274331748486, + 0.005414807703346014, + 0.0032839856576174498, + 0.00245476420968771, + 0.0031149492133408785, + 0.004157766699790955, + 0.002207443816587329, + 0.004738292191177607, + 0.003013974754139781, + 0.0040681432001292706, + 0.02324635721743107, + 0.0331486314535141, + 0.035199861973524094, + 0.046253401786088943, + 0.016717592254281044, + 0.029079247266054153, + 0.018159203231334686, + 0.010202918201684952, + 0.003252132795751095, + 0.0050375559367239475, + 0.004830505698919296, + 0.006009787321090698, + 0.004859744571149349, + 0.003880501724779606, + 0.0038847788237035275, + 0.004041587933897972, + 0.0037163111846894026, + 0.005749564617872238, + 0.0047915359027683735, + 0.003966625779867172, + 0.0075066955760121346, + 0.01357750128954649, + 0.00820930115878582, + 0.007615253794938326, + 0.03869212791323662, + 0.04658088833093643, + 0.03681504353880882, + 0.0331365205347538, + 0.06343952566385269, + 0.06615158170461655, + 0.0634850561618805, + 0.05338701978325844, + 0.003507378976792097, + 0.004871922545135021, + 0.002842520596459508, + 0.005243324674665928, + 0.0050456346943974495, + 0.004492816049605608, + 0.0026470699813216925, + 0.003197845770046115, + 0.003834065981209278, + 0.0030255417805165052, + 0.003460444277152419, + 0.0031539793126285076, + 0.005537698976695538, + 0.004378584213554859, + 0.0033664654474705458, + 0.004636554978787899, + 0.0038725826889276505, + 0.0027219902258366346, + 0.004889602307230234, + 0.004433673340827227, + 0.004245948512107134, + 0.004324947949498892, + 0.0043798815459012985, + 0.004698801785707474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.011072855442762375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038090646266937256 + ], + "opposite_logprob_tokens": [ + -3.288090705871582 + ], + "logprob_reward": -0.038090646266937256, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038090646266937256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06901180744171143, + 0.023244814947247505, + 0.016932843253016472, + 0.07333029061555862, + 0.0432266891002655, + 0.05718443542718887, + 0.08799388259649277, + 0.06879422068595886, + 0.3110954165458679, + 0.24918554723262787 + ], + "credit_map": [ + 0.008515805006027222, + 0.007637463975697756, + 0.004275027196854353, + 0.00641864025965333, + 0.004547974560409784, + 0.012645252980291843, + 0.00834609568119049, + 0.016625545918941498, + 0.00421311566606164, + 0.0026302975602447987, + 0.002869606949388981, + 0.0028073724824935198, + 0.002466247184202075, + 0.0024626769591122866, + 0.002851771889254451, + 0.002943725325167179, + 0.0018200321355834603, + 0.0020783166401088238, + 0.0020298894960433245, + 0.0022823321633040905, + 0.0029361986089497805, + 0.0017312420532107353, + 0.0016531490255147219, + 0.002401683246716857, + 0.005252083297818899, + 0.008508416824042797, + 0.019119199365377426, + 0.012141129933297634, + 0.008049007505178452, + 0.008606144227087498, + 0.006138917058706284, + 0.005515390541404486, + 0.004106665961444378, + 0.0083061084151268, + 0.0031621463131159544, + 0.008752738125622272, + 0.005538794212043285, + 0.0050664301961660385, + 0.004493610002100468, + 0.0038001961074769497, + 0.004260132554918528, + 0.00870993547141552, + 0.00475333770737052, + 0.008755470626056194, + 0.01070472877472639, + 0.006842580158263445, + 0.00674207229167223, + 0.006416178308427334, + 0.010748548433184624, + 0.010845721699297428, + 0.011364701204001904, + 0.007140986155718565, + 0.013846637681126595, + 0.011860148049890995, + 0.012395374476909637, + 0.009791758842766285, + 0.0063055939972400665, + 0.008992975577712059, + 0.0052852020598948, + 0.006763427518308163, + 0.016029084101319313, + 0.008878478780388832, + 0.00983180571347475, + 0.006707656662911177, + 0.029581353068351746, + 0.028126811608672142, + 0.04071203991770744, + 0.0325685478746891, + 0.04719044268131256, + 0.04562044143676758, + 0.0329146608710289, + 0.054381124675273895, + 0.026211949065327644, + 0.01908637396991253, + 0.03340228646993637, + 0.029204145073890686, + 0.03692122548818588, + 0.03390924632549286, + 0.040643565356731415, + 0.02980676293373108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.038090646266937256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, tight close-up captures the viewer's gaze firmly within a chilly, serene dojo. The frame is dominated by a cool, celestial light slanting through a high window. The beam carves a sharp path through the dust-filled air, illuminating the heart of the scene. A mysterious warrior stands there, his face largely obscured by a cold, steel mask carved with intricate ancient patterns, revealing only a pair of eagle-like eyes. He clad in a suit of dark, bladed armor, with sharp metal barbs extending from his shoulders and vambraces. The battle-scarred armor gleams darkly in the dim light. Without warning, he swung his bladed arm with a lightning-fast, gusty whistle, striking a solid wooden training dummy before him with steady, deliberate force. With a dull, explosive thud, the dummy disintegrates, sending countless wood chips and fibers of varying sizes scattering violently in all directions. Several of the largest fragments, in slow motion, streak past the camera, their rough fractured texture clearly visible. At this climax of the unleashing violence, the camera precisely captures a flash of icy coldness in the soldier's eyes through the slit in his mask, his pupils instantly constricting. The entire scene is imbued with a powerful tension, from extreme tranquility to a sudden explosion, conveying a sense of focus, coldness, and overwhelming power.", + "seed": 995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_13/index_9862_seed_995_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04379362384788692, + "vlm_logprob": -0.04379362384788692, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04379362384788692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030030755326151848 + ], + "opposite_logprob_tokens": [ + -3.7800307273864746 + ], + "logprob_reward": -0.030030755326151848, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030030755326151848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28629982471466064, + 0.16989630460739136, + 0.05076003074645996, + 0.057747095823287964, + 0.0996876135468483, + 0.07593640685081482, + 0.03560899570584297, + 0.027864595875144005, + 0.14768706262111664, + 0.04851207137107849 + ], + "credit_map": [ + 0.019811198115348816, + 0.022749755531549454, + 0.06641805917024612, + 0.05211896076798439, + 0.022534798830747604, + 0.044938236474990845, + 0.03393113985657692, + 0.023797690868377686, + 0.013054021634161472, + 0.016583168879151344, + 0.00922081246972084, + 0.007486924063414335, + 0.05057840794324875, + 0.05199304595589638, + 0.00929823238402605, + 0.011681688018143177, + 0.0072902194224298, + 0.006376801989972591, + 0.007665371987968683, + 0.007247908506542444, + 0.003958737011998892, + 0.005902612581849098, + 0.00659326184540987, + 0.005725115537643433, + 0.006720538251101971, + 0.006982250604778528, + 0.006820763926953077, + 0.007705519907176495, + 0.004613586235791445, + 0.00957662332803011, + 0.007214679848402739, + 0.008113132789731026, + 0.014789328910410404, + 0.0075340596958994865, + 0.02195657603442669, + 0.013064311817288399, + 0.006047714501619339, + 0.014125037007033825, + 0.009616870433092117, + 0.012553718872368336, + 0.007058554328978062, + 0.010410413146018982, + 0.007306240499019623, + 0.012550265528261662, + 0.009289036504924297, + 0.00993308424949646, + 0.008246760815382004, + 0.01114205177873373, + 0.0055365124717354774, + 0.004841657821089029, + 0.004694616887718439, + 0.0038866379763931036, + 0.005303816869854927, + 0.0033190937247127295, + 0.0049188281409442425, + 0.0031078322790563107, + 0.0048092082142829895, + 0.0026014421600848436, + 0.004195708781480789, + 0.002793032443150878, + 0.003951238002628088, + 0.002385321306064725, + 0.003957930486649275, + 0.0031707168091088533, + 0.025709714740514755, + 0.007113733794540167, + 0.008716980926692486, + 0.005088878329843283, + 0.0111548425629735, + 0.055613260716199875, + 0.0112785454839468, + 0.0230111051350832, + 0.006048413924872875, + 0.005769018549472094, + 0.0061021242290735245, + 0.004787144251167774, + 0.00607773894444108, + 0.005105185322463512, + 0.008705850690603256, + 0.005916595458984375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.030030755326151848 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025133469607681036 + ], + "opposite_logprob_tokens": [ + -6.002513408660889 + ], + "logprob_reward": -0.0025133469607681036, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025133469607681036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09592057019472122, + 0.12718696892261505, + 0.02948080562055111, + 0.05734071135520935, + 0.06659069657325745, + 0.0631142407655716, + 0.08991765230894089, + 0.05748074874281883, + 0.2969295382499695, + 0.11603806912899017 + ], + "credit_map": [ + 0.006030747666954994, + 0.009124544449150562, + 0.011378702707588673, + 0.02104315347969532, + 0.012743799015879631, + 0.009960749186575413, + 0.013096951879560947, + 0.01254191342741251, + 0.0077205076813697815, + 0.01242215558886528, + 0.008310060016810894, + 0.005925052333623171, + 0.039198070764541626, + 0.03482865169644356, + 0.00960671715438366, + 0.009175747632980347, + 0.0022151346784085035, + 0.0022321378346532583, + 0.006738218013197184, + 0.00466426694765687, + 0.0018785725114867091, + 0.0056336489506065845, + 0.0032939871307462454, + 0.0028248385060578585, + 0.004676878917962313, + 0.004665954504162073, + 0.007072454784065485, + 0.010739290155470371, + 0.005921232979744673, + 0.008191737346351147, + 0.007635278627276421, + 0.008437879383563995, + 0.006241196300834417, + 0.007182389497756958, + 0.00858518946915865, + 0.010709565132856369, + 0.00522785447537899, + 0.008721687830984592, + 0.009476798586547375, + 0.010446012951433659, + 0.008675639517605305, + 0.007717772386968136, + 0.006645083427429199, + 0.007824918255209923, + 0.01026441715657711, + 0.007619271520525217, + 0.0062970565631985664, + 0.008070082403719425, + 0.010451401583850384, + 0.00987344328314066, + 0.019417112693190575, + 0.015016420744359493, + 0.008900354616343975, + 0.008593394421041012, + 0.010001379065215588, + 0.007664144970476627, + 0.008465402759611607, + 0.00592779042199254, + 0.009085095487535, + 0.008183651603758335, + 0.006671835668385029, + 0.004076386336237192, + 0.008971716277301311, + 0.006098864134401083, + 0.04882485046982765, + 0.013089248910546303, + 0.021204428747296333, + 0.008523351512849331, + 0.020808300003409386, + 0.11529068648815155, + 0.020446695387363434, + 0.04874198138713837, + 0.017158333212137222, + 0.010791375301778316, + 0.025635795667767525, + 0.014096148312091827, + 0.011130592785775661, + 0.0121811144053936, + 0.013259333558380604, + 0.011785365641117096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.0025133469607681036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306145757436752 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.04306145757436752, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306145757436752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12364739179611206, + 0.0753953829407692, + 0.0743989571928978, + 0.029781106859445572, + 0.12221644073724747, + 0.12937337160110474, + 0.1301984041929245, + 0.13319770991802216, + 0.08606794476509094, + 0.09572326391935349 + ], + "credit_map": [ + 0.007230817805975676, + 0.009032771922647953, + 0.014835099689662457, + 0.03159009665250778, + 0.010493864305317402, + 0.01115464698523283, + 0.02272195555269718, + 0.01658814214169979, + 0.008389606140553951, + 0.008046114817261696, + 0.009388669393956661, + 0.010824938304722309, + 0.011724645271897316, + 0.006215262226760387, + 0.012379784137010574, + 0.008426360785961151, + 0.004645093809813261, + 0.005651187151670456, + 0.014213666319847107, + 0.014309745281934738, + 0.005345079582184553, + 0.011943187564611435, + 0.011180899105966091, + 0.007110099773854017, + 0.0024770721793174744, + 0.0033760382793843746, + 0.0035199064295738935, + 0.005395670421421528, + 0.003403550712391734, + 0.002712831599637866, + 0.004045022185891867, + 0.0048510171473026276, + 0.013401662930846214, + 0.016746174544095993, + 0.017863964661955833, + 0.016598347574472427, + 0.012420259416103363, + 0.01636148802936077, + 0.01446808036416769, + 0.014356459490954876, + 0.010691984556615353, + 0.014976243488490582, + 0.01693151518702507, + 0.022740496322512627, + 0.016632409766316414, + 0.015265226364135742, + 0.015685727819800377, + 0.016449758782982826, + 0.014037523418664932, + 0.017873212695121765, + 0.02086394838988781, + 0.02271939627826214, + 0.01367967575788498, + 0.013110264204442501, + 0.01700340211391449, + 0.01091097854077816, + 0.017389893531799316, + 0.013999122194945812, + 0.017534395679831505, + 0.01902798004448414, + 0.02128804475069046, + 0.009521838277578354, + 0.022987892851233482, + 0.011448542587459087, + 0.007876676507294178, + 0.012280995957553387, + 0.013027826324105263, + 0.011441844515502453, + 0.00921089667826891, + 0.013029022142291069, + 0.009009241126477718, + 0.010191436856985092, + 0.01167269516736269, + 0.009960714727640152, + 0.015599577687680721, + 0.011876179836690426, + 0.009451819583773613, + 0.016028210520744324, + 0.012767210602760315, + 0.008366868831217289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.04306145757436752 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306042939424515 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.04306042939424515, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306042939424515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11317145824432373, + 0.05777667835354805, + 0.054286889731884, + 0.04358839988708496, + 0.0941288024187088, + 0.1361066848039627, + 0.06861954182386398, + 0.08031894266605377, + 0.20270127058029175, + 0.14930133521556854 + ], + "credit_map": [ + 0.010707247070968151, + 0.0059915389865636826, + 0.03456774726510048, + 0.01209475751966238, + 0.01409233920276165, + 0.01608934812247753, + 0.008890102617442608, + 0.010738377459347248, + 0.0051984586752951145, + 0.00666548078879714, + 0.005415564868599176, + 0.004890793934464455, + 0.012422176077961922, + 0.013445447199046612, + 0.004668848589062691, + 0.0050699058920145035, + 0.004881418310105801, + 0.004842955153435469, + 0.011140810325741768, + 0.00965819414705038, + 0.0038253518287092447, + 0.008270161226391792, + 0.006475130561739206, + 0.005192862823605537, + 0.0034684427082538605, + 0.003192547708749771, + 0.0057728104293346405, + 0.008881877176463604, + 0.0031638352666050196, + 0.0050890520215034485, + 0.006268887314945459, + 0.007750947028398514, + 0.00949557963758707, + 0.009442782029509544, + 0.0135801387950778, + 0.014827815815806389, + 0.007620136719197035, + 0.016345271840691566, + 0.010345193557441235, + 0.012471883557736874, + 0.014199177734553814, + 0.017745522782206535, + 0.01874609664082527, + 0.019524887204170227, + 0.014497144147753716, + 0.014620783738791943, + 0.01754644326865673, + 0.019226636737585068, + 0.007299874909222126, + 0.009838941507041454, + 0.012340642511844635, + 0.01198610756546259, + 0.007159237749874592, + 0.005869584158062935, + 0.008032940328121185, + 0.006092216353863478, + 0.00829511508345604, + 0.010164700448513031, + 0.01062988955527544, + 0.013952548615634441, + 0.011936971917748451, + 0.005646951496601105, + 0.012947549112141132, + 0.006745221558958292, + 0.02441279962658882, + 0.008996553719043732, + 0.024329952895641327, + 0.012141859158873558, + 0.013009866699576378, + 0.07113007456064224, + 0.015300496481359005, + 0.033379655331373215, + 0.01954321376979351, + 0.015008357353508472, + 0.02911899983882904, + 0.019982073456048965, + 0.01280113123357296, + 0.02185683138668537, + 0.01757150888442993, + 0.013419223949313164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.04306042939424515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chilly, serene dojo with a high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing a steel mask with intricate patterns and dark, bladed armor visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior swing a bladed arm to strike a wooden training dummy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dummy disintegrate into wood chips and fibers that scatter violently in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight close-up capturing the warrior's eyes constricting through the mask slit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030212998390198 + ], + "opposite_logprob_tokens": [ + -2.350302219390869 + ], + "logprob_reward": -0.10030212998390198, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030212998390198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09268314391374588, + 0.04336794093251228, + 0.06762660294771194, + 0.07613176852464676, + 0.0755489394068718, + 0.05800211429595947, + 0.0905013158917427, + 0.04501686617732048, + 0.3491837978363037, + 0.10193748772144318 + ], + "credit_map": [ + 0.00651624146848917, + 0.008942791260778904, + 0.014889082871377468, + 0.015187501907348633, + 0.00918056070804596, + 0.013037682510912418, + 0.013908096589148045, + 0.01102119404822588, + 0.0033397104125469923, + 0.004824269562959671, + 0.005977939814329147, + 0.0065536582842469215, + 0.006112381815910339, + 0.005584321916103363, + 0.005210630130022764, + 0.005765029229223728, + 0.005066836718469858, + 0.005398657638579607, + 0.00979586225003004, + 0.014504360035061836, + 0.005364235490560532, + 0.005797485820949078, + 0.012798630632460117, + 0.008900536224246025, + 0.006553557235747576, + 0.006817474029958248, + 0.00877477414906025, + 0.01625790074467659, + 0.009615638293325901, + 0.007749948184937239, + 0.010263556614518166, + 0.010098918341100216, + 0.008637419901788235, + 0.00985315814614296, + 0.009937089867889881, + 0.012124973349273205, + 0.006395371165126562, + 0.012088462710380554, + 0.007551833521574736, + 0.008960628882050514, + 0.0063757519237697124, + 0.007985549978911877, + 0.006632979493588209, + 0.008550846949219704, + 0.00798479001969099, + 0.006275601219385862, + 0.007776523008942604, + 0.006420067511498928, + 0.009948833845555782, + 0.01320762000977993, + 0.01218375377357006, + 0.014639502391219139, + 0.011727307923138142, + 0.009650344029068947, + 0.010042373090982437, + 0.009101582691073418, + 0.006378548219799995, + 0.004848944488912821, + 0.005385236348956823, + 0.004653549287468195, + 0.007282649166882038, + 0.004264266230165958, + 0.007488531060516834, + 0.004715139511972666, + 0.02973870374262333, + 0.02798358164727688, + 0.024577798321843147, + 0.021343162283301353, + 0.039200328290462494, + 0.11025786399841309, + 0.033082351088523865, + 0.06300000846385956, + 0.009833643212914467, + 0.008307114243507385, + 0.014276676811277866, + 0.00914010964334011, + 0.012165813706815243, + 0.013544336892664433, + 0.018909776583313942, + 0.015760019421577454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.10030212998390198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_tensor([5723, 5723])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_tensor([5723, 5723])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59ec7f930ec66e29db2a5f541f628c84a8ce63f8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_tensor([5723, 5723])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a first-person perspective, the camera executes a smooth, gliding forward tracking shot through the sun-drenched garden, gradually bringing a magnificent cluster of large, lush peonies into sharp focus. The flowers are a symphony of pink, with a subtle blush on the outer petals, deepening to a vibrant magenta toward the densely layered core. Each heavy, spherical blossom is composed of countless ruffled, velvety petals, almost translucent. A gentle breeze causes the heavy flower heads to sway rhythmically and tremble gently on their sturdy green stems in a delicate and natural dance. The entire scene is bathed in the warm, inviting golden light of the afternoon, which filters through the surrounding foliage, creating a softly dappled effect and giving the petals an inner glow. The background dissolves into a beautiful bokeh of rich greens and blurred floral colors, creating an atmosphere of profound tranquility and serene beauty, capturing a fleeting, perfect moment in nature.", + "seed": 10870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_5723_seed_10870_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1992964579258114, + "vlm_logprob": -0.1992964579258114, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.1992964579258114, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8432210683822632 + ], + "opposite_logprob_tokens": [ + -0.5932210683822632 + ], + "logprob_reward": -0.8432210683822632, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8432210683822632, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13435272872447968, + 0.14864501357078552, + 0.07478291541337967, + 0.1962655484676361, + 0.12215898185968399, + 0.057932861149311066, + 0.06914642453193665, + 0.05811107158660889, + 0.08196836709976196, + 0.05663616210222244 + ], + "credit_map": [ + 0.014073754660785198, + 0.018176738172769547, + 0.016474714502692223, + 0.01440054178237915, + 0.015311208553612232, + 0.01104785967618227, + 0.015650689601898193, + 0.02921721152961254, + 0.01633790321648121, + 0.022328892722725868, + 0.02871742844581604, + 0.030344203114509583, + 0.02424047887325287, + 0.007177432999014854, + 0.007908920757472515, + 0.011589748784899712, + 0.005964886397123337, + 0.010812029242515564, + 0.010804101824760437, + 0.00698152044788003, + 0.008058680221438408, + 0.007736856583505869, + 0.01287791132926941, + 0.011546927504241467, + 0.01317712664604187, + 0.04493727907538414, + 0.051400598138570786, + 0.03091358207166195, + 0.011656263843178749, + 0.013207545503973961, + 0.014540689066052437, + 0.01643245853483677, + 0.011837378144264221, + 0.012529473751783371, + 0.014707106165587902, + 0.01535933930426836, + 0.01927088387310505, + 0.020617052912712097, + 0.015866350382566452, + 0.011971400119364262, + 0.005350853316485882, + 0.00529659166932106, + 0.012373843230307102, + 0.006409857887774706, + 0.012550875544548035, + 0.00610263692215085, + 0.00472071161493659, + 0.005127491429448128, + 0.004651910159736872, + 0.006112759467214346, + 0.007657533977180719, + 0.01040105801075697, + 0.007368539925664663, + 0.011086891405284405, + 0.007442981004714966, + 0.014424753375351429, + 0.009476475417613983, + 0.00954017136245966, + 0.007384098134934902, + 0.006480303127318621, + 0.006020545959472656, + 0.005865174811333418, + 0.006486422847956419, + 0.006857882719486952, + 0.003735826350748539, + 0.02501264587044716, + 0.026562659069895744, + 0.005754711572080851, + 0.0043357606045901775, + 0.005736886989325285, + 0.005177006125450134, + 0.00565287284553051, + 0.0034728909377008677, + 0.002492601051926613, + 0.020920665934681892, + 0.003620330011472106, + 0.004076289478689432, + 0.015100781805813313, + 0.0038407398387789726, + 0.0031118609476834536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.8432210683822632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297792196273804 + ], + "opposite_logprob_tokens": [ + -2.237977981567383 + ], + "logprob_reward": -0.11297792196273804, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297792196273804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07511848211288452, + 0.13764099776744843, + 0.059609707444906235, + 0.25540345907211304, + 0.15064401924610138, + 0.08603067696094513, + 0.06421664357185364, + 0.0413239523768425, + 0.08663254231214523, + 0.043379541486501694 + ], + "credit_map": [ + 0.011208822019398212, + 0.011171210557222366, + 0.009737849235534668, + 0.014212453737854958, + 0.009596885181963444, + 0.0048370882868766785, + 0.007132841274142265, + 0.007221324369311333, + 0.014337059110403061, + 0.02711993083357811, + 0.02248862199485302, + 0.02658929117023945, + 0.014734080992639065, + 0.007384990807622671, + 0.010106170549988747, + 0.014880841597914696, + 0.007068413309752941, + 0.010591553524136543, + 0.007815634831786156, + 0.009536707773804665, + 0.005249124951660633, + 0.004986544139683247, + 0.007110179401934147, + 0.007251546252518892, + 0.018321050330996513, + 0.07256783545017242, + 0.09472482651472092, + 0.02670787274837494, + 0.012693481519818306, + 0.007709887810051441, + 0.012029719538986683, + 0.010648781433701515, + 0.00969285424798727, + 0.012034427374601364, + 0.01648605614900589, + 0.017683541402220726, + 0.021961916238069534, + 0.03144573047757149, + 0.022175442427396774, + 0.019164040684700012, + 0.010400406084954739, + 0.012642144225537777, + 0.015483667142689228, + 0.014219528064131737, + 0.010170611552894115, + 0.006973939947783947, + 0.006556381471455097, + 0.009583990089595318, + 0.004881700035184622, + 0.007101224735379219, + 0.008764746598899364, + 0.00810323841869831, + 0.0062748440541327, + 0.008083509281277657, + 0.008727018721401691, + 0.012280355207622051, + 0.005047729704529047, + 0.005370841361582279, + 0.005395462736487389, + 0.0056557320058345795, + 0.004040445666760206, + 0.0037911341059952974, + 0.005113959312438965, + 0.006908644922077656, + 0.00419299490749836, + 0.021959610283374786, + 0.022690987214446068, + 0.009242206811904907, + 0.008371174335479736, + 0.006664934102445841, + 0.006977550685405731, + 0.006533083040267229, + 0.0026097078807651997, + 0.0030364212580025196, + 0.015720589086413383, + 0.0035271779634058475, + 0.0036298984196037054, + 0.009201336652040482, + 0.003212832845747471, + 0.0024415780790150166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.11297792196273804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013488966505974531 + ], + "opposite_logprob_tokens": [ + -6.626348972320557 + ], + "logprob_reward": -0.0013488966505974531, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013488966505974531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13790081441402435, + 0.058787811547517776, + 0.08223484456539154, + 0.10610625147819519, + 0.06605689227581024, + 0.1825476586818695, + 0.07944051921367645, + 0.07733834534883499, + 0.0546429343521595, + 0.15494391322135925 + ], + "credit_map": [ + 0.021442538127303123, + 0.0185651034116745, + 0.017521237954497337, + 0.021296972408890724, + 0.01175805926322937, + 0.009516277350485325, + 0.017591312527656555, + 0.02020932175219059, + 0.007891733199357986, + 0.00841282494366169, + 0.010533374734222889, + 0.010218985378742218, + 0.010596602223813534, + 0.0031727792229503393, + 0.0037553426809608936, + 0.004206178244203329, + 0.009524728171527386, + 0.013806933537125587, + 0.011763839982450008, + 0.008246159180998802, + 0.010112290270626545, + 0.008548364043235779, + 0.010545284487307072, + 0.009687254205346107, + 0.00745328888297081, + 0.02240612916648388, + 0.03517686948180199, + 0.011851600371301174, + 0.00702258525416255, + 0.0061591388657689095, + 0.009188306517899036, + 0.006848348304629326, + 0.004989199806004763, + 0.005200282670557499, + 0.008093426004052162, + 0.007953211665153503, + 0.008999975398182869, + 0.012283604592084885, + 0.011058208532631397, + 0.007478994317352772, + 0.019934196025133133, + 0.011901712976396084, + 0.05707230791449547, + 0.018969135358929634, + 0.03376148268580437, + 0.013567407615482807, + 0.01545393280684948, + 0.011887500993907452, + 0.007023352663964033, + 0.007467451971024275, + 0.014411665499210358, + 0.00883944146335125, + 0.00737793417647481, + 0.015299500897526741, + 0.00882669910788536, + 0.010194485075771809, + 0.009194849990308285, + 0.011364060454070568, + 0.012121252715587616, + 0.010585351847112179, + 0.006867471616715193, + 0.006877804175019264, + 0.011820378713309765, + 0.008507182821631432, + 0.003245886182412505, + 0.009705573320388794, + 0.009103377349674702, + 0.009026482701301575, + 0.0059824176132678986, + 0.005348968785256147, + 0.005552307236939669, + 0.006677926052361727, + 0.010773235000669956, + 0.007835596799850464, + 0.04951091483235359, + 0.010233230888843536, + 0.013236412778496742, + 0.03967329114675522, + 0.013134920969605446, + 0.010546320118010044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0013488966505974531 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052507114596664906 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.0052507114596664906, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052507114596664906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125990629196167, + 0.0705922544002533, + 0.08560418337583542, + 0.0755753442645073, + 0.1623525619506836, + 0.20625147223472595, + 0.033350784331560135, + 0.054207246750593185, + 0.14045308530330658, + 0.045622389763593674 + ], + "credit_map": [ + 0.01964881271123886, + 0.018360579386353493, + 0.017849376425147057, + 0.015929678454995155, + 0.017309337854385376, + 0.00848589651286602, + 0.011779546737670898, + 0.01662740670144558, + 0.008193104527890682, + 0.011317572556436062, + 0.013741031289100647, + 0.015899790450930595, + 0.006984081584960222, + 0.004011749755591154, + 0.004537314176559448, + 0.0059076049365103245, + 0.009285883978009224, + 0.016374733299016953, + 0.013717281632125378, + 0.00974317453801632, + 0.010371755808591843, + 0.00852235872298479, + 0.010209409520030022, + 0.007379590068012476, + 0.004440925549715757, + 0.022744929417967796, + 0.02674783766269684, + 0.008615214377641678, + 0.004125918727368116, + 0.002136343391612172, + 0.0031417934224009514, + 0.003622376825660467, + 0.016310717910528183, + 0.012117280624806881, + 0.014574536122381687, + 0.017802296206355095, + 0.024103373289108276, + 0.03378598764538765, + 0.02211630530655384, + 0.02154206484556198, + 0.03368939831852913, + 0.016983065754175186, + 0.08188292384147644, + 0.017797334119677544, + 0.02317817509174347, + 0.010799790732562542, + 0.009750633500516415, + 0.012170156463980675, + 0.0028607379645109177, + 0.0021136789582669735, + 0.006510637700557709, + 0.003119087079539895, + 0.003397746477276087, + 0.008276747539639473, + 0.002729888306930661, + 0.004342259839177132, + 0.006493877153843641, + 0.006697503384202719, + 0.007097057998180389, + 0.009696063585579395, + 0.005143263842910528, + 0.004293938633054495, + 0.006737654563039541, + 0.008047886192798615, + 0.00710771419107914, + 0.04840909317135811, + 0.05236801877617836, + 0.007406902965158224, + 0.007632785476744175, + 0.005307047627866268, + 0.0057175373658537865, + 0.006503979209810495, + 0.00302890338934958, + 0.0019278408726677299, + 0.015912342816591263, + 0.0027085430920124054, + 0.0035621169954538345, + 0.012902994640171528, + 0.002623338019475341, + 0.0029563119169324636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.0052507114596664906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033683691173791885 + ], + "opposite_logprob_tokens": [ + -3.4086837768554688 + ], + "logprob_reward": -0.033683691173791885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033683691173791885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10855358839035034, + 0.06209662929177284, + 0.06586823612451553, + 0.06990870833396912, + 0.2791348099708557, + 0.10494200140237808, + 0.11809898167848587, + 0.06936585903167725, + 0.09673269838094711, + 0.025298545137047768 + ], + "credit_map": [ + 0.014164303429424763, + 0.014990519732236862, + 0.014597040601074696, + 0.01641976833343506, + 0.012099614366889, + 0.007375813089311123, + 0.011647161096334457, + 0.01725936308503151, + 0.005442421417683363, + 0.008561003021895885, + 0.008857164531946182, + 0.01620637997984886, + 0.00633464427664876, + 0.0035546242725104094, + 0.005487540271133184, + 0.007652848493307829, + 0.010674871504306793, + 0.012625706382095814, + 0.009142416529357433, + 0.006717185024172068, + 0.008329623378813267, + 0.006150087807327509, + 0.006867839023470879, + 0.0053605083376169205, + 0.0062450203113257885, + 0.009751738980412483, + 0.016137035563588142, + 0.00880455318838358, + 0.0050825621001422405, + 0.007182125933468342, + 0.007078878581523895, + 0.009626790881156921, + 0.01927429437637329, + 0.023275839164853096, + 0.028121422976255417, + 0.029713967815041542, + 0.042519234120845795, + 0.06269368529319763, + 0.04072485864162445, + 0.03281151503324509, + 0.01488177664577961, + 0.008069836534559727, + 0.04019360989332199, + 0.009319608099758625, + 0.015145472250878811, + 0.006255850195884705, + 0.004873636178672314, + 0.00620220648124814, + 0.011823280714452267, + 0.012252142652869225, + 0.020728696137666702, + 0.012540116906166077, + 0.013676155358552933, + 0.008984928019344807, + 0.014273802749812603, + 0.023819850757718086, + 0.007194200064986944, + 0.0092964181676507, + 0.007709130644798279, + 0.010214902460575104, + 0.00813707709312439, + 0.0076779150404036045, + 0.009808402508497238, + 0.009327816776931286, + 0.004934628494083881, + 0.024383913725614548, + 0.027702482417225838, + 0.009831065312027931, + 0.008426415733993053, + 0.00548206502571702, + 0.007551372982561588, + 0.008420748636126518, + 0.0017613645177334547, + 0.00281394156627357, + 0.005257096141576767, + 0.002843258436769247, + 0.0024890603963285685, + 0.003918853122740984, + 0.003373903688043356, + 0.002841066336259246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.033683691173791885 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a first-person perspective, the camera executes a smooth, gliding forward tracking shot through the sun-drenched garden, gradually bringing a magnificent cluster of large, lush peonies into sharp focus. The flowers are a symphony of pink, with a subtle blush on the outer petals, deepening to a vibrant magenta toward the densely layered core. Each heavy, spherical blossom is composed of countless ruffled, velvety petals, almost translucent. A gentle breeze causes the heavy flower heads to sway rhythmically and tremble gently on their sturdy green stems in a delicate and natural dance. The entire scene is bathed in the warm, inviting golden light of the afternoon, which filters through the surrounding foliage, creating a softly dappled effect and giving the petals an inner glow. The background dissolves into a beautiful bokeh of rich greens and blurred floral colors, creating an atmosphere of profound tranquility and serene beauty, capturing a fleeting, perfect moment in nature.", + "seed": 10871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_5723_seed_10871_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18344745410140603, + "vlm_logprob": -0.18344745410140603, + "vlm_yes_no_margin": 3.6249999523162844, + "vlm_token_credit": -0.18344745410140603, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7785050868988037 + ], + "opposite_logprob_tokens": [ + -0.6535050868988037 + ], + "logprob_reward": -0.7785050868988037, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7785050868988037, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2258566915988922, + 0.09649243950843811, + 0.030362321063876152, + 0.08056344091892242, + 0.11270742863416672, + 0.08194085955619812, + 0.09478795528411865, + 0.16630437970161438, + 0.06420309096574783, + 0.046781379729509354 + ], + "credit_map": [ + 0.013182461261749268, + 0.013333623297512531, + 0.025169482454657555, + 0.07136092334985733, + 0.032845038920640945, + 0.011111414059996605, + 0.012464739382266998, + 0.04638902097940445, + 0.009406417608261108, + 0.008299513719975948, + 0.0075722187757492065, + 0.006292636971920729, + 0.030301425606012344, + 0.006446606479585171, + 0.004167274571955204, + 0.024006349965929985, + 0.0019514437299221754, + 0.003483391832560301, + 0.003937183879315853, + 0.004111713729798794, + 0.003928358666598797, + 0.0034859252627938986, + 0.004368465393781662, + 0.005095840897411108, + 0.00605540256947279, + 0.022407278418540955, + 0.006611536722630262, + 0.006800064351409674, + 0.008401218801736832, + 0.007979565300047398, + 0.004778438713401556, + 0.017529934644699097, + 0.011274334974586964, + 0.005812362302094698, + 0.033353738486766815, + 0.00756604690104723, + 0.012675628997385502, + 0.0077328067272901535, + 0.009698032401502132, + 0.02459447830915451, + 0.006918116472661495, + 0.012896059080958366, + 0.010790103115141392, + 0.016275113448500633, + 0.008816387504339218, + 0.00797208584845066, + 0.00817244965583086, + 0.010100551880896091, + 0.007800532970577478, + 0.018060075119137764, + 0.011115442961454391, + 0.01079618465155363, + 0.01765712909400463, + 0.006578123662620783, + 0.009067247621715069, + 0.013713221065700054, + 0.0247661042958498, + 0.022858569398522377, + 0.029268786311149597, + 0.027474353089928627, + 0.014924700371921062, + 0.013613170944154263, + 0.014395307749509811, + 0.019003381952643394, + 0.008917578496038914, + 0.005197552964091301, + 0.008390459232032299, + 0.009884694591164589, + 0.006222808267921209, + 0.010341301560401917, + 0.0051434398628771305, + 0.010105254128575325, + 0.003962832968682051, + 0.004046281334012747, + 0.00499323196709156, + 0.004587843082845211, + 0.004283159505575895, + 0.008743499405682087, + 0.00401543453335762, + 0.01214909553527832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.7785050868988037 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297951638698578 + ], + "opposite_logprob_tokens": [ + -2.2379794120788574 + ], + "logprob_reward": -0.11297951638698578, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297951638698578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20177242159843445, + 0.10226840525865555, + 0.03188890963792801, + 0.08299527317285538, + 0.11316719651222229, + 0.06782086938619614, + 0.038217268884181976, + 0.16877010464668274, + 0.1583314836025238, + 0.03476811572909355 + ], + "credit_map": [ + 0.013719317503273487, + 0.014859476126730442, + 0.021846843883395195, + 0.058455221354961395, + 0.04158834367990494, + 0.01390625350177288, + 0.015578092075884342, + 0.021818866953253746, + 0.01082427054643631, + 0.009639949537813663, + 0.006815528497099876, + 0.009344149380922318, + 0.03331136330962181, + 0.008178716525435448, + 0.005352133885025978, + 0.01880229450762272, + 0.002693836111575365, + 0.0036939107812941074, + 0.005270195659250021, + 0.005419260356575251, + 0.0033678438048809767, + 0.0034982848446816206, + 0.0030835981015115976, + 0.00486197741702199, + 0.004789508413523436, + 0.022502614185214043, + 0.006958942860364914, + 0.006914952769875526, + 0.005933653563261032, + 0.008196468465030193, + 0.005109624937176704, + 0.022589510306715965, + 0.009163029491901398, + 0.007375206798315048, + 0.0404847152531147, + 0.009590709581971169, + 0.013113321736454964, + 0.006832608953118324, + 0.006889559328556061, + 0.019718050956726074, + 0.0061386991292238235, + 0.008281663991510868, + 0.00911643821746111, + 0.012329474091529846, + 0.007747202180325985, + 0.007376799359917641, + 0.006850639823824167, + 0.009979956783354282, + 0.0031297330278903246, + 0.005534430034458637, + 0.004298546351492405, + 0.004993600305169821, + 0.006581481080502272, + 0.0038223685696721077, + 0.004764081910252571, + 0.005093027371913195, + 0.025958361104130745, + 0.01887638494372368, + 0.026229439303278923, + 0.035958267748355865, + 0.01323680579662323, + 0.012517813593149185, + 0.01735103316605091, + 0.01864200085401535, + 0.014539639465510845, + 0.012354649603366852, + 0.02239360846579075, + 0.0172860287129879, + 0.010552417486906052, + 0.02703705057501793, + 0.016913140192627907, + 0.0372549444437027, + 0.0036491751670837402, + 0.003539757337421179, + 0.006877833511680365, + 0.0043519288301467896, + 0.0038278924766927958, + 0.0034616366028785706, + 0.004341848194599152, + 0.0047180429100990295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.11297951638698578 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00083078199531883 + ], + "opposite_logprob_tokens": [ + -7.12583065032959 + ], + "logprob_reward": -0.00083078199531883, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.00083078199531883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1327410340309143, + 0.1268962025642395, + 0.0471455380320549, + 0.06350056082010269, + 0.45574045181274414, + 0.027021532878279686, + 0.04953869432210922, + 0.03785567358136177, + 0.03194735571742058, + 0.027612969279289246 + ], + "credit_map": [ + 0.02167532406747341, + 0.010930280201137066, + 0.015292057767510414, + 0.01487671211361885, + 0.020396223291754723, + 0.016264313831925392, + 0.012874077074229717, + 0.02043204940855503, + 0.012494194321334362, + 0.011964606121182442, + 0.008476868271827698, + 0.00703468406572938, + 0.04476788640022278, + 0.007669036742299795, + 0.005229332484304905, + 0.029259612783789635, + 0.0031904992647469044, + 0.005097406916320324, + 0.006502896081656218, + 0.005792187061160803, + 0.006041180342435837, + 0.004940188955515623, + 0.007711018435657024, + 0.007870168425142765, + 0.0033056342508643866, + 0.014770478941500187, + 0.006302129942923784, + 0.00470951059833169, + 0.005919826216995716, + 0.006429184228181839, + 0.004641974810510874, + 0.017421826720237732, + 0.032765407115221024, + 0.026738008484244347, + 0.16490435600280762, + 0.03793111816048622, + 0.05209378898143768, + 0.026343032717704773, + 0.027923032641410828, + 0.08704173564910889, + 0.0019091073190793395, + 0.0032845509704202414, + 0.003781890030950308, + 0.004242234863340855, + 0.005179929547011852, + 0.0024025030434131622, + 0.0025224725250154734, + 0.003698847023770213, + 0.003175607416778803, + 0.009248150512576103, + 0.0072617786936461926, + 0.005616731476038694, + 0.009647885337471962, + 0.0034799373243004084, + 0.005133686121553183, + 0.005974920466542244, + 0.005386800970882177, + 0.005176176317036152, + 0.006493035703897476, + 0.006237541325390339, + 0.0043738484382629395, + 0.0028641114477068186, + 0.0030683972872793674, + 0.004255763255059719, + 0.003199370577931404, + 0.0037841072771698236, + 0.004064100328832865, + 0.004517631139606237, + 0.0034697232767939568, + 0.0034291043411940336, + 0.0034461345057934523, + 0.0060371882282197475, + 0.0027418474201112986, + 0.002231279853731394, + 0.005081582348793745, + 0.0029796448070555925, + 0.002367408247664571, + 0.003406975883990526, + 0.0032403937075287104, + 0.005563840735703707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.00083078199531883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24059829115867615, + 0.03438173234462738, + 0.03555642068386078, + 0.09224934130907059, + 0.038598671555519104, + 0.018403220921754837, + 0.016136594116687775, + 0.09979788959026337, + 0.2628316581249237, + 0.16144612431526184 + ], + "credit_map": [ + 0.011985844932496548, + 0.012497523799538612, + 0.021441718563437462, + 0.08010192215442657, + 0.03385133296251297, + 0.012118548154830933, + 0.014610249549150467, + 0.053991157561540604, + 0.004509583115577698, + 0.002405926352366805, + 0.0032845656387507915, + 0.002766992663964629, + 0.009514234028756618, + 0.0021671417634934187, + 0.00191908807028085, + 0.007814198732376099, + 0.002564588561654091, + 0.0034150381106883287, + 0.005834247451275587, + 0.0042082262225449085, + 0.005008557345718145, + 0.004472100175917149, + 0.004469018895179033, + 0.005584642756730318, + 0.005863022059202194, + 0.02622120827436447, + 0.008681438863277435, + 0.006720458157360554, + 0.00453601311892271, + 0.0060018994845449924, + 0.00540495291352272, + 0.028820347040891647, + 0.007410023827105761, + 0.003793618641793728, + 0.010211760178208351, + 0.0043282234109938145, + 0.004400560632348061, + 0.0025646709837019444, + 0.0026852465234696865, + 0.0032045673578977585, + 0.0015529432566836476, + 0.0026237978599965572, + 0.002518530236557126, + 0.0036098435521125793, + 0.0024432155769318342, + 0.001614683773368597, + 0.001894444110803306, + 0.002145762788131833, + 0.0012286573182791471, + 0.002427751198410988, + 0.0024339465890079737, + 0.0023207201156765223, + 0.0026678668800741434, + 0.0013706263853237033, + 0.0016790225636214018, + 0.0020080036483705044, + 0.010124241933226585, + 0.01922297663986683, + 0.016694582998752594, + 0.013892187736928463, + 0.011646706610918045, + 0.009107277728617191, + 0.00794216338545084, + 0.011167751625180244, + 0.021203914657235146, + 0.012926748022437096, + 0.05115286633372307, + 0.015382042154669762, + 0.012210763059556484, + 0.04524480178952217, + 0.017800509929656982, + 0.08691002428531647, + 0.009604835882782936, + 0.009267115965485573, + 0.015854142606258392, + 0.007450969889760017, + 0.01055549830198288, + 0.050908513367176056, + 0.009837500751018524, + 0.04796753078699112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.006738912779837847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1408165842294693, + 0.013971447944641113, + 0.05346975475549698, + 0.02851996198296547, + 0.031268563121557236, + 0.025217436254024506, + 0.033082231879234314, + 0.40547576546669006, + 0.11208704859018326, + 0.15609121322631836 + ], + "credit_map": [ + 0.008107775822281837, + 0.008697343990206718, + 0.013182605616748333, + 0.046010978519916534, + 0.02393103577196598, + 0.00772910937666893, + 0.009806247428059578, + 0.02335149422287941, + 0.0022546472027897835, + 0.0011852221796289086, + 0.002152703469619155, + 0.0018655004678294063, + 0.0014557656832039356, + 0.001367067452520132, + 0.0013143701944500208, + 0.0023761719930917025, + 0.0038384385406970978, + 0.005727805197238922, + 0.008316953666508198, + 0.009455790743231773, + 0.006380714010447264, + 0.00645983824506402, + 0.00499463826417923, + 0.008295576088130474, + 0.0027373519260436296, + 0.002804130082949996, + 0.00481896847486496, + 0.003331006970256567, + 0.003914311528205872, + 0.0039543528109788895, + 0.003226973582059145, + 0.0037328661419451237, + 0.004791005980223417, + 0.002958276541903615, + 0.010130598209798336, + 0.0027339875232428312, + 0.003716401755809784, + 0.0016592537285760045, + 0.0017442115349695086, + 0.00353482854552567, + 0.0018449228955432773, + 0.0034981481730937958, + 0.0029440841171890497, + 0.0029510161839425564, + 0.0031820780131965876, + 0.0033057923428714275, + 0.00309204519726336, + 0.0043993499130010605, + 0.0027406937442719936, + 0.002477504312992096, + 0.006903276313096285, + 0.005662357900291681, + 0.0032638623379170895, + 0.002228168770670891, + 0.0019327522022649646, + 0.007873613387346268, + 0.043712567538022995, + 0.07686804980039597, + 0.07307454943656921, + 0.054393965750932693, + 0.049795981496572495, + 0.034870050847530365, + 0.02653154358267784, + 0.046229057013988495, + 0.009882521815598011, + 0.005465231835842133, + 0.01915142312645912, + 0.008746549487113953, + 0.006422590464353561, + 0.019489390775561333, + 0.009423760697245598, + 0.033505577594041824, + 0.011641884222626686, + 0.0071168155409395695, + 0.039436548948287964, + 0.007955004461109638, + 0.00973033532500267, + 0.0451815202832222, + 0.005492400377988815, + 0.029536690562963486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.018182972446084023 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a first-person perspective, the camera executes a smooth, gliding forward tracking shot through the sun-drenched garden, gradually bringing a magnificent cluster of large, lush peonies into sharp focus. The flowers are a symphony of pink, with a subtle blush on the outer petals, deepening to a vibrant magenta toward the densely layered core. Each heavy, spherical blossom is composed of countless ruffled, velvety petals, almost translucent. A gentle breeze causes the heavy flower heads to sway rhythmically and tremble gently on their sturdy green stems in a delicate and natural dance. The entire scene is bathed in the warm, inviting golden light of the afternoon, which filters through the surrounding foliage, creating a softly dappled effect and giving the petals an inner glow. The background dissolves into a beautiful bokeh of rich greens and blurred floral colors, creating an atmosphere of profound tranquility and serene beauty, capturing a fleeting, perfect moment in nature.", + "seed": 10872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_5723_seed_10872_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1788944644620642, + "vlm_logprob": -0.1788944644620642, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.1788944644620642, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7230225801467896 + ], + "opposite_logprob_tokens": [ + -0.7230225801467896 + ], + "logprob_reward": -0.7230225801467896, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7230225801467896, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18181411921977997, + 0.08377545326948166, + 0.2403787076473236, + 0.05155355483293533, + 0.03077385015785694, + 0.036130376160144806, + 0.050208281725645065, + 0.16894608736038208, + 0.057822808623313904, + 0.09859677404165268 + ], + "credit_map": [ + 0.027868090197443962, + 0.019601579755544662, + 0.028783855959773064, + 0.034791748970746994, + 0.015890492126345634, + 0.012326101772487164, + 0.016810217872262, + 0.025742026045918465, + 0.011303972452878952, + 0.010104221291840076, + 0.019249552860856056, + 0.015922430902719498, + 0.004764417186379433, + 0.005007568746805191, + 0.007412681356072426, + 0.010010600090026855, + 0.01819724775850773, + 0.008268977515399456, + 0.08464682102203369, + 0.01879338175058365, + 0.0078898249194026, + 0.008853847160935402, + 0.0194191075861454, + 0.07430949062108994, + 0.006051152013242245, + 0.005292369984090328, + 0.007253978867083788, + 0.007438300177454948, + 0.004781223833560944, + 0.0036354532930999994, + 0.007143754046410322, + 0.00995731446892023, + 0.002801991533488035, + 0.0026543918065726757, + 0.004507888574153185, + 0.005080792587250471, + 0.0022912435233592987, + 0.0024985340423882008, + 0.004369064699858427, + 0.006569940596818924, + 0.006293407175689936, + 0.004322214517742395, + 0.005085355136543512, + 0.006440820638090372, + 0.002904960885643959, + 0.002271875739097595, + 0.004721125587821007, + 0.0040906136855483055, + 0.007993888109922409, + 0.012361975386738777, + 0.010119534097611904, + 0.003954372368752956, + 0.004311185795813799, + 0.002414073795080185, + 0.0046423738822340965, + 0.004410869907587767, + 0.04713067412376404, + 0.010675239376723766, + 0.0390080064535141, + 0.01830054633319378, + 0.011218919418752193, + 0.009217469021677971, + 0.013416890054941177, + 0.019978325814008713, + 0.009776534512639046, + 0.00819720420986414, + 0.012788772583007812, + 0.0054434193298220634, + 0.008054482750594616, + 0.0029617438558489084, + 0.006484649144113064, + 0.004115996416658163, + 0.025052489712834358, + 0.013726437464356422, + 0.017079032957553864, + 0.008807151578366756, + 0.006503045558929443, + 0.005864361766725779, + 0.009319092147052288, + 0.012245154939591885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.7230225801467896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283889532089233 + ], + "opposite_logprob_tokens": [ + -2.017838954925537 + ], + "logprob_reward": -0.14283889532089233, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283889532089233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09854099899530411, + 0.047914791852235794, + 0.23558354377746582, + 0.026294030249118805, + 0.031078770756721497, + 0.04768054559826851, + 0.042125094681978226, + 0.32648584246635437, + 0.0368288978934288, + 0.10746748745441437 + ], + "credit_map": [ + 0.015513051301240921, + 0.00869480986148119, + 0.014338381588459015, + 0.016523409634828568, + 0.010931914672255516, + 0.007135682739317417, + 0.011956878006458282, + 0.013446873053908348, + 0.007065455429255962, + 0.005645876284688711, + 0.008220708929002285, + 0.010421602986752987, + 0.002901277504861355, + 0.003142624394968152, + 0.004630190320312977, + 0.005887064151465893, + 0.017901813611388206, + 0.006672639399766922, + 0.08886732161045074, + 0.018689516931772232, + 0.005641712807118893, + 0.0079653887078166, + 0.01801135390996933, + 0.07183383405208588, + 0.0049753692001104355, + 0.0016626202268525958, + 0.0035133722703903913, + 0.00379705848172307, + 0.002234812593087554, + 0.002314805518835783, + 0.0033322202507406473, + 0.004463774152100086, + 0.005471094511449337, + 0.0028278902173042297, + 0.004545816220343113, + 0.0037884959019720554, + 0.002641132101416588, + 0.0026751458644866943, + 0.0038797513116151094, + 0.00524944718927145, + 0.009461513720452785, + 0.005376380402594805, + 0.006851520389318466, + 0.00806434266269207, + 0.003822088008746505, + 0.0030462846625596285, + 0.006343794986605644, + 0.00471462681889534, + 0.00708641204982996, + 0.004406751599162817, + 0.004993996117264032, + 0.008053574711084366, + 0.004114188253879547, + 0.003048731479793787, + 0.004997409880161285, + 0.00542403245344758, + 0.11288432031869888, + 0.030560635030269623, + 0.08076398074626923, + 0.020232416689395905, + 0.016547292470932007, + 0.011702591553330421, + 0.017559684813022614, + 0.03623494505882263, + 0.005841949488967657, + 0.003937895409762859, + 0.007545690052211285, + 0.004304101690649986, + 0.00472915219143033, + 0.0028211669996380806, + 0.004828001372516155, + 0.0028209462761878967, + 0.029275082051753998, + 0.013492732308804989, + 0.026387913152575493, + 0.009207583963871002, + 0.0060248724184930325, + 0.005010359920561314, + 0.007951682433485985, + 0.010117268189787865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.14283889532089233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006517431465908885 + ], + "opposite_logprob_tokens": [ + -7.375651836395264 + ], + "logprob_reward": -0.0006517431465908885, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006517431465908885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08452723920345306, + 0.048718784004449844, + 0.14325407147407532, + 0.020939182490110397, + 0.05273152515292168, + 0.01876424252986908, + 0.04769419506192207, + 0.49475106596946716, + 0.048268720507621765, + 0.040351033210754395 + ], + "credit_map": [ + 0.009137135930359364, + 0.009227211587131023, + 0.015292306430637836, + 0.01302201859652996, + 0.00813744030892849, + 0.008609505370259285, + 0.008521506562829018, + 0.012580111622810364, + 0.005406167823821306, + 0.004224264994263649, + 0.01042796391993761, + 0.008538149297237396, + 0.005123229697346687, + 0.004481537267565727, + 0.0048794979229569435, + 0.005637969356030226, + 0.010740529745817184, + 0.0038435617461800575, + 0.051241301000118256, + 0.011431491002440453, + 0.0037495698779821396, + 0.004925939254462719, + 0.010585223324596882, + 0.0467364639043808, + 0.0031785129103809595, + 0.002112835180014372, + 0.0025421343743801117, + 0.0028470323886722326, + 0.002141776494681835, + 0.001667601871304214, + 0.0025008819065988064, + 0.0039484077133238316, + 0.006774727255105972, + 0.004864900838583708, + 0.006726226303726435, + 0.007939678616821766, + 0.005553567782044411, + 0.00530595425516367, + 0.007540341001003981, + 0.008026132360100746, + 0.00272729923017323, + 0.0019486809615045786, + 0.0023479547817260027, + 0.002928233239799738, + 0.0021463807206600904, + 0.0020272911060601473, + 0.0027162383776158094, + 0.0019221644615754485, + 0.008497455157339573, + 0.005146434996277094, + 0.005655321758240461, + 0.01331790816038847, + 0.002824468305334449, + 0.0023873657919466496, + 0.004724852740764618, + 0.005140385124832392, + 0.1807328164577484, + 0.033068954944610596, + 0.13674350082874298, + 0.02877868339419365, + 0.019991712644696236, + 0.018036624416708946, + 0.026676518842577934, + 0.05072220042347908, + 0.005240057595074177, + 0.006496094167232513, + 0.0087142838165164, + 0.005550877191126347, + 0.006778812035918236, + 0.004146484192460775, + 0.006528525613248348, + 0.004813586827367544, + 0.005333768203854561, + 0.00685267336666584, + 0.003592376597225666, + 0.008095228113234043, + 0.004122417885810137, + 0.003010756801813841, + 0.005351777654141188, + 0.003992033656686544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0006517431465908885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780227206647396 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780227206647396, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780227206647396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11580224335193634, + 0.05514292046427727, + 0.3865737318992615, + 0.03373608738183975, + 0.1378663331270218, + 0.03818213567137718, + 0.07294966280460358, + 0.07494962960481644, + 0.03627854958176613, + 0.04851868376135826 + ], + "credit_map": [ + 0.02322196029126644, + 0.012554566375911236, + 0.02237183228135109, + 0.015955159440636635, + 0.007596862502396107, + 0.008223133161664009, + 0.010498909279704094, + 0.015379820950329304, + 0.008007936179637909, + 0.004718977026641369, + 0.013829386793076992, + 0.008184856735169888, + 0.003991781733930111, + 0.0035039677750319242, + 0.004456072114408016, + 0.008449940010905266, + 0.028687002137303352, + 0.010105771943926811, + 0.143788143992424, + 0.03199906274676323, + 0.008273455314338207, + 0.012152262963354588, + 0.028989672660827637, + 0.12257838994264603, + 0.005695980042219162, + 0.002600725507363677, + 0.004605392459779978, + 0.00507061742246151, + 0.002987277926877141, + 0.00264586484991014, + 0.0035833243746310472, + 0.006546906195580959, + 0.024007536470890045, + 0.012141945771872997, + 0.020312096923589706, + 0.025580234825611115, + 0.009764032438397408, + 0.012158654630184174, + 0.013377661816775799, + 0.020524168387055397, + 0.005479660350829363, + 0.0027103940956294537, + 0.006175675895065069, + 0.007600570097565651, + 0.002818609355017543, + 0.002299107611179352, + 0.0045707933604717255, + 0.006527327001094818, + 0.009559143334627151, + 0.02000337280333042, + 0.01592850871384144, + 0.007915197871625423, + 0.00583227165043354, + 0.002950926311314106, + 0.006041193846613169, + 0.004719045478850603, + 0.01838834397494793, + 0.01573355309665203, + 0.009978813119232655, + 0.004458336625248194, + 0.007810469251126051, + 0.003997245337814093, + 0.005835027899593115, + 0.00874784030020237, + 0.0038649393245577812, + 0.0054840147495269775, + 0.005509188398718834, + 0.005963598378002644, + 0.004396870266646147, + 0.0024091482628136873, + 0.006095593795180321, + 0.002555198734626174, + 0.004435625858604908, + 0.006472197361290455, + 0.00989840179681778, + 0.012170514091849327, + 0.0033837731461972, + 0.002672637812793255, + 0.00640844227746129, + 0.0030770881567150354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.009780227206647396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11251761019229889, + 0.0382067933678627, + 0.11189242452383041, + 0.05615059286355972, + 0.09119974076747894, + 0.03374522924423218, + 0.034507058560848236, + 0.25845155119895935, + 0.21853069961071014, + 0.044798266142606735 + ], + "credit_map": [ + 0.011987236328423023, + 0.014620911329984665, + 0.0187569297850132, + 0.019954701885581017, + 0.010524467565119267, + 0.010457053780555725, + 0.010998886078596115, + 0.015217429958283901, + 0.005586390849202871, + 0.0035737454891204834, + 0.0054825977422297, + 0.007625012192875147, + 0.0037201468367129564, + 0.003117681248113513, + 0.0042011016048491, + 0.004900116939097643, + 0.007970521226525307, + 0.005045609083026648, + 0.03660779073834419, + 0.012285958044230938, + 0.004028612282127142, + 0.0048239147290587425, + 0.00908338651061058, + 0.03204663097858429, + 0.006899977568536997, + 0.003860171651467681, + 0.007654218934476376, + 0.008121490478515625, + 0.005618487484753132, + 0.004455558490008116, + 0.008914940059185028, + 0.010625752620398998, + 0.00868130475282669, + 0.007286073639988899, + 0.012914723716676235, + 0.015845078974962234, + 0.00670559611171484, + 0.007076949812471867, + 0.009312013164162636, + 0.023377997800707817, + 0.004629176110029221, + 0.0038502071984112263, + 0.004136479925364256, + 0.005775128491222858, + 0.003107756143435836, + 0.0026463945396244526, + 0.004417459946125746, + 0.005182624328881502, + 0.004386499524116516, + 0.005807631183415651, + 0.006109630689024925, + 0.004675410222262144, + 0.00370173086412251, + 0.003189611481502652, + 0.004259752109646797, + 0.002376791089773178, + 0.07301691174507141, + 0.0253168772906065, + 0.06908538192510605, + 0.023257574066519737, + 0.020185546949505806, + 0.011508909054100513, + 0.021046938374638557, + 0.015033427625894547, + 0.02326968125998974, + 0.03684787452220917, + 0.03396746143698692, + 0.03766850009560585, + 0.024668598547577858, + 0.010667044669389725, + 0.03776296600699425, + 0.013678562827408314, + 0.004594151861965656, + 0.00704047828912735, + 0.0059554073959589005, + 0.005905261263251305, + 0.003953929990530014, + 0.004287981893867254, + 0.00546878669410944, + 0.0075922682881355286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.018178876489400864 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a first-person perspective, the camera executes a smooth, gliding forward tracking shot through the sun-drenched garden, gradually bringing a magnificent cluster of large, lush peonies into sharp focus. The flowers are a symphony of pink, with a subtle blush on the outer petals, deepening to a vibrant magenta toward the densely layered core. Each heavy, spherical blossom is composed of countless ruffled, velvety petals, almost translucent. A gentle breeze causes the heavy flower heads to sway rhythmically and tremble gently on their sturdy green stems in a delicate and natural dance. The entire scene is bathed in the warm, inviting golden light of the afternoon, which filters through the surrounding foliage, creating a softly dappled effect and giving the petals an inner glow. The background dissolves into a beautiful bokeh of rich greens and blurred floral colors, creating an atmosphere of profound tranquility and serene beauty, capturing a fleeting, perfect moment in nature.", + "seed": 10873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_14/index_5723_seed_10873_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17360478766495363, + "vlm_logprob": -0.17360478766495363, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.17360478766495363, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.658319354057312 + ], + "opposite_logprob_tokens": [ + -0.783319354057312 + ], + "logprob_reward": -0.658319354057312, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.658319354057312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08445251733064651, + 0.05520638823509216, + 0.06210091710090637, + 0.03862251713871956, + 0.11591403186321259, + 0.05986730754375458, + 0.16490313410758972, + 0.23814858496189117, + 0.13506732881069183, + 0.04571729525923729 + ], + "credit_map": [ + 0.008811774663627148, + 0.011908045038580894, + 0.012370153330266476, + 0.014331048354506493, + 0.007078142371028662, + 0.0074269725009799, + 0.011551846750080585, + 0.010974536649882793, + 0.006301292683929205, + 0.007522040046751499, + 0.006181709468364716, + 0.0068314154632389545, + 0.005880304146558046, + 0.005233997944742441, + 0.008049657568335533, + 0.009205973707139492, + 0.007502655498683453, + 0.00912371277809143, + 0.009747306816279888, + 0.010428612120449543, + 0.006409961264580488, + 0.007312567438930273, + 0.007461125031113625, + 0.004114979412406683, + 0.004102743696421385, + 0.004554287996143103, + 0.005246935877948999, + 0.0038380196783691645, + 0.0029921524692326784, + 0.004223188851028681, + 0.003552894340828061, + 0.010112295858561993, + 0.02463243342936039, + 0.008599610067903996, + 0.010737638920545578, + 0.006785112898796797, + 0.00924405176192522, + 0.012012219987809658, + 0.0066836243495345116, + 0.0372193418443203, + 0.004150845110416412, + 0.007990346290171146, + 0.006570814177393913, + 0.012271840125322342, + 0.00865902192890644, + 0.007142110262066126, + 0.0074391551315784454, + 0.005643174052238464, + 0.013260891661047935, + 0.011302626691758633, + 0.013916908763349056, + 0.01608937978744507, + 0.025564556941390038, + 0.018697813153266907, + 0.03956425562500954, + 0.02650669403374195, + 0.03166933357715607, + 0.014097895473241806, + 0.04066463187336922, + 0.023495152592658997, + 0.019295701757073402, + 0.03108365088701248, + 0.046709463000297546, + 0.03113275021314621, + 0.011392511427402496, + 0.007311329245567322, + 0.011182107962667942, + 0.011011211201548576, + 0.014962919987738132, + 0.034849122166633606, + 0.015174679458141327, + 0.02918344922363758, + 0.012415386736392975, + 0.0037018994335085154, + 0.0037595885805785656, + 0.004085542168468237, + 0.0024849644396454096, + 0.0034581301733851433, + 0.004527370911091566, + 0.011284410953521729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.658319354057312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991454899311066 + ], + "opposite_logprob_tokens": [ + -1.8049145936965942 + ], + "logprob_reward": -0.17991454899311066, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991454899311066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06058501824736595, + 0.056426022201776505, + 0.08040472865104675, + 0.0331549234688282, + 0.06431983411312103, + 0.07527084648609161, + 0.21474415063858032, + 0.07782337069511414, + 0.3096104860305786, + 0.027660615742206573 + ], + "credit_map": [ + 0.005149708595126867, + 0.010781949386000633, + 0.0062012760899960995, + 0.01068783551454544, + 0.004839946981519461, + 0.004681580699980259, + 0.010926144197583199, + 0.007316578645259142, + 0.004247106146067381, + 0.01019609160721302, + 0.006515800952911377, + 0.007860047742724419, + 0.009417246095836163, + 0.00645138556137681, + 0.007303287740796804, + 0.004435055889189243, + 0.010890744626522064, + 0.01049115601927042, + 0.013571752235293388, + 0.014817233197391033, + 0.007525887805968523, + 0.007978834211826324, + 0.008331947028636932, + 0.0067971753887832165, + 0.003998253494501114, + 0.0036754882894456387, + 0.003926158417016268, + 0.003611493157222867, + 0.002456304617226124, + 0.004433764144778252, + 0.0030362093821167946, + 0.008017254061996937, + 0.011124052107334137, + 0.004454442299902439, + 0.005939105059951544, + 0.005739700049161911, + 0.005769598763436079, + 0.00853195134550333, + 0.004814743064343929, + 0.017946241423487663, + 0.006016061641275883, + 0.009587480686604977, + 0.010463403537869453, + 0.013869227841496468, + 0.008428164757788181, + 0.009591379202902317, + 0.01031765341758728, + 0.0069974674843251705, + 0.018850557506084442, + 0.01621178723871708, + 0.020703505724668503, + 0.023552309721708298, + 0.02491428703069687, + 0.024554697796702385, + 0.05275866016745567, + 0.03319835290312767, + 0.006273595150560141, + 0.006555401720106602, + 0.010636587627232075, + 0.012397641316056252, + 0.006221683230251074, + 0.011136523447930813, + 0.015302034094929695, + 0.009299907833337784, + 0.0267610102891922, + 0.009214142337441444, + 0.011810505762696266, + 0.012551059946417809, + 0.013943457044661045, + 0.10029648244380951, + 0.016805212944746017, + 0.11822862178087234, + 0.004099943675100803, + 0.0028317116666585207, + 0.004244078416377306, + 0.003303145058453083, + 0.0026660626754164696, + 0.002381235361099243, + 0.004933738149702549, + 0.0032006993424147367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.17991454899311066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005760917556472123 + ], + "opposite_logprob_tokens": [ + -7.500576019287109 + ], + "logprob_reward": -0.0005760917556472123, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005760917556472123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13662293553352356, + 0.0447624996304512, + 0.12945224344730377, + 0.050902388989925385, + 0.14970265328884125, + 0.06342751532793045, + 0.08193174004554749, + 0.05642799660563469, + 0.21957853436470032, + 0.06719154119491577 + ], + "credit_map": [ + 0.01590961031615734, + 0.01597040519118309, + 0.018341662362217903, + 0.02152462676167488, + 0.010468985885381699, + 0.01221390999853611, + 0.019152211025357246, + 0.02304151840507984, + 0.003941842820495367, + 0.004691889975219965, + 0.007526496425271034, + 0.005599976051598787, + 0.006215947680175304, + 0.006071811076253653, + 0.0065451753325760365, + 0.004169357009232044, + 0.0174372810870409, + 0.012546846643090248, + 0.018026214092969894, + 0.030894458293914795, + 0.017634473741054535, + 0.008761516772210598, + 0.015886155888438225, + 0.008265301585197449, + 0.004970595706254244, + 0.007507397327572107, + 0.0054406640119850636, + 0.005429854616522789, + 0.005194332450628281, + 0.006755135953426361, + 0.005602853372693062, + 0.010001550428569317, + 0.033767733722925186, + 0.010582379065454006, + 0.011145712807774544, + 0.008933781646192074, + 0.010172232054173946, + 0.015136554837226868, + 0.008763235993683338, + 0.051201026886701584, + 0.004778294358402491, + 0.007481003645807505, + 0.008834777399897575, + 0.012119853869080544, + 0.010045557282865047, + 0.00754219526425004, + 0.0062900930643081665, + 0.006335741840302944, + 0.008245840668678284, + 0.005416988395154476, + 0.007777702063322067, + 0.007734184619039297, + 0.007957029156386852, + 0.012282930314540863, + 0.017551809549331665, + 0.014965253882110119, + 0.0077279736287891865, + 0.002935256576165557, + 0.009225391782820225, + 0.005500384606420994, + 0.005617413204163313, + 0.006096981465816498, + 0.01300162635743618, + 0.006322968285530806, + 0.017629409208893776, + 0.0075567904859781265, + 0.009335300885140896, + 0.009708991274237633, + 0.009290122427046299, + 0.07046430557966232, + 0.01191832683980465, + 0.08367527276277542, + 0.019765932112932205, + 0.003776747500523925, + 0.00495797349140048, + 0.004912711679935455, + 0.003885553916916251, + 0.004432511981576681, + 0.007120017893612385, + 0.018340088427066803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.0005760917556472123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05250554904341698, + 0.04946741461753845, + 0.3561744689941406, + 0.0418013334274292, + 0.040446169674396515, + 0.06301373988389969, + 0.04324301704764366, + 0.07464786618947983, + 0.14805357158184052, + 0.13064686954021454 + ], + "credit_map": [ + 0.005346880294382572, + 0.00586958322674036, + 0.005862098652869463, + 0.007373234257102013, + 0.003536722855642438, + 0.0065887910313904285, + 0.007211906369775534, + 0.0107163330540061, + 0.003951894585043192, + 0.005756068974733353, + 0.010659495368599892, + 0.00547953462228179, + 0.004279640037566423, + 0.005147906951606274, + 0.0080017801374197, + 0.00619109021499753, + 0.05361410230398178, + 0.02473401091992855, + 0.07085134834051132, + 0.06910495460033417, + 0.04037449136376381, + 0.019005032256245613, + 0.04356883838772774, + 0.03492169454693794, + 0.004439409822225571, + 0.006010469049215317, + 0.006080469116568565, + 0.005221409257501364, + 0.004055576864629984, + 0.005252590402960777, + 0.006286655552685261, + 0.004454752895981073, + 0.0053536719642579556, + 0.00315567827783525, + 0.007343688514083624, + 0.003686030162498355, + 0.004885230679064989, + 0.006874246522784233, + 0.0035492312163114548, + 0.00559838954359293, + 0.002999846823513508, + 0.003263054881244898, + 0.004383346065878868, + 0.003979395143687725, + 0.005035574082285166, + 0.020943624898791313, + 0.006396576762199402, + 0.01601232774555683, + 0.0034027202054858208, + 0.0027555052656680346, + 0.003187317168340087, + 0.003625167766585946, + 0.006472789216786623, + 0.005510879214853048, + 0.01016373373568058, + 0.00812490563839674, + 0.006055195815861225, + 0.0027259665075689554, + 0.012248304672539234, + 0.005083329044282436, + 0.00600622920319438, + 0.01510858628898859, + 0.013834993354976177, + 0.013585256412625313, + 0.01623392291367054, + 0.0054238238371908665, + 0.006872165948152542, + 0.008659143932163715, + 0.0060194870457053185, + 0.04151330515742302, + 0.012787477113306522, + 0.050544243305921555, + 0.045798640698194504, + 0.004935083445161581, + 0.006502765230834484, + 0.005346421618014574, + 0.003434194019064307, + 0.010542870499193668, + 0.009819066151976585, + 0.04426782950758934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.008639576844871044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, gliding forward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large, lush peonies visible as the main subject with pink outer petals deepening to vibrant magenta in the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, golden afternoon light creating a softly dappled effect on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy flower heads sway rhythmically and tremble gently on their sturdy green stems due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057436667382717 + ], + "opposite_logprob_tokens": [ + -3.8955743312835693 + ], + "logprob_reward": -0.02057436667382717, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057436667382717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09710950404405594, + 0.04691057652235031, + 0.09413585066795349, + 0.05703851953148842, + 0.09330429136753082, + 0.14498844742774963, + 0.07016709446907043, + 0.08084344863891602, + 0.2482592612504959, + 0.06724300235509872 + ], + "credit_map": [ + 0.009275955148041248, + 0.012985175475478172, + 0.011783371679484844, + 0.020393462851643562, + 0.007578220684081316, + 0.009197806939482689, + 0.012994948774576187, + 0.012900566682219505, + 0.00473701860755682, + 0.0048958174884319305, + 0.007975061424076557, + 0.007254110183566809, + 0.0034868272487074137, + 0.0045362552627921104, + 0.006007353775203228, + 0.008018131367862225, + 0.014225817285478115, + 0.009541179984807968, + 0.019896844401955605, + 0.01570807583630085, + 0.008111603558063507, + 0.007573259063065052, + 0.011044641025364399, + 0.008034426718950272, + 0.004945018328726292, + 0.008408907800912857, + 0.0046101282350718975, + 0.007050725165754557, + 0.006655772216618061, + 0.007105187978595495, + 0.00961554516106844, + 0.008647236041724682, + 0.015004727989435196, + 0.010557098314166069, + 0.010510532185435295, + 0.007566132582724094, + 0.011391696520149708, + 0.008410501293838024, + 0.005998948588967323, + 0.023864658549427986, + 0.007399659138172865, + 0.0057361372746527195, + 0.00744849955663085, + 0.008199792355298996, + 0.010238627903163433, + 0.05704544112086296, + 0.008970614522695541, + 0.03994966670870781, + 0.007982520386576653, + 0.004867805168032646, + 0.008160438388586044, + 0.007868390530347824, + 0.009562534280121326, + 0.010341127403080463, + 0.009801200591027737, + 0.011583073996007442, + 0.007795318495482206, + 0.00581470038741827, + 0.01401967741549015, + 0.010604465380311012, + 0.008439424447715282, + 0.010916415601968765, + 0.014288301579654217, + 0.008965149521827698, + 0.026933586224913597, + 0.008222728036344051, + 0.013919701799750328, + 0.012792034074664116, + 0.013878118246793747, + 0.0653926208615303, + 0.01825052872300148, + 0.08886992931365967, + 0.010873342864215374, + 0.00666743703186512, + 0.0087122842669487, + 0.008939908817410469, + 0.005991138983517885, + 0.005486351903527975, + 0.011035261675715446, + 0.00953727774322033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.02057436667382717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_tensor([13, 13])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_tensor([13, 13])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8e622c5e7908cd1ea107bd7479e3977108cc0029 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_tensor([13, 13])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic cinematic panorama captures a surreal vision of Earth, suspended in the infinite, velvety black of space, sparsely dotted with distant, glittering stars. The camera remains completely static, allowing the planet to undergo a slow, deliberate, and continuous rotation on its axis. Bright, unfiltered sunlight from an off-camera source illuminates one hemisphere, casting the other in deep shadow, with a sharp terminal line slowly sliding across the surface. The ochre sands of the Sahara Desert and the deep emerald green of the Amazon River are rendered in stunning detail, gracefully gliding into view, partially obscured by a complex, swirling system of realistic white clouds that drift and evolve in the upper atmosphere. The deep blue ocean reflects dazzling specular highlights, while a thin, vibrant cyan atmospheric halo beautifully outlines the planet's curved edge against the void. As the rotation continues, the nighttime city grid begins to shimmer, creating a spectacle of wonder and awe: a living, breathing world.", + "seed": 12267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_13_seed_12267_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2534253083635122, + "vlm_logprob": -0.2534253083635122, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.2534253083635122, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008889386430382729 + ], + "opposite_logprob_tokens": [ + -5.258889198303223 + ], + "logprob_reward": -0.008889386430382729, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008889386430382729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2173008918762207, + 0.09297805279493332, + 0.02656145766377449, + 0.07231540977954865, + 0.03148510307073593, + 0.08368159085512161, + 0.08572894334793091, + 0.12730839848518372, + 0.21509774029254913, + 0.04754237085580826 + ], + "credit_map": [ + 0.036631546914577484, + 0.044743239879608154, + 0.01652456447482109, + 0.027716735377907753, + 0.02076347917318344, + 0.031887564808130264, + 0.026493389159440994, + 0.012540368363261223, + 0.016160473227500916, + 0.023882200941443443, + 0.006251696031540632, + 0.009451099671423435, + 0.011438575573265553, + 0.012757471762597561, + 0.00840124860405922, + 0.004635289777070284, + 0.0039388565346598625, + 0.004252887796610594, + 0.002181757939979434, + 0.002698748605325818, + 0.0030117405112832785, + 0.0037709439639002085, + 0.0033288050908595324, + 0.0033777181524783373, + 0.008500528521835804, + 0.006031012628227472, + 0.009654413908720016, + 0.009410930797457695, + 0.005004932172596455, + 0.009151479229331017, + 0.012556405737996101, + 0.012005710043013096, + 0.002873077290132642, + 0.002580194268375635, + 0.003590885316953063, + 0.006096642930060625, + 0.0026905604172497988, + 0.004337069112807512, + 0.00568463234230876, + 0.0036320416256785393, + 0.007712756749242544, + 0.007198099046945572, + 0.010601142421364784, + 0.012820929288864136, + 0.013742547482252121, + 0.009873955510556698, + 0.011170245707035065, + 0.010561916045844555, + 0.009168821386992931, + 0.004651912022382021, + 0.010869873687624931, + 0.006290451157838106, + 0.006348591763526201, + 0.01566167362034321, + 0.01249670423567295, + 0.020240914076566696, + 0.014332365244626999, + 0.007981661707162857, + 0.011417967267334461, + 0.0073901270516216755, + 0.0180829968303442, + 0.0386393740773201, + 0.012166180647909641, + 0.01729772239923477, + 0.02369115874171257, + 0.013980335555970669, + 0.02085379883646965, + 0.06214181333780289, + 0.015725253149867058, + 0.0542043000459671, + 0.014711144380271435, + 0.009789927862584591, + 0.004057836253196001, + 0.006502960808575153, + 0.006706717424094677, + 0.006251547951251268, + 0.007139013614505529, + 0.004866484086960554, + 0.007576739881187677, + 0.004441072233021259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.008889386430382729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296886950731277 + ], + "opposite_logprob_tokens": [ + -2.237968921661377 + ], + "logprob_reward": -0.11296886950731277, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296886950731277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11782253533601761, + 0.09953732043504715, + 0.03694875165820122, + 0.04759780690073967, + 0.06121431291103363, + 0.14759191870689392, + 0.11829418689012527, + 0.15140755474567413, + 0.13987867534160614, + 0.07970693707466125 + ], + "credit_map": [ + 0.018196240067481995, + 0.015388334169983864, + 0.011917161755263805, + 0.017021337524056435, + 0.012462303973734379, + 0.018727341666817665, + 0.015804432332515717, + 0.008305404335260391, + 0.010999980382621288, + 0.019409244880080223, + 0.007213755510747433, + 0.01321706734597683, + 0.008021202869713306, + 0.012309405952692032, + 0.01876160129904747, + 0.009605072438716888, + 0.0046723997220396996, + 0.008421835489571095, + 0.0034073798451572657, + 0.003443631576374173, + 0.0040800319984555244, + 0.004944380838423967, + 0.004457545466721058, + 0.003521550912410021, + 0.005661400966346264, + 0.005081872455775738, + 0.004412428941577673, + 0.006980340462177992, + 0.003689400153234601, + 0.006729917135089636, + 0.00782889686524868, + 0.007213552948087454, + 0.00660975044593215, + 0.0046915351413190365, + 0.004950474016368389, + 0.008871607482433319, + 0.004870598204433918, + 0.012009143829345703, + 0.012804671190679073, + 0.0064065414480865, + 0.012532497756183147, + 0.012618792243301868, + 0.021891813725233078, + 0.024068139493465424, + 0.0242754053324461, + 0.018328335136175156, + 0.021491385996341705, + 0.012385554611682892, + 0.012427162379026413, + 0.009673019871115685, + 0.017822861671447754, + 0.015545226633548737, + 0.009281696751713753, + 0.010498418472707272, + 0.019420407712459564, + 0.02362540178000927, + 0.015422160737216473, + 0.01366944145411253, + 0.022945769131183624, + 0.018452953547239304, + 0.016000010073184967, + 0.01790006458759308, + 0.023384476080536842, + 0.02363269031047821, + 0.010888570919632912, + 0.013281209394335747, + 0.019305014982819557, + 0.02820250205695629, + 0.017242247238755226, + 0.02473047375679016, + 0.013454216532409191, + 0.012774455361068249, + 0.00821725558489561, + 0.01502901315689087, + 0.008882713504135609, + 0.011032993905246258, + 0.011821825057268143, + 0.007138281594961882, + 0.01022481732070446, + 0.007360038813203573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.11296886950731277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539323091506958 + ], + "opposite_logprob_tokens": [ + -0.4289323389530182 + ], + "logprob_reward": -1.0539323091506958, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539323091506958, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08296070992946625, + 0.01996237225830555, + 0.03782980144023895, + 0.03270832449197769, + 0.10444748401641846, + 0.11824352294206619, + 0.08386973291635513, + 0.236477330327034, + 0.22242596745491028, + 0.061074767261743546 + ], + "credit_map": [ + 0.009788721799850464, + 0.010969467461109161, + 0.009224582463502884, + 0.010321653448045254, + 0.006353701930493116, + 0.012529746629297733, + 0.01605285331606865, + 0.007719985209405422, + 0.0020615807734429836, + 0.004328486509621143, + 0.0014368332922458649, + 0.0027643097564578056, + 0.0023626552429050207, + 0.0027286403346806765, + 0.0023949930910021067, + 0.0018848712788894773, + 0.004416916985064745, + 0.005691522732377052, + 0.00393842114135623, + 0.005434577818959951, + 0.0041285548359155655, + 0.004835488274693489, + 0.005625168327242136, + 0.00375915109179914, + 0.004290605429559946, + 0.003754647448658943, + 0.002589109819382429, + 0.004503223579376936, + 0.0031578335911035538, + 0.003992033656686544, + 0.004948449321091175, + 0.005472423508763313, + 0.01237514615058899, + 0.014372443780303001, + 0.014222540892660618, + 0.014897900633513927, + 0.010567075572907925, + 0.01659446209669113, + 0.013558945618569851, + 0.00785896833986044, + 0.011070159263908863, + 0.011581982485949993, + 0.014840984717011452, + 0.012760035693645477, + 0.02567042037844658, + 0.015564188361167908, + 0.014052090235054493, + 0.012703664600849152, + 0.010748835280537605, + 0.003837501397356391, + 0.01037243939936161, + 0.004741230979561806, + 0.007866241969168186, + 0.022189687937498093, + 0.009075531736016273, + 0.015038263984024525, + 0.035031143575906754, + 0.014992805197834969, + 0.01873641274869442, + 0.014565187506377697, + 0.03174132481217384, + 0.07839661091566086, + 0.015090446919202805, + 0.027923399582505226, + 0.023462707176804543, + 0.01379301492124796, + 0.023235177621245384, + 0.06931377202272415, + 0.01760103553533554, + 0.0366654172539711, + 0.018601493909955025, + 0.019753336906433105, + 0.0077064186334609985, + 0.010958795435726643, + 0.004787498619407415, + 0.009775957092642784, + 0.008446765132248402, + 0.00485025392845273, + 0.007889720611274242, + 0.006659356877207756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -1.0539323091506958 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08913029730319977 + ], + "opposite_logprob_tokens": [ + -2.464130401611328 + ], + "logprob_reward": -0.08913029730319977, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08913029730319977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05719630792737007, + 0.09285320341587067, + 0.05414063110947609, + 0.08101168274879456, + 0.10845310986042023, + 0.17194955050945282, + 0.0713246539235115, + 0.05424166098237038, + 0.20010536909103394, + 0.10872383415699005 + ], + "credit_map": [ + 0.008696841076016426, + 0.007382587064057589, + 0.004339737351983786, + 0.008157004602253437, + 0.005370599217712879, + 0.007767821196466684, + 0.00933668203651905, + 0.006145041435956955, + 0.010205082595348358, + 0.019601386040449142, + 0.006380962207913399, + 0.010843812488019466, + 0.007990935817360878, + 0.012034333311021328, + 0.017002012580633163, + 0.008794684894382954, + 0.004665955435484648, + 0.008118742145597935, + 0.004847564268857241, + 0.01043318584561348, + 0.004220946226269007, + 0.0065263365395367146, + 0.00749597093090415, + 0.007831934839487076, + 0.008998594246804714, + 0.006389009766280651, + 0.009718211367726326, + 0.009402292780578136, + 0.0059277368709445, + 0.00909784808754921, + 0.016668297350406647, + 0.014809689484536648, + 0.010179363191127777, + 0.007866584695875645, + 0.00978140253573656, + 0.018631035462021828, + 0.010333219543099403, + 0.01431390829384327, + 0.02026284858584404, + 0.017084753140807152, + 0.0147623922675848, + 0.013830522075295448, + 0.022560324519872665, + 0.018562009558081627, + 0.03158142417669296, + 0.026917744427919388, + 0.018488306552171707, + 0.025246839970350266, + 0.007000998593866825, + 0.006134169641882181, + 0.007410543505102396, + 0.00832341704517603, + 0.005639085080474615, + 0.00676528038457036, + 0.012967264279723167, + 0.0170839074999094, + 0.006455537863075733, + 0.004089280962944031, + 0.00841710064560175, + 0.008154311217367649, + 0.005597555544227362, + 0.004305980633944273, + 0.009277714416384697, + 0.00794418528676033, + 0.020902642980217934, + 0.0163860023021698, + 0.01945331320166588, + 0.06494482606649399, + 0.013476410880684853, + 0.04142516478896141, + 0.012897673062980175, + 0.010619348846375942, + 0.011745700612664223, + 0.02053063176572323, + 0.010479770600795746, + 0.013706817291676998, + 0.014945869334042072, + 0.009052672423422337, + 0.014867004938423634, + 0.013395378366112709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.08913029730319977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00220567942596972 + ], + "opposite_logprob_tokens": [ + -6.127205848693848 + ], + "logprob_reward": -0.00220567942596972, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00220567942596972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0835803747177124, + 0.06906013935804367, + 0.03348821774125099, + 0.0363401398062706, + 0.07096804678440094, + 0.07405300438404083, + 0.04545417055487633, + 0.18290576338768005, + 0.14390133321285248, + 0.2602488398551941 + ], + "credit_map": [ + 0.009954161010682583, + 0.007549412548542023, + 0.009347883984446526, + 0.014491608366370201, + 0.008282199501991272, + 0.011131740175187588, + 0.014196610078215599, + 0.008626757189631462, + 0.008274156600236893, + 0.014316784217953682, + 0.0048512243665754795, + 0.008444284088909626, + 0.006552851293236017, + 0.008939490653574467, + 0.011065909639000893, + 0.006615438032895327, + 0.0037132911384105682, + 0.004920886363834143, + 0.002624337561428547, + 0.005372030194848776, + 0.0035883202217519283, + 0.003938581328839064, + 0.005449641030281782, + 0.0038811296690255404, + 0.005180351436138153, + 0.0037580349016934633, + 0.003898082533851266, + 0.0037813768722116947, + 0.0033336584456264973, + 0.004086502827703953, + 0.006378134246915579, + 0.005924000404775143, + 0.007860131561756134, + 0.008110785856842995, + 0.009239798411726952, + 0.011045380495488644, + 0.005581601057201624, + 0.010045374743640423, + 0.01023879274725914, + 0.008846180513501167, + 0.007673242129385471, + 0.007149912882596254, + 0.008957888931035995, + 0.01076540071517229, + 0.009009788744151592, + 0.014207940548658371, + 0.008774678222835064, + 0.007514148950576782, + 0.005377982743084431, + 0.005053996574133635, + 0.007127862889319658, + 0.00702572800219059, + 0.003666359232738614, + 0.007253855001181364, + 0.005674999672919512, + 0.004273387137800455, + 0.01902616024017334, + 0.013871234841644764, + 0.018184464424848557, + 0.012700894847512245, + 0.02487976849079132, + 0.05134378373622894, + 0.019111480563879013, + 0.023787979036569595, + 0.0185952540487051, + 0.008354208432137966, + 0.02550831250846386, + 0.01701316423714161, + 0.014684509485960007, + 0.028913326561450958, + 0.014074123464524746, + 0.016758428886532784, + 0.031245851889252663, + 0.03607909381389618, + 0.026133539155125618, + 0.0404738150537014, + 0.04073603078722954, + 0.020916765555739403, + 0.03416823968291283, + 0.03049551323056221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.00220567942596972 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic cinematic panorama captures a surreal vision of Earth, suspended in the infinite, velvety black of space, sparsely dotted with distant, glittering stars. The camera remains completely static, allowing the planet to undergo a slow, deliberate, and continuous rotation on its axis. Bright, unfiltered sunlight from an off-camera source illuminates one hemisphere, casting the other in deep shadow, with a sharp terminal line slowly sliding across the surface. The ochre sands of the Sahara Desert and the deep emerald green of the Amazon River are rendered in stunning detail, gracefully gliding into view, partially obscured by a complex, swirling system of realistic white clouds that drift and evolve in the upper atmosphere. The deep blue ocean reflects dazzling specular highlights, while a thin, vibrant cyan atmospheric halo beautifully outlines the planet's curved edge against the void. As the rotation continues, the nighttime city grid begins to shimmer, creating a spectacle of wonder and awe: a living, breathing world.", + "seed": 12268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_13_seed_12268_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4302057626657188, + "vlm_logprob": -0.4302057626657188, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.4302057626657188, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0050278655253350735 + ], + "opposite_logprob_tokens": [ + -6.130027770996094 + ], + "logprob_reward": -0.0050278655253350735, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0050278655253350735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1268409937620163, + 0.05781274288892746, + 0.06728130578994751, + 0.10916748642921448, + 0.0547991544008255, + 0.017747532576322556, + 0.05711956322193146, + 0.42442384362220764, + 0.047761838883161545, + 0.037045542150735855 + ], + "credit_map": [ + 0.012551463209092617, + 0.017793018370866776, + 0.017838334664702415, + 0.022892868146300316, + 0.010066602379083633, + 0.013759524561464787, + 0.01842143014073372, + 0.013517752289772034, + 0.005867070518434048, + 0.01072889007627964, + 0.006607010494917631, + 0.010773010551929474, + 0.007223775144666433, + 0.005537421442568302, + 0.0055514234118163586, + 0.005524143110960722, + 0.004887565039098263, + 0.004662117920815945, + 0.0075301541946828365, + 0.00656347069889307, + 0.008115614764392376, + 0.018339084461331367, + 0.004965726751834154, + 0.012217570096254349, + 0.005644569173455238, + 0.004511329811066389, + 0.004053507931530476, + 0.0032509006559848785, + 0.005546753294765949, + 0.040870774537324905, + 0.004261256195604801, + 0.04102839529514313, + 0.00443927152082324, + 0.012864677235484123, + 0.004058779217302799, + 0.006442162208259106, + 0.006045909598469734, + 0.008206033147871494, + 0.005629344377666712, + 0.007112971041351557, + 0.001531252171844244, + 0.00214696628972888, + 0.002496186411008239, + 0.0022239391691982746, + 0.001990237971767783, + 0.0019495480228215456, + 0.002954091178253293, + 0.002455309731885791, + 0.004884912166744471, + 0.007764121517539024, + 0.00527332304045558, + 0.0075943791307508945, + 0.012411627918481827, + 0.005610184744000435, + 0.006791474297642708, + 0.006789545528590679, + 0.0325084887444973, + 0.09475523978471756, + 0.09274030476808548, + 0.033512018620967865, + 0.045388102531433105, + 0.03257524222135544, + 0.019461248070001602, + 0.07348320633172989, + 0.005169278476387262, + 0.006954281125217676, + 0.00617595948278904, + 0.006357494276016951, + 0.009032875299453735, + 0.004365867003798485, + 0.004288545809686184, + 0.005417536478489637, + 0.0025403830222785473, + 0.005236364901065826, + 0.003774678334593773, + 0.0036496836692094803, + 0.003907023463398218, + 0.003426966955885291, + 0.009321358054876328, + 0.0051890830509364605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0050278655253350735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062075287103652954 + ], + "opposite_logprob_tokens": [ + -2.81207537651062 + ], + "logprob_reward": -0.062075287103652954, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062075287103652954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1042231097817421, + 0.06492146849632263, + 0.255656361579895, + 0.06170697137713432, + 0.1363503336906433, + 0.09982854872941971, + 0.07651360332965851, + 0.1589142084121704, + 0.018503671512007713, + 0.023381639271974564 + ], + "credit_map": [ + 0.010419046506285667, + 0.018826764076948166, + 0.013314669951796532, + 0.022717077285051346, + 0.007549063768237829, + 0.007934192195534706, + 0.01385719794780016, + 0.00960509292781353, + 0.0057769292034208775, + 0.007980636321008205, + 0.009989326819777489, + 0.010318473912775517, + 0.006455718539655209, + 0.006974897347390652, + 0.009086708538234234, + 0.008338776417076588, + 0.019399579614400864, + 0.013234275393188, + 0.019654272124171257, + 0.016897359862923622, + 0.03504183143377304, + 0.09416262060403824, + 0.013434650376439095, + 0.04383178427815437, + 0.007260838523507118, + 0.008972877636551857, + 0.006214532069861889, + 0.007020233199000359, + 0.0048264688812196255, + 0.00918701384216547, + 0.007457387167960405, + 0.010767619125545025, + 0.015222570858895779, + 0.03130147233605385, + 0.011980139650404453, + 0.017181100323796272, + 0.011264408007264137, + 0.014538200572133064, + 0.017053147777915, + 0.017809296026825905, + 0.0076056006364524364, + 0.01436087116599083, + 0.0158357173204422, + 0.018974075093865395, + 0.009622184559702873, + 0.010747754946351051, + 0.007737046107649803, + 0.01494530588388443, + 0.007940632291138172, + 0.01133018359541893, + 0.005578073672950268, + 0.008689867332577705, + 0.016995592042803764, + 0.009348065592348576, + 0.006309906020760536, + 0.010321281850337982, + 0.011659189127385616, + 0.0321543850004673, + 0.0365254208445549, + 0.014017240144312382, + 0.016967210918664932, + 0.01260434091091156, + 0.006509049329906702, + 0.028477376326918602, + 0.0017252671532332897, + 0.0025925147347152233, + 0.003166424809023738, + 0.0027126939967274666, + 0.002461799420416355, + 0.00197139591909945, + 0.0017787775723263621, + 0.0020947963930666447, + 0.0013083966914564371, + 0.004330550320446491, + 0.002142820041626692, + 0.002808991586789489, + 0.0031371908262372017, + 0.001864301273599267, + 0.004224459175020456, + 0.003564929822459817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.062075287103652954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.0177648067474365 + ], + "opposite_logprob_tokens": [ + -0.14276480674743652 + ], + "logprob_reward": -2.0177648067474365, + "yes_no_margin_reward": -1.875, + "token_credit_reward": -2.0177648067474365, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1117246076464653, + 0.08233724534511566, + 0.09359515458345413, + 0.1166757345199585, + 0.10600830614566803, + 0.14277273416519165, + 0.05308822914958, + 0.11299816519021988, + 0.15287546813488007, + 0.027924342080950737 + ], + "credit_map": [ + 0.010878821834921837, + 0.0246772151440382, + 0.010107964277267456, + 0.02411222644150257, + 0.009525952860713005, + 0.009212590754032135, + 0.012272014282643795, + 0.010937816463410854, + 0.0075352913700044155, + 0.01123156026005745, + 0.010396408848464489, + 0.012689885683357716, + 0.012279309332370758, + 0.007917394861578941, + 0.01327996701002121, + 0.007007428910583258, + 0.006804454140365124, + 0.004683209583163261, + 0.007033800706267357, + 0.007369776722043753, + 0.011944929137825966, + 0.032309114933013916, + 0.007280419580638409, + 0.01616944558918476, + 0.0067521012388169765, + 0.009250598028302193, + 0.007594112306833267, + 0.008446967229247093, + 0.009267335757613182, + 0.03557316213846207, + 0.007053595036268234, + 0.032737862318754196, + 0.009810530580580235, + 0.024947306141257286, + 0.00759240984916687, + 0.01625078357756138, + 0.012568420730531216, + 0.00938294269144535, + 0.012616795487701893, + 0.01283911895006895, + 0.013121104799211025, + 0.01899539865553379, + 0.023811275139451027, + 0.028028441593050957, + 0.0158845204859972, + 0.013131731189787388, + 0.012372822500765324, + 0.017427435144782066, + 0.004805980250239372, + 0.008253170177340508, + 0.004658179823309183, + 0.008333193138241768, + 0.00983280036598444, + 0.00569088663905859, + 0.004373617470264435, + 0.007140402216464281, + 0.009394126944243908, + 0.018441783264279366, + 0.02467181347310543, + 0.019524483010172844, + 0.008477241732180119, + 0.008809426799416542, + 0.006904937792569399, + 0.01677434705197811, + 0.013245658949017525, + 0.0179547518491745, + 0.032248642295598984, + 0.02528628334403038, + 0.020771000534296036, + 0.01366612408310175, + 0.009607505984604359, + 0.020095504820346832, + 0.001883559045381844, + 0.005581547040492296, + 0.002289135241881013, + 0.002808179473504424, + 0.004831444937735796, + 0.00268613756634295, + 0.0037077716551721096, + 0.00413656746968627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -2.0177648067474365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05506960675120354 + ], + "opposite_logprob_tokens": [ + -2.9300696849823 + ], + "logprob_reward": -0.05506960675120354, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05506960675120354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11744730919599533, + 0.03741554543375969, + 0.18908728659152985, + 0.0778084471821785, + 0.06649764627218246, + 0.05439816415309906, + 0.02407374233007431, + 0.3734996020793915, + 0.03426568955183029, + 0.02550654485821724 + ], + "credit_map": [ + 0.01114199124276638, + 0.016297776252031326, + 0.017208609730005264, + 0.02020493522286415, + 0.009984712116420269, + 0.013264182955026627, + 0.017642121762037277, + 0.011702985502779484, + 0.0032013673335313797, + 0.005726642441004515, + 0.005200347397476435, + 0.005549828987568617, + 0.0051978398114442825, + 0.0034828975331038237, + 0.005169469863176346, + 0.003887159051373601, + 0.012282541021704674, + 0.007804933004081249, + 0.013155320659279823, + 0.010261611081659794, + 0.024376273155212402, + 0.07175222784280777, + 0.009254073724150658, + 0.040200330317020416, + 0.005811169743537903, + 0.003937275614589453, + 0.00390825979411602, + 0.0026703067123889923, + 0.0053663295693695545, + 0.02663571573793888, + 0.0033688738476485014, + 0.026110531762242317, + 0.006872080732136965, + 0.015036402270197868, + 0.004626162350177765, + 0.008333227597177029, + 0.007275543175637722, + 0.007997469045221806, + 0.007386100012809038, + 0.008970667608082294, + 0.0048179603181779385, + 0.007672186475247145, + 0.007589585613459349, + 0.00944268424063921, + 0.007377977482974529, + 0.005772738717496395, + 0.0046870955266058445, + 0.007037941366434097, + 0.0022711018100380898, + 0.0044032814912498, + 0.002176809823140502, + 0.004267263226211071, + 0.0030463687144219875, + 0.0017146477475762367, + 0.0022437574807554483, + 0.003950514830648899, + 0.026395762339234352, + 0.08527569472789764, + 0.08101711422204971, + 0.03140242025256157, + 0.037366967648267746, + 0.02805611491203308, + 0.01618005707859993, + 0.06780550628900528, + 0.0032255009282380342, + 0.004144034348428249, + 0.006061067339032888, + 0.005671550985425711, + 0.005203492008149624, + 0.002870723372325301, + 0.00257323426194489, + 0.004516088869422674, + 0.0016889836406335235, + 0.0034062801860272884, + 0.00244426098652184, + 0.0032579172402620316, + 0.0033993280958384275, + 0.0025310362689197063, + 0.005473973695188761, + 0.003304768120869994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.05506960675120354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091247200965881 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091247200965881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091247200965881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11982239037752151, + 0.0575420968234539, + 0.13884451985359192, + 0.07412797957658768, + 0.06944160163402557, + 0.08603968471288681, + 0.16763783991336823, + 0.18820486962795258, + 0.0830625593662262, + 0.0152764692902565 + ], + "credit_map": [ + 0.011491992510855198, + 0.023945314809679985, + 0.012098926119506359, + 0.025377998128533363, + 0.009671750478446484, + 0.0108800008893013, + 0.014790760353207588, + 0.011565646156668663, + 0.005447959061712027, + 0.008460217155516148, + 0.007681726012378931, + 0.009957612492144108, + 0.008464397862553596, + 0.004350189585238695, + 0.009133956395089626, + 0.0040460387244820595, + 0.00889567844569683, + 0.005234554875642061, + 0.0060445093549788, + 0.00802967045456171, + 0.019467025995254517, + 0.05615315958857536, + 0.006280212663114071, + 0.028739718720316887, + 0.007010501343756914, + 0.005397376138716936, + 0.004404077306389809, + 0.005451028235256672, + 0.005936203058809042, + 0.020423704758286476, + 0.005029367748647928, + 0.0204757172614336, + 0.005409091245383024, + 0.01501502562314272, + 0.005580620840191841, + 0.012105068191885948, + 0.008241399191319942, + 0.006708404514938593, + 0.008180242031812668, + 0.008201749995350838, + 0.007112402934581041, + 0.011313985101878643, + 0.014979523606598377, + 0.01639118418097496, + 0.010347423143684864, + 0.008070433512330055, + 0.007810960989445448, + 0.010013768449425697, + 0.014081752859055996, + 0.02961619757115841, + 0.011880403384566307, + 0.023488570004701614, + 0.035073067992925644, + 0.019614489749073982, + 0.010862651281058788, + 0.023020712658762932, + 0.014485864900052547, + 0.04207870364189148, + 0.04233623668551445, + 0.018105803057551384, + 0.015731053426861763, + 0.013394939713180065, + 0.009080677293241024, + 0.03299158066511154, + 0.007323117461055517, + 0.011437810026109219, + 0.013676749542355537, + 0.01569106988608837, + 0.011944870464503765, + 0.006447216030210257, + 0.005188611336052418, + 0.011353113688528538, + 0.0015485031763091683, + 0.002558123553171754, + 0.001628836034797132, + 0.0014990052441135049, + 0.0028835830744355917, + 0.0020072071347385645, + 0.0014418839709833264, + 0.0017093271017074585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.011091247200965881 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic cinematic panorama captures a surreal vision of Earth, suspended in the infinite, velvety black of space, sparsely dotted with distant, glittering stars. The camera remains completely static, allowing the planet to undergo a slow, deliberate, and continuous rotation on its axis. Bright, unfiltered sunlight from an off-camera source illuminates one hemisphere, casting the other in deep shadow, with a sharp terminal line slowly sliding across the surface. The ochre sands of the Sahara Desert and the deep emerald green of the Amazon River are rendered in stunning detail, gracefully gliding into view, partially obscured by a complex, swirling system of realistic white clouds that drift and evolve in the upper atmosphere. The deep blue ocean reflects dazzling specular highlights, while a thin, vibrant cyan atmospheric halo beautifully outlines the planet's curved edge against the void. As the rotation continues, the nighttime city grid begins to shimmer, creating a spectacle of wonder and awe: a living, breathing world.", + "seed": 12269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_13_seed_12269_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.25256986934691666, + "vlm_logprob": -0.25256986934691666, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.25256986934691666, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007304153870791197 + ], + "opposite_logprob_tokens": [ + -5.5073041915893555 + ], + "logprob_reward": -0.007304153870791197, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007304153870791197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15147678554058075, + 0.04839235171675682, + 0.33022090792655945, + 0.028083939105272293, + 0.16785210371017456, + 0.07073307782411575, + 0.05119488015770912, + 0.05600149929523468, + 0.07954194396734238, + 0.016502514481544495 + ], + "credit_map": [ + 0.018185768276453018, + 0.01717645674943924, + 0.022641751915216446, + 0.03304378688335419, + 0.01336623728275299, + 0.01013279054313898, + 0.022371763363480568, + 0.014558243565261364, + 0.007043882738798857, + 0.006107938941568136, + 0.0064088949002325535, + 0.008481884375214577, + 0.0035232456866651773, + 0.0046035186387598515, + 0.006825947668403387, + 0.005397041793912649, + 0.017841750755906105, + 0.018593095242977142, + 0.07515714317560196, + 0.029501939192414284, + 0.014485899358987808, + 0.013311807066202164, + 0.1305992156267166, + 0.030730068683624268, + 0.003111736848950386, + 0.0031888766679912806, + 0.004677509423345327, + 0.005820346064865589, + 0.0026765356305986643, + 0.002019508508965373, + 0.0026718496810644865, + 0.003917580470442772, + 0.017966311424970627, + 0.017200132831931114, + 0.032807912677526474, + 0.045315567404031754, + 0.012954230420291424, + 0.006899870932102203, + 0.02170773595571518, + 0.013000357896089554, + 0.008361047133803368, + 0.008794221095740795, + 0.011741052381694317, + 0.011392020620405674, + 0.006415437441319227, + 0.007809378672391176, + 0.008409031666815281, + 0.007810897193849087, + 0.004052906297147274, + 0.004960819613188505, + 0.00947606936097145, + 0.010355079546570778, + 0.004017843399196863, + 0.004655161406844854, + 0.009333357214927673, + 0.004343645181506872, + 0.006815526634454727, + 0.005576481577008963, + 0.010566993616521358, + 0.01026963535696268, + 0.005111347883939743, + 0.0039125229232013226, + 0.007616469636559486, + 0.006132525857537985, + 0.0031930445693433285, + 0.006886253599077463, + 0.018291719257831573, + 0.027774684131145477, + 0.002692743204534054, + 0.0021394770592451096, + 0.014314859174191952, + 0.004249176010489464, + 0.0014173409435898066, + 0.0018439196282997727, + 0.002901315689086914, + 0.00442906329408288, + 0.0015292633324861526, + 0.0007894205045886338, + 0.0024433601647615433, + 0.001148833311162889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.007304153870791197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869156703352928 + ], + "opposite_logprob_tokens": [ + -3.048691511154175 + ], + "logprob_reward": -0.04869156703352928, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869156703352928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1444132924079895, + 0.06383393704891205, + 0.09846843779087067, + 0.06479493528604507, + 0.1119583249092102, + 0.05984848737716675, + 0.07431656867265701, + 0.11445142328739166, + 0.20691534876823425, + 0.060999274253845215 + ], + "credit_map": [ + 0.015787046402692795, + 0.01657078228890896, + 0.018520522862672806, + 0.026216475293040276, + 0.012181590311229229, + 0.016687702387571335, + 0.018446216359734535, + 0.020002955570816994, + 0.006862045265734196, + 0.011095944792032242, + 0.008439850062131882, + 0.010482287965714931, + 0.006496471352875233, + 0.005969885271042585, + 0.007347219157963991, + 0.00714023457840085, + 0.005625490564852953, + 0.005689538549631834, + 0.020792251452803612, + 0.00950818881392479, + 0.005941573530435562, + 0.005734662991017103, + 0.034715574234724045, + 0.010461149737238884, + 0.008246336132287979, + 0.0068714008666574955, + 0.011018653400242329, + 0.01369401440024376, + 0.0038176632951945066, + 0.0046919253654778, + 0.010202745907008648, + 0.006252198480069637, + 0.014097931794822216, + 0.013665222562849522, + 0.01587757095694542, + 0.02009292133152485, + 0.010320129804313183, + 0.010585753247141838, + 0.014271145686507225, + 0.013047643005847931, + 0.005713270977139473, + 0.00543358875438571, + 0.013119665905833244, + 0.01317519135773182, + 0.005533508490771055, + 0.004204649943858385, + 0.007635569199919701, + 0.005033046007156372, + 0.007550963666290045, + 0.006173662841320038, + 0.014446656219661236, + 0.012940378859639168, + 0.006519918330013752, + 0.011423508636653423, + 0.008103106170892715, + 0.007158367894589901, + 0.01078333891928196, + 0.009939203038811684, + 0.023668957874178886, + 0.02605629712343216, + 0.008868076838552952, + 0.007153761573135853, + 0.016812484711408615, + 0.011169307865202427, + 0.01627887226641178, + 0.02199978195130825, + 0.04105747118592262, + 0.05651574209332466, + 0.012887893244624138, + 0.007827804423868656, + 0.03774354234337807, + 0.012604240328073502, + 0.004789895378053188, + 0.007381184957921505, + 0.012575704604387283, + 0.013953839428722858, + 0.0063668969087302685, + 0.0028688462916761637, + 0.008350313641130924, + 0.004712594207376242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.04869156703352928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137092113494873 + ], + "opposite_logprob_tokens": [ + -0.38709211349487305 + ], + "logprob_reward": -1.137092113494873, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137092113494873, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07370104640722275, + 0.034300658851861954, + 0.16848184168338776, + 0.04043896123766899, + 0.17333301901817322, + 0.08425705879926682, + 0.06292131543159485, + 0.12532015144824982, + 0.18916529417037964, + 0.04808058217167854 + ], + "credit_map": [ + 0.009876716881990433, + 0.009440260007977486, + 0.009721851907670498, + 0.009481280110776424, + 0.007009383291006088, + 0.01010370533913374, + 0.007830129936337471, + 0.010237712413072586, + 0.005520802456885576, + 0.004982477985322475, + 0.004781283438205719, + 0.003936638589948416, + 0.003118049819022417, + 0.0038848677650094032, + 0.004109024070203304, + 0.003967512398958206, + 0.009856883436441422, + 0.009798207320272923, + 0.03854207694530487, + 0.01519013475626707, + 0.009032772853970528, + 0.006626246031373739, + 0.06458792090415955, + 0.014847595244646072, + 0.0044023762457072735, + 0.003018643707036972, + 0.007119523826986551, + 0.00858023390173912, + 0.003426598384976387, + 0.003937710542231798, + 0.004677439574152231, + 0.005276432726532221, + 0.026664897799491882, + 0.015789488330483437, + 0.02390938252210617, + 0.02575218863785267, + 0.01605396717786789, + 0.019468732178211212, + 0.022298680618405342, + 0.023395681753754616, + 0.012331332080066204, + 0.011306459084153175, + 0.013029137626290321, + 0.011549990624189377, + 0.00603331858292222, + 0.007636304944753647, + 0.013329051434993744, + 0.00904146395623684, + 0.007108259946107864, + 0.0061244722455739975, + 0.013391121290624142, + 0.01285328809171915, + 0.004716968163847923, + 0.0056487442925572395, + 0.007658457849174738, + 0.005420007277280092, + 0.014489271678030491, + 0.018127409741282463, + 0.02177044376730919, + 0.02415549010038376, + 0.010659248568117619, + 0.008355909027159214, + 0.017039870843291283, + 0.010722513310611248, + 0.014677083119750023, + 0.01780790649354458, + 0.040001772344112396, + 0.05100661516189575, + 0.01109594851732254, + 0.006764628458768129, + 0.03691967576742172, + 0.01089166384190321, + 0.0035732369869947433, + 0.006083074025809765, + 0.008929047733545303, + 0.01148221641778946, + 0.004576925653964281, + 0.0022736366372555494, + 0.007622763980180025, + 0.003539683297276497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -1.137092113494873 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06212570145726204 + ], + "opposite_logprob_tokens": [ + -2.8121256828308105 + ], + "logprob_reward": -0.06212570145726204, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06212570145726204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0737219825387001, + 0.09370951354503632, + 0.5423963069915771, + 0.030865386128425598, + 0.06316985934972763, + 0.024777304381132126, + 0.039578814059495926, + 0.04117761179804802, + 0.058552443981170654, + 0.03205087408423424 + ], + "credit_map": [ + 0.008249985054135323, + 0.008862974122166634, + 0.01226027961820364, + 0.013756834901869297, + 0.006506452802568674, + 0.005192872602492571, + 0.011211137287318707, + 0.007681427989155054, + 0.009739617817103863, + 0.012788322754204273, + 0.012661963701248169, + 0.018195686861872673, + 0.007685461547225714, + 0.009506221860647202, + 0.012765585444867611, + 0.010366642847657204, + 0.029281750321388245, + 0.027698399499058723, + 0.12375380843877792, + 0.049900081008672714, + 0.02512701228260994, + 0.02177366055548191, + 0.2148474156856537, + 0.05001411214470863, + 0.0029355864971876144, + 0.003933468367904425, + 0.006150481291115284, + 0.006351215299218893, + 0.002032961929216981, + 0.0019193803891539574, + 0.0042688664980232716, + 0.0032734207343310118, + 0.006721909623593092, + 0.006335995625704527, + 0.01180863007903099, + 0.016487767919898033, + 0.004763637203723192, + 0.0030052345246076584, + 0.007945454679429531, + 0.006101224105805159, + 0.00483191292732954, + 0.0032713639084249735, + 0.003497494151815772, + 0.0034977176692336798, + 0.001949332538060844, + 0.00217397790402174, + 0.0033352933824062347, + 0.002220208290964365, + 0.003257286734879017, + 0.0034979151096194983, + 0.007699132896959782, + 0.008240998722612858, + 0.0032701794989407063, + 0.004651608411222696, + 0.005759523715823889, + 0.0032021619845181704, + 0.0037417085841298103, + 0.00405028136447072, + 0.0071500870399177074, + 0.009244512766599655, + 0.0030491137877106667, + 0.0034292738419026136, + 0.006011680234223604, + 0.004500948823988438, + 0.00406028563156724, + 0.00637673819437623, + 0.011281694285571575, + 0.01646261289715767, + 0.002886340022087097, + 0.0022357963025569916, + 0.010820932686328888, + 0.004428036976605654, + 0.002317853271961212, + 0.004015644080936909, + 0.00622639711946249, + 0.007893525995314121, + 0.0031648888252675533, + 0.0018722818931564689, + 0.004175263457000256, + 0.0023850149009376764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06212570145726204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635810878127813 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635810878127813, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635810878127813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06709636002779007, + 0.053067561239004135, + 0.4415851831436157, + 0.04267098009586334, + 0.20234188437461853, + 0.035766515880823135, + 0.025421224534511566, + 0.03460260480642319, + 0.070711150765419, + 0.02673652023077011 + ], + "credit_map": [ + 0.008905649185180664, + 0.006853290367871523, + 0.009235422126948833, + 0.01011901069432497, + 0.006300101522356272, + 0.007490518037229776, + 0.009231332689523697, + 0.008961034007370472, + 0.007201715372502804, + 0.007051701657474041, + 0.007375938352197409, + 0.007977820001542568, + 0.004988749511539936, + 0.005113072227686644, + 0.007887954823672771, + 0.0054706125520169735, + 0.02313503623008728, + 0.022401845082640648, + 0.10122557729482651, + 0.038463227450847626, + 0.020485538989305496, + 0.018198713660240173, + 0.17682518064975739, + 0.04085008427500725, + 0.004773099906742573, + 0.004379925783723593, + 0.00892467051744461, + 0.00673364382237196, + 0.0028524533845484257, + 0.00269392691552639, + 0.007847590371966362, + 0.004465668462216854, + 0.020764224231243134, + 0.018754232674837112, + 0.04061109945178032, + 0.05293814837932587, + 0.010307835415005684, + 0.015562954358756542, + 0.02627914398908615, + 0.01712425798177719, + 0.0046026441268622875, + 0.004874704405665398, + 0.00565507635474205, + 0.0046556913293898106, + 0.002158217132091522, + 0.003824872663244605, + 0.005340409465134144, + 0.004654894582927227, + 0.0023419230710715055, + 0.002661963691934943, + 0.005296146497130394, + 0.004960232879966497, + 0.002322302432730794, + 0.002188717946410179, + 0.0034094913862645626, + 0.0022404449991881847, + 0.003934131935238838, + 0.0029933725018054247, + 0.006714451592415571, + 0.006333023309707642, + 0.0029863647650927305, + 0.0024702169466763735, + 0.005366784520447254, + 0.003804258769378066, + 0.004707572050392628, + 0.006813788786530495, + 0.015071188099682331, + 0.021041691303253174, + 0.003158045932650566, + 0.002256251871585846, + 0.01349396537989378, + 0.0041686468757689, + 0.0026447626296430826, + 0.002967932727187872, + 0.005204124376177788, + 0.005755389109253883, + 0.002535057719796896, + 0.002100859535858035, + 0.002853518119081855, + 0.0026748753152787685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.007635810878127813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic cinematic panorama captures a surreal vision of Earth, suspended in the infinite, velvety black of space, sparsely dotted with distant, glittering stars. The camera remains completely static, allowing the planet to undergo a slow, deliberate, and continuous rotation on its axis. Bright, unfiltered sunlight from an off-camera source illuminates one hemisphere, casting the other in deep shadow, with a sharp terminal line slowly sliding across the surface. The ochre sands of the Sahara Desert and the deep emerald green of the Amazon River are rendered in stunning detail, gracefully gliding into view, partially obscured by a complex, swirling system of realistic white clouds that drift and evolve in the upper atmosphere. The deep blue ocean reflects dazzling specular highlights, while a thin, vibrant cyan atmospheric halo beautifully outlines the planet's curved edge against the void. As the rotation continues, the nighttime city grid begins to shimmer, creating a spectacle of wonder and awe: a living, breathing world.", + "seed": 12270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_15/index_13_seed_12270_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.32537248618900777, + "vlm_logprob": -0.32537248618900777, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.32537248618900777, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02309262566268444 + ], + "opposite_logprob_tokens": [ + -3.89809250831604 + ], + "logprob_reward": -0.02309262566268444, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02309262566268444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07426602393388748, + 0.05152478441596031, + 0.0616624541580677, + 0.0325346514582634, + 0.16011790931224823, + 0.05400598421692848, + 0.1409507691860199, + 0.2607905864715576, + 0.11351728439331055, + 0.05062956362962723 + ], + "credit_map": [ + 0.008883772417902946, + 0.007125361356884241, + 0.005760733503848314, + 0.01200185064226389, + 0.010479996912181377, + 0.009252083487808704, + 0.011305702850222588, + 0.009456522762775421, + 0.003489058930426836, + 0.0037227855063974857, + 0.0048597161658108234, + 0.005903374869376421, + 0.013789912685751915, + 0.005528382491320372, + 0.006040961481630802, + 0.008190595544874668, + 0.003776618279516697, + 0.003350290469825268, + 0.0038850747514516115, + 0.005523313768208027, + 0.005579296965152025, + 0.01917368546128273, + 0.005478152073919773, + 0.014896025881171227, + 0.00302480929531157, + 0.0022415886633098125, + 0.006880294531583786, + 0.004541568923741579, + 0.0031121906358748674, + 0.003395369276404381, + 0.005107252858579159, + 0.004231576807796955, + 0.012420487590134144, + 0.01467637624591589, + 0.018600143492221832, + 0.022390248253941536, + 0.028616247698664665, + 0.015938086435198784, + 0.02026442065834999, + 0.027211906388401985, + 0.0032655883114784956, + 0.004239934030920267, + 0.014643242582678795, + 0.007675550878047943, + 0.004000488203018904, + 0.007546236738562584, + 0.0055525656789541245, + 0.007082375697791576, + 0.009538399055600166, + 0.007501658052206039, + 0.02794243022799492, + 0.01311984658241272, + 0.015467630699276924, + 0.028169332072138786, + 0.019496627151966095, + 0.019714845344424248, + 0.011779813095927238, + 0.0067530483938753605, + 0.09169359505176544, + 0.05578364431858063, + 0.009308167733252048, + 0.010233837179839611, + 0.022116603329777718, + 0.05312187597155571, + 0.00468655489385128, + 0.006464342121034861, + 0.030851155519485474, + 0.04375539720058441, + 0.00366243370808661, + 0.010662195272743702, + 0.007898718118667603, + 0.005536498036235571, + 0.003908548038452864, + 0.004575187806040049, + 0.010797872208058834, + 0.008112729527056217, + 0.003950361628085375, + 0.002502567134797573, + 0.008130455389618874, + 0.008651843294501305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.02309262566268444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023327670991420746 + ], + "opposite_logprob_tokens": [ + -3.773327589035034 + ], + "logprob_reward": -0.023327670991420746, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023327670991420746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1394316852092743, + 0.038028743118047714, + 0.058820247650146484, + 0.03059500642120838, + 0.09042669832706451, + 0.0512005090713501, + 0.10402039438486099, + 0.1660364717245102, + 0.3005421459674835, + 0.02089810185134411 + ], + "credit_map": [ + 0.011513998731970787, + 0.0066259680315852165, + 0.0077027445659041405, + 0.01394817978143692, + 0.011028670705854893, + 0.02201428823173046, + 0.0372375063598156, + 0.02936033345758915, + 0.002477864967659116, + 0.002808942459523678, + 0.0025973492302000523, + 0.004856152925640345, + 0.010259767062962055, + 0.005419730208814144, + 0.004125125706195831, + 0.005483814515173435, + 0.004219680558890104, + 0.0029809961561113596, + 0.01089749950915575, + 0.005305449012666941, + 0.005281716585159302, + 0.011073198169469833, + 0.006348358932882547, + 0.0127133559435606, + 0.0028380376752465963, + 0.001993652433156967, + 0.00450114905834198, + 0.0045489720068871975, + 0.00286403251811862, + 0.0035170214250683784, + 0.005421970039606094, + 0.004910172428935766, + 0.007566031999886036, + 0.009330850094556808, + 0.0132343340665102, + 0.01205702405422926, + 0.011344612576067448, + 0.00864493940025568, + 0.011501353234052658, + 0.01674756035208702, + 0.003791664494201541, + 0.005028644111007452, + 0.008558123372495174, + 0.007392156403511763, + 0.004636532627046108, + 0.007721805479377508, + 0.00547842588275671, + 0.008593164384365082, + 0.007779211737215519, + 0.003319495590403676, + 0.03499936684966087, + 0.009090355597436428, + 0.0058434191159904, + 0.018008122220635414, + 0.015433454886078835, + 0.009546974673867226, + 0.00909667368978262, + 0.004388840403407812, + 0.057118091732263565, + 0.03140445798635483, + 0.007847857661545277, + 0.029092062264680862, + 0.01533924788236618, + 0.011749261058866978, + 0.011756541207432747, + 0.014242974109947681, + 0.08262088149785995, + 0.11891591548919678, + 0.008131805807352066, + 0.02771538309752941, + 0.024243902415037155, + 0.01291475910693407, + 0.0019644428975880146, + 0.0029271747916936874, + 0.0041121868416666985, + 0.0034582861699163914, + 0.0022777188569307327, + 0.0012832630891352892, + 0.002611835254356265, + 0.0022631948813796043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.023327670991420746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502095103263855 + ], + "opposite_logprob_tokens": [ + -0.2520950734615326 + ], + "logprob_reward": -1.502095103263855, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502095103263855, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04218798503279686, + 0.05762489140033722, + 0.06262770295143127, + 0.025752205401659012, + 0.12293338775634766, + 0.1207302063703537, + 0.23377695679664612, + 0.10567712783813477, + 0.19036297500133514, + 0.038326580077409744 + ], + "credit_map": [ + 0.004900151863694191, + 0.003368489211425185, + 0.003015662543475628, + 0.005209483206272125, + 0.004781530238687992, + 0.010363315232098103, + 0.004566449671983719, + 0.005982902832329273, + 0.00389990140683949, + 0.0046857804991304874, + 0.006555508356541395, + 0.010424052365124226, + 0.006357199978083372, + 0.00595957413315773, + 0.00862525962293148, + 0.011117614805698395, + 0.005297341383993626, + 0.005914634093642235, + 0.008470767177641392, + 0.007550753653049469, + 0.006895651109516621, + 0.012361538596451283, + 0.006551675032824278, + 0.009585336782038212, + 0.0023450336884707212, + 0.0019396463176235557, + 0.0039789751172065735, + 0.00439765490591526, + 0.0031938524916768074, + 0.002145888051018119, + 0.0042687663808465, + 0.0034823878668248653, + 0.012077318504452705, + 0.013388263992965221, + 0.016165193170309067, + 0.01586128957569599, + 0.01850711554288864, + 0.011398251168429852, + 0.014649000018835068, + 0.02088695392012596, + 0.005711941514164209, + 0.005120399408042431, + 0.04796793311834335, + 0.013123983517289162, + 0.006328491028398275, + 0.014784694649279118, + 0.012940420769155025, + 0.014752351678907871, + 0.015438668429851532, + 0.0052750492468476295, + 0.0769096240401268, + 0.019042445346713066, + 0.015590835362672806, + 0.04993331804871559, + 0.0326993465423584, + 0.018887672573328018, + 0.006785954348742962, + 0.00437273271381855, + 0.008277707733213902, + 0.014139569364488125, + 0.007012534886598587, + 0.02669462375342846, + 0.013515057973563671, + 0.02487894892692566, + 0.007752145174890757, + 0.009660433977842331, + 0.04879413917660713, + 0.07170054316520691, + 0.006750626023858786, + 0.02206842415034771, + 0.015202995389699936, + 0.008433645591139793, + 0.0033499987330287695, + 0.0027914915699511766, + 0.0085010239854455, + 0.006403956562280655, + 0.0031878433655947447, + 0.002912837080657482, + 0.004945628345012665, + 0.006233799271285534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -1.502095103263855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05504591390490532 + ], + "opposite_logprob_tokens": [ + -2.9300458431243896 + ], + "logprob_reward": -0.05504591390490532, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05504591390490532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15567053854465485, + 0.06962651759386063, + 0.26755303144454956, + 0.01689547300338745, + 0.042255863547325134, + 0.046765539795160294, + 0.1779562085866928, + 0.04179886728525162, + 0.11992859095335007, + 0.06154944747686386 + ], + "credit_map": [ + 0.011362548917531967, + 0.00993912573903799, + 0.009794825688004494, + 0.013031651265919209, + 0.01584305800497532, + 0.03603601083159447, + 0.03884051367640495, + 0.020822787657380104, + 0.006799677386879921, + 0.0062471902929246426, + 0.008889663964509964, + 0.010915758088231087, + 0.010780208744108677, + 0.006959096062928438, + 0.009182249195873737, + 0.009852665476500988, + 0.01641392521560192, + 0.01406288892030716, + 0.019649121910333633, + 0.023197170346975327, + 0.023769589141011238, + 0.08115462958812714, + 0.020621996372938156, + 0.0686836913228035, + 0.0016243314603343606, + 0.0013079483760520816, + 0.003072055522352457, + 0.0021731897722929716, + 0.0021969059016555548, + 0.001672498881816864, + 0.002632926916703582, + 0.002215614542365074, + 0.004223851021379232, + 0.0042248619720339775, + 0.00491302739828825, + 0.00491125974804163, + 0.008395982906222343, + 0.003919404931366444, + 0.005360051523894072, + 0.0063074198551476, + 0.0031137417536228895, + 0.0035426332615315914, + 0.014088745228946209, + 0.0072909691371023655, + 0.0036608048249036074, + 0.004149318672716618, + 0.005316105671226978, + 0.005603217054158449, + 0.010500980541110039, + 0.005155007820576429, + 0.05349290370941162, + 0.013618427328765392, + 0.014029034413397312, + 0.044743604958057404, + 0.021568456664681435, + 0.014847783371806145, + 0.0028263672720640898, + 0.0029928816948086023, + 0.006676503457129002, + 0.010708742775022984, + 0.003941032569855452, + 0.005548421293497086, + 0.004794808104634285, + 0.004310104064643383, + 0.00484453747048974, + 0.006161515135318041, + 0.0315220020711422, + 0.045451872050762177, + 0.0034919041208922863, + 0.013120437040925026, + 0.009672206826508045, + 0.005664100870490074, + 0.004373067058622837, + 0.007214057259261608, + 0.014733729884028435, + 0.011188195087015629, + 0.005916909780353308, + 0.003380743321031332, + 0.00825321115553379, + 0.006489529740065336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.05504591390490532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a planet Earth suspended in space with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet undergoing a slow, continuous rotation on its axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera completely static while the planet rotates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are specific geographic features like ochre Sahara sands, emerald Amazon River, and deep blue ocean visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thin, vibrant cyan atmospheric halo visible outlining the planet's curved edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330111712217331 + ], + "opposite_logprob_tokens": [ + -3.773301124572754 + ], + "logprob_reward": -0.02330111712217331, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330111712217331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22913514077663422, + 0.06808561831712723, + 0.053562991321086884, + 0.047709520906209946, + 0.04739244282245636, + 0.12059874832630157, + 0.07050064206123352, + 0.24504277110099792, + 0.08478189259767532, + 0.033190254122018814 + ], + "credit_map": [ + 0.019833708181977272, + 0.01335526816546917, + 0.017087960615754128, + 0.038897302001714706, + 0.02234809845685959, + 0.05715911090373993, + 0.029715631157159805, + 0.03073805384337902, + 0.004990922287106514, + 0.0043988958932459354, + 0.006596459075808525, + 0.008308385498821735, + 0.014672876335680485, + 0.009495866484940052, + 0.008269010111689568, + 0.011353205889463425, + 0.003901891177520156, + 0.004044339992105961, + 0.00601828470826149, + 0.005933333653956652, + 0.0069856224581599236, + 0.00743927201256156, + 0.008881691843271255, + 0.010358558036386967, + 0.005306354258209467, + 0.005459662061184645, + 0.0067527759820222855, + 0.005176820792257786, + 0.007268041837960482, + 0.0041122231632471085, + 0.006896702107042074, + 0.0067369407042860985, + 0.0038875648751854897, + 0.004143208265304565, + 0.005058895796537399, + 0.004822095390409231, + 0.007114671170711517, + 0.005086071789264679, + 0.008013227954506874, + 0.009266708046197891, + 0.006744528654962778, + 0.006951999384909868, + 0.035671133548021317, + 0.016947994008660316, + 0.00921698659658432, + 0.018537191674113274, + 0.01372295431792736, + 0.012805965729057789, + 0.004594928119331598, + 0.004149646498262882, + 0.015381856821477413, + 0.008630309253931046, + 0.005803175736218691, + 0.011942067183554173, + 0.009174969978630543, + 0.0108236875385046, + 0.01262043695896864, + 0.008165904320776463, + 0.03424600139260292, + 0.01826631836593151, + 0.019006885588169098, + 0.06298867613077164, + 0.03246331214904785, + 0.057285234332084656, + 0.009274822659790516, + 0.008532825857400894, + 0.01689733937382698, + 0.020591342821717262, + 0.006931869778782129, + 0.005663896445184946, + 0.00767585588619113, + 0.009213934652507305, + 0.0028193609323352575, + 0.0029647352639585733, + 0.00628025084733963, + 0.005185679066926241, + 0.0027113272808492184, + 0.001935320789925754, + 0.004998663906008005, + 0.006294917315244675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.02330111712217331 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_tensor([1907, 1907])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_tensor([1907, 1907])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..daea42e6bc1df75e369d4e86f561539cfcdc0a53 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_tensor([1907, 1907])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, extreme long shot, parallel to the horizon, gazes silently across a boundless, white, snow-covered wasteland. Shrouded by thick, leaden clouds, the world takes on a chilling blue hue. A single, seemingly endless, heavily armored train barrels through at thundering speed, the sole dynamic stroke on this deadened canvas. Its body is riveted from weathered, dark-gray, matte-finish steel. Its windows are as narrow as fortress embrasures, and faint red warning lights dot the hull, like dying embers in the twilight. Its massive front shovel violently pushes the snow to the sides, while its wheels race along the tracks, whipping up a fierce blizzard of ice crystals and powdery snow. This white storm follows the train's wake, a trail that gradually blurs and merges with the horizon. In the entire picture, apart from the train's own unstoppable movement, everything is silent. This extreme contrast between movement and stillness creates a grand, lonely and doomsday aesthetic.", + "seed": 6791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_1907_seed_6791_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07809861535206437, + "vlm_logprob": -0.07809861535206437, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07809861535206437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20864273607730865 + ], + "opposite_logprob_tokens": [ + -2.0836427211761475 + ], + "logprob_reward": -0.20864273607730865, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.20864273607730865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10451731830835342, + 0.047841738909482956, + 0.03729156777262688, + 0.03921947255730629, + 0.07829603552818298, + 0.0641525611281395, + 0.4174705147743225, + 0.04796530678868294, + 0.11027052998542786, + 0.05297497659921646 + ], + "credit_map": [ + 0.011814446188509464, + 0.011279192753136158, + 0.018922992050647736, + 0.010587913915514946, + 0.009024652652442455, + 0.011212676763534546, + 0.011484045535326004, + 0.020191393792629242, + 0.0049077728763222694, + 0.0035613770596683025, + 0.012444619089365005, + 0.005101203918457031, + 0.0037644582334905863, + 0.0071303281001746655, + 0.00448036752641201, + 0.006451611407101154, + 0.004064174368977547, + 0.0029437856283038855, + 0.006399976089596748, + 0.003970343619585037, + 0.002303134649991989, + 0.004685969091951847, + 0.00509320292621851, + 0.007830977439880371, + 0.004444397520273924, + 0.0032770060934126377, + 0.007614850532263517, + 0.0027708555571734905, + 0.002628745511174202, + 0.0054814014583826065, + 0.0026975038927048445, + 0.010304712690412998, + 0.009016106836497784, + 0.004009373020380735, + 0.025695083662867546, + 0.006235914304852486, + 0.0034739121329039335, + 0.015202502720057964, + 0.004734930116683245, + 0.0099282031878829, + 0.006377573125064373, + 0.0080531295388937, + 0.008585178293287754, + 0.007833176292479038, + 0.00830419734120369, + 0.007332601584494114, + 0.010129057802259922, + 0.007537638768553734, + 0.06270848959684372, + 0.023450098931789398, + 0.07162987440824509, + 0.03436572849750519, + 0.035964079201221466, + 0.058420851826667786, + 0.023794950917363167, + 0.10713642090559006, + 0.006763502489775419, + 0.004177545662969351, + 0.005841375328600407, + 0.004502164199948311, + 0.005293671041727066, + 0.0064942603930830956, + 0.004631235729902983, + 0.01026154775172472, + 0.01042967103421688, + 0.012354383245110512, + 0.013628050684928894, + 0.01503918319940567, + 0.014201669953763485, + 0.012610092759132385, + 0.013583607040345669, + 0.018423857167363167, + 0.006614629179239273, + 0.006087328307330608, + 0.006600784137845039, + 0.005346570163965225, + 0.00862360280007124, + 0.005217746831476688, + 0.009050781838595867, + 0.005433526821434498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.20864273607730865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004680273588746786 + ], + "opposite_logprob_tokens": [ + -5.379680156707764 + ], + "logprob_reward": -0.004680273588746786, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004680273588746786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07482954114675522, + 0.05506925284862518, + 0.1692153811454773, + 0.05130341649055481, + 0.17073792219161987, + 0.032977402210235596, + 0.3367118835449219, + 0.04095108062028885, + 0.01899857632815838, + 0.049205511808395386 + ], + "credit_map": [ + 0.008119389414787292, + 0.007833091542124748, + 0.014102395623922348, + 0.006042763590812683, + 0.0076322234235703945, + 0.00963627640157938, + 0.006950159091502428, + 0.014513233676552773, + 0.006614177022129297, + 0.004774684086441994, + 0.008447016589343548, + 0.006313151214271784, + 0.0038548209704458714, + 0.007997204549610615, + 0.006747798528522253, + 0.010320398025214672, + 0.01870650239288807, + 0.011611035093665123, + 0.03708953410387039, + 0.015090841799974442, + 0.01133803091943264, + 0.02232242003083229, + 0.017869098111987114, + 0.03518791124224663, + 0.006583528593182564, + 0.004902644548565149, + 0.009292489849030972, + 0.005814646370708942, + 0.00441716518253088, + 0.0070116822607815266, + 0.004928932059556246, + 0.008352326229214668, + 0.018255189061164856, + 0.01078036893159151, + 0.052433643490076065, + 0.014868472702801228, + 0.009138049557805061, + 0.029688997194170952, + 0.01316916849464178, + 0.022404035553336143, + 0.003172503551468253, + 0.0037111735437065363, + 0.005013533867895603, + 0.004087838809937239, + 0.004250156693160534, + 0.0038776695728302, + 0.004435637965798378, + 0.004428889602422714, + 0.04977702721953392, + 0.017289530485868454, + 0.07649993896484375, + 0.026268431916832924, + 0.025864984840154648, + 0.0441623330116272, + 0.019887102767825127, + 0.07696252316236496, + 0.007745649665594101, + 0.0034054238349199295, + 0.004460487049072981, + 0.0034703079145401716, + 0.003269853303208947, + 0.00591275142505765, + 0.003253253875300288, + 0.009433353319764137, + 0.0034038915764540434, + 0.0014585739700123668, + 0.0026500963140279055, + 0.0018309245351701975, + 0.0018260871293023229, + 0.0024014925584197044, + 0.0015424119774252176, + 0.003885099198669195, + 0.006651883479207754, + 0.005424316972494125, + 0.007305802311748266, + 0.004251115955412388, + 0.007089126389473677, + 0.004935030359774828, + 0.0078760189935565, + 0.005672219675034285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.004680273588746786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087591759860516 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087591759860516, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087591759860516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18845698237419128, + 0.03720661997795105, + 0.08745936304330826, + 0.05585112050175667, + 0.06128430739045143, + 0.049144092947244644, + 0.41494154930114746, + 0.05997169017791748, + 0.01634341850876808, + 0.029340896755456924 + ], + "credit_map": [ + 0.02269524335861206, + 0.015077103860676289, + 0.06289703398942947, + 0.013620459474623203, + 0.013429508544504642, + 0.02743528038263321, + 0.014887870289385319, + 0.01841447502374649, + 0.004636303521692753, + 0.003682894865050912, + 0.004949340596795082, + 0.0036089587956666946, + 0.002687430940568447, + 0.005127571523189545, + 0.0039045431185513735, + 0.008609574288129807, + 0.009818414226174355, + 0.005953080020844936, + 0.0162152461707592, + 0.00838522519916296, + 0.006153046153485775, + 0.012214389629662037, + 0.009667486883699894, + 0.01905246637761593, + 0.0076233395375311375, + 0.00408250791952014, + 0.011060542427003384, + 0.0059062642976641655, + 0.004977267235517502, + 0.007762279361486435, + 0.005195807199925184, + 0.009243103675544262, + 0.005963895469903946, + 0.00494795199483633, + 0.01645084097981453, + 0.006340321619063616, + 0.003979931585490704, + 0.009370164945721626, + 0.006345366593450308, + 0.007885831408202648, + 0.004685455933213234, + 0.0058280047960579395, + 0.006358667742460966, + 0.006308907177299261, + 0.006003753747791052, + 0.0056566824205219746, + 0.008752665482461452, + 0.005549950059503317, + 0.06557934731245041, + 0.025504296645522118, + 0.07095157355070114, + 0.03210395947098732, + 0.029871435835957527, + 0.06283700466156006, + 0.0224054753780365, + 0.1056884378194809, + 0.006221582647413015, + 0.007701965514570475, + 0.008263739757239819, + 0.006444551516324282, + 0.008369069546461105, + 0.00807519443333149, + 0.007102187257260084, + 0.007793394383043051, + 0.0017483099363744259, + 0.0023516130167990923, + 0.0022257824894040823, + 0.0017642028396949172, + 0.003053016494959593, + 0.0019616568461060524, + 0.0014170997310429811, + 0.0018217358738183975, + 0.003493351163342595, + 0.003093893639743328, + 0.0037781589198857546, + 0.005493425764143467, + 0.0036651126574724913, + 0.0031942061614245176, + 0.003105448093265295, + 0.00351729872636497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.011087591759860516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293081670999527 + ], + "opposite_logprob_tokens": [ + -3.7732930183410645 + ], + "logprob_reward": -0.023293081670999527, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293081670999527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17955134809017181, + 0.03882169350981712, + 0.04083484411239624, + 0.029841691255569458, + 0.05731099471449852, + 0.07951793074607849, + 0.4818044602870941, + 0.02658843621611595, + 0.013829737901687622, + 0.05189886689186096 + ], + "credit_map": [ + 0.019538775086402893, + 0.01181816402822733, + 0.04874282702803612, + 0.011060020886361599, + 0.009021362289786339, + 0.02849513106048107, + 0.014633946120738983, + 0.03624111786484718, + 0.004665104206651449, + 0.0030154024716466665, + 0.006180372554808855, + 0.0031269388273358345, + 0.002326244255527854, + 0.005521679297089577, + 0.0037794578820466995, + 0.0102064935490489, + 0.0047200568951666355, + 0.003202444640919566, + 0.007491228170692921, + 0.004210156854242086, + 0.0021976823918521404, + 0.005638505332171917, + 0.005106611642986536, + 0.008268160745501518, + 0.0032678875140845776, + 0.00216117175295949, + 0.005902004428207874, + 0.003706845687702298, + 0.003010914660990238, + 0.003859503660351038, + 0.0036996405106037855, + 0.0042337230406701565, + 0.0071295942179858685, + 0.004890971817076206, + 0.010931695811450481, + 0.0054136766120791435, + 0.004119339864701033, + 0.010002345778048038, + 0.00506315054371953, + 0.009760222397744656, + 0.010404768399894238, + 0.008185076527297497, + 0.01845928467810154, + 0.007432799320667982, + 0.005912178661674261, + 0.010903862304985523, + 0.007780870888382196, + 0.010439089499413967, + 0.06925361603498459, + 0.026865320280194283, + 0.13734832406044006, + 0.038055673241615295, + 0.03059958480298519, + 0.06377561390399933, + 0.030453763902187347, + 0.08545256406068802, + 0.002988271415233612, + 0.003005025675520301, + 0.003197059268131852, + 0.002721562748774886, + 0.004342340864241123, + 0.003405685070902109, + 0.0036756384652107954, + 0.00325285317376256, + 0.0021839183755218983, + 0.0011067686136811972, + 0.0018924741307273507, + 0.0014804234961047769, + 0.00172489567194134, + 0.0017811370780691504, + 0.0014700181782245636, + 0.0021901021245867014, + 0.005641045980155468, + 0.006914461497217417, + 0.006915008183568716, + 0.005775037687271833, + 0.008068540133535862, + 0.0051970891654491425, + 0.007644910831004381, + 0.005742772947996855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.023293081670999527 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278939366340637 + ], + "opposite_logprob_tokens": [ + -2.017789363861084 + ], + "logprob_reward": -0.14278939366340637, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278939366340637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27292293310165405, + 0.051920708268880844, + 0.08525000512599945, + 0.04462289437651634, + 0.05002143234014511, + 0.045050691813230515, + 0.18492305278778076, + 0.08090855181217194, + 0.05939605087041855, + 0.12498367577791214 + ], + "credit_map": [ + 0.03020799718797207, + 0.016981692984700203, + 0.09557244181632996, + 0.016832999885082245, + 0.016296733170747757, + 0.04109130799770355, + 0.02236572839319706, + 0.03357401490211487, + 0.006629853043705225, + 0.004458054434508085, + 0.007139948662370443, + 0.005373742431402206, + 0.0035148882307112217, + 0.006797926500439644, + 0.005203649401664734, + 0.012802643701434135, + 0.010118333622813225, + 0.005740367341786623, + 0.014419863931834698, + 0.008170245215296745, + 0.006228674203157425, + 0.013029657304286957, + 0.008655345998704433, + 0.018887508660554886, + 0.005355844274163246, + 0.00447132857516408, + 0.0071715074591338634, + 0.006151545327156782, + 0.004651973955333233, + 0.00557726202532649, + 0.005840466357767582, + 0.0054029677994549274, + 0.0065352036617696285, + 0.0034968876279890537, + 0.012531646527349949, + 0.003688165685161948, + 0.003047601319849491, + 0.008983785286545753, + 0.003809555433690548, + 0.007928590290248394, + 0.005455328617244959, + 0.004694586619734764, + 0.006135700736194849, + 0.006329420953989029, + 0.005879610776901245, + 0.004999635741114616, + 0.007678595837205648, + 0.0038778113666921854, + 0.025923753157258034, + 0.013104691170156002, + 0.04907358065247536, + 0.014159291982650757, + 0.009720118716359138, + 0.029298990964889526, + 0.011549046263098717, + 0.03209357336163521, + 0.012380938045680523, + 0.007969990372657776, + 0.011589225381612778, + 0.008042522706091404, + 0.009220583364367485, + 0.010681264102458954, + 0.007730004843324423, + 0.013294022530317307, + 0.004983778577297926, + 0.007322458550333977, + 0.007881342433393002, + 0.007257523480802774, + 0.009256486780941486, + 0.006553006824105978, + 0.007928203791379929, + 0.008213247172534466, + 0.01198308914899826, + 0.0170734990388155, + 0.015756957232952118, + 0.014905769377946854, + 0.01925947144627571, + 0.012914439663290977, + 0.01736263744533062, + 0.015727810561656952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.14278939366340637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, extreme long shot, parallel to the horizon, gazes silently across a boundless, white, snow-covered wasteland. Shrouded by thick, leaden clouds, the world takes on a chilling blue hue. A single, seemingly endless, heavily armored train barrels through at thundering speed, the sole dynamic stroke on this deadened canvas. Its body is riveted from weathered, dark-gray, matte-finish steel. Its windows are as narrow as fortress embrasures, and faint red warning lights dot the hull, like dying embers in the twilight. Its massive front shovel violently pushes the snow to the sides, while its wheels race along the tracks, whipping up a fierce blizzard of ice crystals and powdery snow. This white storm follows the train's wake, a trail that gradually blurs and merges with the horizon. In the entire picture, apart from the train's own unstoppable movement, everything is silent. This extreme contrast between movement and stillness creates a grand, lonely and doomsday aesthetic.", + "seed": 6792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_1907_seed_6792_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10483154915273189, + "vlm_logprob": -0.10483154915273189, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.10483154915273189, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3169228732585907 + ], + "opposite_logprob_tokens": [ + -1.691922903060913 + ], + "logprob_reward": -0.3169228732585907, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.3169228732585907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10633403062820435, + 0.043869588524103165, + 0.12875689566135406, + 0.18754708766937256, + 0.07757218182086945, + 0.06793185323476791, + 0.061770688742399216, + 0.048592276871204376, + 0.25369778275489807, + 0.023927561938762665 + ], + "credit_map": [ + 0.012261344119906425, + 0.01193260122090578, + 0.01404213160276413, + 0.01348453015089035, + 0.015987137332558632, + 0.01649606227874756, + 0.01132077258080244, + 0.010809461586177349, + 0.00473825866356492, + 0.00609746715053916, + 0.006780909840017557, + 0.004829218611121178, + 0.005198673810809851, + 0.006354822311550379, + 0.004687872249633074, + 0.005182368215173483, + 0.015400768257677555, + 0.020491356030106544, + 0.017311133444309235, + 0.012045369483530521, + 0.015217632986605167, + 0.018935389816761017, + 0.017444564029574394, + 0.011910693719983101, + 0.009027577936649323, + 0.006607723422348499, + 0.005100557114928961, + 0.006389241199940443, + 0.00483224680647254, + 0.07518935203552246, + 0.00492270989343524, + 0.07547769695520401, + 0.007581762503832579, + 0.008985048159956932, + 0.008254452608525753, + 0.007492369972169399, + 0.009128115139901638, + 0.01246198546141386, + 0.012732823379337788, + 0.010935633443295956, + 0.009212006814777851, + 0.007098604459315538, + 0.005920002236962318, + 0.008468134328722954, + 0.010261059738695621, + 0.009492778219282627, + 0.007597764953970909, + 0.009881511330604553, + 0.0072987568564713, + 0.006653790362179279, + 0.010539953596889973, + 0.00921799149364233, + 0.009231310337781906, + 0.006613090168684721, + 0.006941358093172312, + 0.005274438764899969, + 0.006759608164429665, + 0.00563177140429616, + 0.004479658789932728, + 0.005555613432079554, + 0.006751942913979292, + 0.006080944556742907, + 0.006674607750028372, + 0.006658129394054413, + 0.02775241807103157, + 0.026926692575216293, + 0.02390463277697563, + 0.026152964681386948, + 0.03863240405917168, + 0.0394466370344162, + 0.030377598479390144, + 0.04050447419285774, + 0.003049813909456134, + 0.002589915180578828, + 0.00280945491977036, + 0.0024135583080351353, + 0.004566227551549673, + 0.0023918733932077885, + 0.0033416040241718292, + 0.002765113953500986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.3169228732585907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004685612861067057 + ], + "opposite_logprob_tokens": [ + -5.379685401916504 + ], + "logprob_reward": -0.004685612861067057, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004685612861067057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11433771997690201, + 0.12355747818946838, + 0.18089330196380615, + 0.17421641945838928, + 0.06187918409705162, + 0.05415012314915657, + 0.07675585895776749, + 0.08420415967702866, + 0.07167903333902359, + 0.058326736092567444 + ], + "credit_map": [ + 0.010526560246944427, + 0.013259288854897022, + 0.015165231190621853, + 0.01440416369587183, + 0.019032880663871765, + 0.016036618500947952, + 0.014560096897184849, + 0.01135287992656231, + 0.007621342316269875, + 0.007894027046859264, + 0.006960813421756029, + 0.005423565395176411, + 0.006438547279685736, + 0.04283458739519119, + 0.00694481423124671, + 0.03943978622555733, + 0.026403840631246567, + 0.02478097379207611, + 0.023741325363516808, + 0.01687527447938919, + 0.016992703080177307, + 0.033127665519714355, + 0.019675767049193382, + 0.01929575577378273, + 0.011294753290712833, + 0.010844252072274685, + 0.00442394707351923, + 0.006940438412129879, + 0.004053570330142975, + 0.06406468898057938, + 0.005238664336502552, + 0.06735609471797943, + 0.0065219225361943245, + 0.007358126807957888, + 0.008053966797888279, + 0.006691292393952608, + 0.008346853777766228, + 0.008579110726714134, + 0.008750922046601772, + 0.0075769866816699505, + 0.004460884723812342, + 0.005642245523631573, + 0.005529134068638086, + 0.009523885324597359, + 0.007411222904920578, + 0.00715376902371645, + 0.007823007181286812, + 0.006605973467230797, + 0.011021665297448635, + 0.008320065215229988, + 0.016581038013100624, + 0.009343275800347328, + 0.009223850443959236, + 0.009429756551980972, + 0.007424213457852602, + 0.005411992780864239, + 0.014533878304064274, + 0.008814448490738869, + 0.00697212154045701, + 0.008424703031778336, + 0.010172043927013874, + 0.012156695127487183, + 0.010111968964338303, + 0.013018300756812096, + 0.006408402230590582, + 0.007471288088709116, + 0.005885480437427759, + 0.0074772764928638935, + 0.012890425510704517, + 0.01132434606552124, + 0.011182494461536407, + 0.009039324708282948, + 0.005485258065164089, + 0.007178056053817272, + 0.006304208654910326, + 0.0065942611545324326, + 0.013223589397966862, + 0.005704192910343409, + 0.007795432582497597, + 0.0060417368076741695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.004685612861067057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108916502445936 + ], + "opposite_logprob_tokens": [ + -5.504108905792236 + ], + "logprob_reward": -0.004108916502445936, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108916502445936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3943624794483185, + 0.08530960232019424, + 0.14129984378814697, + 0.05179916322231293, + 0.047286443412303925, + 0.03369579464197159, + 0.05448465794324875, + 0.05953964591026306, + 0.06548574566841125, + 0.0667366310954094 + ], + "credit_map": [ + 0.04523749649524689, + 0.04540601745247841, + 0.04526127874851227, + 0.04432704299688339, + 0.05757570639252663, + 0.06380698084831238, + 0.043976061046123505, + 0.048771873116493225, + 0.007265792693942785, + 0.010271317325532436, + 0.00936929788440466, + 0.0052225226536393166, + 0.00661786925047636, + 0.020277347415685654, + 0.007715949323028326, + 0.018569504842162132, + 0.016926679760217667, + 0.022212078794836998, + 0.01824595220386982, + 0.017378222197294235, + 0.012154570780694485, + 0.020208850502967834, + 0.02091195434331894, + 0.013261544518172741, + 0.008346552960574627, + 0.007339038420468569, + 0.0044885617680847645, + 0.004963553976267576, + 0.005876824259757996, + 0.008690908551216125, + 0.005057747475802898, + 0.007035975344479084, + 0.0060870591551065445, + 0.005330183077603579, + 0.005121434573084116, + 0.004809088073670864, + 0.004880616441369057, + 0.006567164324223995, + 0.009119854308664799, + 0.005371042527258396, + 0.004029726609587669, + 0.004367246758192778, + 0.0029841416981071234, + 0.004013401456177235, + 0.00436783442273736, + 0.004672553390264511, + 0.004284271504729986, + 0.004976620897650719, + 0.007964900694787502, + 0.006031641736626625, + 0.007291081827133894, + 0.006837928667664528, + 0.006434314418584108, + 0.007798539940267801, + 0.006349063012748957, + 0.0057771895080804825, + 0.006294384133070707, + 0.00835860799998045, + 0.0054904562421143055, + 0.006881777662783861, + 0.008594618178904057, + 0.00847573857754469, + 0.0077763888984918594, + 0.00766766956076026, + 0.006096698343753815, + 0.008144542574882507, + 0.010506114922463894, + 0.007831953465938568, + 0.0064359791576862335, + 0.008492772467434406, + 0.007815973833203316, + 0.010161709040403366, + 0.00753964064642787, + 0.008692928589880466, + 0.007936225272715092, + 0.007682666182518005, + 0.010698516853153706, + 0.006936502177268267, + 0.00921378843486309, + 0.008036368526518345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004108916502445936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097530603408813 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097530603408813, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097530603408813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03027154877781868, + 0.016596730798482895, + 0.023268355056643486, + 0.06743888556957245, + 0.05304291471838951, + 0.015047227963805199, + 0.024156613275408745, + 0.022487077862024307, + 0.7268377542495728, + 0.02085292525589466 + ], + "credit_map": [ + 0.004840141162276268, + 0.003295968985185027, + 0.0036492343060672283, + 0.003790029091760516, + 0.0042502060532569885, + 0.004111301153898239, + 0.0039164782501757145, + 0.0024181853514164686, + 0.0018769607413560152, + 0.002606899943202734, + 0.002416956005617976, + 0.0018453762168064713, + 0.0025085131637752056, + 0.0023219885770231485, + 0.0018396347295492887, + 0.0011803987435996532, + 0.002722659148275852, + 0.003384788054972887, + 0.003033455228433013, + 0.0025720770936459303, + 0.0022850672248750925, + 0.00324335228651762, + 0.003854634938761592, + 0.0021723187528550625, + 0.00370418862439692, + 0.00498685147613287, + 0.0027764816768467426, + 0.002435375237837434, + 0.0032177045941352844, + 0.023942535743117332, + 0.0032721615862101316, + 0.02310357801616192, + 0.004755015019327402, + 0.005119174253195524, + 0.006707938853651285, + 0.004873447585850954, + 0.008014514110982418, + 0.007025259546935558, + 0.009657283313572407, + 0.006890273652970791, + 0.0016055313171818852, + 0.0017556656384840608, + 0.0015848141629248857, + 0.0021153746638447046, + 0.0017846105620265007, + 0.0020960450638085604, + 0.001900189439766109, + 0.002204995369538665, + 0.004024638328701258, + 0.0023278763983398676, + 0.00403377553448081, + 0.0030204784125089645, + 0.0030386110302060843, + 0.00308758276514709, + 0.002686323132365942, + 0.001937324763275683, + 0.0033372098114341497, + 0.002396307187154889, + 0.0017426784615963697, + 0.002435750560835004, + 0.003089794423431158, + 0.0030712021980434656, + 0.0031412094831466675, + 0.003272922709584236, + 0.07322815805673599, + 0.06644612550735474, + 0.06110566854476929, + 0.0741288959980011, + 0.11452370136976242, + 0.11555890738964081, + 0.1045430526137352, + 0.11730318516492844, + 0.0018104856135323644, + 0.0023202646989375353, + 0.0028490633703768253, + 0.0027495166286826134, + 0.003775476012378931, + 0.00196097488515079, + 0.0033083385787904263, + 0.002078803488984704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.038097530603408813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034281253814697 + ], + "opposite_logprob_tokens": [ + -1.910342812538147 + ], + "logprob_reward": -0.16034281253814697, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034281253814697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18689753115177155, + 0.07265686243772507, + 0.0890926867723465, + 0.2377714067697525, + 0.020241713151335716, + 0.03710899502038956, + 0.04697040095925331, + 0.08406977355480194, + 0.12931391596794128, + 0.09587673097848892 + ], + "credit_map": [ + 0.022852271795272827, + 0.020337430760264397, + 0.02088167518377304, + 0.021597102284431458, + 0.025654369965195656, + 0.027791842818260193, + 0.02304886281490326, + 0.024733973667025566, + 0.006706241983920336, + 0.010430838912725449, + 0.005651642102748156, + 0.005053682718425989, + 0.005732723046094179, + 0.019062086939811707, + 0.0045052929781377316, + 0.015514352358877659, + 0.008410628885030746, + 0.013241985812783241, + 0.010599935427308083, + 0.013758420012891293, + 0.010728086344897747, + 0.009577005170285702, + 0.014598960988223553, + 0.008177661336958408, + 0.013793685473501682, + 0.009634888730943203, + 0.006855334155261517, + 0.0050009205006062984, + 0.005847719497978687, + 0.09376025944948196, + 0.006801164709031582, + 0.09607743471860886, + 0.00245476933196187, + 0.0023544246796518564, + 0.001766175962984562, + 0.0018804074497893453, + 0.0025391948875039816, + 0.0028215087950229645, + 0.004199661780148745, + 0.002225569449365139, + 0.004919570870697498, + 0.0029281629249453545, + 0.003317032242193818, + 0.005288202315568924, + 0.004314000718295574, + 0.00544720096513629, + 0.0048738811165094376, + 0.0060209427028894424, + 0.005825927015393972, + 0.003749515162780881, + 0.006220174953341484, + 0.006119161378592253, + 0.005234700161963701, + 0.005417793523520231, + 0.009589169174432755, + 0.0048139626160264015, + 0.013464315794408321, + 0.010422329418361187, + 0.007886289618909359, + 0.009067005477845669, + 0.010307028889656067, + 0.011136115528643131, + 0.010528425686061382, + 0.0112582603469491, + 0.013446788303554058, + 0.013356871902942657, + 0.010732297785580158, + 0.013827906921505928, + 0.019548770040273666, + 0.01974831335246563, + 0.018804503604769707, + 0.019848467782139778, + 0.0068994201719760895, + 0.011281239800155163, + 0.009776264429092407, + 0.009257503785192966, + 0.028612956404685974, + 0.00878176186233759, + 0.012004764750599861, + 0.00926282163709402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.16034281253814697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, extreme long shot, parallel to the horizon, gazes silently across a boundless, white, snow-covered wasteland. Shrouded by thick, leaden clouds, the world takes on a chilling blue hue. A single, seemingly endless, heavily armored train barrels through at thundering speed, the sole dynamic stroke on this deadened canvas. Its body is riveted from weathered, dark-gray, matte-finish steel. Its windows are as narrow as fortress embrasures, and faint red warning lights dot the hull, like dying embers in the twilight. Its massive front shovel violently pushes the snow to the sides, while its wheels race along the tracks, whipping up a fierce blizzard of ice crystals and powdery snow. This white storm follows the train's wake, a trail that gradually blurs and merges with the horizon. In the entire picture, apart from the train's own unstoppable movement, everything is silent. This extreme contrast between movement and stillness creates a grand, lonely and doomsday aesthetic.", + "seed": 6793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_1907_seed_6793_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12014019316993654, + "vlm_logprob": -0.12014019316993654, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.12014019316993654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17593957483768463 + ], + "opposite_logprob_tokens": [ + -2.4259395599365234 + ], + "logprob_reward": -0.17593957483768463, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.17593957483768463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1963280439376831, + 0.12962450087070465, + 0.11013232916593552, + 0.12385664880275726, + 0.03863970562815666, + 0.048789799213409424, + 0.05034109205007553, + 0.17646506428718567, + 0.06116091459989548, + 0.06466187536716461 + ], + "credit_map": [ + 0.02652834914624691, + 0.035261236131191254, + 0.029333116486668587, + 0.020464595407247543, + 0.01964452490210533, + 0.0296930018812418, + 0.014557036571204662, + 0.020846163854002953, + 0.013924314640462399, + 0.011509803123772144, + 0.023472923785448074, + 0.01192272175103426, + 0.017619868740439415, + 0.017393123358488083, + 0.017903484404087067, + 0.015878256410360336, + 0.009970449842512608, + 0.008968031033873558, + 0.01607862487435341, + 0.014350276440382004, + 0.013721278868615627, + 0.013960545882582664, + 0.01917828992009163, + 0.013904831372201443, + 0.04380420967936516, + 0.006922726985067129, + 0.047011684626340866, + 0.005989027675241232, + 0.0036219479516148567, + 0.006900717504322529, + 0.0043531074188649654, + 0.00525323860347271, + 0.0063729663379490376, + 0.010019482113420963, + 0.005182531196624041, + 0.003577700350433588, + 0.003758197883144021, + 0.0037222879473119974, + 0.0032051685266196728, + 0.002801374299451709, + 0.007001800462603569, + 0.008605322800576687, + 0.007142247166484594, + 0.006397097371518612, + 0.004332272335886955, + 0.004859047010540962, + 0.005670882761478424, + 0.004781128838658333, + 0.008842195384204388, + 0.004831249359995127, + 0.008992115035653114, + 0.006405807565897703, + 0.006053498946130276, + 0.0061102453619241714, + 0.00541889201849699, + 0.0036870853509753942, + 0.027562329545617104, + 0.012686951085925102, + 0.046732403337955475, + 0.016392970457673073, + 0.011385120451450348, + 0.013675570487976074, + 0.018373101949691772, + 0.02965662255883217, + 0.00978325679898262, + 0.008925622329115868, + 0.006533374078571796, + 0.00875531230121851, + 0.0045012920163571835, + 0.005465987604111433, + 0.010129407979547977, + 0.007066661957651377, + 0.007973899133503437, + 0.005836899857968092, + 0.008380274288356304, + 0.013543659821152687, + 0.004749673418700695, + 0.0057564168237149715, + 0.00908453669399023, + 0.009336510673165321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.17593957483768463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00365256960503757 + ], + "opposite_logprob_tokens": [ + -5.628652572631836 + ], + "logprob_reward": -0.00365256960503757, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00365256960503757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1278695911169052, + 0.11625269055366516, + 0.10367937386035919, + 0.07380883395671844, + 0.03691122680902481, + 0.0952727422118187, + 0.10131147503852844, + 0.10814917087554932, + 0.1333652287721634, + 0.10337965190410614 + ], + "credit_map": [ + 0.013913745060563087, + 0.01336174551397562, + 0.020633818581700325, + 0.016064198687672615, + 0.015176938846707344, + 0.017580708488821983, + 0.015084021724760532, + 0.016054419800639153, + 0.016246916726231575, + 0.018049558624625206, + 0.018839150667190552, + 0.01001760084182024, + 0.010023005306720734, + 0.014757317490875721, + 0.015788784250617027, + 0.012530356645584106, + 0.013325578533113003, + 0.00993229728192091, + 0.01893560402095318, + 0.011856009252369404, + 0.008425780571997166, + 0.01527292188256979, + 0.011095266789197922, + 0.01483591552823782, + 0.010391529649496078, + 0.00970985647290945, + 0.011025365442037582, + 0.008973507210612297, + 0.008704849518835545, + 0.008247584104537964, + 0.0076790968887507915, + 0.009077041409909725, + 0.004179697483778, + 0.007619784213602543, + 0.0041909595020115376, + 0.003828715533018112, + 0.005067198537290096, + 0.0038873266894370317, + 0.00485010864213109, + 0.003287436906248331, + 0.01395394466817379, + 0.02208811603486538, + 0.011690104380249977, + 0.01462217140942812, + 0.009286373853683472, + 0.007231462746858597, + 0.008168470114469528, + 0.00823209248483181, + 0.017750144004821777, + 0.013022135943174362, + 0.02382383681833744, + 0.012688959948718548, + 0.007815579883754253, + 0.011814288794994354, + 0.006854248698800802, + 0.0075422837398946285, + 0.016691476106643677, + 0.011334629729390144, + 0.024395689368247986, + 0.013264385983347893, + 0.01004527322947979, + 0.00898818951100111, + 0.01178246084600687, + 0.011647072620689869, + 0.01580180414021015, + 0.014882289804518223, + 0.011449310928583145, + 0.017223946750164032, + 0.0074200681410729885, + 0.010049095377326012, + 0.03684493899345398, + 0.01969376765191555, + 0.020743675529956818, + 0.013983608223497868, + 0.015039199963212013, + 0.019339630380272865, + 0.00641059223562479, + 0.007607171777635813, + 0.010141437873244286, + 0.010114340111613274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00365256960503757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650803938508034 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650803938508034, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650803938508034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15304404497146606, + 0.07190744578838348, + 0.05731233209371567, + 0.1870099902153015, + 0.1407821625471115, + 0.03426273539662361, + 0.09379010647535324, + 0.1313095986843109, + 0.08580575883388519, + 0.04477579891681671 + ], + "credit_map": [ + 0.02666899748146534, + 0.026676371693611145, + 0.01739399880170822, + 0.016695816069841385, + 0.013642583042383194, + 0.023398909717798233, + 0.010667262598872185, + 0.017900103703141212, + 0.00854841060936451, + 0.009069294668734074, + 0.008488844148814678, + 0.007436149287968874, + 0.011819005943834782, + 0.009178977459669113, + 0.010662863031029701, + 0.0067038992419838905, + 0.01020811591297388, + 0.006357132922858, + 0.011601949110627174, + 0.004333214368671179, + 0.004589323420077562, + 0.008768857456743717, + 0.003839760320261121, + 0.0076139746233820915, + 0.05380856245756149, + 0.01345804799348116, + 0.05779779329895973, + 0.012888778932392597, + 0.011531851254403591, + 0.012089130468666553, + 0.015404860489070415, + 0.010030966252088547, + 0.013281782157719135, + 0.016208862885832787, + 0.017647825181484222, + 0.01611514575779438, + 0.02234802581369877, + 0.01715916581451893, + 0.02247726544737816, + 0.015544092282652855, + 0.003600562456995249, + 0.004681815858930349, + 0.00329335848800838, + 0.004369890317320824, + 0.004364808555692434, + 0.004882265813648701, + 0.005333340261131525, + 0.00373669620603323, + 0.015035550110042095, + 0.010036037303507328, + 0.021137462928891182, + 0.010894414968788624, + 0.006867891643196344, + 0.014959254302084446, + 0.007359357085078955, + 0.007500137202441692, + 0.015290997922420502, + 0.011882706545293331, + 0.018552834168076515, + 0.01723560504615307, + 0.01662271097302437, + 0.014130856841802597, + 0.016588270664215088, + 0.021005624905228615, + 0.008725774474442005, + 0.006064843852072954, + 0.009232299402356148, + 0.010667625814676285, + 0.0051411790773272514, + 0.006566674914211035, + 0.024791056290268898, + 0.014616301283240318, + 0.0061183227226138115, + 0.004945437889546156, + 0.004325712565332651, + 0.005877731367945671, + 0.005871676839888096, + 0.006917082704603672, + 0.00599038228392601, + 0.004729451611638069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.008650803938508034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521148920059204 + ], + "opposite_logprob_tokens": [ + -1.5021148920059204 + ], + "logprob_reward": -0.2521148920059204, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521148920059204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18252822756767273, + 0.10863082855939865, + 0.14883053302764893, + 0.0911368578672409, + 0.11492088437080383, + 0.08234516531229019, + 0.12720903754234314, + 0.039039209485054016, + 0.05686423182487488, + 0.048495106399059296 + ], + "credit_map": [ + 0.03176126629114151, + 0.03459078073501587, + 0.021860864013433456, + 0.020279860123991966, + 0.018360909074544907, + 0.028138844296336174, + 0.012972637079656124, + 0.014563052915036678, + 0.014901197515428066, + 0.012651159428060055, + 0.020227715373039246, + 0.009733090177178383, + 0.012332447804510593, + 0.014071611687541008, + 0.013654384762048721, + 0.011059208773076534, + 0.020459575578570366, + 0.015298509038984776, + 0.030769435688853264, + 0.011297430843114853, + 0.011957166716456413, + 0.021298402920365334, + 0.014383764006197453, + 0.023366250097751617, + 0.011891908943653107, + 0.01034552976489067, + 0.011294113472104073, + 0.012311757542192936, + 0.011634191498160362, + 0.01013906765729189, + 0.011410432867705822, + 0.012109849601984024, + 0.012445736676454544, + 0.01739536225795746, + 0.01384990755468607, + 0.011440462432801723, + 0.01799772121012211, + 0.012905115261673927, + 0.01687612384557724, + 0.012010455131530762, + 0.010788215324282646, + 0.012941097840666771, + 0.010666359215974808, + 0.011800000444054604, + 0.008858553133904934, + 0.01094026304781437, + 0.009733784012496471, + 0.006616884842514992, + 0.02161644771695137, + 0.014853978529572487, + 0.03560663014650345, + 0.01709684170782566, + 0.00738875987008214, + 0.011153214611113071, + 0.009615341201424599, + 0.009877800941467285, + 0.0058634537272155285, + 0.0033456841483712196, + 0.007163350936025381, + 0.004288536962121725, + 0.004811401944607496, + 0.0031438402365893126, + 0.005500949919223785, + 0.004921987187117338, + 0.007576949428766966, + 0.0056775351986289024, + 0.004954018630087376, + 0.009911651723086834, + 0.006746527273207903, + 0.006499988958239555, + 0.008708879351615906, + 0.006788678467273712, + 0.009230579249560833, + 0.006533913314342499, + 0.006500753574073315, + 0.008806254714727402, + 0.003507504239678383, + 0.00491699343547225, + 0.004284269642084837, + 0.004714833106845617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.2521148920059204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034312546253204 + ], + "opposite_logprob_tokens": [ + -1.9103431701660156 + ], + "logprob_reward": -0.16034312546253204, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034312546253204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1894107162952423, + 0.09778188914060593, + 0.06620389968156815, + 0.08479608595371246, + 0.04657050967216492, + 0.12540380656719208, + 0.068898506462574, + 0.08841847628355026, + 0.12261085957288742, + 0.10990525037050247 + ], + "credit_map": [ + 0.026059089228510857, + 0.028973961248993874, + 0.018947729840874672, + 0.024234844371676445, + 0.020068777725100517, + 0.025390909984707832, + 0.028614530339837074, + 0.017120884731411934, + 0.012199965305626392, + 0.01021448615938425, + 0.011675841175019741, + 0.010663197375833988, + 0.017065977677702904, + 0.010999009013175964, + 0.016035668551921844, + 0.008927746675908566, + 0.009600818157196045, + 0.008057103492319584, + 0.010782540775835514, + 0.006027277559041977, + 0.007635758724063635, + 0.007584377191960812, + 0.008234890177845955, + 0.008281135000288486, + 0.0191592238843441, + 0.010261539369821548, + 0.01912332884967327, + 0.00706194993108511, + 0.006797094363719225, + 0.0060722799971699715, + 0.009490042924880981, + 0.006830619182437658, + 0.005933769978582859, + 0.010544564574956894, + 0.0054382761009037495, + 0.0056077828630805016, + 0.005704025737941265, + 0.0040944283828139305, + 0.00583833921700716, + 0.0034093207214027643, + 0.016338948160409927, + 0.019890759140253067, + 0.014795897528529167, + 0.020494332537055016, + 0.01518343761563301, + 0.012794338166713715, + 0.015330540016293526, + 0.010575558990240097, + 0.010968676768243313, + 0.00823821872472763, + 0.017745086923241615, + 0.00782571267336607, + 0.005788362119346857, + 0.007311718538403511, + 0.0060690962709486485, + 0.004951626993715763, + 0.009078027680516243, + 0.0070571391843259335, + 0.010081082582473755, + 0.013584750704467297, + 0.011230873875319958, + 0.009749483317136765, + 0.012991636991500854, + 0.014645486138761044, + 0.009869378991425037, + 0.007196289021521807, + 0.009219910018146038, + 0.014911447651684284, + 0.007682919967919588, + 0.011997750960290432, + 0.035975560545921326, + 0.025757605209946632, + 0.013573046773672104, + 0.010769454762339592, + 0.01606840081512928, + 0.021030927076935768, + 0.010520415380597115, + 0.009332437999546528, + 0.013409579172730446, + 0.015200987458229065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.16034312546253204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, extreme long shot, parallel to the horizon, gazes silently across a boundless, white, snow-covered wasteland. Shrouded by thick, leaden clouds, the world takes on a chilling blue hue. A single, seemingly endless, heavily armored train barrels through at thundering speed, the sole dynamic stroke on this deadened canvas. Its body is riveted from weathered, dark-gray, matte-finish steel. Its windows are as narrow as fortress embrasures, and faint red warning lights dot the hull, like dying embers in the twilight. Its massive front shovel violently pushes the snow to the sides, while its wheels race along the tracks, whipping up a fierce blizzard of ice crystals and powdery snow. This white storm follows the train's wake, a trail that gradually blurs and merges with the horizon. In the entire picture, apart from the train's own unstoppable movement, everything is silent. This extreme contrast between movement and stillness creates a grand, lonely and doomsday aesthetic.", + "seed": 6794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_2/index_1907_seed_6794_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07751059825532139, + "vlm_logprob": -0.07751059825532139, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.07751059825532139, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14909318089485168 + ], + "opposite_logprob_tokens": [ + -2.5240931510925293 + ], + "logprob_reward": -0.14909318089485168, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.14909318089485168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13377927243709564, + 0.07314573228359222, + 0.05942637473344803, + 0.34596142172813416, + 0.03838922083377838, + 0.07289434224367142, + 0.11756397038698196, + 0.04565127566456795, + 0.035340242087841034, + 0.07784818112850189 + ], + "credit_map": [ + 0.023496119305491447, + 0.013427289202809334, + 0.017975537106394768, + 0.020168013870716095, + 0.011377239599823952, + 0.01720709353685379, + 0.016765888780355453, + 0.013362093828618526, + 0.009474515914916992, + 0.006464867852628231, + 0.013245931826531887, + 0.01002551056444645, + 0.009044806472957134, + 0.009389975108206272, + 0.008196486160159111, + 0.0073036435060203075, + 0.00937680248171091, + 0.009068273939192295, + 0.008839740417897701, + 0.008255254477262497, + 0.006486544385552406, + 0.00543462997302413, + 0.006512023042887449, + 0.005453105084598064, + 0.12432433664798737, + 0.017019104212522507, + 0.14986306428909302, + 0.018387384712696075, + 0.010424909181892872, + 0.008493411354720592, + 0.008031935431063175, + 0.009417257271707058, + 0.004572667181491852, + 0.005785324610769749, + 0.006449851207435131, + 0.004353802651166916, + 0.0038987542502582073, + 0.008026727475225925, + 0.0022647972218692303, + 0.003037298796698451, + 0.012731439433991909, + 0.007734033744782209, + 0.01585494354367256, + 0.0077204955741763115, + 0.007226312067359686, + 0.008174281567335129, + 0.005696381442248821, + 0.007756453473120928, + 0.020117124542593956, + 0.01346516888588667, + 0.026360396295785904, + 0.016411280259490013, + 0.009104600176215172, + 0.010288427583873272, + 0.00980337057262659, + 0.012013604864478111, + 0.00844111479818821, + 0.00660816440358758, + 0.008007551543414593, + 0.0055983783677220345, + 0.004262112081050873, + 0.00434355903416872, + 0.004781280178576708, + 0.00360911525785923, + 0.004019126296043396, + 0.004340287763625383, + 0.00521722761914134, + 0.006965252570807934, + 0.003710217773914337, + 0.004064974375069141, + 0.0031568859703838825, + 0.003866268554702401, + 0.013975074514746666, + 0.011429116129875183, + 0.014049292542040348, + 0.006713871844112873, + 0.011611289344727993, + 0.008783550001680851, + 0.005393933039158583, + 0.005892056040465832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.14909318089485168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005281898658722639 + ], + "opposite_logprob_tokens": [ + -5.255281925201416 + ], + "logprob_reward": -0.005281898658722639, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005281898658722639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07400403171777725, + 0.08021464943885803, + 0.020249441266059875, + 0.7100802659988403, + 0.031553007662296295, + 0.011490881443023682, + 0.020895572379231453, + 0.02247558906674385, + 0.012968173250555992, + 0.0160683561116457 + ], + "credit_map": [ + 0.01259464118629694, + 0.007086918223649263, + 0.008171280845999718, + 0.009076641872525215, + 0.006350204348564148, + 0.009401260875165462, + 0.007800844497978687, + 0.013522236607968807, + 0.012422780506312847, + 0.006738091818988323, + 0.018257014453411102, + 0.009578010998666286, + 0.006426667794585228, + 0.010093867778778076, + 0.008376232348382473, + 0.008321985602378845, + 0.0035000089555978775, + 0.0025356144178658724, + 0.004386188928037882, + 0.0025747492909431458, + 0.001990503864362836, + 0.0018579033203423023, + 0.0014966977760195732, + 0.0019077742472290993, + 0.258543461561203, + 0.03012024238705635, + 0.31693047285079956, + 0.03769329562783241, + 0.01647367514669895, + 0.016279855743050575, + 0.016248375177383423, + 0.01779083162546158, + 0.0047797407023608685, + 0.005782678723335266, + 0.00569140212610364, + 0.003413862781599164, + 0.0035531886387616396, + 0.0030436788219958544, + 0.002103331033140421, + 0.0031851250678300858, + 0.0016718761762604117, + 0.0011906776344403625, + 0.0015287610003724694, + 0.0013949045678600669, + 0.0011703362688422203, + 0.0016542328521609306, + 0.0011465802090242505, + 0.0017335116863250732, + 0.0030508951749652624, + 0.0017452772008255124, + 0.004231824539601803, + 0.004819120280444622, + 0.0016910125268623233, + 0.0020158193074166775, + 0.0017994302324950695, + 0.0015421914868056774, + 0.004159624222666025, + 0.004680393263697624, + 0.00380318914540112, + 0.002816232852637768, + 0.0019781463779509068, + 0.0015452260849997401, + 0.002230682410299778, + 0.0012620947090908885, + 0.0016495819436386228, + 0.001957249827682972, + 0.0015598232857882977, + 0.0018166209338232875, + 0.0021372076589614153, + 0.0017926058499142528, + 0.001107470248825848, + 0.0009476125123910606, + 0.0037341059651225805, + 0.0015503889881074429, + 0.003975642379373312, + 0.0016436252044513822, + 0.0012262850068509579, + 0.0014373344602063298, + 0.0011141039431095123, + 0.0013868700480088592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.005281898658722639 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002826031530275941 + ], + "opposite_logprob_tokens": [ + -5.87782621383667 + ], + "logprob_reward": -0.002826031530275941, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002826031530275941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1299249231815338, + 0.05817379057407379, + 0.1306711733341217, + 0.15369080007076263, + 0.1264929473400116, + 0.057315487414598465, + 0.08381325751543045, + 0.08383721113204956, + 0.07340572029352188, + 0.10267461836338043 + ], + "credit_map": [ + 0.012024183757603168, + 0.012731008231639862, + 0.014136508107185364, + 0.023470189422369003, + 0.01725577376782894, + 0.018325962126255035, + 0.015463846735656261, + 0.016517464071512222, + 0.008630006574094296, + 0.0042916759848594666, + 0.00819492619484663, + 0.006963835097849369, + 0.0059061357751488686, + 0.007151629310101271, + 0.006984542589634657, + 0.010051037184894085, + 0.02400270104408264, + 0.029621385037899017, + 0.015689002349972725, + 0.014107909053564072, + 0.014873097650706768, + 0.01372822280973196, + 0.010306821204721928, + 0.008342042565345764, + 0.03770717978477478, + 0.010272778570652008, + 0.04289169982075691, + 0.01544537115842104, + 0.012293510138988495, + 0.01959061436355114, + 0.007327577099204063, + 0.00816208217293024, + 0.019547438248991966, + 0.014991081319749355, + 0.0265981275588274, + 0.014152525924146175, + 0.014012344181537628, + 0.015152736566960812, + 0.012577212415635586, + 0.009461493231356144, + 0.008317328058183193, + 0.009031802415847778, + 0.011890470050275326, + 0.006511195562779903, + 0.005523010157048702, + 0.00515992846339941, + 0.00497142318636179, + 0.005910330917686224, + 0.012489435262978077, + 0.009740257635712624, + 0.013156336732208729, + 0.012385780923068523, + 0.008884771727025509, + 0.008970869705080986, + 0.008226471021771431, + 0.00995933823287487, + 0.011533565819263458, + 0.012988429516553879, + 0.009290165267884731, + 0.011646048165857792, + 0.010312525555491447, + 0.009523809887468815, + 0.010796226561069489, + 0.0077464403584599495, + 0.010279380716383457, + 0.009437848813831806, + 0.008864388801157475, + 0.009322487749159336, + 0.009435036219656467, + 0.01205614022910595, + 0.0068177348002791405, + 0.007192700169980526, + 0.016234789043664932, + 0.013075657188892365, + 0.014622824266552925, + 0.009556086733937263, + 0.016693979501724243, + 0.013590076938271523, + 0.007977301254868507, + 0.010923903435468674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.002826031530275941 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001759111881256 + ], + "opposite_logprob_tokens": [ + -2.6950175762176514 + ], + "logprob_reward": -0.07001759111881256, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001759111881256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11313716322183609, + 0.08715416491031647, + 0.04998425766825676, + 0.11785873025655746, + 0.12241437286138535, + 0.0936959981918335, + 0.04978233575820923, + 0.03754667192697525, + 0.03079691156744957, + 0.2976294159889221 + ], + "credit_map": [ + 0.014552456326782703, + 0.011137028224766254, + 0.012095073238015175, + 0.017377017065882683, + 0.013407452031970024, + 0.015386405400931835, + 0.014620394445955753, + 0.014561343006789684, + 0.01004793494939804, + 0.006623638328164816, + 0.01589805632829666, + 0.014438510872423649, + 0.011728838086128235, + 0.009712203405797482, + 0.010776100680232048, + 0.0079288799315691, + 0.008891250938177109, + 0.004802627023309469, + 0.008327215909957886, + 0.005002941936254501, + 0.008788825944066048, + 0.006721928250044584, + 0.0034597362391650677, + 0.003989731892943382, + 0.0376022607088089, + 0.00591396214440465, + 0.045748550444841385, + 0.007607220206409693, + 0.007431309670209885, + 0.0035706195048987865, + 0.004417434334754944, + 0.005567376036196947, + 0.022446490824222565, + 0.019724180921912193, + 0.032603468745946884, + 0.013304653577506542, + 0.010488354600965977, + 0.008244608528912067, + 0.007877374067902565, + 0.007725234143435955, + 0.018579311668872833, + 0.012837626039981842, + 0.028186121955513954, + 0.01039834599941969, + 0.005378568544983864, + 0.005849634762853384, + 0.0069722565822303295, + 0.005494129378348589, + 0.008020102046430111, + 0.005458706058561802, + 0.012813241221010685, + 0.008722824044525623, + 0.0030485778115689754, + 0.004631682299077511, + 0.004036678001284599, + 0.0030505252070724964, + 0.005571021232753992, + 0.005234916694462299, + 0.005291460081934929, + 0.004821969196200371, + 0.0038788188248872757, + 0.00442487932741642, + 0.004094255156815052, + 0.004229350481182337, + 0.0054243868216872215, + 0.00353979947976768, + 0.005667915102094412, + 0.003738192142918706, + 0.004203952848911285, + 0.003198752412572503, + 0.002418453572317958, + 0.002605457790195942, + 0.07229749113321304, + 0.02884582243859768, + 0.08268573880195618, + 0.022875700145959854, + 0.026040347293019295, + 0.024334995076060295, + 0.0185860525816679, + 0.02196325734257698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.07001759111881256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme long shot parallel to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless, white, snow-covered wasteland under thick, leaden clouds with a chilling blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily armored train with weathered, dark-gray, matte-finish steel and narrow fortress-like windows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train move at high speed with a massive front shovel pushing snow and wheels whipping up a blizzard of ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are faint red warning lights visible dotting the train's hull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603342890739441 + ], + "opposite_logprob_tokens": [ + -1.9103343486785889 + ], + "logprob_reward": -0.1603342890739441, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603342890739441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07218951731920242, + 0.11088992655277252, + 0.07930724322795868, + 0.3658643066883087, + 0.07107029855251312, + 0.052375439554452896, + 0.05842410773038864, + 0.0357135571539402, + 0.06548748165369034, + 0.08867812156677246 + ], + "credit_map": [ + 0.010755096562206745, + 0.0057703424245119095, + 0.008005167357623577, + 0.009242638014256954, + 0.011939792893826962, + 0.009050056338310242, + 0.010182112455368042, + 0.007244305219501257, + 0.015464778058230877, + 0.008393846452236176, + 0.02270258404314518, + 0.016072891652584076, + 0.009947988204658031, + 0.013205369003117085, + 0.014444779604673386, + 0.010657679289579391, + 0.011941214092075825, + 0.016283150762319565, + 0.006987544242292643, + 0.008484588004648685, + 0.009712384082376957, + 0.01561329048126936, + 0.005109324119985104, + 0.005175739526748657, + 0.10845905542373657, + 0.01877615787088871, + 0.1325041502714157, + 0.03120872937142849, + 0.025491798296570778, + 0.014644443988800049, + 0.016743090003728867, + 0.018036868423223495, + 0.009880888275802135, + 0.007875633426010609, + 0.007611893117427826, + 0.008864405564963818, + 0.008483320474624634, + 0.012866311706602573, + 0.00878625176846981, + 0.006701588165014982, + 0.008736035786569118, + 0.008548184297978878, + 0.014241144992411137, + 0.005813290365040302, + 0.004909456241875887, + 0.004057744517922401, + 0.0032998002134263515, + 0.002769777085632086, + 0.007224337197840214, + 0.008698451332747936, + 0.009412439540028572, + 0.012898407876491547, + 0.004425328690558672, + 0.005771256983280182, + 0.004925242159515619, + 0.0050686378963291645, + 0.006080518011003733, + 0.005637426860630512, + 0.005250610411167145, + 0.004254188388586044, + 0.004349404014647007, + 0.00394307030364871, + 0.003342979121953249, + 0.002855355618521571, + 0.00825791247189045, + 0.008135437965393066, + 0.008018145337700844, + 0.007851710543036461, + 0.006076999474316835, + 0.009477143175899982, + 0.006826432887464762, + 0.010843686759471893, + 0.016316469758749008, + 0.011542466469109058, + 0.017354877665638924, + 0.008667814545333385, + 0.010076330043375492, + 0.00944606028497219, + 0.006272599101066589, + 0.009001498110592365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.1603342890739441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_tensor([9760, 9760])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_tensor([9760, 9760])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a4b4bbcbbe2163c18e56b3e07ac88cbb14bc31bb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_tensor([9760, 9760])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Low-angle tracking shots, stylized in a style reminiscent of high-speed sports documentaries, glide smoothly from left to right, just above the ground, immersing the viewer in the sun-drenched fields of Monarda di Dima. The camera captures their vibrant red, spiky flower heads, each a chaotic crown of tubular petals swaying animatedly atop tall, slender green stems. In the brilliant, high-contrast light of the midday sun, every detail is rendered with hyper-real clarity; petals shimmer, while sharp, dark shadows define their complex forms. A persistent breeze sweeps across the field, inspiring a chaotic yet synchronized dance, with the flowers flowing across the frame in powerful, rhythmic waves, dipping, swaying, and bouncing. The camera's fluid, purposeful movements, combined with a high-definition aesthetic, fuse this natural spectacle with the dynamic energy and concentrated intensity of a major sporting event, creating a vibrant and exciting atmosphere.", + "seed": 12811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_9760_seed_12811_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1505558688659221, + "vlm_logprob": -0.1505558688659221, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1505558688659221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05413895845413208 + ], + "opposite_logprob_tokens": [ + -3.9291388988494873 + ], + "logprob_reward": -0.05413895845413208, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05413895845413208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025760900229215622, + 0.67552649974823, + 0.03997575119137764, + 0.022468935698270798, + 0.013662303797900677, + 0.005670615006238222, + 0.010687015019357204, + 0.19212140142917633, + 0.009150690399110317, + 0.004975873976945877 + ], + "credit_map": [ + 0.002359668491408229, + 0.004402243532240391, + 0.0043595763854682446, + 0.002917325356975198, + 0.0024583879858255386, + 0.0033942991867661476, + 0.002564546652138233, + 0.003304853104054928, + 0.03418935462832451, + 0.27067023515701294, + 0.018932342529296875, + 0.022243421524763107, + 0.03202012926340103, + 0.25759923458099365, + 0.014382819645106792, + 0.02548900619149208, + 0.006623001303523779, + 0.0034011169336736202, + 0.004099240060895681, + 0.003079686313867569, + 0.011773651465773582, + 0.0053300755098462105, + 0.003514984855428338, + 0.0021539931185543537, + 0.0016969323623925447, + 0.00408858573064208, + 0.0012564295902848244, + 0.005188720300793648, + 0.00512797012925148, + 0.0026331120170652866, + 0.0014109746553003788, + 0.0010662110289558768, + 0.0009901141747832298, + 0.0037860169541090727, + 0.000641866005025804, + 0.004714494105428457, + 0.0009040076402015984, + 0.0005812130984850228, + 0.001111190184019506, + 0.0009334018686786294, + 0.0005354881286621094, + 0.0011162045411765575, + 0.0006255346816033125, + 0.0012572156265377998, + 0.0006378886755555868, + 0.00041917586349882185, + 0.0005480043473653495, + 0.000531103229150176, + 0.0009621226927265525, + 0.0016253986395895481, + 0.0012942301109433174, + 0.0015768242301419377, + 0.0010731532238423824, + 0.0015990709653124213, + 0.0012386261951178312, + 0.00131758872885257, + 0.011308558285236359, + 0.0719475969672203, + 0.0046720923855900764, + 0.0061081754975020885, + 0.07500895857810974, + 0.014433137141168118, + 0.004808587953448296, + 0.0038342850748449564, + 0.0013641403056681156, + 0.0012475481489673257, + 0.0010399766033515334, + 0.0009079331648536026, + 0.001617278903722763, + 0.0010748935164883733, + 0.000759099202696234, + 0.0011398199712857604, + 0.0003516323340591043, + 0.0006958413287065923, + 0.0006743510020896792, + 0.0007128959987312555, + 0.0007995772757567465, + 0.0006106143118813634, + 0.0006372684729285538, + 0.0004936937475576997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.05413895845413208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486398458480835 + ], + "opposite_logprob_tokens": [ + -1.2236398458480835 + ], + "logprob_reward": -0.3486398458480835, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486398458480835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02779546007514, + 0.4228837192058563, + 0.028142787516117096, + 0.017729157581925392, + 0.02502240613102913, + 0.01707005873322487, + 0.031621333211660385, + 0.40325549244880676, + 0.006550435442477465, + 0.01992921531200409 + ], + "credit_map": [ + 0.0032801954075694084, + 0.0028788589406758547, + 0.0051168352365493774, + 0.0031157524790614843, + 0.0023359728511422873, + 0.0030870309565216303, + 0.004748651757836342, + 0.003232161747291684, + 0.02169012650847435, + 0.16895076632499695, + 0.011837486177682877, + 0.014379140920937061, + 0.02027982659637928, + 0.16124680638313293, + 0.009359247051179409, + 0.015140321105718613, + 0.0027729489374905825, + 0.0029569861944764853, + 0.0032398279290646315, + 0.003385762684047222, + 0.005780460778623819, + 0.005269330460578203, + 0.002797663677483797, + 0.0019398079020902514, + 0.0013219766551628709, + 0.002236476866528392, + 0.0009416762040928006, + 0.0043087671510875225, + 0.0020432542078197002, + 0.002271524863317609, + 0.0011946854647248983, + 0.0034107952378690243, + 0.0014450792223215103, + 0.006198283284902573, + 0.0015411135973408818, + 0.007499230559915304, + 0.002390841022133827, + 0.0012962102191522717, + 0.0028473727870732546, + 0.0018042774172499776, + 0.0016718131955713034, + 0.0034593227319419384, + 0.0019161718664690852, + 0.0017742523923516273, + 0.0029973869677633047, + 0.0014836692716926336, + 0.0022673902567476034, + 0.0015000526327639818, + 0.003993962425738573, + 0.006651473697274923, + 0.0018545325146988034, + 0.006042861845344305, + 0.002205977216362953, + 0.0031882186885923147, + 0.00501042976975441, + 0.0026738790329545736, + 0.024036962538957596, + 0.14894601702690125, + 0.010005266405642033, + 0.010173844173550606, + 0.15910948812961578, + 0.031594716012477875, + 0.010478408075869083, + 0.008910768665373325, + 0.0007375605055131018, + 0.0008694540010765195, + 0.0009403042495250702, + 0.0010086148977279663, + 0.0007717267144471407, + 0.0008915960206650198, + 0.0007691231439821422, + 0.0005620555020868778, + 0.001610533450730145, + 0.004269870929419994, + 0.001906102173961699, + 0.005056386813521385, + 0.0027304862160235643, + 0.0015129846287891269, + 0.0014642009045928717, + 0.0013786490308120847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.3486398458480835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028306676540523767 + ], + "opposite_logprob_tokens": [ + -5.877830505371094 + ], + "logprob_reward": -0.0028306676540523767, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028306676540523767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053472209721803665, + 0.7466031312942505, + 0.032687682658433914, + 0.014198335818946362, + 0.05577486753463745, + 0.00825099740177393, + 0.016272976994514465, + 0.05159307271242142, + 0.011216357350349426, + 0.00993034802377224 + ], + "credit_map": [ + 0.006342108827084303, + 0.012162267230451107, + 0.006769945845007896, + 0.004618881735950708, + 0.007191742304712534, + 0.006926952861249447, + 0.004863939713686705, + 0.004596374463289976, + 0.03645564988255501, + 0.29962828755378723, + 0.02095843106508255, + 0.024579597637057304, + 0.03555859997868538, + 0.28630560636520386, + 0.015825310721993446, + 0.02729174867272377, + 0.004572277422994375, + 0.0031487217638641596, + 0.004115718882530928, + 0.0025435625575482845, + 0.007613783236593008, + 0.004290517885237932, + 0.0038539913948625326, + 0.0025491106789559126, + 0.0008955838275142014, + 0.0018573375418782234, + 0.000785553827881813, + 0.002802755683660507, + 0.001767315436154604, + 0.0012143600033596158, + 0.0009806951275095344, + 0.003894735826179385, + 0.0029938961379230022, + 0.017522208392620087, + 0.002468451391905546, + 0.01671251282095909, + 0.006658875849097967, + 0.0021813157945871353, + 0.003959648311138153, + 0.0032779648900032043, + 0.0009409100748598576, + 0.0008447038126178086, + 0.0012549984967336059, + 0.001153807621449232, + 0.0010301978327333927, + 0.001342162024229765, + 0.0007291474612429738, + 0.0009550710674375296, + 0.0018762463005259633, + 0.003198192687705159, + 0.0011663037585094571, + 0.0032852592412382364, + 0.0017284146742895246, + 0.001634026411920786, + 0.0019122229423373938, + 0.0014723128406330943, + 0.0036204007919877768, + 0.01294164452701807, + 0.0035337796434760094, + 0.00599326379597187, + 0.013698253780603409, + 0.005761751905083656, + 0.003343223826959729, + 0.002700763288885355, + 0.0019739605486392975, + 0.0013536869082599878, + 0.0026041094679385424, + 0.0009265280095860362, + 0.0012138945749029517, + 0.0011726767988875508, + 0.0008170103537850082, + 0.0011544915614649653, + 0.0005615600966848433, + 0.0020453170873224735, + 0.0009728990262374282, + 0.002015948761254549, + 0.001672833925113082, + 0.0009213573066517711, + 0.000973843561951071, + 0.0007665896555408835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0028306676540523767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134702742099762 + ], + "opposite_logprob_tokens": [ + -1.3134702444076538 + ], + "logprob_reward": -0.3134702742099762, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134702742099762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014874829910695553, + 0.8052052855491638, + 0.01782059296965599, + 0.008737445808947086, + 0.046740446239709854, + 0.005956349428743124, + 0.005740372464060783, + 0.06119878590106964, + 0.01970231533050537, + 0.014023511670529842 + ], + "credit_map": [ + 0.0019335615215823054, + 0.0018627102253958583, + 0.002271803095936775, + 0.0017069842433556914, + 0.00117783946916461, + 0.0016499594785273075, + 0.0021132442634552717, + 0.0021587281953543425, + 0.04000470042228699, + 0.32320475578308105, + 0.023018062114715576, + 0.026574818417429924, + 0.03808078542351723, + 0.3078341782093048, + 0.017214516177773476, + 0.029273489490151405, + 0.0028127271216362715, + 0.0016156474594026804, + 0.001812163507565856, + 0.0017576228128746152, + 0.0050266217440366745, + 0.00238641444593668, + 0.0013695606030523777, + 0.00103983492590487, + 0.0006572140264324844, + 0.0009899868164211512, + 0.0006505921483039856, + 0.0023349933326244354, + 0.0012326466385275126, + 0.0011870471062138677, + 0.0005436618812382221, + 0.0011413046158850193, + 0.00258627743460238, + 0.014526533894240856, + 0.0019366355845704675, + 0.015287174843251705, + 0.004425546620041132, + 0.0016742823645472527, + 0.003525612410157919, + 0.0027783820405602455, + 0.00037952925777062774, + 0.0015105122001841664, + 0.0006024704198352993, + 0.0007756553241051733, + 0.0010268664918839931, + 0.0005309979896992445, + 0.0006734196795150638, + 0.00045689818216487765, + 0.0006606322713196278, + 0.0006417249678634107, + 0.0006316878134384751, + 0.001291124615818262, + 0.0005915501969866455, + 0.0006127975648269057, + 0.0007014125003479421, + 0.0006094428827054799, + 0.003572121961042285, + 0.020666414871811867, + 0.0021122132893651724, + 0.003274887567386031, + 0.02201099693775177, + 0.005811648443341255, + 0.0020383039955049753, + 0.0017122025601565838, + 0.0024264852982014418, + 0.0023316428996622562, + 0.004708689637482166, + 0.002627689391374588, + 0.0020235038828104734, + 0.00245614699088037, + 0.0017313010757789016, + 0.0013968553394079208, + 0.0010536324698477983, + 0.002227941295132041, + 0.001670940313488245, + 0.0029819977935403585, + 0.0016060874331742525, + 0.0012854979140684009, + 0.002159543801099062, + 0.001037870068103075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.3134702742099762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369959816336632 + ], + "opposite_logprob_tokens": [ + -3.4086995124816895 + ], + "logprob_reward": -0.03369959816336632, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369959816336632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04190460219979286, + 0.7084529399871826, + 0.1316409856081009, + 0.012795154936611652, + 0.03003464825451374, + 0.013430659659206867, + 0.010889431461691856, + 0.032971497625112534, + 0.007439898326992989, + 0.010440248996019363 + ], + "credit_map": [ + 0.004857673309743404, + 0.00578629644587636, + 0.0051275985315442085, + 0.004272917751222849, + 0.005242999643087387, + 0.006743362173438072, + 0.004237115848809481, + 0.005636637099087238, + 0.034737251698970795, + 0.2878405451774597, + 0.019755013287067413, + 0.02513636089861393, + 0.033742137253284454, + 0.26767048239707947, + 0.015182768926024437, + 0.024388387799263, + 0.02237078920006752, + 0.011584562249481678, + 0.012992200441658497, + 0.011444847099483013, + 0.03712180256843567, + 0.019089382141828537, + 0.009756194427609444, + 0.007281211204826832, + 0.001120489789173007, + 0.002532240003347397, + 0.0008552828803658485, + 0.001427177106961608, + 0.0026573033537715673, + 0.001718534855172038, + 0.0014402713859453797, + 0.0010438555618748069, + 0.0014568292535841465, + 0.0091986283659935, + 0.0014728999231010675, + 0.008186624385416508, + 0.0037577326875180006, + 0.00139511504676193, + 0.0024154188577085733, + 0.0021513979882001877, + 0.0009316274663433433, + 0.003512952011078596, + 0.0025331215001642704, + 0.0011540378909558058, + 0.0020450917072594166, + 0.0015150534454733133, + 0.0008617441053502262, + 0.0008770313579589128, + 0.0010749161010608077, + 0.00157555949408561, + 0.0007883802172727883, + 0.0028027205262333155, + 0.0015050163492560387, + 0.001103327958844602, + 0.0011515088845044374, + 0.0008880012901499867, + 0.0033296546898782253, + 0.0032255924306809902, + 0.0038155780639499426, + 0.007332042325288057, + 0.003342473180964589, + 0.005584545899182558, + 0.003353782696649432, + 0.0029878271743655205, + 0.0007798906299285591, + 0.0013014116557314992, + 0.0016482924111187458, + 0.0007797603029757738, + 0.001086801989004016, + 0.0008872065227478743, + 0.0005728723481297493, + 0.0003836624382529408, + 0.0005809037829749286, + 0.0018711108714342117, + 0.0012336900690570474, + 0.0025696405209600925, + 0.0014629389625042677, + 0.0007863835198804736, + 0.0012686204863712192, + 0.0006669607246294618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03369959816336632 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Low-angle tracking shots, stylized in a style reminiscent of high-speed sports documentaries, glide smoothly from left to right, just above the ground, immersing the viewer in the sun-drenched fields of Monarda di Dima. The camera captures their vibrant red, spiky flower heads, each a chaotic crown of tubular petals swaying animatedly atop tall, slender green stems. In the brilliant, high-contrast light of the midday sun, every detail is rendered with hyper-real clarity; petals shimmer, while sharp, dark shadows define their complex forms. A persistent breeze sweeps across the field, inspiring a chaotic yet synchronized dance, with the flowers flowing across the frame in powerful, rhythmic waves, dipping, swaying, and bouncing. The camera's fluid, purposeful movements, combined with a high-definition aesthetic, fuse this natural spectacle with the dynamic energy and concentrated intensity of a major sporting event, creating a vibrant and exciting atmosphere.", + "seed": 12812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_9760_seed_12812_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11625511501915753, + "vlm_logprob": -0.11625511501915753, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.11625511501915753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05222855508327484 + ], + "opposite_logprob_tokens": [ + -4.3022284507751465 + ], + "logprob_reward": -0.05222855508327484, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.05222855508327484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06901855766773224, + 0.01827901042997837, + 0.1426471471786499, + 0.023260122165083885, + 0.02700495906174183, + 0.055870071053504944, + 0.44274792075157166, + 0.04415864869952202, + 0.13753105700016022, + 0.03948243334889412 + ], + "credit_map": [ + 0.01653522066771984, + 0.007172322832047939, + 0.007849875837564468, + 0.00532778212800622, + 0.013899842277169228, + 0.009358780458569527, + 0.00474834768101573, + 0.004126391373574734, + 0.0023144737351685762, + 0.002083238447085023, + 0.0028361177537590265, + 0.002065521664917469, + 0.0028475031722337008, + 0.0020619418937712908, + 0.0019947546534240246, + 0.0020754605066031218, + 0.0076182931661605835, + 0.007410246413201094, + 0.007911124266684055, + 0.005461026448756456, + 0.012572593055665493, + 0.04259001091122627, + 0.04685061797499657, + 0.012233252637088299, + 0.002092324197292328, + 0.002956234384328127, + 0.0031238484662026167, + 0.003438504645600915, + 0.0034040571190416813, + 0.002849181881174445, + 0.0025749094784259796, + 0.00282106245867908, + 0.002433453919366002, + 0.007709484547376633, + 0.0023230609949678183, + 0.0025058614555746317, + 0.0041273534297943115, + 0.0035614417865872383, + 0.0024827546440064907, + 0.0018615495646372437, + 0.0056433058343827724, + 0.015074054710566998, + 0.004476484842598438, + 0.005133247468620539, + 0.008918432518839836, + 0.006434848066419363, + 0.005831360816955566, + 0.004358336329460144, + 0.0182667076587677, + 0.18290676176548004, + 0.006955298129469156, + 0.013704205863177776, + 0.17492538690567017, + 0.0268685445189476, + 0.010913755744695663, + 0.008207292295992374, + 0.003795196767896414, + 0.004301639273762703, + 0.004796199966222048, + 0.007163433823734522, + 0.005628238897770643, + 0.004001748748123646, + 0.0069642807357013226, + 0.0075079151429235935, + 0.026221703737974167, + 0.020161373540759087, + 0.014195808209478855, + 0.013266954571008682, + 0.03217695280909538, + 0.012828524224460125, + 0.008933363482356071, + 0.009746380150318146, + 0.004909060895442963, + 0.010722551494836807, + 0.0036968765780329704, + 0.0044666170142591, + 0.0045776329934597015, + 0.006197695154696703, + 0.0026819859631359577, + 0.0022300162818282843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.05222855508327484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134714961051941 + ], + "opposite_logprob_tokens": [ + -1.3134715557098389 + ], + "logprob_reward": -0.3134714961051941, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134714961051941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06807185709476471, + 0.02977708913385868, + 0.05784761533141136, + 0.06339620798826218, + 0.06816596537828445, + 0.09734801948070526, + 0.4347900450229645, + 0.06702613085508347, + 0.07982434332370758, + 0.0337526910007 + ], + "credit_map": [ + 0.013711052015423775, + 0.0055604418739676476, + 0.02152867801487446, + 0.0074237012304365635, + 0.004609627649188042, + 0.00647730054333806, + 0.005124043673276901, + 0.0036370176821947098, + 0.002683029742911458, + 0.0038250500801950693, + 0.0026306684594601393, + 0.006886606570333242, + 0.0033623490016907454, + 0.004516796674579382, + 0.003005167469382286, + 0.002867422765120864, + 0.0037939257454127073, + 0.004191800951957703, + 0.005752264522016048, + 0.0033480345737189054, + 0.005477841477841139, + 0.015738731250166893, + 0.015653446316719055, + 0.003891577012836933, + 0.009708794765174389, + 0.006119270343333483, + 0.006078762002289295, + 0.006799446418881416, + 0.009868199005723, + 0.008062808774411678, + 0.005973084829747677, + 0.010785848833620548, + 0.00605032267048955, + 0.021011268720030785, + 0.006575531791895628, + 0.005540064536035061, + 0.01005532592535019, + 0.007903934456408024, + 0.005825459025800228, + 0.005204064771533012, + 0.00955883041024208, + 0.02635662630200386, + 0.008300891146063805, + 0.00949741993099451, + 0.015697726979851723, + 0.011121154762804508, + 0.009485523216426373, + 0.007329848594963551, + 0.01954534836113453, + 0.17784690856933594, + 0.007090842816978693, + 0.014130250550806522, + 0.17033909261226654, + 0.024899516254663467, + 0.012347922660410404, + 0.008590190671384335, + 0.006459256634116173, + 0.009306243620812893, + 0.006199798546731472, + 0.00871828943490982, + 0.011423262767493725, + 0.0070937941782176495, + 0.00952926930040121, + 0.008296222425997257, + 0.016862794756889343, + 0.01121929194778204, + 0.007599278818815947, + 0.005977144930511713, + 0.019568515941500664, + 0.007299779914319515, + 0.005050142761319876, + 0.006247405428439379, + 0.003418693318963051, + 0.0071985917165875435, + 0.00352575839497149, + 0.004536351654678583, + 0.004285270813852549, + 0.0049362946301698685, + 0.00306383497081697, + 0.0027878987602889538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.3134714961051941 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001958120847120881 + ], + "opposite_logprob_tokens": [ + -6.251957893371582 + ], + "logprob_reward": -0.001958120847120881, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001958120847120881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16955937445163727, + 0.09245425462722778, + 0.25958478450775146, + 0.037695374339818954, + 0.03711610659956932, + 0.038428470492362976, + 0.2426781952381134, + 0.06805328279733658, + 0.0229665357619524, + 0.03146364912390709 + ], + "credit_map": [ + 0.0315370038151741, + 0.016035132110118866, + 0.049813881516456604, + 0.014983451925218105, + 0.011644228361546993, + 0.018881019204854965, + 0.01206375565379858, + 0.01460090558975935, + 0.016692129895091057, + 0.009943039156496525, + 0.009741686284542084, + 0.017553137615323067, + 0.009908890351653099, + 0.009270296432077885, + 0.00962644349783659, + 0.009718632325530052, + 0.014847527258098125, + 0.018471822142601013, + 0.01804746501147747, + 0.016660310328006744, + 0.023578962311148643, + 0.06881064921617508, + 0.0770193338394165, + 0.02214871346950531, + 0.0036373715847730637, + 0.004497947171330452, + 0.0030387372244149446, + 0.0038615716621279716, + 0.00645331060513854, + 0.005926618818193674, + 0.0033167575020343065, + 0.006963060237467289, + 0.0031634618062525988, + 0.011375532485544682, + 0.0033688999246805906, + 0.002967288950458169, + 0.005263154394924641, + 0.004653315991163254, + 0.0031964131630957127, + 0.003128036856651306, + 0.004047869239002466, + 0.009039039723575115, + 0.003056569956243038, + 0.0038700976874679327, + 0.00745646096765995, + 0.00408667279407382, + 0.003988496493548155, + 0.0028832678217440844, + 0.009750070050358772, + 0.09582174569368362, + 0.006022250279784203, + 0.009767243638634682, + 0.09135150909423828, + 0.01678052730858326, + 0.007736672647297382, + 0.005448176059871912, + 0.007786298170685768, + 0.008653903380036354, + 0.00697143143042922, + 0.010700947605073452, + 0.00998807791620493, + 0.007079704664647579, + 0.008599094115197659, + 0.008273830637335777, + 0.0035449357237666845, + 0.0029759169556200504, + 0.0019341144943609834, + 0.0032674772664904594, + 0.003474954515695572, + 0.0030951437074691057, + 0.0020324834622442722, + 0.0026415095198899508, + 0.002667535562068224, + 0.0067373840138316154, + 0.003487556939944625, + 0.003475017147138715, + 0.0029503335244953632, + 0.008039138279855251, + 0.0020562876015901566, + 0.0020503958221524954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.001958120847120881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992103099822998 + ], + "opposite_logprob_tokens": [ + -1.80492103099823 + ], + "logprob_reward": -0.17992103099822998, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992103099822998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07614915072917938, + 0.034087639302015305, + 0.08857223391532898, + 0.04871285706758499, + 0.07582496106624603, + 0.08207015693187714, + 0.4182691276073456, + 0.04250529035925865, + 0.09865694493055344, + 0.03515161946415901 + ], + "credit_map": [ + 0.008061339147388935, + 0.006961116101592779, + 0.019389964640140533, + 0.009114737622439861, + 0.009937286376953125, + 0.0069926222786307335, + 0.007165729533880949, + 0.008526355028152466, + 0.0049616489559412, + 0.002733614295721054, + 0.0036490196362137794, + 0.006938721053302288, + 0.0046387724578380585, + 0.004697698168456554, + 0.003354659304022789, + 0.003113505896180868, + 0.005707471631467342, + 0.005213480442762375, + 0.0068281544372439384, + 0.006139467936009169, + 0.007883589714765549, + 0.021743228659033775, + 0.025807436555624008, + 0.009249399416148663, + 0.003915256820619106, + 0.004961366765201092, + 0.0049471743404865265, + 0.00425676628947258, + 0.011998003348708153, + 0.006863441783934832, + 0.005008960608392954, + 0.006761886645108461, + 0.00533529045060277, + 0.011581335216760635, + 0.006313628517091274, + 0.006698292214423418, + 0.011172838509082794, + 0.0190825667232275, + 0.0081925168633461, + 0.007448490709066391, + 0.00929897278547287, + 0.022445108741521835, + 0.006656815763562918, + 0.008428004570305347, + 0.013604044914245605, + 0.008362519554793835, + 0.0076499502174556255, + 0.0056247347965836525, + 0.01919356919825077, + 0.16730473935604095, + 0.00940294936299324, + 0.019401274621486664, + 0.15757021307945251, + 0.024523312225937843, + 0.011280682869255543, + 0.009592429734766483, + 0.0048171537928283215, + 0.0055177719332277775, + 0.005635637324303389, + 0.004129885230213404, + 0.006153912283480167, + 0.005472296383231878, + 0.005800552200525999, + 0.004978081677109003, + 0.02022898569703102, + 0.014180080033838749, + 0.010012669488787651, + 0.008459271863102913, + 0.022909779101610184, + 0.009135534055531025, + 0.00627987552434206, + 0.007450748700648546, + 0.0039961370639503, + 0.008190315216779709, + 0.0035190414637327194, + 0.004118174314498901, + 0.005315588321536779, + 0.004687346983700991, + 0.002889922820031643, + 0.0024350914172828197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.17992103099822998 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369637206196785 + ], + "opposite_logprob_tokens": [ + -3.408696413040161 + ], + "logprob_reward": -0.03369637206196785, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369637206196785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15087561309337616, + 0.08696810156106949, + 0.08196183294057846, + 0.06376981735229492, + 0.030060365796089172, + 0.03529795631766319, + 0.31216365098953247, + 0.05648836866021156, + 0.15511825680732727, + 0.02729606069624424 + ], + "credit_map": [ + 0.017949294298887253, + 0.015737038105726242, + 0.02669910527765751, + 0.019252484664320946, + 0.02175925299525261, + 0.01739858277142048, + 0.01527194119989872, + 0.01680791564285755, + 0.011521856300532818, + 0.00782422348856926, + 0.010456464253365993, + 0.018968038260936737, + 0.010088276118040085, + 0.010214724577963352, + 0.010507981292903423, + 0.007386540528386831, + 0.006340960506349802, + 0.008109093643724918, + 0.0062633780762553215, + 0.007765761576592922, + 0.007585183251649141, + 0.01612074486911297, + 0.021341484040021896, + 0.008435229770839214, + 0.008441679179668427, + 0.00719842966645956, + 0.0056146904826164246, + 0.006004024296998978, + 0.014347219839692116, + 0.009210768155753613, + 0.005777857732027769, + 0.007175151724368334, + 0.0030327271670103073, + 0.007753267418593168, + 0.0027492663357406855, + 0.0025336151011288166, + 0.0047968425787985325, + 0.0041065034456551075, + 0.0023249711375683546, + 0.002763171913102269, + 0.0037535633891820908, + 0.008406075648963451, + 0.0037851815577596426, + 0.003863684367388487, + 0.005297897849231958, + 0.0037262781988829374, + 0.003535372903570533, + 0.0029299012385308743, + 0.01256021298468113, + 0.12877240777015686, + 0.005330362357199192, + 0.009916937910020351, + 0.12256351858377457, + 0.019451597705483437, + 0.007984546944499016, + 0.005584060680121183, + 0.005884818732738495, + 0.008964341133832932, + 0.005150451324880123, + 0.006197158247232437, + 0.008533366955816746, + 0.006144988350570202, + 0.006452178116887808, + 0.00916106253862381, + 0.03209792822599411, + 0.021805744618177414, + 0.015922192484140396, + 0.014027934521436691, + 0.03541523218154907, + 0.015143482945859432, + 0.010359245352447033, + 0.010346491821110249, + 0.0033950209617614746, + 0.005678663961589336, + 0.0022672952618449926, + 0.003530357964336872, + 0.003502526553347707, + 0.004947409965097904, + 0.0017738803289830685, + 0.0022009059321135283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.03369637206196785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Low-angle tracking shots, stylized in a style reminiscent of high-speed sports documentaries, glide smoothly from left to right, just above the ground, immersing the viewer in the sun-drenched fields of Monarda di Dima. The camera captures their vibrant red, spiky flower heads, each a chaotic crown of tubular petals swaying animatedly atop tall, slender green stems. In the brilliant, high-contrast light of the midday sun, every detail is rendered with hyper-real clarity; petals shimmer, while sharp, dark shadows define their complex forms. A persistent breeze sweeps across the field, inspiring a chaotic yet synchronized dance, with the flowers flowing across the frame in powerful, rhythmic waves, dipping, swaying, and bouncing. The camera's fluid, purposeful movements, combined with a high-definition aesthetic, fuse this natural spectacle with the dynamic energy and concentrated intensity of a major sporting event, creating a vibrant and exciting atmosphere.", + "seed": 12813, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_9760_seed_12813_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17055000998079778, + "vlm_logprob": -0.17055000998079778, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.17055000998079778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06414856761693954 + ], + "opposite_logprob_tokens": [ + -3.6891486644744873 + ], + "logprob_reward": -0.06414856761693954, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.06414856761693954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2123032510280609, + 0.2382814735174179, + 0.06650342792272568, + 0.025837305933237076, + 0.05448710545897484, + 0.06142686679959297, + 0.09801143407821655, + 0.05251404270529747, + 0.14870582520961761, + 0.0419292226433754 + ], + "credit_map": [ + 0.019416527822613716, + 0.022119777277112007, + 0.03532920032739639, + 0.037885770201683044, + 0.020079784095287323, + 0.018290990963578224, + 0.029312938451766968, + 0.02986828051507473, + 0.028052518144249916, + 0.011177613399922848, + 0.09264946728944778, + 0.014390794560313225, + 0.011147118173539639, + 0.016152577474713326, + 0.05512785539031029, + 0.009583563543856144, + 0.014813434332609177, + 0.008624127134680748, + 0.0058010113425552845, + 0.006465649697929621, + 0.006076284684240818, + 0.010370252653956413, + 0.008845306001603603, + 0.005507373251020908, + 0.0029708368238061666, + 0.001181200030259788, + 0.002506387885659933, + 0.0025416191201657057, + 0.0015206563984975219, + 0.0038691333029419184, + 0.004547136835753918, + 0.0067003387957811356, + 0.004593206569552422, + 0.005464127752929926, + 0.0067475237883627415, + 0.010933786630630493, + 0.004952163901180029, + 0.006151286885142326, + 0.008103782311081886, + 0.00754123367369175, + 0.005669679492712021, + 0.00616232817992568, + 0.010125440545380116, + 0.00915312860161066, + 0.0039881388656795025, + 0.008490883745253086, + 0.010816173627972603, + 0.007021100725978613, + 0.01013324223458767, + 0.010917723178863525, + 0.011282811872661114, + 0.01069005485624075, + 0.012116624042391777, + 0.013649274595081806, + 0.01329846866428852, + 0.01592324860394001, + 0.004286197014153004, + 0.00538833299651742, + 0.01078473124653101, + 0.005637096706777811, + 0.005051928572356701, + 0.009148009121418, + 0.006364570930600166, + 0.005853179842233658, + 0.00877682026475668, + 0.020624961704015732, + 0.010144870728254318, + 0.05196046829223633, + 0.012149311602115631, + 0.01517558190971613, + 0.014324519783258438, + 0.015549307689070702, + 0.0039081452414393425, + 0.005015346221625805, + 0.0049049085937440395, + 0.004746740683913231, + 0.00859004259109497, + 0.004219132009893656, + 0.005355602130293846, + 0.005189310293644667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.06414856761693954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34864866733551025 + ], + "opposite_logprob_tokens": [ + -1.2236486673355103 + ], + "logprob_reward": -0.34864866733551025, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34864866733551025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05467779561877251, + 0.1502426564693451, + 0.021113678812980652, + 0.04168371111154556, + 0.11251188814640045, + 0.062361281365156174, + 0.14526601135730743, + 0.1357918083667755, + 0.2121180146932602, + 0.06423310935497284 + ], + "credit_map": [ + 0.005353371147066355, + 0.005327795632183552, + 0.0071266572922468185, + 0.007246929220855236, + 0.00768292834982276, + 0.0073913573287427425, + 0.008503571152687073, + 0.006045193411409855, + 0.047540560364723206, + 0.0072820535860955715, + 0.038240011781454086, + 0.011768911965191364, + 0.006264216266572475, + 0.00961320661008358, + 0.016904423013329506, + 0.012629276141524315, + 0.004037810489535332, + 0.0028572697192430496, + 0.002157064387574792, + 0.003267211141064763, + 0.0013974873581901193, + 0.003125355113297701, + 0.0024624578654766083, + 0.0018090219236910343, + 0.0020123410504311323, + 0.0019360334845259786, + 0.002117657568305731, + 0.00207883189432323, + 0.002541055902838707, + 0.015555739402770996, + 0.0022624207194894552, + 0.01317963469773531, + 0.009652581997215748, + 0.014903220348060131, + 0.012636909261345863, + 0.017577728256583214, + 0.012264752760529518, + 0.017939120531082153, + 0.015477621927857399, + 0.01205996423959732, + 0.004965085070580244, + 0.006638537626713514, + 0.007484289351850748, + 0.009362490847706795, + 0.00411105016246438, + 0.007960290648043156, + 0.010593688115477562, + 0.011245853267610073, + 0.015362289734184742, + 0.016909243538975716, + 0.018624017015099525, + 0.017426032572984695, + 0.015253126621246338, + 0.01875806786119938, + 0.01978297345340252, + 0.02315026894211769, + 0.007227828726172447, + 0.009333307854831219, + 0.018936442211270332, + 0.011082557961344719, + 0.015446416102349758, + 0.024376582354307175, + 0.02582991123199463, + 0.023558765649795532, + 0.013819596730172634, + 0.030304493382573128, + 0.022031618282198906, + 0.05171043053269386, + 0.02983219549059868, + 0.020186875015497208, + 0.02031775377690792, + 0.023915058001875877, + 0.004499480593949556, + 0.008036771789193153, + 0.00456780381500721, + 0.008562631905078888, + 0.013934523798525333, + 0.006270973011851311, + 0.00911384541541338, + 0.009247084148228168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.34864866733551025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798524901270866 + ], + "opposite_logprob_tokens": [ + -4.634798526763916 + ], + "logprob_reward": -0.009798524901270866, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798524901270866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11888209730386734, + 0.2394367754459381, + 0.03488076478242874, + 0.05518271401524544, + 0.10509738326072693, + 0.08327263593673706, + 0.05314348265528679, + 0.2245015799999237, + 0.048403311520814896, + 0.037199296057224274 + ], + "credit_map": [ + 0.010783945210278034, + 0.013962960802018642, + 0.015179279260337353, + 0.01557352114468813, + 0.024639880284667015, + 0.012007071636617184, + 0.010888016782701015, + 0.015847425907850266, + 0.07551715523004532, + 0.019223367795348167, + 0.04919736459851265, + 0.013490783050656319, + 0.018659045919775963, + 0.01734348013997078, + 0.02872310020029545, + 0.01728246733546257, + 0.006701791658997536, + 0.005281329154968262, + 0.003009103238582611, + 0.0034685558639466763, + 0.0034261122345924377, + 0.005700637120753527, + 0.0035598003305494785, + 0.00373343494720757, + 0.004391964990645647, + 0.003925619181245565, + 0.005387422163039446, + 0.004205111414194107, + 0.004429478198289871, + 0.012139258906245232, + 0.008398167788982391, + 0.012305693700909615, + 0.009724300354719162, + 0.013987009413540363, + 0.010152310132980347, + 0.013045129366219044, + 0.01400347426533699, + 0.01643325388431549, + 0.014175607822835445, + 0.013576298952102661, + 0.007944884710013866, + 0.011191071942448616, + 0.010550743900239468, + 0.007203946355730295, + 0.0077324057929217815, + 0.013725902885198593, + 0.01415910106152296, + 0.01076457928866148, + 0.0041271476075053215, + 0.006709437817335129, + 0.006515194196254015, + 0.007135553751140833, + 0.00759533466771245, + 0.00799557100981474, + 0.006925043649971485, + 0.006140199955552816, + 0.018782801926136017, + 0.01616072840988636, + 0.04534244164824486, + 0.022606229409575462, + 0.032821450382471085, + 0.03234719857573509, + 0.03307299688458443, + 0.023367727175354958, + 0.0036609817761927843, + 0.006303593050688505, + 0.006084244232624769, + 0.0042237103916704655, + 0.009102442301809788, + 0.0045199268497526646, + 0.007533707190304995, + 0.006974705029278994, + 0.003628369187936187, + 0.004797704517841339, + 0.004074456170201302, + 0.004878987558186054, + 0.006770312320441008, + 0.004273023921996355, + 0.004116409458220005, + 0.0046600340865552425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.009798524901270866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870881497859955 + ], + "opposite_logprob_tokens": [ + -1.1370881795883179 + ], + "logprob_reward": -0.3870881497859955, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870881497859955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1513994187116623, + 0.13909120857715607, + 0.1071925014257431, + 0.11193384975194931, + 0.07083974033594131, + 0.10694438964128494, + 0.04146745800971985, + 0.13764473795890808, + 0.07302858680486679, + 0.060458146035671234 + ], + "credit_map": [ + 0.014502107165753841, + 0.01634693704545498, + 0.018980175256729126, + 0.02136869914829731, + 0.017235741019248962, + 0.018097568303346634, + 0.023426199331879616, + 0.02144198678433895, + 0.04889658838510513, + 0.0074843731708824635, + 0.011783892288804054, + 0.00973571464419365, + 0.00763612100854516, + 0.008531885221600533, + 0.03692066669464111, + 0.00810196902602911, + 0.018825169652700424, + 0.01644819602370262, + 0.013108714483678341, + 0.010377406142652035, + 0.00824730470776558, + 0.015312375500798225, + 0.016234155744314194, + 0.008639182895421982, + 0.00917134527117014, + 0.007601907476782799, + 0.010505635291337967, + 0.012160249054431915, + 0.00816735066473484, + 0.030442669987678528, + 0.010982298292219639, + 0.02290239930152893, + 0.006121049169450998, + 0.0113004669547081, + 0.007221264764666557, + 0.00857642199844122, + 0.009332441724836826, + 0.011103452183306217, + 0.008315550163388252, + 0.008869093842804432, + 0.00964383501559496, + 0.011257712729275227, + 0.012940151616930962, + 0.016720805317163467, + 0.0061229257844388485, + 0.015683377161622047, + 0.013866669498383999, + 0.02070891484618187, + 0.003586679697036743, + 0.0027483024168759584, + 0.00921783596277237, + 0.004916644189506769, + 0.002839344786480069, + 0.004007660783827305, + 0.004661395214498043, + 0.009489595890045166, + 0.009314344264566898, + 0.008894559927284718, + 0.013324366882443428, + 0.01392016839236021, + 0.015359927900135517, + 0.02545115537941456, + 0.028591806069016457, + 0.022788405418395996, + 0.004802043084055185, + 0.012245585210621357, + 0.005140097811818123, + 0.021395325660705566, + 0.007193200755864382, + 0.008200757205486298, + 0.008390243165194988, + 0.0056613353081047535, + 0.0068045686930418015, + 0.005311949178576469, + 0.00734720891341567, + 0.00784866139292717, + 0.009161186404526234, + 0.007269124034792185, + 0.008966638706624508, + 0.007748807314783335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.3870881497859955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306614026427269 + ], + "opposite_logprob_tokens": [ + -3.1680660247802734 + ], + "logprob_reward": -0.04306614026427269, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306614026427269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15616190433502197, + 0.18637531995773315, + 0.05334371700882912, + 0.036666251718997955, + 0.0924062579870224, + 0.07689861953258514, + 0.048288147896528244, + 0.17824634909629822, + 0.12175791710615158, + 0.0498555451631546 + ], + "credit_map": [ + 0.013708393089473248, + 0.016786424443125725, + 0.0242100041359663, + 0.023468518629670143, + 0.024602102115750313, + 0.01243365928530693, + 0.01833667978644371, + 0.022616133093833923, + 0.06337114423513412, + 0.010787542909383774, + 0.01735772006213665, + 0.015665246173739433, + 0.011626253835856915, + 0.011234561912715435, + 0.04571923241019249, + 0.010613596998155117, + 0.010517044924199581, + 0.005257640033960342, + 0.005533581133931875, + 0.005743136163800955, + 0.0059834630228579044, + 0.008711223490536213, + 0.006109153386205435, + 0.005488470196723938, + 0.0033138361759483814, + 0.004071755334734917, + 0.0048269121907651424, + 0.004237837623804808, + 0.004580206703394651, + 0.005235414952039719, + 0.004059473518282175, + 0.006340809166431427, + 0.00912524200975895, + 0.012287188321352005, + 0.010950262658298016, + 0.01413275022059679, + 0.009201277047395706, + 0.013792471028864384, + 0.012287321500480175, + 0.010629743337631226, + 0.008500322699546814, + 0.009689871221780777, + 0.006712278351187706, + 0.013277865014970303, + 0.006033428478986025, + 0.011575287207961082, + 0.008002061396837234, + 0.013107502833008766, + 0.0053044334053993225, + 0.005369571968913078, + 0.005680239759385586, + 0.005399617366492748, + 0.006374091375619173, + 0.008085490204393864, + 0.005107339937239885, + 0.006967364810407162, + 0.012841428630053997, + 0.010939083062112331, + 0.03049345873296261, + 0.021615376695990562, + 0.02124663256108761, + 0.025691628456115723, + 0.03154183551669121, + 0.023876886814832687, + 0.009086287580430508, + 0.017581729218363762, + 0.011047354899346828, + 0.03036055713891983, + 0.014737413264811039, + 0.017296042293310165, + 0.00925427209585905, + 0.012394251301884651, + 0.0041509694419801235, + 0.005656423978507519, + 0.004527355078607798, + 0.007874695584177971, + 0.009744000621140003, + 0.005129276309162378, + 0.006229052320122719, + 0.006543769035488367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.04306614026427269 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Low-angle tracking shots, stylized in a style reminiscent of high-speed sports documentaries, glide smoothly from left to right, just above the ground, immersing the viewer in the sun-drenched fields of Monarda di Dima. The camera captures their vibrant red, spiky flower heads, each a chaotic crown of tubular petals swaying animatedly atop tall, slender green stems. In the brilliant, high-contrast light of the midday sun, every detail is rendered with hyper-real clarity; petals shimmer, while sharp, dark shadows define their complex forms. A persistent breeze sweeps across the field, inspiring a chaotic yet synchronized dance, with the flowers flowing across the frame in powerful, rhythmic waves, dipping, swaying, and bouncing. The camera's fluid, purposeful movements, combined with a high-definition aesthetic, fuse this natural spectacle with the dynamic energy and concentrated intensity of a major sporting event, creating a vibrant and exciting atmosphere.", + "seed": 12814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_3/index_9760_seed_12814_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11620637241285295, + "vlm_logprob": -0.11620637241285295, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11620637241285295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06329235434532166 + ], + "opposite_logprob_tokens": [ + -3.4382922649383545 + ], + "logprob_reward": -0.06329235434532166, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.06329235434532166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13103054463863373, + 0.045105911791324615, + 0.3579971194267273, + 0.04810350388288498, + 0.11792665719985962, + 0.040255945175886154, + 0.03068768046796322, + 0.14161568880081177, + 0.04341629147529602, + 0.04386070743203163 + ], + "credit_map": [ + 0.01219878252595663, + 0.012971125543117523, + 0.012603759765625, + 0.013732311315834522, + 0.022357920184731483, + 0.010988418944180012, + 0.021473363041877747, + 0.02470484748482704, + 0.004157865885645151, + 0.004037455655634403, + 0.0073148468509316444, + 0.004833062179386616, + 0.005721069872379303, + 0.010649028234183788, + 0.00499396026134491, + 0.0033986179623752832, + 0.025141162797808647, + 0.020571183413267136, + 0.023525219410657883, + 0.12061918526887894, + 0.030474020168185234, + 0.10223089158535004, + 0.020351383835077286, + 0.015084028244018555, + 0.006471051834523678, + 0.011571628972887993, + 0.005404523573815823, + 0.004267850890755653, + 0.00822293758392334, + 0.004476598929613829, + 0.0033314486499875784, + 0.004357460420578718, + 0.01038892287760973, + 0.019074315205216408, + 0.020345540717244148, + 0.008832594379782677, + 0.014007613062858582, + 0.02159236930310726, + 0.0144508620724082, + 0.009234425611793995, + 0.004439608659595251, + 0.004266937263309956, + 0.0044861407950520515, + 0.004082940053194761, + 0.006032858043909073, + 0.006711104419082403, + 0.005342534277588129, + 0.004893822129815817, + 0.002995191141963005, + 0.004211351275444031, + 0.0038609961047768593, + 0.002997742500156164, + 0.005231240764260292, + 0.0048790620639920235, + 0.003996671177446842, + 0.0025154214818030596, + 0.016515731811523438, + 0.02084326557815075, + 0.019764341413974762, + 0.014409550465643406, + 0.015459938906133175, + 0.015744559466838837, + 0.022708751261234283, + 0.016169540584087372, + 0.006827250588685274, + 0.006881793960928917, + 0.003697515930980444, + 0.004487654659897089, + 0.006188135128468275, + 0.006041647866368294, + 0.005519063677638769, + 0.003773229196667671, + 0.0046333433128893375, + 0.005126233212649822, + 0.0045616645365953445, + 0.005040181800723076, + 0.007895857095718384, + 0.006330734118819237, + 0.00513117341324687, + 0.005141517613083124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.06329235434532166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255811244249344 + ], + "opposite_logprob_tokens": [ + -1.600581169128418 + ], + "logprob_reward": -0.2255811244249344, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255811244249344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1427033692598343, + 0.0372527651488781, + 0.3930646479129791, + 0.2081703245639801, + 0.0832671970129013, + 0.01946588233113289, + 0.0299907848238945, + 0.04720610752701759, + 0.025590071454644203, + 0.013288907706737518 + ], + "credit_map": [ + 0.009022924117743969, + 0.008103767409920692, + 0.008736787363886833, + 0.011892784386873245, + 0.03952130675315857, + 0.034504108130931854, + 0.019365370273590088, + 0.011556330136954784, + 0.003558041527867317, + 0.004577017389237881, + 0.005916304886341095, + 0.005004580598324537, + 0.005869424436241388, + 0.0048421770334243774, + 0.0044091506861150265, + 0.0030760697554796934, + 0.017516929656267166, + 0.025771601125597954, + 0.02746887132525444, + 0.1408289074897766, + 0.05070106312632561, + 0.09243558347225189, + 0.022766482084989548, + 0.015575200319290161, + 0.015117562375962734, + 0.03559558466076851, + 0.02378513291478157, + 0.020880358293652534, + 0.026725735515356064, + 0.02040683850646019, + 0.028842246159911156, + 0.03681686148047447, + 0.005734710488468409, + 0.008815835230052471, + 0.01939922384917736, + 0.005477081052958965, + 0.00622941181063652, + 0.026374932378530502, + 0.006001511123031378, + 0.005234488286077976, + 0.0023911602329462767, + 0.0021143679041415453, + 0.0022118797060102224, + 0.0022069320548325777, + 0.002963531529530883, + 0.003198642283678055, + 0.00210452894680202, + 0.00227484037168324, + 0.003014292335137725, + 0.00469884742051363, + 0.0037730105686932802, + 0.003135914681479335, + 0.004870986565947533, + 0.003891412867233157, + 0.0037079346366226673, + 0.0028983859810978174, + 0.0050048003904521465, + 0.006022333167493343, + 0.006387652829289436, + 0.004663942847400904, + 0.006783938035368919, + 0.005027075298130512, + 0.006250892765820026, + 0.007065473590046167, + 0.003431786084547639, + 0.005614044610410929, + 0.0020947284065186977, + 0.002915247343480587, + 0.002706959843635559, + 0.0027624524664133787, + 0.003167416900396347, + 0.002897436497732997, + 0.0016550050349906087, + 0.001865134690888226, + 0.0014457119395956397, + 0.00179033144377172, + 0.0019002427579835057, + 0.0016497340984642506, + 0.0013285804307088256, + 0.0016541671939194202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.2255811244249344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019527667900547385 + ], + "opposite_logprob_tokens": [ + -6.251952648162842 + ], + "logprob_reward": -0.0019527667900547385, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019527667900547385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3156881332397461, + 0.16503176093101501, + 0.11753471195697784, + 0.16405418515205383, + 0.06832222640514374, + 0.0440862737596035, + 0.026309145614504814, + 0.020819077268242836, + 0.0528193823993206, + 0.025335097685456276 + ], + "credit_map": [ + 0.01850912533700466, + 0.017896143719553947, + 0.01706048846244812, + 0.01539725624024868, + 0.08821985125541687, + 0.038420818746089935, + 0.018955286592245102, + 0.10122919082641602, + 0.013831572607159615, + 0.021192964166402817, + 0.02734735794365406, + 0.018163476139307022, + 0.035646721720695496, + 0.02128627710044384, + 0.015796612948179245, + 0.011766789481043816, + 0.005573546979576349, + 0.00636287359520793, + 0.007752705831080675, + 0.013894984498620033, + 0.04510190337896347, + 0.02677221968770027, + 0.005700050387531519, + 0.006376434583216906, + 0.01371525228023529, + 0.030285431072115898, + 0.017749998718500137, + 0.01812015101313591, + 0.03254460543394089, + 0.018140142783522606, + 0.016282979398965836, + 0.017215624451637268, + 0.004485988058149815, + 0.006609979551285505, + 0.016160672530531883, + 0.004676053766161203, + 0.004945307970046997, + 0.02186811901628971, + 0.005978965200483799, + 0.0035971435718238354, + 0.005310386419296265, + 0.0055249775759875774, + 0.0058546182699501514, + 0.004660994280129671, + 0.004936575423926115, + 0.005364344455301762, + 0.006950993090867996, + 0.0054833898320794106, + 0.002571780700236559, + 0.003671359270811081, + 0.002825029194355011, + 0.0022209021262824535, + 0.0038531434256583452, + 0.0036432708147913218, + 0.0034561611246317625, + 0.0040675015188753605, + 0.002548104850575328, + 0.002668328583240509, + 0.0025187742430716753, + 0.002579330699518323, + 0.003025852609425783, + 0.002575855003669858, + 0.002288209507241845, + 0.002614624099805951, + 0.006399644538760185, + 0.01392250694334507, + 0.004040462896227837, + 0.004016315098851919, + 0.0069379727356135845, + 0.0053819757886230946, + 0.005005246493965387, + 0.007115265354514122, + 0.0027519718278199434, + 0.003606550395488739, + 0.0019052674761041999, + 0.0026173065416514874, + 0.0049741314724087715, + 0.002951384522020817, + 0.003445585258305073, + 0.0030829021707177162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0019527667900547385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521108388900757 + ], + "opposite_logprob_tokens": [ + -1.5021108388900757 + ], + "logprob_reward": -0.2521108388900757, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521108388900757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31637006998062134, + 0.026832157745957375, + 0.3765706419944763, + 0.15091098845005035, + 0.03604485094547272, + 0.016920028254389763, + 0.02527019940316677, + 0.02211812511086464, + 0.012855266220867634, + 0.01610773243010044 + ], + "credit_map": [ + 0.01762130670249462, + 0.018362777307629585, + 0.008605116978287697, + 0.011581064201891422, + 0.04672252759337425, + 0.09068500995635986, + 0.012828163802623749, + 0.10996409505605698, + 0.0042806100100278854, + 0.002534101950004697, + 0.003606302896514535, + 0.0030054436065256596, + 0.004533506464213133, + 0.0035528307780623436, + 0.003063090145587921, + 0.0022562711965292692, + 0.02756069041788578, + 0.024022169411182404, + 0.021663878113031387, + 0.11671243607997894, + 0.1227395087480545, + 0.0294501893222332, + 0.020098671317100525, + 0.014323093928396702, + 0.0112729761749506, + 0.027790943160653114, + 0.01853467896580696, + 0.016719376668334007, + 0.020111296325922012, + 0.01403894741088152, + 0.01873544231057167, + 0.02370733767747879, + 0.002918825950473547, + 0.003820066573098302, + 0.008411657065153122, + 0.0030550614465028048, + 0.0033132622484117746, + 0.009099853225052357, + 0.003294215304777026, + 0.002131908666342497, + 0.002225554082542658, + 0.0019824584014713764, + 0.0017237007850781083, + 0.0015314636984840035, + 0.002660905010998249, + 0.0032125969883054495, + 0.0016801750753074884, + 0.0019031739793717861, + 0.002801699796691537, + 0.0029065061826258898, + 0.0026646642945706844, + 0.0025115434546023607, + 0.004249064717441797, + 0.004400849342346191, + 0.002511359518393874, + 0.0032245106995105743, + 0.002215601271018386, + 0.0032264278270304203, + 0.003015948459506035, + 0.0023128173779696226, + 0.0031159461941570044, + 0.0022840164601802826, + 0.002246295567601919, + 0.0037010747473686934, + 0.002186983125284314, + 0.002090999623760581, + 0.0011570826172828674, + 0.001567100640386343, + 0.0016078273765742779, + 0.0016931622521951795, + 0.0013230243930593133, + 0.0012290861923247576, + 0.0014284227509051561, + 0.0023832020815461874, + 0.001200423575937748, + 0.0013739541172981262, + 0.0041089775040745735, + 0.0022653283085674047, + 0.001544177532196045, + 0.0018032473744824529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.2521108388900757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, just above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the main subjects vibrant red, spiky flower heads with tubular petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by brilliant, high-contrast midday sun with sharp dark shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flowers move in a chaotic yet synchronized dance of dipping, swaying, and bouncing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809477761387825 + ], + "opposite_logprob_tokens": [ + -3.2880947589874268 + ], + "logprob_reward": -0.03809477761387825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809477761387825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12314856797456741, + 0.04559813439846039, + 0.28962433338165283, + 0.2861877977848053, + 0.08766306191682816, + 0.029663264751434326, + 0.013402801007032394, + 0.04968246817588806, + 0.032931987196207047, + 0.042097609490156174 + ], + "credit_map": [ + 0.0082019604742527, + 0.014271845109760761, + 0.009795445017516613, + 0.009493338875472546, + 0.014639287255704403, + 0.03501056879758835, + 0.010176705196499825, + 0.021559415385127068, + 0.0036886599846184254, + 0.0050431289710104465, + 0.0069852229207754135, + 0.006136101670563221, + 0.008062266744673252, + 0.007662848569452763, + 0.004352223593741655, + 0.003667679615318775, + 0.028425760567188263, + 0.018932851031422615, + 0.018831530585885048, + 0.08168396353721619, + 0.09423355758190155, + 0.02298910915851593, + 0.01497835572808981, + 0.009549212642014027, + 0.022088026627898216, + 0.043795302510261536, + 0.031314946711063385, + 0.03257722407579422, + 0.04150693863630295, + 0.025477569550275803, + 0.04069402068853378, + 0.048733774572610855, + 0.007578226272016764, + 0.010486643761396408, + 0.02009534277021885, + 0.005691241938620806, + 0.007322030141949654, + 0.025321975350379944, + 0.006458889693021774, + 0.004708714317530394, + 0.004289089236408472, + 0.002527720294892788, + 0.003826743923127651, + 0.0027441857382655144, + 0.004397439770400524, + 0.005025513470172882, + 0.0030677090398967266, + 0.003784861881285906, + 0.0012965850764885545, + 0.0014882844407111406, + 0.0018802437698468566, + 0.0016220267862081528, + 0.0016578809591010213, + 0.00246634054929018, + 0.0018225018866360188, + 0.0011689370730891824, + 0.00601805467158556, + 0.007593229878693819, + 0.007057358976453543, + 0.005676313769072294, + 0.005964974407106638, + 0.005503703840076923, + 0.006799767259508371, + 0.005069066304713488, + 0.00432822247967124, + 0.0057631866075098515, + 0.0040970672853291035, + 0.0035520638339221478, + 0.004357863217592239, + 0.003136878367513418, + 0.003716223407536745, + 0.003980481531471014, + 0.0050749508664011955, + 0.006283151917159557, + 0.0032281852327287197, + 0.003144333604723215, + 0.009924233891069889, + 0.005483029875904322, + 0.004566216841340065, + 0.004393508657813072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.03809477761387825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_tensor([2613, 2613])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_tensor([2613, 2613])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..255e550ca84fd65c3fb4806e1128ac86483ccdef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_tensor([2613, 2613])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic 360-degree panning shot, captured in a minimalist, flat illustration style, captures an N-1 starfighter as it pierces a sea of clouds from below at breakneck speed. The fighter boasts a slender, streamlined chrome fuselage composed of smooth, light gray blocks, outlined with hard, white highlights to simulate a metallic sheen. Striking bright yellow accents adorn the nose, engine tips, and wingtips. The sea of clouds it breaks through is not a natural formation, but rather a composite of geometric, layered blocks of pure white, sky blue, and dark blue. These blocks violently explode in all directions as the fighter breaks through, forming countless sharp, fragmented shapes and creating a powerful visual impact. The scene boasts highly saturated colors, with clean lines and no complex shading or gradients. Through pure shapes, high-contrast colors, and the rapid, panoramic camera movement, the starfighter perfectly captures the sheer force and impact of its sudden breakout.", + "seed": 4087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_2613_seed_4087_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05941811529919505, + "vlm_logprob": -0.05941811529919505, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.05941811529919505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008812346495687962 + ], + "opposite_logprob_tokens": [ + -4.883812427520752 + ], + "logprob_reward": -0.008812346495687962, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008812346495687962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20015382766723633, + 0.10810977965593338, + 0.2406352460384369, + 0.09377378970384598, + 0.05359456315636635, + 0.05609612911939621, + 0.05365341156721115, + 0.0626642182469368, + 0.07558175921440125, + 0.05573730543255806 + ], + "credit_map": [ + 0.025455597788095474, + 0.02325778268277645, + 0.030055036768317223, + 0.028606120496988297, + 0.015283877961337566, + 0.025024758651852608, + 0.023973390460014343, + 0.028497250750660896, + 0.018673717975616455, + 0.014599462039768696, + 0.018183793872594833, + 0.008691556751728058, + 0.009875323623418808, + 0.009985045529901981, + 0.015442798845469952, + 0.012658072635531425, + 0.02059010975062847, + 0.04273580387234688, + 0.05733479559421539, + 0.025189081206917763, + 0.023909855633974075, + 0.02026374265551567, + 0.023132817819714546, + 0.027479026466608047, + 0.025439810007810593, + 0.02505199983716011, + 0.006696173455566168, + 0.010678541846573353, + 0.006047314964234829, + 0.007451188750565052, + 0.005246622487902641, + 0.007162122987210751, + 0.0037183696404099464, + 0.008737623691558838, + 0.010347387753427029, + 0.009216115809977055, + 0.005055231507867575, + 0.00769586069509387, + 0.003997395746409893, + 0.0048265764489769936, + 0.007012979593127966, + 0.008554442785680294, + 0.006569407880306244, + 0.008168302476406097, + 0.0051847887225449085, + 0.008596610277891159, + 0.005704198498278856, + 0.006305394694209099, + 0.005765073467046022, + 0.009689914993941784, + 0.004382368177175522, + 0.006368194241076708, + 0.008625083602964878, + 0.007036855444312096, + 0.006218577269464731, + 0.00556733924895525, + 0.009930442087352276, + 0.0040171160362660885, + 0.006744828075170517, + 0.006753872614353895, + 0.007193668745458126, + 0.006888972595334053, + 0.012312420643866062, + 0.008822888135910034, + 0.009229193441569805, + 0.007814603857696056, + 0.018085893243551254, + 0.011316278949379921, + 0.006966428831219673, + 0.005376297049224377, + 0.007208081893622875, + 0.009584976360201836, + 0.005837978329509497, + 0.006981336511671543, + 0.0048842718824744225, + 0.005607277154922485, + 0.009018588811159134, + 0.009334521368145943, + 0.007395229768007994, + 0.006678095553070307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.008812346495687962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263508390635252 + ], + "opposite_logprob_tokens": [ + -3.651350736618042 + ], + "logprob_reward": -0.0263508390635252, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263508390635252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17908012866973877, + 0.08810915052890778, + 0.18648916482925415, + 0.0659828633069992, + 0.12109879404306412, + 0.03205462172627449, + 0.06312776356935501, + 0.04882657527923584, + 0.13232630491256714, + 0.08290465176105499 + ], + "credit_map": [ + 0.017775997519493103, + 0.020713746547698975, + 0.024406544864177704, + 0.027269480749964714, + 0.01579032838344574, + 0.021310683339834213, + 0.029881183058023453, + 0.02193216048181057, + 0.011496251448988914, + 0.01241159625351429, + 0.013522198423743248, + 0.012161772698163986, + 0.006721138954162598, + 0.007490528281778097, + 0.01207277737557888, + 0.0122328856959939, + 0.019361747428774834, + 0.03585533797740936, + 0.04945208132266998, + 0.016402173787355423, + 0.021658223122358322, + 0.012383366003632545, + 0.01564040035009384, + 0.015735840424895287, + 0.009230010211467743, + 0.009768777526915073, + 0.004422156605869532, + 0.009309429675340652, + 0.010186239145696163, + 0.006511616054922342, + 0.009024466387927532, + 0.007530164904892445, + 0.0060813818126916885, + 0.03628142178058624, + 0.0447673574090004, + 0.008467866107821465, + 0.006550936494022608, + 0.006076883990317583, + 0.0071509904228150845, + 0.0057219574227929115, + 0.0026117349043488503, + 0.004977257922291756, + 0.003340555354952812, + 0.004753194283694029, + 0.004100851248949766, + 0.0039055978413671255, + 0.004284126218408346, + 0.004081302788108587, + 0.00712567288428545, + 0.010699504055082798, + 0.004706846084445715, + 0.00831347145140171, + 0.009894547052681446, + 0.00956762582063675, + 0.005802645348012447, + 0.007017452735453844, + 0.005562105216085911, + 0.004941803868860006, + 0.00527285085991025, + 0.004883230663836002, + 0.008434769697487354, + 0.007379140704870224, + 0.006253277882933617, + 0.006099394988268614, + 0.011935514397919178, + 0.009851451963186264, + 0.013789372518658638, + 0.029536966234445572, + 0.017285548150539398, + 0.01817905157804489, + 0.017128368839621544, + 0.014620037749409676, + 0.006893505342304707, + 0.006461295764893293, + 0.011314082890748978, + 0.008906150236725807, + 0.023093104362487793, + 0.009775055572390556, + 0.006902911700308323, + 0.00955854170024395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0263508390635252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251422990113497 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251422990113497, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251422990113497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2752562463283539, + 0.052684955298900604, + 0.053227584809064865, + 0.03609498590230942, + 0.098707415163517, + 0.030905958265066147, + 0.2255338877439499, + 0.05756964534521103, + 0.13544157147407532, + 0.03457771614193916 + ], + "credit_map": [ + 0.0235774964094162, + 0.04022281616926193, + 0.03549788147211075, + 0.0406525619328022, + 0.02537178434431553, + 0.03587720915675163, + 0.04141443222761154, + 0.03264206275343895, + 0.007493702694773674, + 0.007201450876891613, + 0.007274414412677288, + 0.0070513649843633175, + 0.0057360948994755745, + 0.0045504323206841946, + 0.008150284178555012, + 0.005227208603173494, + 0.0058102733455598354, + 0.009238203056156635, + 0.01214279979467392, + 0.00470593711361289, + 0.0047400775365531445, + 0.0038865930400788784, + 0.00796476099640131, + 0.004738934338092804, + 0.004502110183238983, + 0.003897814778611064, + 0.005714106373488903, + 0.006364257074892521, + 0.0037921988405287266, + 0.0036018090322613716, + 0.0037175812758505344, + 0.00450510811060667, + 0.004295593127608299, + 0.03090631403028965, + 0.0323859341442585, + 0.007263998966664076, + 0.007111040875315666, + 0.007559122983366251, + 0.005428565200418234, + 0.0037568516563624144, + 0.0026634943205863237, + 0.0076301125809550285, + 0.002547159092500806, + 0.004538507666438818, + 0.0034111293498426676, + 0.003775818971917033, + 0.003172739641740918, + 0.0031669943127781153, + 0.03135848417878151, + 0.06424856185913086, + 0.01309614721685648, + 0.025579214096069336, + 0.03323886916041374, + 0.01963943988084793, + 0.021982060745358467, + 0.016391107812523842, + 0.00849731732159853, + 0.0028480510227382183, + 0.005841210018843412, + 0.004308812320232391, + 0.009220007807016373, + 0.00808671209961176, + 0.011286991648375988, + 0.007480543572455645, + 0.009703912772238255, + 0.010996349155902863, + 0.016243765130639076, + 0.039735157042741776, + 0.013095961883664131, + 0.0146018136292696, + 0.01641232706606388, + 0.014652284793555737, + 0.003760482184588909, + 0.0035333409905433655, + 0.004607676528394222, + 0.0037155000027269125, + 0.006305560003966093, + 0.005365467164665461, + 0.0034908829256892204, + 0.003798808204010129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.005251422990113497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17000176012516022, + 0.06915713101625443, + 0.08091934025287628, + 0.11206258088350296, + 0.10406941920518875, + 0.026181688532233238, + 0.11041530966758728, + 0.19013474881649017, + 0.08223912119865417, + 0.05481893569231033 + ], + "credit_map": [ + 0.01497127115726471, + 0.02590995840728283, + 0.026759881526231766, + 0.02126665785908699, + 0.0165251437574625, + 0.017571771517395973, + 0.02627016417682171, + 0.02072690986096859, + 0.009485733695328236, + 0.008411077782511711, + 0.012912053614854813, + 0.005812877789139748, + 0.005431989207863808, + 0.008457054384052753, + 0.008703921921551228, + 0.009942417033016682, + 0.010562850162386894, + 0.009698898531496525, + 0.010895279236137867, + 0.00884238164871931, + 0.01289870124310255, + 0.0056996471248567104, + 0.013080893084406853, + 0.009240682236850262, + 0.027794577181339264, + 0.0198927354067564, + 0.010841421782970428, + 0.012473098933696747, + 0.009817466139793396, + 0.00941501371562481, + 0.010156646370887756, + 0.011671619489789009, + 0.005762678571045399, + 0.027687251567840576, + 0.030000483617186546, + 0.008047383278608322, + 0.007911527529358864, + 0.010668640956282616, + 0.0076807108707726, + 0.006310736294835806, + 0.0028763071168214083, + 0.002980895573273301, + 0.003872287692502141, + 0.0038365283980965614, + 0.0034357719123363495, + 0.0025790506042540073, + 0.0032546010334044695, + 0.0033462459687143564, + 0.01491946168243885, + 0.031503669917583466, + 0.005439459346234798, + 0.011171453632414341, + 0.013459796085953712, + 0.012105301953852177, + 0.012621210888028145, + 0.00919495802372694, + 0.025657078251242638, + 0.014511431567370892, + 0.02442467398941517, + 0.02179952897131443, + 0.025643201544880867, + 0.02318205125629902, + 0.030814148485660553, + 0.024102646857500076, + 0.010030138306319714, + 0.00818283949047327, + 0.015468962490558624, + 0.013413150794804096, + 0.008710048161447048, + 0.00856503564864397, + 0.007323260419070721, + 0.010545684024691582, + 0.006129079964011908, + 0.005549245048314333, + 0.0069264317862689495, + 0.005235835909843445, + 0.014255907386541367, + 0.0068571544252336025, + 0.0038245972245931625, + 0.006040684413164854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.004633048549294472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520429193973541 + ], + "opposite_logprob_tokens": [ + -1.5020428895950317 + ], + "logprob_reward": -0.2520429193973541, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520429193973541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19864521920681, + 0.08852888643741608, + 0.16174069046974182, + 0.0648980513215065, + 0.08460699021816254, + 0.03338339924812317, + 0.06290766596794128, + 0.055167123675346375, + 0.11478917300701141, + 0.13533279299736023 + ], + "credit_map": [ + 0.013405400328338146, + 0.025671780109405518, + 0.026044229045510292, + 0.029121365398168564, + 0.015119430609047413, + 0.02320820838212967, + 0.043210405856370926, + 0.02286439575254917, + 0.013398203998804092, + 0.011109668761491776, + 0.01700160652399063, + 0.009689906612038612, + 0.008155010640621185, + 0.009309730492532253, + 0.010921524837613106, + 0.00894323829561472, + 0.013391835615038872, + 0.025945279747247696, + 0.03442561253905296, + 0.01742446981370449, + 0.019635653123259544, + 0.013287446461617947, + 0.019438065588474274, + 0.018192319199442863, + 0.008162475191056728, + 0.00827906746417284, + 0.008937052451074123, + 0.009214982390403748, + 0.009399626404047012, + 0.008320741355419159, + 0.006628162227571011, + 0.005955942906439304, + 0.00608479930087924, + 0.018787844106554985, + 0.01734974980354309, + 0.009322721511125565, + 0.007850964553654194, + 0.009866987355053425, + 0.00931612029671669, + 0.006027797237038612, + 0.003884740639477968, + 0.003550333436578512, + 0.004765783436596394, + 0.004722935147583485, + 0.00313078286126256, + 0.004349440336227417, + 0.003887009574100375, + 0.00509237265214324, + 0.008462728001177311, + 0.013869996182620525, + 0.0037086624652147293, + 0.008027547039091587, + 0.009586958214640617, + 0.006677936762571335, + 0.008314674720168114, + 0.004259158857166767, + 0.008142411708831787, + 0.0038561082910746336, + 0.004920981824398041, + 0.00414906395599246, + 0.007690190803259611, + 0.0092550627887249, + 0.010195647366344929, + 0.006957655772566795, + 0.012054436840116978, + 0.013055812567472458, + 0.016919028013944626, + 0.015550612471997738, + 0.019482579082250595, + 0.013321615755558014, + 0.0108086122199893, + 0.013596468605101109, + 0.019620636478066444, + 0.015474552288651466, + 0.011747029609978199, + 0.01106218621134758, + 0.016285937279462814, + 0.034402914345264435, + 0.012051235884428024, + 0.01468830555677414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.2520429193973541 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic 360-degree panning shot, captured in a minimalist, flat illustration style, captures an N-1 starfighter as it pierces a sea of clouds from below at breakneck speed. The fighter boasts a slender, streamlined chrome fuselage composed of smooth, light gray blocks, outlined with hard, white highlights to simulate a metallic sheen. Striking bright yellow accents adorn the nose, engine tips, and wingtips. The sea of clouds it breaks through is not a natural formation, but rather a composite of geometric, layered blocks of pure white, sky blue, and dark blue. These blocks violently explode in all directions as the fighter breaks through, forming countless sharp, fragmented shapes and creating a powerful visual impact. The scene boasts highly saturated colors, with clean lines and no complex shading or gradients. Through pure shapes, high-contrast colors, and the rapid, panoramic camera movement, the starfighter perfectly captures the sheer force and impact of its sudden breakout.", + "seed": 4088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_2613_seed_4088_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05516908490099013, + "vlm_logprob": -0.05516908490099013, + "vlm_yes_no_margin": 3.649999976158142, + "vlm_token_credit": -0.05516908490099013, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08045530319213867 + ], + "opposite_logprob_tokens": [ + -2.5804553031921387 + ], + "logprob_reward": -0.08045530319213867, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08045530319213867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17451152205467224, + 0.13655132055282593, + 0.056965626776218414, + 0.10328146815299988, + 0.07383687794208527, + 0.04419992119073868, + 0.10592648386955261, + 0.045806244015693665, + 0.22948023676872253, + 0.029440317302942276 + ], + "credit_map": [ + 0.012320626527071, + 0.01626669615507126, + 0.023156745359301567, + 0.02453433722257614, + 0.02673932909965515, + 0.014931912533938885, + 0.028563203290104866, + 0.027998674660921097, + 0.03089672327041626, + 0.019665906205773354, + 0.01727292500436306, + 0.010991448536515236, + 0.013192341662943363, + 0.018945544958114624, + 0.013612471520900726, + 0.011973956599831581, + 0.006305210292339325, + 0.007899138145148754, + 0.006799505092203617, + 0.006053925957530737, + 0.005589705426245928, + 0.007866517640650272, + 0.00994354858994484, + 0.006508071441203356, + 0.007715108338743448, + 0.010906071402132511, + 0.012695125304162502, + 0.013858302496373653, + 0.009606614708900452, + 0.011625380255281925, + 0.02523929812014103, + 0.011635567992925644, + 0.010158631950616837, + 0.0077844527550041676, + 0.010507533326745033, + 0.00965110957622528, + 0.009576322510838509, + 0.008991834707558155, + 0.00968190748244524, + 0.007485080044716597, + 0.004286181181669235, + 0.006096433382481337, + 0.006974073592573404, + 0.0074450368992984295, + 0.004141107201576233, + 0.00411413935944438, + 0.005249353125691414, + 0.005893596448004246, + 0.008506551384925842, + 0.007676148321479559, + 0.01971879042685032, + 0.011379255913197994, + 0.009394475258886814, + 0.01264428161084652, + 0.020180171355605125, + 0.016426807269454002, + 0.007003110833466053, + 0.00452977092936635, + 0.006927062291651964, + 0.0039510116912424564, + 0.0048540919087827206, + 0.008320932276546955, + 0.004598298110067844, + 0.005621967371553183, + 0.043004829436540604, + 0.009207249619066715, + 0.03789762035012245, + 0.010252031497657299, + 0.020070208236575127, + 0.0602826327085495, + 0.01533592026680708, + 0.03342975676059723, + 0.004847223870456219, + 0.0028176819905638695, + 0.004190423991531134, + 0.003902799217030406, + 0.002667252440005541, + 0.0035853837616741657, + 0.002446057042106986, + 0.004983493592590094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.08045530319213867 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307185113430023 + ], + "opposite_logprob_tokens": [ + -3.168071746826172 + ], + "logprob_reward": -0.04307185113430023, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307185113430023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10161945223808289, + 0.07672633975744247, + 0.028576672077178955, + 0.04593057557940483, + 0.05704392120242119, + 0.1552666872739792, + 0.20289653539657593, + 0.09411809593439102, + 0.19957952201366425, + 0.038242194801568985 + ], + "credit_map": [ + 0.009602752514183521, + 0.012981538660824299, + 0.013875400647521019, + 0.012390699237585068, + 0.00963075552135706, + 0.019165003672242165, + 0.00849079992622137, + 0.01548249926418066, + 0.008931344375014305, + 0.014478553086519241, + 0.013036295771598816, + 0.008113167248666286, + 0.007547815330326557, + 0.011290285736322403, + 0.007727793417870998, + 0.005601080600172281, + 0.0030411037150770426, + 0.005223499611020088, + 0.0035799015313386917, + 0.004061106592416763, + 0.0024448561016470194, + 0.003516829339787364, + 0.003970910329371691, + 0.002738462993875146, + 0.0038916978519409895, + 0.004292809870094061, + 0.004918849561363459, + 0.005854479968547821, + 0.005801297724246979, + 0.005622224882245064, + 0.010368823073804379, + 0.005180394276976585, + 0.0067407358437776566, + 0.0042799063958227634, + 0.007995128631591797, + 0.005304013378918171, + 0.010715433396399021, + 0.011493522673845291, + 0.004164352081716061, + 0.0063508301973342896, + 0.017888016998767853, + 0.02487141080200672, + 0.02082846872508526, + 0.022156961262226105, + 0.016210954636335373, + 0.01088328193873167, + 0.02135412022471428, + 0.021073484793305397, + 0.014348693192005157, + 0.005470724776387215, + 0.07504364103078842, + 0.011135437525808811, + 0.008104726672172546, + 0.013857860118150711, + 0.007900428026914597, + 0.0670350193977356, + 0.01346205547451973, + 0.00750849861651659, + 0.016790291294455528, + 0.007588627748191357, + 0.008399233222007751, + 0.017516395077109337, + 0.008938326500356197, + 0.013914665207266808, + 0.03369596600532532, + 0.013418236747384071, + 0.035630516707897186, + 0.012262807227671146, + 0.019691798835992813, + 0.04266316443681717, + 0.01591334491968155, + 0.02630368247628212, + 0.00541321188211441, + 0.00356140173971653, + 0.006007039453834295, + 0.003415450919419527, + 0.004284825176000595, + 0.005754152778536081, + 0.0053834570571780205, + 0.004422655329108238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.04307185113430023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036178866866976023 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036178866866976023, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036178866866976023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.199040949344635, + 0.16498050093650818, + 0.14944231510162354, + 0.030725792050361633, + 0.05255676433444023, + 0.07931835204362869, + 0.09075664728879929, + 0.0614083893597126, + 0.14052702486515045, + 0.031243277713656425 + ], + "credit_map": [ + 0.015843257308006287, + 0.023365890607237816, + 0.02807551622390747, + 0.026441039517521858, + 0.027967700734734535, + 0.01953156664967537, + 0.03240340203046799, + 0.02541257254779339, + 0.026105554774403572, + 0.030874621123075485, + 0.024977337568998337, + 0.011163857765495777, + 0.01589428447186947, + 0.03241090849041939, + 0.012459835037589073, + 0.01109409797936678, + 0.016791533678770065, + 0.028352811932563782, + 0.015965702012181282, + 0.025934040546417236, + 0.012241553515195847, + 0.017886484041810036, + 0.01730644330382347, + 0.014963739551603794, + 0.0028354707174003124, + 0.0027942266315221786, + 0.0039047093596309423, + 0.003567929845303297, + 0.004377266392111778, + 0.005710790399461985, + 0.0038295902777463198, + 0.0037058100569993258, + 0.0047759609296917915, + 0.0070204115472733974, + 0.007471022661775351, + 0.007531764451414347, + 0.007162924390286207, + 0.00727192685008049, + 0.00663423165678978, + 0.004688523709774017, + 0.010733403265476227, + 0.011646045371890068, + 0.010306993499398232, + 0.011238516308367252, + 0.008239097893238068, + 0.005800423678010702, + 0.009653388522565365, + 0.011700484901666641, + 0.006567030213773251, + 0.005469961557537317, + 0.02481076307594776, + 0.00782856810837984, + 0.005369187798351049, + 0.010470109060406685, + 0.007313417736440897, + 0.022927608340978622, + 0.008927643299102783, + 0.00584468012675643, + 0.008382776752114296, + 0.00551301334053278, + 0.007216277997940779, + 0.011914478614926338, + 0.005973325110971928, + 0.00763619365170598, + 0.02385324239730835, + 0.007770496420562267, + 0.028411541134119034, + 0.008109994232654572, + 0.012117602862417698, + 0.029293814674019814, + 0.009575924836099148, + 0.02139439433813095, + 0.005782066844403744, + 0.002545130904763937, + 0.0046006496995687485, + 0.0028545022942125797, + 0.0027087810449302197, + 0.0056579820811748505, + 0.003291931701824069, + 0.0038022343069314957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0036178866866976023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944785661995411 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944785661995411, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944785661995411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14744409918785095, + 0.22262009978294373, + 0.0549268014729023, + 0.08188238739967346, + 0.03957948088645935, + 0.05647733435034752, + 0.1127186045050621, + 0.12513189017772675, + 0.13102410733699799, + 0.02819524146616459 + ], + "credit_map": [ + 0.009733924642205238, + 0.014656254090368748, + 0.01512365322560072, + 0.02178841643035412, + 0.023358888924121857, + 0.01402041595429182, + 0.025313399732112885, + 0.023449141532182693, + 0.03934956341981888, + 0.029404547065496445, + 0.03764290362596512, + 0.01750703901052475, + 0.01999528706073761, + 0.038586556911468506, + 0.020269086584448814, + 0.019865117967128754, + 0.006757048424333334, + 0.004209748934954405, + 0.006897539831697941, + 0.005996914580464363, + 0.005673194769769907, + 0.007836004719138145, + 0.009592889808118343, + 0.007963461801409721, + 0.009102233685553074, + 0.01071722898632288, + 0.013318516314029694, + 0.008528562262654305, + 0.011613289825618267, + 0.012577015906572342, + 0.007448811549693346, + 0.008576726540923119, + 0.0037843007594347, + 0.0032916669733822346, + 0.0052507114596664906, + 0.003963657654821873, + 0.006299635861068964, + 0.0086497338488698, + 0.003955709282308817, + 0.0043840655125677586, + 0.00676016416400671, + 0.007076830603182316, + 0.007848970592021942, + 0.005946213845163584, + 0.0068476758897304535, + 0.004825871903449297, + 0.00858069583773613, + 0.008590913377702236, + 0.011971917003393173, + 0.009410379454493523, + 0.021496225148439407, + 0.011640598997473717, + 0.009077569469809532, + 0.02039865031838417, + 0.010368827730417252, + 0.018354428932070732, + 0.0186675563454628, + 0.0070925368927419186, + 0.027473319321870804, + 0.006671187933534384, + 0.008414368145167828, + 0.029740797355771065, + 0.00842039380222559, + 0.01865173503756523, + 0.017767447978258133, + 0.005022455472499132, + 0.03281876817345619, + 0.006866061594337225, + 0.009768853895366192, + 0.021619200706481934, + 0.008559425361454487, + 0.02860189601778984, + 0.005055083427578211, + 0.0026463891845196486, + 0.003936817403882742, + 0.0027283926028758287, + 0.0025685324799269438, + 0.0057102395221591, + 0.0023711062967777252, + 0.003178680781275034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005944785661995411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275559782981873 + ], + "opposite_logprob_tokens": [ + -2.0177555084228516 + ], + "logprob_reward": -0.14275559782981873, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14275559782981873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1211075410246849, + 0.09173644334077835, + 0.06517408043146133, + 0.07273488491773605, + 0.10402564704418182, + 0.1697693169116974, + 0.2972749173641205, + 0.03241525962948799, + 0.02944261021912098, + 0.016319269314408302 + ], + "credit_map": [ + 0.008372715674340725, + 0.015738191083073616, + 0.014471092261373997, + 0.01541067473590374, + 0.01630263589322567, + 0.013072998262941837, + 0.021651381626725197, + 0.0160878486931324, + 0.016038257628679276, + 0.016300154849886894, + 0.01331125758588314, + 0.00969876628369093, + 0.00840427540242672, + 0.012576977722346783, + 0.007817013189196587, + 0.007589738816022873, + 0.007042497396469116, + 0.005532057955861092, + 0.008738177828490734, + 0.0074851359240710735, + 0.005658960901200771, + 0.009701142087578773, + 0.01402461901307106, + 0.006991489790380001, + 0.005282843019813299, + 0.010232485830783844, + 0.008084320463240147, + 0.008601559326052666, + 0.006945919711142778, + 0.006455509923398495, + 0.01911233551800251, + 0.008019907400012016, + 0.011666705831885338, + 0.014765501022338867, + 0.012307512573897839, + 0.01417767908424139, + 0.013916346244513988, + 0.014476201497018337, + 0.012411640025675297, + 0.010304054245352745, + 0.027869783341884613, + 0.018605465069413185, + 0.023534011095762253, + 0.02055160515010357, + 0.021222945302724838, + 0.015846531838178635, + 0.020922118797898293, + 0.021216848865151405, + 0.020800232887268066, + 0.018017519265413284, + 0.08264710009098053, + 0.023125244304537773, + 0.021173587068915367, + 0.027379650622606277, + 0.02612043172121048, + 0.07801114767789841, + 0.004796974360942841, + 0.002094317926093936, + 0.004821560811251402, + 0.0027936045080423355, + 0.0030213650315999985, + 0.006954740732908249, + 0.002769551007077098, + 0.005163147579878569, + 0.004376549273729324, + 0.0021836827509105206, + 0.0047412291169166565, + 0.0025109145790338516, + 0.0030757649801671505, + 0.004825382959097624, + 0.0033994156401604414, + 0.004329672083258629, + 0.0020923574920743704, + 0.0014991643838584423, + 0.0021786633878946304, + 0.0018055401742458344, + 0.0017261331668123603, + 0.0026024526450783014, + 0.0020597034599632025, + 0.0023552540224045515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.14275559782981873 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic 360-degree panning shot, captured in a minimalist, flat illustration style, captures an N-1 starfighter as it pierces a sea of clouds from below at breakneck speed. The fighter boasts a slender, streamlined chrome fuselage composed of smooth, light gray blocks, outlined with hard, white highlights to simulate a metallic sheen. Striking bright yellow accents adorn the nose, engine tips, and wingtips. The sea of clouds it breaks through is not a natural formation, but rather a composite of geometric, layered blocks of pure white, sky blue, and dark blue. These blocks violently explode in all directions as the fighter breaks through, forming countless sharp, fragmented shapes and creating a powerful visual impact. The scene boasts highly saturated colors, with clean lines and no complex shading or gradients. Through pure shapes, high-contrast colors, and the rapid, panoramic camera movement, the starfighter perfectly captures the sheer force and impact of its sudden breakout.", + "seed": 4089, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_2613_seed_4089_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1541564249433577, + "vlm_logprob": -0.1541564249433577, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1541564249433577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.144687220454216 + ], + "opposite_logprob_tokens": [ + -2.0196871757507324 + ], + "logprob_reward": -0.144687220454216, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.144687220454216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07327812910079956, + 0.11709016561508179, + 0.09559622406959534, + 0.0515487864613533, + 0.20127259194850922, + 0.05898166820406914, + 0.05149006098508835, + 0.04002397879958153, + 0.2678897976875305, + 0.04282863810658455 + ], + "credit_map": [ + 0.009153514169156551, + 0.006428019143640995, + 0.010967081412672997, + 0.009465141221880913, + 0.007735623978078365, + 0.011398558504879475, + 0.00803348608314991, + 0.01009670365601778, + 0.007087234873324633, + 0.026357825845479965, + 0.004613090772181749, + 0.01532062329351902, + 0.04837903380393982, + 0.007618653122335672, + 0.004090528003871441, + 0.003623180091381073, + 0.018856201320886612, + 0.010654238052666187, + 0.012797806411981583, + 0.015744442120194435, + 0.008892812766134739, + 0.011671886779367924, + 0.008246885612607002, + 0.008731944486498833, + 0.005034783389419317, + 0.010517320595681667, + 0.00478529604151845, + 0.006088501773774624, + 0.007056663278490305, + 0.007371273823082447, + 0.006175803020596504, + 0.004519145004451275, + 0.010536277666687965, + 0.008923801593482494, + 0.01164204254746437, + 0.009588846936821938, + 0.07091252505779266, + 0.06740846484899521, + 0.011259530670940876, + 0.011001103557646275, + 0.008421537466347218, + 0.010465184226632118, + 0.008442394435405731, + 0.008167917840182781, + 0.0054091219790279865, + 0.006998856086283922, + 0.005116569809615612, + 0.0059600877575576305, + 0.007789941504597664, + 0.0059026507660746574, + 0.005764218512922525, + 0.004659782629460096, + 0.0043174647726118565, + 0.00812052097171545, + 0.0035365738440304995, + 0.011398905888199806, + 0.00381510891020298, + 0.004285299219191074, + 0.005278575234115124, + 0.005253758281469345, + 0.005302173085510731, + 0.006672909948974848, + 0.003972612787038088, + 0.005443541333079338, + 0.013590673916041851, + 0.009324200451374054, + 0.09985098242759705, + 0.1080530434846878, + 0.007888279855251312, + 0.006866696756333113, + 0.01168771367520094, + 0.010628199204802513, + 0.003917159512639046, + 0.004023145884275436, + 0.006502733565866947, + 0.007357378490269184, + 0.004937343765050173, + 0.005097771994769573, + 0.0065939282067120075, + 0.004399178549647331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.144687220454216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033720459789037704 + ], + "opposite_logprob_tokens": [ + -3.4087204933166504 + ], + "logprob_reward": -0.033720459789037704, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033720459789037704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10600731521844864, + 0.12198560684919357, + 0.04022669047117233, + 0.040432557463645935, + 0.2984849214553833, + 0.0282015111297369, + 0.13645783066749573, + 0.06793559342622757, + 0.11412975192070007, + 0.04613824561238289 + ], + "credit_map": [ + 0.012956454418599606, + 0.010712956078350544, + 0.014710444025695324, + 0.01488145999610424, + 0.017802679911255836, + 0.017185702919960022, + 0.009421547874808311, + 0.008336071856319904, + 0.006881451699882746, + 0.031047213822603226, + 0.006649279035627842, + 0.011457334272563457, + 0.04773136228322983, + 0.008546161465346813, + 0.005545365624129772, + 0.004127451218664646, + 0.0063957772217690945, + 0.004000652581453323, + 0.004703918006271124, + 0.002845063339918852, + 0.00412964541465044, + 0.009904088452458382, + 0.0030528453644365072, + 0.005194698926061392, + 0.0035061754751950502, + 0.0063327886164188385, + 0.0037347187753766775, + 0.006488928571343422, + 0.004464166704565287, + 0.0067026689648628235, + 0.004626181907951832, + 0.004576929844915867, + 0.015882067382335663, + 0.03037288971245289, + 0.009373540990054607, + 0.009190749377012253, + 0.10087503492832184, + 0.09795835614204407, + 0.021968351677060127, + 0.01286393590271473, + 0.003336380235850811, + 0.0032297512516379356, + 0.00370203610509634, + 0.0025778061244636774, + 0.003065415658056736, + 0.004906822927296162, + 0.0030480169225484133, + 0.0043352828361094, + 0.015715371817350388, + 0.02156645804643631, + 0.017450610175728798, + 0.01659846492111683, + 0.01235439907759428, + 0.014604133553802967, + 0.016178447753190994, + 0.021989954635500908, + 0.005609835032373667, + 0.011966608464717865, + 0.009390758350491524, + 0.011444023810327053, + 0.006221665535122156, + 0.007683470379561186, + 0.008547564968466759, + 0.007071672007441521, + 0.007328642997890711, + 0.007722033187747002, + 0.038231559097766876, + 0.03819955512881279, + 0.005926379002630711, + 0.005110948346555233, + 0.006251132115721703, + 0.005359492730349302, + 0.005096155684441328, + 0.006262837443500757, + 0.0054094260558485985, + 0.006114678457379341, + 0.008244732394814491, + 0.005554038565605879, + 0.0051727378740906715, + 0.004283639136701822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.033720459789037704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631788495928049 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631788495928049, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631788495928049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09970571100711823, + 0.11444097757339478, + 0.048272501677274704, + 0.07447772473096848, + 0.1882074624300003, + 0.11445217579603195, + 0.1794818639755249, + 0.038790758699178696, + 0.09299775213003159, + 0.04917304590344429 + ], + "credit_map": [ + 0.012525713071227074, + 0.012456455267965794, + 0.01471076998859644, + 0.008886330761015415, + 0.013170612044632435, + 0.016648340970277786, + 0.009184643626213074, + 0.012122857384383678, + 0.004865488037467003, + 0.03715845197439194, + 0.006800535134971142, + 0.007001563441008329, + 0.041475098580121994, + 0.007149404846131802, + 0.005934086628258228, + 0.004056361969560385, + 0.005053373519331217, + 0.005676375702023506, + 0.005343955010175705, + 0.007349940482527018, + 0.007247551344335079, + 0.007235965225845575, + 0.006598861422389746, + 0.0037664854899048805, + 0.007233453448861837, + 0.014471279457211494, + 0.00456652045249939, + 0.008333292789757252, + 0.010844460688531399, + 0.013307617977261543, + 0.008835908956825733, + 0.006885198876261711, + 0.008601914159953594, + 0.01197890006005764, + 0.005972930695861578, + 0.005193747114390135, + 0.06536366045475006, + 0.06186484545469284, + 0.015367230400443077, + 0.013864233158528805, + 0.013061201199889183, + 0.012388541363179684, + 0.013184195384383202, + 0.016589757055044174, + 0.009429018013179302, + 0.01729309745132923, + 0.013448907993733883, + 0.019057465717196465, + 0.025858653709292412, + 0.026836924254894257, + 0.02213629148900509, + 0.020732922479510307, + 0.014936132356524467, + 0.024606874212622643, + 0.021442309021949768, + 0.022931769490242004, + 0.00433309655636549, + 0.00517138559371233, + 0.0061179352924227715, + 0.0042218174785375595, + 0.0044440277852118015, + 0.005845608655363321, + 0.0046692658215761185, + 0.003987626172602177, + 0.004740640055388212, + 0.008458222262561321, + 0.024213619530200958, + 0.020360052585601807, + 0.010357857681810856, + 0.009528576396405697, + 0.007183876354247332, + 0.008154919371008873, + 0.006493942346423864, + 0.00804673321545124, + 0.006937421392649412, + 0.006003261543810368, + 0.006244427990168333, + 0.005716622807085514, + 0.005841644015163183, + 0.003888995386660099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.007631788495928049 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634140715003014 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634140715003014, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634140715003014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0997353047132492, + 0.11545252799987793, + 0.04193984344601631, + 0.056214869022369385, + 0.2598443329334259, + 0.022096768021583557, + 0.019457705318927765, + 0.05342940613627434, + 0.27974647283554077, + 0.05208277329802513 + ], + "credit_map": [ + 0.014445370994508266, + 0.013110809959471226, + 0.014907840639352798, + 0.01106687169522047, + 0.012319274246692657, + 0.012820441275835037, + 0.008996807038784027, + 0.01206788420677185, + 0.006252606399357319, + 0.03187749907374382, + 0.004254598170518875, + 0.011525418609380722, + 0.04734528809785843, + 0.0071927704848349094, + 0.003837241092696786, + 0.0031671011820435524, + 0.0035467869602143764, + 0.00503944605588913, + 0.004226856864988804, + 0.005567648913711309, + 0.007275416515767574, + 0.005346797872334719, + 0.007489135023206472, + 0.003447754541411996, + 0.005698664113879204, + 0.010425354354083538, + 0.004423142410814762, + 0.005748213734477758, + 0.010481203906238079, + 0.00852849893271923, + 0.00549622206017375, + 0.005413568578660488, + 0.013544845394790173, + 0.020404387265443802, + 0.008214873261749744, + 0.006322226952761412, + 0.09752363711595535, + 0.09286695718765259, + 0.011785888113081455, + 0.009181501343846321, + 0.0028798491694033146, + 0.0033080819994211197, + 0.0035572885535657406, + 0.002802804345265031, + 0.002371102338656783, + 0.002569856122136116, + 0.0028164349496364594, + 0.0017913506599143147, + 0.0026799107436090708, + 0.001678191008977592, + 0.0020010308362543583, + 0.002030023606494069, + 0.0017921252874657512, + 0.002471928484737873, + 0.0018091935198754072, + 0.004995300900191069, + 0.007621073164045811, + 0.00799130741506815, + 0.00751880556344986, + 0.005729962140321732, + 0.005343820434063673, + 0.008455177769064903, + 0.006523921154439449, + 0.004245335701853037, + 0.012741707265377045, + 0.013765920884907246, + 0.10146518796682358, + 0.0982523262500763, + 0.01609608717262745, + 0.013141912408173084, + 0.0137347262352705, + 0.0105486074462533, + 0.00717353168874979, + 0.008536497130990028, + 0.007417542859911919, + 0.006623541470617056, + 0.006989373359829187, + 0.005315769929438829, + 0.005377435591071844, + 0.004649082664400339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.008634140715003014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761085152626038 + ], + "opposite_logprob_tokens": [ + -0.8261085152626038 + ], + "logprob_reward": -0.5761085152626038, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761085152626038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050028178840875626, + 0.2597782015800476, + 0.02647584117949009, + 0.06172635406255722, + 0.11061179637908936, + 0.039619285613298416, + 0.03850390017032623, + 0.03565135970711708, + 0.29158198833465576, + 0.08602306991815567 + ], + "credit_map": [ + 0.006184885278344154, + 0.005975842010229826, + 0.006740496028214693, + 0.005990875884890556, + 0.0067937602289021015, + 0.007834973745048046, + 0.005838852841407061, + 0.004668495152145624, + 0.01195582840591669, + 0.07647059857845306, + 0.00806952640414238, + 0.026001714169979095, + 0.10857660323381424, + 0.015253204852342606, + 0.007061010226607323, + 0.006389740854501724, + 0.0037820423021912575, + 0.003167837392538786, + 0.004588383249938488, + 0.0032821563072502613, + 0.003446538932621479, + 0.0023208754137158394, + 0.0036642360500991344, + 0.0022237717639654875, + 0.0060874950140714645, + 0.012146225199103355, + 0.004505726508796215, + 0.006116349250078201, + 0.010388808324933052, + 0.009875460527837276, + 0.007189044263213873, + 0.005417250096797943, + 0.006520997732877731, + 0.01602713204920292, + 0.008822151459753513, + 0.006203803233802319, + 0.02902630716562271, + 0.027418477460741997, + 0.008704857900738716, + 0.00788807775825262, + 0.005771677475422621, + 0.006361760664731264, + 0.00523372832685709, + 0.004683857783675194, + 0.004033410456031561, + 0.0053420113399624825, + 0.0040319133549928665, + 0.004160929471254349, + 0.0035662874579429626, + 0.007168069947510958, + 0.00566874211654067, + 0.006725166458636522, + 0.0032409136183559895, + 0.0028019524179399014, + 0.006231838371604681, + 0.003100935136899352, + 0.0035166677553206682, + 0.003927199169993401, + 0.005130647215992212, + 0.004323186352849007, + 0.0039741783402860165, + 0.006750116590410471, + 0.003732413286343217, + 0.0042969542555511, + 0.014123184606432915, + 0.01220761053264141, + 0.10881166160106659, + 0.1165861263871193, + 0.0077619594521820545, + 0.00709437346085906, + 0.011814712546765804, + 0.013182360678911209, + 0.012814845889806747, + 0.01531977392733097, + 0.010664421133697033, + 0.010054362006485462, + 0.010445957072079182, + 0.009210154414176941, + 0.009041458368301392, + 0.008472099900245667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.5761085152626038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic 360-degree panning shot, captured in a minimalist, flat illustration style, captures an N-1 starfighter as it pierces a sea of clouds from below at breakneck speed. The fighter boasts a slender, streamlined chrome fuselage composed of smooth, light gray blocks, outlined with hard, white highlights to simulate a metallic sheen. Striking bright yellow accents adorn the nose, engine tips, and wingtips. The sea of clouds it breaks through is not a natural formation, but rather a composite of geometric, layered blocks of pure white, sky blue, and dark blue. These blocks violently explode in all directions as the fighter breaks through, forming countless sharp, fragmented shapes and creating a powerful visual impact. The scene boasts highly saturated colors, with clean lines and no complex shading or gradients. Through pure shapes, high-contrast colors, and the rapid, panoramic camera movement, the starfighter perfectly captures the sheer force and impact of its sudden breakout.", + "seed": 4090, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_4/index_2613_seed_4090_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17313195033930243, + "vlm_logprob": -0.17313195033930243, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.17313195033930243, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004767716862261295 + ], + "opposite_logprob_tokens": [ + -5.879767894744873 + ], + "logprob_reward": -0.004767716862261295, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004767716862261295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19913142919540405, + 0.06215851008892059, + 0.12472718954086304, + 0.05779292434453964, + 0.07244423776865005, + 0.23195165395736694, + 0.025589099153876305, + 0.038447991013526917, + 0.12239763140678406, + 0.06535941362380981 + ], + "credit_map": [ + 0.0220111645758152, + 0.022731587290763855, + 0.027851099148392677, + 0.030678318813443184, + 0.02055004984140396, + 0.027429310604929924, + 0.013699191622436047, + 0.03418070077896118, + 0.01068344246596098, + 0.008709321729838848, + 0.010351057164371014, + 0.007511841133236885, + 0.00582844065502286, + 0.00742353405803442, + 0.005737310275435448, + 0.005913558416068554, + 0.011054346337914467, + 0.012147420085966587, + 0.017958417534828186, + 0.019542492926120758, + 0.019626526162028313, + 0.01731937751173973, + 0.012758699245750904, + 0.01431988924741745, + 0.007439224049448967, + 0.00850764662027359, + 0.008124067448079586, + 0.004980713129043579, + 0.005903295241296291, + 0.007015181705355644, + 0.008095858618617058, + 0.007726930547505617, + 0.009276974014937878, + 0.00953307282179594, + 0.012814081273972988, + 0.007081171963363886, + 0.004678124096244574, + 0.008326335810124874, + 0.007046407554298639, + 0.013688058592379093, + 0.01933148130774498, + 0.012627877295017242, + 0.02050800435245037, + 0.023948460817337036, + 0.018032390624284744, + 0.05721503496170044, + 0.02366751991212368, + 0.056620869785547256, + 0.0022062938660383224, + 0.003295677015557885, + 0.004020748194307089, + 0.004229182843118906, + 0.003635125933215022, + 0.0030269816052168608, + 0.0029512043111026287, + 0.002223880495876074, + 0.003057145280763507, + 0.004428884945809841, + 0.0035467292182147503, + 0.004785179626196623, + 0.004591105505824089, + 0.005772791337221861, + 0.007394112180918455, + 0.004872041288763285, + 0.0173285324126482, + 0.011193164624273777, + 0.012646655552089214, + 0.011418829672038555, + 0.013253943994641304, + 0.023695368319749832, + 0.01669021137058735, + 0.016170915216207504, + 0.008174767717719078, + 0.008684653788805008, + 0.007850691676139832, + 0.009867774322628975, + 0.01317218691110611, + 0.007349965162575245, + 0.006593091879040003, + 0.0036662754137068987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.004767716862261295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355484500527382 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355484500527382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355484500527382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1576172113418579, + 0.0999588668346405, + 0.06155940145254135, + 0.04529719799757004, + 0.06548558175563812, + 0.2878637909889221, + 0.02948652021586895, + 0.034281037747859955, + 0.18731790781021118, + 0.031132489442825317 + ], + "credit_map": [ + 0.015224414877593517, + 0.02479003742337227, + 0.01625659316778183, + 0.0249481238424778, + 0.01061315555125475, + 0.019200609996914864, + 0.018068952485919, + 0.028515322133898735, + 0.021405884996056557, + 0.010519939474761486, + 0.017172962427139282, + 0.01004988607019186, + 0.011865769512951374, + 0.009582393802702427, + 0.0069429934956133366, + 0.012419032864272594, + 0.004127316642552614, + 0.004670127760618925, + 0.004380629863590002, + 0.010181115940213203, + 0.015664810314774513, + 0.004911857191473246, + 0.009889108128845692, + 0.007734436076134443, + 0.010478640906512737, + 0.0053649903275072575, + 0.007022919598966837, + 0.0035571104381233454, + 0.004178440198302269, + 0.0036027254536747932, + 0.005145711358636618, + 0.005946662276983261, + 0.007445550989359617, + 0.010295507498085499, + 0.014009756036102772, + 0.004588138312101364, + 0.0030746497213840485, + 0.00705557968467474, + 0.007612643297761679, + 0.011403755284845829, + 0.023839348927140236, + 0.019554773345589638, + 0.02580072544515133, + 0.04206627979874611, + 0.029845528304576874, + 0.061606377363204956, + 0.022136172279715538, + 0.06301459670066833, + 0.0024455601815134287, + 0.003467926289886236, + 0.0031706015579402447, + 0.004981266334652901, + 0.0037930170074105263, + 0.003350616665557027, + 0.004883121699094772, + 0.0033944109454751015, + 0.003597419708967209, + 0.004149061627686024, + 0.005802621133625507, + 0.005157897248864174, + 0.004333796910941601, + 0.003635859815403819, + 0.003960007801651955, + 0.0036443741992115974, + 0.027265336364507675, + 0.019080741330981255, + 0.018477199599146843, + 0.019547605887055397, + 0.024686794728040695, + 0.038994692265987396, + 0.018798869103193283, + 0.020466666668653488, + 0.003971821628510952, + 0.004421166144311428, + 0.004458657931536436, + 0.0029287245124578476, + 0.003957946784794331, + 0.003966291435062885, + 0.004113894421607256, + 0.003313989145681262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.026355484500527382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031932813581079245 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031932813581079245, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031932813581079245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1553051620721817, + 0.15401940047740936, + 0.04211646318435669, + 0.05131637305021286, + 0.0639125183224678, + 0.298267662525177, + 0.023541687056422234, + 0.023165229707956314, + 0.1560528725385666, + 0.03230265900492668 + ], + "credit_map": [ + 0.014309043064713478, + 0.024624988436698914, + 0.019037874415516853, + 0.02496735192835331, + 0.0119777237996459, + 0.019160805270075798, + 0.01310122013092041, + 0.028126155957579613, + 0.0249618049710989, + 0.018854763358831406, + 0.017609471455216408, + 0.023802347481250763, + 0.012847788631916046, + 0.01751832291483879, + 0.01651797629892826, + 0.02190692350268364, + 0.0036082479637116194, + 0.0038056576158851385, + 0.004155643749982119, + 0.005338230170309544, + 0.011729327961802483, + 0.0037264693528413773, + 0.004538341891020536, + 0.005214544478803873, + 0.005253985989838839, + 0.004436926916241646, + 0.004635897930711508, + 0.006501613650470972, + 0.007634617388248444, + 0.008247528225183487, + 0.007196949794888496, + 0.0074088540859520435, + 0.00710745994001627, + 0.010005542077124119, + 0.012242438271641731, + 0.0062410649843513966, + 0.003946255426853895, + 0.006796521134674549, + 0.0060561299324035645, + 0.011517110280692577, + 0.023914512246847153, + 0.01882106624543667, + 0.025485636666417122, + 0.03828861936926842, + 0.0294230617582798, + 0.06988460570573807, + 0.024067871272563934, + 0.06838228553533554, + 0.002469229744747281, + 0.0027844798751175404, + 0.003238906618207693, + 0.0030251997523009777, + 0.002740806434303522, + 0.0032566345762461424, + 0.0028984101954847574, + 0.0031280198600143194, + 0.0022291787900030613, + 0.0036886283196508884, + 0.002142776269465685, + 0.0025883330963552, + 0.002454233355820179, + 0.0024943859316408634, + 0.0023303816560655832, + 0.005237312521785498, + 0.01719043217599392, + 0.01366441696882248, + 0.013421515002846718, + 0.016290603205561638, + 0.019670257344841957, + 0.037992533296346664, + 0.018254393711686134, + 0.019568707793951035, + 0.004021530505269766, + 0.004775505047291517, + 0.005100043956190348, + 0.0033139411825686693, + 0.004465778358280659, + 0.004106012638658285, + 0.003814395982772112, + 0.0027054534293711185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0031932813581079245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247628316283226 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.005247628316283226, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247628316283226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12955321371555328, + 0.07064472883939743, + 0.09231818467378616, + 0.05851782113313675, + 0.04601135849952698, + 0.3691452145576477, + 0.0376056544482708, + 0.0283522866666317, + 0.09621823579072952, + 0.07163333147764206 + ], + "credit_map": [ + 0.01725889928638935, + 0.020242860540747643, + 0.01763191446661949, + 0.014152693562209606, + 0.008790932595729828, + 0.019416863098740578, + 0.010967747308313847, + 0.02109130471944809, + 0.016216741874814034, + 0.00920809619128704, + 0.011274714954197407, + 0.008362195454537868, + 0.007392917759716511, + 0.006280957721173763, + 0.005444343667477369, + 0.006464758422225714, + 0.008917662315070629, + 0.005693142767995596, + 0.012465311214327812, + 0.015235435217618942, + 0.01142835058271885, + 0.009029076434671879, + 0.017703546211123466, + 0.011845658533275127, + 0.006763987708836794, + 0.009084414690732956, + 0.007524404209107161, + 0.007358614355325699, + 0.0047361841425299644, + 0.006743902340531349, + 0.007730995304882526, + 0.0085753183811903, + 0.0070235226303339005, + 0.004030250944197178, + 0.00523545453324914, + 0.007618478965014219, + 0.004451422952115536, + 0.006445297040045261, + 0.005135028623044491, + 0.00607190141454339, + 0.031962525099515915, + 0.021197061985731125, + 0.033172670751810074, + 0.03939247876405716, + 0.029109027236700058, + 0.08987285196781158, + 0.03221466392278671, + 0.09222393482923508, + 0.0035819460172206163, + 0.004928187932819128, + 0.004446917679160833, + 0.004625476431101561, + 0.004087172448635101, + 0.005603292025625706, + 0.005053633823990822, + 0.005279030650854111, + 0.0023883595131337643, + 0.0028745941817760468, + 0.004063841886818409, + 0.004300669301301241, + 0.0036554064135998487, + 0.004429608583450317, + 0.0030733076855540276, + 0.0035665014293044806, + 0.012382333166897297, + 0.00909394957125187, + 0.009360948577523232, + 0.010391494259238243, + 0.011850716546177864, + 0.020891187712550163, + 0.010212255641818047, + 0.01203534472733736, + 0.0071196951903402805, + 0.009355025365948677, + 0.009277258068323135, + 0.009433797560632229, + 0.011370540596544743, + 0.008155211806297302, + 0.010381544008851051, + 0.006540253758430481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.005247628316283226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist flat illustration with clean lines and no complex shading or gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an N-1 starfighter with a slender, streamlined chrome fuselage made of smooth light gray blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the starfighter feature bright yellow accents on the nose, engine tips, and wingtips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sea of clouds composed of geometric layered blocks in pure white, sky blue, and dark blue that explode into sharp fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a dynamic 360-degree panning shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8260956406593323 + ], + "opposite_logprob_tokens": [ + -0.5760956406593323 + ], + "logprob_reward": -0.8260956406593323, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8260956406593323, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15381017327308655, + 0.11272616684436798, + 0.09834510087966919, + 0.048941705375909805, + 0.05679270997643471, + 0.07714460045099258, + 0.03120451793074608, + 0.02849552407860756, + 0.373906672000885, + 0.01863282360136509 + ], + "credit_map": [ + 0.017825480550527573, + 0.028484784066677094, + 0.02176262065768242, + 0.01839805394411087, + 0.012957638129591942, + 0.022217098623514175, + 0.010179613716900349, + 0.021984882652759552, + 0.022953784093260765, + 0.012313496321439743, + 0.022611889988183975, + 0.010768369771540165, + 0.012296091765165329, + 0.010207940824329853, + 0.007464524358510971, + 0.014110064134001732, + 0.011704438365995884, + 0.00838115718215704, + 0.012088957242667675, + 0.011900787241756916, + 0.014552869834005833, + 0.010146410204470158, + 0.017711671069264412, + 0.011858806945383549, + 0.006919125095009804, + 0.005883068777620792, + 0.005055811256170273, + 0.004545061849057674, + 0.008520593866705894, + 0.005860273260623217, + 0.005538986995816231, + 0.006618785671889782, + 0.005883437115699053, + 0.006627758499234915, + 0.00898914597928524, + 0.008780485019087791, + 0.006127355620265007, + 0.0056469435803592205, + 0.0057759592309594154, + 0.008961623534560204, + 0.007652709260582924, + 0.004680714104324579, + 0.007403580471873283, + 0.00821998156607151, + 0.007832963019609451, + 0.0174527820199728, + 0.00557945528998971, + 0.01832241378724575, + 0.003147477749735117, + 0.0037563492078334093, + 0.003628745209425688, + 0.005361644551157951, + 0.005444572307169437, + 0.00403855973854661, + 0.002896589459851384, + 0.0029305799398571253, + 0.0022631429601460695, + 0.0030158793088048697, + 0.0029957706574350595, + 0.0032018902711570263, + 0.003488333895802498, + 0.005335183348506689, + 0.004662893712520599, + 0.0035324289929121733, + 0.03640613332390785, + 0.03434762731194496, + 0.03713542968034744, + 0.042120836675167084, + 0.04911302030086517, + 0.08201014250516891, + 0.04956812411546707, + 0.04320535808801651, + 0.002137713134288788, + 0.0020030755549669266, + 0.0016284303274005651, + 0.0031570340506732464, + 0.003490528091788292, + 0.0021114845294505358, + 0.0019287563627585769, + 0.002175800735130906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.8260956406593323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_tensor([5553, 5553])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_tensor([5553, 5553])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee72caf7b0b00b8c8ebe64a856be30a7f0c6ac72 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_tensor([5553, 5553])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot energetically follows a golden retriever as it runs with boundless enthusiasm down the steep slope of a massive sand dune. The dog's long golden fur flutters in the wind, its ears flap happily, and its pink tongue sticks out from the corner of its mouth in a blissful smile. Its entire body is a picture centered on movement, its powerful leg muscles propelling it forward, its paws digging into the fine, pale sand, and its every step sending up small clouds of dust. The camera remains low, emphasizing the dog's speed and determination as it chases a bright yellow tennis ball, which tumbles and bounces precariously just ahead of it. The entire scene is bathed in the warm golden light of the afternoon sun, casting long, soft shadows and making every grain of sand shimmer. In the background, the deep blue of the ocean meets the hazy light blue of the sky on the horizon, creating an expansive atmosphere of pure, carefree tranquility.", + "seed": 20814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_5553_seed_20814_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1155347143765539, + "vlm_logprob": -0.1155347143765539, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.1155347143765539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5291491150856018 + ], + "opposite_logprob_tokens": [ + -0.9041491150856018 + ], + "logprob_reward": -0.5291491150856018, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5291491150856018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12446419894695282, + 0.04438814893364906, + 0.04313259944319725, + 0.06577238440513611, + 0.07187174260616302, + 0.1856764853000641, + 0.04233899340033531, + 0.08958540111780167, + 0.1782103180885315, + 0.15455979108810425 + ], + "credit_map": [ + 0.015104283578693867, + 0.01156670693308115, + 0.009210114367306232, + 0.014956788159906864, + 0.01575564034283161, + 0.01940627582371235, + 0.020367609336972237, + 0.01809678226709366, + 0.00545377004891634, + 0.005362874362617731, + 0.003880258882418275, + 0.004385849926620722, + 0.007476257625967264, + 0.005546450149267912, + 0.005872153677046299, + 0.006410535424947739, + 0.0055880420841276646, + 0.007491074036806822, + 0.003135673003271222, + 0.0042753685265779495, + 0.006202880758792162, + 0.00744717987254262, + 0.004842192400246859, + 0.00415018992498517, + 0.008952723816037178, + 0.008989540860056877, + 0.0062964665703475475, + 0.007566629443317652, + 0.008339572697877884, + 0.0054264673963189125, + 0.013851932249963284, + 0.006349049508571625, + 0.01351351197808981, + 0.006446852348744869, + 0.005002234596759081, + 0.007825388573110104, + 0.012860521674156189, + 0.011370126157999039, + 0.010422173887491226, + 0.00443093990907073, + 0.02327726036310196, + 0.017806081101298332, + 0.01401987299323082, + 0.014262293465435505, + 0.04010738059878349, + 0.03833292797207832, + 0.016974441707134247, + 0.020896224305033684, + 0.005834674462676048, + 0.005935432855039835, + 0.003079958027228713, + 0.005514880642294884, + 0.006399061065167189, + 0.004270145203918219, + 0.0076173581182956696, + 0.0036874832585453987, + 0.025211235508322716, + 0.010161085985600948, + 0.005757824517786503, + 0.00552199874073267, + 0.006090953480452299, + 0.007265238557010889, + 0.007061086595058441, + 0.022515974938869476, + 0.0382491797208786, + 0.022045675665140152, + 0.010800659656524658, + 0.016132310032844543, + 0.040332019329071045, + 0.0111311674118042, + 0.02908152900636196, + 0.010437770746648312, + 0.035126011818647385, + 0.015046607702970505, + 0.009241662919521332, + 0.012329533696174622, + 0.04860563576221466, + 0.014084771275520325, + 0.010910526849329472, + 0.009215043857693672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.5291491150856018 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813551366329193 + ], + "opposite_logprob_tokens": [ + -3.288135528564453 + ], + "logprob_reward": -0.03813551366329193, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813551366329193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16476857662200928, + 0.03636816889047623, + 0.02305898256599903, + 0.05164307728409767, + 0.10375463962554932, + 0.15574537217617035, + 0.029057780280709267, + 0.03776407241821289, + 0.23825585842132568, + 0.15958352386951447 + ], + "credit_map": [ + 0.024486688897013664, + 0.01922130212187767, + 0.010133720934391022, + 0.015591411851346493, + 0.036391016095876694, + 0.014390652067959309, + 0.028496194630861282, + 0.016057584434747696, + 0.004515511449426413, + 0.005399928893893957, + 0.0037140336353331804, + 0.004092979710549116, + 0.005863005295395851, + 0.0029491654131561518, + 0.006331829819828272, + 0.0035017170011997223, + 0.0024933123495429754, + 0.0027412152849137783, + 0.0026716399006545544, + 0.0028319398406893015, + 0.003608263796195388, + 0.002498829970136285, + 0.003797240788117051, + 0.0024165387731045485, + 0.007724287919700146, + 0.008302574045956135, + 0.004133088048547506, + 0.005962924566119909, + 0.006985753774642944, + 0.00497450539842248, + 0.008768066763877869, + 0.00479187211021781, + 0.01891431026160717, + 0.011241991072893143, + 0.006776758469641209, + 0.016629712656140327, + 0.012993015348911285, + 0.010127834975719452, + 0.0206146202981472, + 0.0064563993364572525, + 0.019573384895920753, + 0.013608976267278194, + 0.01429368369281292, + 0.012077564373612404, + 0.027959689497947693, + 0.02951970137655735, + 0.01568741910159588, + 0.023024965077638626, + 0.0037292405031621456, + 0.004120444413274527, + 0.001948905410245061, + 0.00390465441159904, + 0.003608603263273835, + 0.002629968337714672, + 0.00538260955363512, + 0.0037333534564822912, + 0.005074542947113514, + 0.0039608050137758255, + 0.005872249603271484, + 0.004111968446522951, + 0.003705422393977642, + 0.003911874257028103, + 0.0059002176858484745, + 0.0052269925363361835, + 0.04130559787154198, + 0.028228536248207092, + 0.018579760566353798, + 0.023413125425577164, + 0.04537659510970116, + 0.018145108595490456, + 0.044034283608198166, + 0.019172850996255875, + 0.04398149624466896, + 0.01142099779099226, + 0.01291743665933609, + 0.011252136901021004, + 0.04845810681581497, + 0.009171543642878532, + 0.012341685593128204, + 0.010040131397545338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03813551366329193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101555794477463 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101555794477463, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101555794477463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14671443402767181, + 0.02745579183101654, + 0.05439768359065056, + 0.13355113565921783, + 0.117977075278759, + 0.1711307317018509, + 0.06220535561442375, + 0.14804702997207642, + 0.07148424535989761, + 0.06703649461269379 + ], + "credit_map": [ + 0.018243800848722458, + 0.01656591147184372, + 0.015938565135002136, + 0.013889255933463573, + 0.027362247928977013, + 0.0146309994161129, + 0.023080065846443176, + 0.017003599554300308, + 0.003639183472841978, + 0.003232678398489952, + 0.0025007780641317368, + 0.0033238192554563284, + 0.0031490058172494173, + 0.0029501765966415405, + 0.0044057066552340984, + 0.004254447296261787, + 0.007434326224029064, + 0.005443177185952663, + 0.004309814423322678, + 0.007847756147384644, + 0.009303429163992405, + 0.006451491266489029, + 0.008545073680579662, + 0.005062619224190712, + 0.015481415204703808, + 0.017222048714756966, + 0.01473710685968399, + 0.01577805168926716, + 0.015840942040085793, + 0.013138017617166042, + 0.02532346360385418, + 0.016030099242925644, + 0.02576248161494732, + 0.005885343533009291, + 0.00682988902553916, + 0.016183050349354744, + 0.007992704398930073, + 0.007529116701334715, + 0.03917795047163963, + 0.008616553619503975, + 0.021455684676766396, + 0.01575889065861702, + 0.011628618463873863, + 0.010828133672475815, + 0.039465613663196564, + 0.03941347450017929, + 0.017905831336975098, + 0.014674508012831211, + 0.00571563420817256, + 0.008779716677963734, + 0.00561289768666029, + 0.010575884953141212, + 0.0060384757816791534, + 0.00446560001000762, + 0.01515022199600935, + 0.005866930354386568, + 0.0450822152197361, + 0.006793263368308544, + 0.009288355708122253, + 0.009819784201681614, + 0.008035048842430115, + 0.010901525616645813, + 0.010887828655540943, + 0.04723901301622391, + 0.01497742161154747, + 0.015117179602384567, + 0.004348021931946278, + 0.007375491317361593, + 0.013974770903587341, + 0.005311548709869385, + 0.006241644732654095, + 0.0041381740011274815, + 0.012698144651949406, + 0.008208232931792736, + 0.004043367691338062, + 0.006629221141338348, + 0.019577855244278908, + 0.005200913641601801, + 0.006973119918256998, + 0.0037056398577988148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.004101555794477463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021946171764284372 + ], + "opposite_logprob_tokens": [ + -6.127194404602051 + ], + "logprob_reward": -0.0021946171764284372, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021946171764284372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12035984545946121, + 0.12414972484111786, + 0.1478450894355774, + 0.05257636681199074, + 0.15392456948757172, + 0.046405211091041565, + 0.04333451762795448, + 0.047841113060712814, + 0.16084615886211395, + 0.1027173399925232 + ], + "credit_map": [ + 0.012982058338820934, + 0.010224837809801102, + 0.009899851866066456, + 0.014032719656825066, + 0.016438007354736328, + 0.01460681576281786, + 0.01960076577961445, + 0.02257479913532734, + 0.014193790964782238, + 0.019401703029870987, + 0.012331278063356876, + 0.010557400062680244, + 0.012235109694302082, + 0.028567081317305565, + 0.011487770825624466, + 0.01537559274584055, + 0.017162134870886803, + 0.02089766599237919, + 0.013451196253299713, + 0.015304086729884148, + 0.019752364605665207, + 0.030211804434657097, + 0.017883511260151863, + 0.013182342983782291, + 0.006241708993911743, + 0.007953881286084652, + 0.004669876303523779, + 0.006359781604260206, + 0.006980523932725191, + 0.005506670102477074, + 0.010067746974527836, + 0.004796181805431843, + 0.017664290964603424, + 0.010096701793372631, + 0.011179194785654545, + 0.010294703766703606, + 0.031906574964523315, + 0.04070325568318367, + 0.02141674980521202, + 0.010663116350769997, + 0.006295736879110336, + 0.004266063682734966, + 0.00443778932094574, + 0.005512109957635403, + 0.005662263371050358, + 0.006721205078065395, + 0.005962706170976162, + 0.007547338027507067, + 0.005834911018610001, + 0.005473753437399864, + 0.0039485241286456585, + 0.004199149552732706, + 0.00798088125884533, + 0.006364563945680857, + 0.006007954478263855, + 0.0035247867926955223, + 0.009545166045427322, + 0.0036357741337269545, + 0.007421219255775213, + 0.004662846680730581, + 0.0038754353299736977, + 0.004363539628684521, + 0.0058994353748857975, + 0.00843769870698452, + 0.0408160574734211, + 0.016121089458465576, + 0.008205314166843891, + 0.011648150160908699, + 0.042254526168107986, + 0.007860247045755386, + 0.025199705734848976, + 0.008741072379052639, + 0.01465463638305664, + 0.012920538894832134, + 0.012188434600830078, + 0.01137601025402546, + 0.016911722719669342, + 0.009121616370975971, + 0.012787749990820885, + 0.012756634503602982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.0021946171764284372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092770162969828 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092770162969828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092770162969828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1072472557425499, + 0.03524375706911087, + 0.019278893247246742, + 0.03494587913155556, + 0.05106320604681969, + 0.0382336787879467, + 0.02352100796997547, + 0.10236458480358124, + 0.06944632530212402, + 0.51865553855896 + ], + "credit_map": [ + 0.013925326056778431, + 0.008691484108567238, + 0.008041665889322758, + 0.01369596179574728, + 0.012129228562116623, + 0.014130416326224804, + 0.016904255375266075, + 0.019728904590010643, + 0.004997174721211195, + 0.004768146201968193, + 0.0035326445940881968, + 0.004575512371957302, + 0.00461542746052146, + 0.0040074787102639675, + 0.004941168706864119, + 0.003806200809776783, + 0.003116954118013382, + 0.0030715903267264366, + 0.002010427648201585, + 0.0015848391922190785, + 0.0024225476663559675, + 0.0025978849735111, + 0.0019680983386933804, + 0.0025065478403121233, + 0.0042532458901405334, + 0.004863161593675613, + 0.0030765447299927473, + 0.004010744858533144, + 0.004372332710772753, + 0.003162022214382887, + 0.008235905319452286, + 0.0029719159938395023, + 0.009641807526350021, + 0.0029356968589127064, + 0.0030585243366658688, + 0.006825266871601343, + 0.004207966849207878, + 0.004194387700408697, + 0.015955451875925064, + 0.0042440989054739475, + 0.00506183784455061, + 0.003906071186065674, + 0.0030276530887931585, + 0.0032932821195572615, + 0.00846915878355503, + 0.007189732510596514, + 0.004030819516628981, + 0.0032551202457398176, + 0.004045289009809494, + 0.0024868196342140436, + 0.0020645621698349714, + 0.0023146322928369045, + 0.0042420197278261185, + 0.002925178036093712, + 0.0034770381171256304, + 0.001965465024113655, + 0.032140087336301804, + 0.0060937535017728806, + 0.005662971641868353, + 0.0060203042812645435, + 0.0062585375271737576, + 0.007470684591680765, + 0.007146554067730904, + 0.031571682542562485, + 0.017550475895404816, + 0.010342758148908615, + 0.0038703237660229206, + 0.005569676868617535, + 0.017027949914336205, + 0.003742329077795148, + 0.007552660536020994, + 0.003790144808590412, + 0.14040185511112213, + 0.02527526579797268, + 0.040701042860746384, + 0.032863445580005646, + 0.1731327474117279, + 0.03284354880452156, + 0.04206264019012451, + 0.03137490153312683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.004092770162969828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot energetically follows a golden retriever as it runs with boundless enthusiasm down the steep slope of a massive sand dune. The dog's long golden fur flutters in the wind, its ears flap happily, and its pink tongue sticks out from the corner of its mouth in a blissful smile. Its entire body is a picture centered on movement, its powerful leg muscles propelling it forward, its paws digging into the fine, pale sand, and its every step sending up small clouds of dust. The camera remains low, emphasizing the dog's speed and determination as it chases a bright yellow tennis ball, which tumbles and bounces precariously just ahead of it. The entire scene is bathed in the warm golden light of the afternoon sun, casting long, soft shadows and making every grain of sand shimmer. In the background, the deep blue of the ocean meets the hazy light blue of the sky on the horizon, creating an expansive atmosphere of pure, carefree tranquility.", + "seed": 20815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_5553_seed_20815_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018854211550205946, + "vlm_logprob": -0.018854211550205946, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.018854211550205946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05145322531461716 + ], + "opposite_logprob_tokens": [ + -3.051453113555908 + ], + "logprob_reward": -0.05145322531461716, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05145322531461716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10313623398542404, + 0.09268876165151596, + 0.041717734187841415, + 0.46761593222618103, + 0.078083336353302, + 0.05434606596827507, + 0.02178526110947132, + 0.05773096904158592, + 0.031125202775001526, + 0.05177044868469238 + ], + "credit_map": [ + 0.030450548976659775, + 0.03292056545615196, + 0.006880538072437048, + 0.005798311438411474, + 0.006910071242600679, + 0.006878085434436798, + 0.00872389692813158, + 0.004574230406433344, + 0.03287854418158531, + 0.03090643137693405, + 0.007373723201453686, + 0.004024098161607981, + 0.003787966910749674, + 0.005080947652459145, + 0.0035457895137369633, + 0.005091275554150343, + 0.008126308210194111, + 0.007166013121604919, + 0.004389300476759672, + 0.008213027380406857, + 0.0030443540308624506, + 0.0032140782568603754, + 0.004022762179374695, + 0.003541897516697645, + 0.19753900170326233, + 0.19977343082427979, + 0.008814392611384392, + 0.007024105172604322, + 0.012742241844534874, + 0.025669528171420097, + 0.011871501803398132, + 0.004181769676506519, + 0.008600134402513504, + 0.010754132643342018, + 0.0051994952373206615, + 0.007178852800279856, + 0.015878016129136086, + 0.007236496079713106, + 0.015954485163092613, + 0.0072817327454686165, + 0.016301296651363373, + 0.016295239329338074, + 0.0023563809227198362, + 0.002367747016251087, + 0.006774296052753925, + 0.003127689938992262, + 0.0044751414097845554, + 0.0026482800021767616, + 0.007312469650059938, + 0.0074109104461967945, + 0.00121870799921453, + 0.0012731876922771335, + 0.0012529438827186823, + 0.0009327038424089551, + 0.0013862946070730686, + 0.0009980449685826898, + 0.01894480735063553, + 0.01927526667714119, + 0.0033051716163754463, + 0.003502566833049059, + 0.004180296324193478, + 0.002285499358549714, + 0.004155202768743038, + 0.002082165563479066, + 0.0079108364880085, + 0.006939734797924757, + 0.0014066157164052129, + 0.0032571854535490274, + 0.0029379501938819885, + 0.0027188004460185766, + 0.0033909145276993513, + 0.0025631689932197332, + 0.00859361607581377, + 0.0072648534551262856, + 0.004018454812467098, + 0.0044817146845161915, + 0.005572678055614233, + 0.005529811140149832, + 0.006547839846462011, + 0.009761489927768707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.05145322531461716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331031858921051 + ], + "opposite_logprob_tokens": [ + -3.773310422897339 + ], + "logprob_reward": -0.02331031858921051, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331031858921051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03155674412846565, + 0.028821540996432304, + 0.013871492817997932, + 0.7014248371124268, + 0.03665207326412201, + 0.09393544495105743, + 0.023218708112835884, + 0.01324065774679184, + 0.02656887285411358, + 0.030709628015756607 + ], + "credit_map": [ + 0.006038124207407236, + 0.0035845856182277203, + 0.0072147436439991, + 0.0032304178457707167, + 0.0026593187358230352, + 0.0028228124137967825, + 0.003784275846555829, + 0.0022224686108529568, + 0.006642259657382965, + 0.008376525714993477, + 0.002228605095297098, + 0.0017480506794527173, + 0.002677246229723096, + 0.0020645211916416883, + 0.0029047098942101, + 0.002179626375436783, + 0.002456624759361148, + 0.0027954119723290205, + 0.0008010851452127099, + 0.0013904750812798738, + 0.0018580022733658552, + 0.0011713390704244375, + 0.0024205194786190987, + 0.0009780362015590072, + 0.29764270782470703, + 0.30114567279815674, + 0.010048172436654568, + 0.0104650454595685, + 0.019948551431298256, + 0.03878466412425041, + 0.017129698768258095, + 0.006260366179049015, + 0.004592372104525566, + 0.005606841761618853, + 0.0036379522643983364, + 0.0038541334215551615, + 0.0068530552089214325, + 0.003355153603479266, + 0.005916517693549395, + 0.002836050232872367, + 0.0400976724922657, + 0.040101058781147, + 0.001715460792183876, + 0.0014684931375086308, + 0.0037645120173692703, + 0.003111125435680151, + 0.002228346187621355, + 0.0014487913576886058, + 0.008167505264282227, + 0.008407567627727985, + 0.0012676542392000556, + 0.0013540638610720634, + 0.0010666291927918792, + 0.0011535707162693143, + 0.000997608876787126, + 0.000804110721219331, + 0.0024592687841504812, + 0.0023530679754912853, + 0.0014081781264394522, + 0.0014418720966205, + 0.0017924741841852665, + 0.001454056822694838, + 0.0012110881507396698, + 0.0011206527706235647, + 0.007681302726268768, + 0.0073125422932207584, + 0.0012467551277950406, + 0.001445234869606793, + 0.0020213481038808823, + 0.0022898081224411726, + 0.002125111175701022, + 0.0024467725306749344, + 0.00469435378909111, + 0.0054112267680466175, + 0.004030154552310705, + 0.0027744388207793236, + 0.003291986184194684, + 0.002918574959039688, + 0.0033936204854398966, + 0.004195274319499731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.02331031858921051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078868061304092 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078868061304092, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078868061304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08159682899713516, + 0.06406720727682114, + 0.030794493854045868, + 0.24838976562023163, + 0.022515537217259407, + 0.4212867021560669, + 0.025237003341317177, + 0.0669848844408989, + 0.011067220009863377, + 0.028060371056199074 + ], + "credit_map": [ + 0.012328237295150757, + 0.01662597991526127, + 0.014692086726427078, + 0.006456643342971802, + 0.007271887734532356, + 0.005063280928879976, + 0.013020414859056473, + 0.006138294469565153, + 0.019550886005163193, + 0.017496537417173386, + 0.00515417056158185, + 0.0033058898989111185, + 0.004652320872992277, + 0.003832936519756913, + 0.006154893431812525, + 0.003919569309800863, + 0.0029345876537263393, + 0.004249121528118849, + 0.0039509618654847145, + 0.006046214606612921, + 0.003487537382170558, + 0.0027370485477149487, + 0.004590977448970079, + 0.0027980466838926077, + 0.09325253963470459, + 0.09635161608457565, + 0.004887024872004986, + 0.00632407097145915, + 0.01483597420156002, + 0.013487527146935463, + 0.012262692674994469, + 0.006988320033997297, + 0.0032474917825311422, + 0.003978163935244083, + 0.0014840713702142239, + 0.0017980176489800215, + 0.003316479967907071, + 0.0025941834319382906, + 0.004596739076077938, + 0.0015003887237980962, + 0.18178319931030273, + 0.1811448335647583, + 0.0086906049400568, + 0.005874543450772762, + 0.01566423662006855, + 0.013300486840307713, + 0.008456232026219368, + 0.006372583098709583, + 0.006909248419106007, + 0.006837684661149979, + 0.0014109520707279444, + 0.0019712094217538834, + 0.0024796673096716404, + 0.002328419592231512, + 0.0021283056121319532, + 0.0011715171858668327, + 0.025555899366736412, + 0.025407938286662102, + 0.001778970006853342, + 0.0018364118877798319, + 0.00453262310475111, + 0.0017273903358727694, + 0.0039926874451339245, + 0.0021529649384319782, + 0.002123734913766384, + 0.002002251334488392, + 0.0006033288082107902, + 0.0008645369671285152, + 0.0018526872154325247, + 0.0016021569026634097, + 0.0011321228230372071, + 0.0008864006958901882, + 0.00449830899015069, + 0.004848584067076445, + 0.0033238830510526896, + 0.0024702264927327633, + 0.0030984298791736364, + 0.002516809618100524, + 0.0032913926988840103, + 0.0040127369575202465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.011078868061304092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002484808210283518 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.002484808210283518, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002484808210283518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448608934879303, + 0.11597062647342682, + 0.008639921434223652, + 0.6342409253120422, + 0.009071001783013344, + 0.05621035769581795, + 0.0030980040319263935, + 0.002897916128858924, + 0.007741620764136314, + 0.017268668860197067 + ], + "credit_map": [ + 0.04807808995246887, + 0.05431260168552399, + 0.013255059719085693, + 0.004692432004958391, + 0.0067527699284255505, + 0.007486053742468357, + 0.006614221725612879, + 0.003669671481475234, + 0.04750131815671921, + 0.04827532172203064, + 0.003480415791273117, + 0.0026422510854899883, + 0.003586667124181986, + 0.0066029466688632965, + 0.0017858792562037706, + 0.002095836214721203, + 0.0018194079166278243, + 0.0018199966289103031, + 0.000889408343937248, + 0.001477581448853016, + 0.0005025786231271923, + 0.0005805190885439515, + 0.0008400533115491271, + 0.0007103767129592597, + 0.26847198605537415, + 0.2721521854400635, + 0.010712703689932823, + 0.009227295406162739, + 0.017643772065639496, + 0.03510809689760208, + 0.015633026137948036, + 0.0052919210866093636, + 0.0011100940173491836, + 0.001496270764619112, + 0.0008977839024737477, + 0.0012902456801384687, + 0.0013923805672675371, + 0.0012349248863756657, + 0.001029035309329629, + 0.0006202672957442701, + 0.024956846609711647, + 0.02496768906712532, + 0.0010016593150794506, + 0.0006928340881131589, + 0.0014317836612462997, + 0.0017953726928681135, + 0.0007036026800051332, + 0.0006605760427191854, + 0.0009253335301764309, + 0.0009241777006536722, + 0.0001679684064583853, + 0.00019848387455567718, + 0.0002576732076704502, + 0.00017100128752645105, + 0.00028332258807495236, + 0.00017004368419293314, + 0.00034495568252168596, + 0.0003406425821594894, + 0.000340199883794412, + 0.00044467218685895205, + 0.0003957039734814316, + 0.0001886952668428421, + 0.0005638310685753822, + 0.00027921574655920267, + 0.002553344704210758, + 0.0023935886565595865, + 0.0004952398012392223, + 0.0005749590927734971, + 0.0004161910619586706, + 0.0004277797706890851, + 0.00035956094507128, + 0.000520957401022315, + 0.0024124840274453163, + 0.0034613090101629496, + 0.002508700592443347, + 0.0017122174613177776, + 0.0024556315038353205, + 0.0012834739172831178, + 0.00168520153965801, + 0.0017496537184342742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.002484808210283518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943837575614452 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943837575614452, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943837575614452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16764529049396515, + 0.07949754595756531, + 0.034321293234825134, + 0.2930266559123993, + 0.04545366391539574, + 0.2605016827583313, + 0.03884970396757126, + 0.036329545080661774, + 0.020364409312605858, + 0.024010205641388893 + ], + "credit_map": [ + 0.0411798320710659, + 0.05350526422262192, + 0.024883022531867027, + 0.007826242595911026, + 0.011670115403831005, + 0.00987339112907648, + 0.011238274164497852, + 0.007469143718481064, + 0.028915653005242348, + 0.027243519201874733, + 0.005702223628759384, + 0.003447320545092225, + 0.002628104528412223, + 0.0042624706402421, + 0.0023705700878053904, + 0.004927685484290123, + 0.0049386643804609776, + 0.00487325806170702, + 0.0036285631358623505, + 0.006627979222685099, + 0.004149416461586952, + 0.0031791036017239094, + 0.003501154249534011, + 0.00342315505258739, + 0.12301574647426605, + 0.12497192621231079, + 0.004375871736556292, + 0.004693107213824987, + 0.009129902347922325, + 0.01638077385723591, + 0.007386265322566032, + 0.003073063911870122, + 0.006566138006746769, + 0.008150149136781693, + 0.004270169883966446, + 0.0057209935039281845, + 0.008568103425204754, + 0.005543319042772055, + 0.004064066801220179, + 0.002570721786469221, + 0.11537618935108185, + 0.11505299806594849, + 0.004403791390359402, + 0.0032949221786111593, + 0.0073418947868049145, + 0.008246437646448612, + 0.00376871507614851, + 0.0030167268123477697, + 0.012688410468399525, + 0.012811179272830486, + 0.0014243690529838204, + 0.002433398738503456, + 0.002767845056951046, + 0.0025970737915486097, + 0.0025425488129258156, + 0.0015848764451220632, + 0.010442905128002167, + 0.01052138116210699, + 0.002043404383584857, + 0.0018744217231869698, + 0.0035533250775188208, + 0.0020734835416078568, + 0.0036532417871057987, + 0.0021673832088708878, + 0.005846551153808832, + 0.005411786027252674, + 0.0011295666918158531, + 0.0017195116961374879, + 0.001515597919933498, + 0.0013100339565426111, + 0.0018984103808179498, + 0.0015329503221437335, + 0.0033393483608961105, + 0.0039005063008517027, + 0.002666173269972205, + 0.002431347733363509, + 0.003412345889955759, + 0.0022076477762311697, + 0.00272026052698493, + 0.0033325757831335068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.005943837575614452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot energetically follows a golden retriever as it runs with boundless enthusiasm down the steep slope of a massive sand dune. The dog's long golden fur flutters in the wind, its ears flap happily, and its pink tongue sticks out from the corner of its mouth in a blissful smile. Its entire body is a picture centered on movement, its powerful leg muscles propelling it forward, its paws digging into the fine, pale sand, and its every step sending up small clouds of dust. The camera remains low, emphasizing the dog's speed and determination as it chases a bright yellow tennis ball, which tumbles and bounces precariously just ahead of it. The entire scene is bathed in the warm golden light of the afternoon sun, casting long, soft shadows and making every grain of sand shimmer. In the background, the deep blue of the ocean meets the hazy light blue of the sky on the horizon, creating an expansive atmosphere of pure, carefree tranquility.", + "seed": 20816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_5553_seed_20816_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026084855012595654, + "vlm_logprob": -0.026084855012595654, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.026084855012595654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0672641173005104 + ], + "opposite_logprob_tokens": [ + -2.8172640800476074 + ], + "logprob_reward": -0.0672641173005104, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0672641173005104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22954301536083221, + 0.08081987500190735, + 0.13941121101379395, + 0.10594234615564346, + 0.06301595270633698, + 0.11128533631563187, + 0.1253339797258377, + 0.02656700275838375, + 0.04296324402093887, + 0.07511799782514572 + ], + "credit_map": [ + 0.033015716820955276, + 0.053123027086257935, + 0.015763649716973305, + 0.012113061733543873, + 0.03152371197938919, + 0.021992672234773636, + 0.027460936456918716, + 0.03455023095011711, + 0.008571869693696499, + 0.009782393462955952, + 0.007665177807211876, + 0.010274332016706467, + 0.01222739927470684, + 0.009216098114848137, + 0.010724162682890892, + 0.012358443811535835, + 0.010870199650526047, + 0.03360520303249359, + 0.008781537413597107, + 0.010789899155497551, + 0.017121806740760803, + 0.03507666289806366, + 0.011764573864638805, + 0.011401322670280933, + 0.015450750477612019, + 0.033233124762773514, + 0.006407837383449078, + 0.006359023042023182, + 0.009602433070540428, + 0.007108753081411123, + 0.007166725117713213, + 0.02061370201408863, + 0.008317163214087486, + 0.007068098522722721, + 0.0054875644855201244, + 0.005034192930907011, + 0.011112869717180729, + 0.008872494101524353, + 0.008854953572154045, + 0.008268619887530804, + 0.023804733529686928, + 0.020001281052827835, + 0.0103581752628088, + 0.006730858236551285, + 0.014973673969507217, + 0.008834551088511944, + 0.011678232811391354, + 0.014903835952281952, + 0.014334578067064285, + 0.017234453931450844, + 0.015101911500096321, + 0.010525127872824669, + 0.01696178875863552, + 0.026806585490703583, + 0.013599302619695663, + 0.010770237073302269, + 0.0029328460805118084, + 0.0028019058518111706, + 0.0037669821176677942, + 0.0029931513126939535, + 0.004412079229950905, + 0.003424535971134901, + 0.003507953602820635, + 0.002727548126131296, + 0.0047251200303435326, + 0.004351959563791752, + 0.006678817793726921, + 0.004837154410779476, + 0.0065839942544698715, + 0.004793972708284855, + 0.005370693281292915, + 0.0056215268559753895, + 0.011227922514081001, + 0.007580538280308247, + 0.009509224444627762, + 0.009013078175485134, + 0.010672411881387234, + 0.010889632627367973, + 0.0065259807743132114, + 0.00969921238720417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0672641173005104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982412464916706 + ], + "opposite_logprob_tokens": [ + -3.5298240184783936 + ], + "logprob_reward": -0.02982412464916706, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982412464916706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15707425773143768, + 0.05310186743736267, + 0.0708329975605011, + 0.1939903348684311, + 0.07190608978271484, + 0.14172391593456268, + 0.09054315835237503, + 0.05002154782414436, + 0.10734666138887405, + 0.06345918029546738 + ], + "credit_map": [ + 0.021281566470861435, + 0.024984177201986313, + 0.018503522500395775, + 0.010952797718346119, + 0.02722538448870182, + 0.012364805676043034, + 0.021702781319618225, + 0.020059218630194664, + 0.009725992567837238, + 0.007763443514704704, + 0.006749289575964212, + 0.00488665234297514, + 0.0053809285163879395, + 0.006446254905313253, + 0.005890928208827972, + 0.0062583801336586475, + 0.008770275861024857, + 0.012610291130840778, + 0.006023290567100048, + 0.007471682503819466, + 0.012423352338373661, + 0.009176995605230331, + 0.007272455375641584, + 0.0070846546441316605, + 0.02734195627272129, + 0.053710613399744034, + 0.012097805738449097, + 0.01315270271152258, + 0.02200288511812687, + 0.018428761512041092, + 0.014070137403905392, + 0.03318547084927559, + 0.008245368488132954, + 0.00919633824378252, + 0.006916685961186886, + 0.007837781682610512, + 0.008734680712223053, + 0.01230187900364399, + 0.0087983263656497, + 0.009875030256807804, + 0.030087321996688843, + 0.023373158648610115, + 0.02298825979232788, + 0.014194462448358536, + 0.013958806172013283, + 0.01258962694555521, + 0.01282661221921444, + 0.011705664917826653, + 0.01378311961889267, + 0.012412330135703087, + 0.010931394062936306, + 0.00716099888086319, + 0.011400916613638401, + 0.01306877750903368, + 0.013165872544050217, + 0.008619754575192928, + 0.0053352792747318745, + 0.005383035633713007, + 0.007266353815793991, + 0.004930519498884678, + 0.00908998865634203, + 0.005985240451991558, + 0.006234858650714159, + 0.005796272307634354, + 0.01890392415225506, + 0.007682772818952799, + 0.0107452142983675, + 0.010000335052609444, + 0.014356146566569805, + 0.022682199254631996, + 0.011128815822303295, + 0.01184725109487772, + 0.008031966164708138, + 0.007986604236066341, + 0.0060057383961975574, + 0.009312007576227188, + 0.008824465796351433, + 0.00791462603956461, + 0.007802177686244249, + 0.007581592071801424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.02982412464916706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329622581601143 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.02329622581601143, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329622581601143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1853870004415512, + 0.0727548748254776, + 0.45711013674736023, + 0.10361170023679733, + 0.03264224901795387, + 0.03952684625983238, + 0.04373624920845032, + 0.012916824780404568, + 0.028306366875767708, + 0.024007735773921013 + ], + "credit_map": [ + 0.028751127421855927, + 0.04559117555618286, + 0.01598382368683815, + 0.011559201404452324, + 0.023858055472373962, + 0.0190172977745533, + 0.018182555213570595, + 0.022443756461143494, + 0.014057227410376072, + 0.019195539876818657, + 0.005176833365112543, + 0.007429828401654959, + 0.006743150297552347, + 0.005844781640917063, + 0.007054926827549934, + 0.007252588868141174, + 0.038673270493745804, + 0.11677027493715286, + 0.029416749253869057, + 0.03908628597855568, + 0.03833259269595146, + 0.12455742806196213, + 0.03238902986049652, + 0.037884533405303955, + 0.015647821128368378, + 0.031079240143299103, + 0.006744762416929007, + 0.00796997919678688, + 0.009884413331747055, + 0.008001623675227165, + 0.006665743887424469, + 0.01761811599135399, + 0.0044662742875516415, + 0.004328811075538397, + 0.0034658429212868214, + 0.0034131528809666634, + 0.0034588465932756662, + 0.004852184560149908, + 0.004357028286904097, + 0.004300106782466173, + 0.008340366184711456, + 0.005656067747622728, + 0.0050622690469026566, + 0.003025618614628911, + 0.004167135804891586, + 0.004652227275073528, + 0.004382421262562275, + 0.004240740090608597, + 0.00587314460426569, + 0.00494837062433362, + 0.004885945934802294, + 0.004112157505005598, + 0.006423547398298979, + 0.008200425654649734, + 0.004262386821210384, + 0.005030269268900156, + 0.0012268491555005312, + 0.0015328563749790192, + 0.0012179454788565636, + 0.0011740056797862053, + 0.0025568169075995684, + 0.0012428410118445754, + 0.0022267017047852278, + 0.001738809165544808, + 0.004576093517243862, + 0.0029449304565787315, + 0.001924585085362196, + 0.0026577210519462824, + 0.00501544214785099, + 0.005938610527664423, + 0.0026255138218402863, + 0.0026234702672809362, + 0.0031926482915878296, + 0.002573091071099043, + 0.0025506396777927876, + 0.004463597666472197, + 0.0028647035360336304, + 0.003016455564647913, + 0.0029004232492297888, + 0.0024461757857352495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02329622581601143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.147135391831398, + 0.04386362060904503, + 0.11828422546386719, + 0.325913667678833, + 0.08595868200063705, + 0.08684086054563522, + 0.08076319843530655, + 0.03127703815698624, + 0.02579737827181816, + 0.05416599661111832 + ], + "credit_map": [ + 0.01976785808801651, + 0.01974870078265667, + 0.02422652766108513, + 0.01660585217177868, + 0.00934769120067358, + 0.017892468720674515, + 0.016240473836660385, + 0.023305818438529968, + 0.007952016778290272, + 0.004963270854204893, + 0.005642443895339966, + 0.004844055976718664, + 0.005444481037557125, + 0.004240385722368956, + 0.005713204853236675, + 0.005063759163022041, + 0.008037343621253967, + 0.03649528697133064, + 0.003898278810083866, + 0.004528854042291641, + 0.009976951405405998, + 0.04206584393978119, + 0.006311772856861353, + 0.006969886366277933, + 0.05791127309203148, + 0.11387618631124496, + 0.021457133814692497, + 0.015099232085049152, + 0.013063600286841393, + 0.022717691957950592, + 0.014353657141327858, + 0.06743490695953369, + 0.014505256898701191, + 0.015890924260020256, + 0.00930715911090374, + 0.006335714366286993, + 0.009296117350459099, + 0.013342018239200115, + 0.00766846863552928, + 0.009613021276891232, + 0.013272003270685673, + 0.012917863205075264, + 0.0120417894795537, + 0.012939016334712505, + 0.009479845874011517, + 0.006041348446160555, + 0.006830692756921053, + 0.01331829372793436, + 0.012315971776843071, + 0.012358181178569794, + 0.009248885326087475, + 0.006243638228625059, + 0.00866879615932703, + 0.016343679279088974, + 0.008121181279420853, + 0.007462867069989443, + 0.00446077948436141, + 0.0050839828327298164, + 0.003605750622227788, + 0.0026425605174154043, + 0.0040807221084833145, + 0.004413036163896322, + 0.0030941369477659464, + 0.0038960669189691544, + 0.00318376743234694, + 0.002284520072862506, + 0.0031519040931016207, + 0.0035610732156783342, + 0.0034746804740279913, + 0.003360418137162924, + 0.0033793796319514513, + 0.0034016345161944628, + 0.009777169674634933, + 0.005291983485221863, + 0.006706045009195805, + 0.006397962104529142, + 0.005260766949504614, + 0.010043811984360218, + 0.004858091473579407, + 0.005830168724060059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2611599564552307, + 0.05526429042220116, + 0.20506691932678223, + 0.1972087323665619, + 0.060878921300172806, + 0.07833588123321533, + 0.059082888066768646, + 0.020085537806153297, + 0.03182867914438248, + 0.03108810819685459 + ], + "credit_map": [ + 0.03945405036211014, + 0.0642140731215477, + 0.02061866782605648, + 0.018650658428668976, + 0.03434652090072632, + 0.018766777589917183, + 0.024073056876659393, + 0.041036173701286316, + 0.011611195281147957, + 0.01033917535096407, + 0.004701103083789349, + 0.006761711556464434, + 0.006143228616565466, + 0.00568891316652298, + 0.004483808297663927, + 0.005535165313631296, + 0.015435570850968361, + 0.07422066479921341, + 0.004904625006020069, + 0.004683579783886671, + 0.008245372213423252, + 0.08140446990728378, + 0.004837131593376398, + 0.01133554708212614, + 0.03281109407544136, + 0.06318959593772888, + 0.014276872389018536, + 0.010635721497237682, + 0.009914259426295757, + 0.014062562957406044, + 0.009021005593240261, + 0.04329763352870941, + 0.010154466144740582, + 0.009345566853880882, + 0.006081358063966036, + 0.005709875840693712, + 0.00762308482080698, + 0.008606571704149246, + 0.005841315723955631, + 0.007516695186495781, + 0.016240470111370087, + 0.014007307589054108, + 0.01067476999014616, + 0.007815084420144558, + 0.007521712686866522, + 0.008468490093946457, + 0.00616219500079751, + 0.007445862051099539, + 0.008203326724469662, + 0.008338592946529388, + 0.006734292488545179, + 0.005201432388275862, + 0.007311656605452299, + 0.008358299732208252, + 0.0084256362169981, + 0.006509661674499512, + 0.0024085857439786196, + 0.0027884889859706163, + 0.0027472088113427162, + 0.002258606953546405, + 0.002908692928031087, + 0.0018228014232590795, + 0.0025945056695491076, + 0.002556651597842574, + 0.0038257953710854053, + 0.003429124131798744, + 0.002110387198626995, + 0.0036991967353969812, + 0.0055622076615691185, + 0.0034109661355614662, + 0.005166489630937576, + 0.004624518565833569, + 0.003975338768213987, + 0.0031990993302315474, + 0.004326492082327604, + 0.00495376531034708, + 0.004011118784546852, + 0.003378581488505006, + 0.0031866151839494705, + 0.004057100974023342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.005947155877947807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot energetically follows a golden retriever as it runs with boundless enthusiasm down the steep slope of a massive sand dune. The dog's long golden fur flutters in the wind, its ears flap happily, and its pink tongue sticks out from the corner of its mouth in a blissful smile. Its entire body is a picture centered on movement, its powerful leg muscles propelling it forward, its paws digging into the fine, pale sand, and its every step sending up small clouds of dust. The camera remains low, emphasizing the dog's speed and determination as it chases a bright yellow tennis ball, which tumbles and bounces precariously just ahead of it. The entire scene is bathed in the warm golden light of the afternoon sun, casting long, soft shadows and making every grain of sand shimmer. In the background, the deep blue of the ocean meets the hazy light blue of the sky on the horizon, creating an expansive atmosphere of pure, carefree tranquility.", + "seed": 20817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_5/index_5553_seed_20817_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04100253083743155, + "vlm_logprob": -0.04100253083743155, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.04100253083743155, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16540275514125824 + ], + "opposite_logprob_tokens": [ + -1.9154027700424194 + ], + "logprob_reward": -0.16540275514125824, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16540275514125824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09371510148048401, + 0.0795237123966217, + 0.042256247252225876, + 0.12411297857761383, + 0.16892731189727783, + 0.07532183080911636, + 0.07530131936073303, + 0.20520219206809998, + 0.0921381339430809, + 0.04350120574235916 + ], + "credit_map": [ + 0.012438744306564331, + 0.013298437930643559, + 0.01036227960139513, + 0.013477987609803677, + 0.013173201121389866, + 0.009363597258925438, + 0.013409916311502457, + 0.008190941996872425, + 0.009626668877899647, + 0.013497769832611084, + 0.01083020307123661, + 0.011851174756884575, + 0.010987997055053711, + 0.008559861220419407, + 0.007074382156133652, + 0.007095652632415295, + 0.004698717035353184, + 0.006115024443715811, + 0.0048459856770932674, + 0.007337141782045364, + 0.0063402047380805016, + 0.0046747829765081406, + 0.005008989945054054, + 0.0032354004215449095, + 0.010688209906220436, + 0.012016762979328632, + 0.01195467822253704, + 0.02978568524122238, + 0.01865251548588276, + 0.009398418478667736, + 0.02126363292336464, + 0.010353075340390205, + 0.018481871113181114, + 0.018927935510873795, + 0.04993937537074089, + 0.03716173395514488, + 0.015964604914188385, + 0.006595152895897627, + 0.01382808480411768, + 0.00802855659276247, + 0.008848502300679684, + 0.009980401024222374, + 0.006052926182746887, + 0.019234567880630493, + 0.007588515989482403, + 0.007357328198850155, + 0.009657356888055801, + 0.0066022262908518314, + 0.01173414383083582, + 0.008151935413479805, + 0.01505536399781704, + 0.010616115294396877, + 0.006929617840796709, + 0.007898970507085323, + 0.006582156289368868, + 0.008333017118275166, + 0.020002273842692375, + 0.025131208822131157, + 0.01881629228591919, + 0.05650986731052399, + 0.015066568739712238, + 0.01738615520298481, + 0.02469613030552864, + 0.02759367786347866, + 0.00914775487035513, + 0.02076301909983158, + 0.008267654106020927, + 0.016442762687802315, + 0.008850809186697006, + 0.009276614524424076, + 0.008265244774520397, + 0.011124269105494022, + 0.0046516782604157925, + 0.007083819713443518, + 0.007292298134416342, + 0.00715249078348279, + 0.0050253295339643955, + 0.0052847000770270824, + 0.0031660906970500946, + 0.0038447987753897905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.16540275514125824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600058138370514 + ], + "opposite_logprob_tokens": [ + -3.8956000804901123 + ], + "logprob_reward": -0.020600058138370514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600058138370514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.206501305103302, + 0.19375570118427277, + 0.047256749123334885, + 0.05839388072490692, + 0.10247817635536194, + 0.03905752673745155, + 0.13541410863399506, + 0.08504731953144073, + 0.06889429688453674, + 0.0632009357213974 + ], + "credit_map": [ + 0.025077596306800842, + 0.031598616391420364, + 0.021922627463936806, + 0.049207352101802826, + 0.022492438554763794, + 0.023019321262836456, + 0.015897322446107864, + 0.017286034300923347, + 0.017241060733795166, + 0.017447227612137794, + 0.06623044610023499, + 0.05134334787726402, + 0.01242739800363779, + 0.009628583677113056, + 0.011992146261036396, + 0.007445481605827808, + 0.005331358872354031, + 0.00681333988904953, + 0.005643072072416544, + 0.008724645711481571, + 0.005981634836643934, + 0.006306938361376524, + 0.0050927698612213135, + 0.003362989751622081, + 0.0045477524399757385, + 0.005900874268263578, + 0.007076576817780733, + 0.011999061331152916, + 0.009916485287249088, + 0.003975982777774334, + 0.009453336708247662, + 0.005523813422769308, + 0.019144723191857338, + 0.013137973845005035, + 0.025253871455788612, + 0.01303824596107006, + 0.011199708096683025, + 0.005073898937553167, + 0.009366761893033981, + 0.00626299437135458, + 0.004329144023358822, + 0.007004579994827509, + 0.003429136471822858, + 0.00788384210318327, + 0.003754208330065012, + 0.0027585586067289114, + 0.005928336177021265, + 0.003969721030443907, + 0.020102182403206825, + 0.013995004817843437, + 0.04112435132265091, + 0.021550146862864494, + 0.007475169375538826, + 0.006796484347432852, + 0.008242425508797169, + 0.016128335148096085, + 0.01135803572833538, + 0.010015437379479408, + 0.010390558280050755, + 0.016361193731427193, + 0.009269403293728828, + 0.008030197583138943, + 0.0077796210534870625, + 0.011842872016131878, + 0.0071249050088226795, + 0.011795523576438427, + 0.00881052203476429, + 0.010065636597573757, + 0.005893454886972904, + 0.008138180710375309, + 0.004992804490029812, + 0.012073268182575703, + 0.007929853163659573, + 0.010125355795025826, + 0.007970569655299187, + 0.00755100604146719, + 0.008533066138625145, + 0.009447219781577587, + 0.005581070203334093, + 0.006062795873731375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020600058138370514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979037955403328 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.00979037955403328, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979037955403328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3631960451602936, + 0.09408847242593765, + 0.07573706656694412, + 0.1647035926580429, + 0.05776021629571915, + 0.06799907982349396, + 0.038945090025663376, + 0.039640750735998154, + 0.060480788350105286, + 0.03744891285896301 + ], + "credit_map": [ + 0.0458124615252018, + 0.05108923465013504, + 0.04589192569255829, + 0.10502984374761581, + 0.03209724649786949, + 0.0188046433031559, + 0.038859281688928604, + 0.02561143785715103, + 0.010968375019729137, + 0.009148837067186832, + 0.011007556691765785, + 0.017047950997948647, + 0.01160347182303667, + 0.021943815052509308, + 0.006731494329869747, + 0.005636974703520536, + 0.008632528595626354, + 0.0069890148006379604, + 0.009464085102081299, + 0.016023878008127213, + 0.010496587492525578, + 0.007171470206230879, + 0.012278798036277294, + 0.004680701065808535, + 0.01568836160004139, + 0.016533801332116127, + 0.018325896933674812, + 0.033441461622714996, + 0.028084823861718178, + 0.012834245339035988, + 0.02850508876144886, + 0.011289916932582855, + 0.008873428218066692, + 0.009788471274077892, + 0.00595243601128459, + 0.010585551150143147, + 0.009093476459383965, + 0.003791045630350709, + 0.00596650829538703, + 0.00370929972268641, + 0.006678121630102396, + 0.012791397981345654, + 0.005617422051727772, + 0.011252093128859997, + 0.008544676005840302, + 0.004736012779176235, + 0.012728143483400345, + 0.00565121416002512, + 0.005313923582434654, + 0.004962476436048746, + 0.007975608110427856, + 0.005611506290733814, + 0.00311113428324461, + 0.003325557801872492, + 0.004728383384644985, + 0.003916500136256218, + 0.00554246548563242, + 0.007492488250136375, + 0.004182306118309498, + 0.006224209442734718, + 0.005117942579090595, + 0.0031611209269613028, + 0.004385232925415039, + 0.003534984076395631, + 0.00732884556055069, + 0.012654098682105541, + 0.007196338847279549, + 0.01049975585192442, + 0.004205881617963314, + 0.0055918446741998196, + 0.004090932197868824, + 0.00891309604048729, + 0.004441471770405769, + 0.006262961775064468, + 0.00533438241109252, + 0.005749959964305162, + 0.004475944675505161, + 0.0041870963759720325, + 0.003858261276036501, + 0.003138837171718478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00979037955403328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024856405798345804 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024856405798345804, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024856405798345804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35721755027770996, + 0.08994031697511673, + 0.0521363765001297, + 0.051749661564826965, + 0.06311878561973572, + 0.0981941670179367, + 0.06610970944166183, + 0.088997982442379, + 0.04035140573978424, + 0.09218401461839676 + ], + "credit_map": [ + 0.02666817419230938, + 0.033828798681497574, + 0.03413737192749977, + 0.08030841499567032, + 0.05578762665390968, + 0.066432423889637, + 0.027157630771398544, + 0.03289711847901344, + 0.012092015706002712, + 0.007292123511433601, + 0.013685465790331364, + 0.01620238646864891, + 0.009966937825083733, + 0.017532482743263245, + 0.007228502072393894, + 0.005940399598330259, + 0.0084318732842803, + 0.003973728511482477, + 0.005718237720429897, + 0.0075070420280098915, + 0.005990607663989067, + 0.0101862084120512, + 0.0061707827262580395, + 0.004157892893999815, + 0.00653246883302927, + 0.0044846078380942345, + 0.005929008591920137, + 0.009787931106984615, + 0.0062326304614543915, + 0.007055932190269232, + 0.007022184319794178, + 0.00470490287989378, + 0.011187350377440453, + 0.005194713827222586, + 0.01927110366523266, + 0.010539846494793892, + 0.003744442481547594, + 0.003816429991275072, + 0.005253723356872797, + 0.004111167974770069, + 0.007846822030842304, + 0.02141251042485237, + 0.00809906143695116, + 0.018550843000411987, + 0.008564217947423458, + 0.008179150521755219, + 0.015536405146121979, + 0.010005153715610504, + 0.011826353147625923, + 0.005145215895026922, + 0.0241545457392931, + 0.008941128849983215, + 0.0026574404910206795, + 0.00324936187826097, + 0.002783034695312381, + 0.007352626882493496, + 0.010125072672963142, + 0.01177835650742054, + 0.009038873016834259, + 0.015625029802322388, + 0.0054427385330200195, + 0.016960857436060905, + 0.006103345192968845, + 0.013923711143434048, + 0.00465393578633666, + 0.006386617198586464, + 0.004931324627250433, + 0.004977627191692591, + 0.0043821982108056545, + 0.005863686092197895, + 0.0034198302309960127, + 0.005736187566071749, + 0.01614115945994854, + 0.008506431244313717, + 0.01946962997317314, + 0.00915741827338934, + 0.007685391698032618, + 0.013374820351600647, + 0.005284401588141918, + 0.012564754113554955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0024856405798345804 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever running down a steep sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle and tracking the dog's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright yellow tennis ball visible tumbling ahead of the dog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background show a deep blue ocean meeting a hazy light blue sky on the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sunlight with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733820773661137 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006733820773661137, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733820773661137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2746482491493225, + 0.3068443238735199, + 0.025853294879198074, + 0.03307342901825905, + 0.06364460289478302, + 0.06514789909124374, + 0.07624255865812302, + 0.040385305881500244, + 0.04047841578722, + 0.07368193566799164 + ], + "credit_map": [ + 0.02798192948102951, + 0.046972114592790604, + 0.031101444736123085, + 0.05443151295185089, + 0.030706046149134636, + 0.03368043154478073, + 0.02560456469655037, + 0.02417019195854664, + 0.0285955797880888, + 0.021905507892370224, + 0.11602096259593964, + 0.06825149804353714, + 0.01373204868286848, + 0.03125258907675743, + 0.015014225617051125, + 0.012071925215423107, + 0.0032931827008724213, + 0.0030246679671108723, + 0.003542664460837841, + 0.003853860544040799, + 0.003182650078088045, + 0.0031224568374454975, + 0.0035654339008033276, + 0.0022683788556605577, + 0.0029184301383793354, + 0.004388164728879929, + 0.0031412469688802958, + 0.005694701801985502, + 0.006360242143273354, + 0.003857042407616973, + 0.004262872040271759, + 0.002450731350108981, + 0.012000663205981255, + 0.009813733398914337, + 0.009329468011856079, + 0.0078009371645748615, + 0.008921010419726372, + 0.004703589715063572, + 0.005476712714880705, + 0.005598493851721287, + 0.005385491531342268, + 0.013664945028722286, + 0.006473430432379246, + 0.011755947954952717, + 0.006378242280334234, + 0.004799323622137308, + 0.010713422670960426, + 0.005977099295705557, + 0.012996761128306389, + 0.00606826925650239, + 0.02200997993350029, + 0.008734269067645073, + 0.005068739876151085, + 0.006369069218635559, + 0.0071224006824195385, + 0.007873067632317543, + 0.006788912229239941, + 0.004615353420376778, + 0.007019018288701773, + 0.0056798807345330715, + 0.0031110835261642933, + 0.006173479370772839, + 0.002505440032109618, + 0.00449213944375515, + 0.0049748108722269535, + 0.006970161572098732, + 0.0052639879286289215, + 0.0052361758425831795, + 0.003887733444571495, + 0.0051306504756212234, + 0.00213273405097425, + 0.006882161367684603, + 0.008534064516425133, + 0.009931642562150955, + 0.011555111035704613, + 0.008665169589221478, + 0.007978487759828568, + 0.007369447965174913, + 0.0075604477897286415, + 0.012087561190128326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.006733820773661137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_tensor([504, 504])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_tensor([504, 504])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3b86bc78eea5787f475b0140c958c63c0388152d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_tensor([504, 504])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a long shot that pays homage to the black-and-white photography of Hiroshi Sugimoto, the image begins at a nearly static, horizontal angle and pans rightward at a glacially slow pace, almost imperceptible to the naked eye. The entire scene is rendered in an extremely high-contrast black and white, richly textured and immeasurably gradated. Above, heavy, swirling clouds cast a deep, charcoal black, a rare double rainbow hangs majestically. This double rainbow, manifesting not in color but as two pure, ethereal arcs of light: the inner rainbow bridge is bright and solid, radiating a searing white glow, while the outer rainbow is a fainter, ghostly, softer halo, their presence defined by the stark contrast against the darkness of the sky behind them. Beneath this sacred arc of light, a landscape of withered vegetation, frozen in place by an icestorm, lies a landscape now enveloped by a thick, crystalline crust of ice. Once withered branches and leaves, delicate grass stems, and the skeletal forms of trees, it is now enveloped by a thick, translucent shell of ice, as if the entire world has been instantly sealed within a vast glass artwork. As the camera pans with an almost meditative rhythm, low-angle light penetrating the clouds refracts across the facets of countless ice crystals, creating a slow, hypnotic ripple of flickering light. Light and shadow stretch and deform across the frozen ground with a slowness that borders on geological time, ultimately creating a timeless, solemn, and profound sense of historical presence, as if the viewer is witnessing a timeless moment rather than a fleeting meteorological spectacle.", + "seed": 20303, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_504_seed_20303_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5725831894204021, + "vlm_logprob": -0.5725831894204021, + "vlm_yes_no_margin": 1.55, + "vlm_token_credit": -0.5725831894204021, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.128406524658203 + ], + "opposite_logprob_tokens": [ + -0.12840649485588074 + ], + "logprob_reward": -2.128406524658203, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.128406524658203, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19091813266277313, + 0.04967289790511131, + 0.028429554775357246, + 0.07271547615528107, + 0.5201034545898438, + 0.036457303911447525, + 0.03308868035674095, + 0.014776092022657394, + 0.030954068526625633, + 0.022884318605065346 + ], + "credit_map": [ + 0.021815050393342972, + 0.017599865794181824, + 0.03280346468091011, + 0.023044539615511894, + 0.016259372234344482, + 0.03566993027925491, + 0.030034270137548447, + 0.013691636733710766, + 0.0047323438338935375, + 0.004942151252180338, + 0.008253789506852627, + 0.005342827644199133, + 0.005453723948448896, + 0.008217747323215008, + 0.007168678566813469, + 0.0055616386234760284, + 0.0028921414632350206, + 0.00267050345428288, + 0.004247763659805059, + 0.003545224666595459, + 0.003434613812714815, + 0.005322491750121117, + 0.003609877312555909, + 0.0027069412171840668, + 0.006063013803213835, + 0.007369330618530512, + 0.013692587614059448, + 0.011219855397939682, + 0.006693663075566292, + 0.008131934329867363, + 0.012111670337617397, + 0.007433425635099411, + 0.02897954359650612, + 0.051108431071043015, + 0.01991494558751583, + 0.13078242540359497, + 0.027843371033668518, + 0.023321043699979782, + 0.20714768767356873, + 0.031006038188934326, + 0.003090675687417388, + 0.003722133580595255, + 0.004222092684358358, + 0.005756346508860588, + 0.003923967946320772, + 0.0029210341162979603, + 0.007137797307223082, + 0.005683256778866053, + 0.0033711125142872334, + 0.004116868134588003, + 0.0034701540134847164, + 0.003334203502163291, + 0.0034271094482392073, + 0.006583327893167734, + 0.004945525899529457, + 0.003840383607894182, + 0.001312660751864314, + 0.002071361755952239, + 0.0027496819384396076, + 0.0019376816926524043, + 0.0015921939630061388, + 0.0017516135703772306, + 0.0013621883699670434, + 0.001998709747567773, + 0.0029450778383761644, + 0.004793247673660517, + 0.0037112184800207615, + 0.0034773682709783316, + 0.0033434187062084675, + 0.006075975950807333, + 0.003385969204828143, + 0.0032217944972217083, + 0.0015252851881086826, + 0.0025178883224725723, + 0.0039031263440847397, + 0.0039031440392136574, + 0.002152410801500082, + 0.002934846794232726, + 0.00241793692111969, + 0.003529680659994483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -2.128406524658203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16047178208827972 + ], + "opposite_logprob_tokens": [ + -1.910471796989441 + ], + "logprob_reward": -0.16047178208827972, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16047178208827972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33257386088371277, + 0.05050094053149223, + 0.03839948773384094, + 0.08133036643266678, + 0.3416682183742523, + 0.027672629803419113, + 0.03645313158631325, + 0.02023470588028431, + 0.038029782474040985, + 0.03313685581088066 + ], + "credit_map": [ + 0.029457032680511475, + 0.040355708450078964, + 0.03873785585165024, + 0.04056992009282112, + 0.02508721873164177, + 0.03893809765577316, + 0.05782310292124748, + 0.06160491332411766, + 0.004652878735214472, + 0.0048506008461117744, + 0.007978057488799095, + 0.006627736147493124, + 0.004399476572871208, + 0.007117916829884052, + 0.006990830413997173, + 0.007883443497121334, + 0.0034769552294164896, + 0.003668425837531686, + 0.004935470875352621, + 0.005287157371640205, + 0.003502191510051489, + 0.0056620994582772255, + 0.0043885186314582825, + 0.007478670217096806, + 0.007511260453611612, + 0.007759183645248413, + 0.011503162793815136, + 0.013021904043853283, + 0.008740920573472977, + 0.010811591520905495, + 0.012998328544199467, + 0.008984019979834557, + 0.01862705685198307, + 0.03164898231625557, + 0.01206982508301735, + 0.0894281268119812, + 0.01902902126312256, + 0.014279801398515701, + 0.13793455064296722, + 0.018650872632861137, + 0.002501168753951788, + 0.003069623839110136, + 0.003238953184336424, + 0.003966642078012228, + 0.003099768655374646, + 0.0029189852066338062, + 0.004806132987141609, + 0.004071356263011694, + 0.0033483945298939943, + 0.004191246815025806, + 0.005350842606276274, + 0.004758939146995544, + 0.003470836905762553, + 0.006704481318593025, + 0.004507150966674089, + 0.004121239297091961, + 0.0019642014522105455, + 0.0026206611655652523, + 0.0029512352775782347, + 0.0025579947978258133, + 0.001783389481715858, + 0.00334699684754014, + 0.0020145645830780268, + 0.0029956616926938295, + 0.00346257328055799, + 0.005408457480370998, + 0.004380539990961552, + 0.004153456073254347, + 0.0045256298035383224, + 0.0070970915257930756, + 0.003803943283855915, + 0.0051980894058942795, + 0.0024398525711148977, + 0.003463139059022069, + 0.006721602287143469, + 0.0055616325698792934, + 0.003582154866307974, + 0.0036452230997383595, + 0.003153616562485695, + 0.004569635726511478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.16047178208827972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556425295770168 + ], + "opposite_logprob_tokens": [ + -4.387556552886963 + ], + "logprob_reward": -0.012556425295770168, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556425295770168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18543799221515656, + 0.0600125826895237, + 0.030706122517585754, + 0.06567181646823883, + 0.5175496935844421, + 0.03516296669840813, + 0.0427490659058094, + 0.019827019423246384, + 0.0249620471149683, + 0.017920738086104393 + ], + "credit_map": [ + 0.01901555247604847, + 0.018608666956424713, + 0.02695745788514614, + 0.023516472429037094, + 0.016868481412529945, + 0.031136969104409218, + 0.03289863094687462, + 0.016435740515589714, + 0.0061562806367874146, + 0.0058866217732429504, + 0.009698844514787197, + 0.010693581774830818, + 0.005310144275426865, + 0.00875886157155037, + 0.009061859920620918, + 0.004446382634341717, + 0.0029139306861907244, + 0.0026926295831799507, + 0.005277109332382679, + 0.004728479310870171, + 0.003226542379707098, + 0.004522777162492275, + 0.004572924226522446, + 0.0027717275079339743, + 0.00543154776096344, + 0.005857551470398903, + 0.012574960477650166, + 0.011103897355496883, + 0.005806182976812124, + 0.005850627552717924, + 0.01117973867803812, + 0.007867306470870972, + 0.02908891998231411, + 0.04925377294421196, + 0.017221370711922646, + 0.13433416187763214, + 0.028157804161310196, + 0.02161911502480507, + 0.20957142114639282, + 0.0283031165599823, + 0.0028189944569021463, + 0.0036417266819626093, + 0.003992823418229818, + 0.005507235880941153, + 0.0038117514923214912, + 0.0026249040383845568, + 0.006848245393484831, + 0.005917280446738005, + 0.0029540678951889277, + 0.003123822156339884, + 0.008149668574333191, + 0.008735585026443005, + 0.003015531925484538, + 0.004694214090704918, + 0.005513246636837721, + 0.006562924478203058, + 0.0016334397951141, + 0.002162371063604951, + 0.002898241626098752, + 0.0035662841983139515, + 0.0018699768697842956, + 0.0021842815913259983, + 0.0021377529483288527, + 0.003374669700860977, + 0.0022804862819612026, + 0.004085282329469919, + 0.0030635870061814785, + 0.0028278157114982605, + 0.0031343181617558002, + 0.004079749807715416, + 0.0024001009296625853, + 0.0030907040927559137, + 0.0011105840094387531, + 0.0018828100292012095, + 0.003805752145126462, + 0.0037951108533889055, + 0.001714192098006606, + 0.001593716093339026, + 0.0020434274338185787, + 0.001975143328309059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.012556425295770168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233608484268188 + ], + "opposite_logprob_tokens": [ + -0.8983608484268188 + ], + "logprob_reward": -0.5233608484268188, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233608484268188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06629548966884613, + 0.03420313075184822, + 0.025130124762654305, + 0.05371370539069176, + 0.6542460322380066, + 0.08656004071235657, + 0.015309558250010014, + 0.03731541335582733, + 0.012697244994342327, + 0.014529213309288025 + ], + "credit_map": [ + 0.006135611794888973, + 0.009003037586808205, + 0.008177850395441055, + 0.010094009339809418, + 0.00832238420844078, + 0.010509122163057327, + 0.006721427198499441, + 0.007332046050578356, + 0.003246566979214549, + 0.0038030336145311594, + 0.004735465627163649, + 0.004728877451270819, + 0.00404052110388875, + 0.006411813665181398, + 0.0034512230195105076, + 0.0037856290582567453, + 0.0020988283213227987, + 0.002687976462766528, + 0.004349416121840477, + 0.0031823895405977964, + 0.002612228039652109, + 0.0032661634031683207, + 0.0022928949911147356, + 0.004640226252377033, + 0.004491068422794342, + 0.006233704276382923, + 0.007018527016043663, + 0.008641986176371574, + 0.006198099814355373, + 0.006230026483535767, + 0.008479112759232521, + 0.006421181373298168, + 0.03542134538292885, + 0.055548299103975296, + 0.024177132174372673, + 0.17024989426136017, + 0.03802606090903282, + 0.02751370146870613, + 0.26440712809562683, + 0.03890251740813255, + 0.005954246502369642, + 0.007922515273094177, + 0.014187270775437355, + 0.013456864282488823, + 0.009154792875051498, + 0.00728854862973094, + 0.012471655383706093, + 0.016124149784445763, + 0.0011157267726957798, + 0.0015270595904439688, + 0.001921715447679162, + 0.002079090103507042, + 0.001303764758631587, + 0.0013923323713243008, + 0.0035126341972500086, + 0.002457235474139452, + 0.003127592382952571, + 0.004761143587529659, + 0.007587060797959566, + 0.006089205853641033, + 0.003083267481997609, + 0.004537201952189207, + 0.0034955772571265697, + 0.004634364042431116, + 0.0011796997860074043, + 0.0013097679475322366, + 0.0019463654607534409, + 0.0026696568820625544, + 0.0012704160762950778, + 0.0014452053001150489, + 0.0017595426179468632, + 0.0011165922041982412, + 0.0015216241590678692, + 0.0017573196673765779, + 0.002008352894335985, + 0.0019479361362755299, + 0.0014980051200836897, + 0.002629287773743272, + 0.0014881431125104427, + 0.0016785450279712677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.5233608484268188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038120366632938385 + ], + "opposite_logprob_tokens": [ + -3.2881202697753906 + ], + "logprob_reward": -0.038120366632938385, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038120366632938385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23140949010849, + 0.07465636730194092, + 0.03542295843362808, + 0.08630257844924927, + 0.35621708631515503, + 0.0365607887506485, + 0.08331329375505447, + 0.027004962787032127, + 0.0416506789624691, + 0.027461843565106392 + ], + "credit_map": [ + 0.020342478528618813, + 0.022908760234713554, + 0.03018820285797119, + 0.03785430267453194, + 0.022177699953317642, + 0.02917390875518322, + 0.03014560230076313, + 0.038618508726358414, + 0.005626271944493055, + 0.006875859573483467, + 0.013583171181380749, + 0.01637834683060646, + 0.005907277576625347, + 0.009030165150761604, + 0.00954204797744751, + 0.0077132247388362885, + 0.003023316152393818, + 0.0028958087787032127, + 0.005387900862842798, + 0.005249769426882267, + 0.0041615962982177734, + 0.005530725698918104, + 0.0049261110834777355, + 0.004247730597853661, + 0.006767967715859413, + 0.011676196940243244, + 0.014800853095948696, + 0.009148688986897469, + 0.011192869395017624, + 0.011111217550933361, + 0.008856331929564476, + 0.012748446315526962, + 0.02025536634027958, + 0.03537210077047348, + 0.014357563108205795, + 0.08651207387447357, + 0.020209260284900665, + 0.01897246204316616, + 0.1379513293504715, + 0.02258692868053913, + 0.003080197609961033, + 0.004515297245234251, + 0.005248303059488535, + 0.004947463981807232, + 0.004548502154648304, + 0.003067741170525551, + 0.005422213580459356, + 0.0057310666888952255, + 0.007787272334098816, + 0.010721486061811447, + 0.013175380416214466, + 0.009146357886493206, + 0.007467948365956545, + 0.014022384770214558, + 0.009711550548672676, + 0.011280909180641174, + 0.002648739144206047, + 0.0027431072667241096, + 0.0036897659301757812, + 0.004272900987416506, + 0.002641969593241811, + 0.004152343608438969, + 0.0033757544588297606, + 0.003480380168184638, + 0.0035153029020875692, + 0.006041686050593853, + 0.0070662484504282475, + 0.007337489165365696, + 0.0030552498064935207, + 0.0069678546860814095, + 0.004322355147451162, + 0.003344491124153137, + 0.0018353472696617246, + 0.0032608972396701574, + 0.005713022314012051, + 0.005657013971358538, + 0.0032629924826323986, + 0.00221725064329803, + 0.003095776541158557, + 0.002419541822746396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.038120366632938385 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a long shot that pays homage to the black-and-white photography of Hiroshi Sugimoto, the image begins at a nearly static, horizontal angle and pans rightward at a glacially slow pace, almost imperceptible to the naked eye. The entire scene is rendered in an extremely high-contrast black and white, richly textured and immeasurably gradated. Above, heavy, swirling clouds cast a deep, charcoal black, a rare double rainbow hangs majestically. This double rainbow, manifesting not in color but as two pure, ethereal arcs of light: the inner rainbow bridge is bright and solid, radiating a searing white glow, while the outer rainbow is a fainter, ghostly, softer halo, their presence defined by the stark contrast against the darkness of the sky behind them. Beneath this sacred arc of light, a landscape of withered vegetation, frozen in place by an icestorm, lies a landscape now enveloped by a thick, crystalline crust of ice. Once withered branches and leaves, delicate grass stems, and the skeletal forms of trees, it is now enveloped by a thick, translucent shell of ice, as if the entire world has been instantly sealed within a vast glass artwork. As the camera pans with an almost meditative rhythm, low-angle light penetrating the clouds refracts across the facets of countless ice crystals, creating a slow, hypnotic ripple of flickering light. Light and shadow stretch and deform across the frozen ground with a slowness that borders on geological time, ultimately creating a timeless, solemn, and profound sense of historical presence, as if the viewer is witnessing a timeless moment rather than a fleeting meteorological spectacle.", + "seed": 20304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_504_seed_20304_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16491980776190757, + "vlm_logprob": -0.16491980776190757, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.16491980776190757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06010636314749718 + ], + "opposite_logprob_tokens": [ + -2.9351062774658203 + ], + "logprob_reward": -0.06010636314749718, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06010636314749718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06740790605545044, + 0.01980462856590748, + 0.018064036965370178, + 0.2668183743953705, + 0.09522565454244614, + 0.01704373024404049, + 0.016750242561101913, + 0.032048068940639496, + 0.3972705006599426, + 0.06956689804792404 + ], + "credit_map": [ + 0.01784696988761425, + 0.004745516926050186, + 0.016140589490532875, + 0.005830163136124611, + 0.006204468663781881, + 0.007578754331916571, + 0.004060478415340185, + 0.0050009614787995815, + 0.004328084643930197, + 0.001983095658943057, + 0.004161874298006296, + 0.0018621530616655946, + 0.0017072901828214526, + 0.0026563196443021297, + 0.0016596075147390366, + 0.0014462030958384275, + 0.001789096393622458, + 0.0029055930208414793, + 0.0022868108935654163, + 0.0022738005500286818, + 0.0022730401251465082, + 0.0023178616538643837, + 0.0016849328530952334, + 0.002532901242375374, + 0.033705856651067734, + 0.0696108341217041, + 0.019295809790492058, + 0.02548026852309704, + 0.05557357892394066, + 0.02731061540544033, + 0.020383214578032494, + 0.015458175912499428, + 0.01350712962448597, + 0.014340739697217941, + 0.008154019713401794, + 0.005487008020281792, + 0.03331274166703224, + 0.010492525063455105, + 0.006004439201205969, + 0.003927051089704037, + 0.001640027156099677, + 0.003092744154855609, + 0.0012458049459382892, + 0.0015792216872796416, + 0.0035787667147815228, + 0.0017332417191937566, + 0.0024022499565035105, + 0.0017716723959892988, + 0.003836345626041293, + 0.0011028925655409694, + 0.0013172756880521774, + 0.0007616679649800062, + 0.0058930423110723495, + 0.0014864270342513919, + 0.0015983607154339552, + 0.0007542293751612306, + 0.009183567017316818, + 0.001933117862790823, + 0.006310745608061552, + 0.0019107810221612453, + 0.007594434078782797, + 0.0021147867664694786, + 0.00147713883779943, + 0.0015234973980113864, + 0.03205081447958946, + 0.025568213313817978, + 0.036477889865636826, + 0.11191152781248093, + 0.09526032209396362, + 0.04220404848456383, + 0.029780376702547073, + 0.02401731349527836, + 0.007882138714194298, + 0.015749771147966385, + 0.005017837975174189, + 0.007851035334169865, + 0.017306596040725708, + 0.0069207013584673405, + 0.005139309912919998, + 0.003699506865814328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.06010636314749718 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992639541625977 + ], + "opposite_logprob_tokens": [ + -1.8049263954162598 + ], + "logprob_reward": -0.17992639541625977, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992639541625977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04582110047340393, + 0.028311025351285934, + 0.05018126964569092, + 0.12034361809492111, + 0.0641598254442215, + 0.004933204036206007, + 0.012264130637049675, + 0.019149743020534515, + 0.6014785170555115, + 0.053357526659965515 + ], + "credit_map": [ + 0.010321435518562794, + 0.003779519582167268, + 0.01254061609506607, + 0.00404047966003418, + 0.0027660017367452383, + 0.0060073137283325195, + 0.0031890440732240677, + 0.0031766919419169426, + 0.006147489417344332, + 0.0020304962527006865, + 0.007518561091274023, + 0.0028235670179128647, + 0.002700046868994832, + 0.0026468660216778517, + 0.002420614706352353, + 0.0020233860705047846, + 0.0052878339774906635, + 0.011556270532310009, + 0.005304738879203796, + 0.005737703759223223, + 0.009279845282435417, + 0.005375525448471308, + 0.004389363341033459, + 0.003249992849305272, + 0.016042102128267288, + 0.03142242133617401, + 0.008295434527099133, + 0.011613242328166962, + 0.025085043162107468, + 0.012148491106927395, + 0.00907618086785078, + 0.006660704500973225, + 0.009601850062608719, + 0.009337423369288445, + 0.004796764813363552, + 0.003636036766692996, + 0.023218179121613503, + 0.007139432709664106, + 0.003853457048535347, + 0.0025766759645193815, + 0.00045804836554452777, + 0.0009513844270259142, + 0.0005007931613363326, + 0.0003942031762562692, + 0.0011598088312894106, + 0.00043772420031018555, + 0.000588702445384115, + 0.0004425389051903039, + 0.0016210833564400673, + 0.0009335267823189497, + 0.001881632604636252, + 0.0008364772074855864, + 0.0021988770458847284, + 0.0017414850881323218, + 0.001452038181014359, + 0.0015990104293450713, + 0.004230991471558809, + 0.0012937933206558228, + 0.0018251239089295268, + 0.0005799286882393062, + 0.007540415972471237, + 0.001900997944176197, + 0.0010171208996325731, + 0.0007613710477016866, + 0.048335906118154526, + 0.039147522300481796, + 0.05475088581442833, + 0.16908730566501617, + 0.14550071954727173, + 0.06411244720220566, + 0.04484343156218529, + 0.035700347274541855, + 0.006108204368501902, + 0.010022792033851147, + 0.005252568516880274, + 0.007778859231621027, + 0.011815818026661873, + 0.004280343651771545, + 0.004547872580587864, + 0.003551066853106022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.17992639541625977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819748803973198 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.01819748803973198, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819748803973198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05232779309153557, + 0.049381088465452194, + 0.12280865758657455, + 0.02157996967434883, + 0.36109259724617004, + 0.07052164524793625, + 0.029782285913825035, + 0.02380671724677086, + 0.12363579124212265, + 0.14506348967552185 + ], + "credit_map": [ + 0.006933398544788361, + 0.0048920875415205956, + 0.010043771006166935, + 0.005018187686800957, + 0.007216221187263727, + 0.006928642746061087, + 0.00452979002147913, + 0.006765695288777351, + 0.00983367022126913, + 0.0036521998699754477, + 0.003313384484499693, + 0.005207741633057594, + 0.01331147737801075, + 0.004499079193919897, + 0.006222128868103027, + 0.0033414061181247234, + 0.013551604934036732, + 0.01188699435442686, + 0.015160092152655125, + 0.012569728307425976, + 0.03649996221065521, + 0.01606195978820324, + 0.007001163437962532, + 0.010077162645757198, + 0.0031084308866411448, + 0.002470439998432994, + 0.003185543231666088, + 0.003240516409277916, + 0.0033070791978389025, + 0.002082981402054429, + 0.0021850320044904947, + 0.001999946078285575, + 0.051404718309640884, + 0.0550931841135025, + 0.028597792610526085, + 0.020155848935246468, + 0.12934240698814392, + 0.041149381548166275, + 0.02105124294757843, + 0.01429803017526865, + 0.007367237005382776, + 0.013595258817076683, + 0.004731081426143646, + 0.006334213539958, + 0.015593868680298328, + 0.006968026515096426, + 0.00965060479938984, + 0.006281353533267975, + 0.008064504712820053, + 0.0018845174927264452, + 0.00245035532861948, + 0.0017462767427787185, + 0.009270334616303444, + 0.0019007170340046287, + 0.003146129660308361, + 0.0013194509083405137, + 0.005447415169328451, + 0.0014433921314775944, + 0.005339261144399643, + 0.001947831129655242, + 0.0024900950957089663, + 0.002332407282665372, + 0.003278111107647419, + 0.0015282046515494585, + 0.009913462214171886, + 0.008047530427575111, + 0.011325671337544918, + 0.03476103022694588, + 0.029647646471858025, + 0.013040757738053799, + 0.009262796491384506, + 0.007636891212314367, + 0.01956970803439617, + 0.031684305518865585, + 0.008168362081050873, + 0.013827770948410034, + 0.04160943627357483, + 0.013757946901023388, + 0.009696630761027336, + 0.006749314721673727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.01819748803973198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.52330082654953 + ], + "opposite_logprob_tokens": [ + -0.89830082654953 + ], + "logprob_reward": -0.52330082654953, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.52330082654953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014368790201842785, + 0.009076201356947422, + 0.04539542645215988, + 0.08160476386547089, + 0.06471657752990723, + 0.03329618647694588, + 0.06781259924173355, + 0.23309257626533508, + 0.42908957600593567, + 0.02154732123017311 + ], + "credit_map": [ + 0.002015090547502041, + 0.001195484190247953, + 0.0028699408285319805, + 0.0011303526116535068, + 0.0019423647318035364, + 0.0017446657875552773, + 0.0018926826305687428, + 0.001578209106810391, + 0.0016858555609360337, + 0.0007019889308139682, + 0.0010219315299764276, + 0.0008389679132960737, + 0.0017256588907912374, + 0.0013328434433788061, + 0.0010516787879168987, + 0.0007172751938924193, + 0.004765294026583433, + 0.010318273678421974, + 0.004035658203065395, + 0.005083238705992699, + 0.008436581119894981, + 0.0057010953314602375, + 0.0038598282262682915, + 0.00319545972160995, + 0.010687686502933502, + 0.021222952753305435, + 0.005692788399755955, + 0.007685289718210697, + 0.017118720337748528, + 0.008458524011075497, + 0.006144856568425894, + 0.004593938123434782, + 0.009329397231340408, + 0.009746886789798737, + 0.005347182508558035, + 0.00363755039870739, + 0.022882437333464622, + 0.0072462595999240875, + 0.003822325263172388, + 0.002704538404941559, + 0.0035948557779192924, + 0.006301450077444315, + 0.002311739372089505, + 0.003020049538463354, + 0.0069306050427258015, + 0.0031427384819835424, + 0.004933691583573818, + 0.0030610566027462482, + 0.01412496529519558, + 0.0044813696295022964, + 0.0060209776274859905, + 0.00426699360832572, + 0.01959610916674137, + 0.0064807310700416565, + 0.008862919174134731, + 0.003978535998612642, + 0.02914227917790413, + 0.011585252359509468, + 0.0776621401309967, + 0.018990634009242058, + 0.04244847223162651, + 0.026803480461239815, + 0.012484738603234291, + 0.01397557370364666, + 0.03352803364396095, + 0.028148414567112923, + 0.03934759274125099, + 0.11775267869234085, + 0.10735524445772171, + 0.045899536460638046, + 0.03152914345264435, + 0.0255289226770401, + 0.0025691883638501167, + 0.0047768657095730305, + 0.001460813800804317, + 0.0019907886162400246, + 0.00557942409068346, + 0.002043972723186016, + 0.0018445999594405293, + 0.0012816678499802947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.52330082654953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067965656518936 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.043067965656518936, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067965656518936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031491588801145554, + 0.010563608258962631, + 0.037316691130399704, + 0.13901296257972717, + 0.036302171647548676, + 0.012056421488523483, + 0.07639608532190323, + 0.0396495945751667, + 0.5903187394142151, + 0.02689214237034321 + ], + "credit_map": [ + 0.008696368895471096, + 0.0016187062719836831, + 0.006134654860943556, + 0.0023888631258159876, + 0.004655590746551752, + 0.003994181286543608, + 0.0016436257865279913, + 0.002359601203352213, + 0.0009587255772203207, + 0.0009646783582866192, + 0.0013769111828878522, + 0.0012820390984416008, + 0.002205003285780549, + 0.0011414231266826391, + 0.0016191238537430763, + 0.0010157034266740084, + 0.0034328526817262173, + 0.0029660379514098167, + 0.0054207067005336285, + 0.004469900391995907, + 0.009799288585782051, + 0.003616828005760908, + 0.0035421012435108423, + 0.0040689753368496895, + 0.020602300763130188, + 0.03467295691370964, + 0.008228673599660397, + 0.014501167461276054, + 0.030684227123856544, + 0.01338035985827446, + 0.009820281527936459, + 0.007122991606593132, + 0.0052065723575651646, + 0.00502807879820466, + 0.003162131179124117, + 0.0019915143493562937, + 0.013037923723459244, + 0.004210647661238909, + 0.0022256856318563223, + 0.0014396192273125052, + 0.0015438527334481478, + 0.001942044124007225, + 0.0007562405662611127, + 0.0009404734009876847, + 0.002928841160610318, + 0.0012342677218839526, + 0.0015967607032507658, + 0.0011139408452436328, + 0.018768051639199257, + 0.005206693895161152, + 0.0034592386800795794, + 0.004312737379223108, + 0.026636645197868347, + 0.005968533921986818, + 0.009390345774590969, + 0.002653829986229539, + 0.0038697910495102406, + 0.002266769064590335, + 0.009064605459570885, + 0.0022383141331374645, + 0.012935233302414417, + 0.004500776994973421, + 0.003012823173776269, + 0.0017612783703953028, + 0.04783570393919945, + 0.038578420877456665, + 0.053982242941856384, + 0.16663692891597748, + 0.1408814787864685, + 0.06265076994895935, + 0.044287923723459244, + 0.03546525165438652, + 0.0032366968225687742, + 0.004169448744505644, + 0.003514515934512019, + 0.0038298543076962233, + 0.0052072759717702866, + 0.0025469763204455376, + 0.0020233867689967155, + 0.0023639884311705828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.043067965656518936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a long shot that pays homage to the black-and-white photography of Hiroshi Sugimoto, the image begins at a nearly static, horizontal angle and pans rightward at a glacially slow pace, almost imperceptible to the naked eye. The entire scene is rendered in an extremely high-contrast black and white, richly textured and immeasurably gradated. Above, heavy, swirling clouds cast a deep, charcoal black, a rare double rainbow hangs majestically. This double rainbow, manifesting not in color but as two pure, ethereal arcs of light: the inner rainbow bridge is bright and solid, radiating a searing white glow, while the outer rainbow is a fainter, ghostly, softer halo, their presence defined by the stark contrast against the darkness of the sky behind them. Beneath this sacred arc of light, a landscape of withered vegetation, frozen in place by an icestorm, lies a landscape now enveloped by a thick, crystalline crust of ice. Once withered branches and leaves, delicate grass stems, and the skeletal forms of trees, it is now enveloped by a thick, translucent shell of ice, as if the entire world has been instantly sealed within a vast glass artwork. As the camera pans with an almost meditative rhythm, low-angle light penetrating the clouds refracts across the facets of countless ice crystals, creating a slow, hypnotic ripple of flickering light. Light and shadow stretch and deform across the frozen ground with a slowness that borders on geological time, ultimately creating a timeless, solemn, and profound sense of historical presence, as if the viewer is witnessing a timeless moment rather than a fleeting meteorological spectacle.", + "seed": 20305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_504_seed_20305_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20867054983973504, + "vlm_logprob": -0.20867054983973504, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.20867054983973504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05384428799152374 + ], + "opposite_logprob_tokens": [ + -3.0538442134857178 + ], + "logprob_reward": -0.05384428799152374, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05384428799152374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37308579683303833, + 0.1245359480381012, + 0.04565510153770447, + 0.09323930740356445, + 0.05531953647732735, + 0.040326666086912155, + 0.037887535989284515, + 0.12508536875247955, + 0.04820222407579422, + 0.05666254833340645 + ], + "credit_map": [ + 0.0395100973546505, + 0.06396862119436264, + 0.07493921369314194, + 0.05593276768922806, + 0.03905114904046059, + 0.03131337836384773, + 0.038688331842422485, + 0.029682228341698647, + 0.01995089463889599, + 0.03600632771849632, + 0.009432041086256504, + 0.011218445375561714, + 0.017534106969833374, + 0.010006221942603588, + 0.010965371504426003, + 0.00942254438996315, + 0.005713245831429958, + 0.008691294118762016, + 0.006387477274984121, + 0.005261572543531656, + 0.006942973472177982, + 0.002958143362775445, + 0.004673409275710583, + 0.005026981700211763, + 0.011781646870076656, + 0.016794933006167412, + 0.015641316771507263, + 0.009662963449954987, + 0.01086542196571827, + 0.00712565379217267, + 0.011139001697301865, + 0.010228375904262066, + 0.008316051214933395, + 0.012420549988746643, + 0.011331966146826744, + 0.00461374968290329, + 0.005030974745750427, + 0.00512919994071126, + 0.004060255829244852, + 0.004416788928210735, + 0.006075084675103426, + 0.007202162407338619, + 0.003752132412046194, + 0.003065011464059353, + 0.007130141835659742, + 0.0034489233512431383, + 0.00469925394281745, + 0.0049539594911038876, + 0.004322461783885956, + 0.007369502447545528, + 0.006319758947938681, + 0.003959150519222021, + 0.005061523988842964, + 0.0037683250848203897, + 0.00358986877836287, + 0.003496942576020956, + 0.014344848692417145, + 0.020265206694602966, + 0.0157022625207901, + 0.025439802557229996, + 0.012390817515552044, + 0.009832489304244518, + 0.015051626600325108, + 0.012058318592607975, + 0.007021899800747633, + 0.007360059302300215, + 0.006954104173928499, + 0.008117611519992352, + 0.00570685276761651, + 0.0034828116185963154, + 0.005171864293515682, + 0.0043870252557098866, + 0.005947688594460487, + 0.006296662148088217, + 0.0068727293983101845, + 0.007128708064556122, + 0.010175243951380253, + 0.005901468452066183, + 0.008611546829342842, + 0.005728498101234436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.05384428799152374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865185618400574 + ], + "opposite_logprob_tokens": [ + -1.2236518859863281 + ], + "logprob_reward": -0.34865185618400574, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865185618400574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26246607303619385, + 0.1450357884168625, + 0.18179647624492645, + 0.08690549433231354, + 0.05068352445960045, + 0.05246986821293831, + 0.06077469885349274, + 0.07831528782844543, + 0.038425106555223465, + 0.043127696961164474 + ], + "credit_map": [ + 0.03416113927960396, + 0.036904703825712204, + 0.040197454392910004, + 0.03302624076604843, + 0.05377532169222832, + 0.020105816423892975, + 0.02555452287197113, + 0.018740849569439888, + 0.021032216027379036, + 0.039541903883218765, + 0.016361789777874947, + 0.011505650356411934, + 0.023109667003154755, + 0.010117407888174057, + 0.015466198325157166, + 0.007900949567556381, + 0.018920041620731354, + 0.021651607006788254, + 0.023509623482823372, + 0.021749574691057205, + 0.02977929823100567, + 0.01467361394315958, + 0.03541510924696922, + 0.016097603365778923, + 0.01066033449023962, + 0.009442522190511227, + 0.012614079751074314, + 0.008022435940802097, + 0.018189841881394386, + 0.008282829076051712, + 0.011808526702225208, + 0.007884927093982697, + 0.0063983057625591755, + 0.0075325630605220795, + 0.005505333188921213, + 0.005030137952417135, + 0.009268173016607761, + 0.004796945955604315, + 0.007228424772620201, + 0.004923637490719557, + 0.005632312037050724, + 0.008973248302936554, + 0.00884836632758379, + 0.004524323623627424, + 0.006959105841815472, + 0.003586407285183668, + 0.00646632444113493, + 0.007479779422283173, + 0.007825927808880806, + 0.010106201283633709, + 0.006263360846787691, + 0.007559319958090782, + 0.010439980775117874, + 0.005550089757889509, + 0.007592935115098953, + 0.005436880048364401, + 0.00849766656756401, + 0.011941477656364441, + 0.013740035705268383, + 0.008778776042163372, + 0.012232374399900436, + 0.006308757700026035, + 0.010186314582824707, + 0.006629879120737314, + 0.004376388620585203, + 0.00836491771042347, + 0.006616700440645218, + 0.004159897565841675, + 0.00536204781383276, + 0.0029628421179950237, + 0.0036273279692977667, + 0.002954979194328189, + 0.004118372220546007, + 0.006188233382999897, + 0.008244779892265797, + 0.005352392327040434, + 0.006988221779465675, + 0.0033452073112130165, + 0.005037969909608364, + 0.0038525196723639965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.34865185618400574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073763370513916 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073763370513916, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073763370513916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3730165958404541, + 0.059858642518520355, + 0.0690956711769104, + 0.0690767914056778, + 0.015014533884823322, + 0.04169248417019844, + 0.04333212226629257, + 0.2602849304676056, + 0.03724367544054985, + 0.03138458728790283 + ], + "credit_map": [ + 0.04437485337257385, + 0.07214602082967758, + 0.06475973129272461, + 0.05901520326733589, + 0.03372654691338539, + 0.026944894343614578, + 0.02954503335058689, + 0.042504314333200455, + 0.008309596218168736, + 0.009691515006124973, + 0.008356102742254734, + 0.006552458740770817, + 0.008772681467235088, + 0.005709002260118723, + 0.005935072898864746, + 0.006532210856676102, + 0.008796516805887222, + 0.012649334035813808, + 0.007917049340903759, + 0.00746488431468606, + 0.010394074954092503, + 0.005153657402843237, + 0.011220250278711319, + 0.005499906372278929, + 0.0077659678645431995, + 0.008428012020885944, + 0.015046167187392712, + 0.007863454520702362, + 0.008966365829110146, + 0.004821025766432285, + 0.008113195188343525, + 0.008072602562606335, + 0.002132541500031948, + 0.002842071233317256, + 0.0015204792143777013, + 0.0011632411042228341, + 0.0027653833385556936, + 0.0012941773748025298, + 0.001820321544073522, + 0.0014763189246878028, + 0.00532536581158638, + 0.006724142003804445, + 0.005129815544933081, + 0.0034675924107432365, + 0.007242289371788502, + 0.0035007318947464228, + 0.006048190873116255, + 0.004254356492310762, + 0.005853868555277586, + 0.007697283290326595, + 0.0053398036397993565, + 0.004897598177194595, + 0.006691797636449337, + 0.0034423843026161194, + 0.0052763912826776505, + 0.004132996313273907, + 0.02871793694794178, + 0.04468318074941635, + 0.03314730152487755, + 0.056314025074243546, + 0.02472485601902008, + 0.01945926807820797, + 0.027675440534949303, + 0.02556290477514267, + 0.00415059644728899, + 0.009030862711369991, + 0.007065517827868462, + 0.004012596793472767, + 0.004154120571911335, + 0.0023188041523098946, + 0.003737018909305334, + 0.002774153370410204, + 0.004212734289467335, + 0.0061967340297997, + 0.003553869901224971, + 0.004793134517967701, + 0.004348039627075195, + 0.0024070674553513527, + 0.0028346471954137087, + 0.0030383621342480183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.016073763370513916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761136412620544 + ], + "opposite_logprob_tokens": [ + -0.8261136412620544 + ], + "logprob_reward": -0.5761136412620544, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761136412620544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21115553379058838, + 0.14964377880096436, + 0.19819973409175873, + 0.05512189120054245, + 0.08555445075035095, + 0.03678635135293007, + 0.03400591015815735, + 0.0836632028222084, + 0.03563578799366951, + 0.1102333590388298 + ], + "credit_map": [ + 0.016990674659609795, + 0.027833759784698486, + 0.033834051340818405, + 0.03976377844810486, + 0.0252008605748415, + 0.014206756837666035, + 0.0327749028801918, + 0.020550759509205818, + 0.016698529943823814, + 0.02441091276705265, + 0.02463248372077942, + 0.022975215688347816, + 0.020752402022480965, + 0.008935604244470596, + 0.016680117696523666, + 0.014558523893356323, + 0.023835178464651108, + 0.03208940103650093, + 0.04193442314863205, + 0.02808956243097782, + 0.016415653750300407, + 0.013019280508160591, + 0.023106642067432404, + 0.019709594547748566, + 0.007785605266690254, + 0.007558777462691069, + 0.007767886854708195, + 0.006634111050516367, + 0.007613220252096653, + 0.004541137255728245, + 0.00772777060046792, + 0.00549338199198246, + 0.01107245497405529, + 0.014776144176721573, + 0.014103671535849571, + 0.008966079913079739, + 0.009312652982771397, + 0.006664659827947617, + 0.007044978439807892, + 0.013613814488053322, + 0.004710337147116661, + 0.006516735535115004, + 0.005631421692669392, + 0.003353071166202426, + 0.00620194710791111, + 0.0031731019262224436, + 0.003969882149249315, + 0.0032298536971211433, + 0.003512348746880889, + 0.0058275386691093445, + 0.00531125208362937, + 0.003989618271589279, + 0.004531637765467167, + 0.003498370759189129, + 0.004008756950497627, + 0.0033263862133026123, + 0.00864911824464798, + 0.012714006006717682, + 0.009611005894839764, + 0.012815078720450401, + 0.01118115521967411, + 0.006692915689200163, + 0.014839563518762589, + 0.00716035719960928, + 0.003871372202411294, + 0.004336242564022541, + 0.006794589571654797, + 0.00490600848570466, + 0.004366734996438026, + 0.0027957649435847998, + 0.0048890747129917145, + 0.003676000516861677, + 0.011155972257256508, + 0.010178948752582073, + 0.017880622297525406, + 0.029645051807165146, + 0.01287265494465828, + 0.008278215304017067, + 0.012663168832659721, + 0.00755872530862689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.5761136412620544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669200390577316 + ], + "opposite_logprob_tokens": [ + -3.0486690998077393 + ], + "logprob_reward": -0.048669200390577316, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669200390577316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2275671660900116, + 0.10670791566371918, + 0.11777174472808838, + 0.1145971268415451, + 0.05613304302096367, + 0.047306180000305176, + 0.0980352908372879, + 0.12169219553470612, + 0.06999693065881729, + 0.040192361921072006 + ], + "credit_map": [ + 0.023769386112689972, + 0.026471566408872604, + 0.03202499821782112, + 0.026944631710648537, + 0.04429972916841507, + 0.028109222650527954, + 0.026304932311177254, + 0.01964271068572998, + 0.011846574954688549, + 0.015771711245179176, + 0.013925367034971714, + 0.00867476686835289, + 0.020183026790618896, + 0.011087482795119286, + 0.01743035949766636, + 0.007788627874106169, + 0.014555358327925205, + 0.013076170347630978, + 0.013528530485928059, + 0.015008900314569473, + 0.021985482424497604, + 0.009287020191550255, + 0.01631089672446251, + 0.01401938684284687, + 0.014783293008804321, + 0.019942229613661766, + 0.02113606035709381, + 0.010164856910705566, + 0.014651080593466759, + 0.009011123329401016, + 0.010994917713105679, + 0.013913562521338463, + 0.00787134375423193, + 0.008534369990229607, + 0.008658261969685555, + 0.004343864507973194, + 0.008501204662024975, + 0.004597511142492294, + 0.005109754391014576, + 0.008516734465956688, + 0.00703444005921483, + 0.009180951863527298, + 0.005388982594013214, + 0.004176768474280834, + 0.0070102764293551445, + 0.0037480772007256746, + 0.006867972202599049, + 0.0038987125735729933, + 0.010343308560550213, + 0.015616782009601593, + 0.014608187600970268, + 0.011261486448347569, + 0.013059268705546856, + 0.007976851426064968, + 0.013403423130512238, + 0.011765986680984497, + 0.013608619570732117, + 0.017091641202569008, + 0.007944593206048012, + 0.010678689926862717, + 0.029024839401245117, + 0.011236869730055332, + 0.019963521510362625, + 0.012143423780798912, + 0.010206001810729504, + 0.01595260761678219, + 0.00805365014821291, + 0.010020737536251545, + 0.0087208840996027, + 0.004833668936043978, + 0.00703976396471262, + 0.005169615615159273, + 0.004606612026691437, + 0.005002729594707489, + 0.005917233414947987, + 0.005534370429813862, + 0.007116124499589205, + 0.003549325279891491, + 0.004441398195922375, + 0.004024570342153311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.048669200390577316 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a long shot that pays homage to the black-and-white photography of Hiroshi Sugimoto, the image begins at a nearly static, horizontal angle and pans rightward at a glacially slow pace, almost imperceptible to the naked eye. The entire scene is rendered in an extremely high-contrast black and white, richly textured and immeasurably gradated. Above, heavy, swirling clouds cast a deep, charcoal black, a rare double rainbow hangs majestically. This double rainbow, manifesting not in color but as two pure, ethereal arcs of light: the inner rainbow bridge is bright and solid, radiating a searing white glow, while the outer rainbow is a fainter, ghostly, softer halo, their presence defined by the stark contrast against the darkness of the sky behind them. Beneath this sacred arc of light, a landscape of withered vegetation, frozen in place by an icestorm, lies a landscape now enveloped by a thick, crystalline crust of ice. Once withered branches and leaves, delicate grass stems, and the skeletal forms of trees, it is now enveloped by a thick, translucent shell of ice, as if the entire world has been instantly sealed within a vast glass artwork. As the camera pans with an almost meditative rhythm, low-angle light penetrating the clouds refracts across the facets of countless ice crystals, creating a slow, hypnotic ripple of flickering light. Light and shadow stretch and deform across the frozen ground with a slowness that borders on geological time, ultimately creating a timeless, solemn, and profound sense of historical presence, as if the viewer is witnessing a timeless moment rather than a fleeting meteorological spectacle.", + "seed": 20306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_6/index_504_seed_20306_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17139032818377017, + "vlm_logprob": -0.17139032818377017, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.17139032818377017, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2063755840063095 + ], + "opposite_logprob_tokens": [ + -1.7063755989074707 + ], + "logprob_reward": -0.2063755840063095, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2063755840063095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4282754361629486, + 0.13797301054000854, + 0.07655824720859528, + 0.05982702225446701, + 0.02648819237947464, + 0.04366717487573624, + 0.0450395904481411, + 0.05211600661277771, + 0.04899081215262413, + 0.08106447756290436 + ], + "credit_map": [ + 0.03356201946735382, + 0.034537602216005325, + 0.0761500671505928, + 0.03688012436032295, + 0.03840705379843712, + 0.03588353097438812, + 0.11499106138944626, + 0.05786397308111191, + 0.017299076542258263, + 0.016540061682462692, + 0.020686544477939606, + 0.016726437956094742, + 0.015160086564719677, + 0.01584157533943653, + 0.021447371691465378, + 0.014271850697696209, + 0.005747444462031126, + 0.010843499563634396, + 0.011930668726563454, + 0.00895539578050375, + 0.012582875788211823, + 0.006832797080278397, + 0.010128102265298367, + 0.009537464007735252, + 0.006169472821056843, + 0.006798761896789074, + 0.007904858328402042, + 0.006565960124135017, + 0.0086830398067832, + 0.006367428228259087, + 0.0077275182120501995, + 0.009609980508685112, + 0.0024176568258553743, + 0.0031172193121165037, + 0.0034415898844599724, + 0.003879385069012642, + 0.0024890333879739046, + 0.0026447854470461607, + 0.003671653103083372, + 0.004826868884265423, + 0.0033542742021381855, + 0.004759024828672409, + 0.007709430996328592, + 0.007421308662742376, + 0.003248897148296237, + 0.00525038642808795, + 0.005259381141513586, + 0.006664469838142395, + 0.004091319628059864, + 0.004052222706377506, + 0.005859977100044489, + 0.008596217259764671, + 0.0049527897499501705, + 0.004301285371184349, + 0.0066257864236831665, + 0.006559988018125296, + 0.006817674729973078, + 0.005275148898363113, + 0.006843941286206245, + 0.00907795038074255, + 0.004863531328737736, + 0.006464743055403233, + 0.007026652339845896, + 0.0057463617995381355, + 0.005437665153294802, + 0.005527660716325045, + 0.00630485825240612, + 0.007066071033477783, + 0.004890416283160448, + 0.006689570378512144, + 0.005742039065808058, + 0.0073325298726558685, + 0.006268211640417576, + 0.006591747514903545, + 0.006442926824092865, + 0.02341204695403576, + 0.005142152309417725, + 0.006195953115820885, + 0.006887777242809534, + 0.020123664289712906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.2063755840063095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521563172340393 + ], + "opposite_logprob_tokens": [ + -1.5021562576293945 + ], + "logprob_reward": -0.2521563172340393, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521563172340393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19886595010757446, + 0.1178320124745369, + 0.12077067047357559, + 0.05358687788248062, + 0.03833336383104324, + 0.059472162276506424, + 0.017627036198973656, + 0.02063082531094551, + 0.06752023845911026, + 0.30536088347435 + ], + "credit_map": [ + 0.013545223511755466, + 0.023919377475976944, + 0.033361367881298065, + 0.02436799556016922, + 0.023346463218331337, + 0.014760185964405537, + 0.036106765270233154, + 0.02945856563746929, + 0.010047866962850094, + 0.01680227927863598, + 0.016968777403235435, + 0.012214873917400837, + 0.015974219888448715, + 0.012820674106478691, + 0.019523702561855316, + 0.013479608111083508, + 0.010627469047904015, + 0.01903902366757393, + 0.012546508572995663, + 0.014403323642909527, + 0.02442919835448265, + 0.013423864729702473, + 0.012958609499037266, + 0.013342669233679771, + 0.0043968623504042625, + 0.008361711166799068, + 0.005047512240707874, + 0.007343203295022249, + 0.009045674465596676, + 0.006968427915126085, + 0.006315827369689941, + 0.006107662804424763, + 0.0040405262261629105, + 0.006249577738344669, + 0.004176826681941748, + 0.004152865614742041, + 0.007448328193277121, + 0.005188744980841875, + 0.003232954302802682, + 0.0038435393944382668, + 0.004647043067961931, + 0.006061123218387365, + 0.0067038992419838905, + 0.011652890592813492, + 0.009733942337334156, + 0.005401890724897385, + 0.00933416374027729, + 0.005937208887189627, + 0.0018939372384920716, + 0.0019224457209929824, + 0.002172591397538781, + 0.0026238001883029938, + 0.001888061291538179, + 0.00228249398060143, + 0.002269117161631584, + 0.0025745893362909555, + 0.002053725067526102, + 0.002828880911692977, + 0.0029150540940463543, + 0.00251634675078094, + 0.0020334103610366583, + 0.0022552483715116978, + 0.0037106871604919434, + 0.002317471895366907, + 0.006231391802430153, + 0.007414168678224087, + 0.010503515601158142, + 0.010255762375891209, + 0.005465332418680191, + 0.006528802216053009, + 0.010288463905453682, + 0.010832802392542362, + 0.028302226215600967, + 0.03559707850217819, + 0.023936504498124123, + 0.06556472927331924, + 0.04599567502737045, + 0.02948283776640892, + 0.030249210074543953, + 0.04623262211680412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.2521563172340393 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082914546132088 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082914546132088, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082914546132088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2562694251537323, + 0.17338529229164124, + 0.0677885040640831, + 0.04892789199948311, + 0.05859481170773506, + 0.1718021035194397, + 0.056620750576257706, + 0.014315446838736534, + 0.062300316989421844, + 0.08999548852443695 + ], + "credit_map": [ + 0.018036458641290665, + 0.020345717668533325, + 0.033832550048828125, + 0.017713112756609917, + 0.02006562054157257, + 0.013832276687026024, + 0.07475798577070236, + 0.05768570303916931, + 0.015817733481526375, + 0.019747266545891762, + 0.02901339717209339, + 0.02250879630446434, + 0.024801179766654968, + 0.015536174178123474, + 0.026364946737885475, + 0.019595803692936897, + 0.006761686876416206, + 0.008584312163293362, + 0.006651158910244703, + 0.007375645916908979, + 0.011358201503753662, + 0.007468312978744507, + 0.011866506189107895, + 0.007722674403339624, + 0.004244576673954725, + 0.005882366560399532, + 0.005307595245540142, + 0.007227565627545118, + 0.009019882418215275, + 0.0047484650276601315, + 0.005916572641581297, + 0.006580868270248175, + 0.00553479790687561, + 0.006696182303130627, + 0.007921959273517132, + 0.007596263661980629, + 0.006020879838615656, + 0.007831069640815258, + 0.00811161007732153, + 0.008882049471139908, + 0.011617431417107582, + 0.021157586947083473, + 0.024695126339793205, + 0.048984594643116, + 0.01195567287504673, + 0.010957240127027035, + 0.025832096114754677, + 0.01660233549773693, + 0.005598371848464012, + 0.007246044930070639, + 0.006220647133886814, + 0.008598883636295795, + 0.008665927685797215, + 0.005982871633023024, + 0.006753652822226286, + 0.007554348092526197, + 0.0012392838252708316, + 0.0019461066694930196, + 0.003047127276659012, + 0.001994517631828785, + 0.0014077634550631046, + 0.0011856736382469535, + 0.0019561771769076586, + 0.0015387969324365258, + 0.007009301800280809, + 0.0053946287371218204, + 0.009846796281635761, + 0.007632766850292683, + 0.0044837999157607555, + 0.00689401850104332, + 0.009726868011057377, + 0.011312136426568031, + 0.009114806540310383, + 0.012170127592980862, + 0.012395921163260937, + 0.010302701033651829, + 0.016508163884282112, + 0.009493201039731503, + 0.011087216436862946, + 0.008923349902033806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016082914546132088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348626047372818 + ], + "opposite_logprob_tokens": [ + -1.2236260175704956 + ], + "logprob_reward": -0.348626047372818, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.348626047372818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15027709305286407, + 0.06387844681739807, + 0.07970134913921356, + 0.18750697374343872, + 0.016952157020568848, + 0.0362439900636673, + 0.021540261805057526, + 0.02412041835486889, + 0.043452680110931396, + 0.3763265609741211 + ], + "credit_map": [ + 0.012048033997416496, + 0.021899059414863586, + 0.022480566054582596, + 0.023116789758205414, + 0.024489393457770348, + 0.015835678204894066, + 0.016443315893411636, + 0.013964274898171425, + 0.005121325608342886, + 0.0054440987296402454, + 0.01004913542419672, + 0.007834640331566334, + 0.00805946160107851, + 0.005378049798309803, + 0.012906764633953571, + 0.009084976278245449, + 0.007123389281332493, + 0.011230550706386566, + 0.010110204108059406, + 0.008389965631067753, + 0.01530586089938879, + 0.007857069373130798, + 0.011682524345815182, + 0.00800179410725832, + 0.016973387449979782, + 0.024969233199954033, + 0.02446499653160572, + 0.02963622286915779, + 0.038360994309186935, + 0.017011668533086777, + 0.023125067353248596, + 0.012965415604412556, + 0.0018194139702245593, + 0.002844229806214571, + 0.0015246329130604863, + 0.001718352665193379, + 0.00213890359736979, + 0.0020652892999351025, + 0.0025678991805762053, + 0.002273434540256858, + 0.0026873305905610323, + 0.00377373187802732, + 0.004480074159801006, + 0.009152556769549847, + 0.0032632234506309032, + 0.0029769614338874817, + 0.005604103207588196, + 0.004306009970605373, + 0.0023149175103753805, + 0.002887461334466934, + 0.0023611043579876423, + 0.002883122069761157, + 0.0034118774347007275, + 0.0024848254397511482, + 0.002843020483851433, + 0.002353935968130827, + 0.002878971165046096, + 0.003219104604795575, + 0.002802949631586671, + 0.0037145144306123257, + 0.002068161964416504, + 0.003098526271060109, + 0.00432945042848587, + 0.0020087400916963816, + 0.0044914428144693375, + 0.0057545313611626625, + 0.005533794406801462, + 0.006555882282555103, + 0.003893427550792694, + 0.005556161981076002, + 0.005601944401860237, + 0.006065498571842909, + 0.02495916560292244, + 0.021694155409932137, + 0.03167827054858208, + 0.11757143586874008, + 0.011609277687966824, + 0.03311517834663391, + 0.028690602630376816, + 0.10700849443674088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.348626047372818 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in high-contrast black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a double rainbow visible in the sky, appearing as two distinct arcs of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape covered in a thick, translucent shell of ice with frozen vegetation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a very slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible flickering ripple of light refracting across the ice crystals on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710777759552 + ], + "opposite_logprob_tokens": [ + -3.4087107181549072 + ], + "logprob_reward": -0.033710777759552, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710777759552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1355845332145691, + 0.059836193919181824, + 0.06561922281980515, + 0.14677219092845917, + 0.04845331609249115, + 0.1506670117378235, + 0.04798782616853714, + 0.06459832936525345, + 0.10262749344110489, + 0.17785392701625824 + ], + "credit_map": [ + 0.012466413900256157, + 0.01120743341743946, + 0.016830971464514732, + 0.014064086601138115, + 0.015206464566290379, + 0.01242324523627758, + 0.03273046389222145, + 0.020655464380979538, + 0.005751661490648985, + 0.007858430966734886, + 0.011476503685116768, + 0.005819480400532484, + 0.006788416299968958, + 0.005770017392933369, + 0.007739033550024033, + 0.008632650598883629, + 0.005695335566997528, + 0.009006954729557037, + 0.007717638276517391, + 0.0095711974427104, + 0.01295126136392355, + 0.006656006909906864, + 0.007869049906730652, + 0.006151773501187563, + 0.012523898854851723, + 0.02032998949289322, + 0.018997596576809883, + 0.020497925579547882, + 0.03160235285758972, + 0.014824994839727879, + 0.016674958169460297, + 0.011320479214191437, + 0.005109175108373165, + 0.004617431201040745, + 0.0057205501943826675, + 0.0062875146977603436, + 0.006034668534994125, + 0.005723877344280481, + 0.00820900872349739, + 0.006751085165888071, + 0.009763971902430058, + 0.01897800713777542, + 0.020698007196187973, + 0.04066130518913269, + 0.010751397348940372, + 0.010524829849600792, + 0.023646298795938492, + 0.015643199905753136, + 0.005115428939461708, + 0.006755192298442125, + 0.004801145289093256, + 0.006228162441402674, + 0.007802647538483143, + 0.005789740476757288, + 0.005433288868516684, + 0.006062218453735113, + 0.004893281497061253, + 0.006125058047473431, + 0.010335315950214863, + 0.008495786227285862, + 0.00693116057664156, + 0.005411618854850531, + 0.013761699199676514, + 0.008644410409033298, + 0.00896298885345459, + 0.012659742496907711, + 0.014677106402814388, + 0.014200001955032349, + 0.009944835677742958, + 0.009678415954113007, + 0.015358799137175083, + 0.017145603895187378, + 0.012457690201699734, + 0.012697059661149979, + 0.01424162182956934, + 0.05091071128845215, + 0.010114566422998905, + 0.01859823428094387, + 0.01385358814150095, + 0.04498044773936272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.033710777759552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_tensor([2634, 2634])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_tensor([2634, 2634])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc62de642ba8e168a7be6463446e9b00ced3d94e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_tensor([2634, 2634])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking close-up shot from an overhead angle, the camera slowly but steadily pans forward, focusing on a beast composed entirely of fire. This lion-like behemoth is in the grip of a silent yet explosive roar. Its body is composed entirely of flowing, incandescent magma and raging flames. Thick, powerful black outlines outline its taut, muscular form, imbued with a passionate, comic-book style. Its mane is a blazing, ever-brighter tongue of flame, soaring skyward. Its eyes are blazing white cores, radiating boundless fury. Its maw opens to its limit, revealing translucent flaming fangs condensed from pure energy. Simultaneously, its sharp front claws, formed from twisted flames, snap open, their tips gleaming with arcing sparks. The beast is the sole source of light, standing atop a crumbling, cracked black stone floor, from which strands of red energy permeate the cracks. Behind it, a deep, chaotic dark background is dotted with flying sparks and ripples of air distorted by intense heat. The dominant tones of the entire scene are scorching orange, crimson, and bright yellow, creating a high-contrast visual impact against the pure black of the background, creating an extreme atmosphere filled with overwhelming power, fury, and explosive energy.", + "seed": 11410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11410_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1640974273905158, + "vlm_logprob": -0.1640974273905158, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1640974273905158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20782747864723206 + ], + "opposite_logprob_tokens": [ + -1.7078274488449097 + ], + "logprob_reward": -0.20782747864723206, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20782747864723206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05699444189667702, + 0.07444595545530319, + 0.03790893778204918, + 0.02513684146106243, + 0.11867015063762665, + 0.07234539836645126, + 0.10626692324876785, + 0.3066081702709198, + 0.14739403128623962, + 0.05422915518283844 + ], + "credit_map": [ + 0.008297696709632874, + 0.003728396026417613, + 0.01507817767560482, + 0.0039379470981657505, + 0.005204780492931604, + 0.007512773387134075, + 0.00436365744099021, + 0.00887101236730814, + 0.009978328831493855, + 0.0058183809742331505, + 0.010706722736358643, + 0.008032041601836681, + 0.008201118558645248, + 0.013518419116735458, + 0.009023602120578289, + 0.009167342446744442, + 0.005973408464342356, + 0.0031941235065460205, + 0.006331663578748703, + 0.002897245343774557, + 0.004269514698535204, + 0.0062590474262833595, + 0.003588633378967643, + 0.005395298823714256, + 0.0030041527934372425, + 0.0022844995837658644, + 0.004239436239004135, + 0.0023822709918022156, + 0.00336945871822536, + 0.0035194410011172295, + 0.0024682257790118456, + 0.0038693544920533895, + 0.014425452798604965, + 0.009408998303115368, + 0.02754659578204155, + 0.008453797549009323, + 0.010693132877349854, + 0.015179111622273922, + 0.009080158546566963, + 0.023882905021309853, + 0.005571823567152023, + 0.008510750718414783, + 0.005634707398712635, + 0.007336297072470188, + 0.018957594409585, + 0.008506936021149158, + 0.005762663669884205, + 0.012064620852470398, + 0.016575461253523827, + 0.010300825349986553, + 0.011652340181171894, + 0.007024839986115694, + 0.010740986093878746, + 0.017259176820516586, + 0.007485777139663696, + 0.025227516889572144, + 0.07993435114622116, + 0.012196054682135582, + 0.013104017823934555, + 0.006871726363897324, + 0.01319956686347723, + 0.12146750092506409, + 0.00876005832105875, + 0.05107489973306656, + 0.015420865267515182, + 0.03277067467570305, + 0.01100339088588953, + 0.0151781952008605, + 0.021235713735222816, + 0.02850845828652382, + 0.009640907868742943, + 0.013635830022394657, + 0.0032317289151251316, + 0.008300917223095894, + 0.00613888306543231, + 0.004133215639740229, + 0.018315456807613373, + 0.003821582067757845, + 0.005067937541753054, + 0.005219436250627041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.20782747864723206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761353373527527 + ], + "opposite_logprob_tokens": [ + -0.8261353373527527 + ], + "logprob_reward": -0.5761353373527527, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761353373527527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11483623832464218, + 0.139005646109581, + 0.04876324161887169, + 0.03851412609219551, + 0.12807126343250275, + 0.03997042402625084, + 0.042547762393951416, + 0.19809262454509735, + 0.17918266355991364, + 0.07101599872112274 + ], + "credit_map": [ + 0.012462736107409, + 0.007705990225076675, + 0.026302825659513474, + 0.009342600591480732, + 0.012716971337795258, + 0.018823670223355293, + 0.008392253890633583, + 0.019089192152023315, + 0.019077494740486145, + 0.010117482393980026, + 0.023073848336935043, + 0.018051497638225555, + 0.014653915539383888, + 0.02120564877986908, + 0.016215072944760323, + 0.016610680148005486, + 0.007115103304386139, + 0.004514436703175306, + 0.007402446586638689, + 0.005067192483693361, + 0.006667762063443661, + 0.00706199137493968, + 0.005153728183358908, + 0.0057805790565907955, + 0.004823199473321438, + 0.004691277630627155, + 0.00631954800337553, + 0.002717109862715006, + 0.004847563803195953, + 0.0053150849416852, + 0.003861644072458148, + 0.005938698537647724, + 0.012551623396575451, + 0.015610996633768082, + 0.017881326377391815, + 0.01675901561975479, + 0.016259044408798218, + 0.01523654442280531, + 0.02095457911491394, + 0.01281813532114029, + 0.003919069189578295, + 0.005331205669790506, + 0.004581757355481386, + 0.0035008529666811228, + 0.00828730408102274, + 0.004673822317272425, + 0.00363657483831048, + 0.006039837375283241, + 0.00650605745613575, + 0.004922495689243078, + 0.004470780026167631, + 0.0031353184022009373, + 0.004894200712442398, + 0.005461391061544418, + 0.004004810471087694, + 0.009152708575129509, + 0.05235271900892258, + 0.007460267283022404, + 0.010193580761551857, + 0.004993770737200975, + 0.0076417988166213036, + 0.07688506692647934, + 0.005534899886697531, + 0.033030521124601364, + 0.019385453313589096, + 0.029061296954751015, + 0.019343357533216476, + 0.01881646364927292, + 0.024714872241020203, + 0.03669513016939163, + 0.01354198157787323, + 0.017624128609895706, + 0.004736603237688541, + 0.012135152705013752, + 0.008748908527195454, + 0.004572286736220121, + 0.02737005241215229, + 0.004536075983196497, + 0.004585521761327982, + 0.004331397823989391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.5761353373527527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978860817849636 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.00978860817849636, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978860817849636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08172225952148438, + 0.0974714532494545, + 0.061749909073114395, + 0.036403801292181015, + 0.09929579496383667, + 0.03611589968204498, + 0.169862300157547, + 0.348209023475647, + 0.027824316173791885, + 0.041345298290252686 + ], + "credit_map": [ + 0.009669767692685127, + 0.003638578811660409, + 0.02835700660943985, + 0.004343579988926649, + 0.005337853915989399, + 0.01218399964272976, + 0.004593220539391041, + 0.01359824649989605, + 0.008052760735154152, + 0.010846142657101154, + 0.013494131155312061, + 0.010926937684416771, + 0.016805769875645638, + 0.012153933756053448, + 0.01602274365723133, + 0.009169030003249645, + 0.005813311785459518, + 0.007468970492482185, + 0.007336518261581659, + 0.006558259949088097, + 0.008767046965658665, + 0.009539167396724224, + 0.0092548169195652, + 0.007011813577264547, + 0.005213335622102022, + 0.002632755320519209, + 0.008867112919688225, + 0.002209789352491498, + 0.0035524489358067513, + 0.0049735470674932, + 0.002302802400663495, + 0.0066520050168037415, + 0.011470946483314037, + 0.010988022200763226, + 0.016568422317504883, + 0.010214676149189472, + 0.010997893288731575, + 0.013143762946128845, + 0.01494495291262865, + 0.010967114008963108, + 0.0031616007909178734, + 0.004892656113952398, + 0.00396203575655818, + 0.0036678302567452192, + 0.007462581619620323, + 0.004607758484780788, + 0.0031240996904671192, + 0.0052373348735272884, + 0.026049751788377762, + 0.016436737030744553, + 0.02077801525592804, + 0.012867596931755543, + 0.019159244373440742, + 0.02355203591287136, + 0.013029373250901699, + 0.037989530712366104, + 0.09173688292503357, + 0.013798908330500126, + 0.015078100375831127, + 0.007428862154483795, + 0.015213477425277233, + 0.13829168677330017, + 0.009691732935607433, + 0.05696934834122658, + 0.0028552638832479715, + 0.0047983210533857346, + 0.0036028206814080477, + 0.0034417614806443453, + 0.0046554505825042725, + 0.002435475355014205, + 0.0035926473792642355, + 0.002442574594169855, + 0.0025082388892769814, + 0.007389867678284645, + 0.0039150407537817955, + 0.0026157591491937637, + 0.015633516013622284, + 0.0027038399130105972, + 0.0030871760100126266, + 0.0034918568562716246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.00978860817849636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196592383086681 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196592383086681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196592383086681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07403697073459625, + 0.06312227994203568, + 0.04501412436366081, + 0.02211039513349533, + 0.1766960322856903, + 0.10580971837043762, + 0.12566909193992615, + 0.13395611941814423, + 0.046570487320423126, + 0.2070147544145584 + ], + "credit_map": [ + 0.009016929194331169, + 0.002954827155917883, + 0.02858094498515129, + 0.0038302738685160875, + 0.004453537054359913, + 0.011283252388238907, + 0.0037555352319031954, + 0.010161670856177807, + 0.006272246595472097, + 0.006612842436879873, + 0.007801638916134834, + 0.007746991701424122, + 0.010021999478340149, + 0.007454242557287216, + 0.008507225662469864, + 0.008705096319317818, + 0.0053449971601367, + 0.003480383427813649, + 0.006575519219040871, + 0.003495819866657257, + 0.006934888660907745, + 0.0073367999866604805, + 0.00654236925765872, + 0.005303344689309597, + 0.0025768675841391087, + 0.002547543030232191, + 0.0027502132579684258, + 0.00234935968182981, + 0.0034867580980062485, + 0.0029549142345786095, + 0.002801477210596204, + 0.002643261570483446, + 0.02335653454065323, + 0.010028858669102192, + 0.03851032257080078, + 0.011285762302577496, + 0.01066640019416809, + 0.027919866144657135, + 0.010690249502658844, + 0.04423803463578224, + 0.008075624704360962, + 0.012318816967308521, + 0.008759250864386559, + 0.008899563923478127, + 0.029144393280148506, + 0.012896673753857613, + 0.01087843906134367, + 0.014836952090263367, + 0.02071145549416542, + 0.011897602118551731, + 0.01369170006364584, + 0.008215507492423058, + 0.014179333113133907, + 0.021108487620949745, + 0.008660613559186459, + 0.02720438502728939, + 0.03626532480120659, + 0.004841523710638285, + 0.00562824122607708, + 0.002842912683263421, + 0.00505846505984664, + 0.05335625633597374, + 0.00403602933511138, + 0.02192736230790615, + 0.004909247159957886, + 0.009561807848513126, + 0.004453603643923998, + 0.0042277490720152855, + 0.009196055121719837, + 0.006063245702534914, + 0.0038746073842048645, + 0.004284174647182226, + 0.01437978446483612, + 0.035163603723049164, + 0.020979400724172592, + 0.0100539680570364, + 0.08709774166345596, + 0.012610731646418571, + 0.014678820967674255, + 0.012050707824528217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.014196592383086681 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539120391011238 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012539120391011238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539120391011238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061233039945364, + 0.04233579710125923, + 0.02241002768278122, + 0.016035988926887512, + 0.34900951385498047, + 0.034436699002981186, + 0.20479507744312286, + 0.21223773062229156, + 0.014626062475144863, + 0.04288002476096153 + ], + "credit_map": [ + 0.0064654042944312096, + 0.004678299184888601, + 0.013449269346892834, + 0.004761912394315004, + 0.005426350515335798, + 0.00918655376881361, + 0.00463690422475338, + 0.012628347612917423, + 0.0036682370118796825, + 0.0041458094492554665, + 0.0064873285591602325, + 0.005018331576138735, + 0.006969603709876537, + 0.005146404728293419, + 0.005271212197840214, + 0.005628867540508509, + 0.0026433130260556936, + 0.002253475831821561, + 0.003906698897480965, + 0.0020540824625641108, + 0.003054837230592966, + 0.0032321219332516193, + 0.0018207168905064464, + 0.0034447801299393177, + 0.0020248445216566324, + 0.0012688806746155024, + 0.003426775336265564, + 0.0010813503758981824, + 0.0020186954643577337, + 0.0022597957868129015, + 0.0011241715401411057, + 0.002831474645063281, + 0.04454641044139862, + 0.024616826325654984, + 0.08245780318975449, + 0.023473480716347694, + 0.0295086856931448, + 0.0468820184469223, + 0.023815475404262543, + 0.07370883971452713, + 0.002835102379322052, + 0.004666794557124376, + 0.0029607629403471947, + 0.0033395225182175636, + 0.009081130847334862, + 0.004238225053995848, + 0.003453426528722048, + 0.0038617367390543222, + 0.033140044659376144, + 0.019732100889086723, + 0.017599666491150856, + 0.016086269170045853, + 0.02521158568561077, + 0.04125341400504112, + 0.01349664106965065, + 0.038275353610515594, + 0.05489635840058327, + 0.009761528111994267, + 0.009640593081712723, + 0.005025472026318312, + 0.010459206067025661, + 0.08202522993087769, + 0.006451872643083334, + 0.03397747501730919, + 0.0017561055719852448, + 0.0013786527561023831, + 0.0016031041741371155, + 0.002354363212361932, + 0.0019325551111251116, + 0.002057276200503111, + 0.0019222334958612919, + 0.0016217728843912482, + 0.003324690042063594, + 0.005878317169845104, + 0.003508873749524355, + 0.0034516719169914722, + 0.015132665634155273, + 0.0031160530634224415, + 0.004750702064484358, + 0.003717047395184636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.012539120391011238 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up shot from an overhead angle, the camera slowly but steadily pans forward, focusing on a beast composed entirely of fire. This lion-like behemoth is in the grip of a silent yet explosive roar. Its body is composed entirely of flowing, incandescent magma and raging flames. Thick, powerful black outlines outline its taut, muscular form, imbued with a passionate, comic-book style. Its mane is a blazing, ever-brighter tongue of flame, soaring skyward. Its eyes are blazing white cores, radiating boundless fury. Its maw opens to its limit, revealing translucent flaming fangs condensed from pure energy. Simultaneously, its sharp front claws, formed from twisted flames, snap open, their tips gleaming with arcing sparks. The beast is the sole source of light, standing atop a crumbling, cracked black stone floor, from which strands of red energy permeate the cracks. Behind it, a deep, chaotic dark background is dotted with flying sparks and ripples of air distorted by intense heat. The dominant tones of the entire scene are scorching orange, crimson, and bright yellow, creating a high-contrast visual impact against the pure black of the background, creating an extreme atmosphere filled with overwhelming power, fury, and explosive energy.", + "seed": 11411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11411_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0489318392239511, + "vlm_logprob": -0.0489318392239511, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.0489318392239511, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041393768042325974 + ], + "opposite_logprob_tokens": [ + -3.416393756866455 + ], + "logprob_reward": -0.041393768042325974, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.041393768042325974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1224360391497612, + 0.06228208169341087, + 0.09277931600809097, + 0.09385120868682861, + 0.03619905188679695, + 0.12788797914981842, + 0.11813170462846756, + 0.12389053404331207, + 0.05166936665773392, + 0.17087267339229584 + ], + "credit_map": [ + 0.00823198538273573, + 0.01559498906135559, + 0.010721386410295963, + 0.009481626562774181, + 0.012567512691020966, + 0.02418198063969612, + 0.022123923525214195, + 0.019532639533281326, + 0.0059653683565557, + 0.00965966284275055, + 0.00438818521797657, + 0.00901842676103115, + 0.007778304163366556, + 0.009368124417960644, + 0.008368209935724735, + 0.007735801860690117, + 0.011588575318455696, + 0.01280707772821188, + 0.010532336309552193, + 0.011066662147641182, + 0.010975443758070469, + 0.011349386535584927, + 0.012578773312270641, + 0.011881056241691113, + 0.009931600652635098, + 0.01590799167752266, + 0.010804734192788601, + 0.010123737156391144, + 0.008269352838397026, + 0.011648532003164291, + 0.008928229101002216, + 0.018237030133605003, + 0.0038781228940933943, + 0.003411073237657547, + 0.003150308970361948, + 0.004051502328366041, + 0.0042227511294186115, + 0.004768868442624807, + 0.005699717905372381, + 0.007016704883426428, + 0.014539064839482307, + 0.01596834696829319, + 0.015540854074060917, + 0.012493861839175224, + 0.02039562724530697, + 0.014986895956099033, + 0.014879723079502583, + 0.01908360980451107, + 0.01257454976439476, + 0.016004787757992744, + 0.011496704071760178, + 0.014288428239524364, + 0.017811337485909462, + 0.01430678740143776, + 0.013602250255644321, + 0.01804685965180397, + 0.013675320893526077, + 0.01600281149148941, + 0.020259279757738113, + 0.013566303998231888, + 0.015138238668441772, + 0.011524476110935211, + 0.016193559393286705, + 0.017530551180243492, + 0.004686553031206131, + 0.004244989715516567, + 0.008746500127017498, + 0.005348691716790199, + 0.005932002328336239, + 0.004417811054736376, + 0.011759020388126373, + 0.0065337978303432465, + 0.021984610706567764, + 0.043626196682453156, + 0.014752046205103397, + 0.04134836047887802, + 0.01376586128026247, + 0.010792963206768036, + 0.014837333932518959, + 0.009765318594872952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.041393768042325974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989283800125122 + ], + "opposite_logprob_tokens": [ + -1.8048927783966064 + ], + "logprob_reward": -0.17989283800125122, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989283800125122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053374286741018295, + 0.0667843446135521, + 0.07376214861869812, + 0.08111684769392014, + 0.04645920172333717, + 0.0628577470779419, + 0.22831900417804718, + 0.14161109924316406, + 0.06556253135204315, + 0.18015283346176147 + ], + "credit_map": [ + 0.004906658083200455, + 0.005800710991024971, + 0.005126268602907658, + 0.005399766843765974, + 0.005856626201421022, + 0.00568791851401329, + 0.00958007387816906, + 0.011016265489161015, + 0.005952600389719009, + 0.007228946313261986, + 0.004912774544209242, + 0.005338240880519152, + 0.012988378293812275, + 0.009005039930343628, + 0.009523957967758179, + 0.011834407225251198, + 0.007650952320545912, + 0.007237638346850872, + 0.007592439651489258, + 0.006285194307565689, + 0.006374727934598923, + 0.009483005851507187, + 0.009017879143357277, + 0.02012030966579914, + 0.007609674707055092, + 0.010349126532673836, + 0.008354558609426022, + 0.00775544298812747, + 0.00715820724144578, + 0.011799830012023449, + 0.007629201747477055, + 0.020460795611143112, + 0.003663604613393545, + 0.0054781935177743435, + 0.004018363542854786, + 0.005122349597513676, + 0.00440222630277276, + 0.006040663458406925, + 0.008060296066105366, + 0.009673506021499634, + 0.0074556805193424225, + 0.010793517343699932, + 0.008092432282865047, + 0.008058582432568073, + 0.01066597830504179, + 0.006012680009007454, + 0.006018212530761957, + 0.005760667845606804, + 0.02572418935596943, + 0.02121267095208168, + 0.039068304002285004, + 0.03370983153581619, + 0.029216913506388664, + 0.02279805578291416, + 0.03077523410320282, + 0.025813786312937737, + 0.016354430466890335, + 0.020495198667049408, + 0.0321529358625412, + 0.011694110929965973, + 0.009893493726849556, + 0.014295455999672413, + 0.018414566293358803, + 0.01831090822815895, + 0.0057324133813381195, + 0.005924510769546032, + 0.007643015589565039, + 0.006502896547317505, + 0.00883127748966217, + 0.007010388653725386, + 0.012224065139889717, + 0.011693961918354034, + 0.016870027408003807, + 0.02831086702644825, + 0.023734671995043755, + 0.022186126559972763, + 0.025275874882936478, + 0.014471085742115974, + 0.03413571044802666, + 0.015168465673923492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.17989283800125122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09512169659137726, + 0.07030772417783737, + 0.04479110985994339, + 0.04655764624476433, + 0.10440699756145477, + 0.12572279572486877, + 0.17371484637260437, + 0.09183743596076965, + 0.052416734397411346, + 0.19512301683425903 + ], + "credit_map": [ + 0.006869640666991472, + 0.011168604716658592, + 0.007427077274769545, + 0.006905205547809601, + 0.012492704205214977, + 0.017738932743668556, + 0.012316621840000153, + 0.02020290307700634, + 0.0059588453732430935, + 0.006975087802857161, + 0.007521974388509989, + 0.0075774784199893475, + 0.009354637004435062, + 0.009454233571887016, + 0.011975650675594807, + 0.01148981787264347, + 0.004768624901771545, + 0.005593124311417341, + 0.005648157559335232, + 0.004084943328052759, + 0.004481662064790726, + 0.0055813719518482685, + 0.005747671239078045, + 0.008885557763278484, + 0.0039998251013457775, + 0.0075117116793990135, + 0.0044651273638010025, + 0.0055038221180438995, + 0.0060279169119894505, + 0.005755435675382614, + 0.004232877865433693, + 0.009060928598046303, + 0.008356400765478611, + 0.008791700936853886, + 0.007995194755494595, + 0.011584028601646423, + 0.009413395076990128, + 0.014914801344275475, + 0.018126020208001137, + 0.025225454941391945, + 0.012362091802060604, + 0.013137212023139, + 0.01804896630346775, + 0.014646883122622967, + 0.009567683562636375, + 0.017763612791895866, + 0.01607692241668701, + 0.02411942183971405, + 0.02019701898097992, + 0.02132878266274929, + 0.03412088751792908, + 0.019053621217608452, + 0.023906521499156952, + 0.01744106411933899, + 0.018605677410960197, + 0.01906127668917179, + 0.009364062920212746, + 0.010568500496447086, + 0.013561584986746311, + 0.01054974552243948, + 0.008149238303303719, + 0.011369203217327595, + 0.01289482694119215, + 0.015380275435745716, + 0.00594301987439394, + 0.006720745004713535, + 0.006987132132053375, + 0.005462069995701313, + 0.004701628349721432, + 0.006112347822636366, + 0.007619749289005995, + 0.008870040997862816, + 0.01793546788394451, + 0.03251960501074791, + 0.016606301069259644, + 0.04762609302997589, + 0.025177327916026115, + 0.012823452241718769, + 0.029878702014684677, + 0.012556062079966068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537237256765366 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537237256765366, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537237256765366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14600074291229248, + 0.1351090520620346, + 0.06329397112131119, + 0.062032781541347504, + 0.08776303380727768, + 0.061775851994752884, + 0.16416418552398682, + 0.09270276874303818, + 0.0678623616695404, + 0.11929525434970856 + ], + "credit_map": [ + 0.012606032192707062, + 0.012577940709888935, + 0.012213725596666336, + 0.012041335925459862, + 0.01992008090019226, + 0.030617041513323784, + 0.01816132478415966, + 0.02786327339708805, + 0.01187221147119999, + 0.011812688782811165, + 0.009500212036073208, + 0.01563205011188984, + 0.015984388068318367, + 0.01817876659333706, + 0.019532237201929092, + 0.032596491277217865, + 0.006651407107710838, + 0.008849944919347763, + 0.006369559094309807, + 0.007180421147495508, + 0.007200533989816904, + 0.007803987246006727, + 0.009320017881691456, + 0.00991810392588377, + 0.007298918906599283, + 0.0076590473763644695, + 0.008242973126471043, + 0.007011826150119305, + 0.005259019788354635, + 0.009001399390399456, + 0.007584365550428629, + 0.009975230321288109, + 0.008030098862946033, + 0.014127957634627819, + 0.008418322540819645, + 0.01401306502521038, + 0.01148128416389227, + 0.009249012917280197, + 0.012313472107052803, + 0.010129819624125957, + 0.008285919204354286, + 0.007508906070142984, + 0.011062159202992916, + 0.006215072702616453, + 0.007848713546991348, + 0.006771788001060486, + 0.007763749919831753, + 0.006319545675069094, + 0.020622117444872856, + 0.017287716269493103, + 0.029980605468153954, + 0.019656730815768242, + 0.02188972197473049, + 0.017004668712615967, + 0.017874954268336296, + 0.019847668707370758, + 0.010763644240796566, + 0.008955944329500198, + 0.018773773685097694, + 0.009308476001024246, + 0.011297674849629402, + 0.008937263861298561, + 0.014371722936630249, + 0.010294261388480663, + 0.006429262459278107, + 0.007216278463602066, + 0.00741646159440279, + 0.0086772246286273, + 0.007161167915910482, + 0.005899458192288876, + 0.016643118113279343, + 0.00841938890516758, + 0.015471029095351696, + 0.022593820467591286, + 0.012317459098994732, + 0.015710365027189255, + 0.01427464745938778, + 0.012490615248680115, + 0.013911915943026543, + 0.012525403872132301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.012537237256765366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14115580916404724, + 0.041505664587020874, + 0.038944028317928314, + 0.05321547016501427, + 0.06303355097770691, + 0.4088818430900574, + 0.11626648902893066, + 0.02953716553747654, + 0.027630094438791275, + 0.07982996106147766 + ], + "credit_map": [ + 0.011640641838312149, + 0.012234529480338097, + 0.017229080200195312, + 0.011476467363536358, + 0.014792823232710361, + 0.016002727672457695, + 0.017603600397706032, + 0.04017592966556549, + 0.0032311829272657633, + 0.003973017912358046, + 0.002824140712618828, + 0.004555145278573036, + 0.007130580022931099, + 0.005106854252517223, + 0.0057958257384598255, + 0.008888914249837399, + 0.0035736423451453447, + 0.00635928800329566, + 0.0034124834928661585, + 0.0033981858287006617, + 0.005524619948118925, + 0.0050336262211203575, + 0.004889759235084057, + 0.006752421148121357, + 0.004406818654388189, + 0.004883888177573681, + 0.006554936990141869, + 0.004642036277800798, + 0.0038977654185146093, + 0.008619106374680996, + 0.004746341612190008, + 0.015464571304619312, + 0.006715643219649792, + 0.007234609220176935, + 0.00524344015866518, + 0.00880961213260889, + 0.006276065483689308, + 0.007333171553909779, + 0.010713845491409302, + 0.010707157664000988, + 0.03992386907339096, + 0.033561449497938156, + 0.05564373359084129, + 0.0419946126639843, + 0.04098183289170265, + 0.059514258056879044, + 0.0557391494512558, + 0.08152289688587189, + 0.012795859947800636, + 0.014353126287460327, + 0.023325515910983086, + 0.015217551961541176, + 0.015424150042235851, + 0.011745251715183258, + 0.012410934083163738, + 0.01099409069865942, + 0.003997153136879206, + 0.003588789375498891, + 0.00502941245213151, + 0.002715348731726408, + 0.0027688671834766865, + 0.0036595831625163555, + 0.0033003573771566153, + 0.004477653186768293, + 0.002630148082971573, + 0.003249562345445156, + 0.0030658990144729614, + 0.003369240090250969, + 0.003824748331680894, + 0.002295228885486722, + 0.006168209481984377, + 0.0030270563438534737, + 0.007488425821065903, + 0.014448333531618118, + 0.011539200320839882, + 0.02134544588625431, + 0.007492364384233952, + 0.005300234537571669, + 0.005368346348404884, + 0.006847603712230921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0040924143977463245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up shot from an overhead angle, the camera slowly but steadily pans forward, focusing on a beast composed entirely of fire. This lion-like behemoth is in the grip of a silent yet explosive roar. Its body is composed entirely of flowing, incandescent magma and raging flames. Thick, powerful black outlines outline its taut, muscular form, imbued with a passionate, comic-book style. Its mane is a blazing, ever-brighter tongue of flame, soaring skyward. Its eyes are blazing white cores, radiating boundless fury. Its maw opens to its limit, revealing translucent flaming fangs condensed from pure energy. Simultaneously, its sharp front claws, formed from twisted flames, snap open, their tips gleaming with arcing sparks. The beast is the sole source of light, standing atop a crumbling, cracked black stone floor, from which strands of red energy permeate the cracks. Behind it, a deep, chaotic dark background is dotted with flying sparks and ripples of air distorted by intense heat. The dominant tones of the entire scene are scorching orange, crimson, and bright yellow, creating a high-contrast visual impact against the pure black of the background, creating an extreme atmosphere filled with overwhelming power, fury, and explosive energy.", + "seed": 11412, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11412_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08180660270154476, + "vlm_logprob": -0.08180660270154476, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.08180660270154476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09647755324840546 + ], + "opposite_logprob_tokens": [ + -2.471477508544922 + ], + "logprob_reward": -0.09647755324840546, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09647755324840546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24197699129581451, + 0.09638696908950806, + 0.08122498542070389, + 0.11083173751831055, + 0.038490455597639084, + 0.043553151190280914, + 0.10418792814016342, + 0.16221751272678375, + 0.06586868315935135, + 0.05526158958673477 + ], + "credit_map": [ + 0.0269208624958992, + 0.03256106749176979, + 0.035478923469781876, + 0.030893594026565552, + 0.025106744840741158, + 0.027975402772426605, + 0.03272487595677376, + 0.030315514653921127, + 0.010739831253886223, + 0.008238336071372032, + 0.01670503243803978, + 0.011657225899398327, + 0.006395151372998953, + 0.010363729670643806, + 0.018661165609955788, + 0.013626503758132458, + 0.011725805699825287, + 0.006929940078407526, + 0.012573756277561188, + 0.00835486501455307, + 0.00487273745238781, + 0.011599374935030937, + 0.010746912099421024, + 0.014421590603888035, + 0.011905125342309475, + 0.011969107203185558, + 0.013413859531283379, + 0.01615007407963276, + 0.009862018749117851, + 0.01139264740049839, + 0.028452899307012558, + 0.007686004042625427, + 0.005644928198307753, + 0.004258118569850922, + 0.005630601197481155, + 0.0038487918209284544, + 0.0041710990481078625, + 0.0046788048930466175, + 0.0036171767860651016, + 0.006640932988375425, + 0.005556175485253334, + 0.004961758852005005, + 0.005958267953246832, + 0.005372983403503895, + 0.00456335349008441, + 0.005635941866785288, + 0.005910622421652079, + 0.005594046786427498, + 0.010800358839333057, + 0.011053882539272308, + 0.016296960413455963, + 0.012608950026333332, + 0.010245300829410553, + 0.009481457993388176, + 0.024975761771202087, + 0.008725257590413094, + 0.013885878026485443, + 0.010855423286557198, + 0.04819590598344803, + 0.021170703694224358, + 0.014030597172677517, + 0.01780165173113346, + 0.020046815276145935, + 0.016230542212724686, + 0.005103468429297209, + 0.008472355082631111, + 0.010540671646595001, + 0.010818599723279476, + 0.006063522771000862, + 0.007436717394739389, + 0.009470748715102673, + 0.007962603121995926, + 0.00454761041328311, + 0.005394854117184877, + 0.008017080835998058, + 0.006655348930507898, + 0.0064728823490440845, + 0.005882358178496361, + 0.00975483562797308, + 0.008536618202924728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.09647755324840546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813360095024109 + ], + "opposite_logprob_tokens": [ + -1.4063360691070557 + ], + "logprob_reward": -0.2813360095024109, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813360095024109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12717173993587494, + 0.07807662338018417, + 0.0891382247209549, + 0.058366820216178894, + 0.07384870201349258, + 0.04122946411371231, + 0.07014389336109161, + 0.36575114727020264, + 0.030277298763394356, + 0.06599606573581696 + ], + "credit_map": [ + 0.021829266101121902, + 0.015685226768255234, + 0.016889411956071854, + 0.014622669667005539, + 0.014438243582844734, + 0.020816072821617126, + 0.011194300837814808, + 0.011696555651724339, + 0.009728322736918926, + 0.011325763538479805, + 0.010843931697309017, + 0.010411320254206657, + 0.01071145385503769, + 0.009910667315125465, + 0.008252138271927834, + 0.006893027573823929, + 0.012143155559897423, + 0.006249954923987389, + 0.01232613530009985, + 0.007163350936025381, + 0.005002065561711788, + 0.011909638531506062, + 0.008264073170721531, + 0.026079846546053886, + 0.00812232680618763, + 0.006178716197609901, + 0.009652524255216122, + 0.006746733095496893, + 0.0035427953116595745, + 0.007496845908463001, + 0.006572023034095764, + 0.010054858401417732, + 0.01300404779613018, + 0.0064281863160431385, + 0.010036895982921124, + 0.008673286996781826, + 0.005926790181547403, + 0.01241710502654314, + 0.009830090217292309, + 0.007532297167927027, + 0.006431687157601118, + 0.004522831179201603, + 0.004362887237221003, + 0.004425998777151108, + 0.005067563150078058, + 0.005077493377029896, + 0.0058326637372374535, + 0.005508342757821083, + 0.006177907343953848, + 0.007358104921877384, + 0.011994928121566772, + 0.009266439825296402, + 0.006603146903216839, + 0.006517152301967144, + 0.01563582755625248, + 0.00659038033336401, + 0.03492458537220955, + 0.03108792006969452, + 0.10532984882593155, + 0.042618248611688614, + 0.0364500917494297, + 0.044796332716941833, + 0.03876617178320885, + 0.031777940690517426, + 0.003464051987975836, + 0.003617681097239256, + 0.005138792097568512, + 0.003921082708984613, + 0.0037085760850459337, + 0.0036477975081652403, + 0.0037919559981673956, + 0.0029873603489249945, + 0.007130531128495932, + 0.009537621401250362, + 0.009874807670712471, + 0.008715040050446987, + 0.00798353087157011, + 0.004985901061445475, + 0.008579973131418228, + 0.009188663214445114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.2813360095024109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081107892096043 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081107892096043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081107892096043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21586646139621735, + 0.11473774909973145, + 0.11215310543775558, + 0.039139725267887115, + 0.08015187829732895, + 0.049007344990968704, + 0.07582566142082214, + 0.1957833468914032, + 0.037665385752916336, + 0.07966936379671097 + ], + "credit_map": [ + 0.032186780124902725, + 0.01815793663263321, + 0.03214513510465622, + 0.023467037826776505, + 0.013778946362435818, + 0.03418084606528282, + 0.04115033522248268, + 0.020799441263079643, + 0.012504815123975277, + 0.011463381350040436, + 0.017005829140543938, + 0.013291720300912857, + 0.00796628836542368, + 0.014471677131950855, + 0.01540729496628046, + 0.022626737132668495, + 0.021068725734949112, + 0.006851703394204378, + 0.011128151789307594, + 0.007860617712140083, + 0.005638693459331989, + 0.0171910859644413, + 0.011357799172401428, + 0.031056329607963562, + 0.006674160249531269, + 0.0026138313114643097, + 0.006674195174127817, + 0.0031728926114737988, + 0.002340570092201233, + 0.004793746862560511, + 0.007449550088495016, + 0.005420778412371874, + 0.013076776638627052, + 0.006358990911394358, + 0.010554133914411068, + 0.010068641044199467, + 0.003991738893091679, + 0.0111690117046237, + 0.01163304504007101, + 0.013299538753926754, + 0.008306728675961494, + 0.005622113589197397, + 0.0057578301057219505, + 0.004449565429240465, + 0.005896523594856262, + 0.007151318714022636, + 0.006279690656810999, + 0.005543570499867201, + 0.007844056002795696, + 0.0057218740694224834, + 0.011172998696565628, + 0.00993324164301157, + 0.0061068604700267315, + 0.007127820048481226, + 0.02066102996468544, + 0.00725778192281723, + 0.017453571781516075, + 0.01158281322568655, + 0.05750428885221481, + 0.024190010502934456, + 0.01890687271952629, + 0.020968392491340637, + 0.0242915116250515, + 0.020885884761810303, + 0.0028587328270077705, + 0.004210269078612328, + 0.004934192635118961, + 0.005445745773613453, + 0.0037144096568226814, + 0.005995366722345352, + 0.0058923992328345776, + 0.0046142702922225, + 0.005564356222748756, + 0.006209417711943388, + 0.008890938945114613, + 0.009980029426515102, + 0.012554190121591091, + 0.005001070909202099, + 0.019108664244413376, + 0.012360699474811554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.011081107892096043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191068708896637 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191068708896637, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191068708896637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2226797193288803, + 0.14379331469535828, + 0.06122341379523277, + 0.11342266201972961, + 0.04707450419664383, + 0.03875458613038063, + 0.06360594928264618, + 0.12936322391033173, + 0.07202760875225067, + 0.10805509239435196 + ], + "credit_map": [ + 0.031097039580345154, + 0.03397023677825928, + 0.029305793344974518, + 0.02188378944993019, + 0.02400565706193447, + 0.030730200931429863, + 0.03102334588766098, + 0.020663656294345856, + 0.016502123326063156, + 0.008508141152560711, + 0.01666186936199665, + 0.010965057648718357, + 0.0062994686886668205, + 0.01769919879734516, + 0.022653400897979736, + 0.044504038989543915, + 0.012036213651299477, + 0.005146373528987169, + 0.009779248386621475, + 0.005561786703765392, + 0.0028702584095299244, + 0.010170198045670986, + 0.0059871980920434, + 0.009672137908637524, + 0.01391332596540451, + 0.014555075205862522, + 0.01358327642083168, + 0.011686738580465317, + 0.010761751793324947, + 0.011266659013926983, + 0.025991979986429214, + 0.011663851328194141, + 0.005786587484180927, + 0.005078342277556658, + 0.008368698880076408, + 0.004989163484424353, + 0.004091773647814989, + 0.005408548284322023, + 0.004731395281851292, + 0.008619996719062328, + 0.005466675851494074, + 0.0037559957709163427, + 0.0050420439802110195, + 0.003992215730249882, + 0.004589760210365057, + 0.004788908176124096, + 0.006006268784403801, + 0.005112717859447002, + 0.006239084992557764, + 0.0047711217775940895, + 0.008885656483471394, + 0.008808368816971779, + 0.005163645837455988, + 0.005940570030361414, + 0.01827378384768963, + 0.005523716099560261, + 0.011326678097248077, + 0.009478997439146042, + 0.0317973718047142, + 0.014390035532414913, + 0.012716569937765598, + 0.01773531548678875, + 0.016032114624977112, + 0.015886131674051285, + 0.00861863698810339, + 0.009181088767945766, + 0.007837529294192791, + 0.008834191597998142, + 0.009901275858283043, + 0.007937395013868809, + 0.011647733859717846, + 0.00806975457817316, + 0.0074663301929831505, + 0.011141855269670486, + 0.01291990652680397, + 0.013115580193698406, + 0.018297815695405006, + 0.010012181475758553, + 0.02062392607331276, + 0.014477494172751904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.014191068708896637 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20879928767681122, + 0.12494771927595139, + 0.09353573620319366, + 0.08127213269472122, + 0.043854378163814545, + 0.07996682077646255, + 0.08159784972667694, + 0.11200332641601562, + 0.02745007909834385, + 0.14657269418239594 + ], + "credit_map": [ + 0.028631584718823433, + 0.023932062089443207, + 0.02854529581964016, + 0.030461840331554413, + 0.016078710556030273, + 0.026812516152858734, + 0.03433002904057503, + 0.020007245242595673, + 0.019514331594109535, + 0.00929603073745966, + 0.015600210055708885, + 0.008379455655813217, + 0.007448283024132252, + 0.018313296139240265, + 0.013329064473509789, + 0.03306705877184868, + 0.01787017472088337, + 0.009960616938769817, + 0.012783198617398739, + 0.010704901069402695, + 0.007254441734403372, + 0.012071890756487846, + 0.014350940473377705, + 0.008539577014744282, + 0.011886704713106155, + 0.008040805347263813, + 0.014367705211043358, + 0.009197845123708248, + 0.005725228693336248, + 0.009560639970004559, + 0.013374458998441696, + 0.009118737652897835, + 0.006823956035077572, + 0.005190966185182333, + 0.006517637986689806, + 0.004556657746434212, + 0.004183175042271614, + 0.005331133957952261, + 0.005511977709829807, + 0.005738875363022089, + 0.011614426970481873, + 0.0061186826787889, + 0.008837743662297726, + 0.010166028514504433, + 0.006889537908136845, + 0.011658457107841969, + 0.011217567138373852, + 0.013464372605085373, + 0.009948265738785267, + 0.007850639522075653, + 0.015307699330151081, + 0.009172046557068825, + 0.006329592317342758, + 0.008910560980439186, + 0.014907870441675186, + 0.009171174839138985, + 0.009626203216612339, + 0.0067711579613387585, + 0.02906312234699726, + 0.012734204530715942, + 0.010994336567819118, + 0.013300389982759953, + 0.014838675037026405, + 0.014675241895020008, + 0.00302161555737257, + 0.003511533373966813, + 0.003899895353242755, + 0.0034682557452470064, + 0.0019539371132850647, + 0.004516740795224905, + 0.004268713761121035, + 0.0028093867003917694, + 0.010526208207011223, + 0.017170170322060585, + 0.02183513157069683, + 0.016065865755081177, + 0.016108939424157143, + 0.016803348436951637, + 0.026009317487478256, + 0.02205372229218483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0059472741559147835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up shot from an overhead angle, the camera slowly but steadily pans forward, focusing on a beast composed entirely of fire. This lion-like behemoth is in the grip of a silent yet explosive roar. Its body is composed entirely of flowing, incandescent magma and raging flames. Thick, powerful black outlines outline its taut, muscular form, imbued with a passionate, comic-book style. Its mane is a blazing, ever-brighter tongue of flame, soaring skyward. Its eyes are blazing white cores, radiating boundless fury. Its maw opens to its limit, revealing translucent flaming fangs condensed from pure energy. Simultaneously, its sharp front claws, formed from twisted flames, snap open, their tips gleaming with arcing sparks. The beast is the sole source of light, standing atop a crumbling, cracked black stone floor, from which strands of red energy permeate the cracks. Behind it, a deep, chaotic dark background is dotted with flying sparks and ripples of air distorted by intense heat. The dominant tones of the entire scene are scorching orange, crimson, and bright yellow, creating a high-contrast visual impact against the pure black of the background, creating an extreme atmosphere filled with overwhelming power, fury, and explosive energy.", + "seed": 11413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_7/index_2634_seed_11413_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09541971115395427, + "vlm_logprob": -0.09541971115395427, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.09541971115395427, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13217534124851227 + ], + "opposite_logprob_tokens": [ + -2.1321754455566406 + ], + "logprob_reward": -0.13217534124851227, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13217534124851227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14014722406864166, + 0.11172502487897873, + 0.32194238901138306, + 0.03787447512149811, + 0.07866068184375763, + 0.09734432399272919, + 0.07248363643884659, + 0.031056886538863182, + 0.05863678827881813, + 0.050128523260354996 + ], + "credit_map": [ + 0.01315227709710598, + 0.013508707284927368, + 0.017984215170145035, + 0.01980539970099926, + 0.02464972622692585, + 0.011973309330642223, + 0.02045268937945366, + 0.01862090826034546, + 0.0074586644768714905, + 0.01707209274172783, + 0.010893098078668118, + 0.017154434695839882, + 0.022724661976099014, + 0.009162516333162785, + 0.010929721407592297, + 0.01632983796298504, + 0.021549472585320473, + 0.014019472524523735, + 0.02887878194451332, + 0.09807340055704117, + 0.014547065831720829, + 0.08648445457220078, + 0.03753991425037384, + 0.020849809050559998, + 0.004642099142074585, + 0.0028378956485539675, + 0.0058394731022417545, + 0.008129176683723927, + 0.005769710056483746, + 0.0031348024494946003, + 0.0044751656241714954, + 0.0030461549758911133, + 0.014704477973282337, + 0.007138220127671957, + 0.012265385128557682, + 0.009689786471426487, + 0.006663763429969549, + 0.012225967831909657, + 0.007611087989062071, + 0.008361992426216602, + 0.00935755018144846, + 0.007548018358647823, + 0.023166080936789513, + 0.023150784894824028, + 0.007906650193035603, + 0.009065445512533188, + 0.010635766200721264, + 0.006514023523777723, + 0.0036278204061090946, + 0.011560210958123207, + 0.006322378758341074, + 0.00812311377376318, + 0.018433529883623123, + 0.009372886270284653, + 0.00859201792627573, + 0.0064516752026975155, + 0.002444290556013584, + 0.0034009632654488087, + 0.0026571948546916246, + 0.004210378043353558, + 0.005437307991087437, + 0.004481212701648474, + 0.004142189398407936, + 0.004283351823687553, + 0.004297490231692791, + 0.008160348981618881, + 0.009223053231835365, + 0.006990301888436079, + 0.011917877942323685, + 0.003079250454902649, + 0.007307811174541712, + 0.007660652045160532, + 0.0034243224654346704, + 0.007184087298810482, + 0.0037355211097747087, + 0.007983686402440071, + 0.007821342907845974, + 0.00374816358089447, + 0.011534362100064754, + 0.004697038792073727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.13217534124851227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134750425815582 + ], + "opposite_logprob_tokens": [ + -1.3134750127792358 + ], + "logprob_reward": -0.3134750425815582, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134750425815582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09460752457380295, + 0.11823750287294388, + 0.3377722203731537, + 0.07481905072927475, + 0.056713346391916275, + 0.04062920808792114, + 0.06027872487902641, + 0.07283978909254074, + 0.07318232208490372, + 0.07092027366161346 + ], + "credit_map": [ + 0.008664931170642376, + 0.009746194817125797, + 0.008825760334730148, + 0.016504324972629547, + 0.022941922768950462, + 0.006809598300606012, + 0.011245139874517918, + 0.009869658388197422, + 0.018417587503790855, + 0.012507028877735138, + 0.008488781750202179, + 0.013721720315515995, + 0.009640908800065517, + 0.01600624993443489, + 0.010368435643613338, + 0.029086800292134285, + 0.023825110867619514, + 0.015884611755609512, + 0.02558586560189724, + 0.10211282223463058, + 0.016906190663576126, + 0.09203691780567169, + 0.03864826634526253, + 0.022772476077079773, + 0.010052026249468327, + 0.007185396738350391, + 0.012798926793038845, + 0.008470830507576466, + 0.010864474810659885, + 0.008114973083138466, + 0.008883512578904629, + 0.008448916487395763, + 0.009163858368992805, + 0.006095048505812883, + 0.009054519236087799, + 0.009419292211532593, + 0.006360574159771204, + 0.004713570699095726, + 0.006107775028795004, + 0.005798711907118559, + 0.0033811444882303476, + 0.004018088337033987, + 0.00772918900474906, + 0.007263914216309786, + 0.005353859160095453, + 0.003643968142569065, + 0.005390028469264507, + 0.003849019994959235, + 0.0044754017144441605, + 0.006690634414553642, + 0.005251901224255562, + 0.009411110542714596, + 0.012861348688602448, + 0.004842537920922041, + 0.00773574085906148, + 0.009010056965053082, + 0.008739739656448364, + 0.007050576154142618, + 0.005825681146234274, + 0.019018813967704773, + 0.007712159305810928, + 0.011653066612780094, + 0.008528665639460087, + 0.004311094060540199, + 0.005142606794834137, + 0.011644470505416393, + 0.013062657788395882, + 0.008419782854616642, + 0.012055504135787487, + 0.004363635089248419, + 0.00813590083271265, + 0.010357772000133991, + 0.004858511500060558, + 0.012806457467377186, + 0.007966292090713978, + 0.012820705771446228, + 0.008877243846654892, + 0.006623881869018078, + 0.0083778603002429, + 0.008589329198002815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.3134750425815582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674246484413743 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.00674246484413743, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674246484413743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26002711057662964, + 0.11623673141002655, + 0.1603507399559021, + 0.14232416450977325, + 0.04591476544737816, + 0.04064419865608215, + 0.06088682636618614, + 0.08601222187280655, + 0.025978703051805496, + 0.06162446364760399 + ], + "credit_map": [ + 0.020963329821825027, + 0.03004922904074192, + 0.03411967307329178, + 0.026844944804906845, + 0.06556616723537445, + 0.021407071501016617, + 0.028253206983208656, + 0.03282349929213524, + 0.008444961160421371, + 0.016652874648571014, + 0.012261842377483845, + 0.01903126947581768, + 0.023051900789141655, + 0.008748350664973259, + 0.012580299749970436, + 0.015465235337615013, + 0.01147556770592928, + 0.014119153842329979, + 0.023850709199905396, + 0.046790797263383865, + 0.012072166427969933, + 0.022462159395217896, + 0.019251903519034386, + 0.010328284464776516, + 0.027385685592889786, + 0.010400712490081787, + 0.015428539365530014, + 0.04735409840941429, + 0.007252097129821777, + 0.00717437919229269, + 0.016970885917544365, + 0.010357768274843693, + 0.004680524580180645, + 0.006456122267991304, + 0.005600099451839924, + 0.007643522694706917, + 0.0058473385870456696, + 0.005776591133326292, + 0.0058049289509654045, + 0.004105638712644577, + 0.003084271913394332, + 0.005985255353152752, + 0.007159478962421417, + 0.006485858000814915, + 0.006576467305421829, + 0.0037335853558033705, + 0.004560848698019981, + 0.003058433998376131, + 0.0032449925784021616, + 0.00803790707141161, + 0.004033308941870928, + 0.008239207789301872, + 0.017498809844255447, + 0.004585865419358015, + 0.007408157456666231, + 0.00783857423812151, + 0.008381528779864311, + 0.01002285536378622, + 0.008946352638304234, + 0.017854804173111916, + 0.019840290769934654, + 0.004462457727640867, + 0.011323025450110435, + 0.005180912092328072, + 0.0019416998839005828, + 0.004885524045675993, + 0.0028159336652606726, + 0.003286229446530342, + 0.005244194995611906, + 0.0017380141653120518, + 0.003711545607075095, + 0.002355561126023531, + 0.007926619611680508, + 0.0086671132594347, + 0.005754299461841583, + 0.007461602799594402, + 0.0104610500857234, + 0.008475132286548615, + 0.005608708132058382, + 0.007269938010722399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.00674246484413743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1806337833404541, + 0.13831572234630585, + 0.25655826926231384, + 0.11536309868097305, + 0.05591263249516487, + 0.02636469341814518, + 0.02263232320547104, + 0.06539613753557205, + 0.05886569619178772, + 0.07995769381523132 + ], + "credit_map": [ + 0.015206909738481045, + 0.02133568748831749, + 0.017125502228736877, + 0.02261086367070675, + 0.05278465896844864, + 0.012140555307269096, + 0.01774521917104721, + 0.02168438583612442, + 0.021684108301997185, + 0.015723999589681625, + 0.011333930306136608, + 0.015556922182440758, + 0.023006385192275047, + 0.01244338694959879, + 0.012258816510438919, + 0.026308152824640274, + 0.018269676715135574, + 0.012558997608721256, + 0.023772159591317177, + 0.07757692039012909, + 0.012330001220107079, + 0.06488145142793655, + 0.029602786526083946, + 0.017566276714205742, + 0.02043461985886097, + 0.008375614881515503, + 0.010996835306286812, + 0.039119113236665726, + 0.009185639210045338, + 0.004928650334477425, + 0.01474681869149208, + 0.007575799711048603, + 0.011175954714417458, + 0.005250883288681507, + 0.006676308810710907, + 0.006011837627738714, + 0.006167685147374868, + 0.009240511804819107, + 0.0047949738800525665, + 0.006594477221369743, + 0.002175745787099004, + 0.003155709942802787, + 0.005385152529925108, + 0.003101135604083538, + 0.0042008594609797, + 0.002704304875805974, + 0.0034230330493301153, + 0.002218750538304448, + 0.0013241162523627281, + 0.003037119749933481, + 0.0026629262138158083, + 0.0030192823614925146, + 0.005648444872349501, + 0.0015322340186685324, + 0.0027233771979808807, + 0.0026848239358514547, + 0.003758182981982827, + 0.007156601641327143, + 0.0038693337701261044, + 0.01469211746007204, + 0.016073666512966156, + 0.005345573648810387, + 0.008925847709178925, + 0.00557481124997139, + 0.0047351298853755, + 0.010554655455052853, + 0.0053947241976857185, + 0.006627021357417107, + 0.012522279284894466, + 0.003161921165883541, + 0.007942022755742073, + 0.007927942089736462, + 0.008594987913966179, + 0.012767331674695015, + 0.006763246841728687, + 0.013373853638768196, + 0.012870080769062042, + 0.007500919979065657, + 0.011402762494981289, + 0.006684509571641684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.008638513274490833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a striking close-up from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the beast composed entirely of flowing magma and flames with thick black comic-book style outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast roaring with its maw fully open, revealing translucent flaming fangs and snapping front claws with arcing sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a crumbling black stone floor with red energy in the cracks against a dark background with flying sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21892692148685455, + 0.1041308343410492, + 0.3395579755306244, + 0.14446766674518585, + 0.06507518142461777, + 0.02218286134302616, + 0.025882961228489876, + 0.026236435398459435, + 0.025279007852077484, + 0.028260134160518646 + ], + "credit_map": [ + 0.017371879890561104, + 0.015566333197057247, + 0.03476153686642647, + 0.02635148912668228, + 0.02379440702497959, + 0.018945662304759026, + 0.0348685197532177, + 0.04726709425449371, + 0.009431557729840279, + 0.010572809725999832, + 0.01319810375571251, + 0.011653700843453407, + 0.015890732407569885, + 0.01803716830909252, + 0.009004132822155952, + 0.016342636197805405, + 0.02738347090780735, + 0.015258532017469406, + 0.03648749366402626, + 0.09592510014772415, + 0.016688412055373192, + 0.08648396283388138, + 0.03877337649464607, + 0.022557653486728668, + 0.02957666851580143, + 0.009864133782684803, + 0.013320419006049633, + 0.05371517688035965, + 0.005634242203086615, + 0.005601638928055763, + 0.017432210966944695, + 0.009323188103735447, + 0.008247824385762215, + 0.006650514435023069, + 0.010912550613284111, + 0.011343847028911114, + 0.0058821733109653, + 0.00881969928741455, + 0.007292615715414286, + 0.005925954319536686, + 0.002142171375453472, + 0.002846900373697281, + 0.00349711487069726, + 0.0031965405214577913, + 0.003773595904931426, + 0.0020251674577593803, + 0.002204488730058074, + 0.002496882574632764, + 0.002028690418228507, + 0.004218702204525471, + 0.0018866495229303837, + 0.0045104483142495155, + 0.006010036915540695, + 0.002005454618483782, + 0.002679933561012149, + 0.0025430459063500166, + 0.0020447333808988333, + 0.0025386083871126175, + 0.00470962468534708, + 0.007360301446169615, + 0.00226100767031312, + 0.002309258794412017, + 0.0029507840517908335, + 0.002062117448076606, + 0.0021895875688642263, + 0.0033409185707569122, + 0.005835606716573238, + 0.0025025876238942146, + 0.003403875045478344, + 0.0014184170868247747, + 0.0031021698378026485, + 0.0034858458675444126, + 0.0018152749398723245, + 0.005397731903940439, + 0.002226900076493621, + 0.0043341475538909435, + 0.003758694278076291, + 0.002134776208549738, + 0.005505750887095928, + 0.0030868572648614645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01606719382107258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_tensor([4544, 4544])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_tensor([4544, 4544])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..55bf99ece91eb2fe8533496b729c3154e07684d6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_tensor([4544, 4544])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly overhead angle, the camera tilts steadily and slowly upward as the lid of a traditional bamboo steamer is carefully lifted. The moment the lid is removed, a thick, scalding white steam billows out, instantly filling the frame and creating a hazy, lingering atmosphere. As the steam slowly dissipates, a tray of plump, round dumplings, neatly arranged, gradually emerges, resting on emerald green leaves at the bottom of the steamer. Illuminated by warm, soft overhead light, each dumpling's wrapper appears nearly translucent, shimmering with crystal clarity, revealing the delicate folds of the wrapper's delicate, clear texture. As the steam continues to rise from below, the dumplings seem to come alive, trembling ever so slightly in the vapor, their smooth surfaces reflecting an alluring, moist sheen. The entire scene creates a warm, comforting, and appetizing atmosphere of home.", + "seed": 6925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_4544_seed_6925_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.006168307131156325, + "vlm_logprob": -0.006168307131156325, + "vlm_yes_no_margin": 5.95, + "vlm_token_credit": -0.006168307131156325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009556971490383148 + ], + "opposite_logprob_tokens": [ + -7.009556770324707 + ], + "logprob_reward": -0.009556971490383148, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.009556971490383148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11492300778627396, + 0.33036425709724426, + 0.07981686294078827, + 0.06971580535173416, + 0.08163868635892868, + 0.11460152268409729, + 0.025446610525250435, + 0.07340756058692932, + 0.02963418699800968, + 0.08045148849487305 + ], + "credit_map": [ + 0.009056980721652508, + 0.010866641066968441, + 0.014954011887311935, + 0.012038576416671276, + 0.011250786483287811, + 0.009705347940325737, + 0.014736057259142399, + 0.03231460601091385, + 0.024469774216413498, + 0.07068429887294769, + 0.028515923768281937, + 0.015817297622561455, + 0.031032275408506393, + 0.016230029985308647, + 0.053749021142721176, + 0.08986562490463257, + 0.010436784476041794, + 0.010133454576134682, + 0.005144084338098764, + 0.007399583701044321, + 0.01703982800245285, + 0.011034105904400349, + 0.009951034560799599, + 0.008677982725203037, + 0.009942891076207161, + 0.009363108314573765, + 0.010804335586726665, + 0.005845924839377403, + 0.011333010159432888, + 0.008359306491911411, + 0.005206932779401541, + 0.008860297501087189, + 0.011329494416713715, + 0.008484783582389355, + 0.013063691556453705, + 0.009149816818535328, + 0.009981511160731316, + 0.01370509434491396, + 0.007358565926551819, + 0.00856572762131691, + 0.01747662015259266, + 0.014254198409616947, + 0.015167459845542908, + 0.012609328143298626, + 0.012611865997314453, + 0.012435280717909336, + 0.008666686713695526, + 0.021380087360739708, + 0.002234256360679865, + 0.0031397314742207527, + 0.002266139490529895, + 0.0032626446336507797, + 0.004643476102501154, + 0.002748525235801935, + 0.003232777351513505, + 0.003919058945029974, + 0.004550341051071882, + 0.017658675089478493, + 0.006612235214561224, + 0.00519848195835948, + 0.008081706240773201, + 0.022497378289699554, + 0.0038749242667108774, + 0.004933817777782679, + 0.0024944350589066744, + 0.004270598292350769, + 0.002915686462074518, + 0.004485362209379673, + 0.004410217981785536, + 0.0034829326905310154, + 0.004760014824569225, + 0.0028149387799203396, + 0.007536663208156824, + 0.011691552586853504, + 0.007825958542525768, + 0.011035330593585968, + 0.01108126062899828, + 0.010611946694552898, + 0.010359754785895348, + 0.01030901912599802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.009556971490383148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014220921322703362 + ], + "opposite_logprob_tokens": [ + -4.264220714569092 + ], + "logprob_reward": -0.014220921322703362, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014220921322703362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10586310923099518, + 0.4477406442165375, + 0.1421831250190735, + 0.0605633407831192, + 0.09204015135765076, + 0.03693082183599472, + 0.010101648047566414, + 0.059152375906705856, + 0.022098835557699203, + 0.02332593873143196 + ], + "credit_map": [ + 0.007645325269550085, + 0.010733040980994701, + 0.016433458775281906, + 0.012186684645712376, + 0.012277924455702305, + 0.011408992111682892, + 0.010740792378783226, + 0.02443689852952957, + 0.03973741456866264, + 0.1262010931968689, + 0.0484498105943203, + 0.02979929931461811, + 0.050703518092632294, + 0.026763303205370903, + 0.022681040689349174, + 0.1034051775932312, + 0.016594162210822105, + 0.02411828190088272, + 0.009831197559833527, + 0.013491826131939888, + 0.025617487728595734, + 0.022037368267774582, + 0.013672970235347748, + 0.01681983657181263, + 0.011616203002631664, + 0.0082740793004632, + 0.005838110577315092, + 0.006050705909729004, + 0.009945064783096313, + 0.006408775225281715, + 0.005130157805979252, + 0.007300243247300386, + 0.01328283827751875, + 0.009854315780103207, + 0.013377455063164234, + 0.007940473034977913, + 0.008915764279663563, + 0.02501847594976425, + 0.007391537073999643, + 0.006259292364120483, + 0.005223290994763374, + 0.004526973236352205, + 0.004436591174453497, + 0.005101068876683712, + 0.0046014790423214436, + 0.005351431202143431, + 0.0033244325313717127, + 0.004365553148090839, + 0.0014950799522921443, + 0.0012022097362205386, + 0.0015008855843916535, + 0.001148445880971849, + 0.001419422565959394, + 0.000794626132119447, + 0.0007829574169591069, + 0.001758020487613976, + 0.003663029056042433, + 0.012776886112987995, + 0.006269697565585375, + 0.004508140031248331, + 0.0051400503143668175, + 0.018118780106306076, + 0.005059744697064161, + 0.003616049885749817, + 0.002060223836451769, + 0.0028472603298723698, + 0.0022615690249949694, + 0.00307834311388433, + 0.003077123314142227, + 0.0034964748192578554, + 0.0029234972316771746, + 0.0023543420247733593, + 0.0023918773513287306, + 0.0020759121980518103, + 0.002190097002312541, + 0.004470436368137598, + 0.0025084547232836485, + 0.003871732158586383, + 0.003026685444638133, + 0.0027907416224479675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.014220921322703362 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010483728256076574 + ], + "opposite_logprob_tokens": [ + -6.876048564910889 + ], + "logprob_reward": -0.0010483728256076574, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010483728256076574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06130461022257805, + 0.32072195410728455, + 0.1833096742630005, + 0.06023913249373436, + 0.07088607549667358, + 0.09568620473146439, + 0.02045835368335247, + 0.12160371989011765, + 0.029596427455544472, + 0.03619387000799179 + ], + "credit_map": [ + 0.005832529626786709, + 0.0052824304439127445, + 0.010149635374546051, + 0.006223912350833416, + 0.004898658953607082, + 0.006091780960559845, + 0.006909973919391632, + 0.01591568998992443, + 0.01492941752076149, + 0.04344048351049423, + 0.028353720903396606, + 0.018271554261446, + 0.019600873813033104, + 0.015025507658720016, + 0.07679541409015656, + 0.10430498421192169, + 0.020794790238142014, + 0.0283803753554821, + 0.017842503264546394, + 0.018359659239649773, + 0.025892604142427444, + 0.023714352399110794, + 0.02500039152801037, + 0.02332499250769615, + 0.0068945977836847305, + 0.007114355452358723, + 0.005228112451732159, + 0.00951374601572752, + 0.012096687220036983, + 0.0063462029211223125, + 0.005197386257350445, + 0.007848042994737625, + 0.011435233987867832, + 0.007318601012229919, + 0.009035171009600163, + 0.007843144237995148, + 0.006760716438293457, + 0.0185487512499094, + 0.005583782680332661, + 0.004360669758170843, + 0.013931852765381336, + 0.012448879890143871, + 0.012226143851876259, + 0.012941816821694374, + 0.010642094537615776, + 0.015285352244973183, + 0.008518614806234837, + 0.009691447019577026, + 0.001871201558969915, + 0.00237179733812809, + 0.002699163043871522, + 0.0018538235453888774, + 0.0031503632199019194, + 0.0027598028536885977, + 0.0016748373163864017, + 0.004077365156263113, + 0.009095669724047184, + 0.02549675665795803, + 0.016010155901312828, + 0.009601123631000519, + 0.009972923435270786, + 0.033755913376808167, + 0.010625065304338932, + 0.007046113256365061, + 0.0030781833920627832, + 0.0030421875417232513, + 0.0027641693595796824, + 0.004762895405292511, + 0.004120586905628443, + 0.00420012790709734, + 0.003568320767953992, + 0.004059957340359688, + 0.004037862177938223, + 0.004915096331387758, + 0.0033046614844352007, + 0.006663132458925247, + 0.004333505406975746, + 0.005122706759721041, + 0.003327181562781334, + 0.004489721730351448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.0010483728256076574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196133067831397 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196133067831397, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196133067831397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07521221041679382, + 0.46207767724990845, + 0.11469557136297226, + 0.039239369332790375, + 0.09820129722356796, + 0.043634701520204544, + 0.026069264858961105, + 0.08790148794651031, + 0.014632036909461021, + 0.038336411118507385 + ], + "credit_map": [ + 0.007490857504308224, + 0.007731874473392963, + 0.011390860192477703, + 0.008094563148915768, + 0.007736498024314642, + 0.0081176171079278, + 0.009003139100968838, + 0.0156467966735363, + 0.02512088604271412, + 0.08549000322818756, + 0.04382362961769104, + 0.02029533125460148, + 0.037174638360738754, + 0.022142408415675163, + 0.08111219853162766, + 0.14691859483718872, + 0.009905940853059292, + 0.015732692554593086, + 0.013317301869392395, + 0.012367887422442436, + 0.016793403774499893, + 0.012588507495820522, + 0.02095072716474533, + 0.01303910743445158, + 0.005315606016665697, + 0.005492022726684809, + 0.004834039602428675, + 0.0038395849987864494, + 0.008097682148218155, + 0.004364163149148226, + 0.003198346123099327, + 0.004097919911146164, + 0.017050061374902725, + 0.007645247504115105, + 0.016111427918076515, + 0.009204737842082977, + 0.009693346917629242, + 0.02434900775551796, + 0.00607515312731266, + 0.008072311989963055, + 0.006181459408253431, + 0.005290365312248468, + 0.00536971865221858, + 0.005436258856207132, + 0.005151038058102131, + 0.006198142189532518, + 0.003232610644772649, + 0.006775104906409979, + 0.0022933403961360455, + 0.0027844009455293417, + 0.002476876135915518, + 0.003522815415635705, + 0.006333553232252598, + 0.004138736054301262, + 0.0017738969763740897, + 0.0027456439565867186, + 0.009808871895074844, + 0.01669139601290226, + 0.010996272787451744, + 0.00950833410024643, + 0.012844443321228027, + 0.010474460199475288, + 0.009055739268660545, + 0.0085219731554389, + 0.0014966471353545785, + 0.0018876923713833094, + 0.001503966748714447, + 0.0022701288107782602, + 0.0016666192095726728, + 0.0016999435611069202, + 0.0019594798795878887, + 0.002147558843716979, + 0.0038191797211766243, + 0.004380781203508377, + 0.0030604361090809107, + 0.007275260053575039, + 0.0060880728997290134, + 0.005562999751418829, + 0.0037203594110906124, + 0.004429319873452187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.003196133067831397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028191369492560625 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028191369492560625, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028191369492560625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15018318593502045, + 0.5038221478462219, + 0.07680913060903549, + 0.039144691079854965, + 0.06502629816532135, + 0.041251592338085175, + 0.020046416670084, + 0.034804996103048325, + 0.021929694339632988, + 0.046981774270534515 + ], + "credit_map": [ + 0.012210079468786716, + 0.017401622608304024, + 0.02161276526749134, + 0.017187999561429024, + 0.014511401765048504, + 0.018285823985934258, + 0.015591646544635296, + 0.03338185325264931, + 0.033084023743867874, + 0.10931585729122162, + 0.05125411972403526, + 0.023175043985247612, + 0.04490462318062782, + 0.024968504905700684, + 0.06425414979457855, + 0.15286588668823242, + 0.0074625276029109955, + 0.009151631034910679, + 0.008388478308916092, + 0.009712251834571362, + 0.011702802032232285, + 0.00920550525188446, + 0.0125492624938488, + 0.008636679500341415, + 0.003725703340023756, + 0.00426822155714035, + 0.005821182858198881, + 0.0037867571227252483, + 0.00560461450368166, + 0.004353044554591179, + 0.003714251797646284, + 0.007870922796428204, + 0.017598388716578484, + 0.0035697428975254297, + 0.00505656935274601, + 0.0046640560030937195, + 0.004319270607084036, + 0.02241380326449871, + 0.002030791947618127, + 0.005373680964112282, + 0.004982451908290386, + 0.005114338360726833, + 0.0035621593706309795, + 0.005778294522315264, + 0.00715682515874505, + 0.005947887431830168, + 0.0029582977294921875, + 0.005751342978328466, + 0.002452282002195716, + 0.0021443262230604887, + 0.0026504178531467915, + 0.0025273330975323915, + 0.0032505951821804047, + 0.0031270827166736126, + 0.0020774395670741796, + 0.0018169416580349207, + 0.0035943828988820314, + 0.004644761327654123, + 0.005351702682673931, + 0.004828497301787138, + 0.004104172345250845, + 0.004932216834276915, + 0.004247438628226519, + 0.0031018266454339027, + 0.002635876415297389, + 0.002171766944229603, + 0.0032454486936330795, + 0.0028840445447713137, + 0.0024294666945934296, + 0.002895379438996315, + 0.0025438042357563972, + 0.0031239090021699667, + 0.004442246630787849, + 0.005005857907235622, + 0.0041539533995091915, + 0.007254335097968578, + 0.007653181441128254, + 0.0069243572652339935, + 0.0060686515644192696, + 0.005479196086525917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0028191369492560625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera tilts steadily and slowly upward as the lid of a traditional bamboo steamer is carefully lifted. The moment the lid is removed, a thick, scalding white steam billows out, instantly filling the frame and creating a hazy, lingering atmosphere. As the steam slowly dissipates, a tray of plump, round dumplings, neatly arranged, gradually emerges, resting on emerald green leaves at the bottom of the steamer. Illuminated by warm, soft overhead light, each dumpling's wrapper appears nearly translucent, shimmering with crystal clarity, revealing the delicate folds of the wrapper's delicate, clear texture. As the steam continues to rise from below, the dumplings seem to come alive, trembling ever so slightly in the vapor, their smooth surfaces reflecting an alluring, moist sheen. The entire scene creates a warm, comforting, and appetizing atmosphere of home.", + "seed": 6926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_4544_seed_6926_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01101355031132698, + "vlm_logprob": -0.01101355031132698, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.01101355031132698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025852955877780914 + ], + "opposite_logprob_tokens": [ + -5.275853157043457 + ], + "logprob_reward": -0.025852955877780914, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.025852955877780914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4822884500026703, + 0.030776897445321083, + 0.04557335004210472, + 0.33456680178642273, + 0.012463429011404514, + 0.0171084962785244, + 0.02384260483086109, + 0.016120772808790207, + 0.020206356421113014, + 0.01705286279320717 + ], + "credit_map": [ + 0.014089756645262241, + 0.013858162797987461, + 0.025806687772274017, + 0.06777375191450119, + 0.02322268672287464, + 0.125803142786026, + 0.17903585731983185, + 0.032698411494493484, + 0.004133810289204121, + 0.004445296712219715, + 0.0028874725103378296, + 0.004422723315656185, + 0.0041561792604625225, + 0.004518499597907066, + 0.0032423187512904406, + 0.0029705953784286976, + 0.00698921037837863, + 0.012513495050370693, + 0.0036133662797510624, + 0.005141933914273977, + 0.005553262308239937, + 0.004811958875507116, + 0.002162716817110777, + 0.00478740781545639, + 0.027363676577806473, + 0.13125237822532654, + 0.0036099941935390234, + 0.004498105496168137, + 0.14383795857429504, + 0.013700561597943306, + 0.004637441597878933, + 0.005666683427989483, + 0.0018908015917986631, + 0.0017520979745313525, + 0.0018011803040280938, + 0.0016114762984216213, + 0.0015976331196725368, + 0.001191969495266676, + 0.0011861782986670732, + 0.001432090881280601, + 0.002441240241751075, + 0.0028367568738758564, + 0.0026027788408100605, + 0.0018816090887412429, + 0.001925870543345809, + 0.0021913156379014254, + 0.0013790115481242537, + 0.0018499133875593543, + 0.0036596208810806274, + 0.004412975627928972, + 0.0020245483610779047, + 0.0036879354156553745, + 0.00205334834754467, + 0.002664006780833006, + 0.0029344791546463966, + 0.00240569026209414, + 0.002343758475035429, + 0.002840477041900158, + 0.001994223101064563, + 0.002663407474756241, + 0.001320456387475133, + 0.0010390715906396508, + 0.0012723504332825541, + 0.0026470290031284094, + 0.0031055475119501352, + 0.003490824019536376, + 0.0020884505938738585, + 0.0025975191965699196, + 0.001706992625258863, + 0.0012802635319530964, + 0.0028733883518725634, + 0.0030633704736828804, + 0.0029135006479918957, + 0.003347687888890505, + 0.0013520423090085387, + 0.0016523763770237565, + 0.0013635995564982295, + 0.002223771531134844, + 0.002319094492122531, + 0.0018807894084602594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.025852955877780914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018231317400932312 + ], + "opposite_logprob_tokens": [ + -4.018231391906738 + ], + "logprob_reward": -0.018231317400932312, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018231317400932312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26317745447158813, + 0.03289036452770233, + 0.043747175484895706, + 0.507773220539093, + 0.014760913327336311, + 0.022244108840823174, + 0.026180481538176537, + 0.022817818447947502, + 0.0527859590947628, + 0.0136225251480937 + ], + "credit_map": [ + 0.009864680469036102, + 0.007983325980603695, + 0.016359860077500343, + 0.01917247101664543, + 0.013574965298175812, + 0.09850574284791946, + 0.0772327408194542, + 0.020483676344156265, + 0.004974666051566601, + 0.004619952756911516, + 0.0028141234070062637, + 0.00503695709630847, + 0.004531120881438255, + 0.004546497482806444, + 0.0033243524376302958, + 0.0030426953453570604, + 0.005981604102998972, + 0.007893199101090431, + 0.004367935936897993, + 0.005633966531604528, + 0.005960594397038221, + 0.005436233710497618, + 0.004065017215907574, + 0.004408626351505518, + 0.04239515960216522, + 0.19850118458271027, + 0.00693756015971303, + 0.006447082385420799, + 0.2178039401769638, + 0.02028893306851387, + 0.006921742111444473, + 0.008477583527565002, + 0.002335131401196122, + 0.003338493872433901, + 0.003083209041506052, + 0.001691063167527318, + 0.0013560057850554585, + 0.0010205203434452415, + 0.0008900095708668232, + 0.0010464813094586134, + 0.0030080676078796387, + 0.005678246263414621, + 0.0028266299050301313, + 0.0023823564406484365, + 0.0022046291269361973, + 0.0028563174419105053, + 0.0010865131625905633, + 0.0022013497073203325, + 0.005068328697234392, + 0.00493917940184474, + 0.0029834811575710773, + 0.0032820829655975103, + 0.00221498915925622, + 0.001960660330951214, + 0.0026383274234831333, + 0.003093431703746319, + 0.003556466195732355, + 0.004544042516499758, + 0.0030011183116585016, + 0.0027279553469270468, + 0.0020679591689258814, + 0.0015505641931667924, + 0.0022921913769096136, + 0.0030775200575590134, + 0.0072789969854056835, + 0.014465280808508396, + 0.005568219348788261, + 0.006054676603525877, + 0.004105860833078623, + 0.004058538004755974, + 0.003933717496693134, + 0.00732066947966814, + 0.002185370307415724, + 0.0027563900221139193, + 0.0009964953642338514, + 0.0011915776412934065, + 0.0012787824962288141, + 0.0022722051944583654, + 0.0013931539142504334, + 0.0015485506737604737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.018231317400932312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002206393051892519 + ], + "opposite_logprob_tokens": [ + -6.127206325531006 + ], + "logprob_reward": -0.002206393051892519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002206393051892519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15622374415397644, + 0.054011277854442596, + 0.14133921265602112, + 0.4501917362213135, + 0.029084665700793266, + 0.04058435186743736, + 0.03273457661271095, + 0.03740619868040085, + 0.03709571063518524, + 0.021328497678041458 + ], + "credit_map": [ + 0.006111584138125181, + 0.006471103522926569, + 0.011159026063978672, + 0.0222136452794075, + 0.01135934516787529, + 0.03730648383498192, + 0.048640038818120956, + 0.01296253688633442, + 0.012811078689992428, + 0.006259816698729992, + 0.005923355929553509, + 0.0068276855163276196, + 0.006813433486968279, + 0.006533532869070768, + 0.0040952907875180244, + 0.004747087135910988, + 0.02601609379053116, + 0.02968057245016098, + 0.01314463559538126, + 0.01616491749882698, + 0.015455173328518867, + 0.016915619373321533, + 0.009914330206811428, + 0.01404788438230753, + 0.03642387315630913, + 0.17705774307250977, + 0.004786073230206966, + 0.00616065738722682, + 0.19300827383995056, + 0.01880273036658764, + 0.006180431228131056, + 0.0077719963155686855, + 0.004238766152411699, + 0.00611259788274765, + 0.004526752978563309, + 0.0037556004244834185, + 0.003515697317197919, + 0.001994943479076028, + 0.0022714908700436354, + 0.0026688219513744116, + 0.00637952983379364, + 0.010666318237781525, + 0.0063031455501914024, + 0.0035137678496539593, + 0.0035336397122591734, + 0.004884958732873201, + 0.002420893870294094, + 0.0028821034356951714, + 0.005794140975922346, + 0.007478269282728434, + 0.004002122208476067, + 0.003592513967305422, + 0.002697149058803916, + 0.0024750723969191313, + 0.0025014893617480993, + 0.004193824250251055, + 0.005680931732058525, + 0.006886899005621672, + 0.003568948246538639, + 0.004170761443674564, + 0.0037279007956385612, + 0.004667268134653568, + 0.004677243530750275, + 0.004026248585432768, + 0.005539662204682827, + 0.009581584483385086, + 0.0041229622438549995, + 0.003933409228920937, + 0.0028834666591137648, + 0.0027157773729413748, + 0.002527408767491579, + 0.005791439674794674, + 0.002708209678530693, + 0.0037773156072944403, + 0.0032467134296894073, + 0.0018326739082112908, + 0.0021288751158863306, + 0.0025710200425237417, + 0.002861633663997054, + 0.002202058909460902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.002206393051892519 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955214146524668 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955214146524668, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955214146524668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13821391761302948, + 0.04238181561231613, + 0.0469316765666008, + 0.4761635661125183, + 0.04165242239832878, + 0.03966192156076431, + 0.06598212569952011, + 0.029956340789794922, + 0.10546077787876129, + 0.013595422729849815 + ], + "credit_map": [ + 0.006914609577506781, + 0.011162656359374523, + 0.02161405049264431, + 0.028025580570101738, + 0.012438072822988033, + 0.019617589190602303, + 0.02526114322245121, + 0.013180208392441273, + 0.006599002983421087, + 0.004692797549068928, + 0.005150208715349436, + 0.006045400630682707, + 0.00646027410402894, + 0.005588536150753498, + 0.0036241712514311075, + 0.004221423529088497, + 0.005849112290889025, + 0.008558936417102814, + 0.004540211986750364, + 0.007012718822807074, + 0.0058519612066447735, + 0.0059477644972503185, + 0.0040869563817977905, + 0.005084015429019928, + 0.03804166987538338, + 0.18347421288490295, + 0.007939185947179794, + 0.007612465415149927, + 0.20317131280899048, + 0.020097075030207634, + 0.006731792818754911, + 0.009095830842852592, + 0.007012905087321997, + 0.0075738620944321156, + 0.008359772153198719, + 0.0046264417469501495, + 0.004788623657077551, + 0.0026863068342208862, + 0.0022445295471698046, + 0.004359977785497904, + 0.005467017646878958, + 0.010016563348472118, + 0.0034765347372740507, + 0.002821172820404172, + 0.00499405711889267, + 0.005314800422638655, + 0.004012580960988998, + 0.0035591935738921165, + 0.009746121242642403, + 0.013352768495678902, + 0.00800687912851572, + 0.005717780441045761, + 0.005209037102758884, + 0.010165316052734852, + 0.004111585207283497, + 0.009672638960182667, + 0.004431597422808409, + 0.004867101088166237, + 0.003351545426994562, + 0.0033705674577504396, + 0.002731313230469823, + 0.003705353010445833, + 0.004020565189421177, + 0.003478298196569085, + 0.02073313109576702, + 0.020217614248394966, + 0.009401732124388218, + 0.012812945060431957, + 0.010297276079654694, + 0.01033659465610981, + 0.00940541923046112, + 0.012256072834134102, + 0.001839833683334291, + 0.0026093812193721533, + 0.001618575886823237, + 0.0017880031373351812, + 0.0011709253303706646, + 0.0013573975302278996, + 0.0016118698986247182, + 0.001599436509422958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.005955214146524668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00282187107950449 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.00282187107950449, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00282187107950449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3263828754425049, + 0.030521506443619728, + 0.17527896165847778, + 0.380206435918808, + 0.012836410664021969, + 0.012236102484166622, + 0.01799294538795948, + 0.017667653039097786, + 0.013159546069800854, + 0.01371763925999403 + ], + "credit_map": [ + 0.010166964493691921, + 0.00952513050287962, + 0.018009882420301437, + 0.04407453164458275, + 0.016211796551942825, + 0.0878600925207138, + 0.12221623957157135, + 0.018318256363272667, + 0.003826005617156625, + 0.003196809208020568, + 0.004371670540422201, + 0.005416885484009981, + 0.003977645188570023, + 0.0038157477974891663, + 0.002841130131855607, + 0.0030756143387407064, + 0.029949085786938667, + 0.052670832723379135, + 0.011356192640960217, + 0.013587304390966892, + 0.02356499619781971, + 0.02251458913087845, + 0.007643391378223896, + 0.013992558233439922, + 0.030942119657993317, + 0.14951546490192413, + 0.00421550590544939, + 0.004915117751806974, + 0.16348077356815338, + 0.015478415414690971, + 0.005147225689142942, + 0.006511812098324299, + 0.0019256146624684334, + 0.0030820516403764486, + 0.0026549811009317636, + 0.001438676961697638, + 0.001378410728648305, + 0.0007139789522625506, + 0.0005708454409614205, + 0.001071851933375001, + 0.001807946595363319, + 0.002871354576200247, + 0.0014592667575925589, + 0.0011791670694947243, + 0.0014248201623558998, + 0.0010848887031897902, + 0.0013841018080711365, + 0.0010245569283142686, + 0.002878560684621334, + 0.0048250723630189896, + 0.001535374904051423, + 0.0019202780677005649, + 0.0015805462608113885, + 0.0017572803189978004, + 0.0018532773246988654, + 0.0016425548819825053, + 0.0023915087804198265, + 0.0029026055708527565, + 0.0014245775528252125, + 0.0017864661058411002, + 0.0016993321478366852, + 0.0031405044719576836, + 0.0024172039702534676, + 0.0019054547883570194, + 0.0016800804296508431, + 0.003272128291428089, + 0.001125552342273295, + 0.0014636121923103929, + 0.00120069261174649, + 0.0014272448606789112, + 0.0011348705738782883, + 0.0018553647678345442, + 0.0015852410579100251, + 0.003077110042795539, + 0.0012593147112056613, + 0.0010997122153639793, + 0.0014931746991351247, + 0.0009067688952200115, + 0.0031054492574185133, + 0.0011908678570762277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.00282187107950449 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera tilts steadily and slowly upward as the lid of a traditional bamboo steamer is carefully lifted. The moment the lid is removed, a thick, scalding white steam billows out, instantly filling the frame and creating a hazy, lingering atmosphere. As the steam slowly dissipates, a tray of plump, round dumplings, neatly arranged, gradually emerges, resting on emerald green leaves at the bottom of the steamer. Illuminated by warm, soft overhead light, each dumpling's wrapper appears nearly translucent, shimmering with crystal clarity, revealing the delicate folds of the wrapper's delicate, clear texture. As the steam continues to rise from below, the dumplings seem to come alive, trembling ever so slightly in the vapor, their smooth surfaces reflecting an alluring, moist sheen. The entire scene creates a warm, comforting, and appetizing atmosphere of home.", + "seed": 6927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_4544_seed_6927_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006957777356728911, + "vlm_logprob": -0.006957777356728911, + "vlm_yes_no_margin": 5.675, + "vlm_token_credit": -0.006957777356728911, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016176989302039146 + ], + "opposite_logprob_tokens": [ + -5.641177177429199 + ], + "logprob_reward": -0.016176989302039146, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.016176989302039146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22328203916549683, + 0.11569471657276154, + 0.15371081233024597, + 0.02168106660246849, + 0.06321150809526443, + 0.033449769020080566, + 0.11789834499359131, + 0.134832501411438, + 0.09525617957115173, + 0.040983088314533234 + ], + "credit_map": [ + 0.024516351521015167, + 0.018550464883446693, + 0.02262965589761734, + 0.038621559739112854, + 0.018770257011055946, + 0.026443401351571083, + 0.04210413992404938, + 0.031646184623241425, + 0.010793612338602543, + 0.041919514536857605, + 0.004492626991122961, + 0.003541270038112998, + 0.022680455818772316, + 0.020736558362841606, + 0.005726523697376251, + 0.00580414105206728, + 0.025533536449074745, + 0.04592800512909889, + 0.013176104985177517, + 0.009140941314399242, + 0.01019616611301899, + 0.02928420715034008, + 0.007540919352322817, + 0.012910904362797737, + 0.0032034467440098524, + 0.004093047697097063, + 0.00119319639634341, + 0.0016543681267648935, + 0.004638922866433859, + 0.003364268224686384, + 0.0019143184181302786, + 0.0016194951022043824, + 0.009439450688660145, + 0.014861470088362694, + 0.008486367762088776, + 0.0055000316351652145, + 0.006064599845558405, + 0.00627212505787611, + 0.005043504294008017, + 0.0075439577922225, + 0.004604743327945471, + 0.01029095146805048, + 0.0017806788673624396, + 0.0027523941826075315, + 0.0034246323630213737, + 0.003222616156563163, + 0.003639122238382697, + 0.00373462843708694, + 0.023328209295868874, + 0.024520710110664368, + 0.009003964252769947, + 0.009510336443781853, + 0.009123379364609718, + 0.01188375148922205, + 0.01914193294942379, + 0.01138604898005724, + 0.032858218997716904, + 0.05667071044445038, + 0.004807265009731054, + 0.009320255368947983, + 0.009050308726727962, + 0.01030584517866373, + 0.005183116998523474, + 0.0066367643885314465, + 0.014873942360281944, + 0.03618628904223442, + 0.005431181751191616, + 0.009377206675708294, + 0.007528387941420078, + 0.007811134215444326, + 0.00537628959864378, + 0.008671741001307964, + 0.010003452189266682, + 0.0036193858832120895, + 0.0022782846353948116, + 0.003639256814494729, + 0.0034157244954258204, + 0.0073685189709067345, + 0.003792295465245843, + 0.0068661668337881565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.016176989302039146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009813870303332806 + ], + "opposite_logprob_tokens": [ + -4.6348137855529785 + ], + "logprob_reward": -0.009813870303332806, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009813870303332806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3449888527393341, + 0.06752009689807892, + 0.20816448330879211, + 0.03222738206386566, + 0.05960017442703247, + 0.05614297837018967, + 0.09236358851194382, + 0.07175233215093613, + 0.04771767929196358, + 0.01952243223786354 + ], + "credit_map": [ + 0.015622484497725964, + 0.02216421253979206, + 0.02905539609491825, + 0.10284321755170822, + 0.028358494862914085, + 0.02740495279431343, + 0.07172234356403351, + 0.047817766666412354, + 0.0055852010846138, + 0.023549770936369896, + 0.0028822575695812702, + 0.002531565958634019, + 0.010227146558463573, + 0.013966687023639679, + 0.003640098264440894, + 0.005137366242706776, + 0.032892338931560516, + 0.06181458383798599, + 0.018432805314660072, + 0.012773885391652584, + 0.013366790488362312, + 0.04043000563979149, + 0.010227859020233154, + 0.01822623237967491, + 0.004691052716225386, + 0.007382795214653015, + 0.0037325965240597725, + 0.002060837112367153, + 0.0035330760292708874, + 0.005406613927334547, + 0.0023431447334587574, + 0.003077267436310649, + 0.010461155325174332, + 0.00984192080795765, + 0.008976712822914124, + 0.005388089455664158, + 0.006720282603055239, + 0.005806414410471916, + 0.004936703015118837, + 0.007468901574611664, + 0.007630960550159216, + 0.013925767503678799, + 0.004687990993261337, + 0.0059241787530481815, + 0.005841961596161127, + 0.005468323361128569, + 0.006463870406150818, + 0.006199925672262907, + 0.01012906339019537, + 0.019637467339634895, + 0.005134436767548323, + 0.010526653379201889, + 0.010408109053969383, + 0.009348351508378983, + 0.018207069486379623, + 0.008972441777586937, + 0.010701066814363003, + 0.027824416756629944, + 0.002782544819638133, + 0.00460071349516511, + 0.006047366186976433, + 0.009018157608807087, + 0.003457180457189679, + 0.007320889737457037, + 0.005104845389723778, + 0.012246714904904366, + 0.0034313087817281485, + 0.006285122595727444, + 0.004012241493910551, + 0.005482983309775591, + 0.005377672147005796, + 0.005776793230324984, + 0.002750974614173174, + 0.001446034642867744, + 0.0012267829151824117, + 0.0058398377150297165, + 0.0018357147928327322, + 0.002367505105212331, + 0.0019641518592834473, + 0.002091433387249708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.009813870303332806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010507544502615929 + ], + "opposite_logprob_tokens": [ + -6.87605094909668 + ], + "logprob_reward": -0.0010507544502615929, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010507544502615929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1263917088508606, + 0.06579262763261795, + 0.08456166088581085, + 0.05913560092449188, + 0.12599588930606842, + 0.023363305255770683, + 0.10582707077264786, + 0.06634578853845596, + 0.3036471903324127, + 0.03893917798995972 + ], + "credit_map": [ + 0.013679101131856441, + 0.01121964305639267, + 0.008830361999571323, + 0.0206555537879467, + 0.013117303140461445, + 0.014140412211418152, + 0.02687378227710724, + 0.01787555404007435, + 0.008182112127542496, + 0.009356969967484474, + 0.004485247656702995, + 0.005699177272617817, + 0.011856717057526112, + 0.0076517281122505665, + 0.006020095199346542, + 0.012540582567453384, + 0.012772366404533386, + 0.021741122007369995, + 0.00595965376123786, + 0.005797161255031824, + 0.008435703814029694, + 0.015012508258223534, + 0.00824857410043478, + 0.006594568956643343, + 0.00663297763094306, + 0.008446921594440937, + 0.003896778915077448, + 0.0067372615449130535, + 0.013267695903778076, + 0.007140414323657751, + 0.008700194768607616, + 0.0043133581057190895, + 0.019309550523757935, + 0.029699033126235008, + 0.016009006649255753, + 0.009660439565777779, + 0.015190799720585346, + 0.017274875193834305, + 0.008687194436788559, + 0.010164990089833736, + 0.0031144164968281984, + 0.003579326206818223, + 0.002000081818550825, + 0.0022654300555586815, + 0.004287079442292452, + 0.0026410999707877636, + 0.002596447477117181, + 0.0028794221580028534, + 0.025979172438383102, + 0.03035108931362629, + 0.0060609388165175915, + 0.006354810670018196, + 0.00911655928939581, + 0.014122460968792439, + 0.006561806425452232, + 0.007280235178768635, + 0.010930337011814117, + 0.010319491848349571, + 0.005247982684522867, + 0.012188663706183434, + 0.004593213088810444, + 0.008005490526556969, + 0.009177344851195812, + 0.005883263424038887, + 0.07957746833562851, + 0.12908071279525757, + 0.01204476784914732, + 0.026674821972846985, + 0.018824664875864983, + 0.016287164762616158, + 0.01072025392204523, + 0.01043733861297369, + 0.008828124031424522, + 0.002914819633588195, + 0.0048624868504703045, + 0.00498725101351738, + 0.005309184081852436, + 0.003711747471243143, + 0.0031013903208076954, + 0.005224172957241535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0010507544502615929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257708020508289 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257708020508289, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257708020508289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21442608535289764, + 0.09085660427808762, + 0.048968810588121414, + 0.07852295786142349, + 0.07699092477560043, + 0.04050721600651741, + 0.09141980111598969, + 0.2317039966583252, + 0.08987810462713242, + 0.036725472658872604 + ], + "credit_map": [ + 0.017875373363494873, + 0.018770789727568626, + 0.026371309533715248, + 0.029490606859326363, + 0.0206427164375782, + 0.027101833373308182, + 0.04364128038287163, + 0.03053218498826027, + 0.010487651452422142, + 0.015506933443248272, + 0.005643544718623161, + 0.008931993506848812, + 0.009986789897084236, + 0.014670917764306068, + 0.008916259743273258, + 0.016712510958313942, + 0.00644643371924758, + 0.0062869940884411335, + 0.004609427880495787, + 0.005988904275000095, + 0.006664250511676073, + 0.008374463766813278, + 0.006301997695118189, + 0.004296336323022842, + 0.012887425720691681, + 0.014000426977872849, + 0.006220294162631035, + 0.007112330757081509, + 0.01437119860202074, + 0.009094109758734703, + 0.009310184046626091, + 0.005526989232748747, + 0.011927840299904346, + 0.015739986672997475, + 0.011215153150260448, + 0.007269200868904591, + 0.008600444532930851, + 0.007269641384482384, + 0.005344667937606573, + 0.00962398573756218, + 0.005503433756530285, + 0.006322598550468683, + 0.0027400702238082886, + 0.0037527752574533224, + 0.005932612810283899, + 0.004790232051163912, + 0.007025709375739098, + 0.004439784213900566, + 0.023001525551080704, + 0.020177315920591354, + 0.004217988811433315, + 0.006015300750732422, + 0.008443058468401432, + 0.012413518503308296, + 0.007876564748585224, + 0.009274528361856937, + 0.06541763246059418, + 0.09715383499860764, + 0.007873348891735077, + 0.014054518193006516, + 0.013960680924355984, + 0.018492551520466805, + 0.007787575479596853, + 0.006963859777897596, + 0.015650426968932152, + 0.027865581214427948, + 0.0046096243895590305, + 0.007785650435835123, + 0.005944900680333376, + 0.007457200437784195, + 0.012076526880264282, + 0.008488195948302746, + 0.0038950343150645494, + 0.003393144579604268, + 0.005670984275639057, + 0.006040299776941538, + 0.0048204148188233376, + 0.0037637583445757627, + 0.004102917388081551, + 0.005038916599005461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.005257708020508289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024895647075027227 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024895647075027227, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024895647075027227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3877628445625305, + 0.08449923992156982, + 0.04095922410488129, + 0.018604731187224388, + 0.031060371547937393, + 0.01485742162913084, + 0.0234611127525568, + 0.19068829715251923, + 0.16778366267681122, + 0.040323108434677124 + ], + "credit_map": [ + 0.020063864067196846, + 0.02579808421432972, + 0.040153659880161285, + 0.09217708557844162, + 0.03776096925139427, + 0.033650998026132584, + 0.08507808297872543, + 0.053080085664987564, + 0.007200292311608791, + 0.02276492491364479, + 0.0036635668948292732, + 0.005300912074744701, + 0.018278302624821663, + 0.008950291201472282, + 0.005710102152079344, + 0.01263083703815937, + 0.005737811326980591, + 0.006049165967851877, + 0.003196821780875325, + 0.0038124697748571634, + 0.00583828566595912, + 0.008643337525427341, + 0.0032126381993293762, + 0.004468688275665045, + 0.0020289153326302767, + 0.002922049490734935, + 0.0012056805426254869, + 0.0019165194826200604, + 0.0035233772359788418, + 0.002492289524525404, + 0.0022600002121180296, + 0.002255897270515561, + 0.00580188911408186, + 0.004231398459523916, + 0.004758988507091999, + 0.0035045575350522995, + 0.0024297190830111504, + 0.0030880437698215246, + 0.0023334103170782328, + 0.004912364296615124, + 0.0020729219540953636, + 0.0029120291583240032, + 0.0012397485552355647, + 0.0021564620546996593, + 0.0015167564852163196, + 0.0017610552022233605, + 0.0017130698543041945, + 0.0014853768516331911, + 0.0027100429870188236, + 0.005015110597014427, + 0.002200422575697303, + 0.0020525238942354918, + 0.0035123114939779043, + 0.0026703255716711283, + 0.0023986969608813524, + 0.0029016758780926466, + 0.054458584636449814, + 0.08063986897468567, + 0.006154106464236975, + 0.011089406907558441, + 0.01111194584518671, + 0.015088277868926525, + 0.006498328875750303, + 0.005647757090628147, + 0.03988085687160492, + 0.07495936751365662, + 0.006398431025445461, + 0.01532249990850687, + 0.009406103752553463, + 0.009222804568707943, + 0.006166714709252119, + 0.006426866631954908, + 0.009087404236197472, + 0.0026489398442208767, + 0.003666294040158391, + 0.010808639228343964, + 0.003318769158795476, + 0.0036955387331545353, + 0.00392612349241972, + 0.003171396441757679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0024895647075027227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera tilts steadily and slowly upward as the lid of a traditional bamboo steamer is carefully lifted. The moment the lid is removed, a thick, scalding white steam billows out, instantly filling the frame and creating a hazy, lingering atmosphere. As the steam slowly dissipates, a tray of plump, round dumplings, neatly arranged, gradually emerges, resting on emerald green leaves at the bottom of the steamer. Illuminated by warm, soft overhead light, each dumpling's wrapper appears nearly translucent, shimmering with crystal clarity, revealing the delicate folds of the wrapper's delicate, clear texture. As the steam continues to rise from below, the dumplings seem to come alive, trembling ever so slightly in the vapor, their smooth surfaces reflecting an alluring, moist sheen. The entire scene creates a warm, comforting, and appetizing atmosphere of home.", + "seed": 6928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_8/index_4544_seed_6928_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.005528674996457994, + "vlm_logprob": -0.005528674996457994, + "vlm_yes_no_margin": 6.225, + "vlm_token_credit": -0.005528674996457994, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015367317944765091 + ], + "opposite_logprob_tokens": [ + -6.76536750793457 + ], + "logprob_reward": -0.015367317944765091, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.015367317944765091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23303359746932983, + 0.12069842219352722, + 0.06453965604305267, + 0.05378448963165283, + 0.05651308223605156, + 0.04612347111105919, + 0.04507893696427345, + 0.22889678180217743, + 0.04071742296218872, + 0.11061418056488037 + ], + "credit_map": [ + 0.01256033219397068, + 0.015379661694169044, + 0.026824790984392166, + 0.039111778140068054, + 0.013928545638918877, + 0.00858766958117485, + 0.07007480412721634, + 0.04656600579619408, + 0.009159225970506668, + 0.01037356723099947, + 0.01872449740767479, + 0.026880502700805664, + 0.018418530002236366, + 0.008206212893128395, + 0.015239457599818707, + 0.013696418143808842, + 0.008823668584227562, + 0.009754987433552742, + 0.00889067817479372, + 0.009052879177033901, + 0.005142356269061565, + 0.007947759702801704, + 0.00788002461194992, + 0.007047301158308983, + 0.006097286939620972, + 0.012239161878824234, + 0.00517628388479352, + 0.006638762075453997, + 0.00644637318328023, + 0.005708293057978153, + 0.004488717764616013, + 0.006989608518779278, + 0.0051159062422811985, + 0.016057152301073074, + 0.006514606531709433, + 0.0034114569425582886, + 0.005317260045558214, + 0.013528029434382915, + 0.0030895909294486046, + 0.003479075152426958, + 0.003296419745311141, + 0.004743113182485104, + 0.006094749551266432, + 0.004103837534785271, + 0.007800000719726086, + 0.0074220020323991776, + 0.007471339777112007, + 0.005192004609853029, + 0.006604515481740236, + 0.002869835589081049, + 0.004287781193852425, + 0.0038933681789785624, + 0.004296989180147648, + 0.008435957133769989, + 0.006700363475829363, + 0.007990125566720963, + 0.04625203460454941, + 0.01805020309984684, + 0.02114950679242611, + 0.015570396557450294, + 0.04651208221912384, + 0.04949681833386421, + 0.019384348765015602, + 0.012481365352869034, + 0.005042436998337507, + 0.0039028541650623083, + 0.002731643384322524, + 0.004324404988437891, + 0.009177519008517265, + 0.004674741066992283, + 0.0077600483782589436, + 0.0031037754379212856, + 0.011364609003067017, + 0.013108366169035435, + 0.00944859255105257, + 0.010781027376651764, + 0.027950119227170944, + 0.017691822722554207, + 0.010167053900659084, + 0.010102586820721626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.015367317944765091 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753950379788876 + ], + "opposite_logprob_tokens": [ + -5.006753921508789 + ], + "logprob_reward": -0.006753950379788876, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753950379788876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26566436886787415, + 0.15035013854503632, + 0.10672146826982498, + 0.04880550503730774, + 0.033013712614774704, + 0.027494505047798157, + 0.023568257689476013, + 0.2688333988189697, + 0.052197106182575226, + 0.023351578041911125 + ], + "credit_map": [ + 0.014123487286269665, + 0.016147488728165627, + 0.029420893639326096, + 0.034188076853752136, + 0.019773351028561592, + 0.009125025011599064, + 0.08818148076534271, + 0.05470455437898636, + 0.00887984037399292, + 0.014395501464605331, + 0.027005448937416077, + 0.03402557224035263, + 0.014339042827486992, + 0.00871894508600235, + 0.021788088604807854, + 0.021197674795985222, + 0.012159647420048714, + 0.016564970836043358, + 0.015377585776150227, + 0.01882365345954895, + 0.007407436612993479, + 0.010109517723321915, + 0.015353989787399769, + 0.010924660600721836, + 0.0071712154895067215, + 0.01552764605730772, + 0.003811383619904518, + 0.0026985390577465296, + 0.004239607602357864, + 0.00960974395275116, + 0.003005084116011858, + 0.002742283744737506, + 0.002659812569618225, + 0.004811951424926519, + 0.004013553727418184, + 0.0036218927707523108, + 0.004918822553008795, + 0.005047600716352463, + 0.003940559457987547, + 0.0039995163679122925, + 0.0023897243663668633, + 0.002704878570511937, + 0.003960755653679371, + 0.0027953980024904013, + 0.004072198644280434, + 0.0044400980696082115, + 0.0038875408936291933, + 0.0032439068891108036, + 0.004418508615344763, + 0.0015572101110592484, + 0.0020877241622656584, + 0.002213381929323077, + 0.0027788165025413036, + 0.005741964094340801, + 0.0027611132245510817, + 0.0020095370709896088, + 0.053607527166604996, + 0.021811947226524353, + 0.026697169989347458, + 0.01848044991493225, + 0.05227424576878548, + 0.05818834900856018, + 0.023605581372976303, + 0.014168106019496918, + 0.008970934897661209, + 0.005420851521193981, + 0.00407679146155715, + 0.004992811940610409, + 0.011004473082721233, + 0.00690156314522028, + 0.0047776224091649055, + 0.006052053067833185, + 0.0027056538965553045, + 0.002328988630324602, + 0.0017545301234349608, + 0.0026135523803532124, + 0.0045536113902926445, + 0.0028375941328704357, + 0.0038999232929199934, + 0.0026577224489301443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.006753950379788876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008163695456460118 + ], + "opposite_logprob_tokens": [ + -7.125816345214844 + ], + "logprob_reward": -0.0008163695456460118, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008163695456460118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37760990858078003, + 0.08683471381664276, + 0.10708445310592651, + 0.051682233810424805, + 0.035464607179164886, + 0.05000592768192291, + 0.03992203250527382, + 0.09369008243083954, + 0.10527347028255463, + 0.05243261903524399 + ], + "credit_map": [ + 0.01220413763076067, + 0.012159999459981918, + 0.022051380947232246, + 0.059586577117443085, + 0.014047292992472649, + 0.007342785131186247, + 0.15624597668647766, + 0.09397174417972565, + 0.006461873650550842, + 0.010230610147118568, + 0.017016902565956116, + 0.011766103096306324, + 0.00951333437114954, + 0.006348083261400461, + 0.014385787770152092, + 0.01111201848834753, + 0.010174895636737347, + 0.027856560423970222, + 0.017313243821263313, + 0.014077836647629738, + 0.008544971235096455, + 0.008799170143902302, + 0.012748577632009983, + 0.007569191511720419, + 0.005547833163291216, + 0.013203395530581474, + 0.003740238957107067, + 0.007162022870033979, + 0.004734774585813284, + 0.0059922426007688046, + 0.008683301508426666, + 0.002618426689878106, + 0.0037604544777423143, + 0.006280536763370037, + 0.0032842098735272884, + 0.0035750523675233126, + 0.004449063446372747, + 0.00859775859862566, + 0.0033423418644815683, + 0.002175187226384878, + 0.003852301277220249, + 0.00513492850586772, + 0.0038273888640105724, + 0.006367059424519539, + 0.008865938521921635, + 0.00977601669728756, + 0.009202245622873306, + 0.002980046672746539, + 0.006121991667896509, + 0.0024023940786719322, + 0.002590389456599951, + 0.0037261939141899347, + 0.006239664740860462, + 0.01182275079190731, + 0.0034096427261829376, + 0.0036090079229325056, + 0.021076912060379982, + 0.007744236849248409, + 0.009962212294340134, + 0.009514573030173779, + 0.013839597813785076, + 0.017623327672481537, + 0.007441976107656956, + 0.0064872498624026775, + 0.016999626532197, + 0.011449670419096947, + 0.007787132635712624, + 0.008076606318354607, + 0.02448565512895584, + 0.013614214025437832, + 0.013995482586324215, + 0.008865082636475563, + 0.007578314747661352, + 0.005635692737996578, + 0.004051997791975737, + 0.004713856149464846, + 0.012746085412800312, + 0.007141753099858761, + 0.005024167243391275, + 0.005540754646062851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0008163695456460118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193400101736188 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.003193400101736188, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193400101736188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15122872591018677, + 0.09678315371274948, + 0.06427770853042603, + 0.053688887506723404, + 0.05496927723288536, + 0.0706792026758194, + 0.17599572241306305, + 0.11134807765483856, + 0.1261076182126999, + 0.09492155909538269 + ], + "credit_map": [ + 0.012924005277454853, + 0.014255342073738575, + 0.022619035094976425, + 0.025979215279221535, + 0.009272023104131222, + 0.005490225274115801, + 0.03377794846892357, + 0.026910951361060143, + 0.0067384447902441025, + 0.010882479138672352, + 0.014336622320115566, + 0.010203985497355461, + 0.011907828971743584, + 0.00805626343935728, + 0.02206609584391117, + 0.012591447681188583, + 0.008516555652022362, + 0.006799790542572737, + 0.010171637870371342, + 0.011575073003768921, + 0.005056166090071201, + 0.005834896117448807, + 0.009434182196855545, + 0.006889406591653824, + 0.006684239022433758, + 0.011575503274798393, + 0.00542542664334178, + 0.0046077026054263115, + 0.007784583605825901, + 0.007613106165081263, + 0.00454961322247982, + 0.0054487185552716255, + 0.005459521897137165, + 0.010534255765378475, + 0.005522334948182106, + 0.0041836961172521114, + 0.007397135719656944, + 0.013885286636650562, + 0.003700743429362774, + 0.004286305978894234, + 0.006834386847913265, + 0.008618156425654888, + 0.0075738499872386456, + 0.006270389072597027, + 0.012883778661489487, + 0.015989167615771294, + 0.005329534411430359, + 0.007179943844676018, + 0.03873233124613762, + 0.008997020311653614, + 0.007647160440683365, + 0.009885664097964764, + 0.0226639024913311, + 0.070867158472538, + 0.009231415577232838, + 0.007971067912876606, + 0.012386620044708252, + 0.009965641424059868, + 0.014493273571133614, + 0.015622451901435852, + 0.014259479939937592, + 0.020234325900673866, + 0.015443875454366207, + 0.008942412212491035, + 0.023339802399277687, + 0.014534424990415573, + 0.009514261968433857, + 0.010655926540493965, + 0.029259653761982918, + 0.015633486211299896, + 0.010848615318536758, + 0.012321466580033302, + 0.012102892622351646, + 0.009814189746975899, + 0.007993115112185478, + 0.0076203057542443275, + 0.021769026294350624, + 0.012317627668380737, + 0.008216160349547863, + 0.015088238753378391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.003193400101736188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a close-up shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily and slowly upward as a traditional bamboo steamer lid is lifted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are plump, round dumplings arranged on emerald green leaves visible inside the steamer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dumpling wrappers appear nearly translucent with a smooth, moist sheen reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, soft overhead light creating a hazy atmosphere from rising white steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015123370103538036 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015123370103538036, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015123370103538036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14817757904529572, + 0.12276613712310791, + 0.08872582018375397, + 0.04629816487431526, + 0.04432305321097374, + 0.07944748550653458, + 0.023851655423641205, + 0.39923974871635437, + 0.020150478929281235, + 0.027019904926419258 + ], + "credit_map": [ + 0.013347098603844643, + 0.01637529395520687, + 0.026067761704325676, + 0.02197241224348545, + 0.013624869287014008, + 0.0071238502860069275, + 0.025182314217090607, + 0.024483969435095787, + 0.00932422373443842, + 0.016852226108312607, + 0.020599549636244774, + 0.015455033630132675, + 0.014834568835794926, + 0.00846133567392826, + 0.021600186824798584, + 0.01563902385532856, + 0.009718162938952446, + 0.021677928045392036, + 0.014538739807903767, + 0.01016567274928093, + 0.0067536053247749805, + 0.009434796869754791, + 0.010353446006774902, + 0.006083470769226551, + 0.005923247430473566, + 0.01118903886526823, + 0.004241629503667355, + 0.004269108641892672, + 0.004526302684098482, + 0.008176276460289955, + 0.0038220516871660948, + 0.004150508902966976, + 0.005274534225463867, + 0.006316839251667261, + 0.004118833690881729, + 0.00338377570733428, + 0.007702630013227463, + 0.011059924028813839, + 0.0033371527679264545, + 0.0031293623615056276, + 0.008325833827257156, + 0.010524681769311428, + 0.009854009374976158, + 0.006531891878694296, + 0.018724629655480385, + 0.010555190034210682, + 0.009342106990516186, + 0.005589139647781849, + 0.005279070232063532, + 0.0012137512676417828, + 0.001485961489379406, + 0.0014060784596949816, + 0.0023570838384330273, + 0.007890193723142147, + 0.002347698900848627, + 0.0018718182109296322, + 0.09639377146959305, + 0.03162308409810066, + 0.03235155716538429, + 0.01908307708799839, + 0.07288507372140884, + 0.10051664710044861, + 0.02780112624168396, + 0.01858539506793022, + 0.0026350729167461395, + 0.002070364076644182, + 0.0013498057378455997, + 0.0014492060290649533, + 0.005642589647322893, + 0.0028539560735225677, + 0.0024734900798648596, + 0.001675994717516005, + 0.0033098715357482433, + 0.002232735976576805, + 0.0030788101721554995, + 0.003105075331404805, + 0.005931155290454626, + 0.0033624516800045967, + 0.002985710045322776, + 0.003014094429090619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0015123370103538036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_tensor([5992, 5992])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_tensor([5992, 5992])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..25567a136124fbb5bc385ab624c37d54bd914724 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_tensor([5992, 5992])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly unstable handheld medium shot, shaky and volatile, shakes violently, as if the photographer is running across uneven ground, desperately trying to catch up and lock the subject in the center of the frame. Across the vast wasteland, bleached white by the glaring midday sun, a weathered wooden cart drifts at a breathtaking speed that defies all physics. The cart is made of rough, varying shades of wood, riddled with cracked grain and deep scratches, its corners reinforced with rusted iron sheets and coarse hemp rope. The body of the cart is bare. Its two massive, solid wooden wheels, worn to the point of losing their roundness, screech with the intense friction of the ground, whipping its tail wildly, carving an exaggerated, graceful arc across the ochre, dry, polygonal earth. As the wheels slid sideways, a vast, swirling cloud of ochre-yellow dust instantly rose, obscuring the sky like a thick fog and nearly engulfing the cart itself, leaving only a blurry silhouette darting wildly through the dust. The entire image, shaken violently, presents a chaotic yet powerful dynamic beauty. Combined with the desolate background and surreal movement, it creates a post-apocalyptic aesthetic that is both desolate and shocking.", + "seed": 21098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_5992_seed_21098_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24610746279358864, + "vlm_logprob": -0.24610746279358864, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.24610746279358864, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.980385959148407 + ], + "opposite_logprob_tokens": [ + -0.480385959148407 + ], + "logprob_reward": -0.980385959148407, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.980385959148407, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.010284182615578175, + 0.019713884219527245, + 0.007876677438616753, + 0.05013458430767059, + 0.6988532543182373, + 0.03423277661204338, + 0.03211022913455963, + 0.05775211378931999, + 0.06402512639760971, + 0.02501722238957882 + ], + "credit_map": [ + 0.001511452253907919, + 0.0014071398181840777, + 0.0011622459860518575, + 0.0013123275712132454, + 0.0012460961006581783, + 0.0015664170496165752, + 0.0007987710414454341, + 0.0012797318631783128, + 0.002133337315171957, + 0.003660176880657673, + 0.0021902320440858603, + 0.0016100980574265122, + 0.003689828095957637, + 0.001613824162632227, + 0.0013762225862592459, + 0.0034401623997837305, + 0.0007999374065548182, + 0.001078554429113865, + 0.001288433326408267, + 0.0008929708856157959, + 0.0008102341089397669, + 0.0011180617148056626, + 0.001037722104229033, + 0.0008507629972882569, + 0.002608560724183917, + 0.0016613580519333482, + 0.0019498916808515787, + 0.0018529262160882354, + 0.0022025650832802057, + 0.017408574000000954, + 0.003950899932533503, + 0.01849980279803276, + 0.01443941704928875, + 0.009718996472656727, + 0.013085484504699707, + 0.00679387291893363, + 0.03935757651925087, + 0.3038085401058197, + 0.022540336474776268, + 0.2891090214252472, + 0.0025253277271986008, + 0.0029392377473413944, + 0.0022716224193573, + 0.003154369769617915, + 0.004622326232492924, + 0.004283312242478132, + 0.007230008486658335, + 0.007206570357084274, + 0.0019563876558095217, + 0.002041390398517251, + 0.004453644156455994, + 0.006497665774077177, + 0.003594831097871065, + 0.0035654036328196526, + 0.00611340394243598, + 0.0038874999154359102, + 0.004329954273998737, + 0.012110471725463867, + 0.00501701096072793, + 0.010257714428007603, + 0.015817172825336456, + 0.0033480229321867228, + 0.003619495080783963, + 0.0032522687688469887, + 0.004781370982527733, + 0.00888774637132883, + 0.008974714204668999, + 0.011370662599802017, + 0.006548508070409298, + 0.007346222177147865, + 0.008724232204258442, + 0.0073916614055633545, + 0.001983117777854204, + 0.002711037639528513, + 0.0034032980911433697, + 0.006581723224371672, + 0.002316064201295376, + 0.002736335853114724, + 0.003252115799114108, + 0.0020335279405117035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.980385959148407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594336092472076 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.020594336092472076, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594336092472076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02114374376833439, + 0.018478641286492348, + 0.02417578548192978, + 0.03376069664955139, + 0.5356227159500122, + 0.052254483103752136, + 0.09505365043878555, + 0.110324926674366, + 0.07641297578811646, + 0.03277232497930527 + ], + "credit_map": [ + 0.0042691659182310104, + 0.002799843205139041, + 0.0025493470020592213, + 0.0023443950340151787, + 0.0024206587113440037, + 0.002544775139540434, + 0.0018976895371451974, + 0.002317868871614337, + 0.002851091790944338, + 0.0026619723066687584, + 0.002210421022027731, + 0.0019551159348338842, + 0.0022149826399981976, + 0.0017444950062781572, + 0.0018316946225240827, + 0.0030088690109550953, + 0.002711305860430002, + 0.003197171026840806, + 0.00321842054836452, + 0.001905178651213646, + 0.0021168452221900225, + 0.0048403991386294365, + 0.003531681140884757, + 0.002654783660545945, + 0.003766497829928994, + 0.0021205514203757048, + 0.0023309921380132437, + 0.0022117604967206717, + 0.0023807487450540066, + 0.009303638711571693, + 0.0033754019532352686, + 0.008271107450127602, + 0.0110020125284791, + 0.007461068220436573, + 0.010023015551269054, + 0.00564250722527504, + 0.030050402507185936, + 0.23239168524742126, + 0.01760166697204113, + 0.2214503437280655, + 0.0028637107461690903, + 0.002990418579429388, + 0.0025856285355985165, + 0.0022957096807658672, + 0.014652995392680168, + 0.015542712062597275, + 0.004874007310718298, + 0.006449299864470959, + 0.004957692697644234, + 0.005978780798614025, + 0.01144222542643547, + 0.019746772944927216, + 0.010660417377948761, + 0.010293800383806229, + 0.018367936834692955, + 0.013606026768684387, + 0.006700792349874973, + 0.015558848157525063, + 0.009903033263981342, + 0.019748512655496597, + 0.023505400866270065, + 0.009303118102252483, + 0.011588726192712784, + 0.01401649508625269, + 0.005827948916703463, + 0.009674903005361557, + 0.010162892751395702, + 0.013449180871248245, + 0.00832870602607727, + 0.007302071899175644, + 0.010529622435569763, + 0.01113765686750412, + 0.003784819273278117, + 0.0028959871269762516, + 0.0033893927466124296, + 0.0023984925355762243, + 0.005645319353789091, + 0.004022958222776651, + 0.004403376020491123, + 0.006231979466974735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.020594336092472076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865443706512451 + ], + "opposite_logprob_tokens": [ + -3.048654556274414 + ], + "logprob_reward": -0.04865443706512451, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865443706512451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031532466411590576, + 0.16769784688949585, + 0.06749801337718964, + 0.1285121887922287, + 0.3670826554298401, + 0.023456832394003868, + 0.03752671927213669, + 0.09047242999076843, + 0.06463849544525146, + 0.021582327783107758 + ], + "credit_map": [ + 0.002546811942011118, + 0.005280553828924894, + 0.004014276433736086, + 0.004490778781473637, + 0.004194970708340406, + 0.0038881422951817513, + 0.0028435506392270327, + 0.00427338108420372, + 0.029174549505114555, + 0.029450438916683197, + 0.018929464742541313, + 0.012832949869334698, + 0.02384386956691742, + 0.01403994020074606, + 0.013315452262759209, + 0.026111183688044548, + 0.00633338512852788, + 0.006288257893174887, + 0.014520812779664993, + 0.008634439669549465, + 0.006042524240911007, + 0.009426163509488106, + 0.008133656345307827, + 0.008118773810565472, + 0.006375977303832769, + 0.0061947316862642765, + 0.004345759283751249, + 0.004802987445145845, + 0.0068743363954126835, + 0.04363761097192764, + 0.011446814984083176, + 0.04483397677540779, + 0.00808725506067276, + 0.0050213392823934555, + 0.0068898587487638, + 0.0035769811365753412, + 0.020851846784353256, + 0.15907417237758636, + 0.011987327598035336, + 0.151593878865242, + 0.0017525997245684266, + 0.0014636139385402203, + 0.002684862120077014, + 0.0023641616571694613, + 0.0036542394664138556, + 0.0033332668244838715, + 0.0036273181904107332, + 0.004576769657433033, + 0.0032247668132185936, + 0.0023961032275110483, + 0.0033332041930407286, + 0.0050382488407194614, + 0.006944968365132809, + 0.007006568368524313, + 0.0051543861627578735, + 0.00442847516387701, + 0.005242107901722193, + 0.014229262247681618, + 0.007428737357258797, + 0.014816033653914928, + 0.019647173583507538, + 0.008183625526726246, + 0.008179436437785625, + 0.0127460528165102, + 0.005032301414757967, + 0.008178775198757648, + 0.009346425533294678, + 0.011723733507096767, + 0.0067166900262236595, + 0.006686884444206953, + 0.008941838517785072, + 0.008011853322386742, + 0.0025887128431349993, + 0.002329942537471652, + 0.0032359007745981216, + 0.004540364257991314, + 0.0018384754657745361, + 0.0022354889661073685, + 0.0026141253765672445, + 0.002199318027123809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.04865443706512451 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057763561606407 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.02057763561606407, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057763561606407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013357004150748253, + 0.011532127857208252, + 0.04685966297984123, + 0.014227024279534817, + 0.4527447521686554, + 0.029567977413535118, + 0.2800895869731903, + 0.11012126505374908, + 0.025442879647016525, + 0.016057677567005157 + ], + "credit_map": [ + 0.0018588511738926172, + 0.0013523371890187263, + 0.001916611916385591, + 0.0018665612442418933, + 0.0016268433537334204, + 0.0017791602294892073, + 0.0011479138629510999, + 0.0018087258795276284, + 0.0016764129977673292, + 0.0021633717697113752, + 0.0014606176409870386, + 0.001135871629230678, + 0.0015966790961101651, + 0.0008567231707274914, + 0.0010993201285600662, + 0.0015431320061907172, + 0.005282080732285976, + 0.008185218088328838, + 0.005227834451943636, + 0.003433969570323825, + 0.004225501324981451, + 0.007951564155519009, + 0.007559632416814566, + 0.004993865732103586, + 0.0013829648960381746, + 0.0008383755339309573, + 0.0009311455069109797, + 0.0009435866377316415, + 0.001081374124623835, + 0.003673702711239457, + 0.0019531731959432364, + 0.003422702196985483, + 0.009242946282029152, + 0.006468133069574833, + 0.008378011174499989, + 0.004451941233128309, + 0.025422178208827972, + 0.1968316286802292, + 0.014644994400441647, + 0.18730492889881134, + 0.00178715109359473, + 0.0022622195538133383, + 0.0029709632508456707, + 0.002515910193324089, + 0.004615996964275837, + 0.005631871521472931, + 0.004697015974670649, + 0.0050868503749370575, + 0.015507102012634277, + 0.01717032492160797, + 0.033039648085832596, + 0.05982458218932152, + 0.031185714527964592, + 0.027407370507717133, + 0.054845381528139114, + 0.041109513491392136, + 0.0073637268505990505, + 0.021942945197224617, + 0.0085678705945611, + 0.01840672641992569, + 0.02990027703344822, + 0.006742989178746939, + 0.007093303371220827, + 0.010103424079716206, + 0.0023047341965138912, + 0.0034322894643992186, + 0.0025547512341290712, + 0.0034707975573837757, + 0.002850674092769623, + 0.003212926210835576, + 0.0037794802337884903, + 0.0038372306153178215, + 0.0014083997812122107, + 0.003320473711937666, + 0.001774223055690527, + 0.001990768825635314, + 0.0025853668339550495, + 0.002206531586125493, + 0.0016670775366947055, + 0.0011048372834920883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.02057763561606407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032494604587555 + ], + "opposite_logprob_tokens": [ + -1.9103249311447144 + ], + "logprob_reward": -0.16032494604587555, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032494604587555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01367229875177145, + 0.026505909860134125, + 0.014865272678434849, + 0.043019723147153854, + 0.47844451665878296, + 0.01683798059821129, + 0.028736863285303116, + 0.14606986939907074, + 0.08047442138195038, + 0.15137314796447754 + ], + "credit_map": [ + 0.0015988238155841827, + 0.0019406435312703252, + 0.001821506186388433, + 0.0016908746911212802, + 0.001648214994929731, + 0.0019236357184126973, + 0.0009318964439444244, + 0.0021167038939893246, + 0.003994626924395561, + 0.00502462824806571, + 0.0028273104690015316, + 0.001836670096963644, + 0.004634523764252663, + 0.0019965320825576782, + 0.0017686107894405723, + 0.00442300783470273, + 0.001475078403018415, + 0.0016424680361524224, + 0.002421767683699727, + 0.0012920721201226115, + 0.0014086227165535092, + 0.0029921759851276875, + 0.0019849149975925684, + 0.0016481723869219422, + 0.0028800200670957565, + 0.001622846000827849, + 0.0013432508567348123, + 0.001346383010968566, + 0.0018757637590169907, + 0.015946650877594948, + 0.0022575785405933857, + 0.015747230499982834, + 0.010314191691577435, + 0.006567572709172964, + 0.009090213105082512, + 0.004656902514398098, + 0.027023814618587494, + 0.20729438960552216, + 0.016071822494268417, + 0.19742561876773834, + 0.001197732868604362, + 0.0012494188267737627, + 0.0007809546659700572, + 0.0008173573878593743, + 0.004145495127886534, + 0.004096677992492914, + 0.0022210227325558662, + 0.002329321578145027, + 0.0023683616891503334, + 0.0021511137019842863, + 0.0036506180185824633, + 0.003424149239435792, + 0.00531643396243453, + 0.0052112252451479435, + 0.003238455392420292, + 0.0033765058033168316, + 0.008664245717227459, + 0.021736107766628265, + 0.010911297053098679, + 0.0216127447783947, + 0.02422013320028782, + 0.018815934658050537, + 0.017171425744891167, + 0.022937988862395287, + 0.006016546860337257, + 0.019872091710567474, + 0.006228641141206026, + 0.009363451041281223, + 0.015026194974780083, + 0.01008076686412096, + 0.006867597345262766, + 0.007019130513072014, + 0.016375476494431496, + 0.012923851609230042, + 0.02177984081208706, + 0.028663014993071556, + 0.013620742596685886, + 0.016245244070887566, + 0.01973412185907364, + 0.022030849009752274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.16032494604587555 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly unstable handheld medium shot, shaky and volatile, shakes violently, as if the photographer is running across uneven ground, desperately trying to catch up and lock the subject in the center of the frame. Across the vast wasteland, bleached white by the glaring midday sun, a weathered wooden cart drifts at a breathtaking speed that defies all physics. The cart is made of rough, varying shades of wood, riddled with cracked grain and deep scratches, its corners reinforced with rusted iron sheets and coarse hemp rope. The body of the cart is bare. Its two massive, solid wooden wheels, worn to the point of losing their roundness, screech with the intense friction of the ground, whipping its tail wildly, carving an exaggerated, graceful arc across the ochre, dry, polygonal earth. As the wheels slid sideways, a vast, swirling cloud of ochre-yellow dust instantly rose, obscuring the sky like a thick fog and nearly engulfing the cart itself, leaving only a blurry silhouette darting wildly through the dust. The entire image, shaken violently, presents a chaotic yet powerful dynamic beauty. Combined with the desolate background and surreal movement, it creates a post-apocalyptic aesthetic that is both desolate and shocking.", + "seed": 21099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_5992_seed_21099_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22345964536070823, + "vlm_logprob": -0.22345964536070823, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.22345964536070823, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9793574213981628 + ], + "opposite_logprob_tokens": [ + -0.47935742139816284 + ], + "logprob_reward": -0.9793574213981628, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9793574213981628, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11269200593233109, + 0.07111262530088425, + 0.07435599714517593, + 0.1283206045627594, + 0.09804778546094894, + 0.2612247169017792, + 0.0492062047123909, + 0.0628589391708374, + 0.07807398587465286, + 0.06410714983940125 + ], + "credit_map": [ + 0.01603079028427601, + 0.013833905570209026, + 0.017067808657884598, + 0.018164964392781258, + 0.011584973894059658, + 0.013893700204789639, + 0.008077702485024929, + 0.014038166962563992, + 0.010165059939026833, + 0.00917841400951147, + 0.014971110969781876, + 0.0061179096810519695, + 0.006286117248237133, + 0.0071559203788638115, + 0.009266174398362637, + 0.00797192007303238, + 0.009882164187729359, + 0.009933795779943466, + 0.008325211703777313, + 0.007657654583454132, + 0.0075250728987157345, + 0.010366271249949932, + 0.009730970486998558, + 0.01093485951423645, + 0.012791830115020275, + 0.01397929061204195, + 0.01019223965704441, + 0.0224198866635561, + 0.015431658364832401, + 0.011630869470536709, + 0.023861147463321686, + 0.018013671040534973, + 0.010224558413028717, + 0.011408972553908825, + 0.012282723560929298, + 0.015298508107662201, + 0.012171367183327675, + 0.00988932978361845, + 0.01542764063924551, + 0.01134469173848629, + 0.017762713134288788, + 0.016229307278990746, + 0.02507631853222847, + 0.019245922565460205, + 0.020915523171424866, + 0.04117845371365547, + 0.04530259966850281, + 0.07551385462284088, + 0.00234791892580688, + 0.008097554557025433, + 0.005728716496378183, + 0.00423757778480649, + 0.005577269941568375, + 0.006995361763983965, + 0.006533621344715357, + 0.009688182733952999, + 0.005795889068394899, + 0.008744391612708569, + 0.009232992306351662, + 0.006982618477195501, + 0.007295619230717421, + 0.006754057481884956, + 0.010467568412423134, + 0.00758580444380641, + 0.007679864764213562, + 0.0108871478587389, + 0.008107097819447517, + 0.011483725160360336, + 0.0097473980858922, + 0.010257137939333916, + 0.010555637069046497, + 0.009355979971587658, + 0.006319237872958183, + 0.004613921511918306, + 0.007768841926008463, + 0.005382133182138205, + 0.011983000673353672, + 0.011173475533723831, + 0.009061707183718681, + 0.007804824970662594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.9793574213981628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329820580780506 + ], + "opposite_logprob_tokens": [ + -3.7732982635498047 + ], + "logprob_reward": -0.02329820580780506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329820580780506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15694153308868408, + 0.07050856202840805, + 0.06884140521287918, + 0.07949955016374588, + 0.14711220562458038, + 0.08524509519338608, + 0.08283055573701859, + 0.04642821475863457, + 0.10485491901636124, + 0.15773802995681763 + ], + "credit_map": [ + 0.02245694398880005, + 0.015301547013223171, + 0.023691752925515175, + 0.03771063685417175, + 0.01188050489872694, + 0.013469856232404709, + 0.01578967086970806, + 0.016640618443489075, + 0.011204206384718418, + 0.007593505084514618, + 0.01063520647585392, + 0.00806825514882803, + 0.007626829203218222, + 0.005122204776853323, + 0.011816118843853474, + 0.00844222865998745, + 0.006286106538027525, + 0.00865984708070755, + 0.008430213667452335, + 0.007581992074847221, + 0.009874257259070873, + 0.006225762888789177, + 0.01394541747868061, + 0.007837806828320026, + 0.007991524413228035, + 0.01005021296441555, + 0.013177366927266121, + 0.007937113754451275, + 0.007241870276629925, + 0.009570818394422531, + 0.009415389969944954, + 0.01411523949354887, + 0.010794022120535374, + 0.013816778548061848, + 0.02062109485268593, + 0.01807192526757717, + 0.012810960412025452, + 0.019049208611249924, + 0.022272951900959015, + 0.029675256460905075, + 0.007275322452187538, + 0.013781829737126827, + 0.00853933859616518, + 0.007689299061894417, + 0.007404221221804619, + 0.010962010361254215, + 0.012839826755225658, + 0.01675323210656643, + 0.005856181960552931, + 0.012172123417258263, + 0.007112437393516302, + 0.009404397569596767, + 0.008608432486653328, + 0.009633214212954044, + 0.016184808686375618, + 0.013858948834240437, + 0.0031381279695779085, + 0.004543730057775974, + 0.0041779386810958385, + 0.008150353096425533, + 0.0033253103028982878, + 0.004919377155601978, + 0.00823482871055603, + 0.009938547387719154, + 0.0059480504132807255, + 0.009683635085821152, + 0.008164770901203156, + 0.014356137253344059, + 0.017079394310712814, + 0.020541775971651077, + 0.014942547306418419, + 0.014138593338429928, + 0.016369938850402832, + 0.023987775668501854, + 0.021458212286233902, + 0.020344004034996033, + 0.02243763394653797, + 0.022324511781334877, + 0.013766812160611153, + 0.017049133777618408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.02329820580780506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13211609423160553, + 0.10342857986688614, + 0.1747620701789856, + 0.10874707251787186, + 0.07551159709692001, + 0.10025633871555328, + 0.05846700072288513, + 0.08072210103273392, + 0.08194638788700104, + 0.08404276520013809 + ], + "credit_map": [ + 0.017086373642086983, + 0.01306312158703804, + 0.01327210757881403, + 0.02356201596558094, + 0.014126059599220753, + 0.017750877887010574, + 0.008271900936961174, + 0.024983642622828484, + 0.010025480762124062, + 0.009907080791890621, + 0.012935624457895756, + 0.015886597335338593, + 0.008973157964646816, + 0.011262434534728527, + 0.012202201411128044, + 0.022235997021198273, + 0.015500581823289394, + 0.02188396453857422, + 0.024453260004520416, + 0.018802424892783165, + 0.022263333201408386, + 0.01673896238207817, + 0.031210782006382942, + 0.02390875294804573, + 0.014922992326319218, + 0.013434609398245811, + 0.014102177694439888, + 0.0124084847047925, + 0.015154683031141758, + 0.012019173242151737, + 0.013883206993341446, + 0.012821749784052372, + 0.00597828533500433, + 0.007005457766354084, + 0.00861065462231636, + 0.010970154777169228, + 0.007924077101051807, + 0.009389604441821575, + 0.012577442452311516, + 0.013055921532213688, + 0.0069833374582231045, + 0.011576221324503422, + 0.00883621908724308, + 0.008594108745455742, + 0.010419716127216816, + 0.013271191157400608, + 0.018712297081947327, + 0.021863248199224472, + 0.004406144376844168, + 0.00928803626447916, + 0.005542990751564503, + 0.006160170305520296, + 0.006161794997751713, + 0.007743999361991882, + 0.006967993453145027, + 0.012195870280265808, + 0.0058309948071837425, + 0.013384203426539898, + 0.011824130080640316, + 0.010047274641692638, + 0.007401687558740377, + 0.008044248446822166, + 0.010255980305373669, + 0.013933579437434673, + 0.0068794251419603825, + 0.010309260338544846, + 0.007404489442706108, + 0.011121179908514023, + 0.00991632416844368, + 0.01174635998904705, + 0.009436961263418198, + 0.015132386237382889, + 0.0057493108324706554, + 0.01176357176154852, + 0.010786443948745728, + 0.012130530551075935, + 0.012520236894488335, + 0.01418362557888031, + 0.007339419797062874, + 0.0095696235075593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.011075684800744057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486430823802948 + ], + "opposite_logprob_tokens": [ + -3.048643112182617 + ], + "logprob_reward": -0.0486430823802948, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486430823802948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10984726250171661, + 0.10131990164518356, + 0.0785616934299469, + 0.13455098867416382, + 0.18356570601463318, + 0.05798131600022316, + 0.10686402767896652, + 0.04103579744696617, + 0.08126743137836456, + 0.10500594973564148 + ], + "credit_map": [ + 0.01337068248540163, + 0.015313134528696537, + 0.014674186706542969, + 0.01778503879904747, + 0.015233820304274559, + 0.012805375270545483, + 0.007051811087876558, + 0.0136132026091218, + 0.009846200235188007, + 0.011956539936363697, + 0.012738200835883617, + 0.01599854975938797, + 0.009275242686271667, + 0.0113835409283638, + 0.012265278026461601, + 0.01785634458065033, + 0.006023169495165348, + 0.006749087478965521, + 0.015296229161322117, + 0.0049582310020923615, + 0.010038159787654877, + 0.006663153879344463, + 0.01003425382077694, + 0.018799401819705963, + 0.011826753616333008, + 0.01101393811404705, + 0.01862172596156597, + 0.012951776385307312, + 0.010420947335660458, + 0.018626198172569275, + 0.018844682723283768, + 0.03224495053291321, + 0.012003428302705288, + 0.014503746293485165, + 0.027068667113780975, + 0.021962130442261696, + 0.016394779086112976, + 0.02440117485821247, + 0.028936658054590225, + 0.03829509764909744, + 0.006824724841862917, + 0.012805377133190632, + 0.00833614356815815, + 0.005261392332613468, + 0.0058552902191877365, + 0.006139081437140703, + 0.004943622276186943, + 0.007815676741302013, + 0.008948908187448978, + 0.011803526431322098, + 0.012454199604690075, + 0.012834114022552967, + 0.013390446081757545, + 0.013072846457362175, + 0.017106320708990097, + 0.017253657802939415, + 0.0034525874070823193, + 0.006059355568140745, + 0.004178434610366821, + 0.004898614715784788, + 0.0051608094945549965, + 0.004996724892407656, + 0.005670919548720121, + 0.006618347018957138, + 0.006566754542291164, + 0.013018549419939518, + 0.006819697562605143, + 0.012935829348862171, + 0.012153862975537777, + 0.013237553648650646, + 0.008503608405590057, + 0.008031566627323627, + 0.010842547751963139, + 0.01510304119437933, + 0.010706226341426373, + 0.013702088966965675, + 0.018815938383340836, + 0.019259819760918617, + 0.008341625332832336, + 0.008234649896621704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0486430823802948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054923832416534424 + ], + "opposite_logprob_tokens": [ + -2.9299237728118896 + ], + "logprob_reward": -0.054923832416534424, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054923832416534424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21725153923034668, + 0.06382252275943756, + 0.0495133213698864, + 0.03165362775325775, + 0.13287034630775452, + 0.05970030277967453, + 0.09149126708507538, + 0.04804284870624542, + 0.09439857304096222, + 0.21125560998916626 + ], + "credit_map": [ + 0.03137211501598358, + 0.023257233202457428, + 0.025884291157126427, + 0.05125793442130089, + 0.016250861808657646, + 0.020831413567066193, + 0.016511160880327225, + 0.03188655897974968, + 0.0067679621279239655, + 0.007184614427387714, + 0.00856146588921547, + 0.007720417808741331, + 0.005580213386565447, + 0.01117476262152195, + 0.006980119273066521, + 0.009852972812950611, + 0.004375402815639973, + 0.005021213088184595, + 0.011144114658236504, + 0.0033178641460835934, + 0.0054307375103235245, + 0.004219697322696447, + 0.005013776011765003, + 0.010990520939230919, + 0.0027227471582591534, + 0.003590376116335392, + 0.002971828216686845, + 0.005470961332321167, + 0.004200302064418793, + 0.002809216035529971, + 0.005914054811000824, + 0.00397414481267333, + 0.010098947212100029, + 0.009797763079404831, + 0.015046562068164349, + 0.012248383834958076, + 0.01238950714468956, + 0.021142389625310898, + 0.019975412636995316, + 0.03217138722538948, + 0.003359165508300066, + 0.004885971546173096, + 0.00776619603857398, + 0.006476516369730234, + 0.005961362272500992, + 0.006792461033910513, + 0.013877945020794868, + 0.010580689646303654, + 0.00834602303802967, + 0.009527018293738365, + 0.009516672231256962, + 0.00835656188428402, + 0.009619185701012611, + 0.012110400013625622, + 0.015434443019330502, + 0.018580976873636246, + 0.004980977158993483, + 0.0037451419048011303, + 0.004677013959735632, + 0.003893026616424322, + 0.0034781349822878838, + 0.00591243477538228, + 0.01065357681363821, + 0.010702547617256641, + 0.003938829526305199, + 0.0071841636672616005, + 0.006591662764549255, + 0.009164983406662941, + 0.018767643719911575, + 0.02109491638839245, + 0.014228595420718193, + 0.013427781872451305, + 0.026909176260232925, + 0.027106238529086113, + 0.029496951028704643, + 0.01840383931994438, + 0.028662094846367836, + 0.03334477171301842, + 0.018514208495616913, + 0.02881833165884018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.054923832416534424 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly unstable handheld medium shot, shaky and volatile, shakes violently, as if the photographer is running across uneven ground, desperately trying to catch up and lock the subject in the center of the frame. Across the vast wasteland, bleached white by the glaring midday sun, a weathered wooden cart drifts at a breathtaking speed that defies all physics. The cart is made of rough, varying shades of wood, riddled with cracked grain and deep scratches, its corners reinforced with rusted iron sheets and coarse hemp rope. The body of the cart is bare. Its two massive, solid wooden wheels, worn to the point of losing their roundness, screech with the intense friction of the ground, whipping its tail wildly, carving an exaggerated, graceful arc across the ochre, dry, polygonal earth. As the wheels slid sideways, a vast, swirling cloud of ochre-yellow dust instantly rose, obscuring the sky like a thick fog and nearly engulfing the cart itself, leaving only a blurry silhouette darting wildly through the dust. The entire image, shaken violently, presents a chaotic yet powerful dynamic beauty. Combined with the desolate background and surreal movement, it creates a post-apocalyptic aesthetic that is both desolate and shocking.", + "seed": 21100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_5992_seed_21100_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.31149540953338145, + "vlm_logprob": -0.31149540953338145, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.31149540953338145, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3176382780075073 + ], + "opposite_logprob_tokens": [ + -0.31763824820518494 + ], + "logprob_reward": -1.3176382780075073, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3176382780075073, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0649680346250534, + 0.12948483228683472, + 0.04650179669260979, + 0.20384371280670166, + 0.17516809701919556, + 0.050443556159734726, + 0.11983150243759155, + 0.08223724365234375, + 0.06848366558551788, + 0.05903756618499756 + ], + "credit_map": [ + 0.008686518296599388, + 0.00901670940220356, + 0.009637034498155117, + 0.005764728877693415, + 0.008275185711681843, + 0.007800733670592308, + 0.00819789245724678, + 0.007589232176542282, + 0.01620650291442871, + 0.008973319083452225, + 0.018940508365631104, + 0.014083626680076122, + 0.015948429703712463, + 0.011071249842643738, + 0.02612142823636532, + 0.018139773979783058, + 0.005327498074620962, + 0.0045365216210484505, + 0.0089340191334486, + 0.005397904198616743, + 0.006251756101846695, + 0.003950233105570078, + 0.006467956583946943, + 0.0056359125301241875, + 0.01168757677078247, + 0.01954382285475731, + 0.013784217648208141, + 0.053413644433021545, + 0.016579587012529373, + 0.022739822044968605, + 0.028442280367016792, + 0.037652768194675446, + 0.006362526211887598, + 0.0071786148473620415, + 0.004159655421972275, + 0.010784127749502659, + 0.010935047641396523, + 0.05747045576572418, + 0.015003781765699387, + 0.06327388435602188, + 0.004578768275678158, + 0.004490127321332693, + 0.005057441536337137, + 0.009375366382300854, + 0.00423779059201479, + 0.005001787096261978, + 0.008404458872973919, + 0.009297812357544899, + 0.01127914059907198, + 0.012610171921551228, + 0.008959494531154633, + 0.022775564342737198, + 0.010120004415512085, + 0.011395442299544811, + 0.023103050887584686, + 0.019588638097047806, + 0.00805372092872858, + 0.013042519800364971, + 0.006292521953582764, + 0.015388701111078262, + 0.009565294720232487, + 0.009202094748616219, + 0.011079858057200909, + 0.009612535126507282, + 0.006274127401411533, + 0.009884537197649479, + 0.007148323580622673, + 0.012442464008927345, + 0.007229416631162167, + 0.008770189248025417, + 0.006984252016991377, + 0.009750355035066605, + 0.004265057388693094, + 0.0076465667225420475, + 0.00454129371792078, + 0.006685700733214617, + 0.010166165418922901, + 0.012528426013886929, + 0.007268805988132954, + 0.005935554392635822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -1.3176382780075073 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020605195313692093 + ], + "opposite_logprob_tokens": [ + -3.8956050872802734 + ], + "logprob_reward": -0.020605195313692093, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020605195313692093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09110631793737411, + 0.08488570153713226, + 0.05844764783978462, + 0.08378994464874268, + 0.20310533046722412, + 0.18329229950904846, + 0.07745925337076187, + 0.13123995065689087, + 0.046869099140167236, + 0.03980438411235809 + ], + "credit_map": [ + 0.008765547536313534, + 0.011707238852977753, + 0.016376538202166557, + 0.014845525845885277, + 0.009253256022930145, + 0.0071553741581737995, + 0.012867949903011322, + 0.010134887881577015, + 0.0072664800100028515, + 0.006813480984419584, + 0.014022323302924633, + 0.00890425220131874, + 0.012685894034802914, + 0.004957521799951792, + 0.01795872114598751, + 0.012277031317353249, + 0.005901857744902372, + 0.005978328175842762, + 0.008262794464826584, + 0.009085778146982193, + 0.007183695212006569, + 0.0058083366602659225, + 0.007976303808391094, + 0.008250555023550987, + 0.011306400410830975, + 0.00873970054090023, + 0.006316995248198509, + 0.015502103604376316, + 0.008688999339938164, + 0.005721720866858959, + 0.016415856778621674, + 0.0110981659963727, + 0.010539257898926735, + 0.008726616390049458, + 0.005748938303440809, + 0.007400876376777887, + 0.011879500932991505, + 0.06936006993055344, + 0.014461897313594818, + 0.07498817145824432, + 0.023662569001317024, + 0.01497688889503479, + 0.016295773908495903, + 0.038094229996204376, + 0.01661333441734314, + 0.01595032587647438, + 0.027338838204741478, + 0.030360359698534012, + 0.007670126389712095, + 0.007803341839462519, + 0.00618285546079278, + 0.018658405169844627, + 0.0077765099704265594, + 0.00835982896387577, + 0.01074987556785345, + 0.010258310474455357, + 0.006991144269704819, + 0.01467923168092966, + 0.0076835015788674355, + 0.012716925702989101, + 0.013540620915591717, + 0.017733938992023468, + 0.025499261915683746, + 0.032395318150520325, + 0.005112394690513611, + 0.008486157283186913, + 0.00404712837189436, + 0.0074991206638514996, + 0.006027353927493095, + 0.006742917466908693, + 0.003982681781053543, + 0.0049713472835719585, + 0.003549558576196432, + 0.006067090202122927, + 0.003014108631759882, + 0.0044118803925812244, + 0.005496706813573837, + 0.007111894898116589, + 0.005439113359898329, + 0.004714034032076597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.020605195313692093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299837484955788 + ], + "opposite_logprob_tokens": [ + -3.7732999324798584 + ], + "logprob_reward": -0.023299837484955788, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299837484955788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14638319611549377, + 0.07418715208768845, + 0.05542341247200966, + 0.05369006097316742, + 0.279359370470047, + 0.09102872014045715, + 0.12008251994848251, + 0.03860781341791153, + 0.04077037796378136, + 0.10046737641096115 + ], + "credit_map": [ + 0.013891922309994698, + 0.014607875607907772, + 0.025412943214178085, + 0.01744619943201542, + 0.01895197294652462, + 0.012773599475622177, + 0.024798214435577393, + 0.018500464037060738, + 0.010653248056769371, + 0.006963385734707117, + 0.011516762897372246, + 0.006978216581046581, + 0.007996803149580956, + 0.005428632255643606, + 0.014732255600392818, + 0.009917850606143475, + 0.008713974617421627, + 0.006140814162790775, + 0.005961722228676081, + 0.00859712902456522, + 0.005807714071124792, + 0.006079778075218201, + 0.006246875040233135, + 0.007875403389334679, + 0.005770494695752859, + 0.006833248306065798, + 0.004243537317961454, + 0.012941203080117702, + 0.005032059270888567, + 0.0030995910055935383, + 0.008518585003912449, + 0.007251342758536339, + 0.008142008446156979, + 0.009219706989824772, + 0.004670314025133848, + 0.007985102012753487, + 0.018508654087781906, + 0.10267924517393112, + 0.017302915453910828, + 0.11085140705108643, + 0.014142435044050217, + 0.008025692775845528, + 0.007824636995792389, + 0.01710587739944458, + 0.008958885446190834, + 0.008420754224061966, + 0.012910800985991955, + 0.013639641925692558, + 0.009296233765780926, + 0.007925821468234062, + 0.010399410501122475, + 0.023188142105937004, + 0.01089564524590969, + 0.01131218671798706, + 0.0268674585968256, + 0.020197631791234016, + 0.0034619485959410667, + 0.004381769336760044, + 0.002761326963081956, + 0.008908838033676147, + 0.004752081818878651, + 0.004137556534260511, + 0.0050752186216413975, + 0.005129072815179825, + 0.0034451710525900126, + 0.006394903175532818, + 0.004025743808597326, + 0.006335841957479715, + 0.005647092591971159, + 0.0067484439350664616, + 0.0038874393794685602, + 0.004285741597414017, + 0.006924794055521488, + 0.0056831734254956245, + 0.004133766051381826, + 0.0076206037774682045, + 0.01668672263622284, + 0.02344810962677002, + 0.013307338580489159, + 0.022662866860628128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.023299837484955788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17469190061092377, + 0.08281166106462479, + 0.036169957369565964, + 0.06282355636358261, + 0.0831013172864914, + 0.1012367233633995, + 0.11024434864521027, + 0.17815831303596497, + 0.09573666751384735, + 0.07502555102109909 + ], + "credit_map": [ + 0.018683722242712975, + 0.022181624546647072, + 0.029430072754621506, + 0.02063691057264805, + 0.019424963742494583, + 0.02039807103574276, + 0.022668806836009026, + 0.021267730742692947, + 0.009273442439734936, + 0.006388076115399599, + 0.011946141719818115, + 0.008088884875178337, + 0.01208166591823101, + 0.008675592951476574, + 0.014964515343308449, + 0.011393342167139053, + 0.004469740204513073, + 0.005904390010982752, + 0.004604080691933632, + 0.003962868824601173, + 0.004803026095032692, + 0.0028350043576210737, + 0.005369699560105801, + 0.004221150651574135, + 0.00429069809615612, + 0.005383416078984737, + 0.0035705461632460356, + 0.012412658892571926, + 0.0060196127742528915, + 0.006412360817193985, + 0.014222150668501854, + 0.010512109845876694, + 0.008209292776882648, + 0.007516934536397457, + 0.004795622546225786, + 0.01162638608366251, + 0.0062128156423568726, + 0.014154673554003239, + 0.011869884096086025, + 0.01871570758521557, + 0.01460627093911171, + 0.009596754796802998, + 0.008227016776800156, + 0.019390562549233437, + 0.009539242833852768, + 0.008940092287957668, + 0.015459481626749039, + 0.015477306209504604, + 0.008734217844903469, + 0.00846813339740038, + 0.009102030657231808, + 0.024349750950932503, + 0.00874181929975748, + 0.010006380267441273, + 0.023543361574411392, + 0.017298653721809387, + 0.013752412050962448, + 0.03634744510054588, + 0.011955163441598415, + 0.02760954573750496, + 0.021059080958366394, + 0.02017810381948948, + 0.024154284968972206, + 0.023102277889847755, + 0.010582542978227139, + 0.015048425644636154, + 0.01213701069355011, + 0.016633830964565277, + 0.008628220297396183, + 0.010741914622485638, + 0.009807901456952095, + 0.012156815268099308, + 0.008860379457473755, + 0.010739129967987537, + 0.008028052747249603, + 0.009445812553167343, + 0.010916274040937424, + 0.011594626121222973, + 0.008435104042291641, + 0.0070061772130429745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.016067780554294586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986595630645752 + ], + "opposite_logprob_tokens": [ + -1.8048659563064575 + ], + "logprob_reward": -0.17986595630645752, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986595630645752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11906636506319046, + 0.07680786401033401, + 0.04600951448082924, + 0.0829290822148323, + 0.23412562906742096, + 0.0763196349143982, + 0.18022939562797546, + 0.05090755596756935, + 0.059760697185993195, + 0.07384423166513443 + ], + "credit_map": [ + 0.011291894130408764, + 0.008575920946896076, + 0.014824125915765762, + 0.01319128181785345, + 0.018150899559259415, + 0.013288908638060093, + 0.02205425687134266, + 0.017689069733023643, + 0.011710367165505886, + 0.005080438684672117, + 0.009699837304651737, + 0.007250040303915739, + 0.007978982292115688, + 0.01102451328188181, + 0.013589510694146156, + 0.010474174283444881, + 0.004123733378946781, + 0.0035061475355178118, + 0.003894778899848461, + 0.004591950681060553, + 0.004686822183430195, + 0.008854090236127377, + 0.007500367239117622, + 0.008851624093949795, + 0.006929762661457062, + 0.007070447318255901, + 0.005372123327106237, + 0.018491055816411972, + 0.007926073856651783, + 0.009956893511116505, + 0.01437269989401102, + 0.012810027226805687, + 0.007518517319113016, + 0.00861802976578474, + 0.0056586130522191525, + 0.008818184956908226, + 0.013897626660764217, + 0.08106254041194916, + 0.019380195066332817, + 0.08917192369699478, + 0.008212968707084656, + 0.005035328678786755, + 0.006609671283513308, + 0.010848890990018845, + 0.010640555061399937, + 0.008543970994651318, + 0.012097327038645744, + 0.014330926351249218, + 0.013258435763418674, + 0.01104750856757164, + 0.015041643753647804, + 0.03598039597272873, + 0.015400458127260208, + 0.016833176836371422, + 0.04257146641612053, + 0.030096309259533882, + 0.003725101938471198, + 0.005033955443650484, + 0.0035902236122637987, + 0.007511841133236885, + 0.004956010729074478, + 0.0054825302213430405, + 0.010323170572519302, + 0.0102847246453166, + 0.004740954842418432, + 0.007908393628895283, + 0.006919562350958586, + 0.00952302385121584, + 0.007820931263267994, + 0.007386532612144947, + 0.009103338234126568, + 0.006357958074659109, + 0.00958671048283577, + 0.009139581583440304, + 0.006499755661934614, + 0.007958861067891121, + 0.012019402347505093, + 0.012755095027387142, + 0.008203298784792423, + 0.007681518327444792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.17986595630645752 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly unstable handheld medium shot, shaky and volatile, shakes violently, as if the photographer is running across uneven ground, desperately trying to catch up and lock the subject in the center of the frame. Across the vast wasteland, bleached white by the glaring midday sun, a weathered wooden cart drifts at a breathtaking speed that defies all physics. The cart is made of rough, varying shades of wood, riddled with cracked grain and deep scratches, its corners reinforced with rusted iron sheets and coarse hemp rope. The body of the cart is bare. Its two massive, solid wooden wheels, worn to the point of losing their roundness, screech with the intense friction of the ground, whipping its tail wildly, carving an exaggerated, graceful arc across the ochre, dry, polygonal earth. As the wheels slid sideways, a vast, swirling cloud of ochre-yellow dust instantly rose, obscuring the sky like a thick fog and nearly engulfing the cart itself, leaving only a blurry silhouette darting wildly through the dust. The entire image, shaken violently, presents a chaotic yet powerful dynamic beauty. Combined with the desolate background and surreal movement, it creates a post-apocalyptic aesthetic that is both desolate and shocking.", + "seed": 21101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000086/group_9/index_5992_seed_21101_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1977822294458747, + "vlm_logprob": -0.1977822294458747, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.1977822294458747, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9049320816993713 + ], + "opposite_logprob_tokens": [ + -0.5299320816993713 + ], + "logprob_reward": -0.9049320816993713, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9049320816993713, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05653538182377815, + 0.4701615571975708, + 0.014761114493012428, + 0.032413262873888016, + 0.06573730707168579, + 0.04871182143688202, + 0.07490352541208267, + 0.11175067722797394, + 0.044832825660705566, + 0.0801924616098404 + ], + "credit_map": [ + 0.0033847910817712545, + 0.005670410115271807, + 0.006726201623678207, + 0.005128229968249798, + 0.009618631564080715, + 0.007419903296977282, + 0.011192885227501392, + 0.0073943305760622025, + 0.14096319675445557, + 0.01929103210568428, + 0.18956944346427917, + 0.02806367538869381, + 0.010772446170449257, + 0.014278626069426537, + 0.01751909777522087, + 0.04970405623316765, + 0.001922264345921576, + 0.0014176146360114217, + 0.0033508504275232553, + 0.0011858271900564432, + 0.002324066124856472, + 0.0015479088760912418, + 0.001472192583605647, + 0.0015403908910229802, + 0.003161220345646143, + 0.003477318910881877, + 0.004569505341351032, + 0.0035465324763208628, + 0.005449092015624046, + 0.002887960057705641, + 0.004204762168228626, + 0.005116872955113649, + 0.006349332630634308, + 0.006702670361846685, + 0.00998612679541111, + 0.005476984661072493, + 0.00786084309220314, + 0.0064597236923873425, + 0.009815851226449013, + 0.013085778802633286, + 0.0044008418917655945, + 0.005050182808190584, + 0.009043204598128796, + 0.00732672493904829, + 0.006261487025767565, + 0.004022803157567978, + 0.006869102828204632, + 0.005737480707466602, + 0.008850332349538803, + 0.013846641406416893, + 0.009345498867332935, + 0.010332340374588966, + 0.009789935313165188, + 0.007755594793707132, + 0.007692981511354446, + 0.007290197536349297, + 0.005151562392711639, + 0.00728472089394927, + 0.013022292405366898, + 0.010309403762221336, + 0.014816253446042538, + 0.011345085687935352, + 0.03206248581409454, + 0.017758874222636223, + 0.0035956762731075287, + 0.006144516635686159, + 0.003965682815760374, + 0.007927372120320797, + 0.005657769739627838, + 0.004941214341670275, + 0.007239720318466425, + 0.005360872019082308, + 0.010141091421246529, + 0.008443836122751236, + 0.008667297661304474, + 0.012809821404516697, + 0.009142720140516758, + 0.009303059428930283, + 0.014737699180841446, + 0.006946935318410397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.9049320816993713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208228327333927 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208228327333927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208228327333927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062290582805871964, + 0.47492989897727966, + 0.016745027154684067, + 0.018155615776777267, + 0.0234979260712862, + 0.04213925078511238, + 0.04879138246178627, + 0.18120625615119934, + 0.0371306948363781, + 0.0951133742928505 + ], + "credit_map": [ + 0.00431677233427763, + 0.006354697048664093, + 0.009266803041100502, + 0.005912275053560734, + 0.008977008983492851, + 0.01248573325574398, + 0.007367373444139957, + 0.007609917316585779, + 0.14346379041671753, + 0.019831456243991852, + 0.19182723760604858, + 0.027469174936413765, + 0.010441959835588932, + 0.014403791166841984, + 0.01743389293551445, + 0.05005856975913048, + 0.002663365565240383, + 0.001476568402722478, + 0.004177252296358347, + 0.0013205079594627023, + 0.002457438735291362, + 0.001527694403193891, + 0.0014209182700142264, + 0.0017012812895700336, + 0.0016739745624363422, + 0.002326266374439001, + 0.002403744263574481, + 0.0018466475885361433, + 0.003329407423734665, + 0.001478550024330616, + 0.0028028939850628376, + 0.0022941334173083305, + 0.0020633654203265905, + 0.002343460451811552, + 0.003793737618252635, + 0.003232453251257539, + 0.002704418497160077, + 0.0020605954341590405, + 0.004260257817804813, + 0.0030396373476833105, + 0.003495927434414625, + 0.004898600280284882, + 0.007257605902850628, + 0.00557114090770483, + 0.005581426899880171, + 0.0038656299002468586, + 0.006018707994371653, + 0.005450209602713585, + 0.004556052852421999, + 0.008868128061294556, + 0.006844448857009411, + 0.006566897500306368, + 0.006973535753786564, + 0.00525323860347271, + 0.004920424427837133, + 0.004808655474334955, + 0.014202366583049297, + 0.018029728904366493, + 0.02338387817144394, + 0.0176734309643507, + 0.02411453239619732, + 0.03523051738739014, + 0.023663237690925598, + 0.024908559396862984, + 0.0032089052256196737, + 0.004535655491054058, + 0.003253333270549774, + 0.005404296331107616, + 0.004280965309590101, + 0.0058295209892094135, + 0.00594509718939662, + 0.004672921262681484, + 0.009904777631163597, + 0.009932251647114754, + 0.009251361712813377, + 0.014439822174608707, + 0.011788886040449142, + 0.015373000875115395, + 0.012255235575139523, + 0.012168039567768574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.014208228327333927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543123215436935 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543123215436935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543123215436935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0933675467967987, + 0.4072476625442505, + 0.027599509805440903, + 0.027554359287023544, + 0.027056492865085602, + 0.07861669361591339, + 0.09651527553796768, + 0.06456306576728821, + 0.06422461569309235, + 0.11325475573539734 + ], + "credit_map": [ + 0.0063078575767576694, + 0.008694479241967201, + 0.01538742333650589, + 0.009519807994365692, + 0.01377660222351551, + 0.011486772447824478, + 0.013388275168836117, + 0.014806327410042286, + 0.12105356156826019, + 0.018775098025798798, + 0.1580926924943924, + 0.024188583716750145, + 0.014019913040101528, + 0.011884819716215134, + 0.017725251615047455, + 0.04150775447487831, + 0.0031333842780441046, + 0.0024986271746456623, + 0.0048338063061237335, + 0.002732326043769717, + 0.0039002057164907455, + 0.00260532321408391, + 0.003543475642800331, + 0.0043523614294826984, + 0.002716889139264822, + 0.0037163689266890287, + 0.0035589905455708504, + 0.003078947775065899, + 0.00488342484459281, + 0.0023407903499901295, + 0.003927042242139578, + 0.0033319059293717146, + 0.002190809929743409, + 0.004110636189579964, + 0.003681462025269866, + 0.0030877103563398123, + 0.004601739812642336, + 0.0018946469062939286, + 0.004868846386671066, + 0.0026206413749605417, + 0.007030851673334837, + 0.01107077021151781, + 0.011532711796462536, + 0.009492672048509121, + 0.011822626926004887, + 0.006938462611287832, + 0.009386381134390831, + 0.011342219077050686, + 0.009761612862348557, + 0.01749369129538536, + 0.012921729125082493, + 0.013552544638514519, + 0.014562058262526989, + 0.010062855668365955, + 0.008706373162567616, + 0.009454415179789066, + 0.004861006513237953, + 0.006829974707216024, + 0.0059298803098499775, + 0.010966376401484013, + 0.008134506642818451, + 0.0061701093800365925, + 0.014144009910523891, + 0.007527204230427742, + 0.004400564357638359, + 0.008805366232991219, + 0.006466805003583431, + 0.013835201971232891, + 0.005751557182520628, + 0.007637458387762308, + 0.009508099406957626, + 0.007819564081728458, + 0.0035357053857296705, + 0.003848281456157565, + 0.005824894178658724, + 0.01104976236820221, + 0.025113100185990334, + 0.008291313424706459, + 0.04076085984706879, + 0.01483084075152874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.012543123215436935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189893379807472 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189893379807472, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189893379807472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13135604560375214, + 0.1307590901851654, + 0.02444172464311123, + 0.022225799039006233, + 0.02065778151154518, + 0.02864248678088188, + 0.08717098832130432, + 0.33715420961380005, + 0.04673606529831886, + 0.1708558052778244 + ], + "credit_map": [ + 0.010348250158131123, + 0.012776614166796207, + 0.017074881121516228, + 0.012508190236985683, + 0.024399079382419586, + 0.01720239780843258, + 0.018194271251559258, + 0.018852345645427704, + 0.03609773516654968, + 0.008177991956472397, + 0.0481518879532814, + 0.008962424471974373, + 0.007270634174346924, + 0.0046209655702114105, + 0.00493090134114027, + 0.012546543031930923, + 0.004001032095402479, + 0.00217971159145236, + 0.00450931815430522, + 0.0017184698954224586, + 0.003513319417834282, + 0.002532501704990864, + 0.003025374375283718, + 0.002961995080113411, + 0.002244309987872839, + 0.0016062981449067593, + 0.003987730480730534, + 0.002287183189764619, + 0.00245989253744483, + 0.002344131702557206, + 0.004000612068921328, + 0.0032956376671791077, + 0.0017784343799576163, + 0.002730422653257847, + 0.003312795888632536, + 0.0023937602527439594, + 0.003358409972861409, + 0.0017083347775042057, + 0.003495731158182025, + 0.0018798907985910773, + 0.003383043920621276, + 0.004268990363925695, + 0.004394656978547573, + 0.003522890154272318, + 0.004018611274659634, + 0.00206867721863091, + 0.0034149622078984976, + 0.0035706493072211742, + 0.010795323178172112, + 0.018022293224930763, + 0.012561616487801075, + 0.0110410051420331, + 0.010336178354918957, + 0.008014397695660591, + 0.007829653099179268, + 0.008570517413318157, + 0.027442898601293564, + 0.03086029551923275, + 0.04411030933260918, + 0.030606567859649658, + 0.04644117131829262, + 0.06939668208360672, + 0.04163385182619095, + 0.04666239768266678, + 0.004131844267249107, + 0.004999920725822449, + 0.006095319055020809, + 0.005612804088741541, + 0.004137021955102682, + 0.006604584865272045, + 0.009246546775102615, + 0.005908020306378603, + 0.011412234976887703, + 0.020085597410798073, + 0.015357238240540028, + 0.029692452400922775, + 0.02275252714753151, + 0.020374871790409088, + 0.03227771446108818, + 0.018903153017163277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.014189893379807472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld, shaky, and volatile as if the photographer is running across uneven ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a weathered wooden cart drifting at a high speed across a vast wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart feature rough wood with cracked grain, rusted iron sheets, coarse hemp rope, and two massive worn wooden wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheels slide sideways to create a swirling cloud of ochre-yellow dust that obscures the sky and nearly engulfs the cart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wasteland with bleached white ground under a glaring midday sun and ochre, dry, polygonal earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303782060742378 + ], + "opposite_logprob_tokens": [ + -3.1680378913879395 + ], + "logprob_reward": -0.04303782060742378, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303782060742378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09455814957618713, + 0.726686954498291, + 0.04662362486124039, + 0.008782313205301762, + 0.016337240114808083, + 0.006687627639621496, + 0.014557307586073875, + 0.02556428872048855, + 0.034064941108226776, + 0.02613753266632557 + ], + "credit_map": [ + 0.007466308306902647, + 0.0070673455484211445, + 0.014874674379825592, + 0.00827670469880104, + 0.01112674456089735, + 0.017366943880915642, + 0.013448946177959442, + 0.014930486679077148, + 0.2214948534965515, + 0.028722506016492844, + 0.29491549730300903, + 0.04121583700180054, + 0.015299534425139427, + 0.021831965073943138, + 0.027918284758925438, + 0.07528845965862274, + 0.007144779898226261, + 0.0038345723878592253, + 0.013143978081643581, + 0.0037633159663528204, + 0.0037076743319630623, + 0.0042142439633607864, + 0.004304155241698027, + 0.006510904524475336, + 0.0010301251895725727, + 0.0012544322526082397, + 0.0010636931983754039, + 0.0010227006860077381, + 0.0013917885953560472, + 0.0008370729628950357, + 0.0012704029213637114, + 0.0009120978065766394, + 0.001260465127415955, + 0.002337890677154064, + 0.002180674346163869, + 0.0023985623847693205, + 0.002722696168348193, + 0.0011540099512785673, + 0.0028447932563722134, + 0.001438148319721222, + 0.0007697493419982493, + 0.0009053078247234225, + 0.0011419712100178003, + 0.00094292510766536, + 0.0008844146504998207, + 0.0004628068127203733, + 0.0007920217467471957, + 0.0007884310907684267, + 0.0010304314782842994, + 0.0032960393000394106, + 0.002322650281712413, + 0.0015551518881693482, + 0.0020471978932619095, + 0.0014311055419966578, + 0.0015235269675031304, + 0.0013512036530300975, + 0.0024455254897475243, + 0.0028262243140488863, + 0.0017782165668904781, + 0.0025276995729655027, + 0.0031889250967651606, + 0.004883575718849897, + 0.00429931515827775, + 0.003614805405959487, + 0.002075520344078541, + 0.004204188473522663, + 0.0032688500359654427, + 0.005692470818758011, + 0.002220802940428257, + 0.004485643468797207, + 0.007744037080556154, + 0.004373427480459213, + 0.0018391921184957027, + 0.0023288249503821135, + 0.0028346581384539604, + 0.003726225346326828, + 0.0033844655845314264, + 0.003734640311449766, + 0.0039023570716381073, + 0.004387167748063803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04303782060742378 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_tensor([3774, 3774])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_tensor([3774, 3774])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d2a69e9420c00c709aec858577cf33297ac6bb46 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_tensor([3774, 3774])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a sleek, deep chestnut horse charging at full speed across an open, dusty plain. Slanting in the golden afternoon sun, its powerful body stretches to its fullest, every muscle clearly undulating and tensing in motion. Sweat dapples its coat, and its dark mane and tail stream wildly behind it in the gusting wind. The camera locks on its powerful shoulders and hind legs, capturing the moment its hooves strike the ground in breathtaking rhythm. Each powerful kick sends dry earth flying backward, creating a billowing dust trail that resembles a golden mist in the light. The horse's nostrils flare with intense breathing, its eyes filled with a primal, unwavering focus. The entire scene, rendered in warm tones, is imbued with a heart-pounding sense of movement, power, and unfettered vitality, creating an atmosphere of epic grandeur.", + "seed": 17673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_3774_seed_17673_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054114597477018835, + "vlm_logprob": -0.054114597477018835, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.054114597477018835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09349136799573898 + ], + "opposite_logprob_tokens": [ + -2.468491315841675 + ], + "logprob_reward": -0.09349136799573898, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09349136799573898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10041993856430054, + 0.047291506081819534, + 0.03540275990962982, + 0.04176696017384529, + 0.06971979886293411, + 0.06629917770624161, + 0.22668619453907013, + 0.0856037512421608, + 0.19223998486995697, + 0.1345699429512024 + ], + "credit_map": [ + 0.016085395589470863, + 0.013481291010975838, + 0.013154079206287861, + 0.011266455985605717, + 0.013253061100840569, + 0.016754543408751488, + 0.00924884993582964, + 0.00717626279219985, + 0.01124513242393732, + 0.004231527913361788, + 0.007138466462492943, + 0.002185444114729762, + 0.0031364161986857653, + 0.009847523644566536, + 0.0028431187383830547, + 0.0066638728603720665, + 0.007954471744596958, + 0.004735411610454321, + 0.006242028903216124, + 0.0022669194731861353, + 0.0030950226355344057, + 0.005165358539670706, + 0.002466300269588828, + 0.0034772451035678387, + 0.007125395815819502, + 0.003857883159071207, + 0.008609890006482601, + 0.0031911053229123354, + 0.003515003016218543, + 0.006777069065719843, + 0.003422101493924856, + 0.005268513225018978, + 0.009722759947180748, + 0.007142725400626659, + 0.006803726311773062, + 0.011932103894650936, + 0.0054801939986646175, + 0.008921696804463863, + 0.009734595194458961, + 0.009981995448470116, + 0.011564298532903194, + 0.0058727930299937725, + 0.004738421179354191, + 0.006091003306210041, + 0.006347960792481899, + 0.00889176968485117, + 0.007744759786874056, + 0.015048172324895859, + 0.03329367935657501, + 0.034703973680734634, + 0.022802338004112244, + 0.02589770220220089, + 0.03833791986107826, + 0.025446977466344833, + 0.02453673630952835, + 0.021666880697011948, + 0.014397962018847466, + 0.00983838364481926, + 0.009365078061819077, + 0.008202208206057549, + 0.0121583491563797, + 0.011270633898675442, + 0.00910390168428421, + 0.01126723363995552, + 0.03822106122970581, + 0.01728067360818386, + 0.03194542974233627, + 0.017757084220647812, + 0.02560485154390335, + 0.02445412054657936, + 0.01811966300010681, + 0.018857087939977646, + 0.018539339303970337, + 0.01926744356751442, + 0.01443495973944664, + 0.021965283900499344, + 0.01665181666612625, + 0.014696029946208, + 0.016952021047472954, + 0.01206305343657732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.09349136799573898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296311765909195 + ], + "opposite_logprob_tokens": [ + -2.2379631996154785 + ], + "logprob_reward": -0.11296311765909195, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296311765909195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11720041185617447, + 0.0983760803937912, + 0.08062668144702911, + 0.15670523047447205, + 0.022727515548467636, + 0.13435959815979004, + 0.04624784365296364, + 0.09332937747240067, + 0.14717547595500946, + 0.10325179249048233 + ], + "credit_map": [ + 0.01846429891884327, + 0.011968516744673252, + 0.024045050144195557, + 0.014237894676625729, + 0.015366834588348866, + 0.013464878313243389, + 0.011732331477105618, + 0.00792060699313879, + 0.02267237938940525, + 0.007453657686710358, + 0.02832980453968048, + 0.004179791081696749, + 0.00589459203183651, + 0.012636378407478333, + 0.005193761084228754, + 0.012015718966722488, + 0.012205054052174091, + 0.0068437159061431885, + 0.021587051451206207, + 0.0037224357947707176, + 0.0058160144835710526, + 0.014152975752949715, + 0.004749692976474762, + 0.011549747548997402, + 0.025007350370287895, + 0.01056052464991808, + 0.03797394409775734, + 0.010329979471862316, + 0.007963982410728931, + 0.029019931331276894, + 0.009669718332588673, + 0.026179803535342216, + 0.004785587545484304, + 0.0031739871483296156, + 0.0021099934820085764, + 0.0025071788113564253, + 0.0019509756239131093, + 0.003172836033627391, + 0.0023417652118951082, + 0.0026851899456232786, + 0.024248167872428894, + 0.010812943801283836, + 0.012502850033342838, + 0.00955598521977663, + 0.017711516469717026, + 0.024801695719361305, + 0.010592957958579063, + 0.024133482947945595, + 0.012621451169252396, + 0.004716268740594387, + 0.005526338703930378, + 0.0038160618860274553, + 0.00427095266059041, + 0.0037806013133376837, + 0.004928154870867729, + 0.00658801244571805, + 0.014357517473399639, + 0.01037545409053564, + 0.009412191808223724, + 0.007092003244906664, + 0.013541563414037228, + 0.015815554186701775, + 0.010818777605891228, + 0.011916319839656353, + 0.041506774723529816, + 0.010369757190346718, + 0.03170131891965866, + 0.010003804229199886, + 0.015533575788140297, + 0.013087386265397072, + 0.013343160971999168, + 0.011629704385995865, + 0.015853503718972206, + 0.014064961113035679, + 0.01149564515799284, + 0.015281319618225098, + 0.015699787065386772, + 0.009854286909103394, + 0.012475444935262203, + 0.00852684210985899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.11296311765909195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750516593456268 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.006750516593456268, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750516593456268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16007648408412933, + 0.16394655406475067, + 0.06750255078077316, + 0.11170679330825806, + 0.04361437261104584, + 0.08826575428247452, + 0.06723987311124802, + 0.06163478642702103, + 0.10947386920452118, + 0.126538947224617 + ], + "credit_map": [ + 0.034954678267240524, + 0.01701568067073822, + 0.03093293309211731, + 0.0156620591878891, + 0.020775172859430313, + 0.021159164607524872, + 0.010327373631298542, + 0.009249432012438774, + 0.03916803002357483, + 0.02416226826608181, + 0.02291707880795002, + 0.009813339449465275, + 0.012949340976774693, + 0.026243286207318306, + 0.010538640432059765, + 0.018154576420783997, + 0.012677819468080997, + 0.008573041297495365, + 0.011095434427261353, + 0.004233655985444784, + 0.005975205451250076, + 0.009166063740849495, + 0.005368776619434357, + 0.01041255984455347, + 0.013926816172897816, + 0.018522588536143303, + 0.011005687527358532, + 0.011448362842202187, + 0.012946241535246372, + 0.01847020536661148, + 0.013101733289659023, + 0.01228516735136509, + 0.005383159499615431, + 0.005789846181869507, + 0.006272939033806324, + 0.005790102295577526, + 0.00405176030471921, + 0.007118649315088987, + 0.0044243368320167065, + 0.004783581010997295, + 0.014661197550594807, + 0.008126795291900635, + 0.009064191952347755, + 0.008016662672162056, + 0.012878013774752617, + 0.019633140414953232, + 0.007430877070873976, + 0.008454871363937855, + 0.014274124056100845, + 0.007566411979496479, + 0.007536687422543764, + 0.005680316127836704, + 0.006671079434454441, + 0.006735272239893675, + 0.006778716575354338, + 0.011997274123132229, + 0.011398405767977238, + 0.00648937001824379, + 0.005929876584559679, + 0.005608394742012024, + 0.005927527789026499, + 0.010324128903448582, + 0.0069546569138765335, + 0.009002428501844406, + 0.02242603525519371, + 0.009611358866095543, + 0.022501029074192047, + 0.011786457151174545, + 0.014457833021879196, + 0.01011690218001604, + 0.012648218311369419, + 0.005926040001213551, + 0.02153896540403366, + 0.01844177581369877, + 0.017199622467160225, + 0.017004726454615593, + 0.019192088395357132, + 0.009535731747746468, + 0.013879667967557907, + 0.009746378287672997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.006750516593456268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048729944974184036 + ], + "opposite_logprob_tokens": [ + -3.04872989654541 + ], + "logprob_reward": -0.048729944974184036, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048729944974184036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1314755529165268, + 0.10994809865951538, + 0.07461093366146088, + 0.13690970838069916, + 0.07429979741573334, + 0.18845608830451965, + 0.04725143685936928, + 0.09421677142381668, + 0.10505937039852142, + 0.037772275507450104 + ], + "credit_map": [ + 0.02108735591173172, + 0.016415951773524284, + 0.021305788308382034, + 0.014083816669881344, + 0.017354097217321396, + 0.020359637215733528, + 0.010470442473888397, + 0.010398470796644688, + 0.025339312851428986, + 0.015437244437634945, + 0.012769514694809914, + 0.006564678158611059, + 0.008111456409096718, + 0.020145557820796967, + 0.008970274589955807, + 0.012610062956809998, + 0.010065349750220776, + 0.012271889485418797, + 0.01071680337190628, + 0.004920950159430504, + 0.007865658961236477, + 0.01086159236729145, + 0.005104696378111839, + 0.01280399039387703, + 0.017048567533493042, + 0.022321466356515884, + 0.01662873663008213, + 0.016500508412718773, + 0.011706151068210602, + 0.02182542346417904, + 0.01738688349723816, + 0.01349196583032608, + 0.00898246094584465, + 0.008846488781273365, + 0.009994519874453545, + 0.010968692600727081, + 0.0063121868297457695, + 0.012276388704776764, + 0.008229713886976242, + 0.00868934579193592, + 0.03332008793950081, + 0.017794903367757797, + 0.014329272322356701, + 0.017925279214978218, + 0.019639726728200912, + 0.02720719203352928, + 0.020472200587391853, + 0.03776742145419121, + 0.008209923282265663, + 0.006026380229741335, + 0.0048927608877420425, + 0.0041101123206317425, + 0.005888018757104874, + 0.005548963323235512, + 0.005240548402070999, + 0.007334727328270674, + 0.02003938890993595, + 0.010250032879412174, + 0.009857992641627789, + 0.009771429933607578, + 0.011119849048554897, + 0.010956354439258575, + 0.00942619051784277, + 0.012795526534318924, + 0.018564574420452118, + 0.0091922702267766, + 0.018122248351573944, + 0.00858828891068697, + 0.00897431094199419, + 0.016603518277406693, + 0.013168614357709885, + 0.011845550499856472, + 0.006405478343367577, + 0.0059320637956261635, + 0.0055360631085932255, + 0.004776728339493275, + 0.004360558465123177, + 0.003961392678320408, + 0.004077504388988018, + 0.002722488483414054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.048729944974184036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638040162622929 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008638040162622929, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638040162622929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12873056530952454, + 0.048792317509651184, + 0.10370343178510666, + 0.4016399085521698, + 0.024511344730854034, + 0.1181780993938446, + 0.029918445274233818, + 0.05801185220479965, + 0.05085926130414009, + 0.03565484657883644 + ], + "credit_map": [ + 0.024082891643047333, + 0.013465057127177715, + 0.036395665258169174, + 0.01002790592610836, + 0.01854506880044937, + 0.013267865404486656, + 0.005163377150893211, + 0.007782745640724897, + 0.007537472061812878, + 0.005289948079735041, + 0.010785928927361965, + 0.0031155431643128395, + 0.003268987638875842, + 0.006729062646627426, + 0.004925613757222891, + 0.007139758672565222, + 0.01813754439353943, + 0.007883692160248756, + 0.028056472539901733, + 0.004278209526091814, + 0.005746631417423487, + 0.01923474110662937, + 0.004691405221819878, + 0.015674728900194168, + 0.06757694482803345, + 0.029079055413603783, + 0.10421084612607956, + 0.025263061746954918, + 0.021034300327301025, + 0.0736946240067482, + 0.024176977574825287, + 0.05660409852862358, + 0.005573983769863844, + 0.0025358896236866713, + 0.002921288600191474, + 0.0031397934071719646, + 0.002244648989289999, + 0.0032464887481182814, + 0.00213492545299232, + 0.002714326838031411, + 0.019315136596560478, + 0.009174142964184284, + 0.008316476829349995, + 0.009163609705865383, + 0.015154104679822922, + 0.013638868927955627, + 0.0142705999314785, + 0.029145173728466034, + 0.0074811698868870735, + 0.002361823571845889, + 0.003152296179905534, + 0.001907584723085165, + 0.002649374073371291, + 0.0022369548678398132, + 0.002163405530154705, + 0.00796583667397499, + 0.012324155308306217, + 0.006104603409767151, + 0.00713844271376729, + 0.004592678509652615, + 0.0074928924441337585, + 0.007379342336207628, + 0.006051200907677412, + 0.006928535643965006, + 0.014694283716380596, + 0.003707289695739746, + 0.011050594039261341, + 0.003569205291569233, + 0.005686318967491388, + 0.004896574653685093, + 0.004462983924895525, + 0.0027920114807784557, + 0.005557606462389231, + 0.004656537435948849, + 0.004931170493364334, + 0.005557563155889511, + 0.004358711652457714, + 0.0035907691344618797, + 0.003702407469972968, + 0.0033000819385051727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.008638040162622929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a sleek, deep chestnut horse charging at full speed across an open, dusty plain. Slanting in the golden afternoon sun, its powerful body stretches to its fullest, every muscle clearly undulating and tensing in motion. Sweat dapples its coat, and its dark mane and tail stream wildly behind it in the gusting wind. The camera locks on its powerful shoulders and hind legs, capturing the moment its hooves strike the ground in breathtaking rhythm. Each powerful kick sends dry earth flying backward, creating a billowing dust trail that resembles a golden mist in the light. The horse's nostrils flare with intense breathing, its eyes filled with a primal, unwavering focus. The entire scene, rendered in warm tones, is imbued with a heart-pounding sense of movement, power, and unfettered vitality, creating an atmosphere of epic grandeur.", + "seed": 17674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_3774_seed_17674_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024975600838661193, + "vlm_logprob": -0.024975600838661193, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.024975600838661193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024214359000325203 + ], + "opposite_logprob_tokens": [ + -3.899214267730713 + ], + "logprob_reward": -0.024214359000325203, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024214359000325203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07882346957921982, + 0.14841152727603912, + 0.07759658247232437, + 0.08879473805427551, + 0.07753614336252213, + 0.08598101884126663, + 0.2176133543252945, + 0.11248113214969635, + 0.09169725328683853, + 0.021064823493361473 + ], + "credit_map": [ + 0.008572030812501907, + 0.01234440878033638, + 0.013216584920883179, + 0.007937366142868996, + 0.008403167128562927, + 0.00902552343904972, + 0.006595544051378965, + 0.012728847563266754, + 0.011964119039475918, + 0.015212801285088062, + 0.008384552784264088, + 0.009132335893809795, + 0.04779065027832985, + 0.014201108366250992, + 0.011630474589765072, + 0.03009548969566822, + 0.008436480537056923, + 0.008860576897859573, + 0.0055201249197125435, + 0.006137831602245569, + 0.012615474872291088, + 0.023626701906323433, + 0.004408736247569323, + 0.007990652695298195, + 0.010447807610034943, + 0.014936218969523907, + 0.006647795904427767, + 0.009519249200820923, + 0.010215041227638721, + 0.012561289593577385, + 0.0128165977075696, + 0.01165074110031128, + 0.011890974827110767, + 0.007986847311258316, + 0.007156132720410824, + 0.004134781192988157, + 0.006550310179591179, + 0.013506677933037281, + 0.004904933739453554, + 0.021405480802059174, + 0.012554129585623741, + 0.010350849479436874, + 0.013165300711989403, + 0.010418307036161423, + 0.010059737600386143, + 0.0065752556547522545, + 0.00894780270755291, + 0.013909631408751011, + 0.02745424397289753, + 0.022462578490376472, + 0.029365532100200653, + 0.02529265731573105, + 0.020714163780212402, + 0.021483061835169792, + 0.04397553205490112, + 0.026865586638450623, + 0.00863291509449482, + 0.01295236125588417, + 0.014728203415870667, + 0.010945324786007404, + 0.02042456716299057, + 0.012540611438453197, + 0.015734180808067322, + 0.016522962599992752, + 0.010319781489670277, + 0.010698368772864342, + 0.01162366196513176, + 0.007905211299657822, + 0.008194671012461185, + 0.014758577570319176, + 0.008872664533555508, + 0.019324321299791336, + 0.004683004692196846, + 0.0025642032269388437, + 0.0024900471325963736, + 0.001982681220397353, + 0.001980660017579794, + 0.0029619091656059027, + 0.0020276112481951714, + 0.002374706557020545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.024214359000325203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370694860816002 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370694860816002, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370694860816002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09726472944021225, + 0.1751493215560913, + 0.05427160859107971, + 0.13498729467391968, + 0.06960781663656235, + 0.18651893734931946, + 0.040243376046419144, + 0.12181063741445541, + 0.08151936531066895, + 0.03862687572836876 + ], + "credit_map": [ + 0.007667040452361107, + 0.014140050858259201, + 0.017603596672415733, + 0.013015889562666416, + 0.013083924539387226, + 0.008814586326479912, + 0.009277612902224064, + 0.013662024401128292, + 0.014510287903249264, + 0.013665398582816124, + 0.008734817616641521, + 0.012027385644614697, + 0.05974552780389786, + 0.017790403217077255, + 0.010226830840110779, + 0.038448676466941833, + 0.0072060804814100266, + 0.006316308863461018, + 0.006546661723405123, + 0.0046823471784591675, + 0.008089059963822365, + 0.006377490237355232, + 0.006312783807516098, + 0.008740880526602268, + 0.02796712890267372, + 0.019619429484009743, + 0.014974565245211124, + 0.018248511478304863, + 0.011201162822544575, + 0.011654645204544067, + 0.013651971705257893, + 0.017669884487986565, + 0.008154409006237984, + 0.010356971062719822, + 0.010168550536036491, + 0.009918492287397385, + 0.006587328854948282, + 0.008218399249017239, + 0.00945830624550581, + 0.006745357997715473, + 0.027772920206189156, + 0.02395601198077202, + 0.02623826079070568, + 0.018068453297019005, + 0.01872766762971878, + 0.018007488921284676, + 0.018434008583426476, + 0.03531414270401001, + 0.005370189435780048, + 0.005469516851007938, + 0.00520945293828845, + 0.004273537080734968, + 0.0037668594159185886, + 0.004398941528052092, + 0.005844762548804283, + 0.005910118576139212, + 0.01035206951200962, + 0.01581854373216629, + 0.014899345114827156, + 0.012153858318924904, + 0.014999193139374256, + 0.017017468810081482, + 0.015930211171507835, + 0.020639952272176743, + 0.013813378289341927, + 0.005352415144443512, + 0.013892053626477718, + 0.005558873992413282, + 0.0062025864608585835, + 0.011463717557489872, + 0.00615716353058815, + 0.019079172983765602, + 0.010155917145311832, + 0.004419384989887476, + 0.00500316359102726, + 0.003271648893132806, + 0.0038167890161275864, + 0.004405447281897068, + 0.0038257669657468796, + 0.003728757379576564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.026370694860816002 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636094018816948 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636094018816948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636094018816948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09063240885734558, + 0.12132630497217178, + 0.13307765126228333, + 0.1036592647433281, + 0.10737262666225433, + 0.14075864851474762, + 0.053135946393013, + 0.13118363916873932, + 0.05878432095050812, + 0.06006919592618942 + ], + "credit_map": [ + 0.011389271356165409, + 0.011025897227227688, + 0.009721375070512295, + 0.010431187227368355, + 0.009368968196213245, + 0.012666555121541023, + 0.008543566800653934, + 0.017485585063695908, + 0.009657172486186028, + 0.011541821993887424, + 0.007173730526119471, + 0.008786180056631565, + 0.032531388103961945, + 0.010378113016486168, + 0.009394505061209202, + 0.03186339512467384, + 0.012368258088827133, + 0.019454114139080048, + 0.012351687997579575, + 0.010768800973892212, + 0.02115766331553459, + 0.031077325344085693, + 0.011020779609680176, + 0.014879031106829643, + 0.01767687126994133, + 0.019517695531249046, + 0.009833977557718754, + 0.011179488152265549, + 0.007970897480845451, + 0.012274524196982384, + 0.012308679521083832, + 0.012897123582661152, + 0.008957755751907825, + 0.012125215493142605, + 0.007921135053038597, + 0.006411916576325893, + 0.007970466278493404, + 0.025495614856481552, + 0.00799765344709158, + 0.030492866411805153, + 0.01779530756175518, + 0.019047072157263756, + 0.019668713212013245, + 0.02332334965467453, + 0.011529995128512383, + 0.014099298976361752, + 0.0177784226834774, + 0.017516491934657097, + 0.006324292626231909, + 0.0057660420425236225, + 0.008590707555413246, + 0.006475585978478193, + 0.005702485796064138, + 0.0048667131923139095, + 0.009776854887604713, + 0.005633263383060694, + 0.013930119574069977, + 0.015840552747249603, + 0.018792208284139633, + 0.013973559252917767, + 0.017629103735089302, + 0.01853828877210617, + 0.01682669296860695, + 0.01565311849117279, + 0.010670146904885769, + 0.005742010194808245, + 0.010036105290055275, + 0.004142275080084801, + 0.004642930347472429, + 0.008321590721607208, + 0.005478817969560623, + 0.009750445373356342, + 0.009898174554109573, + 0.009804255329072475, + 0.006652447395026684, + 0.005590400658547878, + 0.007641888223588467, + 0.006271441001445055, + 0.007253335788846016, + 0.006957255769520998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.02636094018816948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815260902047157 + ], + "opposite_logprob_tokens": [ + -3.2881526947021484 + ], + "logprob_reward": -0.03815260902047157, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815260902047157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0673207938671112, + 0.09268322587013245, + 0.06684357672929764, + 0.05184473842382431, + 0.14044831693172455, + 0.16984833776950836, + 0.06930606067180634, + 0.18318040668964386, + 0.10590961575508118, + 0.05261491984128952 + ], + "credit_map": [ + 0.006886539980769157, + 0.007761465851217508, + 0.013557609170675278, + 0.007825098931789398, + 0.0064149643294513226, + 0.006427169777452946, + 0.007500585168600082, + 0.010947360657155514, + 0.008868904784321785, + 0.007913164794445038, + 0.005386339966207743, + 0.004828747361898422, + 0.0300427358597517, + 0.008459567092359066, + 0.006110538262873888, + 0.021073229610919952, + 0.005503993481397629, + 0.0088422242552042, + 0.0050940243527293205, + 0.004823980387300253, + 0.012315583415329456, + 0.019481969997286797, + 0.0051355017349123955, + 0.005646292585879564, + 0.005585408769547939, + 0.006796244997531176, + 0.006561356130987406, + 0.005478875245898962, + 0.0053041186183691025, + 0.0077898139134049416, + 0.008860734291374683, + 0.0054681864567101, + 0.012493053451180458, + 0.015166306868195534, + 0.008150913752615452, + 0.00799999013543129, + 0.007886611856520176, + 0.03781910240650177, + 0.009823606349527836, + 0.04110873118042946, + 0.01691112108528614, + 0.023144524544477463, + 0.017369577661156654, + 0.03225141763687134, + 0.015713641420006752, + 0.01259633805602789, + 0.02974417246878147, + 0.02211754024028778, + 0.008306015282869339, + 0.006392658222466707, + 0.010820546187460423, + 0.00828987080603838, + 0.006129732355475426, + 0.007282872684299946, + 0.01380703691393137, + 0.008277329616248608, + 0.021245406940579414, + 0.01762392744421959, + 0.04346046224236488, + 0.014361795969307423, + 0.02140744961798191, + 0.021615615114569664, + 0.016919083893299103, + 0.026546655222773552, + 0.02128635160624981, + 0.006779382936656475, + 0.01956840790808201, + 0.008014905266463757, + 0.008435296826064587, + 0.016865026205778122, + 0.007576298899948597, + 0.01738394983112812, + 0.010712473653256893, + 0.005741171073168516, + 0.007196699734777212, + 0.0035212161019444466, + 0.006942352745682001, + 0.009663200937211514, + 0.004666637163609266, + 0.004171169362962246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.03815260902047157 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779401123523712 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779401123523712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779401123523712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07482483237981796, + 0.15233618021011353, + 0.07479676604270935, + 0.10683950781822205, + 0.07912518829107285, + 0.07748039811849594, + 0.04833844304084778, + 0.25103047490119934, + 0.05969332903623581, + 0.0755348652601242 + ], + "credit_map": [ + 0.007287169340997934, + 0.008849999867379665, + 0.012687165290117264, + 0.010884531773626804, + 0.0062579624354839325, + 0.0059233373031020164, + 0.012529874220490456, + 0.010404793545603752, + 0.011388205923140049, + 0.015582554042339325, + 0.007616047747433186, + 0.010079586878418922, + 0.043784286826848984, + 0.01196371391415596, + 0.010880758985877037, + 0.04104103520512581, + 0.005558900535106659, + 0.008970390073955059, + 0.005874278489500284, + 0.009303468279540539, + 0.010286637581884861, + 0.01334492489695549, + 0.005857603158801794, + 0.015600559301674366, + 0.017056461423635483, + 0.021587496623396873, + 0.011513689532876015, + 0.01460023783147335, + 0.008150259964168072, + 0.008976349607110023, + 0.01104242168366909, + 0.013912580907344818, + 0.007058062124997377, + 0.009127292782068253, + 0.008405318483710289, + 0.0057000694796442986, + 0.006137515883892775, + 0.01519173663109541, + 0.006750611122697592, + 0.020754583179950714, + 0.01089072972536087, + 0.011144152842462063, + 0.00843083020299673, + 0.013193994760513306, + 0.0076560163870453835, + 0.005753773730248213, + 0.010592762380838394, + 0.009818131104111671, + 0.0060590291395783424, + 0.005827620625495911, + 0.00974573940038681, + 0.0030276800971478224, + 0.00521433399990201, + 0.006668213754892349, + 0.005768678151071072, + 0.00602714903652668, + 0.02639399655163288, + 0.02751963771879673, + 0.0485289990901947, + 0.027083106338977814, + 0.02888302505016327, + 0.03376132249832153, + 0.02725077234208584, + 0.03160960599780083, + 0.011634977534413338, + 0.005173286888748407, + 0.009508082643151283, + 0.004787638317793608, + 0.005693358834832907, + 0.007774217054247856, + 0.00557305384427309, + 0.009548713453114033, + 0.02619978040456772, + 0.0070904940366744995, + 0.013638005591928959, + 0.005904894322156906, + 0.006620957516133785, + 0.006391533184796572, + 0.00492732273414731, + 0.004761871881783009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.009779401123523712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a sleek, deep chestnut horse charging at full speed across an open, dusty plain. Slanting in the golden afternoon sun, its powerful body stretches to its fullest, every muscle clearly undulating and tensing in motion. Sweat dapples its coat, and its dark mane and tail stream wildly behind it in the gusting wind. The camera locks on its powerful shoulders and hind legs, capturing the moment its hooves strike the ground in breathtaking rhythm. Each powerful kick sends dry earth flying backward, creating a billowing dust trail that resembles a golden mist in the light. The horse's nostrils flare with intense breathing, its eyes filled with a primal, unwavering focus. The entire scene, rendered in warm tones, is imbued with a heart-pounding sense of movement, power, and unfettered vitality, creating an atmosphere of epic grandeur.", + "seed": 17675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_3774_seed_17675_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02752819932065904, + "vlm_logprob": -0.02752819932065904, + "vlm_yes_no_margin": 3.924999952316284, + "vlm_token_credit": -0.02752819932065904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05268803611397743 + ], + "opposite_logprob_tokens": [ + -3.0526881217956543 + ], + "logprob_reward": -0.05268803611397743, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05268803611397743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08531954139471054, + 0.07917177677154541, + 0.020512904971837997, + 0.0653320923447609, + 0.041869163513183594, + 0.047905273735523224, + 0.21120096743106842, + 0.2380472719669342, + 0.04119143262505531, + 0.16944962739944458 + ], + "credit_map": [ + 0.015711648389697075, + 0.011526022106409073, + 0.00921002496033907, + 0.007479907479137182, + 0.011654634959995747, + 0.012221374548971653, + 0.00629414664581418, + 0.011221780441701412, + 0.013828235678374767, + 0.0037987055256962776, + 0.012929631397128105, + 0.007495786529034376, + 0.003642308758571744, + 0.013327653519809246, + 0.006104376632720232, + 0.01804507151246071, + 0.0038197182584553957, + 0.003149547381326556, + 0.002608418930321932, + 0.001625335542485118, + 0.0025272739585489035, + 0.0022974477615207434, + 0.002163045108318329, + 0.0023221191950142384, + 0.010172159411013126, + 0.004209177102893591, + 0.01596912555396557, + 0.004998679272830486, + 0.004399372264742851, + 0.011838862672448158, + 0.007544224616140127, + 0.0062004877254366875, + 0.007114970590919256, + 0.0040792739018797874, + 0.007063087075948715, + 0.003332355758175254, + 0.0034425558988004923, + 0.0071827732026577, + 0.003469274379312992, + 0.006184875965118408, + 0.009645906277000904, + 0.003714646678417921, + 0.01055760495364666, + 0.004989694803953171, + 0.0036148084327578545, + 0.004183133598417044, + 0.005981075577437878, + 0.005218404810875654, + 0.04958570748567581, + 0.024863004684448242, + 0.050278086215257645, + 0.018235785886645317, + 0.012453352101147175, + 0.020307864993810654, + 0.01669166050851345, + 0.018785512074828148, + 0.04660751670598984, + 0.030548231676220894, + 0.03011718951165676, + 0.028076715767383575, + 0.02596467360854149, + 0.018617287278175354, + 0.025646908208727837, + 0.03246873989701271, + 0.008412093855440617, + 0.004502685274928808, + 0.006548791192471981, + 0.0051194182597100735, + 0.003501992905512452, + 0.003308236366137862, + 0.0048294817097485065, + 0.004968732129782438, + 0.05614830181002617, + 0.02073090337216854, + 0.026467112824320793, + 0.01732460781931877, + 0.009665868245065212, + 0.0058173248544335365, + 0.007925194688141346, + 0.025370309129357338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.05268803611397743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018222538754343987 + ], + "opposite_logprob_tokens": [ + -4.018222332000732 + ], + "logprob_reward": -0.018222538754343987, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018222538754343987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09503691643476486, + 0.1344207525253296, + 0.040341947227716446, + 0.06665560603141785, + 0.029362553730607033, + 0.0432668961584568, + 0.181891068816185, + 0.038924332708120346, + 0.16593898832798004, + 0.2041609138250351 + ], + "credit_map": [ + 0.017855774611234665, + 0.009571734815835953, + 0.016726957634091377, + 0.007587925996631384, + 0.00802565272897482, + 0.015992002561688423, + 0.006253042258322239, + 0.013023819774389267, + 0.016323326155543327, + 0.008678428828716278, + 0.02659844607114792, + 0.016486819833517075, + 0.011067810468375683, + 0.02368987165391445, + 0.008265750482678413, + 0.023310299962759018, + 0.009012439288198948, + 0.004978166427463293, + 0.00910994503647089, + 0.002801409689709544, + 0.0028511618729680777, + 0.004876659717410803, + 0.0030139859300106764, + 0.00369817903265357, + 0.010280258022248745, + 0.006753915920853615, + 0.0076890792697668076, + 0.009425233118236065, + 0.005961019545793533, + 0.007845793850719929, + 0.010608178563416004, + 0.008092127740383148, + 0.0059566255658864975, + 0.003289049956947565, + 0.0034525468945503235, + 0.0019177499925717711, + 0.003452811622992158, + 0.0037891550455242395, + 0.0023787678219377995, + 0.0051258462481200695, + 0.006989818066358566, + 0.004454548936337233, + 0.005647486541420221, + 0.003996852319687605, + 0.0055353231728076935, + 0.004403702914714813, + 0.005686605349183083, + 0.006552556063979864, + 0.05604272335767746, + 0.010801355354487896, + 0.06293006986379623, + 0.011103696189820766, + 0.008420618250966072, + 0.008480161428451538, + 0.0056978813372552395, + 0.018414562568068504, + 0.006630932446569204, + 0.004543178714811802, + 0.004321952350437641, + 0.0037650589365512133, + 0.0044128261506557465, + 0.004100528080016375, + 0.004307256080210209, + 0.006842600181698799, + 0.017295636236667633, + 0.015480192378163338, + 0.022827081382274628, + 0.020313292741775513, + 0.021418185904622078, + 0.023911964148283005, + 0.02584429457783699, + 0.018848350271582603, + 0.06717497110366821, + 0.026638757437467575, + 0.0316249243915081, + 0.019100235775113106, + 0.011761859990656376, + 0.008198088966310024, + 0.0098663829267025, + 0.0297956932336092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.018222538754343987 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370870277285576 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370870277285576, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370870277285576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09490501880645752, + 0.07972831279039383, + 0.0658746287226677, + 0.2977917790412903, + 0.05066084489226341, + 0.03818969801068306, + 0.05199167877435684, + 0.060987066477537155, + 0.10542663931846619, + 0.15444432199001312 + ], + "credit_map": [ + 0.016925357282161713, + 0.01062002219259739, + 0.011320168152451515, + 0.007602449972182512, + 0.009161252528429031, + 0.012041552923619747, + 0.007073572371155024, + 0.020160647109150887, + 0.011585668660700321, + 0.006836118176579475, + 0.013301023282110691, + 0.010354814119637012, + 0.006525822915136814, + 0.013870097696781158, + 0.004788334481418133, + 0.012466438114643097, + 0.010936296544969082, + 0.008261983282864094, + 0.009085687808692455, + 0.007362987380474806, + 0.0066305482760071754, + 0.008375545032322407, + 0.006189065985381603, + 0.00903251487761736, + 0.06510870903730392, + 0.017582004889845848, + 0.041164420545101166, + 0.026407945901155472, + 0.015359875746071339, + 0.04600675031542778, + 0.026994861662387848, + 0.05916721001267433, + 0.009060417301952839, + 0.005366085097193718, + 0.006948043592274189, + 0.004669856745749712, + 0.0034524821676313877, + 0.007324533071368933, + 0.004645139444619417, + 0.009194288402795792, + 0.00784969236701727, + 0.0030102089513093233, + 0.0064049093052744865, + 0.004318132996559143, + 0.003020709380507469, + 0.004196865949779749, + 0.003875243477523327, + 0.005513934418559074, + 0.005990322679281235, + 0.008040610700845718, + 0.006410632282495499, + 0.005641204304993153, + 0.004756647162139416, + 0.00775372888892889, + 0.006161841098219156, + 0.007236692123115063, + 0.011102037504315376, + 0.0086029302328825, + 0.00728544732555747, + 0.007072813343256712, + 0.00547610130161047, + 0.007049511652439833, + 0.005996053107082844, + 0.008402170613408089, + 0.018527550622820854, + 0.012238251976668835, + 0.01876274310052395, + 0.008639485575258732, + 0.012097225524485111, + 0.009123455733060837, + 0.013376465067267418, + 0.012661469168961048, + 0.0443313829600811, + 0.0209671463817358, + 0.0234702005982399, + 0.016481271013617516, + 0.010433673858642578, + 0.008623473346233368, + 0.009828712791204453, + 0.0203084759414196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.03370870277285576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029831646010279655 + ], + "opposite_logprob_tokens": [ + -3.529831647872925 + ], + "logprob_reward": -0.029831646010279655, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029831646010279655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06835735589265823, + 0.05823758989572525, + 0.057644911110401154, + 0.19731546938419342, + 0.042898181825876236, + 0.02853315696120262, + 0.11764135211706161, + 0.05004260316491127, + 0.05805883929133415, + 0.32127055525779724 + ], + "credit_map": [ + 0.01115881372243166, + 0.009535017423331738, + 0.007998406887054443, + 0.007927151396870613, + 0.007221360225230455, + 0.007235987111926079, + 0.007148009724915028, + 0.010132603347301483, + 0.006702136714011431, + 0.007839825935661793, + 0.006678534671664238, + 0.009034770540893078, + 0.006735882721841335, + 0.00820345338433981, + 0.0048957508988678455, + 0.008147239685058594, + 0.008419735357165337, + 0.00566121144220233, + 0.009680420160293579, + 0.005190075840801001, + 0.005595894530415535, + 0.008475041948258877, + 0.006347437389194965, + 0.008275090716779232, + 0.03235754743218422, + 0.01609421893954277, + 0.032496046274900436, + 0.019332848489284515, + 0.015078665688633919, + 0.033526480197906494, + 0.023966360837221146, + 0.02446330152451992, + 0.007459809072315693, + 0.004569163080304861, + 0.00747529836371541, + 0.003165086265653372, + 0.0033838681410998106, + 0.00827110093086958, + 0.0038018582854419947, + 0.004772001411765814, + 0.005366350989788771, + 0.0026436932384967804, + 0.004587371367961168, + 0.002964512212201953, + 0.002644624561071396, + 0.0027860223781317472, + 0.003933214582502842, + 0.003607369028031826, + 0.03753800690174103, + 0.009778743609786034, + 0.036792486906051636, + 0.008045675233006477, + 0.005254407413303852, + 0.006840316113084555, + 0.005635893438011408, + 0.007755816914141178, + 0.008479996584355831, + 0.007862536236643791, + 0.005715877749025822, + 0.005839590448886156, + 0.005007162224501371, + 0.005980750545859337, + 0.004602739121764898, + 0.00655394559726119, + 0.011646927334368229, + 0.0052377888932824135, + 0.0073094419203698635, + 0.005012890789657831, + 0.008502746932208538, + 0.007062279153615236, + 0.007450669538229704, + 0.005836097523570061, + 0.0945887491106987, + 0.0439092181622982, + 0.047385841608047485, + 0.03413775563240051, + 0.018993351608514786, + 0.017304202541708946, + 0.021624818444252014, + 0.043326638638973236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.029831646010279655 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190072951838374 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003190072951838374, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190072951838374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17337965965270996, + 0.38407987356185913, + 0.029280254617333412, + 0.07023463398218155, + 0.04049238562583923, + 0.05078023672103882, + 0.04113481938838959, + 0.038531579077243805, + 0.03105263225734234, + 0.1410338580608368 + ], + "credit_map": [ + 0.04484640806913376, + 0.016790255904197693, + 0.026895342394709587, + 0.0141819529235363, + 0.011484239250421524, + 0.029198767617344856, + 0.010993864387273788, + 0.018988847732543945, + 0.06321955472230911, + 0.020959191024303436, + 0.08000024408102036, + 0.047099366784095764, + 0.02267700619995594, + 0.07649403810501099, + 0.01618223823606968, + 0.057448286563158035, + 0.007984179072082043, + 0.002419763710349798, + 0.00309010804630816, + 0.0019784944597631693, + 0.001961406087502837, + 0.006602675188332796, + 0.0018553859554231167, + 0.0033882437273859978, + 0.012431853450834751, + 0.006743710022419691, + 0.010404391214251518, + 0.007189401891082525, + 0.006446734070777893, + 0.006894601974636316, + 0.009411695413291454, + 0.010712248273193836, + 0.009146341122686863, + 0.004217641428112984, + 0.007631740067154169, + 0.0028628245927393436, + 0.0020888231229037046, + 0.004102376289665699, + 0.003276875475421548, + 0.007165765855461359, + 0.011039896868169308, + 0.004319620784372091, + 0.01001870445907116, + 0.0050871651619672775, + 0.0043028718791902065, + 0.00441333744674921, + 0.004334115423262119, + 0.007264529354870319, + 0.005064557306468487, + 0.005956132430583239, + 0.0049280873499810696, + 0.006044910289347172, + 0.0040321359410882, + 0.00464426027610898, + 0.0054620904847979546, + 0.005002651363611221, + 0.005790014285594225, + 0.007447860669344664, + 0.003767040092498064, + 0.005274631083011627, + 0.0048913205973804, + 0.003791258204728365, + 0.0033220546320080757, + 0.004247402772307396, + 0.005892537068575621, + 0.003291790373623371, + 0.0037868348881602287, + 0.004350819159299135, + 0.0036017608363181353, + 0.002693576505407691, + 0.003920047543942928, + 0.003515268675982952, + 0.04683300480246544, + 0.01881563477218151, + 0.019205817952752113, + 0.013295674696564674, + 0.008648212067782879, + 0.005575117655098438, + 0.008074432611465454, + 0.020585959777235985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.003190072951838374 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a sleek, deep chestnut horse charging at full speed across an open, dusty plain. Slanting in the golden afternoon sun, its powerful body stretches to its fullest, every muscle clearly undulating and tensing in motion. Sweat dapples its coat, and its dark mane and tail stream wildly behind it in the gusting wind. The camera locks on its powerful shoulders and hind legs, capturing the moment its hooves strike the ground in breathtaking rhythm. Each powerful kick sends dry earth flying backward, creating a billowing dust trail that resembles a golden mist in the light. The horse's nostrils flare with intense breathing, its eyes filled with a primal, unwavering focus. The entire scene, rendered in warm tones, is imbued with a heart-pounding sense of movement, power, and unfettered vitality, creating an atmosphere of epic grandeur.", + "seed": 17676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_0/index_3774_seed_17676_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13408236466348172, + "vlm_logprob": -0.13408236466348172, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.13408236466348172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058518264442682266 + ], + "opposite_logprob_tokens": [ + -2.933518171310425 + ], + "logprob_reward": -0.058518264442682266, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.058518264442682266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17705272138118744, + 0.11254275590181351, + 0.05342646688222885, + 0.04685036838054657, + 0.18528662621974945, + 0.03324722498655319, + 0.06195034086704254, + 0.05869442597031593, + 0.1250169426202774, + 0.1459321528673172 + ], + "credit_map": [ + 0.0207506213337183, + 0.02324483171105385, + 0.01892506517469883, + 0.018565509468317032, + 0.017914071679115295, + 0.02405049093067646, + 0.02726241946220398, + 0.026339687407016754, + 0.012482699938118458, + 0.0023853941820561886, + 0.03004417195916176, + 0.004998036194592714, + 0.0034491880796849728, + 0.008160167373716831, + 0.003412120044231415, + 0.047610972076654434, + 0.008148559369146824, + 0.004628384485840797, + 0.009325051680207253, + 0.004041559528559446, + 0.002951350063085556, + 0.00880322977900505, + 0.004862763453274965, + 0.010665565729141235, + 0.005466258153319359, + 0.006418598350137472, + 0.006299490109086037, + 0.004504665732383728, + 0.004379565827548504, + 0.0068117789924144745, + 0.0036999646108597517, + 0.009270042181015015, + 0.03979583829641342, + 0.014427347108721733, + 0.018068009987473488, + 0.009166309610009193, + 0.01048571802675724, + 0.02851239964365959, + 0.014281123876571655, + 0.05054986849427223, + 0.004188706632703543, + 0.004667752422392368, + 0.003575026523321867, + 0.00390970055013895, + 0.0025215495843440294, + 0.0036031221970915794, + 0.0030537606216967106, + 0.00772760109975934, + 0.0076468284241855145, + 0.00915457308292389, + 0.007016557734459639, + 0.00898662768304348, + 0.005877667106688023, + 0.0075888014398515224, + 0.006903359666466713, + 0.008775920607149601, + 0.007063687779009342, + 0.009788207709789276, + 0.009676468558609486, + 0.007194051519036293, + 0.004702083766460419, + 0.008059795014560223, + 0.0034322512801736593, + 0.008777876384556293, + 0.01933857426047325, + 0.01423492468893528, + 0.01882745511829853, + 0.011406525038182735, + 0.017302438616752625, + 0.016907714307308197, + 0.007992041297256947, + 0.019007254391908646, + 0.024538245052099228, + 0.019360529258847237, + 0.02075548842549324, + 0.013526436872780323, + 0.020730769261717796, + 0.018397366628050804, + 0.01035325787961483, + 0.018270058557391167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.058518264442682266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233327150344849 + ], + "opposite_logprob_tokens": [ + -0.8983327150344849 + ], + "logprob_reward": -0.5233327150344849, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233327150344849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12681874632835388, + 0.40707316994667053, + 0.1390840709209442, + 0.023079169914126396, + 0.12779287993907928, + 0.031658392399549484, + 0.029114607721567154, + 0.03514707088470459, + 0.046099912375211716, + 0.03413199260830879 + ], + "credit_map": [ + 0.026660548523068428, + 0.0126876812428236, + 0.011799812316894531, + 0.008865971118211746, + 0.00889742560684681, + 0.0161064974963665, + 0.013821214437484741, + 0.027979597449302673, + 0.04315308853983879, + 0.007092976476997137, + 0.10864923149347305, + 0.018866444006562233, + 0.010843835771083832, + 0.03156165033578873, + 0.012383229099214077, + 0.17452268302440643, + 0.012241937220096588, + 0.008575160056352615, + 0.021690623834729195, + 0.009337677620351315, + 0.006406120024621487, + 0.039301127195358276, + 0.011800558306276798, + 0.02973087690770626, + 0.003203641390427947, + 0.0019924729131162167, + 0.0035165157169103622, + 0.0017460474045947194, + 0.0018629106925800443, + 0.0032953014597296715, + 0.002818968379870057, + 0.004643313121050596, + 0.022418314591050148, + 0.010500878095626831, + 0.021129382774233818, + 0.0057164267636835575, + 0.007140162400901318, + 0.021335771307349205, + 0.008465436287224293, + 0.0310865119099617, + 0.004032372962683439, + 0.0028875996358692646, + 0.008963806554675102, + 0.00250902958214283, + 0.0012944134650751948, + 0.002624285174533725, + 0.0013570478186011314, + 0.00798984058201313, + 0.003692576428875327, + 0.0033901275601238012, + 0.003076272550970316, + 0.003681732341647148, + 0.002212011255323887, + 0.0031194717157632113, + 0.003420507535338402, + 0.00652190949767828, + 0.009039781987667084, + 0.004383823834359646, + 0.004823016002774239, + 0.0026160317938774824, + 0.0018626770470291376, + 0.0037983297370374203, + 0.0017114037182182074, + 0.006912007927894592, + 0.01002683024853468, + 0.005697780754417181, + 0.006522585637867451, + 0.00483167776837945, + 0.0050970204174518585, + 0.006107155699282885, + 0.002431161468848586, + 0.0053856996819376945, + 0.007924498990178108, + 0.005226725712418556, + 0.0044342526234686375, + 0.003481559455394745, + 0.0037450368981808424, + 0.0039940145798027515, + 0.0017458568327128887, + 0.0035800449550151825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.5233327150344849 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598655566573143 + ], + "opposite_logprob_tokens": [ + -3.8955986499786377 + ], + "logprob_reward": -0.020598655566573143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598655566573143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18984825909137726, + 0.18418097496032715, + 0.05097857490181923, + 0.05253376066684723, + 0.3302459120750427, + 0.02787952870130539, + 0.040670059621334076, + 0.06291447579860687, + 0.03490075096487999, + 0.025847705081105232 + ], + "credit_map": [ + 0.031434234231710434, + 0.017435532063245773, + 0.022462377324700356, + 0.00924031250178814, + 0.011484389193356037, + 0.037221312522888184, + 0.01634654588997364, + 0.044223565608263016, + 0.02054557576775551, + 0.004440436605364084, + 0.047140274196863174, + 0.008938273414969444, + 0.005557951517403126, + 0.013716300949454308, + 0.007366518955677748, + 0.07647562772035599, + 0.006700063124299049, + 0.00526275672018528, + 0.008528674952685833, + 0.004404202103614807, + 0.0037092624697834253, + 0.00885356031358242, + 0.005481087602674961, + 0.008038966916501522, + 0.005839042365550995, + 0.004362278152257204, + 0.007600911892950535, + 0.0036709958221763372, + 0.003960876725614071, + 0.00836225040256977, + 0.004942599218338728, + 0.013794802129268646, + 0.06655507534742355, + 0.02486821822822094, + 0.04600225389003754, + 0.013042147271335125, + 0.01775369979441166, + 0.04806520789861679, + 0.02369767613708973, + 0.09026163071393967, + 0.0039052851498126984, + 0.003333822125568986, + 0.005582144483923912, + 0.002248127246275544, + 0.0016810502856969833, + 0.0025279675610363483, + 0.0020776065066456795, + 0.0065235267393291, + 0.007049221079796553, + 0.005537399090826511, + 0.004930119030177593, + 0.003774050623178482, + 0.003646883415058255, + 0.005017013754695654, + 0.0038108087610453367, + 0.006904561538249254, + 0.013198823668062687, + 0.008119597099721432, + 0.006434321869164705, + 0.00739878648892045, + 0.004271556157618761, + 0.008294079452753067, + 0.003763441229239106, + 0.01143387146294117, + 0.006654474418610334, + 0.0031016827560961246, + 0.00506948446854949, + 0.003147817449644208, + 0.003523214254528284, + 0.005271506030112505, + 0.0019744588062167168, + 0.006158108357340097, + 0.006415282376110554, + 0.003382282331585884, + 0.0028231891337782145, + 0.0018726211274042726, + 0.0031077000312507153, + 0.0034694806672632694, + 0.001456824247725308, + 0.0033203251659870148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.020598655566573143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038162361830472946 + ], + "opposite_logprob_tokens": [ + -3.2881624698638916 + ], + "logprob_reward": -0.038162361830472946, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038162361830472946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12553377449512482, + 0.17225123941898346, + 0.06036887690424919, + 0.08298394083976746, + 0.23625756800174713, + 0.02873288467526436, + 0.03358491510152817, + 0.07602660357952118, + 0.11916247010231018, + 0.06509778648614883 + ], + "credit_map": [ + 0.029892658814787865, + 0.009684366174042225, + 0.016627265140414238, + 0.007395741064101458, + 0.009908242151141167, + 0.02529183402657509, + 0.007031816989183426, + 0.019701845943927765, + 0.01838819310069084, + 0.0039536370895802975, + 0.04647963494062424, + 0.007515440694987774, + 0.00485631451010704, + 0.013312884606420994, + 0.004295668099075556, + 0.07344947010278702, + 0.0077196029014885426, + 0.00794883631169796, + 0.00808696262538433, + 0.0066979555413126945, + 0.0071503478102386, + 0.007272942457348108, + 0.0069823130033910275, + 0.008509919047355652, + 0.009050630033016205, + 0.008151519112288952, + 0.01122218556702137, + 0.0059126815758645535, + 0.00831152219325304, + 0.012392350472509861, + 0.007557976059615612, + 0.020385080948472023, + 0.04922065511345863, + 0.01675400882959366, + 0.030649330466985703, + 0.008683828637003899, + 0.012819681316614151, + 0.034244026988744736, + 0.017148001119494438, + 0.06673800945281982, + 0.004096972290426493, + 0.0037243382539600134, + 0.005514117889106274, + 0.0022427099756896496, + 0.00173189549241215, + 0.002618845086544752, + 0.0020789941772818565, + 0.006725011859089136, + 0.00554592814296484, + 0.005052332766354084, + 0.0038192016072571278, + 0.004311645869165659, + 0.0032292825635522604, + 0.0035002236254513264, + 0.0033004742581397295, + 0.004825826268643141, + 0.018278826028108597, + 0.009160979650914669, + 0.008483324199914932, + 0.007006563246250153, + 0.005251628812402487, + 0.009549255482852459, + 0.004680138546973467, + 0.013615884818136692, + 0.02374321222305298, + 0.011858778074383736, + 0.01812989078462124, + 0.010669401846826077, + 0.009237684309482574, + 0.018765946850180626, + 0.005400774069130421, + 0.021356772631406784, + 0.012933260761201382, + 0.008660535328090191, + 0.007047645282000303, + 0.0051607307977974415, + 0.010526690632104874, + 0.007204410620033741, + 0.003958842717111111, + 0.009605667553842068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.038162361830472946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, deep chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse charging at full speed across an open, dusty plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's nostrils flaring and its eyes showing intense focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a billowing dust trail resembling golden mist created by the horse's hooves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799826443195343 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799826443195343, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799826443195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09370129555463791, + 0.6231123208999634, + 0.14061382412910461, + 0.02220601588487625, + 0.020009126514196396, + 0.014866219833493233, + 0.022275982424616814, + 0.012424321845173836, + 0.022903835400938988, + 0.02788703702390194 + ], + "credit_map": [ + 0.01765691116452217, + 0.01074602548032999, + 0.00971996784210205, + 0.007864445447921753, + 0.007297636941075325, + 0.010736706666648388, + 0.01113207172602415, + 0.018547529354691505, + 0.0668797641992569, + 0.009757580235600471, + 0.16814130544662476, + 0.028546031564474106, + 0.01625029742717743, + 0.04669627547264099, + 0.01771966740489006, + 0.269121378660202, + 0.013034163974225521, + 0.007990827783942223, + 0.025738077238202095, + 0.008609441109001637, + 0.004807669203728437, + 0.038385678082704544, + 0.0101785808801651, + 0.03186937794089317, + 0.0033863913267850876, + 0.0018229549750685692, + 0.004590249620378017, + 0.0014962212881073356, + 0.0014720920007675886, + 0.003457212820649147, + 0.0015934522962197661, + 0.004387442488223314, + 0.0025269524194300175, + 0.0018099063308909535, + 0.0031087480019778013, + 0.0015145861543715, + 0.0009875032119452953, + 0.005392697174102068, + 0.0014885998098179698, + 0.0031801313161849976, + 0.0019019409082829952, + 0.0011111084604635835, + 0.0035157473757863045, + 0.001269491040147841, + 0.0007924073725007474, + 0.0015480155125260353, + 0.0011149590136483312, + 0.0036125509068369865, + 0.0027225716039538383, + 0.00231099221855402, + 0.0024679219350218773, + 0.0028579458594322205, + 0.001729754381813109, + 0.0023853573948144913, + 0.003020839998498559, + 0.004780598916113377, + 0.002188317943364382, + 0.0014114945661276579, + 0.00178617425262928, + 0.0015911782393231988, + 0.0007074200548231602, + 0.001380312256515026, + 0.0008615381084382534, + 0.0024978870060294867, + 0.005520144011825323, + 0.0024138211738318205, + 0.0034897266887128353, + 0.00162231526337564, + 0.0018871570937335491, + 0.0034940019249916077, + 0.001308268983848393, + 0.0031683999113738537, + 0.008308880031108856, + 0.0045071677304804325, + 0.003057579044252634, + 0.002435701433569193, + 0.0023224649485200644, + 0.0029807526152580976, + 0.0010474943555891514, + 0.0032269982621073723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.029799826443195343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_tensor([7032, 7032])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_tensor([7032, 7032])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2c956a7c5b06d9bd0c30336b4c625611b1704ce1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_tensor([7032, 7032])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Static, low-angle, extreme close-ups capture a captivating time-lapse of the complex fungal growth. The entire scene is constructed in a vibrant pointillist style, with each element a tapestry composed of tiny, distinct dots. The sequence begins on the forest floor, a mosaic of dark, damp soil and decaying leaves, delicately cracked and ridged. From this spot, the fungus emerges, its pitted, honeycombed cap, and thick stem transforming into a shimmering array of earthy beige, dark brown, and creamy white. Filmed at a high frame rate and rendered as an ultra-smooth, slow-motion sequence, this growth shows the stem steadily and deliberately elongating upward, while the curled cap gracefully unfurls, revealing its complex, shadowy depths. Spotted sunlight filters through the unseen canopy, its ever-shifting patterns creating dynamic light throughout the scene, with each ray of light and shadow composed of a single luminous dot. This pointillist vision of life unfolding creates a tranquil and hypnotic atmosphere, highlighting the flowing, organic magic of nature with maximum clarity and fluid motion.", + "seed": 33170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_7032_seed_33170_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.47085950593464077, + "vlm_logprob": -0.47085950593464077, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.47085950593464077, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9110870361328125 + ], + "opposite_logprob_tokens": [ + -0.16108709573745728 + ], + "logprob_reward": -1.9110870361328125, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9110870361328125, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13826541602611542, + 0.07979270815849304, + 0.05362485721707344, + 0.14500680565834045, + 0.10419289767742157, + 0.15761810541152954, + 0.11402436345815659, + 0.05466300621628761, + 0.051790717989206314, + 0.1010211855173111 + ], + "credit_map": [ + 0.029685229063034058, + 0.019836096093058586, + 0.02222350612282753, + 0.013630240224301815, + 0.01390247605741024, + 0.010938686318695545, + 0.016216585412621498, + 0.011832593940198421, + 0.013813378289341927, + 0.008939343504607677, + 0.012419351376593113, + 0.01272435300052166, + 0.009235398843884468, + 0.007112397346645594, + 0.009146427735686302, + 0.006402058992534876, + 0.007711142301559448, + 0.00498022697865963, + 0.0059028188697993755, + 0.008790109306573868, + 0.005352618172764778, + 0.005804301705211401, + 0.00706314854323864, + 0.008020486682653427, + 0.024535715579986572, + 0.013568549416959286, + 0.015401818789541721, + 0.019878899678587914, + 0.01889076828956604, + 0.014792908914387226, + 0.018090827390551567, + 0.01984730362892151, + 0.010738502256572247, + 0.006544174626469612, + 0.023921899497509003, + 0.009779720567166805, + 0.005608103703707457, + 0.011248537339270115, + 0.012438323348760605, + 0.023913629353046417, + 0.025437157601118088, + 0.014871174469590187, + 0.024452200159430504, + 0.020776135846972466, + 0.014564305543899536, + 0.013232930563390255, + 0.022772744297981262, + 0.02151143178343773, + 0.012068086303770542, + 0.011066705919802189, + 0.016510844230651855, + 0.014963863417506218, + 0.018547991290688515, + 0.008931746706366539, + 0.02062269113957882, + 0.01131241861730814, + 0.006323777139186859, + 0.005868441425263882, + 0.009204322472214699, + 0.004820014350116253, + 0.0058287158608436584, + 0.006980548147112131, + 0.00543441204354167, + 0.010202772915363312, + 0.004396555479615927, + 0.008858723565936089, + 0.008080242201685905, + 0.006568725686520338, + 0.007050370331853628, + 0.004471132531762123, + 0.008215573616325855, + 0.004149388521909714, + 0.011618644930422306, + 0.012369521893560886, + 0.017419487237930298, + 0.01176364440470934, + 0.010114409029483795, + 0.011091334745287895, + 0.014481506310403347, + 0.012162629514932632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -1.9110870361328125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048694632947444916 + ], + "opposite_logprob_tokens": [ + -3.048694610595703 + ], + "logprob_reward": -0.048694632947444916, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048694632947444916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11172588169574738, + 0.18353891372680664, + 0.09304026514291763, + 0.14965063333511353, + 0.15918561816215515, + 0.09493508189916611, + 0.074045829474926, + 0.03818473592400551, + 0.04484304040670395, + 0.05085000395774841 + ], + "credit_map": [ + 0.024084443226456642, + 0.013620271347463131, + 0.016536325216293335, + 0.012982191517949104, + 0.010998514480888844, + 0.009417868219316006, + 0.015261639840900898, + 0.008824635297060013, + 0.04178845137357712, + 0.022961793467402458, + 0.025085465982556343, + 0.032716140151023865, + 0.017163243144750595, + 0.012832087464630604, + 0.016315922141075134, + 0.014675811864435673, + 0.023959258571267128, + 0.008765483275055885, + 0.014157237485051155, + 0.010843579657375813, + 0.0059928796254098415, + 0.009876427240669727, + 0.010787077248096466, + 0.008658315055072308, + 0.0314064659178257, + 0.01491872314363718, + 0.016369637101888657, + 0.021286992356181145, + 0.01744292862713337, + 0.0142202852293849, + 0.018031243234872818, + 0.015974359586834908, + 0.014522259123623371, + 0.016213953495025635, + 0.027375303208827972, + 0.02034308947622776, + 0.02168887108564377, + 0.018754146993160248, + 0.015607518143951893, + 0.02468045987188816, + 0.010108781047165394, + 0.009758451022207737, + 0.016868075355887413, + 0.011426045559346676, + 0.009982309304177761, + 0.011468561366200447, + 0.010075101628899574, + 0.015247758477926254, + 0.007678357418626547, + 0.006416438613086939, + 0.010375257581472397, + 0.012064068578183651, + 0.012008040212094784, + 0.006204703822731972, + 0.011510930024087429, + 0.0077880327589809895, + 0.0038580559194087982, + 0.003979775123298168, + 0.005510274786502123, + 0.004651689901947975, + 0.005221585743129253, + 0.005282619036734104, + 0.00542492838576436, + 0.004255807958543301, + 0.0035387868992984295, + 0.005398256704211235, + 0.007118727080523968, + 0.00728296022862196, + 0.004704770166426897, + 0.004845982883125544, + 0.008271459490060806, + 0.0036820960231125355, + 0.005481548607349396, + 0.0031836850102990866, + 0.007854054681956768, + 0.006142663303762674, + 0.004039464518427849, + 0.008028280921280384, + 0.009877883829176426, + 0.006242423318326473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.048694632947444916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1476994752883911, + 0.09107760339975357, + 0.04637154936790466, + 0.09730479121208191, + 0.07539249956607819, + 0.17496895790100098, + 0.10121679306030273, + 0.055567897856235504, + 0.04616210609674454, + 0.1642383635044098 + ], + "credit_map": [ + 0.03319884464144707, + 0.019080614671111107, + 0.023851647973060608, + 0.01623549498617649, + 0.01485629752278328, + 0.01048891618847847, + 0.01854451559484005, + 0.011443139053881168, + 0.021483682096004486, + 0.009776240214705467, + 0.015007754787802696, + 0.013955711387097836, + 0.009638131596148014, + 0.007809835951775312, + 0.007858997210860252, + 0.005547257140278816, + 0.007892465218901634, + 0.003836499061435461, + 0.006368367467075586, + 0.0076246438547968864, + 0.004454273264855146, + 0.005274786613881588, + 0.005320959724485874, + 0.005599557422101498, + 0.018126465380191803, + 0.009563377127051353, + 0.011862631887197495, + 0.017893999814987183, + 0.007623101584613323, + 0.008341717533767223, + 0.01583692990243435, + 0.008056573569774628, + 0.009711680002510548, + 0.010209622792899609, + 0.009095914661884308, + 0.011471422389149666, + 0.012905450537800789, + 0.007448197808116674, + 0.009295498952269554, + 0.005254716612398624, + 0.025588370859622955, + 0.017328601330518723, + 0.028288165107369423, + 0.022746417671442032, + 0.017712900415062904, + 0.010787954553961754, + 0.025151308625936508, + 0.027365248650312424, + 0.010431122966110706, + 0.010008247569203377, + 0.01484610978513956, + 0.013362528756260872, + 0.015732292085886, + 0.007427096366882324, + 0.01828944683074951, + 0.01111995056271553, + 0.006652320269495249, + 0.004956459626555443, + 0.0074188378639519215, + 0.006958812475204468, + 0.005685548763722181, + 0.006967248860746622, + 0.00845892820507288, + 0.008469744585454464, + 0.0036798007786273956, + 0.008121623657643795, + 0.006016060244292021, + 0.006191589403897524, + 0.006466189865022898, + 0.00536881759762764, + 0.006135433446615934, + 0.004182594362646341, + 0.01685512065887451, + 0.026389004662632942, + 0.02819063328206539, + 0.019753089174628258, + 0.014676004648208618, + 0.018657095730304718, + 0.02012774348258972, + 0.019589679315686226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021984234917908907 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021984234917908907, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021984234917908907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15128178894519806, + 0.0746808871626854, + 0.04955067113041878, + 0.1419254094362259, + 0.09373847395181656, + 0.14561213552951813, + 0.13861292600631714, + 0.06898076087236404, + 0.04170826077461243, + 0.09390870481729507 + ], + "credit_map": [ + 0.032960761338472366, + 0.019262392073869705, + 0.02411087602376938, + 0.0157542135566473, + 0.01652325876057148, + 0.01079003605991602, + 0.019167669117450714, + 0.01271259505301714, + 0.016733383759856224, + 0.008270053192973137, + 0.012687000446021557, + 0.010043020360171795, + 0.00801597349345684, + 0.006757617462426424, + 0.007674906402826309, + 0.004498932510614395, + 0.009443442337214947, + 0.005289265885949135, + 0.008314399048686028, + 0.007476301863789558, + 0.00471944035962224, + 0.004186338745057583, + 0.0057136560790240765, + 0.004407824017107487, + 0.024693341925740242, + 0.013628441840410233, + 0.015088190324604511, + 0.020794378593564034, + 0.018311619758605957, + 0.01437202375382185, + 0.017444785684347153, + 0.017592620104551315, + 0.0114037049934268, + 0.00556977279484272, + 0.021866438910365105, + 0.008641663007438183, + 0.0041107735596597195, + 0.010026751086115837, + 0.010792657732963562, + 0.021326707676053047, + 0.021089758723974228, + 0.012645361013710499, + 0.02582642249763012, + 0.017701057717204094, + 0.012853551656007767, + 0.012246303260326385, + 0.019200073555111885, + 0.024049600586295128, + 0.014551283791661263, + 0.013937251642346382, + 0.01926065981388092, + 0.02133313938975334, + 0.024845551699399948, + 0.010556304827332497, + 0.023220030590891838, + 0.010908699594438076, + 0.008229444734752178, + 0.0062401373870670795, + 0.011595241725444794, + 0.008896858431398869, + 0.006761165801435709, + 0.007139010354876518, + 0.01082353200763464, + 0.009295366704463959, + 0.003620295086875558, + 0.0070934174582362175, + 0.006416372023522854, + 0.005154722835868597, + 0.004782503936439753, + 0.004749856423586607, + 0.0054781995713710785, + 0.004412893671542406, + 0.010447072796523571, + 0.012147692032158375, + 0.0161336287856102, + 0.010477025993168354, + 0.009790606796741486, + 0.010257004760205746, + 0.013091214001178741, + 0.011564458720386028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0021984234917908907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706174492836 + ], + "opposite_logprob_tokens": [ + -1.1370617151260376 + ], + "logprob_reward": -0.38706174492836, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706174492836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10623694211244583, + 0.22068104147911072, + 0.18991635739803314, + 0.130900576710701, + 0.08804101496934891, + 0.04152785986661911, + 0.03926721215248108, + 0.0779213160276413, + 0.0353945754468441, + 0.07011310756206512 + ], + "credit_map": [ + 0.026407495141029358, + 0.01283759344369173, + 0.014289826154708862, + 0.01766011118888855, + 0.011309771798551083, + 0.007005427964031696, + 0.007777228020131588, + 0.00894949585199356, + 0.0589744858443737, + 0.022709723562002182, + 0.03706095740199089, + 0.030162781476974487, + 0.024353183805942535, + 0.018390491604804993, + 0.01453914400190115, + 0.014490277506411076, + 0.05602246895432472, + 0.016348568722605705, + 0.029270846396684647, + 0.018300799652934074, + 0.014322406612336636, + 0.021424289792776108, + 0.015073785558342934, + 0.019153183326125145, + 0.02065848745405674, + 0.012989035807549953, + 0.016937777400016785, + 0.02433140203356743, + 0.01150920893996954, + 0.010835116729140282, + 0.02063743770122528, + 0.013002104125916958, + 0.0077878013253211975, + 0.00887035671621561, + 0.015846867114305496, + 0.010942787863314152, + 0.010498376563191414, + 0.009888753294944763, + 0.009585880674421787, + 0.014620197005569935, + 0.0062875826843082905, + 0.004671339876949787, + 0.0052111390978097916, + 0.006222481839358807, + 0.004985426086932421, + 0.004462852142751217, + 0.006126641761511564, + 0.0035603928845375776, + 0.00410631624981761, + 0.003450042335316539, + 0.00550356088206172, + 0.0059073083102703094, + 0.006593567319214344, + 0.002946845255792141, + 0.006227053701877594, + 0.004532515536993742, + 0.005296641495078802, + 0.010254910215735435, + 0.011786705814301968, + 0.014006482437252998, + 0.010584326460957527, + 0.006767746526747942, + 0.009872527793049812, + 0.009351975284516811, + 0.0031169794965535402, + 0.003297734074294567, + 0.004277452826499939, + 0.004955875687301159, + 0.002885943278670311, + 0.0055500841699540615, + 0.0063769519329071045, + 0.004933554213494062, + 0.008237346075475216, + 0.009454835206270218, + 0.008522962220013142, + 0.011977054178714752, + 0.006893005687743425, + 0.007660939823836088, + 0.010215303860604763, + 0.007151660975068808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.38706174492836 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, low-angle, extreme close-ups capture a captivating time-lapse of the complex fungal growth. The entire scene is constructed in a vibrant pointillist style, with each element a tapestry composed of tiny, distinct dots. The sequence begins on the forest floor, a mosaic of dark, damp soil and decaying leaves, delicately cracked and ridged. From this spot, the fungus emerges, its pitted, honeycombed cap, and thick stem transforming into a shimmering array of earthy beige, dark brown, and creamy white. Filmed at a high frame rate and rendered as an ultra-smooth, slow-motion sequence, this growth shows the stem steadily and deliberately elongating upward, while the curled cap gracefully unfurls, revealing its complex, shadowy depths. Spotted sunlight filters through the unseen canopy, its ever-shifting patterns creating dynamic light throughout the scene, with each ray of light and shadow composed of a single luminous dot. This pointillist vision of life unfolding creates a tranquil and hypnotic atmosphere, highlighting the flowing, organic magic of nature with maximum clarity and fluid motion.", + "seed": 33171, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_7032_seed_33171_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21601593564264476, + "vlm_logprob": -0.21601593564264476, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.21601593564264476, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28237614035606384 + ], + "opposite_logprob_tokens": [ + -1.4073761701583862 + ], + "logprob_reward": -0.28237614035606384, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28237614035606384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06378994137048721, + 0.27759188413619995, + 0.07082092016935349, + 0.06968972831964493, + 0.04205377772450447, + 0.056035034358501434, + 0.33518338203430176, + 0.0397716760635376, + 0.024383122101426125, + 0.020680496469140053 + ], + "credit_map": [ + 0.007838494144380093, + 0.006692211143672466, + 0.011005522683262825, + 0.008941750042140484, + 0.00549474498257041, + 0.007567564491182566, + 0.008498054929077625, + 0.00775159802287817, + 0.022044774144887924, + 0.014569967053830624, + 0.03791910409927368, + 0.03271479159593582, + 0.014804832637310028, + 0.044610247015953064, + 0.03546690195798874, + 0.07546129077672958, + 0.005700687412172556, + 0.007533403113484383, + 0.011145628988742828, + 0.00867271889001131, + 0.007486366201192141, + 0.005420350935310125, + 0.017118481919169426, + 0.007743277586996555, + 0.006940237712115049, + 0.005290299654006958, + 0.011425265111029148, + 0.011028003878891468, + 0.006699456367641687, + 0.005464000161737204, + 0.013834596611559391, + 0.00900786742568016, + 0.005571672227233648, + 0.004588243551552296, + 0.006406917702406645, + 0.005899510346353054, + 0.005438769236207008, + 0.003042708383873105, + 0.0055230543948709965, + 0.0055829002521932125, + 0.0036546490155160427, + 0.00386437913402915, + 0.008699520491063595, + 0.007427411153912544, + 0.0035951309837400913, + 0.0034221895039081573, + 0.01678801327943802, + 0.00858374685049057, + 0.018668875098228455, + 0.008624766953289509, + 0.12863829731941223, + 0.01670924574136734, + 0.01362813450396061, + 0.019549978896975517, + 0.04260937497019768, + 0.0867547020316124, + 0.004996928386390209, + 0.0042283195070922375, + 0.004672217182815075, + 0.003673331346362829, + 0.0035010443534702063, + 0.002429588232189417, + 0.007427674252539873, + 0.008842574432492256, + 0.0024357903748750687, + 0.003413732396438718, + 0.002548634074628353, + 0.0034922773484140635, + 0.002720578806474805, + 0.0018031932413578033, + 0.00375113426707685, + 0.004217784386128187, + 0.0017164932796731591, + 0.0026454522740095854, + 0.0036408090963959694, + 0.0019959870260208845, + 0.0021472377702593803, + 0.001838166150264442, + 0.0028857984580099583, + 0.003810551716014743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.28237614035606384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035216063261032 + ], + "opposite_logprob_tokens": [ + -2.3503520488739014 + ], + "logprob_reward": -0.10035216063261032, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035216063261032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16051191091537476, + 0.2128002941608429, + 0.037365805357694626, + 0.04485463351011276, + 0.041444260627031326, + 0.07656919956207275, + 0.2626989781856537, + 0.038169652223587036, + 0.02436109073460102, + 0.1012241542339325 + ], + "credit_map": [ + 0.01378015335649252, + 0.01292266882956028, + 0.030222086235880852, + 0.01488984003663063, + 0.009336509741842747, + 0.02829524129629135, + 0.018902702257037163, + 0.032162703573703766, + 0.019187357276678085, + 0.010914906859397888, + 0.028795551508665085, + 0.026196228340268135, + 0.01105410698801279, + 0.03425624594092369, + 0.024308685213327408, + 0.058087196201086044, + 0.004736755508929491, + 0.004560315981507301, + 0.005128737539052963, + 0.003971814177930355, + 0.0035566173028200865, + 0.004680480342358351, + 0.006895539350807667, + 0.0038355467841029167, + 0.004758266266435385, + 0.004273007623851299, + 0.006361063569784164, + 0.006849372759461403, + 0.005031721666455269, + 0.0039265770465135574, + 0.00823619682341814, + 0.005418428685516119, + 0.0051527782343328, + 0.0028681838884949684, + 0.0066934688948094845, + 0.004841316491365433, + 0.0036382859107106924, + 0.006056398153305054, + 0.0057328809052705765, + 0.006460944190621376, + 0.0051346695981919765, + 0.004411923233419657, + 0.022233858704566956, + 0.00857811700552702, + 0.005305576138198376, + 0.0038146041333675385, + 0.012073504738509655, + 0.015016937628388405, + 0.013423141092061996, + 0.00610654940828681, + 0.10252845287322998, + 0.012443773448467255, + 0.010347601026296616, + 0.015050927177071571, + 0.03499876707792282, + 0.06779976934194565, + 0.004904198460280895, + 0.004145249258726835, + 0.006581238005310297, + 0.004927879199385643, + 0.002796957502141595, + 0.003278519958257675, + 0.006038147956132889, + 0.005497461184859276, + 0.0019228234887123108, + 0.0018896579276770353, + 0.004035512916743755, + 0.0023492693435400724, + 0.0019362855236977339, + 0.0016608345322310925, + 0.003953452687710524, + 0.006613255012780428, + 0.009759459644556046, + 0.008355780504643917, + 0.022699248045682907, + 0.012924058362841606, + 0.006094862706959248, + 0.006510536652058363, + 0.014142293483018875, + 0.020737914368510246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.10035216063261032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024911104701459408 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024911104701459408, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024911104701459408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026269759982824326, + 0.042941268533468246, + 0.17168474197387695, + 0.24401594698429108, + 0.10516978800296783, + 0.037094686180353165, + 0.20922498404979706, + 0.07289290428161621, + 0.04229540750384331, + 0.04841054975986481 + ], + "credit_map": [ + 0.0026660331059247255, + 0.0029033171012997627, + 0.0034400303848087788, + 0.0036105273757129908, + 0.002514156047254801, + 0.0034462197218090296, + 0.003208619076758623, + 0.004480856470763683, + 0.004480036441236734, + 0.003567697014659643, + 0.006285479292273521, + 0.0056389025412499905, + 0.0037982913199812174, + 0.0057668015360832214, + 0.005804266314953566, + 0.007599796634167433, + 0.01919974945485592, + 0.016914084553718567, + 0.023242343217134476, + 0.017374509945511818, + 0.018554264679551125, + 0.025559106841683388, + 0.02692708931863308, + 0.023913592100143433, + 0.024265719577670097, + 0.025893576443195343, + 0.04864261671900749, + 0.029547587037086487, + 0.01968131586909294, + 0.021893322467803955, + 0.04685394465923309, + 0.027237877249717712, + 0.012280691415071487, + 0.00872997660189867, + 0.020426511764526367, + 0.017913762480020523, + 0.008740615099668503, + 0.00904097966849804, + 0.0152666587382555, + 0.012770588509738445, + 0.003250137669965625, + 0.0032445460092276335, + 0.006518823094666004, + 0.005526759196072817, + 0.002141557401046157, + 0.0016212842892855406, + 0.009390745311975479, + 0.005400829017162323, + 0.010443667881190777, + 0.00630218768492341, + 0.07504579424858093, + 0.011485815979540348, + 0.007491486147046089, + 0.012486585415899754, + 0.03392273932695389, + 0.05204671621322632, + 0.004167471546679735, + 0.0046632071025669575, + 0.012389960698783398, + 0.009318751282989979, + 0.004636475816369057, + 0.0051582204177975655, + 0.017256878316402435, + 0.01530193816870451, + 0.003529652487486601, + 0.005814841017127037, + 0.005073108244687319, + 0.005223705433309078, + 0.004997374024242163, + 0.003487630980089307, + 0.00678597204387188, + 0.007383126299828291, + 0.003475409932434559, + 0.0047433990985155106, + 0.008035456761717796, + 0.00809727143496275, + 0.004327549133449793, + 0.005715824663639069, + 0.006289997603744268, + 0.007725640200078487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.0024911104701459408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00151305110193789 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.00151305110193789, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00151305110193789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07077766209840775, + 0.048342328518629074, + 0.16330721974372864, + 0.11272023618221283, + 0.0983848124742508, + 0.08816257864236832, + 0.2581136226654053, + 0.06815949082374573, + 0.048179492354393005, + 0.04385259002447128 + ], + "credit_map": [ + 0.006363802123814821, + 0.00893429946154356, + 0.011921031400561333, + 0.01267812680453062, + 0.006213545333594084, + 0.00820899661630392, + 0.01113385334610939, + 0.005324001424014568, + 0.005173390731215477, + 0.004211243242025375, + 0.007421655114740133, + 0.006945797707885504, + 0.0037752308417111635, + 0.004760174546390772, + 0.006483376491814852, + 0.009571456350386143, + 0.020841192454099655, + 0.01653539575636387, + 0.023260897025465965, + 0.015844428911805153, + 0.01954425498843193, + 0.026410818099975586, + 0.025048406794667244, + 0.01582181639969349, + 0.013280821964144707, + 0.008411196060478687, + 0.022250138223171234, + 0.01319381594657898, + 0.01051240786910057, + 0.013999756425619125, + 0.018828250467777252, + 0.012243831530213356, + 0.014153063297271729, + 0.008509526960551739, + 0.01529509387910366, + 0.009532121941447258, + 0.008391152136027813, + 0.012398529797792435, + 0.014915047213435173, + 0.015190266072750092, + 0.01035406906157732, + 0.005560624413192272, + 0.025028694421052933, + 0.01107162144035101, + 0.0042444923892617226, + 0.004224145319312811, + 0.011895721778273582, + 0.01578320376574993, + 0.014998783357441425, + 0.008466610684990883, + 0.09551867842674255, + 0.015066301450133324, + 0.012470031157135963, + 0.015936879441142082, + 0.03464697673916817, + 0.06100933998823166, + 0.004531116224825382, + 0.004206265788525343, + 0.011659872718155384, + 0.008267443627119064, + 0.0043805260211229324, + 0.003976314328610897, + 0.01638457365334034, + 0.014753374271094799, + 0.003068670630455017, + 0.005227430257946253, + 0.005810575094074011, + 0.0074404217302799225, + 0.003609344596043229, + 0.004394723568111658, + 0.008633843623101711, + 0.009994477964937687, + 0.0033483500592410564, + 0.005041881464421749, + 0.004762541502714157, + 0.0054064588621258736, + 0.0064470525830984116, + 0.0037711546756327152, + 0.007053919602185488, + 0.008021224290132523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.00151305110193789 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933472156524658 + ], + "opposite_logprob_tokens": [ + -0.6933472156524658 + ], + "logprob_reward": -0.6933472156524658, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933472156524658, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10657770931720734, + 0.1803029477596283, + 0.08656914532184601, + 0.08994869142770767, + 0.10564512014389038, + 0.1631438434123993, + 0.1539573222398758, + 0.04122128710150719, + 0.04177026078104973, + 0.030863646417856216 + ], + "credit_map": [ + 0.011145709082484245, + 0.013816369697451591, + 0.013323172926902771, + 0.016412680968642235, + 0.00920068472623825, + 0.01156522985547781, + 0.016440121456980705, + 0.014673737809062004, + 0.01360237319022417, + 0.009608590044081211, + 0.02616811916232109, + 0.021480830386281013, + 0.011147309094667435, + 0.029862036928534508, + 0.023719048127532005, + 0.04471464455127716, + 0.009742640890181065, + 0.00758582865819335, + 0.013154596090316772, + 0.00650982279330492, + 0.007599335629492998, + 0.009548187255859375, + 0.01587747596204281, + 0.01655125804245472, + 0.01244437973946333, + 0.008325353264808655, + 0.014737424440681934, + 0.011084471829235554, + 0.012644934467971325, + 0.008784255012869835, + 0.012209431268274784, + 0.009718445129692554, + 0.01729731634259224, + 0.00803865771740675, + 0.020182926207780838, + 0.012989435344934464, + 0.007306116633117199, + 0.010181285440921783, + 0.0154978446662426, + 0.014151528477668762, + 0.019228603690862656, + 0.015333027578890324, + 0.02580556459724903, + 0.020348571240901947, + 0.010833549313247204, + 0.00958025362342596, + 0.03700724616646767, + 0.025007031857967377, + 0.008398515172302723, + 0.013485371135175228, + 0.022703878581523895, + 0.014282145537436008, + 0.014967244118452072, + 0.016627997159957886, + 0.03540382161736488, + 0.02808835171163082, + 0.004308404866605997, + 0.004111255519092083, + 0.00807130429893732, + 0.005729044787585735, + 0.0030626938678324223, + 0.002398683689534664, + 0.006091837305575609, + 0.0074480632320046425, + 0.003519509918987751, + 0.005069822538644075, + 0.005125946365296841, + 0.005366383586078882, + 0.0038497145287692547, + 0.0034495058935135603, + 0.007997524924576283, + 0.007391853723675013, + 0.003313186578452587, + 0.004373508971184492, + 0.004507030826061964, + 0.003441482549533248, + 0.0037263541016727686, + 0.0031317281536757946, + 0.004197925794869661, + 0.004172428045421839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.6933472156524658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "Static, low-angle, extreme close-ups capture a captivating time-lapse of the complex fungal growth. The entire scene is constructed in a vibrant pointillist style, with each element a tapestry composed of tiny, distinct dots. The sequence begins on the forest floor, a mosaic of dark, damp soil and decaying leaves, delicately cracked and ridged. From this spot, the fungus emerges, its pitted, honeycombed cap, and thick stem transforming into a shimmering array of earthy beige, dark brown, and creamy white. Filmed at a high frame rate and rendered as an ultra-smooth, slow-motion sequence, this growth shows the stem steadily and deliberately elongating upward, while the curled cap gracefully unfurls, revealing its complex, shadowy depths. Spotted sunlight filters through the unseen canopy, its ever-shifting patterns creating dynamic light throughout the scene, with each ray of light and shadow composed of a single luminous dot. This pointillist vision of life unfolding creates a tranquil and hypnotic atmosphere, highlighting the flowing, organic magic of nature with maximum clarity and fluid motion.", + "seed": 33172, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_7032_seed_33172_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.4355080961249769, + "vlm_logprob": -0.4355080961249769, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.4355080961249769, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4075069427490234 + ], + "opposite_logprob_tokens": [ + -0.2825070023536682 + ], + "logprob_reward": -1.4075069427490234, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4075069427490234, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15117354691028595, + 0.05415017530322075, + 0.07129590213298798, + 0.04224001616239548, + 0.10479138791561127, + 0.032754674553871155, + 0.4420945942401886, + 0.032093845307826996, + 0.038020554929971695, + 0.03138525038957596 + ], + "credit_map": [ + 0.023152964189648628, + 0.019522113725543022, + 0.024299349635839462, + 0.016903799027204514, + 0.019376713782548904, + 0.014255968853831291, + 0.023209232836961746, + 0.010453425347805023, + 0.007095527369529009, + 0.0071649812161922455, + 0.0054911463521420956, + 0.005731156561523676, + 0.006754453293979168, + 0.008191117085516453, + 0.00833948329091072, + 0.005382311996072531, + 0.00787085760384798, + 0.004272304475307465, + 0.02150350622832775, + 0.005818021018058062, + 0.0058442940935492516, + 0.0038043407257646322, + 0.01755107194185257, + 0.00463150953873992, + 0.00862746313214302, + 0.0026944384444504976, + 0.006296479143202305, + 0.004857370629906654, + 0.0038044617976993322, + 0.003937138710170984, + 0.004979424644261599, + 0.007043241057544947, + 0.010108170099556446, + 0.009686669334769249, + 0.01711091212928295, + 0.015781531110405922, + 0.012960081920027733, + 0.007837964221835136, + 0.017062077298760414, + 0.014243992045521736, + 0.006375032011419535, + 0.0029321920592337847, + 0.0036120687145739794, + 0.004358449950814247, + 0.003049704944714904, + 0.00269404542632401, + 0.004775965120643377, + 0.004957219585776329, + 0.06088746711611748, + 0.03361954912543297, + 0.10543642938137054, + 0.03248121589422226, + 0.06566749513149261, + 0.03829031065106392, + 0.048607729375362396, + 0.05710441246628761, + 0.0030654172878712416, + 0.003610519925132394, + 0.007060007192194462, + 0.0036317938938736916, + 0.0034398010466247797, + 0.003054177388548851, + 0.004389213863760233, + 0.0038429186679422855, + 0.005767428316175938, + 0.004531425889581442, + 0.0044005741365253925, + 0.00498707452788949, + 0.005458998028188944, + 0.003528457833454013, + 0.004868384916335344, + 0.004478211514651775, + 0.003567485371604562, + 0.0038964191917330027, + 0.0036178571172058582, + 0.006030059419572353, + 0.003153278259560466, + 0.0032086018472909927, + 0.004691653419286013, + 0.003219899022951722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -1.4075069427490234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709225714206696 + ], + "opposite_logprob_tokens": [ + -2.1270923614501953 + ], + "logprob_reward": -0.12709225714206696, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709225714206696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23982223868370056, + 0.06965538114309311, + 0.024271076545119286, + 0.024564141407608986, + 0.0077584777027368546, + 0.04673546925187111, + 0.5683767199516296, + 0.0102060716599226, + 0.006276784464716911, + 0.0023336068261414766 + ], + "credit_map": [ + 0.05508040264248848, + 0.026283985003829002, + 0.04546579346060753, + 0.025246182456612587, + 0.026454733684659004, + 0.02411073073744774, + 0.01881265826523304, + 0.018367759883403778, + 0.015074566006660461, + 0.005629129707813263, + 0.010667138732969761, + 0.006398648954927921, + 0.008475253358483315, + 0.008662686683237553, + 0.008300280198454857, + 0.006447683088481426, + 0.003355487948283553, + 0.002216398250311613, + 0.004822293296456337, + 0.0029042446985840797, + 0.002179034985601902, + 0.0011398119386285543, + 0.005504016298800707, + 0.0021497898269444704, + 0.0033875564113259315, + 0.0017071499023586512, + 0.005780814681202173, + 0.002078312449157238, + 0.0018364944262430072, + 0.0020777659956365824, + 0.0016050912672653794, + 0.006090958137065172, + 0.0010796652641147375, + 0.0006180498749017715, + 0.0010849606478586793, + 0.0012007508194074035, + 0.0007481189677491784, + 0.0008534801308996975, + 0.0010145038831979036, + 0.0011589489877223969, + 0.004544881172478199, + 0.0031902201008051634, + 0.010905180126428604, + 0.0037839056458324194, + 0.014633920975029469, + 0.0030792728066444397, + 0.002949371701106429, + 0.003648718586191535, + 0.08076015114784241, + 0.04336579889059067, + 0.1310150921344757, + 0.04307188838720322, + 0.07298702746629715, + 0.051081858575344086, + 0.07041974365711212, + 0.07567519694566727, + 0.0008822626550681889, + 0.0009118380839936435, + 0.002455695765092969, + 0.0017827106639742851, + 0.0007868718821555376, + 0.001155075733549893, + 0.0013951162109151483, + 0.0008365020621567965, + 0.0006151327979750931, + 0.0007463907240889966, + 0.0008314148872159421, + 0.001008450286462903, + 0.000793072918895632, + 0.0005415257182903588, + 0.0011472469195723534, + 0.0005935506778769195, + 0.0003421278961468488, + 0.00020051341562066227, + 0.0003349145990796387, + 0.00031588785350322723, + 0.00024394581851083785, + 0.00030704919481649995, + 0.0003138228494208306, + 0.00027534557739272714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.12709225714206696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076340362429618835 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.0076340362429618835, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076340362429618835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09724432229995728, + 0.17841915786266327, + 0.257750928401947, + 0.0561317540705204, + 0.11898531764745712, + 0.0622154176235199, + 0.10004676133394241, + 0.0391559898853302, + 0.03723258897662163, + 0.05281775817275047 + ], + "credit_map": [ + 0.012848230078816414, + 0.009697694331407547, + 0.01590808853507042, + 0.014332087710499763, + 0.011791493743658066, + 0.009202172979712486, + 0.01252515334635973, + 0.010939393192529678, + 0.04216697812080383, + 0.012671194039285183, + 0.03355466574430466, + 0.014045308344066143, + 0.023113595321774483, + 0.020413970574736595, + 0.017248013988137245, + 0.015205426141619682, + 0.025421490892767906, + 0.009651429019868374, + 0.08653508126735687, + 0.01974746584892273, + 0.014548255130648613, + 0.01105410698801279, + 0.07458094507455826, + 0.016212139278650284, + 0.00640550721436739, + 0.004868668969720602, + 0.007154969964176416, + 0.008044439367949963, + 0.008058263920247555, + 0.005234192591160536, + 0.009941861033439636, + 0.0064238496124744415, + 0.0209786519408226, + 0.01261166762560606, + 0.018935156986117363, + 0.014237146824598312, + 0.012257004156708717, + 0.008031109347939491, + 0.015364598482847214, + 0.01656997948884964, + 0.007074730005115271, + 0.007309675682336092, + 0.007900872267782688, + 0.009707002900540829, + 0.007157187908887863, + 0.004824456758797169, + 0.011093460023403168, + 0.007148037198930979, + 0.015195869840681553, + 0.007328535430133343, + 0.020713912323117256, + 0.008469436317682266, + 0.007616458460688591, + 0.009783653542399406, + 0.016918661072850227, + 0.014020244590938091, + 0.003381964284926653, + 0.004283956252038479, + 0.006219722330570221, + 0.00442484999075532, + 0.003995353821665049, + 0.004989054519683123, + 0.007300912868231535, + 0.004560176748782396, + 0.004176313988864422, + 0.004512674640864134, + 0.005056826397776604, + 0.0050493567250669, + 0.005783434025943279, + 0.003170135896652937, + 0.006748878862708807, + 0.0027349689044058323, + 0.005947108380496502, + 0.006191896740347147, + 0.007767036557197571, + 0.009291940368711948, + 0.005086286924779415, + 0.006028575822710991, + 0.0072540538385510445, + 0.005250861402601004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.0076340362429618835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491943072527647 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.002491943072527647, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491943072527647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09055124223232269, + 0.039571572095155716, + 0.1954195201396942, + 0.037919458001852036, + 0.10635261237621307, + 0.04371955245733261, + 0.40745681524276733, + 0.022497864440083504, + 0.02914038673043251, + 0.027370942756533623 + ], + "credit_map": [ + 0.01581289805471897, + 0.011769793927669525, + 0.014002660289406776, + 0.009452257305383682, + 0.011651662178337574, + 0.007684110198169947, + 0.01263803243637085, + 0.00753982737660408, + 0.005101067014038563, + 0.0049032187089324, + 0.004408452659845352, + 0.004538669250905514, + 0.0052298326045274734, + 0.00466776080429554, + 0.006046749651432037, + 0.0046758223325014114, + 0.019003640860319138, + 0.006899567320942879, + 0.06466671079397202, + 0.014501729048788548, + 0.01030353456735611, + 0.008569663390517235, + 0.058214545249938965, + 0.01326011959463358, + 0.004580228589475155, + 0.0024979799054563046, + 0.005002299789339304, + 0.004895699676126242, + 0.004383517894893885, + 0.003924836870282888, + 0.006540463771671057, + 0.006094435695558786, + 0.010609510354697704, + 0.010092602111399174, + 0.016201233491301537, + 0.016798237338662148, + 0.012886468321084976, + 0.007589958142489195, + 0.017436984926462173, + 0.014737618155777454, + 0.007202968001365662, + 0.004287497606128454, + 0.007354957517236471, + 0.005430135875940323, + 0.0074279760010540485, + 0.0038027935661375523, + 0.004197083413600922, + 0.0040161400102078915, + 0.05995393171906471, + 0.030918754637241364, + 0.09290356934070587, + 0.031529586762189865, + 0.04182085022330284, + 0.03735874593257904, + 0.056597646325826645, + 0.05637376382946968, + 0.002472140360623598, + 0.0029112768825143576, + 0.0038154986687004566, + 0.002201391151174903, + 0.0023036878556013107, + 0.0031841339077800512, + 0.0032110849861055613, + 0.002398650161921978, + 0.004910999443382025, + 0.0029632337391376495, + 0.003493270603939891, + 0.00356080150231719, + 0.004126107320189476, + 0.00271931872703135, + 0.003986458294093609, + 0.0033801994286477566, + 0.0029263226315379143, + 0.003747502574697137, + 0.0027747745625674725, + 0.004769321531057358, + 0.003489625407382846, + 0.002856155391782522, + 0.004027745686471462, + 0.0027794938068836927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.002491943072527647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328153014183044 + ], + "opposite_logprob_tokens": [ + -0.7578153014183044 + ], + "logprob_reward": -0.6328153014183044, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328153014183044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3011927008628845, + 0.0908714085817337, + 0.14925765991210938, + 0.07303877174854279, + 0.052085209637880325, + 0.07116124033927917, + 0.18145209550857544, + 0.03827420994639397, + 0.020826566964387894, + 0.021840158849954605 + ], + "credit_map": [ + 0.06600576639175415, + 0.02934800088405609, + 0.059966716915369034, + 0.03160155564546585, + 0.031619537621736526, + 0.0306550245732069, + 0.026612285524606705, + 0.02538379840552807, + 0.019782287999987602, + 0.008660176768898964, + 0.010789335705339909, + 0.010315794497728348, + 0.012116958387196064, + 0.008086537942290306, + 0.01174598466604948, + 0.009374323301017284, + 0.013867480680346489, + 0.009288405999541283, + 0.042275652289390564, + 0.013253326527774334, + 0.010656051337718964, + 0.009912261739373207, + 0.03828620910644531, + 0.011718258261680603, + 0.006721781566739082, + 0.007478401064872742, + 0.011521507985889912, + 0.008952862583100796, + 0.009715352207422256, + 0.0054672169499099255, + 0.009526730515062809, + 0.013654910959303379, + 0.007897177711129189, + 0.007849004119634628, + 0.0065146554261446, + 0.007227012887597084, + 0.006422670092433691, + 0.005118155851960182, + 0.005492661613970995, + 0.005563866347074509, + 0.008043956011533737, + 0.0078117274679243565, + 0.009912963956594467, + 0.010017606429755688, + 0.0098191536962986, + 0.005225503351539373, + 0.011273476295173168, + 0.009056842885911465, + 0.029307769611477852, + 0.012786059640347958, + 0.04590575024485588, + 0.012150156311690807, + 0.01772427372634411, + 0.01683376356959343, + 0.023649653419852257, + 0.02309466525912285, + 0.004054965917021036, + 0.006138708908110857, + 0.010165220126509666, + 0.0034967747051268816, + 0.002909099915996194, + 0.003323953365907073, + 0.004665491171181202, + 0.003519992344081402, + 0.0019174162298440933, + 0.00232130428776145, + 0.0022425425704568624, + 0.00393617944791913, + 0.0023600400891155005, + 0.0016691610217094421, + 0.003925427328795195, + 0.0024544934276491404, + 0.002149739069864154, + 0.0023718199227005243, + 0.0028914858121424913, + 0.004411858506500721, + 0.0023032594472169876, + 0.0019535308238118887, + 0.0031127280090004206, + 0.002645733766257763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.6328153014183044 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, low-angle, extreme close-ups capture a captivating time-lapse of the complex fungal growth. The entire scene is constructed in a vibrant pointillist style, with each element a tapestry composed of tiny, distinct dots. The sequence begins on the forest floor, a mosaic of dark, damp soil and decaying leaves, delicately cracked and ridged. From this spot, the fungus emerges, its pitted, honeycombed cap, and thick stem transforming into a shimmering array of earthy beige, dark brown, and creamy white. Filmed at a high frame rate and rendered as an ultra-smooth, slow-motion sequence, this growth shows the stem steadily and deliberately elongating upward, while the curled cap gracefully unfurls, revealing its complex, shadowy depths. Spotted sunlight filters through the unseen canopy, its ever-shifting patterns creating dynamic light throughout the scene, with each ray of light and shadow composed of a single luminous dot. This pointillist vision of life unfolding creates a tranquil and hypnotic atmosphere, highlighting the flowing, organic magic of nature with maximum clarity and fluid motion.", + "seed": 33173, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_1/index_7032_seed_33173_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2769165922887623, + "vlm_logprob": -0.2769165922887623, + "vlm_yes_no_margin": 3.299999988079071, + "vlm_token_credit": -0.2769165922887623, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0548471212387085 + ], + "opposite_logprob_tokens": [ + -0.4298471510410309 + ], + "logprob_reward": -1.0548471212387085, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0548471212387085, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09906580299139023, + 0.10951275378465652, + 0.03204352408647537, + 0.10388693958520889, + 0.037201814353466034, + 0.06717351078987122, + 0.2884499728679657, + 0.1487199068069458, + 0.05076088756322861, + 0.06318488717079163 + ], + "credit_map": [ + 0.020378215238451958, + 0.009698877111077309, + 0.006891580298542976, + 0.008510920219123363, + 0.006196093279868364, + 0.004878185223788023, + 0.035667404532432556, + 0.006844516843557358, + 0.018432587385177612, + 0.007646637968719006, + 0.010326916351914406, + 0.008022453635931015, + 0.03408107906579971, + 0.012303996831178665, + 0.011221654713153839, + 0.007477418519556522, + 0.002119392389431596, + 0.0015205440577119589, + 0.00568242184817791, + 0.0027135421987622976, + 0.00264963717199862, + 0.0111406110227108, + 0.002937757410109043, + 0.003279615193605423, + 0.02391408570110798, + 0.004710470791906118, + 0.012383266352117062, + 0.007091546896845102, + 0.005279554054141045, + 0.03025355562567711, + 0.007438415661454201, + 0.012816043570637703, + 0.003042889991775155, + 0.0035612795036286116, + 0.004404930397868156, + 0.004578718915581703, + 0.005996242165565491, + 0.0067866211757063866, + 0.005594127811491489, + 0.003237000899389386, + 0.004451325628906488, + 0.00209451443515718, + 0.022423379123210907, + 0.004103195387870073, + 0.002653444418683648, + 0.026454759761691093, + 0.002458979142829776, + 0.002533908002078533, + 0.01801888644695282, + 0.012135623022913933, + 0.08822804689407349, + 0.015628287568688393, + 0.008624149486422539, + 0.02483552135527134, + 0.013020808808505535, + 0.10795862972736359, + 0.0047257766127586365, + 0.006146061234176159, + 0.01489295158535242, + 0.006783993914723396, + 0.005428474862128496, + 0.05583268776535988, + 0.005339747294783592, + 0.04957020282745361, + 0.006933362223207951, + 0.005053938832134008, + 0.006086447741836309, + 0.007392255589365959, + 0.007599383592605591, + 0.008751355111598969, + 0.005430260207504034, + 0.003513879841193557, + 0.014270774088799953, + 0.005023824516683817, + 0.005497215315699577, + 0.007117308676242828, + 0.004980133380740881, + 0.018787462264299393, + 0.0035736504942178726, + 0.003934512846171856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -1.0548471212387085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674400432035327 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674400432035327, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674400432035327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17578378319740295, + 0.040234409272670746, + 0.027821531519293785, + 0.029367052018642426, + 0.0225945133715868, + 0.07558697462081909, + 0.5154465436935425, + 0.02599503844976425, + 0.020643603056669235, + 0.06652656197547913 + ], + "credit_map": [ + 0.03537163510918617, + 0.050715211778879166, + 0.021937325596809387, + 0.011643261648714542, + 0.014230573549866676, + 0.01335423719137907, + 0.011994702741503716, + 0.01653684675693512, + 0.007478660438209772, + 0.0038422264624387026, + 0.004680995363742113, + 0.003131750039756298, + 0.006683430168777704, + 0.004958317149430513, + 0.006068777292966843, + 0.0033902537543326616, + 0.004433675203472376, + 0.0019536346662789583, + 0.0030105523765087128, + 0.002927391557022929, + 0.0022985839750617743, + 0.006526391953229904, + 0.004236632026731968, + 0.002434668829664588, + 0.004489680752158165, + 0.0024002266582101583, + 0.002842577872797847, + 0.0032184687443077564, + 0.0020971850026398897, + 0.004701114259660244, + 0.0036759874783456326, + 0.005941812414675951, + 0.0016657623928040266, + 0.0016472275601699948, + 0.0037793652154505253, + 0.002447878709062934, + 0.002535539213567972, + 0.005651648156344891, + 0.0032023906242102385, + 0.0016647008014842868, + 0.004156609997153282, + 0.003597817150875926, + 0.024414781481027603, + 0.0069048223085701466, + 0.0041528488509356976, + 0.02362353913486004, + 0.0038070182781666517, + 0.004929534159600735, + 0.031818121671676636, + 0.02163192443549633, + 0.1575920134782791, + 0.025996103882789612, + 0.014579272828996181, + 0.0422479473054409, + 0.023290695622563362, + 0.19829046726226807, + 0.0019380113808438182, + 0.0021598548628389835, + 0.0022448331583291292, + 0.0021145024802535772, + 0.0020328492391854525, + 0.006474405061453581, + 0.002440793439745903, + 0.006589789874851704, + 0.0022062333300709724, + 0.0020509413443505764, + 0.0026726732030510902, + 0.0034872761461883783, + 0.0034303285647183657, + 0.0025644877459853888, + 0.002388934139162302, + 0.0018427271861582994, + 0.01114679966121912, + 0.006347781512886286, + 0.009080221876502037, + 0.010536791756749153, + 0.006333413068205118, + 0.008280017413198948, + 0.009368439204990864, + 0.005433098413050175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00674400432035327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03269149363040924, + 0.016291353851556778, + 0.026333551853895187, + 0.09023553133010864, + 0.015323283150792122, + 0.05541297048330307, + 0.28981709480285645, + 0.23424389958381653, + 0.028666755184531212, + 0.2109839916229248 + ], + "credit_map": [ + 0.0047574955970048904, + 0.003207131754606962, + 0.0029628053307533264, + 0.0035080797970294952, + 0.003396155545488, + 0.0035243660677224398, + 0.008592860773205757, + 0.0027426027227193117, + 0.0019078511977568269, + 0.001527625136077404, + 0.002108779037371278, + 0.0024044334422796965, + 0.0017120612319558859, + 0.002636980963870883, + 0.0028415571432560682, + 0.0011520682601258159, + 0.003067907178774476, + 0.001977083971723914, + 0.004152200650423765, + 0.0021443194709718227, + 0.0020769028924405575, + 0.009445075877010822, + 0.0020480952225625515, + 0.0014219689182937145, + 0.019864864647388458, + 0.004316492006182671, + 0.011506272479891777, + 0.005967105273157358, + 0.004671904724091291, + 0.028181836009025574, + 0.00604963256046176, + 0.009677428752183914, + 0.0011528912000358105, + 0.001253837370313704, + 0.0028335365932434797, + 0.0016369278309866786, + 0.0017747062956914306, + 0.0036508517805486917, + 0.0018856585957109928, + 0.001134874182753265, + 0.0032350115943700075, + 0.0022308568004518747, + 0.018067078664898872, + 0.0037089732941240072, + 0.003037764225155115, + 0.019852396100759506, + 0.002704204758629203, + 0.002576688304543495, + 0.018161185085773468, + 0.012517403811216354, + 0.08622656017541885, + 0.014901732094585896, + 0.00936903990805149, + 0.022885702550411224, + 0.013280653394758701, + 0.11247486621141434, + 0.012211065739393234, + 0.011239065788686275, + 0.019191524013876915, + 0.010910111479461193, + 0.010290736332535744, + 0.07824378460645676, + 0.00902820099145174, + 0.08312942087650299, + 0.005239881109446287, + 0.002905622124671936, + 0.0032050388399511576, + 0.003831158857792616, + 0.003714440856128931, + 0.00389148504473269, + 0.002961518941447139, + 0.002917611040174961, + 0.03522636368870735, + 0.0144881010055542, + 0.019641947001218796, + 0.0238968376070261, + 0.021291421726346016, + 0.06122551113367081, + 0.020528201013803482, + 0.014685632660984993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019402741454541683 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019402741454541683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019402741454541683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061211906373500824, + 0.07688950002193451, + 0.019513262435793877, + 0.08535970747470856, + 0.01806303858757019, + 0.05001622810959816, + 0.18600238859653473, + 0.42902612686157227, + 0.03652846813201904, + 0.03738934546709061 + ], + "credit_map": [ + 0.014787916094064713, + 0.004074301570653915, + 0.003862748621031642, + 0.004172897431999445, + 0.004544309340417385, + 0.0033652218990027905, + 0.021646490320563316, + 0.004758017603307962, + 0.013152996078133583, + 0.005510241258889437, + 0.007602344732731581, + 0.005662075709551573, + 0.023555966094136238, + 0.008455201983451843, + 0.008003800176084042, + 0.004946868866682053, + 0.0014820285141468048, + 0.0013747179182246327, + 0.002365877153351903, + 0.0020019542425870895, + 0.002256012987345457, + 0.005000039003789425, + 0.0027320561930537224, + 0.002300575841218233, + 0.01991419680416584, + 0.004271931014955044, + 0.009796421974897385, + 0.006050906144082546, + 0.004328212235122919, + 0.023873379454016685, + 0.006072656251490116, + 0.01105200219899416, + 0.0015811967896297574, + 0.001919357106089592, + 0.0018936506239697337, + 0.0024185082875192165, + 0.0032136610243469477, + 0.0025736463721841574, + 0.003065187484025955, + 0.0013978307833895087, + 0.003029878018423915, + 0.0012324880808591843, + 0.01767251268029213, + 0.00237468839623034, + 0.0019308759365230799, + 0.020730936899781227, + 0.001400687382556498, + 0.0016441618790850043, + 0.01161443442106247, + 0.007934090681374073, + 0.056298401206731796, + 0.009137257002294064, + 0.0054823788814246655, + 0.014662738889455795, + 0.008688820526003838, + 0.07218427211046219, + 0.017005814239382744, + 0.01956757716834545, + 0.036671027541160583, + 0.018101288005709648, + 0.013793460093438625, + 0.15309886634349823, + 0.015078926458954811, + 0.1557091772556305, + 0.005790721159428358, + 0.0037587331607937813, + 0.004171327687799931, + 0.005206291098147631, + 0.005173915531486273, + 0.006278614979237318, + 0.003638348774984479, + 0.0025105150416493416, + 0.0029612979851663113, + 0.003717185463756323, + 0.004436524119228125, + 0.005931608844548464, + 0.004325435031205416, + 0.005992375314235687, + 0.005679271649569273, + 0.004345650784671307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.0019402741454541683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant pointillist style composed of tiny, distinct dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin on a forest floor with dark, damp soil and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the fungus depicted with a pitted, honeycombed cap and a thick stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fungus display colors of earthy beige, dark brown, and creamy white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the stem elongating upward while the cap unfurls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134215474128723 + ], + "opposite_logprob_tokens": [ + -1.3134214878082275 + ], + "logprob_reward": -0.3134215474128723, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134215474128723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058572907000780106, + 0.04430783912539482, + 0.03926407918334007, + 0.03032519854605198, + 0.017752645537257195, + 0.03628702834248543, + 0.2104567289352417, + 0.44069644808769226, + 0.01375540066510439, + 0.10858176648616791 + ], + "credit_map": [ + 0.00746794231235981, + 0.01273286435753107, + 0.007265524007380009, + 0.0061354730278253555, + 0.005283754784613848, + 0.005127578973770142, + 0.00641288235783577, + 0.008146886713802814, + 0.005203329958021641, + 0.004585164599120617, + 0.005402885843068361, + 0.004661337472498417, + 0.00485609145835042, + 0.007565829437226057, + 0.00798373855650425, + 0.004049461334943771, + 0.005073247477412224, + 0.0017544872825965285, + 0.005423595663160086, + 0.00212037842720747, + 0.0037258050870150328, + 0.016566086560487747, + 0.002308137249201536, + 0.0022923413198441267, + 0.0030285040847957134, + 0.0025225921999663115, + 0.0041217985562980175, + 0.004481013398617506, + 0.002557797124609351, + 0.003649825695902109, + 0.004704526159912348, + 0.005259142257273197, + 0.001381228445097804, + 0.0016165637644007802, + 0.0024812270421534777, + 0.0018731445306912065, + 0.002383651677519083, + 0.004024422261863947, + 0.002444067969918251, + 0.0015483404276892543, + 0.0032335701398551464, + 0.0014637243002653122, + 0.007185571826994419, + 0.0030090303625911474, + 0.0023512959014624357, + 0.012933303602039814, + 0.002701330231502652, + 0.003409202443435788, + 0.013303538784384727, + 0.008839301764965057, + 0.06411517411470413, + 0.010909562930464745, + 0.00621808273717761, + 0.017902469262480736, + 0.009950059466063976, + 0.07921852916479111, + 0.019523482769727707, + 0.02092391438782215, + 0.03648136556148529, + 0.01925639435648918, + 0.015335870906710625, + 0.1531996726989746, + 0.016255537047982216, + 0.15972021222114563, + 0.0017946353182196617, + 0.001411700970493257, + 0.0015677991323173046, + 0.0021999552845954895, + 0.00240110931918025, + 0.0017289469251409173, + 0.0013864367501810193, + 0.0012648169649764895, + 0.02374402806162834, + 0.00841842032968998, + 0.00997255276888609, + 0.014446907676756382, + 0.006229593884199858, + 0.03362482786178589, + 0.007462080102413893, + 0.004683355800807476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.3134215474128723 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_tensor([4564, 4564])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_tensor([4564, 4564])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..741ea9d483e7de0820226c5581238e7fe4b478f2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_tensor([4564, 4564])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly elevated angle, the camera remains still, focused on an antique brass oil lamp. The lamp rests on the worn surface of a dark walnut desk, its base covered in age-worn scratches and a faint patina. Within its smooth glass shade, a teardrop-shaped flame dances rhythmically and subtly, each pulse casting a warm golden glow. Each tremor of the flame sets the surrounding light and shadows in a gentle, hypnotic dance across the desk, the stack of old leather-bound books nearby, and the background wall, making the still objects seem to breathe. The entire scene, illuminated by this single source of warm light, creates an atmosphere of profound tranquility, concentration, and nostalgia. The deep, softly-edged shadows contrast sharply with the bright flame, giving the image the quality of a classical oil painting frozen in time.", + "seed": 22311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_4564_seed_22311_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06011208798736334, + "vlm_logprob": -0.06011208798736334, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.06011208798736334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015030699782073498 + ], + "opposite_logprob_tokens": [ + -4.640030860900879 + ], + "logprob_reward": -0.015030699782073498, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015030699782073498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23666906356811523, + 0.10594990104436874, + 0.0661381259560585, + 0.1367509812116623, + 0.05160734802484512, + 0.05896837264299393, + 0.056504808366298676, + 0.07735133916139603, + 0.10895837843418121, + 0.10110170394182205 + ], + "credit_map": [ + 0.025335898622870445, + 0.04628002643585205, + 0.02667546644806862, + 0.034754738211631775, + 0.0202216524630785, + 0.023859988898038864, + 0.030053116381168365, + 0.029488155618309975, + 0.008662735112011433, + 0.016835926100611687, + 0.01596970483660698, + 0.014021455310285091, + 0.01034222636371851, + 0.014903414994478226, + 0.012564376927912235, + 0.012650061398744583, + 0.006786591373383999, + 0.010960450395941734, + 0.008019662462174892, + 0.010937344282865524, + 0.005740757565945387, + 0.006635911297053099, + 0.009370344690978527, + 0.007687060162425041, + 0.010931795462965965, + 0.020212804898619652, + 0.014675638638436794, + 0.021680457517504692, + 0.012067332863807678, + 0.020577924326062202, + 0.018272463232278824, + 0.018332572653889656, + 0.005471712443977594, + 0.007998893037438393, + 0.006056496873497963, + 0.006038423161953688, + 0.004103069193661213, + 0.0098007433116436, + 0.005279116798192263, + 0.006858893670141697, + 0.005134922452270985, + 0.010774710215628147, + 0.008412566035985947, + 0.006662672385573387, + 0.0046928818337619305, + 0.011163115501403809, + 0.005011767148971558, + 0.007115737535059452, + 0.004461828153580427, + 0.007743953727185726, + 0.006096595898270607, + 0.009289740584790707, + 0.005285000894218683, + 0.005795038305222988, + 0.010058710351586342, + 0.007773938123136759, + 0.008202585391700268, + 0.009653960354626179, + 0.012385333888232708, + 0.009477006271481514, + 0.007895147427916527, + 0.007661515846848488, + 0.01325968001037836, + 0.008816109970211983, + 0.01180251780897379, + 0.0197425689548254, + 0.01693049632012844, + 0.010325278155505657, + 0.007488983683288097, + 0.016637656837701797, + 0.01464137900620699, + 0.01138949953019619, + 0.007960579358041286, + 0.016155080869793892, + 0.00937783531844616, + 0.011294595897197723, + 0.007880560122430325, + 0.02107308618724346, + 0.013105879537761211, + 0.014254087582230568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.015030699782073498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302050307393074 + ], + "opposite_logprob_tokens": [ + -3.7733020782470703 + ], + "logprob_reward": -0.023302050307393074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302050307393074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20896925032138824, + 0.2607164680957794, + 0.12466847151517868, + 0.03339296951889992, + 0.037600480020046234, + 0.09527695924043655, + 0.05426850542426109, + 0.10839374363422394, + 0.04417731240391731, + 0.03253582864999771 + ], + "credit_map": [ + 0.02101266384124756, + 0.035752613097429276, + 0.02416209876537323, + 0.03218892589211464, + 0.012298963032662868, + 0.024179644882678986, + 0.033757347613573074, + 0.025616997852921486, + 0.022942300885915756, + 0.04235953092575073, + 0.04356519132852554, + 0.052350252866744995, + 0.018895136192440987, + 0.021427925676107407, + 0.029255785048007965, + 0.029920347034931183, + 0.014749580062925816, + 0.023783111944794655, + 0.01674337312579155, + 0.0207784753292799, + 0.007926483638584614, + 0.014488629065454006, + 0.012448880821466446, + 0.013749931938946247, + 0.0034198411740362644, + 0.007000671233981848, + 0.004606449976563454, + 0.0036182613112032413, + 0.003018768271431327, + 0.0037991339340806007, + 0.004465694539248943, + 0.0034641509409993887, + 0.0037994086742401123, + 0.005740014836192131, + 0.0040900856256484985, + 0.0053888410329818726, + 0.002650914015248418, + 0.00738133117556572, + 0.003982986323535442, + 0.004566899035125971, + 0.009019880555570126, + 0.018263060599565506, + 0.01464687567204237, + 0.014246875420212746, + 0.006831625010818243, + 0.013685803860425949, + 0.007538706995546818, + 0.011044136248528957, + 0.0045752814039587975, + 0.008473418653011322, + 0.005980534479022026, + 0.008538169786334038, + 0.003879354102537036, + 0.007452280260622501, + 0.00788075476884842, + 0.007488708943128586, + 0.009424833580851555, + 0.01618102751672268, + 0.019321653991937637, + 0.02018210105597973, + 0.008078894577920437, + 0.009796803817152977, + 0.013067369349300861, + 0.012341061607003212, + 0.005082421470433474, + 0.00787676963955164, + 0.006623935420066118, + 0.004272682126611471, + 0.0033017327077686787, + 0.006587035488337278, + 0.005704581271857023, + 0.004728154744952917, + 0.003590803360566497, + 0.006109579466283321, + 0.004775650333613157, + 0.0051897610537707806, + 0.002850298536941409, + 0.0037212425377219915, + 0.0028078800532966852, + 0.00349061144515872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023302050307393074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059398591518402 + ], + "opposite_logprob_tokens": [ + -3.8955938816070557 + ], + "logprob_reward": -0.02059398591518402, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059398591518402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22442373633384705, + 0.1726524829864502, + 0.09083008021116257, + 0.07912615686655045, + 0.07369362562894821, + 0.09095193445682526, + 0.10695929080247879, + 0.05829207971692085, + 0.05754466727375984, + 0.04552590101957321 + ], + "credit_map": [ + 0.024494830518960953, + 0.03792122006416321, + 0.025729840621352196, + 0.037035539746284485, + 0.016284657642245293, + 0.021079767495393753, + 0.031643304973840714, + 0.030234578996896744, + 0.01712125726044178, + 0.026192231103777885, + 0.025578055530786514, + 0.032518304884433746, + 0.013025936670601368, + 0.02009783498942852, + 0.018374072387814522, + 0.01974480226635933, + 0.010507333092391491, + 0.018780384212732315, + 0.009625831618905067, + 0.011191817000508308, + 0.007840290665626526, + 0.012887921184301376, + 0.010192807763814926, + 0.009803696535527706, + 0.008627619594335556, + 0.016882313415408134, + 0.010023935697972775, + 0.008750672452151775, + 0.00718326959758997, + 0.013561036437749863, + 0.005772038828581572, + 0.008325272239744663, + 0.007461567409336567, + 0.012956706807017326, + 0.008340484462678432, + 0.009485002607107162, + 0.0062212091870605946, + 0.014194710180163383, + 0.006051658187061548, + 0.008982288651168346, + 0.011929567903280258, + 0.01801689900457859, + 0.011897251941263676, + 0.012331513687968254, + 0.007177657913416624, + 0.011812081560492516, + 0.00905650481581688, + 0.00873046088963747, + 0.012699450366199017, + 0.018367130309343338, + 0.012309336103498936, + 0.010284791700541973, + 0.00752456858754158, + 0.019748656079173088, + 0.011485663242638111, + 0.014539694413542747, + 0.00541682867333293, + 0.00801534578204155, + 0.0070855082012712955, + 0.008989573456346989, + 0.004042134620249271, + 0.00804703589528799, + 0.009084763005375862, + 0.00761089101433754, + 0.004322386812418699, + 0.007643362507224083, + 0.009982637129724026, + 0.007160946726799011, + 0.003683079732581973, + 0.0076368520967662334, + 0.0059671904891729355, + 0.01114821620285511, + 0.0036618548911064863, + 0.007668141275644302, + 0.006006937474012375, + 0.005547311622649431, + 0.0036266278475522995, + 0.005962833296507597, + 0.006306835450232029, + 0.006745356600731611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02059398591518402 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255602926015854 + ], + "opposite_logprob_tokens": [ + -1.6005603075027466 + ], + "logprob_reward": -0.2255602926015854, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255602926015854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2057497352361679, + 0.14087256789207458, + 0.1772354394197464, + 0.050494544208049774, + 0.044508371502161026, + 0.22060547769069672, + 0.02731984667479992, + 0.03796563297510147, + 0.013373224064707756, + 0.08187516778707504 + ], + "credit_map": [ + 0.023851940408349037, + 0.03954392671585083, + 0.026539919897913933, + 0.03635580465197563, + 0.012395457364618778, + 0.018039971590042114, + 0.021509559825062752, + 0.0275131668895483, + 0.018201570957899094, + 0.02861345000565052, + 0.01869843900203705, + 0.021025264635682106, + 0.011129366233944893, + 0.01563388668000698, + 0.015194195322692394, + 0.012376389466226101, + 0.02584136091172695, + 0.037197962403297424, + 0.022314025089144707, + 0.02720050700008869, + 0.010692862793803215, + 0.020938877016305923, + 0.014110132120549679, + 0.018939698114991188, + 0.004229115787893534, + 0.009073969908058643, + 0.005275047849863768, + 0.005419121589511633, + 0.004228913225233555, + 0.008477993309497833, + 0.006718156393617392, + 0.007072227541357279, + 0.004444110672920942, + 0.0059884279035031796, + 0.00585573585703969, + 0.010372933931648731, + 0.003425164381042123, + 0.004196873400360346, + 0.005508262664079666, + 0.004716862458735704, + 0.02575487270951271, + 0.04599129036068916, + 0.025743305683135986, + 0.02570745162665844, + 0.016262788325548172, + 0.03986535593867302, + 0.016740916296839714, + 0.024539491161704063, + 0.002562364097684622, + 0.004585618618875742, + 0.0027166716754436493, + 0.004117452073842287, + 0.0017826862167567015, + 0.005091172177344561, + 0.0030154280830174685, + 0.003448453964665532, + 0.003569694934412837, + 0.0069212354719638824, + 0.005728802643716335, + 0.005628164857625961, + 0.0035263136960566044, + 0.004699775483459234, + 0.0039096008986234665, + 0.003982042893767357, + 0.001103146467357874, + 0.0024822833947837353, + 0.0016196246724575758, + 0.0019076615571975708, + 0.0014712691772729158, + 0.001316616777330637, + 0.001485094428062439, + 0.001987527357414365, + 0.00888828281313181, + 0.015440460294485092, + 0.010268257930874825, + 0.012146134860813618, + 0.0073675126768648624, + 0.00993852224200964, + 0.009638868272304535, + 0.008187126368284225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.2255602926015854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607341133058071 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607341133058071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607341133058071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643151581287384, + 0.14969627559185028, + 0.05460207909345627, + 0.09801824390888214, + 0.15252560377120972, + 0.045199356973171234, + 0.06074414774775505, + 0.053258076310157776, + 0.050857141613960266, + 0.07078396528959274 + ], + "credit_map": [ + 0.027943933382630348, + 0.046388667076826096, + 0.028418634086847305, + 0.036656077951192856, + 0.0220484621822834, + 0.028193654492497444, + 0.04306617006659508, + 0.03159955143928528, + 0.01239700336009264, + 0.020618891343474388, + 0.014274531044065952, + 0.02579840086400509, + 0.012758047319948673, + 0.02304050326347351, + 0.018269017338752747, + 0.022539883852005005, + 0.005777785554528236, + 0.008304858580231667, + 0.006222449243068695, + 0.006200624164193869, + 0.008220303803682327, + 0.0053848037496209145, + 0.008838287554681301, + 0.005652967374771833, + 0.007051660213619471, + 0.017078228294849396, + 0.011930464766919613, + 0.011048409156501293, + 0.009690782055258751, + 0.018751060590147972, + 0.010103385895490646, + 0.012364254333078861, + 0.011888692155480385, + 0.020921576768159866, + 0.021286997944116592, + 0.042363833636045456, + 0.010287485085427761, + 0.01237806212157011, + 0.018306175246834755, + 0.01509278453886509, + 0.004124398808926344, + 0.008483284153044224, + 0.005830953363329172, + 0.010283684358000755, + 0.0028162484522908926, + 0.004825276788324118, + 0.004171627573668957, + 0.004663882311433554, + 0.006354714278131723, + 0.008375178091228008, + 0.007656115107238293, + 0.007189547643065453, + 0.005142257083207369, + 0.006819921545684338, + 0.011374475434422493, + 0.007831933908164501, + 0.005744371563196182, + 0.00759775098413229, + 0.005917666479945183, + 0.009394685737788677, + 0.003728581825271249, + 0.007635194342583418, + 0.0071489023976027966, + 0.0060909204185009, + 0.005084211938083172, + 0.005923041142523289, + 0.00655306177213788, + 0.005976085085421801, + 0.0038105174899101257, + 0.0075810132548213005, + 0.007327037863433361, + 0.008602170273661613, + 0.008080817759037018, + 0.00934553425759077, + 0.007902017794549465, + 0.011411572806537151, + 0.007410295773297548, + 0.006657267455011606, + 0.00946358498185873, + 0.01051287166774273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01607341133058071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly elevated angle, the camera remains still, focused on an antique brass oil lamp. The lamp rests on the worn surface of a dark walnut desk, its base covered in age-worn scratches and a faint patina. Within its smooth glass shade, a teardrop-shaped flame dances rhythmically and subtly, each pulse casting a warm golden glow. Each tremor of the flame sets the surrounding light and shadows in a gentle, hypnotic dance across the desk, the stack of old leather-bound books nearby, and the background wall, making the still objects seem to breathe. The entire scene, illuminated by this single source of warm light, creates an atmosphere of profound tranquility, concentration, and nostalgia. The deep, softly-edged shadows contrast sharply with the bright flame, giving the image the quality of a classical oil painting frozen in time.", + "seed": 22312, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_4564_seed_22312_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04165285304188728, + "vlm_logprob": -0.04165285304188728, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04165285304188728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016269885003566742 + ], + "opposite_logprob_tokens": [ + -4.891269683837891 + ], + "logprob_reward": -0.016269885003566742, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.016269885003566742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2624664306640625, + 0.12767541408538818, + 0.1159399151802063, + 0.0884755402803421, + 0.0740228220820427, + 0.05083948001265526, + 0.07210945338010788, + 0.13543178141117096, + 0.05035457760095596, + 0.022684575989842415 + ], + "credit_map": [ + 0.03434173762798309, + 0.022973865270614624, + 0.06333500891923904, + 0.026197176426649094, + 0.021569665521383286, + 0.040064841508865356, + 0.0177445150911808, + 0.036239635199308395, + 0.017937136813998222, + 0.010046777315437794, + 0.030415872111916542, + 0.011752072721719742, + 0.00925967562943697, + 0.01921464316546917, + 0.00800009723752737, + 0.021049130707979202, + 0.01188589632511139, + 0.012205432169139385, + 0.021574843674898148, + 0.01491326093673706, + 0.011942355893552303, + 0.015174037776887417, + 0.00976648461073637, + 0.018477607518434525, + 0.007218428421765566, + 0.014155600219964981, + 0.012220649048686028, + 0.015426971949636936, + 0.008388056419789791, + 0.009813779033720493, + 0.009820561856031418, + 0.011431489139795303, + 0.005087910685688257, + 0.012082765810191631, + 0.00929164607077837, + 0.011925607919692993, + 0.009545010514557362, + 0.007962117902934551, + 0.008520888164639473, + 0.009606877341866493, + 0.0054937489330768585, + 0.005811732728034258, + 0.009448195807635784, + 0.006869536824524403, + 0.005299690179526806, + 0.006017944775521755, + 0.004960609599947929, + 0.00693802023306489, + 0.009477644227445126, + 0.0070161293260753155, + 0.015534396283328533, + 0.006593870930373669, + 0.00802659336477518, + 0.01098072063177824, + 0.004892969038337469, + 0.009587128646671772, + 0.018665893003344536, + 0.014455911703407764, + 0.02304726466536522, + 0.012913784943521023, + 0.019403448328375816, + 0.017940906807780266, + 0.011458761990070343, + 0.017545796930789948, + 0.005255203694105148, + 0.008456823416054249, + 0.006715369876474142, + 0.007492346223443747, + 0.004175164271146059, + 0.005910505075007677, + 0.005563369952142239, + 0.006785794626921415, + 0.0031079731415957212, + 0.0023361286148428917, + 0.00436055613681674, + 0.0030054175294935703, + 0.0017809023847803473, + 0.002499058609828353, + 0.002275101374834776, + 0.003319439012557268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.016269885003566742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036305338144302 + ], + "opposite_logprob_tokens": [ + -2.35036301612854 + ], + "logprob_reward": -0.10036305338144302, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036305338144302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39698296785354614, + 0.22174076735973358, + 0.12011729180812836, + 0.0245269313454628, + 0.01754513755440712, + 0.026657709851861, + 0.061991892755031586, + 0.0889359638094902, + 0.02375195547938347, + 0.017749404534697533 + ], + "credit_map": [ + 0.05893411487340927, + 0.02612069994211197, + 0.10841359943151474, + 0.036463797092437744, + 0.02144370786845684, + 0.05841517075896263, + 0.020361263304948807, + 0.0668305978178978, + 0.03205735608935356, + 0.018566813319921494, + 0.057338397949934006, + 0.02435614913702011, + 0.011637664400041103, + 0.03076709620654583, + 0.012080654501914978, + 0.03493664413690567, + 0.015687065199017525, + 0.010492666624486446, + 0.02850378304719925, + 0.013704490847885609, + 0.00913128349930048, + 0.01688656210899353, + 0.007319798227399588, + 0.018391646444797516, + 0.003181933192536235, + 0.002747661666944623, + 0.004811631515622139, + 0.0032134864013642073, + 0.0018549262313172221, + 0.0027359211817383766, + 0.0020476120989769697, + 0.003933758940547705, + 0.0032513788901269436, + 0.0017197397537529469, + 0.0032138258684426546, + 0.002725596074014902, + 0.0011905295541509986, + 0.002000333508476615, + 0.0014079372631385922, + 0.002035796409472823, + 0.003101987298578024, + 0.0028779562562704086, + 0.005236239638179541, + 0.0034577937331050634, + 0.003179732942953706, + 0.0029295848216861486, + 0.0021174708381295204, + 0.0037569445557892323, + 0.011994128115475178, + 0.005392232909798622, + 0.014918332919478416, + 0.005459824576973915, + 0.003995809238404036, + 0.008524361066520214, + 0.003385456744581461, + 0.008321745321154594, + 0.016727199777960777, + 0.007992762140929699, + 0.017842842265963554, + 0.00763721251860261, + 0.008640052750706673, + 0.012104667723178864, + 0.006049922667443752, + 0.011941301636397839, + 0.004380014259368181, + 0.002734170528128743, + 0.004213979002088308, + 0.0029515742789953947, + 0.0020121841225773096, + 0.0023778665345162153, + 0.0018896324327215552, + 0.0031925335060805082, + 0.001969766104593873, + 0.002529809018597007, + 0.0030702867079526186, + 0.002986367093399167, + 0.001546197570860386, + 0.0017100381664931774, + 0.0018122670007869601, + 0.002124673454090953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.10036305338144302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084088012576103 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.016084088012576103, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084088012576103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28142350912094116, + 0.1228204220533371, + 0.09097655862569809, + 0.056653670966625214, + 0.11135195195674896, + 0.0969516709446907, + 0.07193966954946518, + 0.09993112832307816, + 0.046784672886133194, + 0.021166743710637093 + ], + "credit_map": [ + 0.0432431623339653, + 0.025127150118350983, + 0.06736472249031067, + 0.030997177585959435, + 0.016940776258707047, + 0.037005819380283356, + 0.015952041372656822, + 0.04479266703128815, + 0.01988384500145912, + 0.012359959073364735, + 0.029076261445879936, + 0.012311290949583054, + 0.007866304367780685, + 0.017437592148780823, + 0.006569701712578535, + 0.017315465956926346, + 0.009951108135282993, + 0.008791771717369556, + 0.01799783669412136, + 0.01415462326258421, + 0.007088472135365009, + 0.01250461209565401, + 0.00671235378831625, + 0.013775781728327274, + 0.004862836096435785, + 0.00773515272885561, + 0.010406740941107273, + 0.007142944261431694, + 0.006250966340303421, + 0.006691916845738888, + 0.005767819005995989, + 0.007795294281095266, + 0.010759648866951466, + 0.011710540391504765, + 0.016232380643486977, + 0.02112036943435669, + 0.013549614697694778, + 0.010964911431074142, + 0.011578882113099098, + 0.015435602515935898, + 0.009550128132104874, + 0.012394516728818417, + 0.014263410121202469, + 0.018440403044223785, + 0.00805267971009016, + 0.011013285256922245, + 0.011557009071111679, + 0.011680243536829948, + 0.010621930472552776, + 0.007701160851866007, + 0.014139117673039436, + 0.007402010727673769, + 0.007665296085178852, + 0.00838309433311224, + 0.00497302645817399, + 0.011054040864109993, + 0.013696056790649891, + 0.012028581462800503, + 0.020657513290643692, + 0.009081018157303333, + 0.010255909524857998, + 0.010775413364171982, + 0.00831962563097477, + 0.015117010101675987, + 0.008508934639394283, + 0.003613021457567811, + 0.010323203168809414, + 0.006046586204320192, + 0.0037166771944612265, + 0.00465285312384367, + 0.00406680116429925, + 0.005856591742485762, + 0.0023356799501925707, + 0.002107934094965458, + 0.004522675648331642, + 0.0030816125217825174, + 0.0018228652188554406, + 0.0023288552183657885, + 0.001891280640847981, + 0.003075840650126338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.016084088012576103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495745688676834 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495745688676834, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495745688676834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28597018122673035, + 0.19324833154678345, + 0.13391219079494476, + 0.05792029947042465, + 0.05067100748419762, + 0.09608765691518784, + 0.028656532987952232, + 0.0964910089969635, + 0.022661272436380386, + 0.03438156470656395 + ], + "credit_map": [ + 0.04404199868440628, + 0.02442771941423416, + 0.06782063096761703, + 0.02648005448281765, + 0.02120702527463436, + 0.04116511717438698, + 0.015833018347620964, + 0.044994574040174484, + 0.030132843181490898, + 0.015880336984992027, + 0.04879762977361679, + 0.01953776925802231, + 0.012321589514613152, + 0.023999840021133423, + 0.01160728931427002, + 0.030971014872193336, + 0.02263743430376053, + 0.009838848374783993, + 0.034451816231012344, + 0.00956797506660223, + 0.007117868401110172, + 0.01913168653845787, + 0.007817546837031841, + 0.023348992690443993, + 0.008359448984265327, + 0.0055402955040335655, + 0.010459329932928085, + 0.008338204585015774, + 0.004826795309782028, + 0.0069960374385118484, + 0.004718809854239225, + 0.008681370876729488, + 0.008516077883541584, + 0.004953073337674141, + 0.00857927743345499, + 0.007826033048331738, + 0.0049218954518437386, + 0.005459567531943321, + 0.003987055271863937, + 0.00642802007496357, + 0.013154327869415283, + 0.009995890781283379, + 0.018444575369358063, + 0.013181712478399277, + 0.009909402579069138, + 0.009995538741350174, + 0.007521134335547686, + 0.013885068707168102, + 0.003651822218671441, + 0.0022921767085790634, + 0.006131378468126059, + 0.0038557685911655426, + 0.00259873503819108, + 0.003218205878511071, + 0.0022436657454818487, + 0.004664777312427759, + 0.010238363407552242, + 0.008723894134163857, + 0.020016618072986603, + 0.013351327739655972, + 0.007913688197731972, + 0.011880489997565746, + 0.010461807250976562, + 0.013904809020459652, + 0.004016013350337744, + 0.0025266793090850115, + 0.004232348408550024, + 0.002371704438701272, + 0.002045374596491456, + 0.0025651389732956886, + 0.0016010210383683443, + 0.0033029895275831223, + 0.005654685664921999, + 0.002453753724694252, + 0.008788720704615116, + 0.003058559261262417, + 0.0019097771728411317, + 0.004929375369101763, + 0.0019243621500208974, + 0.005662328097969294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05495745688676834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589781925082207 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.020589781925082207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589781925082207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1771794855594635, + 0.11206074059009552, + 0.06811373680830002, + 0.0993862971663475, + 0.1152038425207138, + 0.09588167816400528, + 0.06777391582727432, + 0.19129802286624908, + 0.04196006804704666, + 0.031142225489020348 + ], + "credit_map": [ + 0.021442413330078125, + 0.016016950830817223, + 0.04062261804938316, + 0.023748505860567093, + 0.014796208590269089, + 0.02202918753027916, + 0.013647057116031647, + 0.024876540526747704, + 0.019796928390860558, + 0.01033759769052267, + 0.021097708493471146, + 0.01462989579886198, + 0.006948061287403107, + 0.01454057265073061, + 0.00898237619549036, + 0.01572759635746479, + 0.0130536500364542, + 0.006065262947231531, + 0.014238073490560055, + 0.006460446398705244, + 0.004444973077625036, + 0.009417569264769554, + 0.004932377487421036, + 0.009501384571194649, + 0.014914841391146183, + 0.012590639293193817, + 0.017997698858380318, + 0.012326890602707863, + 0.010764416307210922, + 0.009967584162950516, + 0.006760779768228531, + 0.014063446782529354, + 0.008403396233916283, + 0.013779946602880955, + 0.01703079603612423, + 0.022499915212392807, + 0.012269321829080582, + 0.012860325165092945, + 0.0132267652079463, + 0.015133378095924854, + 0.01704627461731434, + 0.010244426317512989, + 0.018992679193615913, + 0.013350831344723701, + 0.007003923412412405, + 0.010334060527384281, + 0.00783352367579937, + 0.011075963266193867, + 0.011505290865898132, + 0.004807246383279562, + 0.01632293127477169, + 0.0064126052893698215, + 0.004471765831112862, + 0.009626542218029499, + 0.004377470817416906, + 0.010250063613057137, + 0.03636088967323303, + 0.016552088782191277, + 0.05268152058124542, + 0.015612500719726086, + 0.009805179201066494, + 0.020730484277009964, + 0.015327355824410915, + 0.024227993562817574, + 0.0047934576869010925, + 0.0035583856515586376, + 0.007701219525188208, + 0.006352892145514488, + 0.0031644327100366354, + 0.004768495447933674, + 0.0039105843752622604, + 0.007710601203143597, + 0.005148542579263449, + 0.0024802491534501314, + 0.0065267039462924, + 0.003944938071072102, + 0.0019494271837174892, + 0.004150804132223129, + 0.002459133043885231, + 0.0044824290089309216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.020589781925082207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly elevated angle, the camera remains still, focused on an antique brass oil lamp. The lamp rests on the worn surface of a dark walnut desk, its base covered in age-worn scratches and a faint patina. Within its smooth glass shade, a teardrop-shaped flame dances rhythmically and subtly, each pulse casting a warm golden glow. Each tremor of the flame sets the surrounding light and shadows in a gentle, hypnotic dance across the desk, the stack of old leather-bound books nearby, and the background wall, making the still objects seem to breathe. The entire scene, illuminated by this single source of warm light, creates an atmosphere of profound tranquility, concentration, and nostalgia. The deep, softly-edged shadows contrast sharply with the bright flame, giving the image the quality of a classical oil painting frozen in time.", + "seed": 22313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_4564_seed_22313_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022255843691527843, + "vlm_logprob": -0.022255843691527843, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.022255843691527843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012270909734070301 + ], + "opposite_logprob_tokens": [ + -4.887270927429199 + ], + "logprob_reward": -0.012270909734070301, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012270909734070301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24157889187335968, + 0.10748991370201111, + 0.04329454526305199, + 0.12908512353897095, + 0.09555938094854355, + 0.04640856385231018, + 0.05999176949262619, + 0.10555379837751389, + 0.1381302922964096, + 0.03290768340229988 + ], + "credit_map": [ + 0.021813232451677322, + 0.020982181653380394, + 0.0299212709069252, + 0.0356220044195652, + 0.031035078689455986, + 0.04173877090215683, + 0.03723631054162979, + 0.023230046033859253, + 0.011697654612362385, + 0.010924491100013256, + 0.01733408495783806, + 0.013796844519674778, + 0.009122484363615513, + 0.01848701760172844, + 0.012495980598032475, + 0.013631361536681652, + 0.0039273458532989025, + 0.003804645035415888, + 0.00690458482131362, + 0.005241848062723875, + 0.004265842493623495, + 0.0070462701842188835, + 0.0052895816043019295, + 0.006814425811171532, + 0.01240930799394846, + 0.013831757009029388, + 0.023789789527654648, + 0.018657458946108818, + 0.012974259443581104, + 0.013764875009655952, + 0.017153266817331314, + 0.01650441624224186, + 0.008231083862483501, + 0.011604160070419312, + 0.011668434366583824, + 0.015316649340093136, + 0.012878241017460823, + 0.010923516936600208, + 0.014745968393981457, + 0.010191325098276138, + 0.0037738406099379063, + 0.006526952609419823, + 0.004885824862867594, + 0.00675026373937726, + 0.005422432441264391, + 0.007570577319711447, + 0.005139270797371864, + 0.006339403335005045, + 0.00722671439871192, + 0.007804426830261946, + 0.011191782541573048, + 0.006381564773619175, + 0.003344430588185787, + 0.010786185972392559, + 0.004944727756083012, + 0.008311932906508446, + 0.008206226862967014, + 0.013138235546648502, + 0.015564528293907642, + 0.020955931395292282, + 0.011246956884860992, + 0.011581717990338802, + 0.015601573511958122, + 0.009258626960217953, + 0.010688453912734985, + 0.019448135048151016, + 0.016403550282120705, + 0.0238785557448864, + 0.017072021961212158, + 0.015474607236683369, + 0.022152438759803772, + 0.013012530282139778, + 0.0032329184468835592, + 0.0038625856395810843, + 0.006879058200865984, + 0.004019930027425289, + 0.002675887430086732, + 0.003929021768271923, + 0.003173109143972397, + 0.005135175306349993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.012270909734070301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608901470899582 + ], + "opposite_logprob_tokens": [ + -4.141088962554932 + ], + "logprob_reward": -0.01608901470899582, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608901470899582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17528022825717926, + 0.3479602634906769, + 0.1796124279499054, + 0.030681438744068146, + 0.03993353620171547, + 0.102618508040905, + 0.021369876340031624, + 0.02444172278046608, + 0.016534848138689995, + 0.06156719848513603 + ], + "credit_map": [ + 0.015867967158555984, + 0.01907055638730526, + 0.024533767253160477, + 0.028309809044003487, + 0.018138552084565163, + 0.027871068567037582, + 0.02416173927485943, + 0.01732676848769188, + 0.04091483727097511, + 0.02836032025516033, + 0.06870853155851364, + 0.037262193858623505, + 0.02521715685725212, + 0.06841389834880829, + 0.030933862552046776, + 0.0481494702398777, + 0.018185174092650414, + 0.017726074904203415, + 0.03472704067826271, + 0.020445287227630615, + 0.01123527530580759, + 0.03982364386320114, + 0.013627811335027218, + 0.023842131718993187, + 0.003311439882963896, + 0.004256741143763065, + 0.005931731779128313, + 0.005179333500564098, + 0.001966359093785286, + 0.0035277281422168016, + 0.0027360604144632816, + 0.003772046882659197, + 0.004723771940916777, + 0.004601364955306053, + 0.00696609215810895, + 0.005429529119282961, + 0.0031974222511053085, + 0.006023375783115625, + 0.003769462928175926, + 0.005222515668720007, + 0.010015235282480717, + 0.010446932166814804, + 0.014910782687366009, + 0.008840490132570267, + 0.006530395243316889, + 0.02217617817223072, + 0.007606919389218092, + 0.0220915786921978, + 0.00220211292617023, + 0.002878679661080241, + 0.004871248267591, + 0.0030795771162956953, + 0.0014427527785301208, + 0.002263870323076844, + 0.002077639102935791, + 0.0025539950001984835, + 0.00401250459253788, + 0.002632290590554476, + 0.003932944033294916, + 0.0030910109635442495, + 0.002398183336481452, + 0.002689966233447194, + 0.0025521395727992058, + 0.003132682293653488, + 0.002032060641795397, + 0.001816021860577166, + 0.0035698902793228626, + 0.0022779761347919703, + 0.0011606354964897037, + 0.0020454395562410355, + 0.0016906112432479858, + 0.0019422126933932304, + 0.005704924464225769, + 0.005229853559285402, + 0.008942830376327038, + 0.006296513136476278, + 0.004060124047100544, + 0.017485177144408226, + 0.004836782347410917, + 0.00901099294424057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.01608901470899582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197137862443924 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197137862443924, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197137862443924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14243890345096588, + 0.0961342379450798, + 0.08704416453838348, + 0.13214781880378723, + 0.15493179857730865, + 0.2111620008945465, + 0.05656525120139122, + 0.04679211601614952, + 0.0382259339094162, + 0.034557800740003586 + ], + "credit_map": [ + 0.014054455794394016, + 0.01597418636083603, + 0.022516321390867233, + 0.016846008598804474, + 0.011182180605828762, + 0.019900623708963394, + 0.01925124041736126, + 0.022713877260684967, + 0.010309788398444653, + 0.008200587704777718, + 0.013355737552046776, + 0.00973587203770876, + 0.008206844329833984, + 0.022316472604870796, + 0.011018108576536179, + 0.012990816496312618, + 0.006260467693209648, + 0.006636698264628649, + 0.00847628340125084, + 0.009513645432889462, + 0.009990998543798923, + 0.024734243750572205, + 0.01019764319062233, + 0.011234176345169544, + 0.011553109623491764, + 0.015187905170023441, + 0.02202669344842434, + 0.022326087579131126, + 0.012827293016016483, + 0.013241025619208813, + 0.01870136708021164, + 0.01628432422876358, + 0.01512694451957941, + 0.015913615003228188, + 0.024415163323283195, + 0.021692050620913506, + 0.013496327213943005, + 0.026549141854047775, + 0.019539635628461838, + 0.018198899924755096, + 0.021577391773462296, + 0.021366842091083527, + 0.03162664920091629, + 0.03548111021518707, + 0.020769011229276657, + 0.029207497835159302, + 0.02650182694196701, + 0.024631641805171967, + 0.0056582181714475155, + 0.007459996733814478, + 0.012015373446047306, + 0.007824914529919624, + 0.00414633983746171, + 0.006235227920114994, + 0.005871897097676992, + 0.007353275083005428, + 0.004326995927840471, + 0.005120264366269112, + 0.006003229878842831, + 0.00854121707379818, + 0.00485380319878459, + 0.004784927237778902, + 0.007176897954195738, + 0.0059847766533494, + 0.0034602086525410414, + 0.003930103033781052, + 0.006226631347090006, + 0.006475225556641817, + 0.002835894003510475, + 0.006677843164652586, + 0.004232317674905062, + 0.0043877046555280685, + 0.0029455379117280245, + 0.0036671743728220463, + 0.0036168117076158524, + 0.00506934616714716, + 0.0027165815699845552, + 0.009013710543513298, + 0.0034574675373733044, + 0.004071167204529047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.018197137862443924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864569380879402 + ], + "opposite_logprob_tokens": [ + -3.0486457347869873 + ], + "logprob_reward": -0.04864569380879402, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864569380879402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2758984863758087, + 0.06661369651556015, + 0.10549158602952957, + 0.04374249279499054, + 0.31437334418296814, + 0.047896865755319595, + 0.048996396362781525, + 0.04983256757259369, + 0.02785777486860752, + 0.019296718761324883 + ], + "credit_map": [ + 0.027997532859444618, + 0.030042991042137146, + 0.06425449252128601, + 0.0334026962518692, + 0.016038792207837105, + 0.03743097931146622, + 0.026196729391813278, + 0.040534306317567825, + 0.00663694366812706, + 0.006792315747588873, + 0.011050520464777946, + 0.01119625847786665, + 0.006460958160459995, + 0.006696869153529406, + 0.008111064322292805, + 0.00966876745223999, + 0.012054530903697014, + 0.013193817809224129, + 0.019744446501135826, + 0.011603891849517822, + 0.006426817271858454, + 0.01995740830898285, + 0.008977018296718597, + 0.01353366207331419, + 0.0044923326931893826, + 0.0041557298973202705, + 0.008574186824262142, + 0.0067147319205105305, + 0.0034141610376536846, + 0.006333441939204931, + 0.005526435095816851, + 0.004531477577984333, + 0.036025386303663254, + 0.04598082974553108, + 0.07234280556440353, + 0.04321013763546944, + 0.01663493551313877, + 0.03692079707980156, + 0.021271442994475365, + 0.04198702424764633, + 0.005365626886487007, + 0.00599232641980052, + 0.009070279076695442, + 0.005492503754794598, + 0.0029941953253000975, + 0.007879659533500671, + 0.0038059731014072895, + 0.007296302355825901, + 0.004940062295645475, + 0.007227630820125341, + 0.011351141147315502, + 0.005576286930590868, + 0.0024479064159095287, + 0.006287053227424622, + 0.0044304677285254, + 0.006735851522535086, + 0.0050184703432023525, + 0.006303734611719847, + 0.012040672823786736, + 0.008124795742332935, + 0.003534620627760887, + 0.0049590980634093285, + 0.004239373840391636, + 0.005611802451312542, + 0.0027785110287368298, + 0.00381208723410964, + 0.00651853671297431, + 0.0034491312690079212, + 0.0015650899149477482, + 0.003838052274659276, + 0.002216526074334979, + 0.0036798438522964716, + 0.0018386710435152054, + 0.0028966395184397697, + 0.0043333531357347965, + 0.0025035396683961153, + 0.0011836912017315626, + 0.0020034918561577797, + 0.0018640378257259727, + 0.002673295559361577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.04864569380879402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607646234333515 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.01607646234333515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607646234333515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21799932420253754, + 0.16967077553272247, + 0.05125223100185394, + 0.05687127262353897, + 0.04125392809510231, + 0.09236859530210495, + 0.12557224929332733, + 0.10437553375959396, + 0.09493257105350494, + 0.045703478157520294 + ], + "credit_map": [ + 0.02102206088602543, + 0.02393331564962864, + 0.039387162774801254, + 0.0292370468378067, + 0.017538726329803467, + 0.026374150067567825, + 0.03198283910751343, + 0.0285240076482296, + 0.018995201215147972, + 0.014563463628292084, + 0.035083431750535965, + 0.02244226634502411, + 0.009818337857723236, + 0.0248271431773901, + 0.017411373555660248, + 0.02652956172823906, + 0.0050065466202795506, + 0.004839131608605385, + 0.006664840504527092, + 0.004806223325431347, + 0.005305970087647438, + 0.010332969017326832, + 0.006077212747186422, + 0.008219337090849876, + 0.00511933071538806, + 0.008071289397776127, + 0.012247766368091106, + 0.007867332547903061, + 0.004554114770144224, + 0.007016659714281559, + 0.0059657166711986065, + 0.006029063370078802, + 0.003484138520434499, + 0.005206862464547157, + 0.007349213119596243, + 0.00588194839656353, + 0.004564350936561823, + 0.004411458037793636, + 0.005657501518726349, + 0.004698454402387142, + 0.006209141109138727, + 0.006106322631239891, + 0.008763072080910206, + 0.010190834291279316, + 0.007172352634370327, + 0.027739807963371277, + 0.008678920567035675, + 0.01750815473496914, + 0.012601091526448727, + 0.020436005666851997, + 0.024594081565737724, + 0.01905527524650097, + 0.009958658367395401, + 0.014185823500156403, + 0.011703766882419586, + 0.013037549331784248, + 0.012170459143817425, + 0.011055142618715763, + 0.015073128044605255, + 0.016271362081170082, + 0.010236119851469994, + 0.01143384724855423, + 0.01652582362294197, + 0.011609647423028946, + 0.010056701488792896, + 0.011678364127874374, + 0.014570062980055809, + 0.014957156032323837, + 0.009692410007119179, + 0.012060336768627167, + 0.011425046250224113, + 0.010492493398487568, + 0.004475310444831848, + 0.005503518506884575, + 0.007552327122539282, + 0.0063982619903981686, + 0.004339846316725016, + 0.005712063983082771, + 0.005379523150622845, + 0.00634262477979064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01607646234333515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly elevated angle, the camera remains still, focused on an antique brass oil lamp. The lamp rests on the worn surface of a dark walnut desk, its base covered in age-worn scratches and a faint patina. Within its smooth glass shade, a teardrop-shaped flame dances rhythmically and subtly, each pulse casting a warm golden glow. Each tremor of the flame sets the surrounding light and shadows in a gentle, hypnotic dance across the desk, the stack of old leather-bound books nearby, and the background wall, making the still objects seem to breathe. The entire scene, illuminated by this single source of warm light, creates an atmosphere of profound tranquility, concentration, and nostalgia. The deep, softly-edged shadows contrast sharply with the bright flame, giving the image the quality of a classical oil painting frozen in time.", + "seed": 22314, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_10/index_4564_seed_22314_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2810067843645811, + "vlm_logprob": -0.2810067843645811, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.2810067843645811, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018434155732393265 + ], + "opposite_logprob_tokens": [ + -4.643434047698975 + ], + "logprob_reward": -0.018434155732393265, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.018434155732393265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23941007256507874, + 0.057139378041028976, + 0.09513726830482483, + 0.09990546852350235, + 0.041567184031009674, + 0.07102952897548676, + 0.15175634622573853, + 0.061306338757276535, + 0.1419849544763565, + 0.040763456374406815 + ], + "credit_map": [ + 0.029939569532871246, + 0.034202586859464645, + 0.01583557389676571, + 0.025891002267599106, + 0.03984565660357475, + 0.03317704051733017, + 0.030085835605859756, + 0.030432822182774544, + 0.004197266418486834, + 0.007517982739955187, + 0.007131634280085564, + 0.007629267405718565, + 0.007644007448107004, + 0.007876048795878887, + 0.007524444255977869, + 0.0076187290251255035, + 0.011433911509811878, + 0.009669452905654907, + 0.005942953750491142, + 0.009647724218666553, + 0.01815144158899784, + 0.012265408411622047, + 0.015258170664310455, + 0.012768203392624855, + 0.011668750084936619, + 0.0074612475000321865, + 0.00770939514040947, + 0.012528444640338421, + 0.012045762501657009, + 0.009642856195569038, + 0.024948472157120705, + 0.013900548219680786, + 0.0029322027694433928, + 0.004543457645922899, + 0.0025862143374979496, + 0.005140749737620354, + 0.00859011709690094, + 0.006287590134888887, + 0.00521886209025979, + 0.006267991382628679, + 0.007638211362063885, + 0.00850458350032568, + 0.005422838032245636, + 0.006616189144551754, + 0.014787652529776096, + 0.009808853268623352, + 0.008684647269546986, + 0.009566552937030792, + 0.011657160706818104, + 0.016776565462350845, + 0.007927041500806808, + 0.013049941509962082, + 0.04168514534831047, + 0.016899408772587776, + 0.023350229486823082, + 0.02041085995733738, + 0.005540817975997925, + 0.004998325370252132, + 0.004434674512594938, + 0.006370671559125185, + 0.011187740601599216, + 0.010991610586643219, + 0.008246046490967274, + 0.00953645259141922, + 0.01536522340029478, + 0.016283318400382996, + 0.01273024920374155, + 0.016310282051563263, + 0.025899142026901245, + 0.01647871732711792, + 0.022068431600928307, + 0.01684960164129734, + 0.003777014324441552, + 0.0038669188506901264, + 0.003380854381248355, + 0.004093946889042854, + 0.007235412485897541, + 0.004604675807058811, + 0.00829879380762577, + 0.005505839362740517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.018434155732393265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487326204776764 + ], + "opposite_logprob_tokens": [ + -1.223732590675354 + ], + "logprob_reward": -0.3487326204776764, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487326204776764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23704957962036133, + 0.10281393676996231, + 0.13582633435726166, + 0.07874123752117157, + 0.04527021944522858, + 0.04122759774327278, + 0.239414244890213, + 0.06445574015378952, + 0.02003060281276703, + 0.03517051413655281 + ], + "credit_map": [ + 0.01611003279685974, + 0.021122809499502182, + 0.022175822407007217, + 0.030871035531163216, + 0.02683919481933117, + 0.026255972683429718, + 0.06044342368841171, + 0.03323127701878548, + 0.008155707269906998, + 0.010088665410876274, + 0.013294698670506477, + 0.01047293096780777, + 0.015960710123181343, + 0.013262344524264336, + 0.0184317659586668, + 0.013147116638720036, + 0.016830725595355034, + 0.014436495490372181, + 0.008706573396921158, + 0.012639756314456463, + 0.022938139736652374, + 0.0156661719083786, + 0.0262970682233572, + 0.0183113981038332, + 0.012297597713768482, + 0.0073510571382939816, + 0.006782212760299444, + 0.0076215569861233234, + 0.010589894838631153, + 0.009899591095745564, + 0.012787356041371822, + 0.011411971412599087, + 0.005736031569540501, + 0.00455540232360363, + 0.004566001705825329, + 0.0043783606961369514, + 0.006740345153957605, + 0.004538917914032936, + 0.008463497273623943, + 0.006291662342846394, + 0.003836119081825018, + 0.005349092651158571, + 0.004039532970637083, + 0.005447246599942446, + 0.007176111917942762, + 0.0045973085798323154, + 0.00516711687669158, + 0.00561506999656558, + 0.02275640331208706, + 0.027431663125753403, + 0.014905217103660107, + 0.019760530441999435, + 0.05587492138147354, + 0.03167271241545677, + 0.03483116999268532, + 0.03218163922429085, + 0.004662050865590572, + 0.007446181029081345, + 0.002764999633654952, + 0.006892507430166006, + 0.011795013211667538, + 0.00821008812636137, + 0.013664745725691319, + 0.009020153433084488, + 0.0030820555984973907, + 0.0026919047813862562, + 0.0017011654563248158, + 0.0022301783319562674, + 0.0027266647666692734, + 0.0020995247177779675, + 0.00264285271987319, + 0.0028562562074512243, + 0.005568778607994318, + 0.003050256287679076, + 0.0033271184656769037, + 0.0031980499625205994, + 0.003999459091573954, + 0.004053378943353891, + 0.0067079984582960606, + 0.005265474319458008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.3487326204776764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372495621442795 + ], + "opposite_logprob_tokens": [ + -3.4087250232696533 + ], + "logprob_reward": -0.03372495621442795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372495621442795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20315144956111908, + 0.09745101630687714, + 0.1374126672744751, + 0.171430304646492, + 0.15123873949050903, + 0.09286677837371826, + 0.05558827146887779, + 0.036566562950611115, + 0.02739378623664379, + 0.026900457218289375 + ], + "credit_map": [ + 0.01281071174889803, + 0.01229630596935749, + 0.014627029187977314, + 0.02505921758711338, + 0.021075423806905746, + 0.020724905654788017, + 0.06354298442602158, + 0.03301485627889633, + 0.014024440199136734, + 0.011614941991865635, + 0.00932976696640253, + 0.008216800168156624, + 0.014918266795575619, + 0.009159461595118046, + 0.017573444172739983, + 0.012613892555236816, + 0.008635487407445908, + 0.012011829763650894, + 0.009186719544231892, + 0.01600480079650879, + 0.02003319375216961, + 0.01344921626150608, + 0.03775749355554581, + 0.020333921536803246, + 0.02057512477040291, + 0.019120197743177414, + 0.014803207479417324, + 0.018770599737763405, + 0.02256564237177372, + 0.025832775980234146, + 0.023994214832782745, + 0.025768522173166275, + 0.016964618116617203, + 0.019005058333277702, + 0.01087119523435831, + 0.018364783376455307, + 0.022416602820158005, + 0.018083631992340088, + 0.0258255023509264, + 0.019707344472408295, + 0.009307922795414925, + 0.012193325906991959, + 0.007050156593322754, + 0.011292166076600552, + 0.011916237883269787, + 0.009165250696241856, + 0.0199420265853405, + 0.011999684385955334, + 0.006244936492294073, + 0.00713494885712862, + 0.0036616940051317215, + 0.005128193646669388, + 0.01242381613701582, + 0.006783359684050083, + 0.007376877591013908, + 0.006834439001977444, + 0.002402000827714801, + 0.003786266315728426, + 0.0017535663209855556, + 0.004339649807661772, + 0.005577542819082737, + 0.003591056214645505, + 0.009962508454918861, + 0.005153970792889595, + 0.002264837035909295, + 0.003107754746451974, + 0.0014078120002523065, + 0.0025162966921925545, + 0.005755862686783075, + 0.004768257029354572, + 0.0035042124800384045, + 0.004068751819431782, + 0.002056629629805684, + 0.002345622517168522, + 0.0014426358975470066, + 0.002600986510515213, + 0.005248600617051125, + 0.003298388794064522, + 0.005924238357692957, + 0.0039833527989685535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.03372495621442795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743272066116333 + ], + "opposite_logprob_tokens": [ + -0.4743271768093109 + ], + "logprob_reward": -0.9743272066116333, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743272066116333, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.36191526055336, + 0.09267336130142212, + 0.08737077564001083, + 0.0716194286942482, + 0.07247065752744675, + 0.04422062262892723, + 0.10605575144290924, + 0.06305906921625137, + 0.05100029334425926, + 0.0496147945523262 + ], + "credit_map": [ + 0.042071450501680374, + 0.034184087067842484, + 0.023159386590123177, + 0.0423455610871315, + 0.04534110054373741, + 0.03969690948724747, + 0.08386532217264175, + 0.05125140771269798, + 0.008139695972204208, + 0.008979039266705513, + 0.008616756647825241, + 0.010259435512125492, + 0.01674744486808777, + 0.008750065229833126, + 0.018557924777269363, + 0.012622995302081108, + 0.006113845389336348, + 0.013822735287249088, + 0.00582859106361866, + 0.008556749671697617, + 0.018098898231983185, + 0.009155191481113434, + 0.015252282842993736, + 0.01054247748106718, + 0.00756455585360527, + 0.01407301053404808, + 0.004382516257464886, + 0.008215649984776974, + 0.013924396596848965, + 0.00719812698662281, + 0.007980569265782833, + 0.008280599489808083, + 0.00764729268848896, + 0.008980311453342438, + 0.004380343947559595, + 0.007945092394948006, + 0.011967180296778679, + 0.0077380770817399025, + 0.013878552243113518, + 0.009933806955814362, + 0.004209127277135849, + 0.005416312254965305, + 0.004161251243203878, + 0.004851391073316336, + 0.007690944708883762, + 0.005006023216992617, + 0.006018216256052256, + 0.0068673547357320786, + 0.010660060681402683, + 0.012487652711570263, + 0.006931958254426718, + 0.008908738382160664, + 0.02452939562499523, + 0.014536685310304165, + 0.01370124239474535, + 0.014300014823675156, + 0.007443207316100597, + 0.009618619456887245, + 0.004598874598741531, + 0.005835678428411484, + 0.010990433394908905, + 0.006844344548881054, + 0.009883436374366283, + 0.007844471372663975, + 0.006710046902298927, + 0.007637868635356426, + 0.0032376523595303297, + 0.005105841439217329, + 0.010842729359865189, + 0.005703166127204895, + 0.00625284668058157, + 0.005510137416422367, + 0.006713608279824257, + 0.0056833201088011265, + 0.0036532299127429724, + 0.004072767216712236, + 0.009299445897340775, + 0.005744902417063713, + 0.007616016548126936, + 0.006831500679254532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.9743272066116333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique brass oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a dark walnut desk with visible age-worn scratches and patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flame inside the glass shade teardrop-shaped and dancing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are old leather-bound books visible nearby on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and does the camera remain still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814982786774635 + ], + "opposite_logprob_tokens": [ + -3.5298149585723877 + ], + "logprob_reward": -0.029814982786774635, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814982786774635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2547365725040436, + 0.06964736431837082, + 0.09788840264081955, + 0.05846148356795311, + 0.06879273056983948, + 0.09735159575939178, + 0.06925799697637558, + 0.09536290168762207, + 0.15255406498908997, + 0.03594692796468735 + ], + "credit_map": [ + 0.02130277454853058, + 0.03217072784900665, + 0.01850578747689724, + 0.03748564049601555, + 0.03379867970943451, + 0.03364885225892067, + 0.04420403018593788, + 0.03362007066607475, + 0.005265159532427788, + 0.008198834955692291, + 0.005885207559913397, + 0.009003921411931515, + 0.01213756948709488, + 0.011493610218167305, + 0.006877582520246506, + 0.010785472579300404, + 0.005901703145354986, + 0.012026476673781872, + 0.006214100401848555, + 0.009397261776030064, + 0.016212932765483856, + 0.011644807644188404, + 0.02196698822081089, + 0.014524134807288647, + 0.005163420457392931, + 0.009825657121837139, + 0.002894060919061303, + 0.007739348337054253, + 0.009727518074214458, + 0.008997959084808826, + 0.005251641850918531, + 0.008861874230206013, + 0.00858741533011198, + 0.00884626992046833, + 0.005973978899419308, + 0.006737301126122475, + 0.011435222811996937, + 0.006854608189314604, + 0.011545742861926556, + 0.008812185376882553, + 0.011698078364133835, + 0.010862515307962894, + 0.008410283364355564, + 0.007019622717052698, + 0.021886246278882027, + 0.013468751683831215, + 0.011763913556933403, + 0.012242176569998264, + 0.006857621017843485, + 0.009076512418687344, + 0.004975123330950737, + 0.0069261654280126095, + 0.01489089336246252, + 0.009725809097290039, + 0.007920986972749233, + 0.008884879760444164, + 0.012587607838213444, + 0.010300123132765293, + 0.0069008092395961285, + 0.00690041296184063, + 0.021625356748700142, + 0.011891044676303864, + 0.012998662889003754, + 0.012158879078924656, + 0.024131976068019867, + 0.016763951629400253, + 0.013410105369985104, + 0.015006603673100471, + 0.027958257123827934, + 0.01797530986368656, + 0.01909673772752285, + 0.01821111887693405, + 0.0041671800427138805, + 0.0034078950993716717, + 0.00269365799613297, + 0.0031598007772117853, + 0.004354241769760847, + 0.005521457176655531, + 0.00718242023140192, + 0.005460273940116167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.029814982786774635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_tensor([3331, 3331])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_tensor([3331, 3331])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c58d5a2f609f8872558dd97663361a3a14b75b8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_tensor([3331, 3331])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a high-frame-rate, fixed-camera, overhead, ultra-long shot, the deep, velvety night sky is adorned with countless sharp stars. The rugged black silhouettes of distant mountains frame the bottom of the frame, creating a grand and serene atmosphere. Suddenly, an incredibly dazzling meteor shatters this tranquility. In a visually smooth, slow-motion sequence, it gracefully arcs from the upper right to the lower left of the frame, a magnificent arc. The meteor's core is a brilliant, incandescent point of light, so intense that it creates a slight halo effect on the screen. It trails a long, shimmering trail of incandescent particles, graduating from bright gold to orange-red. Captured at high frame rate, each spark detached from the trail is clearly visible, each flickering in its final moments before dissipating into the endless darkness. This fleeting light briefly illuminates a corner of the night sky and the outline of the mountains. The entire scene, with its sharp contrast between grandeur and transience, creates an awe-inspiring atmosphere of extreme tranquility and cosmic wonder.", + "seed": 24674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_3331_seed_24674_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09088604971766472, + "vlm_logprob": -0.09088604971766472, + "vlm_yes_no_margin": 3.4000000238418577, + "vlm_token_credit": -0.09088604971766472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25849124789237976 + ], + "opposite_logprob_tokens": [ + -1.7584912776947021 + ], + "logprob_reward": -0.25849124789237976, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.25849124789237976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0724993646144867, + 0.04235072061419487, + 0.06527180224657059, + 0.024013737216591835, + 0.016465598717331886, + 0.05483667552471161, + 0.029250120744109154, + 0.06130516901612282, + 0.5908496975898743, + 0.04315706342458725 + ], + "credit_map": [ + 0.005401932634413242, + 0.017219753935933113, + 0.007593722082674503, + 0.00618071248754859, + 0.012102569453418255, + 0.010739247314631939, + 0.007685033138841391, + 0.00557640241459012, + 0.0031697270460426807, + 0.007352275308221579, + 0.004726898856461048, + 0.004056814592331648, + 0.007692993152886629, + 0.008328091353178024, + 0.003794496413320303, + 0.003229425987228751, + 0.006849796045571566, + 0.01357545331120491, + 0.0050029149278998375, + 0.006032811477780342, + 0.010665369220077991, + 0.012716921046376228, + 0.005905997008085251, + 0.004522545263171196, + 0.002649676753208041, + 0.0031889877282083035, + 0.0019603483378887177, + 0.002446786966174841, + 0.004069370683282614, + 0.0031579448841512203, + 0.0032950108870863914, + 0.003245612606406212, + 0.0023033826146274805, + 0.002155911410227418, + 0.002562585286796093, + 0.002190215513110161, + 0.001809708192013204, + 0.001651852740906179, + 0.002291968557983637, + 0.001499975216574967, + 0.012607874348759651, + 0.00489412946626544, + 0.015219508670270443, + 0.002703265752643347, + 0.0027845059521496296, + 0.004716277588158846, + 0.007036910392343998, + 0.004874206613749266, + 0.004004102665930986, + 0.002393136266618967, + 0.0030334715265780687, + 0.002643079264089465, + 0.0023411011788994074, + 0.007929205894470215, + 0.0028976250905543566, + 0.0040084002539515495, + 0.0077151944860816, + 0.005519506987184286, + 0.007876137271523476, + 0.005596929229795933, + 0.005371523089706898, + 0.006826470606029034, + 0.010191645473241806, + 0.01220776792615652, + 0.11026091128587723, + 0.017190473154187202, + 0.013383518904447556, + 0.019754935055971146, + 0.01904643513262272, + 0.06239189952611923, + 0.0941295400261879, + 0.254692018032074, + 0.006040241569280624, + 0.0027957544662058353, + 0.002420474076643586, + 0.00282541592605412, + 0.004734721966087818, + 0.005437764804810286, + 0.004780196584761143, + 0.014122496359050274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.25849124789237976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002508828416466713 + ], + "opposite_logprob_tokens": [ + -6.002508640289307 + ], + "logprob_reward": -0.002508828416466713, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002508828416466713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07087504863739014, + 0.02966615930199623, + 0.0535830557346344, + 0.052706874907016754, + 0.05287131667137146, + 0.0326458103954792, + 0.014955224469304085, + 0.05034761130809784, + 0.4754028618335724, + 0.16694599390029907 + ], + "credit_map": [ + 0.005247362423688173, + 0.017131628468632698, + 0.0027386907022446394, + 0.00459259282797575, + 0.021680373698472977, + 0.01155262440443039, + 0.003883801866322756, + 0.00404798798263073, + 0.0028961298521608114, + 0.006701214704662561, + 0.0013396918075159192, + 0.0022698251996189356, + 0.005233428440988064, + 0.006324284709990025, + 0.0020889972802251577, + 0.002812590217217803, + 0.004263411276042461, + 0.00939102377742529, + 0.0038227790500968695, + 0.006111667491495609, + 0.010580147616565228, + 0.009848426096141338, + 0.005210136994719505, + 0.004355464596301317, + 0.005631096661090851, + 0.006002344656735659, + 0.004471464082598686, + 0.004554019775241613, + 0.009201631881296635, + 0.00765438936650753, + 0.007473863661289215, + 0.0077180699445307255, + 0.007308881729841232, + 0.00891722273081541, + 0.005686966702342033, + 0.00595135148614645, + 0.0065305158495903015, + 0.005894496571272612, + 0.0070792073383927345, + 0.005502683110535145, + 0.008817209862172604, + 0.0021041505970060825, + 0.011263212189078331, + 0.0019330488285049796, + 0.0013754951069131494, + 0.0019507503602653742, + 0.0019400431774556637, + 0.003261906560510397, + 0.0018394700018689036, + 0.0014205817133188248, + 0.0015587415546178818, + 0.001915288856253028, + 0.0018082844326272607, + 0.003206979250535369, + 0.0013769845245406032, + 0.001828895416110754, + 0.006727195344865322, + 0.004569657612591982, + 0.00785908941179514, + 0.005433766636997461, + 0.0044511412270367146, + 0.006165049504488707, + 0.006453985348343849, + 0.008687732741236687, + 0.09173906594514847, + 0.013893372379243374, + 0.01054206769913435, + 0.015383509919047356, + 0.014277622103691101, + 0.049268897622823715, + 0.07536975294351578, + 0.20492859184741974, + 0.033042412251234055, + 0.00711902417242527, + 0.004156085196882486, + 0.007383972872048616, + 0.00870511494576931, + 0.016891757026314735, + 0.01697446033358574, + 0.07267317175865173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.002508828416466713 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104988634586334 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.038104988634586334, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104988634586334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03896740451455116, + 0.034715667366981506, + 0.09271342307329178, + 0.023895733058452606, + 0.009147098287940025, + 0.05416109040379524, + 0.039345309138298035, + 0.06225030496716499, + 0.36419904232025146, + 0.2806049883365631 + ], + "credit_map": [ + 0.0033570232335478067, + 0.009797725826501846, + 0.003543838858604431, + 0.0035598184913396835, + 0.007107816636562347, + 0.00560034392401576, + 0.0029296514112502337, + 0.003071184502914548, + 0.0035192095674574375, + 0.005999076180160046, + 0.0027747293934226036, + 0.003624893492087722, + 0.0055017839185893536, + 0.007280393037945032, + 0.0024269975256174803, + 0.003588586812838912, + 0.008156444877386093, + 0.019084185361862183, + 0.005801153369247913, + 0.0076382956467568874, + 0.013054123148322105, + 0.025652091950178146, + 0.006291124038398266, + 0.007035999093204737, + 0.002678046701475978, + 0.003055997658520937, + 0.0017119288677349687, + 0.0016277951654046774, + 0.005043506156653166, + 0.0030417919624596834, + 0.0033953061792999506, + 0.0033413597848266363, + 0.0011940300464630127, + 0.001124399364925921, + 0.0009055319824256003, + 0.0012112964177504182, + 0.0010447752429172397, + 0.0011728891404345632, + 0.001368939527310431, + 0.0011252366239205003, + 0.00949498638510704, + 0.009768643416464329, + 0.005253362003713846, + 0.005041998811066151, + 0.005122685804963112, + 0.008434304967522621, + 0.005292287562042475, + 0.0057528214529156685, + 0.005633053835481405, + 0.0046401238068938255, + 0.0037975474260747433, + 0.004348369315266609, + 0.004100269638001919, + 0.006320909131318331, + 0.004130255896598101, + 0.006374778226017952, + 0.009759240783751011, + 0.004552270285785198, + 0.010281390510499477, + 0.006932754535228014, + 0.005746642127633095, + 0.007855567149817944, + 0.007180511020123959, + 0.009941930882632732, + 0.05240289866924286, + 0.04407496750354767, + 0.031709522008895874, + 0.03725261241197586, + 0.04162416607141495, + 0.03811066597700119, + 0.04157932475209236, + 0.07744486629962921, + 0.05412625893950462, + 0.01343734934926033, + 0.008291641250252724, + 0.015833662822842598, + 0.01572914607822895, + 0.027869800105690956, + 0.02729693055152893, + 0.1180201917886734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.038104988634586334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539708986878395 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539708986878395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539708986878395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06812424212694168, + 0.03008384443819523, + 0.05691365525126457, + 0.037681445479393005, + 0.030090678483247757, + 0.08155408501625061, + 0.05654853582382202, + 0.09743131697177887, + 0.3286471366882324, + 0.2129250466823578 + ], + "credit_map": [ + 0.006284513510763645, + 0.01440734975039959, + 0.0073404996655881405, + 0.0051383087411522865, + 0.013650446198880672, + 0.00997134204953909, + 0.006220871116966009, + 0.005110916215926409, + 0.0029997890815138817, + 0.005122940521687269, + 0.0027481357101351023, + 0.00296552712097764, + 0.006212925538420677, + 0.005066589917987585, + 0.002918061800301075, + 0.0020498742815107107, + 0.00506139500066638, + 0.011413630098104477, + 0.0033947359770536423, + 0.005380817223340273, + 0.00776476226747036, + 0.014725903980433941, + 0.004071973729878664, + 0.0051004416309297085, + 0.004083533305674791, + 0.005180892068892717, + 0.0031284699216485023, + 0.0025945594534277916, + 0.006040224339812994, + 0.004865020513534546, + 0.005013332702219486, + 0.00677541783079505, + 0.0046095638535916805, + 0.00595843093469739, + 0.0027329488657414913, + 0.003028987208381295, + 0.0028037279844284058, + 0.0032105157151818275, + 0.003071435960009694, + 0.004675072152167559, + 0.02267637848854065, + 0.006073472555726767, + 0.02873280644416809, + 0.004143876954913139, + 0.0028940432239323854, + 0.005557188764214516, + 0.004539186134934425, + 0.006937134545296431, + 0.0075576347298920155, + 0.005151338875293732, + 0.009895611554384232, + 0.006311717443168163, + 0.005111142061650753, + 0.009641838259994984, + 0.005494892131537199, + 0.007384368218481541, + 0.014001062139868736, + 0.00890547689050436, + 0.013569529168307781, + 0.010737001895904541, + 0.00919998437166214, + 0.011135980486869812, + 0.012567021884024143, + 0.017315266653895378, + 0.04826398566365242, + 0.020730698481202126, + 0.017362425103783607, + 0.021225210279226303, + 0.022290891036391258, + 0.03376219794154167, + 0.04847542569041252, + 0.11653633415699005, + 0.045492514967918396, + 0.009258794598281384, + 0.0064774975180625916, + 0.008968200534582138, + 0.011717082001268864, + 0.020481832325458527, + 0.02170197293162346, + 0.08882715553045273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.012539708986878395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427854746580124 + ], + "opposite_logprob_tokens": [ + -2.0177855491638184 + ], + "logprob_reward": -0.1427854746580124, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1427854746580124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030580541118979454, + 0.015517855994403362, + 0.023446904495358467, + 0.019399315118789673, + 0.00979594700038433, + 0.019632315263152122, + 0.008923383429646492, + 0.04850766435265541, + 0.29474174976348877, + 0.529454231262207 + ], + "credit_map": [ + 0.0021484040189534426, + 0.006211277563124895, + 0.0029208126943558455, + 0.0019388376967981458, + 0.00739958556368947, + 0.003341438015922904, + 0.004266571719199419, + 0.0023536127991974354, + 0.0012030703946948051, + 0.0022673849016427994, + 0.0010770284570753574, + 0.001765803201124072, + 0.002493842039257288, + 0.003554331371560693, + 0.001662344904616475, + 0.001494051655754447, + 0.0020011994056403637, + 0.004702187608927488, + 0.0014896861976012588, + 0.0021110393572598696, + 0.004544609226286411, + 0.00430421344935894, + 0.001830016728490591, + 0.00246395543217659, + 0.0017830650322139263, + 0.002755936933681369, + 0.0013732460793107748, + 0.0013752710074186325, + 0.003448396921157837, + 0.0023747673258185387, + 0.0023933930788189173, + 0.003895238507539034, + 0.0013237905222922564, + 0.0015698301140218973, + 0.0011672907276079059, + 0.0011500355321913958, + 0.0011761813657358289, + 0.0011223229812458158, + 0.0008669341332279146, + 0.0014195616822689772, + 0.003367203287780285, + 0.004292775876820087, + 0.0021968409419059753, + 0.0023551906924694777, + 0.0016474800650030375, + 0.002186670433729887, + 0.001679284032434225, + 0.0019068700494244695, + 0.0013131513260304928, + 0.0009327956358902156, + 0.0007476652972400188, + 0.001073088962584734, + 0.0010230584302917123, + 0.0015800204128026962, + 0.0011124382726848125, + 0.001141164917498827, + 0.007138917688280344, + 0.004645810462534428, + 0.004828401375561953, + 0.0049628023989498615, + 0.005881702993065119, + 0.006055436562746763, + 0.006053334102034569, + 0.008941256441175938, + 0.05550002679228783, + 0.00890806782990694, + 0.006767128128558397, + 0.009966316632926464, + 0.009974038228392601, + 0.03166670724749565, + 0.04652469605207443, + 0.12543481588363647, + 0.1083117425441742, + 0.021662458777427673, + 0.012132885865867138, + 0.023540876805782318, + 0.027005471289157867, + 0.05205317959189415, + 0.05307689681649208, + 0.23167070746421814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.1427854746580124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-frame-rate, fixed-camera, overhead, ultra-long shot, the deep, velvety night sky is adorned with countless sharp stars. The rugged black silhouettes of distant mountains frame the bottom of the frame, creating a grand and serene atmosphere. Suddenly, an incredibly dazzling meteor shatters this tranquility. In a visually smooth, slow-motion sequence, it gracefully arcs from the upper right to the lower left of the frame, a magnificent arc. The meteor's core is a brilliant, incandescent point of light, so intense that it creates a slight halo effect on the screen. It trails a long, shimmering trail of incandescent particles, graduating from bright gold to orange-red. Captured at high frame rate, each spark detached from the trail is clearly visible, each flickering in its final moments before dissipating into the endless darkness. This fleeting light briefly illuminates a corner of the night sky and the outline of the mountains. The entire scene, with its sharp contrast between grandeur and transience, creates an awe-inspiring atmosphere of extreme tranquility and cosmic wonder.", + "seed": 24675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_3331_seed_24675_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12237984724342824, + "vlm_logprob": -0.12237984724342824, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.12237984724342824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37775853276252747 + ], + "opposite_logprob_tokens": [ + -1.252758502960205 + ], + "logprob_reward": -0.37775853276252747, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.37775853276252747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21477393805980682, + 0.2583056092262268, + 0.0706252008676529, + 0.07231217622756958, + 0.037800516933202744, + 0.030073821544647217, + 0.03815790265798569, + 0.021983882412314415, + 0.05617835372686386, + 0.19978854060173035 + ], + "credit_map": [ + 0.011757832951843739, + 0.03844108805060387, + 0.017785733565688133, + 0.012370152398943901, + 0.046499546617269516, + 0.057982444763183594, + 0.01585046574473381, + 0.01408670749515295, + 0.025736313313245773, + 0.06490112096071243, + 0.029068920761346817, + 0.020484035834670067, + 0.03146645054221153, + 0.030158769339323044, + 0.028574474155902863, + 0.027915528044104576, + 0.00519366841763258, + 0.010029224678874016, + 0.009093573316931725, + 0.008793218992650509, + 0.005754928104579449, + 0.008062317967414856, + 0.011195478960871696, + 0.012502800673246384, + 0.005506531335413456, + 0.00848308578133583, + 0.006246834993362427, + 0.008283963426947594, + 0.01691337488591671, + 0.010384582914412022, + 0.009695407934486866, + 0.006798404734581709, + 0.0028340541757643223, + 0.004930560942739248, + 0.0035795869771391153, + 0.004657282028347254, + 0.006475947331637144, + 0.004384215455502272, + 0.00567846791818738, + 0.005260402802377939, + 0.0035225236788392067, + 0.004924633540213108, + 0.003570632077753544, + 0.003182274056598544, + 0.00248302286490798, + 0.004654044285416603, + 0.0041761151514947414, + 0.003560577519237995, + 0.00408034585416317, + 0.0074422298930585384, + 0.004528507124632597, + 0.004481221083551645, + 0.0029837049078196287, + 0.004806021694093943, + 0.003640418639406562, + 0.006195456255227327, + 0.002447982784360647, + 0.002669939072802663, + 0.002328420989215374, + 0.002583370078355074, + 0.0030353309120982885, + 0.0029955473728477955, + 0.003019223688170314, + 0.0029040691442787647, + 0.0074568637646734715, + 0.007518711034208536, + 0.005945651326328516, + 0.004252552520483732, + 0.006130971945822239, + 0.010925492271780968, + 0.006638367660343647, + 0.007309748325496912, + 0.018271038308739662, + 0.022599535062909126, + 0.03156972676515579, + 0.01813429780304432, + 0.029663657769560814, + 0.028551755473017693, + 0.01930149830877781, + 0.031697068363428116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.37775853276252747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268855020403862 + ], + "opposite_logprob_tokens": [ + -5.2552690505981445 + ], + "logprob_reward": -0.005268855020403862, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268855020403862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.335818350315094, + 0.3425652086734772, + 0.1393575519323349, + 0.03914453834295273, + 0.014881712384521961, + 0.022573422640562057, + 0.013112908229231834, + 0.02240898460149765, + 0.022827722132205963, + 0.04730965569615364 + ], + "credit_map": [ + 0.014034051448106766, + 0.056390054523944855, + 0.02151227369904518, + 0.024936826899647713, + 0.07671478390693665, + 0.08028943091630936, + 0.02889961004257202, + 0.033041298389434814, + 0.030054356902837753, + 0.08649443835020065, + 0.029433000832796097, + 0.02375744841992855, + 0.051449090242385864, + 0.06134014576673508, + 0.03331633284687996, + 0.02672039158642292, + 0.009655297733843327, + 0.018560310825705528, + 0.013537507504224777, + 0.014532538130879402, + 0.02031843736767769, + 0.02500595711171627, + 0.026634663343429565, + 0.01111284177750349, + 0.00359521945938468, + 0.005174304358661175, + 0.003398073837161064, + 0.0036208671517670155, + 0.010099081322550774, + 0.005096131470054388, + 0.004442581441253424, + 0.003718279069289565, + 0.001439661835320294, + 0.0020406560506671667, + 0.0017496367217972875, + 0.0015400031115859747, + 0.0021823199931532145, + 0.0015804065624251962, + 0.0023177161347121, + 0.0020313123241066933, + 0.0025628474541008472, + 0.005775791592895985, + 0.0024999077431857586, + 0.002495318651199341, + 0.0010980587685480714, + 0.00367374112829566, + 0.002039463957771659, + 0.0024282936938107014, + 0.0012823863653466105, + 0.0019341235747560859, + 0.00172831816598773, + 0.0013520150678232312, + 0.001036647823639214, + 0.0021772764157503843, + 0.0014795954339206219, + 0.0021225460804998875, + 0.0022939906921237707, + 0.003611329011619091, + 0.0027427440509200096, + 0.0016519867349416018, + 0.0015859645791351795, + 0.005801078863441944, + 0.0014915409265086055, + 0.0032303486950695515, + 0.0023672194220125675, + 0.0026229070499539375, + 0.0029971697367727757, + 0.0019595578778535128, + 0.003615007968619466, + 0.003429297124966979, + 0.0028048274107277393, + 0.00303173647262156, + 0.004504653625190258, + 0.006975824479013681, + 0.006069442257285118, + 0.0038570177275687456, + 0.005465625785291195, + 0.009065460413694382, + 0.003973919898271561, + 0.007397710811346769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.005268855020403862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001381367444992 + ], + "opposite_logprob_tokens": [ + -2.6950137615203857 + ], + "logprob_reward": -0.07001381367444992, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001381367444992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10379719734191895, + 0.06641072779893875, + 0.07849632203578949, + 0.07777637988328934, + 0.04019796475768089, + 0.028948940336704254, + 0.039210282266139984, + 0.027827830985188484, + 0.0474792942404747, + 0.48985499143600464 + ], + "credit_map": [ + 0.005328429397195578, + 0.01256021298468113, + 0.011078507639467716, + 0.012838857248425484, + 0.01698913425207138, + 0.017182163894176483, + 0.011434155516326427, + 0.016385739669203758, + 0.006036986131221056, + 0.014725818298757076, + 0.007855923846364021, + 0.005413483362644911, + 0.007899348624050617, + 0.007367054931819439, + 0.009050918743014336, + 0.008061202242970467, + 0.005085827317088842, + 0.012590784579515457, + 0.007210359908640385, + 0.009774545207619667, + 0.012857013382017612, + 0.008233519271016121, + 0.01450023427605629, + 0.008244041353464127, + 0.0053032697178423405, + 0.009616812691092491, + 0.005740565247833729, + 0.010689194314181805, + 0.017605010420084, + 0.01096612960100174, + 0.010295887477695942, + 0.007559517864137888, + 0.0034770199563354254, + 0.006247935350984335, + 0.0038396466989070177, + 0.004329707007855177, + 0.006175176240503788, + 0.006001030560582876, + 0.005041581578552723, + 0.0050858743488788605, + 0.0030566074419766665, + 0.006084067281335592, + 0.0021748996805399656, + 0.0028779751155525446, + 0.002362594474107027, + 0.0041128648445010185, + 0.0031749741174280643, + 0.005104959011077881, + 0.0037128806579858065, + 0.006865249946713448, + 0.0036748037673532963, + 0.0050600552931427956, + 0.004192875698208809, + 0.006923056673258543, + 0.004988646134734154, + 0.003792717820033431, + 0.0030421451665461063, + 0.003444841131567955, + 0.0037865545600652695, + 0.002853719750419259, + 0.003468923270702362, + 0.004776017740368843, + 0.003045162186026573, + 0.0034104681108146906, + 0.0043020835146307945, + 0.008406695909798145, + 0.004039057996124029, + 0.0032133781351149082, + 0.006209583021700382, + 0.011709431186318398, + 0.0047666453756392, + 0.0048324186354875565, + 0.04288411885499954, + 0.055864062160253525, + 0.08002614229917526, + 0.04594278708100319, + 0.07129520922899246, + 0.07063505053520203, + 0.04627280682325363, + 0.07693485915660858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07001381367444992 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23436690866947174, + 0.14150077104568481, + 0.14324086904525757, + 0.05172440782189369, + 0.04966501519083977, + 0.06431472301483154, + 0.031471140682697296, + 0.06145867705345154, + 0.10047464072704315, + 0.12178286164999008 + ], + "credit_map": [ + 0.01129755936563015, + 0.036231473088264465, + 0.01827932335436344, + 0.019438497722148895, + 0.04985684156417847, + 0.05044705793261528, + 0.021842198446393013, + 0.026973934844136238, + 0.014648673124611378, + 0.03994212672114372, + 0.013109668157994747, + 0.010691025294363499, + 0.019043568521738052, + 0.01706545799970627, + 0.014619624242186546, + 0.01238061673939228, + 0.009664526209235191, + 0.021202411502599716, + 0.012921716086566448, + 0.01530730165541172, + 0.024558192119002342, + 0.021811427548527718, + 0.0226672925055027, + 0.015107995830476284, + 0.0045083132572472095, + 0.009416240267455578, + 0.0045619807206094265, + 0.005729238968342543, + 0.005976928863674402, + 0.007153208367526531, + 0.009180334396660328, + 0.005198159255087376, + 0.005176709499210119, + 0.00908856000751257, + 0.00497043039649725, + 0.0051105995662510395, + 0.007534031290560961, + 0.00622105598449707, + 0.0066146766766905785, + 0.004948947578668594, + 0.006964664440602064, + 0.01659456081688404, + 0.004673763178288937, + 0.004857961554080248, + 0.006600613240152597, + 0.009350131265819073, + 0.006854979321360588, + 0.008418041281402111, + 0.0033196292351931334, + 0.0055834525264799595, + 0.0035003870725631714, + 0.0031617034692317247, + 0.0036530375946313143, + 0.005320279859006405, + 0.0034261071123182774, + 0.0035065424162894487, + 0.006846560165286064, + 0.006959347054362297, + 0.0069150859490036964, + 0.006145893130451441, + 0.008529876358807087, + 0.00979541800916195, + 0.007676116656512022, + 0.00859037321060896, + 0.009069118648767471, + 0.017232870683073997, + 0.009017331525683403, + 0.006796585395932198, + 0.014399231411516666, + 0.024225492030382156, + 0.0103234788402915, + 0.009410525672137737, + 0.012181811966001987, + 0.020984681323170662, + 0.014061830006539822, + 0.009408220648765564, + 0.01817525364458561, + 0.022009707987308502, + 0.009719401597976685, + 0.015241949819028378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016066256910562515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427917778491974 + ], + "opposite_logprob_tokens": [ + -2.017791748046875 + ], + "logprob_reward": -0.1427917778491974, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427917778491974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18082286417484283, + 0.23551905155181885, + 0.08995722234249115, + 0.05864882469177246, + 0.045994363725185394, + 0.04368153214454651, + 0.04221431538462639, + 0.028182996436953545, + 0.058278750628232956, + 0.21670004725456238 + ], + "credit_map": [ + 0.013547674752771854, + 0.029434261843562126, + 0.022558119148015976, + 0.01705170050263405, + 0.030168041586875916, + 0.033189624547958374, + 0.016543537378311157, + 0.01832989603281021, + 0.021332578733563423, + 0.06110759451985359, + 0.022506892681121826, + 0.017100831493735313, + 0.03721380606293678, + 0.034918416291475296, + 0.022216185927391052, + 0.01912275142967701, + 0.006234733387827873, + 0.013582191430032253, + 0.009406594559550285, + 0.009683125652372837, + 0.013447676785290241, + 0.012721898965537548, + 0.016697082668542862, + 0.008183919824659824, + 0.00464820908382535, + 0.007988931611180305, + 0.005869139451533556, + 0.00906265527009964, + 0.011274224147200584, + 0.006370925810188055, + 0.007123693823814392, + 0.006311040837317705, + 0.004289951641112566, + 0.007964645512402058, + 0.003599470481276512, + 0.006575569976121187, + 0.007179299369454384, + 0.005873372778296471, + 0.00642295042052865, + 0.004089103080332279, + 0.003806665539741516, + 0.009853040799498558, + 0.003960335627198219, + 0.006140495650470257, + 0.003082763170823455, + 0.006189854349941015, + 0.004320103675127029, + 0.006328273098915815, + 0.00428053131327033, + 0.006306132301688194, + 0.005163196474313736, + 0.005674597807228565, + 0.003748863935470581, + 0.005960910581052303, + 0.00597055908292532, + 0.005109522957354784, + 0.004124602768570185, + 0.003584532765671611, + 0.0025336346589028835, + 0.002189984545111656, + 0.0031610424630343914, + 0.005492761731147766, + 0.00296001392416656, + 0.004136422649025917, + 0.0062316604889929295, + 0.00656948983669281, + 0.0066682519391179085, + 0.005665403325110674, + 0.008877256885170937, + 0.009069079533219337, + 0.007320533040910959, + 0.007877076044678688, + 0.020384235307574272, + 0.02494109235703945, + 0.03235159441828728, + 0.019567398354411125, + 0.03558332100510597, + 0.032736342400312424, + 0.02018793299794197, + 0.03094814531505108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.1427917778491974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-frame-rate, fixed-camera, overhead, ultra-long shot, the deep, velvety night sky is adorned with countless sharp stars. The rugged black silhouettes of distant mountains frame the bottom of the frame, creating a grand and serene atmosphere. Suddenly, an incredibly dazzling meteor shatters this tranquility. In a visually smooth, slow-motion sequence, it gracefully arcs from the upper right to the lower left of the frame, a magnificent arc. The meteor's core is a brilliant, incandescent point of light, so intense that it creates a slight halo effect on the screen. It trails a long, shimmering trail of incandescent particles, graduating from bright gold to orange-red. Captured at high frame rate, each spark detached from the trail is clearly visible, each flickering in its final moments before dissipating into the endless darkness. This fleeting light briefly illuminates a corner of the night sky and the outline of the mountains. The entire scene, with its sharp contrast between grandeur and transience, creates an awe-inspiring atmosphere of extreme tranquility and cosmic wonder.", + "seed": 24676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_3331_seed_24676_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12806959431618453, + "vlm_logprob": -0.12806959431618453, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.12806959431618453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.356088250875473 + ], + "opposite_logprob_tokens": [ + -1.3560882806777954 + ], + "logprob_reward": -0.356088250875473, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.356088250875473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1162978932261467, + 0.14908459782600403, + 0.1325971931219101, + 0.07853438705205917, + 0.04933236166834831, + 0.08629077672958374, + 0.16147427260875702, + 0.10964082926511765, + 0.05186142399907112, + 0.06488626450300217 + ], + "credit_map": [ + 0.019953053444623947, + 0.027312923222780228, + 0.016524868085980415, + 0.013166501186788082, + 0.013662365265190601, + 0.011763174086809158, + 0.006542464718222618, + 0.007372542750090361, + 0.02153608575463295, + 0.0253303200006485, + 0.016779297962784767, + 0.01966823823750019, + 0.01919436641037464, + 0.02468615397810936, + 0.012182684615254402, + 0.00970744900405407, + 0.009304728358983994, + 0.045901618897914886, + 0.0049494183622300625, + 0.005378656089305878, + 0.044169217348098755, + 0.01562560722231865, + 0.00419607013463974, + 0.0030718808993697166, + 0.010622452013194561, + 0.0228644460439682, + 0.006362923886626959, + 0.009460460394620895, + 0.013166701421141624, + 0.00816418882459402, + 0.004209619015455246, + 0.003683588933199644, + 0.009985578246414661, + 0.006264230702072382, + 0.006594191305339336, + 0.007210119161754847, + 0.008322346024215221, + 0.005664548836648464, + 0.0027106241323053837, + 0.002580720465630293, + 0.01158664096146822, + 0.009161552414298058, + 0.012583225034177303, + 0.010049959644675255, + 0.01379460096359253, + 0.010987316258251667, + 0.007773338817059994, + 0.010354151017963886, + 0.016934698447585106, + 0.023973967880010605, + 0.02024565078318119, + 0.029386738315224648, + 0.02257293090224266, + 0.022707227617502213, + 0.01549059059470892, + 0.010162469930946827, + 0.009350739419460297, + 0.02236345037817955, + 0.009851044043898582, + 0.010243101045489311, + 0.02019922249019146, + 0.009563094936311245, + 0.022343356162309647, + 0.0057268207892775536, + 0.006896941922605038, + 0.013050848618149757, + 0.004177482798695564, + 0.006665523629635572, + 0.007352862972766161, + 0.005327007733285427, + 0.005120804533362389, + 0.0032699538860470057, + 0.006431355606764555, + 0.019737020134925842, + 0.004772665910422802, + 0.007459649350494146, + 0.011426846496760845, + 0.007341988384723663, + 0.0042941272258758545, + 0.003422605339437723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.356088250875473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002211269922554493 + ], + "opposite_logprob_tokens": [ + -6.127211093902588 + ], + "logprob_reward": -0.002211269922554493, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002211269922554493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15368430316448212, + 0.05212651938199997, + 0.16926245391368866, + 0.06415069103240967, + 0.06667821109294891, + 0.06681083142757416, + 0.05669691041111946, + 0.12757131457328796, + 0.08437506854534149, + 0.15864364802837372 + ], + "credit_map": [ + 0.02662705071270466, + 0.024510947987437248, + 0.0212942473590374, + 0.023724330589175224, + 0.01883690059185028, + 0.019364768639206886, + 0.010229745879769325, + 0.009096311405301094, + 0.006796402391046286, + 0.011617588810622692, + 0.005892328452318907, + 0.007533598691225052, + 0.00621030991896987, + 0.00770058948546648, + 0.003319196868687868, + 0.0030565045308321714, + 0.01615791767835617, + 0.051923446357250214, + 0.00859626941382885, + 0.016881324350833893, + 0.037128712981939316, + 0.024018827825784683, + 0.009061501361429691, + 0.005494448821991682, + 0.010035958141088486, + 0.015725472941994667, + 0.005428565200418234, + 0.007940266281366348, + 0.010003350675106049, + 0.007598333992063999, + 0.004539086949080229, + 0.002879659878090024, + 0.012398785911500454, + 0.008555063046514988, + 0.006994945928454399, + 0.009178919717669487, + 0.01303060632199049, + 0.008726912550628185, + 0.004457623697817326, + 0.0033353555481880903, + 0.011687254533171654, + 0.006843757349997759, + 0.00618803920224309, + 0.008153969421982765, + 0.013494165614247322, + 0.008260051719844341, + 0.007271040230989456, + 0.004912554286420345, + 0.00673446711152792, + 0.01176142506301403, + 0.005075684282928705, + 0.006097310688346624, + 0.009984839707612991, + 0.006721741519868374, + 0.005273906514048576, + 0.005047537852078676, + 0.012970036827027798, + 0.026715246960520744, + 0.011091334745287895, + 0.013639097101986408, + 0.017071062698960304, + 0.01847594603896141, + 0.01836840622127056, + 0.009240183979272842, + 0.009982789866626263, + 0.02292611636221409, + 0.005313714034855366, + 0.011759224347770214, + 0.015037035569548607, + 0.008207501843571663, + 0.0064772022888064384, + 0.004671481437981129, + 0.021892476826906204, + 0.03883703798055649, + 0.01277195569127798, + 0.021770307794213295, + 0.027801934629678726, + 0.01766297034919262, + 0.009647657163441181, + 0.008259328082203865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.002211269922554493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292031407356262 + ], + "opposite_logprob_tokens": [ + -2.2379202842712402 + ], + "logprob_reward": -0.11292031407356262, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292031407356262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10464431345462799, + 0.22583544254302979, + 0.18970075249671936, + 0.10102611035108566, + 0.025350771844387054, + 0.03868686035275459, + 0.05113502964377403, + 0.08721049129962921, + 0.12572947144508362, + 0.05068071186542511 + ], + "credit_map": [ + 0.01876039430499077, + 0.018910039216279984, + 0.015916425734758377, + 0.01322728581726551, + 0.010143461637198925, + 0.0142499515786767, + 0.008035698905587196, + 0.005401064176112413, + 0.030284002423286438, + 0.03979165479540825, + 0.0207989439368248, + 0.02808745577931404, + 0.028452955186367035, + 0.047664374113082886, + 0.01718492992222309, + 0.013571144081652164, + 0.0214238241314888, + 0.04844871535897255, + 0.011827326379716396, + 0.01564057357609272, + 0.032144945114851, + 0.03854244947433472, + 0.012567508965730667, + 0.009105416014790535, + 0.013965770602226257, + 0.02056615799665451, + 0.010759006254374981, + 0.014123890548944473, + 0.018887348473072052, + 0.012512033805251122, + 0.0062739006243646145, + 0.003938005305826664, + 0.00485530961304903, + 0.003249575151130557, + 0.002790748607367277, + 0.0030577031429857016, + 0.004261677619069815, + 0.0027307525742799044, + 0.0020034629851579666, + 0.0024015437811613083, + 0.0072530643083155155, + 0.004560604225844145, + 0.004639334511011839, + 0.004672097973525524, + 0.005820862948894501, + 0.006422988139092922, + 0.0026501945685595274, + 0.002667719963937998, + 0.005555319134145975, + 0.008965877816081047, + 0.004689251538366079, + 0.007296435069292784, + 0.008130006492137909, + 0.00699504092335701, + 0.006289695389568806, + 0.003213403280824423, + 0.009395106695592403, + 0.01201929897069931, + 0.008170059882104397, + 0.009684314019978046, + 0.011572816409170628, + 0.010879095643758774, + 0.018875956535339355, + 0.0066138398833572865, + 0.01735992357134819, + 0.02792433463037014, + 0.013256298378109932, + 0.017869755625724792, + 0.019284965470433235, + 0.012630495242774487, + 0.009433574974536896, + 0.007970144040882587, + 0.006082088686525822, + 0.01341406349092722, + 0.004756314214318991, + 0.006686427164822817, + 0.00695485295727849, + 0.006668703630566597, + 0.0027743102982640266, + 0.003343953052535653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.11292031407356262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263423640280962 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.0263423640280962, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263423640280962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17547108232975006, + 0.23476044833660126, + 0.21339531242847443, + 0.025684408843517303, + 0.022188259288668633, + 0.04764480143785477, + 0.050066981464624405, + 0.0341733917593956, + 0.10425777733325958, + 0.0923575684428215 + ], + "credit_map": [ + 0.03013327717781067, + 0.05508994683623314, + 0.014094239100813866, + 0.01248385850340128, + 0.026943938806653023, + 0.020917948335409164, + 0.00827676709741354, + 0.007531106472015381, + 0.03490396961569786, + 0.05703537166118622, + 0.017361806705594063, + 0.023734034970402718, + 0.03841906413435936, + 0.03895220905542374, + 0.013797386549413204, + 0.010556613095104694, + 0.022239815443754196, + 0.07576761394739151, + 0.008560019545257092, + 0.012740613892674446, + 0.060126546770334244, + 0.021884679794311523, + 0.006714345887303352, + 0.005361685063689947, + 0.0035241637378931046, + 0.004422063007950783, + 0.0029089825693517923, + 0.004126402549445629, + 0.0033841056283563375, + 0.004170238971710205, + 0.0018458290724083781, + 0.0013026213273406029, + 0.0032212671358138323, + 0.003672241000458598, + 0.0018892684020102024, + 0.002341764746233821, + 0.003994187340140343, + 0.003906388767063618, + 0.0017454582266509533, + 0.001417682971805334, + 0.006775236688554287, + 0.005517103243619204, + 0.005113640800118446, + 0.006326286122202873, + 0.008043898269534111, + 0.007543032988905907, + 0.0045113032683730125, + 0.003814305877313018, + 0.0045204684138298035, + 0.008690450340509415, + 0.006388611160218716, + 0.007936039939522743, + 0.005732161924242973, + 0.005885878577828407, + 0.008417353965342045, + 0.002496017375960946, + 0.0038057502824813128, + 0.006827839184552431, + 0.003150745527818799, + 0.004310825373977423, + 0.0047703152522444725, + 0.0059773181565105915, + 0.002560896333307028, + 0.0027697025798261166, + 0.01320298109203577, + 0.021959086880087852, + 0.01082434132695198, + 0.016874318942427635, + 0.01778324879705906, + 0.009678838774561882, + 0.007244194392114878, + 0.006690762471407652, + 0.011504075489938259, + 0.0225298423320055, + 0.008023886010050774, + 0.013893675059080124, + 0.013630863279104233, + 0.011757150292396545, + 0.005747046321630478, + 0.0052710301242768764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0263423640280962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278577268123627 + ], + "opposite_logprob_tokens": [ + -2.0177857875823975 + ], + "logprob_reward": -0.14278577268123627, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278577268123627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14464348554611206, + 0.08334213495254517, + 0.08291510492563248, + 0.055277567356824875, + 0.08033541589975357, + 0.09557562321424484, + 0.09390962868928909, + 0.0694955512881279, + 0.1643304079771042, + 0.13017505407333374 + ], + "credit_map": [ + 0.023771295323967934, + 0.02969752624630928, + 0.019622769206762314, + 0.017718426883220673, + 0.017058638855814934, + 0.016545552760362625, + 0.008647939190268517, + 0.01158135011792183, + 0.010306700132787228, + 0.015470580197870731, + 0.008939479477703571, + 0.012037349864840508, + 0.010856698267161846, + 0.013384531252086163, + 0.0069383615627884865, + 0.005408443510532379, + 0.010404327884316444, + 0.017551176249980927, + 0.006064640358090401, + 0.007495759055018425, + 0.01579318754374981, + 0.015198320150375366, + 0.006151975132524967, + 0.004255726933479309, + 0.008755850605666637, + 0.009743836708366871, + 0.006461470853537321, + 0.0075475606136024, + 0.009133785031735897, + 0.007492090575397015, + 0.003275649156421423, + 0.0028673317283391953, + 0.010186773724853992, + 0.01242650207132101, + 0.008993134833872318, + 0.011278250254690647, + 0.016222767531871796, + 0.010886403732001781, + 0.0042989603243768215, + 0.006042630411684513, + 0.014331718906760216, + 0.010785453021526337, + 0.010553467087447643, + 0.013381247408688068, + 0.017411815002560616, + 0.012398773804306984, + 0.008935151621699333, + 0.007778008934110403, + 0.008962958119809628, + 0.01619596593081951, + 0.012687988579273224, + 0.015398827381432056, + 0.011273853480815887, + 0.013015289790928364, + 0.010190964676439762, + 0.00618379469960928, + 0.0070424191653728485, + 0.015881212428212166, + 0.006679038051515818, + 0.00873121339827776, + 0.009417274035513401, + 0.007305091246962547, + 0.009520409628748894, + 0.004918895196169615, + 0.022665156051516533, + 0.03570158779621124, + 0.015638692304491997, + 0.025134257972240448, + 0.03312716260552406, + 0.01304724719375372, + 0.009809444658458233, + 0.009206865914165974, + 0.016681579872965813, + 0.030155248939990997, + 0.012596548534929752, + 0.021131904795765877, + 0.01975650154054165, + 0.01459667831659317, + 0.007556598167866468, + 0.007700007874518633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.14278577268123627 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-frame-rate, fixed-camera, overhead, ultra-long shot, the deep, velvety night sky is adorned with countless sharp stars. The rugged black silhouettes of distant mountains frame the bottom of the frame, creating a grand and serene atmosphere. Suddenly, an incredibly dazzling meteor shatters this tranquility. In a visually smooth, slow-motion sequence, it gracefully arcs from the upper right to the lower left of the frame, a magnificent arc. The meteor's core is a brilliant, incandescent point of light, so intense that it creates a slight halo effect on the screen. It trails a long, shimmering trail of incandescent particles, graduating from bright gold to orange-red. Captured at high frame rate, each spark detached from the trail is clearly visible, each flickering in its final moments before dissipating into the endless darkness. This fleeting light briefly illuminates a corner of the night sky and the outline of the mountains. The entire scene, with its sharp contrast between grandeur and transience, creates an awe-inspiring atmosphere of extreme tranquility and cosmic wonder.", + "seed": 24677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_11/index_3331_seed_24677_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04598734173923731, + "vlm_logprob": -0.04598734173923731, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04598734173923731, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11169951409101486 + ], + "opposite_logprob_tokens": [ + -2.6116995811462402 + ], + "logprob_reward": -0.11169951409101486, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.11169951409101486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35761168599128723, + 0.10337970405817032, + 0.07433146983385086, + 0.04583628103137016, + 0.0413077287375927, + 0.02799641713500023, + 0.09762591123580933, + 0.04112815484404564, + 0.13808873295783997, + 0.07269379496574402 + ], + "credit_map": [ + 0.02589121274650097, + 0.051711976528167725, + 0.04852158576250076, + 0.06985297799110413, + 0.049752023071050644, + 0.029857398942112923, + 0.05216351896524429, + 0.02986103668808937, + 0.009465995244681835, + 0.020223740488290787, + 0.01425646711140871, + 0.02168787270784378, + 0.006394140422344208, + 0.006494169123470783, + 0.014495654962956905, + 0.01036167610436678, + 0.007241728715598583, + 0.009963360615074635, + 0.011056451126933098, + 0.015475234016776085, + 0.005156893283128738, + 0.004028849303722382, + 0.012091715820133686, + 0.009317253716289997, + 0.004348946735262871, + 0.006134853698313236, + 0.0068708984181284904, + 0.009378508664667606, + 0.0038870086427778006, + 0.0022672759369015694, + 0.00761387636885047, + 0.005334921646863222, + 0.004334408324211836, + 0.007186089642345905, + 0.006250948179513216, + 0.005630046594887972, + 0.005780552513897419, + 0.003919459879398346, + 0.005106680560857058, + 0.003099552122876048, + 0.0029237179551273584, + 0.0022016684524714947, + 0.003884549019858241, + 0.004482803400605917, + 0.0026992952916771173, + 0.0022574467584490776, + 0.0053371889516711235, + 0.004209751728922129, + 0.012833268381655216, + 0.006293900776654482, + 0.018248610198497772, + 0.02274589240550995, + 0.005925120320171118, + 0.006614017765969038, + 0.008337989449501038, + 0.016627127304673195, + 0.006707393564283848, + 0.002769853686913848, + 0.005673352628946304, + 0.004480816423892975, + 0.003627738682553172, + 0.006672749761492014, + 0.0034444115590304136, + 0.00775184528902173, + 0.016770152375102043, + 0.004816585686057806, + 0.006990752182900906, + 0.007453490048646927, + 0.009998446330428123, + 0.03941647708415985, + 0.012962473556399345, + 0.03968038409948349, + 0.008501405827701092, + 0.0063119628466665745, + 0.005744586233049631, + 0.006550485733896494, + 0.01012354250997305, + 0.010669445618987083, + 0.015646519139409065, + 0.009145859628915787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.11169951409101486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006755726411938667 + ], + "opposite_logprob_tokens": [ + -5.006755828857422 + ], + "logprob_reward": -0.006755726411938667, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006755726411938667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2409849315881729, + 0.272684782743454, + 0.05062001943588257, + 0.061229508370161057, + 0.020230459049344063, + 0.0352582186460495, + 0.08242742717266083, + 0.06586077809333801, + 0.036924924701452255, + 0.133778914809227 + ], + "credit_map": [ + 0.016531430184841156, + 0.044595036655664444, + 0.05202013999223709, + 0.044946152716875076, + 0.022784516215324402, + 0.01643107831478119, + 0.02370385453104973, + 0.019972747191786766, + 0.022387316450476646, + 0.04679198935627937, + 0.036179739981889725, + 0.0626240000128746, + 0.017449170351028442, + 0.018640130758285522, + 0.04286915808916092, + 0.025743301957845688, + 0.00477591110393405, + 0.00668039545416832, + 0.0070772855542600155, + 0.00907889660447836, + 0.005630549509078264, + 0.00469538290053606, + 0.007093303371220827, + 0.005588303320109844, + 0.0054966118186712265, + 0.007695907726883888, + 0.010707289911806583, + 0.009371829219162464, + 0.006958214100450277, + 0.005336191970854998, + 0.007222904823720455, + 0.008440564386546612, + 0.002763926750048995, + 0.0023935679346323013, + 0.003242531791329384, + 0.0034852696117013693, + 0.0016586706042289734, + 0.0011851084418594837, + 0.003297544317319989, + 0.0022038414608687162, + 0.004016567021608353, + 0.002597044687718153, + 0.0046454984694719315, + 0.007349121849983931, + 0.003279329976066947, + 0.003523975145071745, + 0.006403815932571888, + 0.0034428660292178392, + 0.011225421912968159, + 0.006470039952546358, + 0.013073258101940155, + 0.013156788423657417, + 0.007235972676426172, + 0.007802837993949652, + 0.008146891370415688, + 0.015316213481128216, + 0.009249177761375904, + 0.002753729932010174, + 0.008083352819085121, + 0.006199862342327833, + 0.004035111516714096, + 0.010380979627370834, + 0.006288071628659964, + 0.018870491534471512, + 0.004992044996470213, + 0.001738287042826414, + 0.0020608927588909864, + 0.0031646136194467545, + 0.003339181188493967, + 0.008536414243280888, + 0.0037665744312107563, + 0.009326918050646782, + 0.012123428285121918, + 0.009165868163108826, + 0.006056112237274647, + 0.010755672119557858, + 0.017614571377635002, + 0.019625205546617508, + 0.03944384679198265, + 0.018994225189089775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.006755726411938667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048645127564668655 + ], + "opposite_logprob_tokens": [ + -3.04864501953125 + ], + "logprob_reward": -0.048645127564668655, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048645127564668655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29593536257743835, + 0.06005220115184784, + 0.03122287429869175, + 0.02940843254327774, + 0.05568685755133629, + 0.045554034411907196, + 0.1520187258720398, + 0.14802192151546478, + 0.11004983633756638, + 0.07204970717430115 + ], + "credit_map": [ + 0.02016294188797474, + 0.05876118317246437, + 0.07265767455101013, + 0.048027582466602325, + 0.023600352928042412, + 0.021311013028025627, + 0.027859512716531754, + 0.023555101826786995, + 0.005314336158335209, + 0.00900813564658165, + 0.008865444920957088, + 0.008171276189386845, + 0.0058103580959141254, + 0.007606175262480974, + 0.008405226282775402, + 0.006871246732771397, + 0.0028158111963421106, + 0.004912679549306631, + 0.003607983933761716, + 0.007189605385065079, + 0.002928928006440401, + 0.002241182839497924, + 0.005221831612288952, + 0.002304854104295373, + 0.003008279949426651, + 0.004845703952014446, + 0.008568869903683662, + 0.004695347044616938, + 0.002161897486075759, + 0.001183229498565197, + 0.0029323147609829903, + 0.0020127887837588787, + 0.006673590280115604, + 0.006093076430261135, + 0.00794840045273304, + 0.01025573443621397, + 0.006481545977294445, + 0.003722376422956586, + 0.009854714386165142, + 0.004657420329749584, + 0.005532025825232267, + 0.003476198762655258, + 0.006343342363834381, + 0.009005471132695675, + 0.003944049123674631, + 0.004259220790117979, + 0.005050876177847385, + 0.007942850701510906, + 0.02770131826400757, + 0.008360056206583977, + 0.03589354455471039, + 0.01991373300552368, + 0.012040073983371258, + 0.014265216886997223, + 0.01628691330552101, + 0.017557866871356964, + 0.023357776924967766, + 0.01753900945186615, + 0.012423321604728699, + 0.02315768040716648, + 0.021826831623911858, + 0.0224310215562582, + 0.011712680570781231, + 0.015573606826364994, + 0.011779249645769596, + 0.005647206678986549, + 0.010646618902683258, + 0.010576131753623486, + 0.013015874661505222, + 0.017061447724699974, + 0.013063912279903889, + 0.028259389102458954, + 0.006459814961999655, + 0.0040575494058430195, + 0.0032854375895112753, + 0.0074694799259305, + 0.008760788477957249, + 0.011358242481946945, + 0.021528810262680054, + 0.009129581972956657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.048645127564668655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195299707353115 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195299707353115, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195299707353115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3006739020347595, + 0.07590003311634064, + 0.026943666860461235, + 0.04557962343096733, + 0.08144010603427887, + 0.09138218313455582, + 0.09837701171636581, + 0.13436469435691833, + 0.054575491696596146, + 0.09076333791017532 + ], + "credit_map": [ + 0.02278982847929001, + 0.04618123918771744, + 0.04178636148571968, + 0.049660831689834595, + 0.03965732827782631, + 0.0316665843129158, + 0.04186052829027176, + 0.027071133255958557, + 0.005587682593613863, + 0.011179172433912754, + 0.0110415518283844, + 0.019419163465499878, + 0.0071188947185873985, + 0.0073842075653374195, + 0.00897553563117981, + 0.005193816963583231, + 0.0020895004272460938, + 0.003485782304778695, + 0.003911735489964485, + 0.0047370027750730515, + 0.0035644613672047853, + 0.002265290357172489, + 0.004122663289308548, + 0.0027672313153743744, + 0.0042549907229840755, + 0.0055511146783828735, + 0.006519153248518705, + 0.005757784005254507, + 0.009227180853486061, + 0.00433440413326025, + 0.004958027042448521, + 0.004976965952664614, + 0.00907547865062952, + 0.013524368405342102, + 0.012293689884245396, + 0.011905747465789318, + 0.010851021856069565, + 0.007437076885253191, + 0.009449097327888012, + 0.006903612986207008, + 0.009464570321142673, + 0.007596230134367943, + 0.012805476784706116, + 0.016728801652789116, + 0.007603784091770649, + 0.008241063915193081, + 0.020389508455991745, + 0.00855273474007845, + 0.0126925278455019, + 0.0076420190744102, + 0.012507669627666473, + 0.015415051020681858, + 0.008890188299119473, + 0.011868779547512531, + 0.009892340749502182, + 0.019468428567051888, + 0.020488817244768143, + 0.011381570249795914, + 0.01032107975333929, + 0.021048497408628464, + 0.02277182973921299, + 0.01938248984515667, + 0.012257725931704044, + 0.01671266369521618, + 0.005920655094087124, + 0.003116591600701213, + 0.0036959999706596136, + 0.004830670077353716, + 0.006486226804554462, + 0.00958856288343668, + 0.0062021189369261265, + 0.014734659343957901, + 0.009365721605718136, + 0.003737747436389327, + 0.004857040476053953, + 0.00761366356164217, + 0.010192899964749813, + 0.015501500107347965, + 0.023461323231458664, + 0.016033438965678215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.014195299707353115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a fixed overhead ultra-long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the background elements a deep night sky with sharp stars and rugged black mountain silhouettes at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a meteor arc smoothly from the upper right to the lower left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor trail show a gradient of incandescent particles from bright gold to orange-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual sparks visible flickering and dissipating due to the high frame-rate capture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864104092121124 + ], + "opposite_logprob_tokens": [ + -3.0486409664154053 + ], + "logprob_reward": -0.04864104092121124, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864104092121124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3052031993865967, + 0.05711391195654869, + 0.05298670381307602, + 0.036817122250795364, + 0.02277430333197117, + 0.04986513406038284, + 0.09082084894180298, + 0.051569920033216476, + 0.057401590049266815, + 0.275447279214859 + ], + "credit_map": [ + 0.01645304076373577, + 0.06688598543405533, + 0.08290544152259827, + 0.058172356337308884, + 0.015484611503779888, + 0.017021363601088524, + 0.02964347042143345, + 0.01863693632185459, + 0.005728036165237427, + 0.00868995115160942, + 0.008128260262310505, + 0.013901041820645332, + 0.0037690228782594204, + 0.005339343566447496, + 0.007243498228490353, + 0.004314759746193886, + 0.004225668031722307, + 0.008494832552969456, + 0.011580592952668667, + 0.006957458797842264, + 0.005461500957608223, + 0.0035763143096119165, + 0.006127180531620979, + 0.006563153583556414, + 0.004237344488501549, + 0.004082281608134508, + 0.007195354904979467, + 0.005563405342400074, + 0.004105669446289539, + 0.002041597617790103, + 0.004274049308151007, + 0.005317419767379761, + 0.002652944764122367, + 0.002404059050604701, + 0.0034067367669194937, + 0.003714342601597309, + 0.002966594649478793, + 0.002239873632788658, + 0.0028954683803021908, + 0.002494282554835081, + 0.006646109279245138, + 0.004394838586449623, + 0.008809354156255722, + 0.007499881088733673, + 0.005624407436698675, + 0.006100168917328119, + 0.005940856412053108, + 0.004849517717957497, + 0.01677495799958706, + 0.006893498357385397, + 0.023060042411088943, + 0.0123292850330472, + 0.005759711377322674, + 0.008072246797382832, + 0.006955177988857031, + 0.010975931771099567, + 0.009906995110213757, + 0.0052251736633479595, + 0.003664204617962241, + 0.006473403889685869, + 0.0055291601456701756, + 0.010757820680737495, + 0.0037384256720542908, + 0.00627473508939147, + 0.0076179890893399715, + 0.003972290083765984, + 0.005128362216055393, + 0.0065896594896912575, + 0.007771245203912258, + 0.008817922323942184, + 0.007895649410784245, + 0.009608475491404533, + 0.021629394963383675, + 0.013622002676129341, + 0.011844044551253319, + 0.02170570008456707, + 0.032912492752075195, + 0.04300402104854584, + 0.08934582024812698, + 0.04138379916548729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.04864104092121124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_tensor([9972, 9972])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_tensor([9972, 9972])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c5cfdfb7b64abddb4e6518ddd699df5120b09b26 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_tensor([9972, 9972])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing medium shot, filmed from a slightly elevated angle, transports us deep into the depths of an ancient military tent. The sole illumination comes from the flickering flame of a simple bronze candlestick beside a sand table. A gaunt, forty-year-old strategist leans over it, clad in a dark, wide-sleeved silk robe. His long, jet-black hair, streaked with silver, is held high in a bun by a simple wooden hairpin. His brow furrowed, his eyes deep as a pool, as he scans the vast wooden sand table before him. Within, miniature mountains and rivers, constructed from fine sand, are clearly visible. Small red and blue flags, representing the opposing armies, are planted at strategic locations. In his right hand, he clutches a fan made of pristine white crane feathers, but he doesn't wave it, simply resting the end of the handle lightly against his lower lip, a gesture of profound thought, as if silently weighing the fate of his troops. The dancing flames cast changing light and shadows on his well-defined face, highlighting his taut jawline. The whole space was filled with a solemn and tense atmosphere, and the air seemed to be frozen. Only the occasional crackling sound of the burning candle suggested that a fateful intellectual storm was brewing in his mind.", + "seed": 4734, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_9972_seed_4734_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041365472599864006, + "vlm_logprob": -0.041365472599864006, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.041365472599864006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10280749201774597 + ], + "opposite_logprob_tokens": [ + -2.4778075218200684 + ], + "logprob_reward": -0.10280749201774597, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10280749201774597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4717128872871399, + 0.1344892829656601, + 0.04231392219662666, + 0.04365009069442749, + 0.04358186572790146, + 0.06079805642366409, + 0.06196489557623863, + 0.0634671300649643, + 0.03521677851676941, + 0.042805131524801254 + ], + "credit_map": [ + 0.03623021021485329, + 0.04043011739850044, + 0.036304276436567307, + 0.03559208661317825, + 0.04330962523818016, + 0.15289293229579926, + 0.036541715264320374, + 0.09041189402341843, + 0.010357597842812538, + 0.005814345553517342, + 0.015897274017333984, + 0.007357052061706781, + 0.013074063695967197, + 0.04687116667628288, + 0.009040453471243382, + 0.02607732266187668, + 0.0028308103792369366, + 0.0020486670546233654, + 0.004805514123290777, + 0.0038683100137859583, + 0.0050382306799292564, + 0.012171789072453976, + 0.004006935749202967, + 0.007543665356934071, + 0.0027163035701960325, + 0.0018779397942125797, + 0.004059141501784325, + 0.002303843619301915, + 0.004712003283202648, + 0.015855809673666954, + 0.0029280048329383135, + 0.009197046980261803, + 0.0034916731528937817, + 0.002174278488382697, + 0.005987498909235001, + 0.0036482319701462984, + 0.005167321301996708, + 0.010835319757461548, + 0.00517010223120451, + 0.007107442244887352, + 0.0054234955459833145, + 0.0034350010100752115, + 0.009967565536499023, + 0.008385795168578625, + 0.010129273869097233, + 0.008001335896551609, + 0.007032287772744894, + 0.008423296734690666, + 0.005574636626988649, + 0.0038744055200368166, + 0.01194092072546482, + 0.0072020115330815315, + 0.00790483970195055, + 0.008250166662037373, + 0.008517008274793625, + 0.008700906299054623, + 0.00476818298920989, + 0.003790956223383546, + 0.00558393495157361, + 0.0050802999176084995, + 0.010550979524850845, + 0.012567860074341297, + 0.009555184282362461, + 0.011569730006158352, + 0.003021450713276863, + 0.003481930820271373, + 0.005606841761618853, + 0.0032025128602981567, + 0.00630123121663928, + 0.003924353513866663, + 0.004777751863002777, + 0.0049007055349648, + 0.0035082909744232893, + 0.0041883657686412334, + 0.005365171004086733, + 0.0041310470551252365, + 0.007930894382297993, + 0.006207641214132309, + 0.004822189453989267, + 0.0066515300422906876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.10280749201774597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026384279131889343 + ], + "opposite_logprob_tokens": [ + -3.6513843536376953 + ], + "logprob_reward": -0.026384279131889343, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026384279131889343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5175955295562744, + 0.10702672600746155, + 0.05959615483880043, + 0.03589673340320587, + 0.023798508569598198, + 0.03585663065314293, + 0.08302952349185944, + 0.08336900919675827, + 0.02428836189210415, + 0.029542740434408188 + ], + "credit_map": [ + 0.03000461496412754, + 0.0649624913930893, + 0.06811679899692535, + 0.039106614887714386, + 0.07665632665157318, + 0.10812656581401825, + 0.057271186262369156, + 0.07335102558135986, + 0.0054511455819010735, + 0.0063955895602703094, + 0.010865908116102219, + 0.0072590382769703865, + 0.024286646395921707, + 0.023655960336327553, + 0.01128561981022358, + 0.01782684214413166, + 0.0043113562278449535, + 0.0031477136071771383, + 0.00736260274425149, + 0.005948742851614952, + 0.011155245825648308, + 0.009120305068790913, + 0.009357547387480736, + 0.00919265951961279, + 0.0033288607373833656, + 0.002236371859908104, + 0.005646286997944117, + 0.004574935883283615, + 0.0047966851852834225, + 0.004486370366066694, + 0.006095652002841234, + 0.004731579218059778, + 0.002073809737339616, + 0.0014151857467368245, + 0.0020323535427451134, + 0.0015370759647339582, + 0.005122987087816, + 0.0034896095748990774, + 0.004318889696151018, + 0.0038086010608822107, + 0.00236206385307014, + 0.001870343228802085, + 0.003403811249881983, + 0.0035514903720468283, + 0.005830878391861916, + 0.008820220828056335, + 0.0038342801854014397, + 0.0061835492961108685, + 0.007826324552297592, + 0.003932280465960503, + 0.01257469691336155, + 0.009807613678276539, + 0.01216763537377119, + 0.00882871076464653, + 0.015238601714372635, + 0.012653671205043793, + 0.005990730598568916, + 0.005570396315306425, + 0.009263097308576107, + 0.010564756579697132, + 0.011802089400589466, + 0.016627945005893707, + 0.011211761273443699, + 0.012338247150182724, + 0.0021707871928811073, + 0.0017525098519399762, + 0.002722774865105748, + 0.0024995303247123957, + 0.003603017656132579, + 0.004062267020344734, + 0.0039789024740457535, + 0.00349857727997005, + 0.00309319281950593, + 0.0023949441965669394, + 0.0033169793896377087, + 0.002427989849820733, + 0.005279756151139736, + 0.004700272809714079, + 0.003558505093678832, + 0.004771106410771608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.026384279131889343 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03375169262290001 + ], + "opposite_logprob_tokens": [ + -3.4087517261505127 + ], + "logprob_reward": -0.03375169262290001, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03375169262290001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47424831986427307, + 0.15026874840259552, + 0.10755881667137146, + 0.03135790303349495, + 0.033260278403759, + 0.0411890372633934, + 0.030816230922937393, + 0.03576694428920746, + 0.05056445673108101, + 0.04496924206614494 + ], + "credit_map": [ + 0.05017049238085747, + 0.03986746445298195, + 0.05717632547020912, + 0.03528473898768425, + 0.057230960577726364, + 0.12149909138679504, + 0.04089755564928055, + 0.07212166488170624, + 0.01816536672413349, + 0.010839087888598442, + 0.017638694494962692, + 0.025893179699778557, + 0.01785380020737648, + 0.021114761009812355, + 0.019010266289114952, + 0.019753573462367058, + 0.010722626931965351, + 0.007868644781410694, + 0.01201984379440546, + 0.01275149080902338, + 0.013714407570660114, + 0.022814292460680008, + 0.012208320200443268, + 0.015459176152944565, + 0.0031747939065098763, + 0.002432688605040312, + 0.0036287710536271334, + 0.0034231902100145817, + 0.006627468857914209, + 0.003198205027729273, + 0.004337191581726074, + 0.004535592161118984, + 0.0035005866084247828, + 0.002211987739428878, + 0.0052177743054926395, + 0.0035477238707244396, + 0.005716992542147636, + 0.0035786847583949566, + 0.004936205223202705, + 0.004550321493297815, + 0.004167966078966856, + 0.0030520514119416475, + 0.007227566093206406, + 0.007471729069948196, + 0.005621566437184811, + 0.0045171696692705154, + 0.004418465308845043, + 0.004712521098554134, + 0.0018570516258478165, + 0.0021013254299759865, + 0.005494854412972927, + 0.0024931812658905983, + 0.004037121310830116, + 0.0054827346466481686, + 0.004374686162918806, + 0.004975274670869112, + 0.004374855197966099, + 0.0027118574362248182, + 0.004331943579018116, + 0.003344527678564191, + 0.0050875237211585045, + 0.006651134695857763, + 0.0040312656201422215, + 0.0052338349632918835, + 0.00505071273073554, + 0.004879739135503769, + 0.010875147767364979, + 0.0053205410949885845, + 0.005495994817465544, + 0.007955243811011314, + 0.004201279487460852, + 0.006785793229937553, + 0.006040665786713362, + 0.004290327895432711, + 0.005898526404052973, + 0.005255394149571657, + 0.0067072343081235886, + 0.005920728202909231, + 0.004575197119265795, + 0.006281164474785328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03375169262290001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299138993024826 + ], + "opposite_logprob_tokens": [ + -3.773299217224121 + ], + "logprob_reward": -0.023299138993024826, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299138993024826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3956240713596344, + 0.15733809769153595, + 0.0878462940454483, + 0.057494208216667175, + 0.050167880952358246, + 0.0709490031003952, + 0.04046931862831116, + 0.0414179265499115, + 0.038434356451034546, + 0.060258861631155014 + ], + "credit_map": [ + 0.029818745329976082, + 0.03215253725647926, + 0.05023600906133652, + 0.04371151700615883, + 0.05637471005320549, + 0.058022916316986084, + 0.06819689273834229, + 0.057110704481601715, + 0.01438406016677618, + 0.007464200723916292, + 0.013682518154382706, + 0.015892144292593002, + 0.01935667172074318, + 0.03407198563218117, + 0.026504727080464363, + 0.025981776416301727, + 0.007232689764350653, + 0.003945055417716503, + 0.007846102118492126, + 0.008925328962504864, + 0.013158770278096199, + 0.015727398917078972, + 0.01618906855583191, + 0.014821875840425491, + 0.0035384742077440023, + 0.00236386526376009, + 0.005585402715951204, + 0.0038740690797567368, + 0.007757337763905525, + 0.01541966199874878, + 0.008703869767487049, + 0.010251526720821857, + 0.0034815024118870497, + 0.002157291630282998, + 0.0034967425744980574, + 0.0039682998321950436, + 0.006830825936049223, + 0.013189488090574741, + 0.007928737439215183, + 0.009114990010857582, + 0.005621143616735935, + 0.004223325755447149, + 0.0116741843521595, + 0.009736542589962482, + 0.01238215807825327, + 0.008270664140582085, + 0.009671634994447231, + 0.009369348175823689, + 0.003026754129678011, + 0.0027587648946791887, + 0.004285650793462992, + 0.00439733499661088, + 0.005146780051290989, + 0.009442020207643509, + 0.004519659094512463, + 0.006892353296279907, + 0.002865635324269533, + 0.002466405974701047, + 0.0037958375178277493, + 0.003664355957880616, + 0.006325825117528439, + 0.006841940339654684, + 0.008050463162362576, + 0.007407459430396557, + 0.003066462464630604, + 0.0027775526978075504, + 0.003280271776020527, + 0.003241707105189562, + 0.00976194441318512, + 0.004006503149867058, + 0.00621638260781765, + 0.006083529908210039, + 0.0062677133828401566, + 0.0037858940195292234, + 0.0063306912779808044, + 0.005166226532310247, + 0.010200174525380135, + 0.010116573423147202, + 0.008066369220614433, + 0.010325217619538307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023299138993024826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058476023375988 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.02058476023375988, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058476023375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41239580512046814, + 0.1175837516784668, + 0.08562721312046051, + 0.0748954564332962, + 0.06445309519767761, + 0.05332820490002632, + 0.040619950741529465, + 0.06601789593696594, + 0.023678692057728767, + 0.061399877071380615 + ], + "credit_map": [ + 0.043218765407800674, + 0.0331147238612175, + 0.04676837846636772, + 0.0455499105155468, + 0.047812774777412415, + 0.09172767400741577, + 0.04353007674217224, + 0.06067350506782532, + 0.013578799553215504, + 0.00546568026766181, + 0.016140487045049667, + 0.013161174952983856, + 0.0177012849599123, + 0.013557509519159794, + 0.017305076122283936, + 0.02067374810576439, + 0.00906131137162447, + 0.004934629891067743, + 0.01406534668058157, + 0.014305824413895607, + 0.01034857239574194, + 0.011047905310988426, + 0.010868111625313759, + 0.010995525866746902, + 0.00749812601134181, + 0.004353129770606756, + 0.010189320892095566, + 0.011584257706999779, + 0.012175099924206734, + 0.0084028085693717, + 0.010290929116308689, + 0.010401785373687744, + 0.00556739280000329, + 0.0031474751885980368, + 0.015890127047896385, + 0.00914972648024559, + 0.006611812859773636, + 0.008912252262234688, + 0.006812622305005789, + 0.008361689746379852, + 0.004636442754417658, + 0.0039904313161969185, + 0.011851482093334198, + 0.006490166764706373, + 0.0074048880487680435, + 0.0054037077352404594, + 0.006367263849824667, + 0.0071838307194411755, + 0.004067676607519388, + 0.0033930549398064613, + 0.005603391211479902, + 0.006429378408938646, + 0.0051437863148748875, + 0.005047069862484932, + 0.005626490339636803, + 0.005309106782078743, + 0.0048621343448758125, + 0.005655226297676563, + 0.008054417558014393, + 0.006096141412854195, + 0.010571658611297607, + 0.010712898336350918, + 0.00936499610543251, + 0.010700427927076817, + 0.003133378690108657, + 0.002073463052511215, + 0.002404086524620652, + 0.0022882833145558834, + 0.003966462332755327, + 0.003366310615092516, + 0.003158085746690631, + 0.003288623644039035, + 0.006984174717217684, + 0.005188187584280968, + 0.008402195759117603, + 0.006591707468032837, + 0.008516531437635422, + 0.011396389454603195, + 0.0053228954784572124, + 0.008997798897325993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02058476023375988 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium shot, filmed from a slightly elevated angle, transports us deep into the depths of an ancient military tent. The sole illumination comes from the flickering flame of a simple bronze candlestick beside a sand table. A gaunt, forty-year-old strategist leans over it, clad in a dark, wide-sleeved silk robe. His long, jet-black hair, streaked with silver, is held high in a bun by a simple wooden hairpin. His brow furrowed, his eyes deep as a pool, as he scans the vast wooden sand table before him. Within, miniature mountains and rivers, constructed from fine sand, are clearly visible. Small red and blue flags, representing the opposing armies, are planted at strategic locations. In his right hand, he clutches a fan made of pristine white crane feathers, but he doesn't wave it, simply resting the end of the handle lightly against his lower lip, a gesture of profound thought, as if silently weighing the fate of his troops. The dancing flames cast changing light and shadows on his well-defined face, highlighting his taut jawline. The whole space was filled with a solemn and tense atmosphere, and the air seemed to be frozen. Only the occasional crackling sound of the burning candle suggested that a fateful intellectual storm was brewing in his mind.", + "seed": 4735, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_9972_seed_4735_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.034047826752066614, + "vlm_logprob": -0.034047826752066614, + "vlm_yes_no_margin": 3.5249999523162843, + "vlm_token_credit": -0.034047826752066614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06730300933122635 + ], + "opposite_logprob_tokens": [ + -2.942302942276001 + ], + "logprob_reward": -0.06730300933122635, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06730300933122635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4675283432006836, + 0.1320044994354248, + 0.08205811679363251, + 0.07380140572786331, + 0.05926113203167915, + 0.04987334832549095, + 0.03339584544301033, + 0.037256691604852676, + 0.02903062291443348, + 0.035789985209703445 + ], + "credit_map": [ + 0.0694967582821846, + 0.07673589140176773, + 0.06203432381153107, + 0.07934194058179855, + 0.060310788452625275, + 0.04561203718185425, + 0.026871345937252045, + 0.047125235199928284, + 0.019324280321598053, + 0.02355841174721718, + 0.019476959481835365, + 0.021556198596954346, + 0.01593557372689247, + 0.01574992947280407, + 0.007424133364111185, + 0.008979013189673424, + 0.011152760125696659, + 0.013715866953134537, + 0.011535264551639557, + 0.01286021526902914, + 0.009184030815958977, + 0.0077578783966600895, + 0.00509691284969449, + 0.010755188763141632, + 0.009508396498858929, + 0.01178777776658535, + 0.010004633106291294, + 0.011750312522053719, + 0.008831367827951908, + 0.009845206514000893, + 0.0055986749939620495, + 0.006475042086094618, + 0.007590153254568577, + 0.008554262109100819, + 0.007625804282724857, + 0.00903306994587183, + 0.007365390658378601, + 0.008142315782606602, + 0.004037967883050442, + 0.006912169978022575, + 0.005781038198620081, + 0.009602208621799946, + 0.00566561846062541, + 0.007603591773658991, + 0.007338934578001499, + 0.0050140125676989555, + 0.0031064555514603853, + 0.0057614874094724655, + 0.004890421871095896, + 0.005643486510962248, + 0.005369143094867468, + 0.006481933407485485, + 0.00276468088850379, + 0.0030597150325775146, + 0.002098230877891183, + 0.00308823655359447, + 0.004247425589710474, + 0.006533738225698471, + 0.004634414799511433, + 0.004946551285684109, + 0.0032899349462240934, + 0.004402996972203255, + 0.0031188426073640585, + 0.0060827890411019325, + 0.003591368440538645, + 0.005907927639782429, + 0.003829085500910878, + 0.0037443754263222218, + 0.003787569236010313, + 0.003355476539582014, + 0.0018531779060140252, + 0.002961644669994712, + 0.006266407668590546, + 0.004560348577797413, + 0.004761949647217989, + 0.004084835760295391, + 0.004393376410007477, + 0.004583540838211775, + 0.0024728733114898205, + 0.004666653927415609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06730300933122635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018213408067822456 + ], + "opposite_logprob_tokens": [ + -4.018213272094727 + ], + "logprob_reward": -0.018213408067822456, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018213408067822456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4852970242500305, + 0.11197303235530853, + 0.08241827040910721, + 0.054765183478593826, + 0.06721752136945724, + 0.03599255159497261, + 0.06629659235477448, + 0.03612026199698448, + 0.0297447070479393, + 0.030174871906638145 + ], + "credit_map": [ + 0.05134977027773857, + 0.08503813296556473, + 0.055832553654909134, + 0.06444710493087769, + 0.09089919179677963, + 0.06301013380289078, + 0.03142230585217476, + 0.043297819793224335, + 0.014017838053405285, + 0.01849835366010666, + 0.013544725254178047, + 0.01562921702861786, + 0.01888648420572281, + 0.01268844772130251, + 0.006203825119882822, + 0.012504142709076405, + 0.00971611961722374, + 0.015019437298178673, + 0.009750686585903168, + 0.011858334764838219, + 0.013266164809465408, + 0.010292315855622292, + 0.004925430752336979, + 0.007589777000248432, + 0.007921037264168262, + 0.007384120021015406, + 0.007133494596928358, + 0.009422839619219303, + 0.007367515005171299, + 0.006300697568804026, + 0.0040378025732934475, + 0.0051976763643324375, + 0.007412595208734274, + 0.014711570926010609, + 0.008201917633414268, + 0.012227104045450687, + 0.010088235139846802, + 0.0066826799884438515, + 0.0037295238580554724, + 0.0041638934053480625, + 0.005284207407385111, + 0.006233370862901211, + 0.004175563808530569, + 0.0051497421227395535, + 0.00476395757868886, + 0.003307092934846878, + 0.0034062203485518694, + 0.003672399092465639, + 0.007934191264212132, + 0.010464584454894066, + 0.008128834888339043, + 0.00839794147759676, + 0.012063329108059406, + 0.008567762561142445, + 0.003362224902957678, + 0.00737772649154067, + 0.004689737223088741, + 0.005787496920675039, + 0.005281633697450161, + 0.006038251798599958, + 0.004247302655130625, + 0.0038672196678817272, + 0.0021839665714651346, + 0.004024652764201164, + 0.0031333647202700377, + 0.005443448666483164, + 0.0036630763206630945, + 0.004325736779719591, + 0.005090790335088968, + 0.0036961061414331198, + 0.002047001849859953, + 0.0023451808374375105, + 0.003651539096608758, + 0.005570114590227604, + 0.003336123889312148, + 0.003878265153616667, + 0.005944524426013231, + 0.0032227770425379276, + 0.0022156424820423126, + 0.0023558861576020718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018213408067822456 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331649139523506 + ], + "opposite_logprob_tokens": [ + -3.7733163833618164 + ], + "logprob_reward": -0.02331649139523506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331649139523506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36267685890197754, + 0.15958578884601593, + 0.06932158023118973, + 0.05980011448264122, + 0.15618254244327545, + 0.037112701684236526, + 0.05292004719376564, + 0.030840305611491203, + 0.031349822878837585, + 0.04021020233631134 + ], + "credit_map": [ + 0.04825305566191673, + 0.05487070977687836, + 0.0473351813852787, + 0.06534045189619064, + 0.046716999262571335, + 0.03626084700226784, + 0.021671714261174202, + 0.04222790524363518, + 0.024623926728963852, + 0.019037365913391113, + 0.020774997770786285, + 0.023968353867530823, + 0.018711727112531662, + 0.01564650982618332, + 0.011932048946619034, + 0.024890845641493797, + 0.0074050091207027435, + 0.008782156743109226, + 0.008423094637691975, + 0.010434040799736977, + 0.009570697322487831, + 0.006259571760892868, + 0.004510580562055111, + 0.013936431147158146, + 0.007257065735757351, + 0.010109295137226582, + 0.008171958848834038, + 0.008928893133997917, + 0.0060792830772697926, + 0.010240785777568817, + 0.0038849893026053905, + 0.005127838347107172, + 0.016090083867311478, + 0.03659989312291145, + 0.019183030351996422, + 0.024471061304211617, + 0.021263178437948227, + 0.016825949773192406, + 0.010505725629627705, + 0.011243617162108421, + 0.0056978208012878895, + 0.005911943037062883, + 0.004610265139490366, + 0.005305781029164791, + 0.00454731797799468, + 0.0038554479833692312, + 0.002791057340800762, + 0.004393068607896566, + 0.0059100789949297905, + 0.009609404020011425, + 0.005716088227927685, + 0.007317617069929838, + 0.010901390574872494, + 0.00558854267001152, + 0.003139666048809886, + 0.004737260285764933, + 0.003100247122347355, + 0.003907216712832451, + 0.0032490172889083624, + 0.0038080932572484016, + 0.0044835517182946205, + 0.004442798905074596, + 0.0022278926335275173, + 0.005621489603072405, + 0.004391565918922424, + 0.005847389809787273, + 0.0036579316947609186, + 0.0045932685025036335, + 0.004111137706786394, + 0.003057239344343543, + 0.002815761836245656, + 0.0028755280654877424, + 0.005640986375510693, + 0.007848418317735195, + 0.004872898571193218, + 0.006586235016584396, + 0.005081158597022295, + 0.003965443931519985, + 0.003007666440680623, + 0.003207392990589142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02331649139523506 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811267763376236 + ], + "opposite_logprob_tokens": [ + -3.2881126403808594 + ], + "logprob_reward": -0.03811267763376236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811267763376236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31595176458358765, + 0.11533046513795853, + 0.10159753262996674, + 0.07100887596607208, + 0.09126753360033035, + 0.04373650997877121, + 0.09787295758724213, + 0.07064378261566162, + 0.0372801348567009, + 0.05531040579080582 + ], + "credit_map": [ + 0.04120839759707451, + 0.04159286245703697, + 0.04524708166718483, + 0.06492511928081512, + 0.03294873237609863, + 0.03030930832028389, + 0.023520441725850105, + 0.03619984909892082, + 0.01866592839360237, + 0.017503459006547928, + 0.014972028322517872, + 0.020459596067667007, + 0.010853035375475883, + 0.010884660296142101, + 0.008054494857788086, + 0.013937272131443024, + 0.013920808210968971, + 0.019997360184788704, + 0.013957804068922997, + 0.01746281050145626, + 0.008207145147025585, + 0.009245076216757298, + 0.006595155224204063, + 0.012211380526423454, + 0.010669511742889881, + 0.015805885195732117, + 0.009041870012879372, + 0.011545861139893532, + 0.00680634006857872, + 0.006035903934389353, + 0.003575376234948635, + 0.0075281341560184956, + 0.009588267654180527, + 0.016275586560368538, + 0.010394645854830742, + 0.013147366233170033, + 0.016897402703762054, + 0.011895515024662018, + 0.006003729999065399, + 0.007065023761242628, + 0.0042329030111432076, + 0.00702027790248394, + 0.0059971995651721954, + 0.0085326312109828, + 0.004159825854003429, + 0.00403566425666213, + 0.00277699320577085, + 0.006981017999351025, + 0.014053570106625557, + 0.01603177934885025, + 0.011749825440347195, + 0.011112391017377377, + 0.017939163371920586, + 0.013655257411301136, + 0.005398542154580355, + 0.007932422682642937, + 0.007653714623302221, + 0.012645392678678036, + 0.009035579860210419, + 0.01094748917967081, + 0.012401707470417023, + 0.007142569869756699, + 0.00428687734529376, + 0.006530454847961664, + 0.004183860495686531, + 0.005983831360936165, + 0.004290658514946699, + 0.006161930039525032, + 0.004839180503040552, + 0.0038267553318291903, + 0.0030381956603378057, + 0.004955723416060209, + 0.005150331184267998, + 0.010892808437347412, + 0.006515876390039921, + 0.01003893744200468, + 0.007489870768040419, + 0.005077330861240625, + 0.0051732673309743404, + 0.0049719917587935925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03811267763376236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293547332286835 + ], + "opposite_logprob_tokens": [ + -3.7732934951782227 + ], + "logprob_reward": -0.023293547332286835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293547332286835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37315425276756287, + 0.12663909792900085, + 0.10931728035211563, + 0.08707257360219955, + 0.06638553738594055, + 0.037715207785367966, + 0.0630868673324585, + 0.057322364300489426, + 0.0408928357064724, + 0.03841399773955345 + ], + "credit_map": [ + 0.04208575561642647, + 0.06697025895118713, + 0.050254937261343, + 0.07212018221616745, + 0.03259927034378052, + 0.043395549058914185, + 0.025040609762072563, + 0.040687691420316696, + 0.012513577938079834, + 0.017876606434583664, + 0.017381202429533005, + 0.023227909579873085, + 0.015632271766662598, + 0.016509266570210457, + 0.00853552483022213, + 0.014962741173803806, + 0.011267613619565964, + 0.017216013744473457, + 0.012884324416518211, + 0.017101911827921867, + 0.015463031828403473, + 0.014062954112887383, + 0.007641049567610025, + 0.013680393807590008, + 0.012541093863546848, + 0.02054128423333168, + 0.010808766819536686, + 0.013882656581699848, + 0.006837788503617048, + 0.008682727813720703, + 0.005270165391266346, + 0.00850809458643198, + 0.006632270757108927, + 0.012790553271770477, + 0.008428533561527729, + 0.010511199943721294, + 0.008704610168933868, + 0.008290230296552181, + 0.003869050880894065, + 0.007159085478633642, + 0.004202407319098711, + 0.006526963785290718, + 0.004943440202623606, + 0.007196780294179916, + 0.004327306989580393, + 0.0032335733994841576, + 0.0027484167367219925, + 0.004536318592727184, + 0.0076486640609800816, + 0.01087617501616478, + 0.007747439667582512, + 0.008039101026952267, + 0.011179503984749317, + 0.008549795486032963, + 0.0034564929082989693, + 0.005589695647358894, + 0.00630115857347846, + 0.009239515289664268, + 0.007444361690431833, + 0.00858213659375906, + 0.009897995740175247, + 0.006249852012842894, + 0.00392033951357007, + 0.005687011871486902, + 0.005392835009843111, + 0.008497634902596474, + 0.004751577973365784, + 0.006038650404661894, + 0.003836359130218625, + 0.004190238658338785, + 0.0028064565267413855, + 0.005379084497690201, + 0.0034549664705991745, + 0.006202246993780136, + 0.0037458918523043394, + 0.005754271522164345, + 0.006610840559005737, + 0.004338900092989206, + 0.00420308206230402, + 0.0041037979535758495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023293547332286835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium shot, filmed from a slightly elevated angle, transports us deep into the depths of an ancient military tent. The sole illumination comes from the flickering flame of a simple bronze candlestick beside a sand table. A gaunt, forty-year-old strategist leans over it, clad in a dark, wide-sleeved silk robe. His long, jet-black hair, streaked with silver, is held high in a bun by a simple wooden hairpin. His brow furrowed, his eyes deep as a pool, as he scans the vast wooden sand table before him. Within, miniature mountains and rivers, constructed from fine sand, are clearly visible. Small red and blue flags, representing the opposing armies, are planted at strategic locations. In his right hand, he clutches a fan made of pristine white crane feathers, but he doesn't wave it, simply resting the end of the handle lightly against his lower lip, a gesture of profound thought, as if silently weighing the fate of his troops. The dancing flames cast changing light and shadows on his well-defined face, highlighting his taut jawline. The whole space was filled with a solemn and tense atmosphere, and the air seemed to be frozen. Only the occasional crackling sound of the burning candle suggested that a fateful intellectual storm was brewing in his mind.", + "seed": 4736, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_9972_seed_4736_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03623136542737484, + "vlm_logprob": -0.03623136542737484, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.03623136542737484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05964061990380287 + ], + "opposite_logprob_tokens": [ + -3.059640645980835 + ], + "logprob_reward": -0.05964061990380287, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05964061990380287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37770432233810425, + 0.08544471114873886, + 0.15384609997272491, + 0.04335976392030716, + 0.0652809739112854, + 0.050946105271577835, + 0.059891242533922195, + 0.06564390659332275, + 0.05761253833770752, + 0.040270350873470306 + ], + "credit_map": [ + 0.06311596184968948, + 0.07541274279356003, + 0.05186938866972923, + 0.04956190288066864, + 0.034112460911273956, + 0.06118651106953621, + 0.020871678367257118, + 0.021573668345808983, + 0.007405778393149376, + 0.01668429747223854, + 0.009984847158193588, + 0.011345365084707737, + 0.011634436435997486, + 0.015269037336111069, + 0.007122625131160021, + 0.005998323671519756, + 0.01907222531735897, + 0.04862994700670242, + 0.012271037325263023, + 0.010252116248011589, + 0.014102404937148094, + 0.034527234733104706, + 0.005837338976562023, + 0.009153801947832108, + 0.005551607348024845, + 0.007999255321919918, + 0.005204050801694393, + 0.005891075357794762, + 0.0041962615214288235, + 0.00867800135165453, + 0.0025572257582098246, + 0.003282286925241351, + 0.014211217872798443, + 0.011652179062366486, + 0.00602927478030324, + 0.005969686899334192, + 0.005109640769660473, + 0.012555508874356747, + 0.0034409966319799423, + 0.006312468089163303, + 0.004342074505984783, + 0.006592972669750452, + 0.007018779404461384, + 0.011750656180083752, + 0.0037368026096373796, + 0.008496501483023167, + 0.005008181091398001, + 0.004000138025730848, + 0.006739967502653599, + 0.011844496242702007, + 0.005599568597972393, + 0.005959610920399427, + 0.008448276668787003, + 0.009867972694337368, + 0.006247207522392273, + 0.005184145178645849, + 0.005041653756052256, + 0.006052155047655106, + 0.008707208558917046, + 0.005293818190693855, + 0.008989050053060055, + 0.009720693342387676, + 0.012165761552751064, + 0.009673567488789558, + 0.004694248549640179, + 0.009835025295615196, + 0.004993578419089317, + 0.00873667374253273, + 0.007469826843589544, + 0.010031864047050476, + 0.0071248216554522514, + 0.004726499319076538, + 0.00366096175275743, + 0.007558364886790514, + 0.0035908392164856195, + 0.0030919071286916733, + 0.004818773828446865, + 0.006831805221736431, + 0.006085461471229792, + 0.0046322401612997055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05964061990380287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023321963846683502 + ], + "opposite_logprob_tokens": [ + -3.7733218669891357 + ], + "logprob_reward": -0.023321963846683502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023321963846683502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44169557094573975, + 0.14678627252578735, + 0.05252263695001602, + 0.051917944103479385, + 0.03380774334073067, + 0.051392167806625366, + 0.05652029812335968, + 0.056282926350831985, + 0.046263888478279114, + 0.06281059235334396 + ], + "credit_map": [ + 0.05648134648799896, + 0.098809614777565, + 0.03881426528096199, + 0.06599383056163788, + 0.05241267383098602, + 0.07831064611673355, + 0.028406625613570213, + 0.022466545924544334, + 0.01799311302602291, + 0.021324029192328453, + 0.016809705644845963, + 0.026957036927342415, + 0.018968069925904274, + 0.02182934246957302, + 0.011051199398934841, + 0.011853775940835476, + 0.006624344736337662, + 0.008389296941459179, + 0.006912728305906057, + 0.0057577029801905155, + 0.008197479881346226, + 0.009512887336313725, + 0.0025595868937671185, + 0.004568606615066528, + 0.007431412115693092, + 0.008862334303557873, + 0.007921731099486351, + 0.004394186660647392, + 0.006382350344210863, + 0.009307637810707092, + 0.0030185726936906576, + 0.004599719308316708, + 0.00454397639259696, + 0.005514499265700579, + 0.004333929158747196, + 0.003687299555167556, + 0.004922413267195225, + 0.0061906552873551846, + 0.0021038881968706846, + 0.0025110796559602022, + 0.004868274088948965, + 0.007138737477362156, + 0.007112620398402214, + 0.010229518637061119, + 0.004609379451721907, + 0.007976992055773735, + 0.005661718547344208, + 0.0037949292454868555, + 0.005554772447794676, + 0.010601704940199852, + 0.005975787527859211, + 0.006850044243037701, + 0.0077985781244933605, + 0.009466785006225109, + 0.006304929964244366, + 0.00396768981590867, + 0.008378306403756142, + 0.006869135890156031, + 0.00917789340019226, + 0.005824027583003044, + 0.006667448673397303, + 0.008912380784749985, + 0.004890590440481901, + 0.0055631413124501705, + 0.0037672300823032856, + 0.008907348848879337, + 0.005291951820254326, + 0.005053150933235884, + 0.006852762773633003, + 0.008780719712376595, + 0.004622205626219511, + 0.002988515654578805, + 0.0044580791145563126, + 0.011006806045770645, + 0.0061816382221877575, + 0.014348058961331844, + 0.006494064349681139, + 0.010909845121204853, + 0.005767580587416887, + 0.0036445176228880882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.023321963846683502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813379257917404 + ], + "opposite_logprob_tokens": [ + -3.2881338596343994 + ], + "logprob_reward": -0.03813379257917404, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813379257917404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4215143322944641, + 0.0922120064496994, + 0.05696233734488487, + 0.06273792684078217, + 0.05202842503786087, + 0.0682230070233345, + 0.08007793128490448, + 0.10168872028589249, + 0.0331185981631279, + 0.0314367450773716 + ], + "credit_map": [ + 0.08568961918354034, + 0.07034669071435928, + 0.09073484688997269, + 0.031484540551900864, + 0.03448987379670143, + 0.06880798935890198, + 0.02040223591029644, + 0.019558532163500786, + 0.02093251794576645, + 0.019299576058983803, + 0.010780656710267067, + 0.00788574293255806, + 0.007912700064480305, + 0.014694682322442532, + 0.005103429313749075, + 0.005602697841823101, + 0.004939264617860317, + 0.01145089603960514, + 0.008419986814260483, + 0.007321421056985855, + 0.006164007354527712, + 0.010411894880235195, + 0.004135051742196083, + 0.004119815304875374, + 0.007170888129621744, + 0.013537433929741383, + 0.006716231815516949, + 0.0058147842064499855, + 0.008738074451684952, + 0.011053385213017464, + 0.005222484469413757, + 0.004484646487981081, + 0.0075517939403653145, + 0.009248904883861542, + 0.007961987517774105, + 0.0060495538637042046, + 0.005349321290850639, + 0.009201651439070702, + 0.0030732834711670876, + 0.0035919295623898506, + 0.0053072781302034855, + 0.00941226165741682, + 0.009792687371373177, + 0.015165841206908226, + 0.005488930735737085, + 0.010646429844200611, + 0.0073124077171087265, + 0.005097172223031521, + 0.008219937793910503, + 0.015573915094137192, + 0.00817001424729824, + 0.007740692235529423, + 0.012881836853921413, + 0.014752905815839767, + 0.005450665019452572, + 0.007287963759154081, + 0.008676794357597828, + 0.013574076816439629, + 0.010748947970569134, + 0.012830141000449657, + 0.015925085172057152, + 0.01487315259873867, + 0.013835780322551727, + 0.011224742978811264, + 0.0031022303737699986, + 0.005673765204846859, + 0.00356930959969759, + 0.0045204199850559235, + 0.004961013328284025, + 0.005777653306722641, + 0.0032687492202967405, + 0.002245456213131547, + 0.002890919102355838, + 0.004750589374452829, + 0.003826623782515526, + 0.003576182061806321, + 0.0035010899882763624, + 0.004946283530443907, + 0.005095604341477156, + 0.00284945429302752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03813379257917404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371170163154602 + ], + "opposite_logprob_tokens": [ + -3.4087116718292236 + ], + "logprob_reward": -0.03371170163154602, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371170163154602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.402353972196579, + 0.09827161580324173, + 0.07824063301086426, + 0.08571738749742508, + 0.06810048967599869, + 0.056113358587026596, + 0.028661515563726425, + 0.045187246054410934, + 0.06334526091814041, + 0.07400842756032944 + ], + "credit_map": [ + 0.04024093598127365, + 0.06873596459627151, + 0.043232113122940063, + 0.049052972346544266, + 0.07158486545085907, + 0.06675407290458679, + 0.027326202020049095, + 0.03542688861489296, + 0.010519854724407196, + 0.020148711279034615, + 0.007983438670635223, + 0.00845204759389162, + 0.01615518145263195, + 0.01729685813188553, + 0.006490020547062159, + 0.011225502006709576, + 0.010371970012784004, + 0.019317451864480972, + 0.006758715957403183, + 0.007999865338206291, + 0.009173715487122536, + 0.014239197596907616, + 0.005423481576144695, + 0.004956232849508524, + 0.009890465997159481, + 0.016227368265390396, + 0.009037768468260765, + 0.0075192684307694435, + 0.013340087607502937, + 0.01481624972075224, + 0.005997006315737963, + 0.008889170363545418, + 0.00745148491114378, + 0.013809135183691978, + 0.0052662743255496025, + 0.006403662730008364, + 0.012583238072693348, + 0.011586717329919338, + 0.005339177791029215, + 0.005660797469317913, + 0.004425923340022564, + 0.010187757201492786, + 0.007496949750930071, + 0.009351613000035286, + 0.007501607295125723, + 0.009152005426585674, + 0.004455245565623045, + 0.0035422532819211483, + 0.002764226868748665, + 0.003716663923114538, + 0.004033552948385477, + 0.00281015457585454, + 0.004377700854092836, + 0.004685072228312492, + 0.0024499252904206514, + 0.0038242186419665813, + 0.005817804019898176, + 0.005994811654090881, + 0.005392924882471561, + 0.004437480587512255, + 0.008162977173924446, + 0.007517654448747635, + 0.002991144545376301, + 0.004872448742389679, + 0.004088155925273895, + 0.009119057096540928, + 0.006247867830097675, + 0.007687018718570471, + 0.011412395164370537, + 0.00987258367240429, + 0.007193973753601313, + 0.007724207825958729, + 0.005261378828436136, + 0.012023856863379478, + 0.007475018966943026, + 0.008381492458283901, + 0.009823865257203579, + 0.011095382273197174, + 0.0113088833168149, + 0.00863854493945837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03371170163154602 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348749175667763 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348749175667763, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348749175667763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2526967227458954, + 0.07720299065113068, + 0.06631455570459366, + 0.11417440325021744, + 0.08038457483053207, + 0.08761660009622574, + 0.07250437140464783, + 0.07169029861688614, + 0.06555743515491486, + 0.11185804009437561 + ], + "credit_map": [ + 0.023954689502716064, + 0.05338657274842262, + 0.040739450603723526, + 0.028481869027018547, + 0.02828664518892765, + 0.04560666158795357, + 0.017909811809659004, + 0.014331015758216381, + 0.005927257239818573, + 0.013973351567983627, + 0.008486938662827015, + 0.01217648759484291, + 0.01068254467099905, + 0.012509982101619244, + 0.005743741057813168, + 0.007702692877501249, + 0.009487414732575417, + 0.01359794195741415, + 0.006527528166770935, + 0.005983898416161537, + 0.008721910417079926, + 0.011976116336882114, + 0.005327341612428427, + 0.0046924082562327385, + 0.010109963826835155, + 0.024894438683986664, + 0.010342627763748169, + 0.010424693115055561, + 0.01681799627840519, + 0.019318193197250366, + 0.011644733138382435, + 0.010621762834489346, + 0.010430512949824333, + 0.014784688130021095, + 0.014254149980843067, + 0.006448831874877214, + 0.009923836216330528, + 0.014978751540184021, + 0.004351614974439144, + 0.005212186835706234, + 0.009072711691260338, + 0.01685146987438202, + 0.012325810268521309, + 0.014468601904809475, + 0.009383114986121655, + 0.014542002230882645, + 0.005633122753351927, + 0.005339764524251223, + 0.005724472925066948, + 0.011305345222353935, + 0.008215120062232018, + 0.009278779849410057, + 0.010520987212657928, + 0.011134539730846882, + 0.00718235457316041, + 0.009142768569290638, + 0.008531808853149414, + 0.010545040480792522, + 0.006962725427001715, + 0.007230496034026146, + 0.011654691770672798, + 0.010719957761466503, + 0.009256776422262192, + 0.006788795348256826, + 0.004412999842315912, + 0.010771489702165127, + 0.007000161800533533, + 0.013647235929965973, + 0.008701130747795105, + 0.01089745108038187, + 0.0058195325545966625, + 0.004307436756789684, + 0.007124335039407015, + 0.020574985072016716, + 0.011162329465150833, + 0.02136717364192009, + 0.01159439142793417, + 0.016333574429154396, + 0.014702551998198032, + 0.008998699486255646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.026348749175667763 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium shot, filmed from a slightly elevated angle, transports us deep into the depths of an ancient military tent. The sole illumination comes from the flickering flame of a simple bronze candlestick beside a sand table. A gaunt, forty-year-old strategist leans over it, clad in a dark, wide-sleeved silk robe. His long, jet-black hair, streaked with silver, is held high in a bun by a simple wooden hairpin. His brow furrowed, his eyes deep as a pool, as he scans the vast wooden sand table before him. Within, miniature mountains and rivers, constructed from fine sand, are clearly visible. Small red and blue flags, representing the opposing armies, are planted at strategic locations. In his right hand, he clutches a fan made of pristine white crane feathers, but he doesn't wave it, simply resting the end of the handle lightly against his lower lip, a gesture of profound thought, as if silently weighing the fate of his troops. The dancing flames cast changing light and shadows on his well-defined face, highlighting his taut jawline. The whole space was filled with a solemn and tense atmosphere, and the air seemed to be frozen. Only the occasional crackling sound of the burning candle suggested that a fateful intellectual storm was brewing in his mind.", + "seed": 4737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_12/index_9972_seed_4737_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03444498293101787, + "vlm_logprob": -0.03444498293101787, + "vlm_yes_no_margin": 3.5249999523162843, + "vlm_token_credit": -0.03444498293101787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06892196089029312 + ], + "opposite_logprob_tokens": [ + -2.9439220428466797 + ], + "logprob_reward": -0.06892196089029312, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06892196089029312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5550835132598877, + 0.101277656853199, + 0.04499215632677078, + 0.06698018312454224, + 0.028185103088617325, + 0.017241232097148895, + 0.04067711904644966, + 0.03381155803799629, + 0.0588952898979187, + 0.05285612866282463 + ], + "credit_map": [ + 0.12938572466373444, + 0.04442396014928818, + 0.08570737391710281, + 0.03307671844959259, + 0.0417371466755867, + 0.12254291772842407, + 0.04106437787413597, + 0.057145338505506516, + 0.023583151400089264, + 0.008603757247328758, + 0.014877580106258392, + 0.005546652711927891, + 0.008259939961135387, + 0.02416253834962845, + 0.006304200738668442, + 0.009939848445355892, + 0.010463208891451359, + 0.004771885462105274, + 0.006809657905250788, + 0.0024180959444493055, + 0.0034386753104627132, + 0.009883319959044456, + 0.0029519416857510805, + 0.004255378618836403, + 0.015017963014543056, + 0.006190705578774214, + 0.010028722696006298, + 0.005776296369731426, + 0.005481617525219917, + 0.014615498483181, + 0.004781122785061598, + 0.005088259931653738, + 0.006292311474680901, + 0.0030562453903257847, + 0.004377890378236771, + 0.00236249016597867, + 0.0024280580691993237, + 0.0054184147156775, + 0.002172196749597788, + 0.002077499870210886, + 0.0030933571979403496, + 0.0026068021543323994, + 0.0027824388816952705, + 0.0022015010472387075, + 0.0013177728978917003, + 0.002460520714521408, + 0.0013932239962741733, + 0.0013856180012226105, + 0.006584381218999624, + 0.0038299576845020056, + 0.006768642459064722, + 0.004600129555910826, + 0.003460798179730773, + 0.006357367616146803, + 0.0039159804582595825, + 0.0051598623394966125, + 0.005478099454194307, + 0.0034882593899965286, + 0.0048732515424489975, + 0.006388673558831215, + 0.002513586077839136, + 0.005555495619773865, + 0.0031100818887352943, + 0.002404113532975316, + 0.012208335101604462, + 0.004530960228294134, + 0.00886484608054161, + 0.008465384133160114, + 0.00553802540525794, + 0.01081403624266386, + 0.0045883674174547195, + 0.0038853404112160206, + 0.010190953500568867, + 0.004636009223759174, + 0.007586185820400715, + 0.002818916691467166, + 0.004205688834190369, + 0.012363881804049015, + 0.0056175347417593, + 0.005436967127025127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.06892196089029312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020622828975319862 + ], + "opposite_logprob_tokens": [ + -3.895622730255127 + ], + "logprob_reward": -0.020622828975319862, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020622828975319862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5585962533950806, + 0.11293451488018036, + 0.048621393740177155, + 0.034305766224861145, + 0.029054146260023117, + 0.03845961019396782, + 0.05073081701993942, + 0.04666268825531006, + 0.02935189940035343, + 0.051282938569784164 + ], + "credit_map": [ + 0.10392384976148605, + 0.056923530995845795, + 0.06688425689935684, + 0.043802935630083084, + 0.04280317202210426, + 0.10667625069618225, + 0.06195894256234169, + 0.07562333345413208, + 0.019346466287970543, + 0.012882625684142113, + 0.011716913431882858, + 0.009686149656772614, + 0.0092534264549613, + 0.02369960956275463, + 0.011401670053601265, + 0.014947645366191864, + 0.009446181356906891, + 0.0066659292206168175, + 0.004984772298485041, + 0.00538119999691844, + 0.0029822729993611574, + 0.0073603312484920025, + 0.005643088836222887, + 0.006157617550343275, + 0.005577169358730316, + 0.005425619892776012, + 0.004106939770281315, + 0.004049291368573904, + 0.0024787662550807, + 0.005482479464262724, + 0.002368962625041604, + 0.004816535394638777, + 0.004444468300789595, + 0.004286616574972868, + 0.004417024087160826, + 0.0029149604961276054, + 0.002702424069866538, + 0.004034628160297871, + 0.0030834958888590336, + 0.0031705289147794247, + 0.006265913601964712, + 0.007105289492756128, + 0.0042213997803628445, + 0.004707351326942444, + 0.003877200186252594, + 0.006650951225310564, + 0.0027431733906269073, + 0.0028883267659693956, + 0.007511475123465061, + 0.005836014170199633, + 0.00794543419033289, + 0.0065116919577121735, + 0.0060606724582612514, + 0.007810989394783974, + 0.003971602767705917, + 0.005082936957478523, + 0.008332920260727406, + 0.0085474643856287, + 0.005093626212328672, + 0.003949400503188372, + 0.004238933324813843, + 0.006807504687458277, + 0.004871122073382139, + 0.004821712151169777, + 0.00549236498773098, + 0.0030440620612353086, + 0.004063758999109268, + 0.0022247671149671078, + 0.0030685137026011944, + 0.004292708355933428, + 0.0028657184448093176, + 0.004300005733966827, + 0.008736115880310535, + 0.005616496782749891, + 0.006677206605672836, + 0.004681313410401344, + 0.0048348079435527325, + 0.009191497229039669, + 0.004762397147715092, + 0.006783099379390478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020622828975319862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813184052705765 + ], + "opposite_logprob_tokens": [ + -3.2881319522857666 + ], + "logprob_reward": -0.03813184052705765, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813184052705765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32457026839256287, + 0.12541265785694122, + 0.09885511547327042, + 0.061756156384944916, + 0.08894072473049164, + 0.0690779909491539, + 0.0274893157184124, + 0.09100589901208878, + 0.07153095304965973, + 0.04136093333363533 + ], + "credit_map": [ + 0.058855947107076645, + 0.0585571750998497, + 0.03170531988143921, + 0.03540649637579918, + 0.027357952669262886, + 0.055643197149038315, + 0.0287555530667305, + 0.028288617730140686, + 0.023605063557624817, + 0.013150447979569435, + 0.020049696788191795, + 0.01389176957309246, + 0.00924032274633646, + 0.023882344365119934, + 0.009262405335903168, + 0.012330611236393452, + 0.018670041114091873, + 0.01068113837391138, + 0.013851091265678406, + 0.009279986843466759, + 0.010560203343629837, + 0.015949735417962074, + 0.00970427319407463, + 0.010158644989132881, + 0.010034467093646526, + 0.007565393578261137, + 0.008527473546564579, + 0.007685000542551279, + 0.006769994273781776, + 0.0076030464842915535, + 0.006984632462263107, + 0.0065861474722623825, + 0.018523963168263435, + 0.009540688246488571, + 0.010666674934327602, + 0.00660793436691165, + 0.008676224388182163, + 0.015752585604786873, + 0.008188730105757713, + 0.010983917862176895, + 0.012675974518060684, + 0.012039709836244583, + 0.008373556658625603, + 0.006483445409685373, + 0.005288770422339439, + 0.009407700039446354, + 0.007240239530801773, + 0.007568594999611378, + 0.003998315893113613, + 0.003892837790772319, + 0.002515506697818637, + 0.002939228666946292, + 0.004002552479505539, + 0.004137761425226927, + 0.0023886808194220066, + 0.0036144296173006296, + 0.01596403867006302, + 0.01018422283232212, + 0.011233494617044926, + 0.009306717664003372, + 0.007942063733935356, + 0.0158825796097517, + 0.011262878775596619, + 0.009229900315403938, + 0.015083247795701027, + 0.008087609894573689, + 0.006358035374432802, + 0.0077966684475541115, + 0.005607393570244312, + 0.012279512360692024, + 0.006818490568548441, + 0.009499991312623024, + 0.006617567501962185, + 0.004135380033403635, + 0.00680265249684453, + 0.004750982392579317, + 0.0025219242088496685, + 0.005068405065685511, + 0.0059584518894553185, + 0.005505568813532591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03813184052705765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356296613812447 + ], + "opposite_logprob_tokens": [ + -3.6513562202453613 + ], + "logprob_reward": -0.026356296613812447, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356296613812447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4191768765449524, + 0.14462065696716309, + 0.06698327511548996, + 0.0693177655339241, + 0.10981106013059616, + 0.03982997313141823, + 0.030955180525779724, + 0.03522820770740509, + 0.0450739786028862, + 0.03900301456451416 + ], + "credit_map": [ + 0.06579276919364929, + 0.07220302522182465, + 0.05002989619970322, + 0.05016833916306496, + 0.02970895729959011, + 0.0662369355559349, + 0.04581782966852188, + 0.03921913355588913, + 0.022826045751571655, + 0.023850124329328537, + 0.018513239920139313, + 0.017467230558395386, + 0.009220244362950325, + 0.01979845017194748, + 0.016923997551202774, + 0.01602132059633732, + 0.009653081186115742, + 0.016157805919647217, + 0.007769004907459021, + 0.007359239738434553, + 0.004532453138381243, + 0.008432858623564243, + 0.006505174562335014, + 0.006573651451617479, + 0.009338312782347202, + 0.009999108500778675, + 0.011114547960460186, + 0.012296493165194988, + 0.005531127564609051, + 0.007856703363358974, + 0.007256971672177315, + 0.00592450425028801, + 0.022751765325665474, + 0.014488433487713337, + 0.014728260226547718, + 0.010896420106291771, + 0.008558708243072033, + 0.01878187246620655, + 0.009065047837793827, + 0.010540557093918324, + 0.008204266428947449, + 0.004747249651700258, + 0.0036504180170595646, + 0.00338690634816885, + 0.0028117273468524218, + 0.0058333175256848335, + 0.004580655600875616, + 0.006615431979298592, + 0.0051867286674678326, + 0.004440756980329752, + 0.003832640592008829, + 0.002847449155524373, + 0.0031923463102430105, + 0.005291474983096123, + 0.0032289663795381784, + 0.0029348174575716257, + 0.005639316514134407, + 0.00539833540096879, + 0.0038808947429060936, + 0.004143569152802229, + 0.003702817251905799, + 0.005272077862173319, + 0.004182703327387571, + 0.003008496016263962, + 0.006853490602225065, + 0.00524505041539669, + 0.0069525535218417645, + 0.00957997515797615, + 0.003534498857334256, + 0.0056664845906198025, + 0.004787891171872616, + 0.002454034984111786, + 0.005783454980701208, + 0.0048156422562897205, + 0.007688922807574272, + 0.00482918368652463, + 0.002580949105322361, + 0.004515145439654589, + 0.0044167437590658665, + 0.0043729753233492374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.026356296613812447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a medium shot filmed from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole illumination provided by a flickering flame from a simple bronze candlestick beside a sand table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a gaunt, forty-year-old strategist visible wearing a dark wide-sleeved silk robe with jet-black hair streaked with silver held in a bun by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the strategist holding a fan made of pristine white crane feathers against his lower lip while scanning a wooden sand table with miniature mountains, rivers, and small red and blue flags?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181919876486063 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181919876486063, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181919876486063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.492963582277298, + 0.11634626239538193, + 0.06355925649404526, + 0.06148585304617882, + 0.08633203059434891, + 0.020942674949765205, + 0.03716684877872467, + 0.034710224717855453, + 0.04262760654091835, + 0.043865617364645004 + ], + "credit_map": [ + 0.10118775069713593, + 0.04766686633229256, + 0.08050024509429932, + 0.037794459611177444, + 0.040728840976953506, + 0.09685822576284409, + 0.04403698071837425, + 0.04419025033712387, + 0.023404225707054138, + 0.014287889935076237, + 0.017387976869940758, + 0.01149802003055811, + 0.011315788142383099, + 0.020009582862257957, + 0.010057085193693638, + 0.00838570948690176, + 0.011550559662282467, + 0.00950819905847311, + 0.006519120652228594, + 0.007659842725843191, + 0.00593246566131711, + 0.007039928343147039, + 0.0066304500214755535, + 0.008718694560229778, + 0.008541149087250233, + 0.008265642449259758, + 0.008388406597077847, + 0.011074624955654144, + 0.00575637212023139, + 0.00776181323453784, + 0.006022871006280184, + 0.0056749796494841576, + 0.01803281344473362, + 0.009666851721704006, + 0.012323982082307339, + 0.008233046159148216, + 0.008146298117935658, + 0.015127523802220821, + 0.007539272774010897, + 0.007262248080223799, + 0.003316611982882023, + 0.0034368992783129215, + 0.002441819990053773, + 0.0017996045062318444, + 0.0028190817683935165, + 0.00336942239664495, + 0.002021349035203457, + 0.0017378877382725477, + 0.00535355182364583, + 0.004834790248423815, + 0.004994851537048817, + 0.005226025823503733, + 0.004225354176014662, + 0.005978087894618511, + 0.0032831027638167143, + 0.0032710880041122437, + 0.005895949900150299, + 0.005196449346840382, + 0.004568941425532103, + 0.0038665556348860264, + 0.0031888799276202917, + 0.005501317325979471, + 0.0033183919731527567, + 0.00317374337464571, + 0.007916679605841637, + 0.006424881983548403, + 0.004740276373922825, + 0.0037527275271713734, + 0.0038280452135950327, + 0.006587407551705837, + 0.004618851002305746, + 0.004758743569254875, + 0.006662349216639996, + 0.006021019537001848, + 0.006994706578552723, + 0.004028341267257929, + 0.0036644069477915764, + 0.005565195810049772, + 0.0048364498652517796, + 0.006093151401728392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0181919876486063 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_tensor([9147, 9147])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_tensor([9147, 9147])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..32822a8d8e9d281070ae97de83c29564b3715488 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_tensor([9147, 9147])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium-wide shot captures the sheer enchantment of a sun-dappled forest clearing. Oliver, a young boy with tousled blond hair and a simple blue and white striped T-shirt, sits cross-legged on a red and white checkered picnic blanket. Next to him sits Sparky, a fluffy golden retriever puppy, his head tilted in curiosity, his tail gently tapping the soft grass. Nearby sits a family of four rabbits: a large brown dad, a gentle white mom, and two small, furry gray bunnies. Between them stands a whimsical tea party, featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups. Oliver, with a look of focused joy on his face, is in action, steadily and deliberately pouring imaginary tea from a small white teapot into a tiny cup, carefully held between the paws of a small rabbit. Another bunny nibbles at a heart-shaped cookie on a central plate, which also contains delicately sliced carrot sticks. The entire scene is bathed in the warm golden glow of the afternoon sun, which filters through the lush green canopy above, creating a beautiful play of light and shadow. The atmosphere is tranquil, warm and whimsical, a magical moment of friendship between children, pets and gentle woodland creatures.", + "seed": 8076, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_9147_seed_8076_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038421569485217334, + "vlm_logprob": -0.038421569485217334, + "vlm_yes_no_margin": 4.474999976158142, + "vlm_token_credit": -0.038421569485217334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0030434264335781336 + ], + "opposite_logprob_tokens": [ + -6.6280436515808105 + ], + "logprob_reward": -0.0030434264335781336, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0030434264335781336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12495918571949005, + 0.0326288603246212, + 0.021736742928624153, + 0.1525275856256485, + 0.0345577634871006, + 0.021335382014513016, + 0.1077333390712738, + 0.08645700663328171, + 0.37344658374786377, + 0.04461760073900223 + ], + "credit_map": [ + 0.01639685221016407, + 0.022397099062800407, + 0.013957920484244823, + 0.014434405602514744, + 0.019875215366482735, + 0.01404586061835289, + 0.01408099103718996, + 0.009770834818482399, + 0.00453426968306303, + 0.005627863109111786, + 0.003906035562977195, + 0.003947368357330561, + 0.006413581315428019, + 0.0033703241497278214, + 0.0025913771241903305, + 0.002238041255623102, + 0.0033787584397941828, + 0.0034853864926844835, + 0.0033347257412970066, + 0.0023925048299133778, + 0.0028327072504907846, + 0.002687546657398343, + 0.0021938905119895935, + 0.0014312228886410594, + 0.025980135425925255, + 0.03133111074566841, + 0.012659533880650997, + 0.013285966590046883, + 0.0517817959189415, + 0.0067873671650886536, + 0.007256744895130396, + 0.0034449223894625902, + 0.00515925046056509, + 0.00440217973664403, + 0.0069657303392887115, + 0.006044409703463316, + 0.004342487547546625, + 0.002521802205592394, + 0.003635930363088846, + 0.0014859717339277267, + 0.0022793954703956842, + 0.0027678629849106073, + 0.0036416565999388695, + 0.0035339752212166786, + 0.0030796544160693884, + 0.002441795775666833, + 0.00206871354021132, + 0.001522327191196382, + 0.03913724422454834, + 0.039965592324733734, + 0.003668589750304818, + 0.003070904640480876, + 0.008594246581196785, + 0.0026686140336096287, + 0.007288711611181498, + 0.003339433576911688, + 0.02872350998222828, + 0.031501881778240204, + 0.004941081628203392, + 0.0036761292722076178, + 0.006329244468361139, + 0.003425173694267869, + 0.00488304952159524, + 0.0029769339598715305, + 0.13929107785224915, + 0.15556831657886505, + 0.009199106134474277, + 0.008941365405917168, + 0.02919885702431202, + 0.005881337448954582, + 0.010269531980156898, + 0.015096936374902725, + 0.00600913492962718, + 0.009997643530368805, + 0.009529227390885353, + 0.005758695304393768, + 0.004108386114239693, + 0.0038983530830591917, + 0.0028733632061630487, + 0.0024427929893136024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0030434264335781336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07554615288972855, + 0.035861484706401825, + 0.04049396142363548, + 0.11414475739002228, + 0.07019363343715668, + 0.037981096655130386, + 0.03128669783473015, + 0.05487268418073654, + 0.48850077390670776, + 0.05111874267458916 + ], + "credit_map": [ + 0.00862559862434864, + 0.011423572897911072, + 0.011995152570307255, + 0.00932235736399889, + 0.009390654042363167, + 0.01003778725862503, + 0.008266019634902477, + 0.006484999321401119, + 0.0039656274020671844, + 0.004075472243130207, + 0.004425720311701298, + 0.003718817373737693, + 0.005081575829535723, + 0.006569747347384691, + 0.004181109834462404, + 0.0038434104062616825, + 0.004286143463104963, + 0.004777878988534212, + 0.006804323755204678, + 0.0040013776160776615, + 0.006141928024590015, + 0.0072245169430971146, + 0.002708669286221266, + 0.0045491172932088375, + 0.024555275216698647, + 0.012519112788140774, + 0.01632803864777088, + 0.01438435260206461, + 0.025189220905303955, + 0.0069451164454221725, + 0.00933996681123972, + 0.004883664194494486, + 0.012377671897411346, + 0.0113037945702672, + 0.01547294296324253, + 0.010436038486659527, + 0.007617956958711147, + 0.004082571715116501, + 0.00598948122933507, + 0.002913165371865034, + 0.0057374825701117516, + 0.006046809256076813, + 0.007917921058833599, + 0.005273465067148209, + 0.0039020490366965532, + 0.002978912089020014, + 0.0042250435799360275, + 0.0018994109705090523, + 0.009094301611185074, + 0.010048716329038143, + 0.002166077494621277, + 0.0019168811850249767, + 0.0027074518147855997, + 0.0016459454782307148, + 0.002594472374767065, + 0.001112847006879747, + 0.017084510996937752, + 0.01923973485827446, + 0.003366707591339946, + 0.0021046032197773457, + 0.003850193927064538, + 0.0032405706588178873, + 0.0033544949255883694, + 0.0026318621821701527, + 0.17721493542194366, + 0.20178872346878052, + 0.015917973592877388, + 0.014313866384327412, + 0.03815467655658722, + 0.00799527857452631, + 0.013686363585293293, + 0.019428923726081848, + 0.00807939376682043, + 0.008862492628395557, + 0.009351237677037716, + 0.0068131196312606335, + 0.00528861116617918, + 0.004710769280791283, + 0.0040007284842431545, + 0.004012384917587042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.00978388637304306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368692100048065 + ], + "opposite_logprob_tokens": [ + -3.408686876296997 + ], + "logprob_reward": -0.03368692100048065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368692100048065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1252640038728714, + 0.041906245052814484, + 0.04217473044991493, + 0.45808783173561096, + 0.0857759565114975, + 0.03798239678144455, + 0.01635015942156315, + 0.10547322779893875, + 0.0627347081899643, + 0.024250736460089684 + ], + "credit_map": [ + 0.015515484847128391, + 0.017388835549354553, + 0.015856454148888588, + 0.015036278404295444, + 0.01875663548707962, + 0.01766778901219368, + 0.015791209414601326, + 0.009251312352716923, + 0.004717677365988493, + 0.009621337987482548, + 0.005040590651333332, + 0.006782830227166414, + 0.0068181948736310005, + 0.003594003850594163, + 0.002797000342980027, + 0.0025346067268401384, + 0.004857045132666826, + 0.003910752944648266, + 0.009978731162846088, + 0.006625249050557613, + 0.0061724018305540085, + 0.004624229855835438, + 0.003349610837176442, + 0.0026567087043076754, + 0.037833843380212784, + 0.10248289257287979, + 0.07402628660202026, + 0.06326740235090256, + 0.09614606946706772, + 0.026151061058044434, + 0.038946423679590225, + 0.019233860075473785, + 0.026228850707411766, + 0.0293484628200531, + 0.00579862343147397, + 0.0043135653249919415, + 0.008338314481079578, + 0.003273563226684928, + 0.004779856652021408, + 0.0036947240587323904, + 0.004758153576403856, + 0.0049591814167797565, + 0.007515539415180683, + 0.005493995733559132, + 0.004166561644524336, + 0.004045437090098858, + 0.004536087159067392, + 0.002507438650354743, + 0.003544688457623124, + 0.0030667344108223915, + 0.002457375405356288, + 0.002003262285143137, + 0.00134250707924366, + 0.0015441746218129992, + 0.0014749296242371202, + 0.0009164887014776468, + 0.033019792288541794, + 0.03441011533141136, + 0.009549585171043873, + 0.0072273630648851395, + 0.007977349683642387, + 0.004157727584242821, + 0.005642443895339966, + 0.0034888472873717546, + 0.016503099352121353, + 0.019919728860259056, + 0.0076704504899680614, + 0.005667840596288443, + 0.004868750460445881, + 0.0029269938822835684, + 0.0025429746601730585, + 0.0026348677929490805, + 0.0033161411993205547, + 0.004736431408673525, + 0.005223451182246208, + 0.003982702270150185, + 0.0023572277277708054, + 0.0018535521812736988, + 0.001426355680450797, + 0.0013548731803894043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.03368692100048065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277946949005127 + ], + "opposite_logprob_tokens": [ + -2.0177793502807617 + ], + "logprob_reward": -0.14277946949005127, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14277946949005127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1700921356678009, + 0.05302075296640396, + 0.05032603070139885, + 0.17216002941131592, + 0.11370273679494858, + 0.0698263868689537, + 0.07865849882364273, + 0.03867975249886513, + 0.21950149536132812, + 0.034032192081213 + ], + "credit_map": [ + 0.017434990033507347, + 0.02064546011388302, + 0.025423787534236908, + 0.028383320197463036, + 0.02972155250608921, + 0.018790530040860176, + 0.018292168155312538, + 0.01140032522380352, + 0.005699323024600744, + 0.011019740253686905, + 0.005617433227598667, + 0.007325987331569195, + 0.008483375422656536, + 0.004963187500834465, + 0.0053594233468174934, + 0.0045522842556238174, + 0.006903204135596752, + 0.005184269975870848, + 0.011878508143126965, + 0.006800615228712559, + 0.007942725904285908, + 0.0038844244554638863, + 0.004476740024983883, + 0.003255543066188693, + 0.055212508887052536, + 0.01479035522788763, + 0.01297931931912899, + 0.014401964843273163, + 0.055992282927036285, + 0.006138267461210489, + 0.008228100836277008, + 0.004417235031723976, + 0.0302784014493227, + 0.033953096717596054, + 0.01157161220908165, + 0.01093609444797039, + 0.010181155987083912, + 0.005440775770694017, + 0.0064927637577056885, + 0.004848829470574856, + 0.010457219555974007, + 0.014137322083115578, + 0.013590458780527115, + 0.009568234905600548, + 0.008097292855381966, + 0.004900103434920311, + 0.005451445002108812, + 0.003624312812462449, + 0.028027182444930077, + 0.028683962300419807, + 0.0030588023364543915, + 0.002866531489416957, + 0.006475171074271202, + 0.0015875882236286998, + 0.0056396727450191975, + 0.002319588325917721, + 0.011366864666342735, + 0.013397973030805588, + 0.002891595708206296, + 0.0022093835286796093, + 0.0029913727194070816, + 0.0016790623776614666, + 0.002571500837802887, + 0.0015720007941126823, + 0.061876438558101654, + 0.07391463220119476, + 0.02496952936053276, + 0.019100984558463097, + 0.01631033420562744, + 0.007032865192741156, + 0.008124493062496185, + 0.008172222413122654, + 0.004556737840175629, + 0.0037498739548027515, + 0.006302399095147848, + 0.005153887905180454, + 0.00433060759678483, + 0.004454227164387703, + 0.0026448569260537624, + 0.0028396015986800194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.14277946949005127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814144128933549 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.002814144128933549, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814144128933549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0654221773147583, + 0.09364665299654007, + 0.029795249924063683, + 0.09649664908647537, + 0.09771429002285004, + 0.06804350763559341, + 0.025706147775053978, + 0.01713273674249649, + 0.47395870089530945, + 0.03208380565047264 + ], + "credit_map": [ + 0.010013194754719734, + 0.008572926744818687, + 0.01029377430677414, + 0.008008500561118126, + 0.010563408955931664, + 0.00710119167342782, + 0.00708521343767643, + 0.003783972468227148, + 0.010428348556160927, + 0.02013993449509144, + 0.013006135821342468, + 0.01660170592367649, + 0.015544290654361248, + 0.006980068515986204, + 0.006698512472212315, + 0.0042476593516767025, + 0.0040523032657802105, + 0.002817936707288027, + 0.0076864175498485565, + 0.004463100805878639, + 0.004001885186880827, + 0.0022070957347750664, + 0.0028182140085846186, + 0.0017483009723946452, + 0.017788590863347054, + 0.03355679661035538, + 0.008373769000172615, + 0.00547076715156436, + 0.015948450192809105, + 0.004837036598473787, + 0.006084316875785589, + 0.004436937160789967, + 0.02010086178779602, + 0.021162960678339005, + 0.0149071691557765, + 0.013283999636769295, + 0.009852697141468525, + 0.00619868841022253, + 0.0077765402384102345, + 0.004431386012583971, + 0.01061104703694582, + 0.013547925278544426, + 0.012403417378664017, + 0.008795425295829773, + 0.006121920887380838, + 0.004570067394524813, + 0.008743338286876678, + 0.0032503728289157152, + 0.008430369198322296, + 0.007920628413558006, + 0.001483075087890029, + 0.0009739965316839516, + 0.002000468783080578, + 0.0018834032816812396, + 0.0019538619089871645, + 0.0010603476548567414, + 0.003144610906019807, + 0.0034033150877803564, + 0.0020209040958434343, + 0.001290623564273119, + 0.002010875614359975, + 0.0023551329504698515, + 0.001350721693597734, + 0.0015565541107207537, + 0.1731308102607727, + 0.19028812646865845, + 0.01499500684440136, + 0.017823349684476852, + 0.035664115101099014, + 0.008106178604066372, + 0.015638723969459534, + 0.018312452360987663, + 0.005048362072557211, + 0.006523859221488237, + 0.00546553498134017, + 0.004628793802112341, + 0.0027610145043581724, + 0.002555281389504671, + 0.0025753991212695837, + 0.0025255654472857714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.002814144128933549 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium-wide shot captures the sheer enchantment of a sun-dappled forest clearing. Oliver, a young boy with tousled blond hair and a simple blue and white striped T-shirt, sits cross-legged on a red and white checkered picnic blanket. Next to him sits Sparky, a fluffy golden retriever puppy, his head tilted in curiosity, his tail gently tapping the soft grass. Nearby sits a family of four rabbits: a large brown dad, a gentle white mom, and two small, furry gray bunnies. Between them stands a whimsical tea party, featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups. Oliver, with a look of focused joy on his face, is in action, steadily and deliberately pouring imaginary tea from a small white teapot into a tiny cup, carefully held between the paws of a small rabbit. Another bunny nibbles at a heart-shaped cookie on a central plate, which also contains delicately sliced carrot sticks. The entire scene is bathed in the warm golden glow of the afternoon sun, which filters through the lush green canopy above, creating a beautiful play of light and shadow. The atmosphere is tranquil, warm and whimsical, a magical moment of friendship between children, pets and gentle woodland creatures.", + "seed": 8077, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_9147_seed_8077_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024593102186918257, + "vlm_logprob": -0.024593102186918257, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.024593102186918257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003821690334007144 + ], + "opposite_logprob_tokens": [ + -6.003821849822998 + ], + "logprob_reward": -0.003821690334007144, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.003821690334007144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30057522654533386, + 0.08821326494216919, + 0.14137676358222961, + 0.09784942865371704, + 0.07136981934309006, + 0.09151915460824966, + 0.05023184418678284, + 0.03631613031029701, + 0.06148630753159523, + 0.06106210872530937 + ], + "credit_map": [ + 0.0457424633204937, + 0.03486678749322891, + 0.053345292806625366, + 0.02846834994852543, + 0.0233471617102623, + 0.04939911887049675, + 0.03272903338074684, + 0.03267701715230942, + 0.012393518351018429, + 0.010249641723930836, + 0.016689378768205643, + 0.009364881552755833, + 0.009139553643763065, + 0.014082779176533222, + 0.0066875284537673, + 0.009605979546904564, + 0.023810679093003273, + 0.023012733086943626, + 0.025865232571959496, + 0.016784939914941788, + 0.018717367202043533, + 0.012849532999098301, + 0.011518609710037708, + 0.008817670866847038, + 0.014629954472184181, + 0.0226993877440691, + 0.017591334879398346, + 0.010363818146288395, + 0.01177557185292244, + 0.008291826583445072, + 0.005339238326996565, + 0.007158295717090368, + 0.011157804168760777, + 0.013570510782301426, + 0.011017503216862679, + 0.007703007198870182, + 0.0082128020003438, + 0.008666342124342918, + 0.00529406638815999, + 0.005747785791754723, + 0.017588773742318153, + 0.018526330590248108, + 0.021075325086712837, + 0.007402229588478804, + 0.007555128540843725, + 0.009070861153304577, + 0.004607755690813065, + 0.005692752078175545, + 0.007224994245916605, + 0.011581718921661377, + 0.0064565567299723625, + 0.004878871608525515, + 0.006615073885768652, + 0.004187172744423151, + 0.0030952547676861286, + 0.006192201282829046, + 0.0054670050740242004, + 0.006778569892048836, + 0.005997844506055117, + 0.0032138172537088394, + 0.004476834088563919, + 0.003774454817175865, + 0.0024877756368368864, + 0.004119826480746269, + 0.00923494528979063, + 0.007229958660900593, + 0.008413010276854038, + 0.01800302043557167, + 0.003717903047800064, + 0.004679691977798939, + 0.005706064403057098, + 0.004501719493418932, + 0.009560891427099705, + 0.011707308702170849, + 0.007423219736665487, + 0.007201808504760265, + 0.006495732814073563, + 0.006629090290516615, + 0.006013468839228153, + 0.006030580960214138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.003821690334007144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2592083811759949, + 0.12150061875581741, + 0.15497925877571106, + 0.10752388834953308, + 0.05753028765320778, + 0.08306721597909927, + 0.025995755568146706, + 0.022684840485453606, + 0.1440010815858841, + 0.023508725687861443 + ], + "credit_map": [ + 0.04158398136496544, + 0.030590882524847984, + 0.03716101869940758, + 0.032250989228487015, + 0.018734892830252647, + 0.03780507296323776, + 0.025863690301775932, + 0.03521781414747238, + 0.018734581768512726, + 0.01095516886562109, + 0.019913313910365105, + 0.013652420602738857, + 0.012852117419242859, + 0.01894146203994751, + 0.013572444207966328, + 0.012879099696874619, + 0.020049210637807846, + 0.01735094003379345, + 0.023601824417710304, + 0.017141707241535187, + 0.01492033526301384, + 0.025279518216848373, + 0.021730247884988785, + 0.014905452728271484, + 0.015211286023259163, + 0.014312401413917542, + 0.012064465321600437, + 0.008016853593289852, + 0.008435552008450031, + 0.020891552790999413, + 0.016817757859826088, + 0.011774007230997086, + 0.00818372517824173, + 0.008397147990763187, + 0.007354614324867725, + 0.004396725911647081, + 0.0047701578587293625, + 0.010764901526272297, + 0.007220186293125153, + 0.006442825775593519, + 0.022205354645848274, + 0.006957365665584803, + 0.03152525797486305, + 0.004944864194840193, + 0.004720883443951607, + 0.006245879922062159, + 0.002791996579617262, + 0.0036756088957190514, + 0.004033396951854229, + 0.004484782926738262, + 0.003944613505154848, + 0.002580211730673909, + 0.002435532631352544, + 0.002952996641397476, + 0.0016090727876871824, + 0.003955147694796324, + 0.0024730151053518057, + 0.002614776836708188, + 0.00487767718732357, + 0.004592433571815491, + 0.0014522550627589226, + 0.0020589325577020645, + 0.0020191234070807695, + 0.0025966253597289324, + 0.019207069650292397, + 0.01729731261730194, + 0.021757736802101135, + 0.03947458788752556, + 0.009718921035528183, + 0.01072123646736145, + 0.014853517524898052, + 0.010970691218972206, + 0.0036989126820117235, + 0.0030354666523635387, + 0.0030833431519567966, + 0.0031798253767192364, + 0.0022924579679965973, + 0.00305348401889205, + 0.002904460998252034, + 0.002260772977024317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.006740807089954615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057238109409809 + ], + "opposite_logprob_tokens": [ + -3.8955724239349365 + ], + "logprob_reward": -0.02057238109409809, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057238109409809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2371937334537506, + 0.08809755742549896, + 0.05836779251694679, + 0.05244329571723938, + 0.039473772048950195, + 0.10933604091405869, + 0.018938183784484863, + 0.02052662894129753, + 0.3581450879573822, + 0.017477864399552345 + ], + "credit_map": [ + 0.03582390770316124, + 0.038194235414266586, + 0.038345567882061005, + 0.020078931003808975, + 0.01894710399210453, + 0.03452923521399498, + 0.024785617366433144, + 0.026489144191145897, + 0.012554741464555264, + 0.011146326549351215, + 0.01231242623180151, + 0.009696436114609241, + 0.009920386597514153, + 0.017851490527391434, + 0.007334243971854448, + 0.00728151248767972, + 0.009509323164820671, + 0.007919729687273502, + 0.00945228710770607, + 0.008871810510754585, + 0.0059270104393363, + 0.005911899730563164, + 0.0050168572925031185, + 0.005758871324360371, + 0.007526798639446497, + 0.006399388425052166, + 0.005642294883728027, + 0.006119078025221825, + 0.004856800194829702, + 0.010957688093185425, + 0.004410381428897381, + 0.006530864629894495, + 0.005584711208939552, + 0.0071022226475179195, + 0.007594078779220581, + 0.004380830097943544, + 0.00486755883321166, + 0.0033315045293420553, + 0.0029934558551758528, + 0.003619411261752248, + 0.03207895904779434, + 0.008675922639667988, + 0.04476447403430939, + 0.006411049515008926, + 0.006326654460281134, + 0.00548376003280282, + 0.002947755390778184, + 0.0026474713813513517, + 0.002171203726902604, + 0.002703009871765971, + 0.003044880460947752, + 0.0037000146694481373, + 0.0018744381377473474, + 0.002563171088695526, + 0.001354631152935326, + 0.00152683281339705, + 0.002908965339884162, + 0.0027571674436330795, + 0.0047154477797448635, + 0.0019498626934364438, + 0.0018035669345408678, + 0.0018912095110863447, + 0.0017533347709104419, + 0.0027470749337226152, + 0.05917488783597946, + 0.048394523561000824, + 0.05245865881443024, + 0.08955759555101395, + 0.024931279942393303, + 0.024299615994095802, + 0.036288753151893616, + 0.023039789870381355, + 0.0033253133296966553, + 0.003420858411118388, + 0.001836409093812108, + 0.002284485613927245, + 0.0017408148851245642, + 0.0016185871791094542, + 0.0018469487549737096, + 0.0014044480631127954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.02057238109409809 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901601284742355 + ], + "opposite_logprob_tokens": [ + -2.4640159606933594 + ], + "logprob_reward": -0.08901601284742355, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901601284742355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3077107071876526, + 0.18851642310619354, + 0.10366237163543701, + 0.07140403985977173, + 0.032471515238285065, + 0.024169154465198517, + 0.05174234136939049, + 0.033779215067625046, + 0.1477721929550171, + 0.03877205774188042 + ], + "credit_map": [ + 0.051369935274124146, + 0.03215603902935982, + 0.046922218054533005, + 0.0345844067633152, + 0.027490410953760147, + 0.047271233052015305, + 0.03184674680233002, + 0.036069732159376144, + 0.027835359796881676, + 0.02130230702459812, + 0.03210672363638878, + 0.020475709810853004, + 0.021140161901712418, + 0.034249044954776764, + 0.0157499760389328, + 0.01565713807940483, + 0.016812225803732872, + 0.016505295410752296, + 0.017649751156568527, + 0.012106183916330338, + 0.014144636690616608, + 0.009400852955877781, + 0.010229483246803284, + 0.006813944783061743, + 0.012514819391071796, + 0.008944815956056118, + 0.009337509982287884, + 0.0067392210476100445, + 0.00663729477673769, + 0.014015141874551773, + 0.00606450904160738, + 0.007150725927203894, + 0.004667769651859999, + 0.005070977844297886, + 0.005412362981587648, + 0.003915265668183565, + 0.003420405788347125, + 0.0037481237668544054, + 0.002661794424057007, + 0.0035748148802667856, + 0.004716598428785801, + 0.0032569614704698324, + 0.005895518697798252, + 0.003113809507340193, + 0.001967142103239894, + 0.0020565909799188375, + 0.0014971945201978087, + 0.0016653399216011167, + 0.0081059280782938, + 0.010906338691711426, + 0.008031852543354034, + 0.006347921676933765, + 0.0049879420548677444, + 0.004651107359677553, + 0.003598620183765888, + 0.005112632643431425, + 0.00469894427806139, + 0.004557141102850437, + 0.004626753274351358, + 0.004004019312560558, + 0.004088317975401878, + 0.005615502595901489, + 0.0028828238137066364, + 0.0033057094551622868, + 0.023652462288737297, + 0.0202353373169899, + 0.02192879468202591, + 0.037605125457048416, + 0.010148616507649422, + 0.010048061609268188, + 0.014478315599262714, + 0.009675491601228714, + 0.005571199581027031, + 0.0056391325779259205, + 0.006150443572551012, + 0.00608464889228344, + 0.004094217903912067, + 0.004198253620415926, + 0.003813265124335885, + 0.003220895305275917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.08901601284742355 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00281461956910789 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.00281461956910789, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00281461956910789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33648237586021423, + 0.1307913362979889, + 0.06034445762634277, + 0.12233278900384903, + 0.05452338233590126, + 0.1125333234667778, + 0.03370630368590355, + 0.027377638965845108, + 0.08587996661663055, + 0.0360284186899662 + ], + "credit_map": [ + 0.04555846005678177, + 0.07141263037919998, + 0.04734385758638382, + 0.04214414209127426, + 0.03014722280204296, + 0.03682848438620567, + 0.031572125852108, + 0.031475428491830826, + 0.01967279054224491, + 0.01711837761104107, + 0.023388506844639778, + 0.022823087871074677, + 0.012443873099982738, + 0.012587559409439564, + 0.01167969312518835, + 0.011077438481152058, + 0.010309469886124134, + 0.01074654795229435, + 0.010382425971329212, + 0.007036266848444939, + 0.00748621579259634, + 0.005547421984374523, + 0.004603978246450424, + 0.0042321281507611275, + 0.018655331805348396, + 0.01519211009144783, + 0.014623146504163742, + 0.010828406549990177, + 0.010206799954175949, + 0.024595579132437706, + 0.013816981576383114, + 0.014414429664611816, + 0.00836541224271059, + 0.008910181000828743, + 0.007602986413985491, + 0.006816101726144552, + 0.005414316896349192, + 0.007878342643380165, + 0.004264226648956537, + 0.005271813832223415, + 0.029040642082691193, + 0.009807005524635315, + 0.04019573703408241, + 0.005511410068720579, + 0.0067939492873847485, + 0.010656028054654598, + 0.004914471879601479, + 0.005614072550088167, + 0.004915422759950161, + 0.0048503871075809, + 0.004223059397190809, + 0.0032120796386152506, + 0.0031708425376564264, + 0.005043116398155689, + 0.0034154169261455536, + 0.004875977523624897, + 0.00355458352714777, + 0.004525455646216869, + 0.004550802521407604, + 0.002578753512352705, + 0.0030801130924373865, + 0.003779088146984577, + 0.002382882172241807, + 0.0029259573202580214, + 0.013509145937860012, + 0.011264009401202202, + 0.0117326145991683, + 0.01948610320687294, + 0.0063407099805772305, + 0.008682827465236187, + 0.008065692149102688, + 0.006798864342272282, + 0.004850631579756737, + 0.005864910315722227, + 0.005131171550601721, + 0.005412423983216286, + 0.003495646407827735, + 0.004924548324197531, + 0.0029738268349319696, + 0.003375258529558778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00281461956910789 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium-wide shot captures the sheer enchantment of a sun-dappled forest clearing. Oliver, a young boy with tousled blond hair and a simple blue and white striped T-shirt, sits cross-legged on a red and white checkered picnic blanket. Next to him sits Sparky, a fluffy golden retriever puppy, his head tilted in curiosity, his tail gently tapping the soft grass. Nearby sits a family of four rabbits: a large brown dad, a gentle white mom, and two small, furry gray bunnies. Between them stands a whimsical tea party, featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups. Oliver, with a look of focused joy on his face, is in action, steadily and deliberately pouring imaginary tea from a small white teapot into a tiny cup, carefully held between the paws of a small rabbit. Another bunny nibbles at a heart-shaped cookie on a central plate, which also contains delicately sliced carrot sticks. The entire scene is bathed in the warm golden glow of the afternoon sun, which filters through the lush green canopy above, creating a beautiful play of light and shadow. The atmosphere is tranquil, warm and whimsical, a magical moment of friendship between children, pets and gentle woodland creatures.", + "seed": 8078, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_9147_seed_8078_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.035806365218013524, + "vlm_logprob": -0.035806365218013524, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.035806365218013524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007777763996273279 + ], + "opposite_logprob_tokens": [ + -5.007777690887451 + ], + "logprob_reward": -0.007777763996273279, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007777763996273279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23373427987098694, + 0.09158452600240707, + 0.05810201168060303, + 0.02434663102030754, + 0.14253543317317963, + 0.07735732942819595, + 0.09599236398935318, + 0.1517675817012787, + 0.08749248832464218, + 0.0370873399078846 + ], + "credit_map": [ + 0.0535954087972641, + 0.05835335701704025, + 0.041938744485378265, + 0.030917560681700706, + 0.014179096557199955, + 0.010255862027406693, + 0.010939840227365494, + 0.013554418459534645, + 0.021355269476771355, + 0.01826438307762146, + 0.018151244148612022, + 0.014504330232739449, + 0.0056120967492461205, + 0.004187386948615313, + 0.00446338951587677, + 0.005046424921602011, + 0.01879480481147766, + 0.015360306017100811, + 0.00779308658093214, + 0.0055905599147081375, + 0.002419875469058752, + 0.002665123902261257, + 0.0025336758699268103, + 0.0029445833060890436, + 0.00501299137249589, + 0.004336890764534473, + 0.003989840392023325, + 0.0032051517628133297, + 0.0017076865769922733, + 0.0018161506159231067, + 0.001989128766581416, + 0.002288789488375187, + 0.03904497250914574, + 0.017090555280447006, + 0.04903849586844444, + 0.01603836566209793, + 0.00497399503365159, + 0.00446668965741992, + 0.005362910684198141, + 0.006519454065710306, + 0.01642918772995472, + 0.013126040808856487, + 0.02120109088718891, + 0.010702395811676979, + 0.003878749441355467, + 0.003058900125324726, + 0.004925124812871218, + 0.004035838879644871, + 0.02587602101266384, + 0.020637672394514084, + 0.017261294648051262, + 0.008537494577467442, + 0.006449909880757332, + 0.006740374490618706, + 0.005858434364199638, + 0.0046311612240970135, + 0.01867208257317543, + 0.04327733814716339, + 0.0108290184289217, + 0.03521379455924034, + 0.011519902385771275, + 0.009990574792027473, + 0.01347927376627922, + 0.008785597048699856, + 0.01041884534060955, + 0.009497120976448059, + 0.015963874757289886, + 0.018163451924920082, + 0.00740483682602644, + 0.007820917293429375, + 0.010099162347614765, + 0.008124271407723427, + 0.00940781831741333, + 0.011861046776175499, + 0.0037498646415770054, + 0.003906079102307558, + 0.001677552587352693, + 0.0018547105137258768, + 0.0020175406243652105, + 0.0026127281598746777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.007777763996273279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097876638174057 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097876638174057, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097876638174057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0986848846077919, + 0.01981079950928688, + 0.008985255844891071, + 0.01900356076657772, + 0.1522878110408783, + 0.059555791318416595, + 0.3454660475254059, + 0.24425512552261353, + 0.03033054992556572, + 0.02162017673254013 + ], + "credit_map": [ + 0.022814380005002022, + 0.02050284668803215, + 0.008844616822898388, + 0.00717909773811698, + 0.014383596368134022, + 0.009305990301072598, + 0.007346202153712511, + 0.008308153599500656, + 0.0034573788288980722, + 0.003249669913202524, + 0.002170329447835684, + 0.0022310016211122274, + 0.0032401755452156067, + 0.001876896363683045, + 0.0018057177076116204, + 0.0017796296160668135, + 0.001753950142301619, + 0.0012973104603588581, + 0.0011921339901164174, + 0.0007593288901261985, + 0.0012977211736142635, + 0.0008481343975290656, + 0.0009626715327613056, + 0.0008740055491216481, + 0.0028804398607462645, + 0.003032330423593521, + 0.0026466313283890486, + 0.0022495752200484276, + 0.001716095837764442, + 0.0018207964021712542, + 0.001991648692637682, + 0.00266604358330369, + 0.05045672133564949, + 0.016356201842427254, + 0.0571732260286808, + 0.011367262341082096, + 0.0035661887377500534, + 0.0032870054710656404, + 0.004853531252592802, + 0.005227671004831791, + 0.016137508675456047, + 0.010254871100187302, + 0.01582903228700161, + 0.005461996886879206, + 0.002615764271467924, + 0.0024161548353731632, + 0.003821505233645439, + 0.003018957795575261, + 0.059046462178230286, + 0.124180369079113, + 0.08912825584411621, + 0.02373827062547207, + 0.011320064775645733, + 0.013409500941634178, + 0.011896084994077682, + 0.012747042812407017, + 0.06269057095050812, + 0.09413541108369827, + 0.0105653777718544, + 0.02518557198345661, + 0.014342736452817917, + 0.014161281287670135, + 0.013446343131363392, + 0.009727810509502888, + 0.0036758657079190016, + 0.003841996192932129, + 0.002354981144890189, + 0.007332028355449438, + 0.0030277788173407316, + 0.003683034097775817, + 0.0038627709727734327, + 0.002552095102146268, + 0.0028514943551272154, + 0.0036594003904610872, + 0.0023913683835417032, + 0.0028320420533418655, + 0.002243676455691457, + 0.002604997018352151, + 0.003055902197957039, + 0.0019812963437289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.0097876638174057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797164723277092 + ], + "opposite_logprob_tokens": [ + -3.529797077178955 + ], + "logprob_reward": -0.029797164723277092, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797164723277092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08036169409751892, + 0.030668603256344795, + 0.019722815603017807, + 0.07611435651779175, + 0.15251432359218597, + 0.09010287374258041, + 0.30343666672706604, + 0.15973205864429474, + 0.05599435418844223, + 0.03135228529572487 + ], + "credit_map": [ + 0.02203834056854248, + 0.0203801728785038, + 0.007334008812904358, + 0.005751022603362799, + 0.008561892434954643, + 0.005291284993290901, + 0.005047779064625502, + 0.0059571899473667145, + 0.00409974530339241, + 0.004475933499634266, + 0.006681658327579498, + 0.005961546208709478, + 0.002776716835796833, + 0.0018185764783993363, + 0.0021407438907772303, + 0.0027136835269629955, + 0.0037271531764417887, + 0.004286638926714659, + 0.0027200793847441673, + 0.0027213844005018473, + 0.001856091432273388, + 0.0014123732689768076, + 0.0013861326733604074, + 0.0016129611758515239, + 0.013228073716163635, + 0.01037952397018671, + 0.018734179437160492, + 0.013180471025407314, + 0.005174161866307259, + 0.004574107937514782, + 0.0060629150830209255, + 0.0047809225507080555, + 0.05039609968662262, + 0.016628442332148552, + 0.05685785785317421, + 0.011395986191928387, + 0.0034620719961822033, + 0.0034359246492385864, + 0.005018657073378563, + 0.005319280084222555, + 0.023937856778502464, + 0.016108132898807526, + 0.0228170994669199, + 0.008081240579485893, + 0.004313907120376825, + 0.003978059161454439, + 0.005945250391960144, + 0.004921325016766787, + 0.04797827824950218, + 0.10635452717542648, + 0.08343078196048737, + 0.02223196253180504, + 0.009740006178617477, + 0.011391944251954556, + 0.010872073471546173, + 0.01143707986921072, + 0.04039547219872475, + 0.04735877364873886, + 0.011802615597844124, + 0.02117645926773548, + 0.008878163062036037, + 0.009111820720136166, + 0.01286147627979517, + 0.008147281594574451, + 0.009421085938811302, + 0.0064477380365133286, + 0.0085642971098423, + 0.009968647733330727, + 0.00420730747282505, + 0.0045107402838766575, + 0.007166721858084202, + 0.00570781622081995, + 0.005063951481133699, + 0.005808690097182989, + 0.004864383954554796, + 0.006054725497961044, + 0.002448663581162691, + 0.0019599085208028555, + 0.0029696007259190083, + 0.0021823588758707047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.029797164723277092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703630328178406 + ], + "opposite_logprob_tokens": [ + -2.1270363330841064 + ], + "logprob_reward": -0.12703630328178406, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703630328178406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11519401520490646, + 0.01982419565320015, + 0.02652856893837452, + 0.01689894311130047, + 0.1391749083995819, + 0.05144611746072769, + 0.2739395201206207, + 0.310472697019577, + 0.02362046018242836, + 0.02290055714547634 + ], + "credit_map": [ + 0.017777200788259506, + 0.017598064616322517, + 0.01695568487048149, + 0.009900175034999847, + 0.01906299777328968, + 0.010518163442611694, + 0.011774593032896519, + 0.01160713192075491, + 0.0026311909314244986, + 0.002398871583864093, + 0.0016995301702991128, + 0.001976561499759555, + 0.0038753056433051825, + 0.0021674050949513912, + 0.002494858345016837, + 0.002580472035333514, + 0.003002244746312499, + 0.006198714952915907, + 0.007028876803815365, + 0.002926501678302884, + 0.002053267089650035, + 0.0017428885912522674, + 0.0018796485383063555, + 0.001696425024420023, + 0.002324581379070878, + 0.0021504575852304697, + 0.0034817976411432028, + 0.003051936626434326, + 0.0014420540537685156, + 0.0016505583189427853, + 0.0012334511848166585, + 0.0015641049249097705, + 0.04588774964213371, + 0.015860028564929962, + 0.05057656764984131, + 0.01038757897913456, + 0.003284833859652281, + 0.0034889401867985725, + 0.0048082806169986725, + 0.004880926571786404, + 0.011859065853059292, + 0.009526848793029785, + 0.011697350069880486, + 0.005594971124082804, + 0.0031945290975272655, + 0.0027480064891278744, + 0.003988842014223337, + 0.0028365044854581356, + 0.04850994423031807, + 0.10030780732631683, + 0.06884679198265076, + 0.017909787595272064, + 0.008442748337984085, + 0.01079097855836153, + 0.009373948909342289, + 0.009757519699633121, + 0.06299342215061188, + 0.11954294145107269, + 0.014325369149446487, + 0.055420126765966415, + 0.01744871214032173, + 0.016625910997390747, + 0.01426850538700819, + 0.009847697801887989, + 0.0034915837459266186, + 0.003083053743466735, + 0.004732448607683182, + 0.003833339549601078, + 0.0016308784252032638, + 0.001908332691527903, + 0.002932750852778554, + 0.0020080721005797386, + 0.0029279289301484823, + 0.0027765671256929636, + 0.0035534449853003025, + 0.004343743901699781, + 0.0018191976705566049, + 0.0019573515746742487, + 0.0036215386353433132, + 0.0019007831579074264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.12703630328178406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632930271327496 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004632930271327496, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632930271327496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3395155668258667, + 0.021284842863678932, + 0.019698908552527428, + 0.04731878638267517, + 0.1390516310930252, + 0.029218006879091263, + 0.1190527081489563, + 0.20126873254776, + 0.05137288570404053, + 0.032217830419540405 + ], + "credit_map": [ + 0.09330783784389496, + 0.09470809251070023, + 0.041496600955724716, + 0.03603922575712204, + 0.01835571601986885, + 0.017527449876070023, + 0.01444044429808855, + 0.023640258237719536, + 0.003761939937248826, + 0.004356088116765022, + 0.002781896386295557, + 0.002498095156624913, + 0.002130110515281558, + 0.0017371946014463902, + 0.001601586234755814, + 0.002417935524135828, + 0.002957736374810338, + 0.0027520449366420507, + 0.005070865154266357, + 0.0016405374044552445, + 0.001817341079004109, + 0.0016191336326301098, + 0.0018032053485512733, + 0.0020380476489663124, + 0.005487252492457628, + 0.00772396195679903, + 0.011203067377209663, + 0.009333034045994282, + 0.003038580296561122, + 0.002749909181147814, + 0.00412197457626462, + 0.003661014139652252, + 0.04574405401945114, + 0.014014467597007751, + 0.05330551415681839, + 0.010603143833577633, + 0.003129007760435343, + 0.003129405900835991, + 0.004336050245910883, + 0.004790022037923336, + 0.006055035162717104, + 0.0032467844430357218, + 0.008416373282670975, + 0.0038627192843705416, + 0.0014707063091918826, + 0.0016459593316540122, + 0.002052025869488716, + 0.002468406455591321, + 0.016124729067087173, + 0.03855632618069649, + 0.03525971248745918, + 0.009879453107714653, + 0.0046347351744771, + 0.0047279116697609425, + 0.004704685881733894, + 0.005165172275155783, + 0.06154775619506836, + 0.06326880306005478, + 0.015013182535767555, + 0.0225180983543396, + 0.008143050596117973, + 0.00864432193338871, + 0.01148289442062378, + 0.010650642216205597, + 0.007748734205961227, + 0.006539612542837858, + 0.006558082066476345, + 0.010074007324874401, + 0.004646517802029848, + 0.0051945531740784645, + 0.006161153316497803, + 0.00445023225620389, + 0.006660881452262402, + 0.007396732456982136, + 0.0037069059908390045, + 0.004152295179665089, + 0.002327831694856286, + 0.002594733377918601, + 0.0027144502382725477, + 0.0026640030555427074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.004632930271327496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium-wide shot captures the sheer enchantment of a sun-dappled forest clearing. Oliver, a young boy with tousled blond hair and a simple blue and white striped T-shirt, sits cross-legged on a red and white checkered picnic blanket. Next to him sits Sparky, a fluffy golden retriever puppy, his head tilted in curiosity, his tail gently tapping the soft grass. Nearby sits a family of four rabbits: a large brown dad, a gentle white mom, and two small, furry gray bunnies. Between them stands a whimsical tea party, featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups. Oliver, with a look of focused joy on his face, is in action, steadily and deliberately pouring imaginary tea from a small white teapot into a tiny cup, carefully held between the paws of a small rabbit. Another bunny nibbles at a heart-shaped cookie on a central plate, which also contains delicately sliced carrot sticks. The entire scene is bathed in the warm golden glow of the afternoon sun, which filters through the lush green canopy above, creating a beautiful play of light and shadow. The atmosphere is tranquil, warm and whimsical, a magical moment of friendship between children, pets and gentle woodland creatures.", + "seed": 8079, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_13/index_9147_seed_8079_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019287812570109964, + "vlm_logprob": -0.019287812570109964, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.019287812570109964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035329563543200493 + ], + "opposite_logprob_tokens": [ + -5.753532886505127 + ], + "logprob_reward": -0.0035329563543200493, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0035329563543200493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2901117205619812, + 0.18754149973392487, + 0.048583630472421646, + 0.08441270887851715, + 0.0477961003780365, + 0.029217414557933807, + 0.02547975443303585, + 0.18737708032131195, + 0.05140150338411331, + 0.04807856306433678 + ], + "credit_map": [ + 0.060858942568302155, + 0.05008423700928688, + 0.03972979262471199, + 0.05029061436653137, + 0.025517094880342484, + 0.019721481949090958, + 0.022575723007321358, + 0.021333837881684303, + 0.0259445458650589, + 0.030643755570054054, + 0.03204961493611336, + 0.03024989925324917, + 0.016101736575365067, + 0.01624237932264805, + 0.014446170069277287, + 0.021863393485546112, + 0.006683163344860077, + 0.005813252180814743, + 0.00609563896432519, + 0.011175919324159622, + 0.004089327994734049, + 0.00557964202016592, + 0.004894415847957134, + 0.004252274055033922, + 0.014556162059307098, + 0.008214307017624378, + 0.010308971628546715, + 0.02529146894812584, + 0.004578469321131706, + 0.009189515374600887, + 0.006607286166399717, + 0.0056665255688130856, + 0.005985180847346783, + 0.007054843008518219, + 0.005166485905647278, + 0.008391422219574451, + 0.005449812859296799, + 0.007033221423625946, + 0.00456327386200428, + 0.004151858855038881, + 0.0038573991041630507, + 0.0032068092841655016, + 0.0036868907045572996, + 0.005274069961160421, + 0.00371229974552989, + 0.004412363283336163, + 0.0026234895922243595, + 0.0024440926499664783, + 0.0036097883712500334, + 0.0031961356289684772, + 0.0027583602350205183, + 0.004898032173514366, + 0.0030407628510147333, + 0.0030466103926301003, + 0.002807692624628544, + 0.0021223712246865034, + 0.022477416321635246, + 0.0259970985352993, + 0.02783225290477276, + 0.054482121020555496, + 0.015621459111571312, + 0.01238242257386446, + 0.012848887592554092, + 0.015735426917672157, + 0.009221991524100304, + 0.008745437487959862, + 0.00671367859467864, + 0.005606650374829769, + 0.005033185239881277, + 0.00521639920771122, + 0.0067756581120193005, + 0.004088501445949078, + 0.009672528132796288, + 0.00972509291023016, + 0.006047497037798166, + 0.0071422625333070755, + 0.00371546414680779, + 0.0040788729675114155, + 0.004104544408619404, + 0.0035923009272664785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0035329563543200493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641822263598442 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641822263598442, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641822263598442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1738612800836563, + 0.12132514268159866, + 0.13897909224033356, + 0.1517355591058731, + 0.13347843289375305, + 0.0680774450302124, + 0.0489032082259655, + 0.06653659790754318, + 0.05352359265089035, + 0.043579645454883575 + ], + "credit_map": [ + 0.02693873643875122, + 0.023766150698065758, + 0.022118376567959785, + 0.020919326692819595, + 0.020643457770347595, + 0.019143473356962204, + 0.02081199176609516, + 0.01951976679265499, + 0.01819157972931862, + 0.012883148156106472, + 0.010591726750135422, + 0.012650908902287483, + 0.018559306859970093, + 0.021487334743142128, + 0.015142831020057201, + 0.011818310245871544, + 0.017719458788633347, + 0.01587015949189663, + 0.019482092931866646, + 0.03929360955953598, + 0.010558520443737507, + 0.011574785225093365, + 0.01357007585465908, + 0.01091039925813675, + 0.022089989855885506, + 0.015550698153674603, + 0.01567106507718563, + 0.04178454354405403, + 0.015105640515685081, + 0.014151690527796745, + 0.015031768940389156, + 0.012350148521363735, + 0.018267331644892693, + 0.011365098878741264, + 0.016086747869849205, + 0.04747267812490463, + 0.010351231321692467, + 0.013873732648789883, + 0.008638706989586353, + 0.007422896102070808, + 0.011004812084138393, + 0.006446304731070995, + 0.007271311711519957, + 0.010620230808854103, + 0.008924486115574837, + 0.010380868799984455, + 0.006315171252936125, + 0.007114260457456112, + 0.006223812699317932, + 0.0043504140339791775, + 0.0065254028886556625, + 0.00913421530276537, + 0.006985293235629797, + 0.00665021687746048, + 0.004759693518280983, + 0.004274159204214811, + 0.009654823690652847, + 0.005667452700436115, + 0.010554678738117218, + 0.012564018368721008, + 0.007018815726041794, + 0.008888090029358864, + 0.005999312736093998, + 0.006189397536218166, + 0.0125199593603611, + 0.0048954118974506855, + 0.004480638541281223, + 0.004212624859064817, + 0.005867741536349058, + 0.007591885980218649, + 0.008759666234254837, + 0.005195665173232555, + 0.007568400353193283, + 0.004719120915979147, + 0.006318274419754744, + 0.004062590654939413, + 0.006281825713813305, + 0.006869514472782612, + 0.003901685820892453, + 0.003858232870697975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008641822263598442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14470648765563965, + 0.07666020840406418, + 0.14674994349479675, + 0.2599394619464874, + 0.156488299369812, + 0.027048494666814804, + 0.07005834579467773, + 0.035648804157972336, + 0.05129331722855568, + 0.031406622380018234 + ], + "credit_map": [ + 0.02169332467019558, + 0.01908678188920021, + 0.015305288136005402, + 0.022314487025141716, + 0.018317991867661476, + 0.015107604674994946, + 0.014321967959403992, + 0.018559029325842857, + 0.01348861213773489, + 0.010227859020233154, + 0.006760251708328724, + 0.008246898651123047, + 0.00860445573925972, + 0.009466142393648624, + 0.008340886794030666, + 0.011525100097060204, + 0.020308522507548332, + 0.017196305096149445, + 0.02040962316095829, + 0.044232603162527084, + 0.01119184773415327, + 0.011385244317352772, + 0.011498681269586086, + 0.010527115315198898, + 0.040255866944789886, + 0.023518702015280724, + 0.030413813889026642, + 0.07754859328269958, + 0.020100221037864685, + 0.028843075037002563, + 0.02132951095700264, + 0.017929669469594955, + 0.02316603809595108, + 0.013280109502375126, + 0.019624654203653336, + 0.05131404101848602, + 0.010270089842379093, + 0.017513062804937363, + 0.01337213534861803, + 0.007948162965476513, + 0.0038115994539111853, + 0.003594588488340378, + 0.004792617168277502, + 0.0061142100021243095, + 0.002280545886605978, + 0.0023023642133921385, + 0.0018700661603361368, + 0.002282502595335245, + 0.011224930174648762, + 0.00751737505197525, + 0.008846783079206944, + 0.02043192833662033, + 0.005896922200918198, + 0.006747338455170393, + 0.005227981135249138, + 0.004165094345808029, + 0.0047049433924257755, + 0.00329335848800838, + 0.006525242235511541, + 0.008350402116775513, + 0.0034053856506943703, + 0.0037220034282654524, + 0.0025588779244571924, + 0.0030885895248502493, + 0.00929592177271843, + 0.005117378663271666, + 0.004925569985061884, + 0.004545085597783327, + 0.006228508893400431, + 0.006918111816048622, + 0.008915956132113934, + 0.005346788093447685, + 0.004008017014712095, + 0.004130009561777115, + 0.005888680461794138, + 0.004060000646859407, + 0.004063841886818409, + 0.003971721976995468, + 0.0026167603209614754, + 0.0026675895787775517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.011071558110415936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700051486492157 + ], + "opposite_logprob_tokens": [ + -2.695005178451538 + ], + "logprob_reward": -0.0700051486492157, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700051486492157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23830975592136383, + 0.071863554418087, + 0.07774641364812851, + 0.07728737592697144, + 0.12731599807739258, + 0.06952323019504547, + 0.12238390743732452, + 0.07337518781423569, + 0.09176738560199738, + 0.05042714625597 + ], + "credit_map": [ + 0.03539707884192467, + 0.034483104944229126, + 0.028329377993941307, + 0.029700934886932373, + 0.031951889395713806, + 0.0327928401529789, + 0.02325773425400257, + 0.022396806627511978, + 0.011172553524374962, + 0.006457554176449776, + 0.006207609083503485, + 0.007162600290030241, + 0.010727635584771633, + 0.012137722223997116, + 0.008296361193060875, + 0.009701521135866642, + 0.011586702428758144, + 0.008847239427268505, + 0.011018140241503716, + 0.0211391169577837, + 0.005561451427638531, + 0.008546518161892891, + 0.005879150237888098, + 0.005168096162378788, + 0.010653294622898102, + 0.008189212530851364, + 0.007877042517066002, + 0.007600727491080761, + 0.011850904673337936, + 0.012722115963697433, + 0.00894036702811718, + 0.009453711099922657, + 0.018219979479908943, + 0.010775317437946796, + 0.016680968925356865, + 0.04550311341881752, + 0.00844397209584713, + 0.012616715393960476, + 0.00874466821551323, + 0.006331278942525387, + 0.009366698563098907, + 0.0066236406564712524, + 0.007398000918328762, + 0.011101772077381611, + 0.010100350715219975, + 0.013214567676186562, + 0.00616751192137599, + 0.005550692789256573, + 0.02402152493596077, + 0.0127935241907835, + 0.013925863429903984, + 0.028985116630792618, + 0.011995768174529076, + 0.014198677614331245, + 0.010026254691183567, + 0.006437188945710659, + 0.012170584872364998, + 0.007474757265299559, + 0.01027174573391676, + 0.012795873917639256, + 0.007739656139165163, + 0.009304962120950222, + 0.006691272370517254, + 0.00692633306607604, + 0.01601077802479267, + 0.01067003421485424, + 0.014353428967297077, + 0.012418846599757671, + 0.008345249108970165, + 0.009708578698337078, + 0.009375249035656452, + 0.010885229334235191, + 0.008027352392673492, + 0.0051603978499770164, + 0.008360374718904495, + 0.00777654442936182, + 0.004237201064825058, + 0.004906967747956514, + 0.005363576114177704, + 0.0065947347320616245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0700051486492157 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest clearing with lush green canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Oliver, a young boy with tousled blond hair and a blue and white striped T-shirt, sitting cross-legged on a red and white checkered picnic blanket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sparky, a fluffy golden retriever puppy with a tilted head and tail tapping the grass, sitting next to Oliver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Oliver pouring imaginary tea from a small white teapot into a tiny cup held by a small rabbit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a whimsical tea party setup featuring a miniature ceramic tea set with pastel pink, sky blue, and lemon yellow teacups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003187577472999692 + ], + "opposite_logprob_tokens": [ + -5.753187656402588 + ], + "logprob_reward": -0.003187577472999692, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003187577472999692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20740646123886108, + 0.10719124227762222, + 0.13397525250911713, + 0.1627730131149292, + 0.06279812753200531, + 0.0850289985537529, + 0.09333562105894089, + 0.04234115779399872, + 0.04358243942260742, + 0.061567746102809906 + ], + "credit_map": [ + 0.029795052483677864, + 0.02042013593018055, + 0.023925388231873512, + 0.026906942948698997, + 0.02739076130092144, + 0.027274789288640022, + 0.024343151599168777, + 0.02735021710395813, + 0.014941269531846046, + 0.01497943326830864, + 0.012171430513262749, + 0.012578804977238178, + 0.014018665067851543, + 0.011487134732306004, + 0.014234865084290504, + 0.012779626995325089, + 0.021324332803487778, + 0.012103859335184097, + 0.01340519730001688, + 0.011999878101050854, + 0.018033456057310104, + 0.020506177097558975, + 0.020540112629532814, + 0.016062229871749878, + 0.0232541523873806, + 0.013032159768044949, + 0.01928156614303589, + 0.02690322883427143, + 0.02204129844903946, + 0.024795260280370712, + 0.016578081995248795, + 0.016887256875634193, + 0.007380988448858261, + 0.0071380361914634705, + 0.007224288769066334, + 0.013261866755783558, + 0.006612380966544151, + 0.0091317193582654, + 0.006031422410160303, + 0.006017419043928385, + 0.012245438061654568, + 0.008191617205739021, + 0.011393333785235882, + 0.013165823183953762, + 0.009076042100787163, + 0.01304267905652523, + 0.008385911583900452, + 0.009528148919343948, + 0.017303699627518654, + 0.008669332601130009, + 0.012403925880789757, + 0.024154046550393105, + 0.00841912068426609, + 0.010865136981010437, + 0.006123526021838188, + 0.005396829918026924, + 0.008481640368700027, + 0.004466000944375992, + 0.0059197633527219296, + 0.005384838208556175, + 0.0046957870945334435, + 0.004862520843744278, + 0.004040302708745003, + 0.004490303341299295, + 0.006960495840758085, + 0.005641452968120575, + 0.00543626956641674, + 0.004492251668125391, + 0.006856519728899002, + 0.006027499213814735, + 0.004101493861526251, + 0.004066454712301493, + 0.008736291900277138, + 0.007475376129150391, + 0.00820342916995287, + 0.0067705498076975346, + 0.00868399627506733, + 0.007914967834949493, + 0.006565155927091837, + 0.007217974867671728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.003187577472999692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_tensor([234, 234])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_tensor([234, 234])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6242dcf89aac370700a553f5c4a2d4cfcac72498 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_tensor([234, 234])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, low-angle tracking shot follows a heavy deep-sea submersible as it trudges through an abyssal plain, a crisscross of cool blue and dark green. The submersible boasts a heavy, matte black metal hull, its surface scarred and scuffed by the immense pressure of the water. A massive, multi-jointed robotic arm folds at one side, poised for action. A single, powerful searchlight on the submersible's front projects a concentrated, cone-shaped beam of cold white light, piercing the endless darkness. As its heavy tracks slowly and laboriously roll over the uneven, silty seabed, each movement stirs up vast swathes of turbid sand. Illuminated by the searchlight's beam, these stirred-up particles drift, swirl, and diffuse through the water in a distinctly slow-motion particle effect, forming a chaotic yet captivating nebula. The entire scene is steeped in a profound sense of isolation and oppressive silence, creating a melancholic and starkly realistic quality.", + "seed": 15244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_234_seed_15244_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08921768516302109, + "vlm_logprob": -0.08921768516302109, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.08921768516302109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12050735950469971 + ], + "opposite_logprob_tokens": [ + -2.37050724029541 + ], + "logprob_reward": -0.12050735950469971, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12050735950469971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214504987001419, + 0.09382885694503784, + 0.09428039193153381, + 0.15477290749549866, + 0.13636638224124908, + 0.08589481562376022, + 0.08935273438692093, + 0.0680115595459938, + 0.03450367972254753, + 0.12153813987970352 + ], + "credit_map": [ + 0.012454100884497166, + 0.02206285297870636, + 0.016267865896224976, + 0.011714152060449123, + 0.010432019829750061, + 0.015425527468323708, + 0.020068051293492317, + 0.013025932013988495, + 0.008394608274102211, + 0.011690239422023296, + 0.01068950816988945, + 0.015947753563523293, + 0.008913276717066765, + 0.007522379979491234, + 0.019887199625372887, + 0.010783908888697624, + 0.0073060267604887486, + 0.0181103702634573, + 0.008521570824086666, + 0.0125276455655694, + 0.00973266176879406, + 0.01228635385632515, + 0.012864265590906143, + 0.012931500561535358, + 0.012032447382807732, + 0.024040069431066513, + 0.016301298514008522, + 0.028352271765470505, + 0.015113999135792255, + 0.01861017569899559, + 0.023973848670721054, + 0.016348810866475105, + 0.014168152585625648, + 0.029543962329626083, + 0.016764825209975243, + 0.01397544052451849, + 0.014578649774193764, + 0.013117791153490543, + 0.022901354357600212, + 0.011316213756799698, + 0.0051595778204500675, + 0.007498357445001602, + 0.011042384430766106, + 0.019607126712799072, + 0.009439237415790558, + 0.008022590540349483, + 0.012597767636179924, + 0.01252778060734272, + 0.007787572685629129, + 0.013926084153354168, + 0.012429972179234028, + 0.015675421804189682, + 0.007184514310210943, + 0.008729887194931507, + 0.012608795426785946, + 0.011010492220520973, + 0.007585709448903799, + 0.012296723201870918, + 0.00849440973252058, + 0.007975790649652481, + 0.00706059904769063, + 0.0064120227470994, + 0.010085229761898518, + 0.00810108333826065, + 0.002704365411773324, + 0.003487624228000641, + 0.00488268630579114, + 0.007028655614703894, + 0.004952887073159218, + 0.0030712406150996685, + 0.00412910059094429, + 0.004247122909873724, + 0.01177592109888792, + 0.010104416869580746, + 0.02103869616985321, + 0.017736855894327164, + 0.01085012499243021, + 0.010984525084495544, + 0.01519309263676405, + 0.023854512721300125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.12050735950469971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043837368488312 + ], + "opposite_logprob_tokens": [ + -1.9104384183883667 + ], + "logprob_reward": -0.16043837368488312, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043837368488312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2664244771003723, + 0.08292027562856674, + 0.07358189672231674, + 0.03563717380166054, + 0.11698091775178909, + 0.05632000416517258, + 0.11428816616535187, + 0.05030330643057823, + 0.08530457317829132, + 0.11823917180299759 + ], + "credit_map": [ + 0.02216014452278614, + 0.030499720945954323, + 0.03684577718377113, + 0.032183438539505005, + 0.02842496708035469, + 0.026875201612710953, + 0.05514484643936157, + 0.03429039567708969, + 0.005991974379867315, + 0.011613867245614529, + 0.010123014450073242, + 0.01237607840448618, + 0.008876685984432697, + 0.00866425409913063, + 0.011964138597249985, + 0.0133102647960186, + 0.0065587712451815605, + 0.014033359475433826, + 0.008243406191468239, + 0.009105958044528961, + 0.010096272453665733, + 0.007811360061168671, + 0.010044917464256287, + 0.007687842007726431, + 0.003223156090825796, + 0.005651871208101511, + 0.004103078506886959, + 0.006306756287813187, + 0.0025963042862713337, + 0.003881138050928712, + 0.005349146667867899, + 0.004525721073150635, + 0.011250374838709831, + 0.023531638085842133, + 0.012340075336396694, + 0.019154775887727737, + 0.011129861697554588, + 0.01123003289103508, + 0.017009060829877853, + 0.011335101909935474, + 0.0035308985970914364, + 0.006383988074958324, + 0.006460538599640131, + 0.012948372401297092, + 0.0036360810045152903, + 0.0054144780151546, + 0.011174900457262993, + 0.00677074259147048, + 0.01292130071669817, + 0.01513825822621584, + 0.011702186428010464, + 0.016625287011265755, + 0.013200119137763977, + 0.0112998615950346, + 0.017773285508155823, + 0.015627872198820114, + 0.004516036249697208, + 0.005765037145465612, + 0.0057814293541014194, + 0.008970445953309536, + 0.00770100811496377, + 0.005592625588178635, + 0.00600172346457839, + 0.005975001957267523, + 0.006436792202293873, + 0.006524714175611734, + 0.00997269805520773, + 0.014665806666016579, + 0.010273482650518417, + 0.008852127008140087, + 0.011877063661813736, + 0.016701892018318176, + 0.011090166866779327, + 0.011883789673447609, + 0.015120056457817554, + 0.015894072130322456, + 0.012972772121429443, + 0.011448672041296959, + 0.01699899695813656, + 0.022830652073025703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.16043837368488312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900618761777878 + ], + "opposite_logprob_tokens": [ + -2.5790061950683594 + ], + "logprob_reward": -0.07900618761777878, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900618761777878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1159159243106842, + 0.10926572978496552, + 0.045934565365314484, + 0.04921354353427887, + 0.15338659286499023, + 0.09414853155612946, + 0.055489249527454376, + 0.034994591027498245, + 0.09180928766727448, + 0.24984200298786163 + ], + "credit_map": [ + 0.01090219896286726, + 0.022535039111971855, + 0.01686493679881096, + 0.014929638244211674, + 0.00933920405805111, + 0.013923615217208862, + 0.015656111761927605, + 0.011765160597860813, + 0.00829810369759798, + 0.01930956542491913, + 0.01561314333230257, + 0.013973685912787914, + 0.010490073822438717, + 0.015006947331130505, + 0.014553102664649487, + 0.012021102011203766, + 0.003446317743510008, + 0.007219524588435888, + 0.003991217352449894, + 0.005989281926304102, + 0.006514155771583319, + 0.004390738904476166, + 0.00713085662573576, + 0.007252472918480635, + 0.004603775218129158, + 0.007159536704421043, + 0.005139546003192663, + 0.006493137683719397, + 0.005339575931429863, + 0.005744690541177988, + 0.00675321277230978, + 0.007980069145560265, + 0.014444021508097649, + 0.03015146404504776, + 0.01576058939099312, + 0.025736916810274124, + 0.013297714293003082, + 0.014210298657417297, + 0.026890534907579422, + 0.012895039282739162, + 0.006031238008290529, + 0.011654382571578026, + 0.013104069977998734, + 0.020172374323010445, + 0.00683288648724556, + 0.008141622878611088, + 0.015558292157948017, + 0.012653659097850323, + 0.00620571244508028, + 0.009426729753613472, + 0.00677144480869174, + 0.009296000935137272, + 0.005313919857144356, + 0.005329744424670935, + 0.008248601108789444, + 0.00489709060639143, + 0.0035930706653743982, + 0.004737544804811478, + 0.003995147533714771, + 0.00536325853317976, + 0.0030066908802837133, + 0.003636748529970646, + 0.005501014646142721, + 0.005161113105714321, + 0.007895146496593952, + 0.007819581776857376, + 0.009557914920151234, + 0.012547130696475506, + 0.011952667497098446, + 0.010714803822338581, + 0.012291230261325836, + 0.019030803814530373, + 0.025770040228962898, + 0.023763952776789665, + 0.03643466904759407, + 0.02997155673801899, + 0.02175913006067276, + 0.022410256788134575, + 0.031807441264390945, + 0.057924941182136536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.07900618761777878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063972145318985 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.043063972145318985, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063972145318985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10136421024799347, + 0.1826397031545639, + 0.05603839457035065, + 0.029442112892866135, + 0.07421060651540756, + 0.09481565654277802, + 0.06375446170568466, + 0.16160279512405396, + 0.13510851562023163, + 0.10102354735136032 + ], + "credit_map": [ + 0.012788627296686172, + 0.015615000389516354, + 0.01589616760611534, + 0.012796073220670223, + 0.00740651972591877, + 0.012368937954306602, + 0.012580075301229954, + 0.011912817135453224, + 0.01426661666482687, + 0.028342202305793762, + 0.019212855026125908, + 0.029003141447901726, + 0.017289170995354652, + 0.016888603568077087, + 0.03520921990275383, + 0.022427882999181747, + 0.006695345044136047, + 0.007805481553077698, + 0.008929776027798653, + 0.006746404804289341, + 0.005733957979828119, + 0.005709330551326275, + 0.005745412781834602, + 0.008672681637108326, + 0.002979028970003128, + 0.004279180895537138, + 0.0032502657268196344, + 0.003616936504840851, + 0.002793462248519063, + 0.002893452998250723, + 0.003913187421858311, + 0.0057165976613759995, + 0.0070331343449652195, + 0.012808882631361485, + 0.007291642017662525, + 0.009378666989505291, + 0.00690927030518651, + 0.006303699221462011, + 0.014842660166323185, + 0.0096426485106349, + 0.006405011285096407, + 0.01243695430457592, + 0.009936156682670116, + 0.01856718398630619, + 0.009373682551085949, + 0.008153212256729603, + 0.017916211858391762, + 0.012027245946228504, + 0.0060994671657681465, + 0.010762807913124561, + 0.007551292423158884, + 0.009658250026404858, + 0.006745109800249338, + 0.006957727484405041, + 0.008312436752021313, + 0.0076673743315041065, + 0.012559034861624241, + 0.016115019097924232, + 0.01599925383925438, + 0.015768570825457573, + 0.03044932521879673, + 0.018247045576572418, + 0.021885862573981285, + 0.03057866543531418, + 0.009711423888802528, + 0.012777529656887054, + 0.014966337941586971, + 0.022840242832899094, + 0.013916078954935074, + 0.012229192070662975, + 0.01917116530239582, + 0.02949654683470726, + 0.009712360799312592, + 0.011233549565076828, + 0.009682904928922653, + 0.01229953859001398, + 0.012767380103468895, + 0.008787453174591064, + 0.01324890460819006, + 0.02329145185649395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.043063972145318985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307253286242485 + ], + "opposite_logprob_tokens": [ + -3.168072462081909 + ], + "logprob_reward": -0.04307253286242485, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307253286242485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19876791536808014, + 0.14942027628421783, + 0.038417648524045944, + 0.06825195997953415, + 0.07161646336317062, + 0.05266273021697998, + 0.061965212225914, + 0.0773715004324913, + 0.07533036917448044, + 0.2061958909034729 + ], + "credit_map": [ + 0.02445760741829872, + 0.03556937724351883, + 0.029810164123773575, + 0.022747503593564034, + 0.015136823058128357, + 0.02299269288778305, + 0.029582373797893524, + 0.018471384420990944, + 0.013343405909836292, + 0.016689535230398178, + 0.017220206558704376, + 0.02797107584774494, + 0.011443333700299263, + 0.01210255641490221, + 0.029091641306877136, + 0.021558521315455437, + 0.004105309024453163, + 0.005718149710446596, + 0.002914365381002426, + 0.00424611009657383, + 0.004412244074046612, + 0.003934124950319529, + 0.004901140462607145, + 0.00818620529025793, + 0.006496729329228401, + 0.010785187594592571, + 0.007374437525868416, + 0.01127209234982729, + 0.0067371162585914135, + 0.007102135103195906, + 0.010280660353600979, + 0.008203605189919472, + 0.006967334542423487, + 0.013670585118234158, + 0.009731817990541458, + 0.00822371430695057, + 0.007399980444461107, + 0.006949849426746368, + 0.008776228874921799, + 0.009896953590214252, + 0.004168362356722355, + 0.005595426075160503, + 0.007277887314558029, + 0.011148656718432903, + 0.004750980995595455, + 0.005241772625595331, + 0.008809669874608517, + 0.0056699723936617374, + 0.005647451151162386, + 0.010394972749054432, + 0.006486722733825445, + 0.0077658104710280895, + 0.006969215348362923, + 0.006658769212663174, + 0.009892357513308525, + 0.008149913512170315, + 0.008817856200039387, + 0.009202949702739716, + 0.00822474341839552, + 0.010550704784691334, + 0.0102896299213171, + 0.008763561025261879, + 0.009028940461575985, + 0.012493115849792957, + 0.0061720493249595165, + 0.006150585133582354, + 0.010515747591853142, + 0.014446745626628399, + 0.009089890867471695, + 0.006916946731507778, + 0.009317469783127308, + 0.012720940634608269, + 0.021270642057061195, + 0.020748749375343323, + 0.027914058417081833, + 0.032476771622896194, + 0.025839433073997498, + 0.019267691299319267, + 0.023137647658586502, + 0.035540904849767685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04307253286242485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle tracking shot follows a heavy deep-sea submersible as it trudges through an abyssal plain, a crisscross of cool blue and dark green. The submersible boasts a heavy, matte black metal hull, its surface scarred and scuffed by the immense pressure of the water. A massive, multi-jointed robotic arm folds at one side, poised for action. A single, powerful searchlight on the submersible's front projects a concentrated, cone-shaped beam of cold white light, piercing the endless darkness. As its heavy tracks slowly and laboriously roll over the uneven, silty seabed, each movement stirs up vast swathes of turbid sand. Illuminated by the searchlight's beam, these stirred-up particles drift, swirl, and diffuse through the water in a distinctly slow-motion particle effect, forming a chaotic yet captivating nebula. The entire scene is steeped in a profound sense of isolation and oppressive silence, creating a melancholic and starkly realistic quality.", + "seed": 15245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_234_seed_15245_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08796413317322731, + "vlm_logprob": -0.08796413317322731, + "vlm_yes_no_margin": 2.649999976158142, + "vlm_token_credit": -0.08796413317322731, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14479242265224457 + ], + "opposite_logprob_tokens": [ + -2.144792318344116 + ], + "logprob_reward": -0.14479242265224457, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.14479242265224457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08476433157920837, + 0.1734960824251175, + 0.07524619996547699, + 0.06209186092019081, + 0.10900114476680756, + 0.16838087141513824, + 0.17134177684783936, + 0.057902637869119644, + 0.02568330243229866, + 0.07209178805351257 + ], + "credit_map": [ + 0.018791770562529564, + 0.013209383934736252, + 0.009175864048302174, + 0.010194892063736916, + 0.011527018621563911, + 0.011283330619335175, + 0.0053415545262396336, + 0.0052405172027647495, + 0.008421155624091625, + 0.008872934617102146, + 0.006797996815294027, + 0.004749921150505543, + 0.07207845896482468, + 0.065299391746521, + 0.004258136730641127, + 0.003018092131242156, + 0.011053727008402348, + 0.015322737395763397, + 0.007877997122704983, + 0.007393976673483849, + 0.014067061245441437, + 0.012688454240560532, + 0.0034121135249733925, + 0.0034301336854696274, + 0.008573050610721111, + 0.011160177178680897, + 0.010142805986106396, + 0.010334765538573265, + 0.007194523233920336, + 0.008554970845580101, + 0.003615827299654484, + 0.0025157418567687273, + 0.011312038637697697, + 0.015213990584015846, + 0.010732508264482021, + 0.009033394046127796, + 0.02148968167603016, + 0.01741177774965763, + 0.01410415954887867, + 0.009703590534627438, + 0.025390710681676865, + 0.027395086362957954, + 0.021800026297569275, + 0.023605670779943466, + 0.023004744201898575, + 0.02215694822371006, + 0.012487543746829033, + 0.012540130876004696, + 0.035880476236343384, + 0.026765193790197372, + 0.02140987291932106, + 0.01704581268131733, + 0.0231938436627388, + 0.023150663822889328, + 0.012767587788403034, + 0.011128315702080727, + 0.005628205835819244, + 0.006819899193942547, + 0.007636881433427334, + 0.008686678484082222, + 0.0061620245687663555, + 0.008467196486890316, + 0.006626191083341837, + 0.007875560782849789, + 0.0034770246129482985, + 0.003451368073001504, + 0.00366584793664515, + 0.0034411109518259764, + 0.002897383412346244, + 0.0032499253284186125, + 0.002850313438102603, + 0.0026503275148570538, + 0.011526803486049175, + 0.010738073848187923, + 0.008499708026647568, + 0.012948811985552311, + 0.00705353356897831, + 0.008358119055628777, + 0.00781241012737155, + 0.005154321435838938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.14479242265224457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14289593696594238 + ], + "opposite_logprob_tokens": [ + -2.0178959369659424 + ], + "logprob_reward": -0.14289593696594238, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14289593696594238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056455060839653015, + 0.13801176846027374, + 0.03897342085838318, + 0.05283070355653763, + 0.20121672749519348, + 0.045533064752817154, + 0.06055823713541031, + 0.15736037492752075, + 0.10068179666996002, + 0.1483788937330246 + ], + "credit_map": [ + 0.010171833448112011, + 0.008910010568797588, + 0.008060604333877563, + 0.007767707575112581, + 0.006606230046600103, + 0.0076850359328091145, + 0.00397046934813261, + 0.0032831679563969374, + 0.006378768477588892, + 0.006074521690607071, + 0.005727540235966444, + 0.005655601155012846, + 0.056818265467882156, + 0.05184787139296532, + 0.0030395605135709047, + 0.0024696362670511007, + 0.00591302989050746, + 0.009348064661026001, + 0.004651164170354605, + 0.003898163791745901, + 0.005238007754087448, + 0.005143861752003431, + 0.0024007754400372505, + 0.0023803566582500935, + 0.0063398429192602634, + 0.008378240279853344, + 0.009406959637999535, + 0.008895049802958965, + 0.005854127928614616, + 0.007928997278213501, + 0.003400644287467003, + 0.002626837231218815, + 0.015631532296538353, + 0.026804668828845024, + 0.012862433679401875, + 0.010938129387795925, + 0.06310273706912994, + 0.05196777731180191, + 0.010475502349436283, + 0.009433942846953869, + 0.006359013728797436, + 0.006001480855047703, + 0.005771306809037924, + 0.00730849290266633, + 0.005852430127561092, + 0.006392118986696005, + 0.00431677745655179, + 0.0035314448177814484, + 0.012168736197054386, + 0.00719124311581254, + 0.005443150643259287, + 0.00608133664354682, + 0.011288253590464592, + 0.009108737111091614, + 0.00410488760098815, + 0.005171894561499357, + 0.020128589123487473, + 0.02869546227157116, + 0.020513541996479034, + 0.019158443436026573, + 0.0226649958640337, + 0.019573619589209557, + 0.016073761507868767, + 0.010551955550909042, + 0.008705656975507736, + 0.012229301035404205, + 0.018178850412368774, + 0.018322443589568138, + 0.012123440392315388, + 0.013543196022510529, + 0.011725593358278275, + 0.0058533078990876675, + 0.02321021258831024, + 0.023752611130475998, + 0.022808054462075233, + 0.01881472021341324, + 0.013113226741552353, + 0.017406277358531952, + 0.01641593873500824, + 0.012857848778367043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.14289593696594238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497325211763382 + ], + "opposite_logprob_tokens": [ + -2.9299733638763428 + ], + "logprob_reward": -0.05497325211763382, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497325211763382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05563169717788696, + 0.08229681849479675, + 0.1386699229478836, + 0.08653254061937332, + 0.09943521022796631, + 0.0711543932557106, + 0.05654894560575485, + 0.11597021669149399, + 0.12204544991254807, + 0.17171478271484375 + ], + "credit_map": [ + 0.013370173051953316, + 0.010977573692798615, + 0.006212609354406595, + 0.00481165898963809, + 0.007013087626546621, + 0.006637577433139086, + 0.002882177708670497, + 0.0037268411833792925, + 0.012828880921006203, + 0.007574756629765034, + 0.006021786481142044, + 0.0049827126786112785, + 0.022837089374661446, + 0.021160418167710304, + 0.0034441782627254725, + 0.0034469941165298223, + 0.018712488934397697, + 0.030678613111376762, + 0.015033735893666744, + 0.015786945819854736, + 0.023778868839144707, + 0.022204602137207985, + 0.0074447584338486195, + 0.005029914900660515, + 0.010633008554577827, + 0.013658395037055016, + 0.015515663661062717, + 0.015129100531339645, + 0.009349439293146133, + 0.012391019612550735, + 0.006067713722586632, + 0.003788200207054615, + 0.010443117469549179, + 0.013053812086582184, + 0.00678429938852787, + 0.006414162460714579, + 0.02851659432053566, + 0.023136530071496964, + 0.005419660359621048, + 0.0056670368649065495, + 0.011190420016646385, + 0.011281709186732769, + 0.009123268537223339, + 0.009191201068460941, + 0.010007748380303383, + 0.010047812946140766, + 0.005655009299516678, + 0.004657224286347628, + 0.01089519914239645, + 0.00887118186801672, + 0.0066041117534041405, + 0.005695009604096413, + 0.007477073930203915, + 0.008385371416807175, + 0.004681079648435116, + 0.0039399233646690845, + 0.010701115243136883, + 0.014169971458613873, + 0.020977269858121872, + 0.02559858374297619, + 0.011938951909542084, + 0.013785943388938904, + 0.01057458110153675, + 0.008223806507885456, + 0.011554908007383347, + 0.01519237831234932, + 0.02653876505792141, + 0.01969342865049839, + 0.011557533405721188, + 0.01693970151245594, + 0.012515854090452194, + 0.008052886463701725, + 0.021796120330691338, + 0.022981449961662292, + 0.028336353600025177, + 0.03851938247680664, + 0.013635341078042984, + 0.018159998580813408, + 0.0180632583796978, + 0.010222884826362133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.05497325211763382 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186133354902267 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186133354902267, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186133354902267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05303294211626053, + 0.1111537367105484, + 0.06324058771133423, + 0.12457585334777832, + 0.3845217227935791, + 0.059257201850414276, + 0.03188261017203331, + 0.04141570255160332, + 0.06382308900356293, + 0.0670965388417244 + ], + "credit_map": [ + 0.010776841081678867, + 0.008396397344768047, + 0.005433585494756699, + 0.005195441655814648, + 0.0072459266521036625, + 0.007606325671076775, + 0.0035024017561227083, + 0.004876023158431053, + 0.005704588256776333, + 0.005407330580055714, + 0.0051465206779539585, + 0.006181271746754646, + 0.043526966124773026, + 0.0372929722070694, + 0.004210202023386955, + 0.0036838797386735678, + 0.01035694032907486, + 0.010470259003341198, + 0.008422181010246277, + 0.007670605089515448, + 0.009695040062069893, + 0.00873580388724804, + 0.004449352156370878, + 0.0034404005855321884, + 0.01630469784140587, + 0.01580371893942356, + 0.014379797503352165, + 0.010921118780970573, + 0.021637650206685066, + 0.021998250856995583, + 0.008028479292988777, + 0.015502138994634151, + 0.031377192586660385, + 0.03609200194478035, + 0.031365714967250824, + 0.031970471143722534, + 0.07043713331222534, + 0.07050871104001999, + 0.04019327089190483, + 0.07257721573114395, + 0.00965832732617855, + 0.010567118413746357, + 0.007695547770708799, + 0.007708650082349777, + 0.00641627749428153, + 0.00796376634389162, + 0.005096651613712311, + 0.0041508604772388935, + 0.005727546289563179, + 0.004617589060217142, + 0.005665719974786043, + 0.0033243861980736256, + 0.0034231117460876703, + 0.0035277639981359243, + 0.00323416106402874, + 0.0023623318411409855, + 0.0037192474119365215, + 0.004600194748491049, + 0.005914807319641113, + 0.008192365989089012, + 0.004037071019411087, + 0.006260963622480631, + 0.0043128300458192825, + 0.004378220997750759, + 0.0059299892745912075, + 0.008132304064929485, + 0.013375775888562202, + 0.01093545462936163, + 0.006684497464448214, + 0.008829329162836075, + 0.006314177997410297, + 0.0036215553991496563, + 0.009026841260492802, + 0.00889639277011156, + 0.010462664999067783, + 0.014028958976268768, + 0.005827941931784153, + 0.0071316189132630825, + 0.006869244854897261, + 0.004852882586419582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.018186133354902267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897292077541351 + ], + "opposite_logprob_tokens": [ + -2.578972816467285 + ], + "logprob_reward": -0.07897292077541351, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897292077541351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06773339956998825, + 0.22606933116912842, + 0.08954587578773499, + 0.06650428473949432, + 0.0995529443025589, + 0.08189020305871964, + 0.07274191081523895, + 0.05976136773824692, + 0.05746312811970711, + 0.1787375807762146 + ], + "credit_map": [ + 0.009317097254097462, + 0.00948911439627409, + 0.007330967113375664, + 0.006847356911748648, + 0.012440470978617668, + 0.013849707320332527, + 0.003933291416615248, + 0.004525387194007635, + 0.008974248543381691, + 0.01004884298890829, + 0.010021365247666836, + 0.006842677015811205, + 0.09421250969171524, + 0.08577559888362885, + 0.006006204057484865, + 0.004187866114079952, + 0.013098263181746006, + 0.01662660390138626, + 0.009725015610456467, + 0.0073690409772098064, + 0.019150391221046448, + 0.018469639122486115, + 0.0028919123578816652, + 0.002215003129094839, + 0.007489966228604317, + 0.010034078732132912, + 0.01089764479547739, + 0.010591473430395126, + 0.00820112507790327, + 0.010594612918794155, + 0.005158640444278717, + 0.003536738222464919, + 0.007769930642098188, + 0.011831714771687984, + 0.008932431228458881, + 0.007657608017325401, + 0.024510763585567474, + 0.020993506535887718, + 0.009050991386175156, + 0.008805986493825912, + 0.010786473751068115, + 0.01215416006743908, + 0.010489000007510185, + 0.010729999281466007, + 0.010209721513092518, + 0.011184045113623142, + 0.009419655427336693, + 0.006917149294167757, + 0.011578120291233063, + 0.00912316795438528, + 0.007314951159060001, + 0.01054409146308899, + 0.010700601153075695, + 0.011082577519118786, + 0.008155989460647106, + 0.004242411348968744, + 0.005784762091934681, + 0.007579794619232416, + 0.008559546433389187, + 0.01166978944092989, + 0.0060998862609267235, + 0.008514193817973137, + 0.00628443481400609, + 0.0052689556032419205, + 0.005830420181155205, + 0.006744211073964834, + 0.010621413588523865, + 0.009309285320341587, + 0.006719628814607859, + 0.0082943644374609, + 0.006028007250279188, + 0.003915797919034958, + 0.02397235855460167, + 0.02388489432632923, + 0.026304028928279877, + 0.032110363245010376, + 0.017473630607128143, + 0.021029045805335045, + 0.021995630115270615, + 0.011967619881033897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.07897292077541351 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle tracking shot follows a heavy deep-sea submersible as it trudges through an abyssal plain, a crisscross of cool blue and dark green. The submersible boasts a heavy, matte black metal hull, its surface scarred and scuffed by the immense pressure of the water. A massive, multi-jointed robotic arm folds at one side, poised for action. A single, powerful searchlight on the submersible's front projects a concentrated, cone-shaped beam of cold white light, piercing the endless darkness. As its heavy tracks slowly and laboriously roll over the uneven, silty seabed, each movement stirs up vast swathes of turbid sand. Illuminated by the searchlight's beam, these stirred-up particles drift, swirl, and diffuse through the water in a distinctly slow-motion particle effect, forming a chaotic yet captivating nebula. The entire scene is steeped in a profound sense of isolation and oppressive silence, creating a melancholic and starkly realistic quality.", + "seed": 15246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_234_seed_15246_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.070651950314641, + "vlm_logprob": -0.070651950314641, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.070651950314641, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09471169859170914 + ], + "opposite_logprob_tokens": [ + -2.5947117805480957 + ], + "logprob_reward": -0.09471169859170914, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09471169859170914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042032063007354736, + 0.05033483728766441, + 0.027635840699076653, + 0.02313641458749771, + 0.3434286117553711, + 0.03461433947086334, + 0.03752904012799263, + 0.03590615093708038, + 0.09537474066019058, + 0.3100079596042633 + ], + "credit_map": [ + 0.006599021144211292, + 0.003704422153532505, + 0.009729296900331974, + 0.0034606943372637033, + 0.0038883748929947615, + 0.006471978034824133, + 0.004246085416525602, + 0.003932190593332052, + 0.011397900059819221, + 0.0029466089326888323, + 0.016454357653856277, + 0.004055195488035679, + 0.0023791235871613026, + 0.0036190173123031855, + 0.004128107335418463, + 0.005354527849704027, + 0.0064621735364198685, + 0.0020931654144078493, + 0.009868706576526165, + 0.0028472121339291334, + 0.0015831778291612864, + 0.0017773892031982541, + 0.0014847948914393783, + 0.0015192223945632577, + 0.004504232667386532, + 0.0015278467908501625, + 0.0073926979675889015, + 0.002462603384628892, + 0.002536038402467966, + 0.0017163282027468085, + 0.001470461254939437, + 0.0015262061497196555, + 0.042136549949645996, + 0.02047405019402504, + 0.06490365415811539, + 0.13571040332317352, + 0.020209133625030518, + 0.02012382075190544, + 0.024653678759932518, + 0.015217323787510395, + 0.0034269464667886496, + 0.002721080556511879, + 0.010168613865971565, + 0.0035146235022693872, + 0.005159592255949974, + 0.003200855804607272, + 0.003519688732922077, + 0.0029029403813183308, + 0.0019425220089033246, + 0.0030510881915688515, + 0.002428706968203187, + 0.0019991560839116573, + 0.010756423696875572, + 0.009701436385512352, + 0.0028098192997276783, + 0.004839890170842409, + 0.0018173932330682874, + 0.0038935565389692783, + 0.0032492117024958134, + 0.002353247720748186, + 0.009488792158663273, + 0.009154529310762882, + 0.0022578504867851734, + 0.0036915699020028114, + 0.00472288578748703, + 0.01078515499830246, + 0.007049445062875748, + 0.006626144051551819, + 0.02632707916200161, + 0.027749767526984215, + 0.00586236035451293, + 0.00625190744176507, + 0.016786666586995125, + 0.01673070713877678, + 0.014803054742515087, + 0.014753232710063457, + 0.10552286356687546, + 0.11313973367214203, + 0.009900158271193504, + 0.01837152987718582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.09471169859170914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12712112069129944 + ], + "opposite_logprob_tokens": [ + -2.1271212100982666 + ], + "logprob_reward": -0.12712112069129944, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12712112069129944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04605785757303238, + 0.014154380187392235, + 0.00904842559248209, + 0.06681117415428162, + 0.06485588103532791, + 0.019413644447922707, + 0.025255177170038223, + 0.02857312746345997, + 0.05480051040649414, + 0.6710298657417297 + ], + "credit_map": [ + 0.01499105803668499, + 0.001981248613446951, + 0.01929960958659649, + 0.0028713871724903584, + 0.0017663517501205206, + 0.002280748449265957, + 0.0012640302302315831, + 0.0016034264117479324, + 0.00278456904925406, + 0.0011269609676674008, + 0.0042114402167499065, + 0.0016109851421788335, + 0.0006009260541759431, + 0.0014483596896752715, + 0.0008636010461486876, + 0.00150753790512681, + 0.0013940987410023808, + 0.0010121650993824005, + 0.002117048017680645, + 0.0013788157375529408, + 0.0007862253114581108, + 0.0008248949307017028, + 0.0006779633113183081, + 0.0008572139777243137, + 0.0153189767152071, + 0.00315229338593781, + 0.026440519839525223, + 0.00783666130155325, + 0.005951996892690659, + 0.002892045071348548, + 0.0025920472107827663, + 0.002626631874591112, + 0.00796069111675024, + 0.003906744532287121, + 0.012010381557047367, + 0.025790361687541008, + 0.004013201221823692, + 0.0034927900414913893, + 0.00477868365123868, + 0.0029030267614871264, + 0.0022636742796748877, + 0.0027207157108932734, + 0.005487489514052868, + 0.0019158958457410336, + 0.0021476128604263067, + 0.0015710354782640934, + 0.001525246538221836, + 0.0017819736385717988, + 0.0012736266944557428, + 0.0023664047475904226, + 0.0017094265203922987, + 0.0015508712967857718, + 0.006577904801815748, + 0.005725829862058163, + 0.0021325566340237856, + 0.003918556962162256, + 0.0017389567801728845, + 0.004812540020793676, + 0.002312521217390895, + 0.0027785201091319323, + 0.006574301049113274, + 0.005080464296042919, + 0.002477138303220272, + 0.002798685571178794, + 0.0021187125239521265, + 0.0020220086444169283, + 0.002513463841751218, + 0.002287788549438119, + 0.02039480395615101, + 0.020768191665410995, + 0.002250140067189932, + 0.0024453981313854456, + 0.035551488399505615, + 0.03427949547767639, + 0.026098890230059624, + 0.024234941229224205, + 0.2396964132785797, + 0.25389981269836426, + 0.016354454681277275, + 0.0409143827855587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.12712112069129944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003993541002274 + ], + "opposite_logprob_tokens": [ + -2.695039987564087 + ], + "logprob_reward": -0.07003993541002274, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003993541002274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5332856178283691, + 0.06719804555177689, + 0.022539906203746796, + 0.05071311816573143, + 0.10469386726617813, + 0.031604740768671036, + 0.016336822882294655, + 0.03982575237751007, + 0.04543951898813248, + 0.0883626714348793 + ], + "credit_map": [ + 0.16766248643398285, + 0.02274266444146633, + 0.21263568103313446, + 0.032153479754924774, + 0.018020978197455406, + 0.04005780071020126, + 0.015274877659976482, + 0.02473762445151806, + 0.017301924526691437, + 0.0036536753177642822, + 0.02399377152323723, + 0.006187588442116976, + 0.002771675353869796, + 0.0046013882383704185, + 0.0033031816128641367, + 0.0053848340176045895, + 0.002606533234938979, + 0.003211322706192732, + 0.004742112942039967, + 0.0038140208926051855, + 0.0024332276079803705, + 0.0022262788843363523, + 0.0019474076107144356, + 0.0015589998802170157, + 0.00904426071792841, + 0.003335413057357073, + 0.014646138995885849, + 0.004691973328590393, + 0.0078031145967543125, + 0.005076361820101738, + 0.0032532322220504284, + 0.0028626148123294115, + 0.011233877390623093, + 0.010127564892172813, + 0.016228606924414635, + 0.0311374980956316, + 0.01062102522701025, + 0.011283112689852715, + 0.007949975319206715, + 0.006112204864621162, + 0.0027308135759085417, + 0.004702831152826548, + 0.005715415347367525, + 0.0029294900596141815, + 0.005058352369815111, + 0.004264930263161659, + 0.0033546737395226955, + 0.002848230767995119, + 0.0009944988414645195, + 0.0016954700695350766, + 0.00118750031106174, + 0.0013425674987956882, + 0.004050636198371649, + 0.003580480581149459, + 0.0014311778359115124, + 0.0020544889848679304, + 0.0027288757264614105, + 0.007392182480543852, + 0.0036537915002554655, + 0.0035644378513097763, + 0.0065346360206604, + 0.005664172116667032, + 0.003975955303758383, + 0.006311699748039246, + 0.0018268452258780599, + 0.002955894684419036, + 0.002918990794569254, + 0.002400279277935624, + 0.014538327232003212, + 0.014908441342413425, + 0.0021732249297201633, + 0.003717513522133231, + 0.004439790267497301, + 0.00470155943185091, + 0.004604227840900421, + 0.005116424988955259, + 0.028899597004055977, + 0.03138688579201698, + 0.0033027969766408205, + 0.005911378189921379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.07003993541002274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289239034056664 + ], + "opposite_logprob_tokens": [ + -3.773289203643799 + ], + "logprob_reward": -0.023289239034056664, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289239034056664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1605326533317566, + 0.045058149844408035, + 0.09810733050107956, + 0.35396498441696167, + 0.1549900323152542, + 0.021451998502016068, + 0.04551983252167702, + 0.015355942770838737, + 0.012442841194570065, + 0.09257612377405167 + ], + "credit_map": [ + 0.05087917298078537, + 0.007761260960251093, + 0.06525404751300812, + 0.010148883797228336, + 0.00468419911339879, + 0.009600275196135044, + 0.004258774220943451, + 0.007946047000586987, + 0.00847647525370121, + 0.003464357927441597, + 0.01275715697556734, + 0.004299953579902649, + 0.003273948561400175, + 0.003758825361728668, + 0.00285525806248188, + 0.00617217505350709, + 0.0135856494307518, + 0.007716446649283171, + 0.02191878855228424, + 0.010944993235170841, + 0.014897864311933517, + 0.013152871280908585, + 0.00912985485047102, + 0.00676086125895381, + 0.07639981806278229, + 0.01738717593252659, + 0.133392333984375, + 0.043633367866277695, + 0.028452308848500252, + 0.020252864807844162, + 0.01682906039059162, + 0.017618052661418915, + 0.01769806630909443, + 0.008911600336432457, + 0.03241139277815819, + 0.06286104768514633, + 0.008113302290439606, + 0.00753813935443759, + 0.011882814578711987, + 0.005573676433414221, + 0.001982499612495303, + 0.004458488430827856, + 0.004698844160884619, + 0.002071379218250513, + 0.002746943151578307, + 0.002438873518258333, + 0.0017137163085862994, + 0.0013412553817033768, + 0.002407487016171217, + 0.0035083526745438576, + 0.0022179370280355215, + 0.002105027437210083, + 0.01633262261748314, + 0.013227146118879318, + 0.0022784294560551643, + 0.003442835295572877, + 0.0009546116925776005, + 0.0023133105132728815, + 0.00199087243527174, + 0.002030974254012108, + 0.002721313387155533, + 0.002264605835080147, + 0.0015531341778114438, + 0.0015271188458427787, + 0.0008159266435541213, + 0.0012950325617566705, + 0.0011616196716204286, + 0.001092818914912641, + 0.002798400353640318, + 0.0027869418263435364, + 0.0009648530394770205, + 0.001527248416095972, + 0.004752205219119787, + 0.005862169899046421, + 0.004579661879688501, + 0.004461810924112797, + 0.030788475647568703, + 0.03242053464055061, + 0.0030489207711070776, + 0.006662351079285145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.023289239034056664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809775784611702 + ], + "opposite_logprob_tokens": [ + -3.288097858428955 + ], + "logprob_reward": -0.03809775784611702, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809775784611702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09805282950401306, + 0.06760118901729584, + 0.015110209584236145, + 0.01730433851480484, + 0.4832090437412262, + 0.04368161782622337, + 0.03818899393081665, + 0.028088241815567017, + 0.0845164954662323, + 0.12424702942371368 + ], + "credit_map": [ + 0.03199565038084984, + 0.004142054356634617, + 0.04048344865441322, + 0.007324117235839367, + 0.0033370491582900286, + 0.004311465658247471, + 0.0021986335050314665, + 0.00426041055470705, + 0.015865761786699295, + 0.004330629948526621, + 0.02309253439307213, + 0.0060804374516010284, + 0.00246189022436738, + 0.00452210521325469, + 0.0033195228315889835, + 0.007928312756121159, + 0.0027082490269094706, + 0.0010148086585104465, + 0.004670286551117897, + 0.0018748161382973194, + 0.0014135909732431173, + 0.0015396384987980127, + 0.0010640159016475081, + 0.0008248049416579306, + 0.0019863785710185766, + 0.0014329921687021852, + 0.0038854971062391996, + 0.0023447899147868156, + 0.0033133579418063164, + 0.0014228395884856582, + 0.0013911675196141005, + 0.001527315704151988, + 0.0587262324988842, + 0.03283068537712097, + 0.08631327748298645, + 0.18370291590690613, + 0.03096570447087288, + 0.03260162100195885, + 0.03501349687576294, + 0.023055074736475945, + 0.003635867265984416, + 0.005569668486714363, + 0.009677782654762268, + 0.004383154679089785, + 0.006973647978156805, + 0.005128341261297464, + 0.004147739615291357, + 0.004165412858128548, + 0.001977956620976329, + 0.0030389355961233377, + 0.002029056428000331, + 0.002046978799626231, + 0.0130175044760108, + 0.010909901931881905, + 0.0018691809382289648, + 0.0032994786743074656, + 0.0018961097812280059, + 0.005181463900953531, + 0.0031056071165949106, + 0.0030287476256489754, + 0.004469076171517372, + 0.003934950102120638, + 0.0033169975504279137, + 0.0031552889849990606, + 0.00419344287365675, + 0.009322317317128181, + 0.0069731492549180984, + 0.00644669821485877, + 0.022892259061336517, + 0.02248489111661911, + 0.0062301140278577805, + 0.005973626859486103, + 0.006707287393510342, + 0.009837000630795956, + 0.00660054013133049, + 0.00656620878726244, + 0.04030497744679451, + 0.04211404547095299, + 0.004102027043700218, + 0.008014940656721592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.03809775784611702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle tracking shot follows a heavy deep-sea submersible as it trudges through an abyssal plain, a crisscross of cool blue and dark green. The submersible boasts a heavy, matte black metal hull, its surface scarred and scuffed by the immense pressure of the water. A massive, multi-jointed robotic arm folds at one side, poised for action. A single, powerful searchlight on the submersible's front projects a concentrated, cone-shaped beam of cold white light, piercing the endless darkness. As its heavy tracks slowly and laboriously roll over the uneven, silty seabed, each movement stirs up vast swathes of turbid sand. Illuminated by the searchlight's beam, these stirred-up particles drift, swirl, and diffuse through the water in a distinctly slow-motion particle effect, forming a chaotic yet captivating nebula. The entire scene is steeped in a profound sense of isolation and oppressive silence, creating a melancholic and starkly realistic quality.", + "seed": 15247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_14/index_234_seed_15247_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06208038404583931, + "vlm_logprob": -0.06208038404583931, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.06208038404583931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08997481316328049 + ], + "opposite_logprob_tokens": [ + -2.589974880218506 + ], + "logprob_reward": -0.08997481316328049, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08997481316328049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7100163698196411, + 0.019951796159148216, + 0.015619542449712753, + 0.023660173639655113, + 0.0173039510846138, + 0.13755342364311218, + 0.03887180984020233, + 0.012804881669580936, + 0.011331071145832539, + 0.012886960059404373 + ], + "credit_map": [ + 0.2078331708908081, + 0.047100987285375595, + 0.31098660826683044, + 0.05228723958134651, + 0.016932442784309387, + 0.04493728280067444, + 0.016064178198575974, + 0.013874492608010769, + 0.0030810139141976833, + 0.00348502304404974, + 0.00451717060059309, + 0.002204886870458722, + 0.00218199728988111, + 0.001580036710947752, + 0.0016221388941630721, + 0.0012795303482562304, + 0.003235119627788663, + 0.0010750873479992151, + 0.00412394292652607, + 0.0017299584578722715, + 0.0020802796352654696, + 0.0008458586526103318, + 0.0015987138031050563, + 0.0009305820567533374, + 0.005884988233447075, + 0.0018721895758062601, + 0.006704017519950867, + 0.0023835778702050447, + 0.0022074906155467033, + 0.0010258755646646023, + 0.0018395187798887491, + 0.001742516877129674, + 0.003480796469375491, + 0.002204270800575614, + 0.0029504485428333282, + 0.0018597656162455678, + 0.002574145793914795, + 0.0010559989605098963, + 0.001528008608147502, + 0.001650518155656755, + 0.047081172466278076, + 0.017736706882715225, + 0.010054637677967548, + 0.0058304863050580025, + 0.006694374606013298, + 0.010775819420814514, + 0.008233225904405117, + 0.031147001311182976, + 0.005259891506284475, + 0.004803794901818037, + 0.00591626763343811, + 0.003930260427296162, + 0.008440227247774601, + 0.0023986664600670338, + 0.004848853684961796, + 0.0032738475129008293, + 0.0018476987024769187, + 0.0012914339313283563, + 0.0033257172908633947, + 0.0022131220903247595, + 0.0010168985463678837, + 0.0008790624560788274, + 0.0012108295923098922, + 0.0010201201075688004, + 0.0011122694704681635, + 0.001075607375241816, + 0.0017782323993742466, + 0.0022453952115029097, + 0.0011687199585139751, + 0.0011349773267284036, + 0.0016719320556148887, + 0.0011439380468800664, + 0.0014821429504081607, + 0.0013066544197499752, + 0.0014223998878151178, + 0.001261304714716971, + 0.003432714845985174, + 0.001575653674080968, + 0.001366430427879095, + 0.001039660070091486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.08997481316328049 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428530514240265 + ], + "opposite_logprob_tokens": [ + -2.017853021621704 + ], + "logprob_reward": -0.1428530514240265, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428530514240265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43911346793174744, + 0.04246363043785095, + 0.06046885624527931, + 0.08560585230588913, + 0.029487883672118187, + 0.22585709393024445, + 0.03984910622239113, + 0.03578347712755203, + 0.014065408147871494, + 0.027305245399475098 + ], + "credit_map": [ + 0.11342158913612366, + 0.032144948840141296, + 0.16707518696784973, + 0.029988490045070648, + 0.02139863930642605, + 0.028457587584853172, + 0.025419050827622414, + 0.021207958459854126, + 0.008035045117139816, + 0.004268513526767492, + 0.01104592252522707, + 0.004392758943140507, + 0.005881402641534805, + 0.0030702175572514534, + 0.003735343459993601, + 0.0020344278309494257, + 0.010264541953802109, + 0.007658446207642555, + 0.012154261581599712, + 0.006863490212708712, + 0.00951464008539915, + 0.003927408717572689, + 0.006250391714274883, + 0.00383567763492465, + 0.02496047504246235, + 0.006075999699532986, + 0.032672539353370667, + 0.006893525831401348, + 0.004136122763156891, + 0.0030095516704022884, + 0.003415655577555299, + 0.00444199051707983, + 0.005310263950377703, + 0.0028387918137013912, + 0.006780159194022417, + 0.003313404740765691, + 0.00229097087867558, + 0.003019785974174738, + 0.002490615239366889, + 0.003443889319896698, + 0.07703849673271179, + 0.02877683937549591, + 0.017322128638625145, + 0.008428167551755905, + 0.010366229340434074, + 0.018091101199388504, + 0.01430957205593586, + 0.051524579524993896, + 0.006398060824722052, + 0.0042234305292367935, + 0.004451473243534565, + 0.0034556216560304165, + 0.006696349009871483, + 0.004940222017467022, + 0.005137930158525705, + 0.004546016920357943, + 0.00438255351036787, + 0.0026812709402292967, + 0.005310677923262119, + 0.0036429865285754204, + 0.0031757522374391556, + 0.0038681100122630596, + 0.004648228175938129, + 0.008073896169662476, + 0.001648982404731214, + 0.0013658319367095828, + 0.0016387144569307566, + 0.0030079877469688654, + 0.0013433683197945356, + 0.0014635092811658978, + 0.0015541347675025463, + 0.00204287888482213, + 0.003789467504248023, + 0.00255954940803349, + 0.003807092783972621, + 0.0025002385955303907, + 0.003959000576287508, + 0.002858719089999795, + 0.0027563406620174646, + 0.005074839107692242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.1428530514240265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370928019285202 + ], + "opposite_logprob_tokens": [ + -3.4087092876434326 + ], + "logprob_reward": -0.03370928019285202, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370928019285202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7474369406700134, + 0.01552184484899044, + 0.03141062706708908, + 0.08620928972959518, + 0.01587192714214325, + 0.030310247093439102, + 0.018534619361162186, + 0.029535667970776558, + 0.011595037765800953, + 0.0135737843811512 + ], + "credit_map": [ + 0.21553298830986023, + 0.050549931824207306, + 0.32232826948165894, + 0.05497654899954796, + 0.023019319400191307, + 0.045592546463012695, + 0.01811814121901989, + 0.017319206148386, + 0.0016801696037873626, + 0.00286768632940948, + 0.0025111306458711624, + 0.0016016324516385794, + 0.002442883560433984, + 0.0013761756708845496, + 0.002016338985413313, + 0.0010258284164592624, + 0.005798641592264175, + 0.0033110259100794792, + 0.006415338721126318, + 0.006795774679630995, + 0.003235179465264082, + 0.0014630674850195646, + 0.002816862892359495, + 0.0015747338766232133, + 0.02457287535071373, + 0.0057412562891840935, + 0.029781809076666832, + 0.006911555305123329, + 0.0057405405677855015, + 0.0039924378506839275, + 0.004527977202087641, + 0.00494083808735013, + 0.0028538918122649193, + 0.0018726239213719964, + 0.002937690820544958, + 0.0021816042717546225, + 0.0017620997969061136, + 0.0013135033659636974, + 0.0016262722201645374, + 0.0013242398854345083, + 0.009329543448984623, + 0.003609456354752183, + 0.0025962430518120527, + 0.0021155327558517456, + 0.0017024428816512227, + 0.002414002316072583, + 0.0022145097609609365, + 0.006328517571091652, + 0.003064175369217992, + 0.0022441863548010588, + 0.0019291890785098076, + 0.001624070922844112, + 0.0032937515061348677, + 0.00205782987177372, + 0.002606690861284733, + 0.0017147265607491136, + 0.0029808059334754944, + 0.0022851789835840464, + 0.004464705940335989, + 0.004331445787101984, + 0.003606372280046344, + 0.0034940922632813454, + 0.003887140192091465, + 0.004485927056521177, + 0.0014361297944560647, + 0.0012534797424450517, + 0.0018236648757010698, + 0.0014547394821420312, + 0.0015670692082494497, + 0.0013357583666220307, + 0.001313904533162713, + 0.0014102909481152892, + 0.0015644881641492248, + 0.0013647929299622774, + 0.0014437113422900438, + 0.001781484461389482, + 0.0029881922528147697, + 0.001424047746695578, + 0.001314807217568159, + 0.001692259800620377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.03370928019285202 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288307711482048 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288307711482048, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288307711482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09868068993091583, + 0.03480469435453415, + 0.0602814257144928, + 0.07764735072851181, + 0.055861230939626694, + 0.532005786895752, + 0.09201636165380478, + 0.014347980730235577, + 0.011607239954173565, + 0.022747235372662544 + ], + "credit_map": [ + 0.01393941231071949, + 0.012664302252233028, + 0.010337325744330883, + 0.010448584333062172, + 0.015512334182858467, + 0.009285799227654934, + 0.0137609438970685, + 0.012731983326375484, + 0.004385989625006914, + 0.0036467702593654394, + 0.004413840360939503, + 0.0031792621593922377, + 0.007072939537465572, + 0.0032918851356953382, + 0.005587969906628132, + 0.003226036438718438, + 0.005687090568244457, + 0.007240278646349907, + 0.006671214941889048, + 0.007136850152164698, + 0.013465724885463715, + 0.0043757022358477116, + 0.009626911021769047, + 0.0060776518657803535, + 0.02004477009177208, + 0.0080689936876297, + 0.02667384408414364, + 0.006660028826445341, + 0.007103698793798685, + 0.0031269220635294914, + 0.003134318394586444, + 0.002834778279066086, + 0.011136233806610107, + 0.0076286024414002895, + 0.009668112732470036, + 0.006127410102635622, + 0.007942057214677334, + 0.003934342414140701, + 0.004556229803711176, + 0.004868237767368555, + 0.16791506111621857, + 0.07110335677862167, + 0.040201541036367416, + 0.032217103987932205, + 0.043908726423978806, + 0.03888743743300438, + 0.03364984318614006, + 0.10412272065877914, + 0.009964842349290848, + 0.008514272049069405, + 0.009185155853629112, + 0.007880935445427895, + 0.01976093091070652, + 0.016360798850655556, + 0.010871597565710545, + 0.00947782676666975, + 0.0016160464147105813, + 0.001269751344807446, + 0.0032939529046416283, + 0.00186430593021214, + 0.001394130871631205, + 0.0012282439274713397, + 0.001530122826807201, + 0.0021514268592000008, + 0.0018676507752388716, + 0.0011445690179243684, + 0.00156106345821172, + 0.0016075298190116882, + 0.0012704164255410433, + 0.0013531337026506662, + 0.001548447529785335, + 0.0012544289929792285, + 0.0030753391329199076, + 0.0025028919335454702, + 0.007004968822002411, + 0.0034396108239889145, + 0.0016348117496818304, + 0.0016087887343019247, + 0.0019446323858574033, + 0.0015361927216872573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.023288307711482048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot following the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible's hull heavy, matte black metal, and visibly scarred or scuffed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, multi-jointed robotic arm visible folded at one side of the submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single searchlight on the front project a concentrated, cone-shaped beam of cold white light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the heavy tracks roll slowly over the silty seabed, stirring up turbid sand that drifts and swirls in a slow-motion particle effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576467737555504 + ], + "opposite_logprob_tokens": [ + -3.8955764770507812 + ], + "logprob_reward": -0.020576467737555504, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576467737555504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7895879745483398, + 0.013256082311272621, + 0.08085208386182785, + 0.03192977234721184, + 0.015488076023757458, + 0.03240630403161049, + 0.013147209770977497, + 0.007279894780367613, + 0.010034032166004181, + 0.006018511019647121 + ], + "credit_map": [ + 0.23037320375442505, + 0.0532749742269516, + 0.34560060501098633, + 0.05816471204161644, + 0.019267210736870766, + 0.049782976508140564, + 0.01781029812991619, + 0.01531408540904522, + 0.0018659456400200725, + 0.0023369607515633106, + 0.0015887825284153223, + 0.0011962103890255094, + 0.0018180407350882888, + 0.0017810476711019874, + 0.0015913667157292366, + 0.00107772892806679, + 0.018151383846998215, + 0.007363299839198589, + 0.023085402324795723, + 0.014433584176003933, + 0.006382890976965427, + 0.002982008969411254, + 0.004939790815114975, + 0.003513726871460676, + 0.006659338716417551, + 0.002537241904065013, + 0.008598842658102512, + 0.005270414520055056, + 0.0024428258184343576, + 0.0016647411976009607, + 0.002598909428343177, + 0.002157460665330291, + 0.0023199107963591814, + 0.002452637767419219, + 0.002324082190170884, + 0.002099965699017048, + 0.0030841080006211996, + 0.001012778957374394, + 0.0012160339392721653, + 0.0009785605361685157, + 0.006231265142560005, + 0.0034494693391025066, + 0.004809287842363119, + 0.005079254973679781, + 0.0031077866442501545, + 0.0023312000557780266, + 0.0036271503195166588, + 0.003770893206819892, + 0.0017767696408554912, + 0.0016049626283347607, + 0.0014071314362809062, + 0.0012607666430994868, + 0.002954039489850402, + 0.00102781702298671, + 0.0018557426519691944, + 0.0012599821202456951, + 0.0008227224461734295, + 0.0006669152062386274, + 0.0019002946792170405, + 0.0008250769460573792, + 0.0008059806423261762, + 0.0004783880722243339, + 0.0009814590448513627, + 0.0007990585290826857, + 0.0011782587971538305, + 0.0010576322674751282, + 0.0011859131045639515, + 0.0019966207910329103, + 0.001031212043017149, + 0.0013083580415695906, + 0.0012589357793331146, + 0.0010171020403504372, + 0.00103831198066473, + 0.0006006529438309371, + 0.0010380946332588792, + 0.0006310517783276737, + 0.0009915921837091446, + 0.0004712970112450421, + 0.0006424200837500393, + 0.0006050910451449454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.020576467737555504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_tensor([1950, 1950])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_tensor([1950, 1950])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a4566f331df0a5be64aece258051f156a72a5e55 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_tensor([1950, 1950])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, filmed from a slightly elevated angle and panning slowly to the right, captures a friendly children's soccer game in full swing in a sunny suburban park. On the neatly manicured, fresh-smelling emerald green lawn, two teams of children, approximately eight to ten years old, dressed in vibrant jerseys—one a dazzling lemon yellow, the other a brilliant royal blue—energize and chase a black and white soccer ball. The warm, golden afternoon sun streams across the field, elongating the soft shadows cast by the children as they run. The camera pans to the sidelines, focusing on a crowd of parents and siblings seated on folding canvas chairs or plaid picnic mats. A father in a gray hoodie excitedly rises from his chair, fists clenched, and cheers on his son on the field. Nearby, a smiling mother gently pats the back of her young daughter as she records the video on her phone. When a young player in a yellow jersey deftly dribbled the ball past the defense, the crowd erupted in cheers mixed with laughter and encouragement. The whole scene was filled with family warmth, healthy vitality, and pure joy, creating a community atmosphere of unity and friendship.", + "seed": 12593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_1950_seed_12593_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03792275208979845, + "vlm_logprob": -0.03792275208979845, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03792275208979845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04079906642436981 + ], + "opposite_logprob_tokens": [ + -3.540799140930176 + ], + "logprob_reward": -0.04079906642436981, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04079906642436981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10008465498685837, + 0.05479109659790993, + 0.2712925970554352, + 0.04447415471076965, + 0.20841465890407562, + 0.09213804453611374, + 0.03277646005153656, + 0.053446244448423386, + 0.08563389629125595, + 0.056948188692331314 + ], + "credit_map": [ + 0.006402200553566217, + 0.006564552430063486, + 0.014906500466167927, + 0.016623640432953835, + 0.007915977388620377, + 0.006804120261222124, + 0.015483043156564236, + 0.025384625419974327, + 0.009242808446288109, + 0.005417565815150738, + 0.006569149438291788, + 0.011591169983148575, + 0.0029946742579340935, + 0.0028988164849579334, + 0.007197462022304535, + 0.00887945108115673, + 0.011257964186370373, + 0.011284537613391876, + 0.09151320159435272, + 0.03105182759463787, + 0.0081391716375947, + 0.01729384995996952, + 0.023821596056222916, + 0.0769304484128952, + 0.0035955626517534256, + 0.003855588613077998, + 0.005884403362870216, + 0.01130723301321268, + 0.0043925135396420956, + 0.0038320664316415787, + 0.006723859813064337, + 0.004882927052676678, + 0.013192269951105118, + 0.021021058782935143, + 0.042745012789964676, + 0.032823700457811356, + 0.046765126287937164, + 0.015316334553062916, + 0.021199481561779976, + 0.015351665206253529, + 0.006287552881985903, + 0.007752484176307917, + 0.015473735518753529, + 0.02777450531721115, + 0.007775755133479834, + 0.00827418640255928, + 0.009433053433895111, + 0.009366767480969429, + 0.0035447529517114162, + 0.003503618761897087, + 0.005453111603856087, + 0.0028383329045027494, + 0.003952894359827042, + 0.0051793125458061695, + 0.0037009308580309153, + 0.004603504668921232, + 0.007347618229687214, + 0.00763223459944129, + 0.005422614980489016, + 0.005200108978897333, + 0.011058283038437366, + 0.006170434411615133, + 0.005754272919148207, + 0.004860675428062677, + 0.009946821257472038, + 0.024424908682703972, + 0.007179625332355499, + 0.009007141925394535, + 0.009191145189106464, + 0.011050510220229626, + 0.006365572568029165, + 0.008468171581625938, + 0.004792364779859781, + 0.016704509034752846, + 0.002234788378700614, + 0.004318437539041042, + 0.005622988101094961, + 0.014101983048021793, + 0.0029055755585432053, + 0.006267542950809002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.04079906642436981 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00765107199549675 + ], + "opposite_logprob_tokens": [ + -4.882650852203369 + ], + "logprob_reward": -0.00765107199549675, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00765107199549675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1698606163263321, + 0.06347493827342987, + 0.19640608131885529, + 0.0784115120768547, + 0.0738741010427475, + 0.03407843038439751, + 0.07381229847669601, + 0.16009487211704254, + 0.1225428357720375, + 0.027444249019026756 + ], + "credit_map": [ + 0.01249153446406126, + 0.008395904675126076, + 0.02207949385046959, + 0.02682342194020748, + 0.007744553033262491, + 0.013251633383333683, + 0.030933696776628494, + 0.04814038798213005, + 0.008054209873080254, + 0.0047694104723632336, + 0.005808010697364807, + 0.015510781668126583, + 0.003062117611989379, + 0.0022913948632776737, + 0.005951058119535446, + 0.018027963116765022, + 0.00764028076082468, + 0.004156591836363077, + 0.07908682525157928, + 0.0136371785774827, + 0.004611755721271038, + 0.007702956907451153, + 0.014191514812409878, + 0.06537898629903793, + 0.003919641952961683, + 0.005053205415606499, + 0.019885210320353508, + 0.0182250514626503, + 0.005304308142513037, + 0.003627283498644829, + 0.0163066778331995, + 0.006090143229812384, + 0.006355156656354666, + 0.006231928244233131, + 0.016052383929491043, + 0.01282639242708683, + 0.01075739972293377, + 0.005877810996025801, + 0.009287392720580101, + 0.006485639605671167, + 0.0031282135751098394, + 0.0032641550060361624, + 0.004871690180152655, + 0.004748613573610783, + 0.004363102372735739, + 0.004463053774088621, + 0.004581144545227289, + 0.0046584634110331535, + 0.009389822371304035, + 0.012080401182174683, + 0.007660032249987125, + 0.006699010729789734, + 0.01035335287451744, + 0.010425064712762833, + 0.010609647259116173, + 0.006594971753656864, + 0.020636865869164467, + 0.02490946836769581, + 0.018899358808994293, + 0.01831241138279438, + 0.022777333855628967, + 0.016131971031427383, + 0.017378954216837883, + 0.021048540249466896, + 0.00937836803495884, + 0.041307419538497925, + 0.007107663433998823, + 0.0166753102093935, + 0.010185682214796543, + 0.01537519320845604, + 0.0082745011895895, + 0.014238699339330196, + 0.0024577556177973747, + 0.002712477697059512, + 0.002134969923645258, + 0.003449169686064124, + 0.002960971323773265, + 0.007843753322958946, + 0.0028499348554760218, + 0.0030352198518812656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.00765107199549675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09195601940155029, + 0.07449321448802948, + 0.050122134387493134, + 0.10251244157552719, + 0.271353155374527, + 0.11097715049982071, + 0.09255826473236084, + 0.04683966562151909, + 0.07015229016542435, + 0.08903566747903824 + ], + "credit_map": [ + 0.012133447453379631, + 0.009012986905872822, + 0.012441328726708889, + 0.011008950881659985, + 0.00646614283323288, + 0.005483077373355627, + 0.012182416394352913, + 0.023227671161293983, + 0.004939645528793335, + 0.006587453186511993, + 0.011339187622070312, + 0.012624907307326794, + 0.0038872931618243456, + 0.003574893344193697, + 0.007068068720400333, + 0.024471763521432877, + 0.004486924968659878, + 0.003491079667583108, + 0.006379218772053719, + 0.005677378736436367, + 0.004011647310107946, + 0.010594368912279606, + 0.00885795895010233, + 0.006623559165745974, + 0.005866744089871645, + 0.004027751740068197, + 0.014471407048404217, + 0.02280035801231861, + 0.01039059367030859, + 0.01788806915283203, + 0.016086529940366745, + 0.010980985127389431, + 0.022090338170528412, + 0.035850681364536285, + 0.039068229496479034, + 0.03603371977806091, + 0.06950373947620392, + 0.020415566861629486, + 0.028524553403258324, + 0.019866321235895157, + 0.006823603995144367, + 0.012652207165956497, + 0.016963033005595207, + 0.01785721816122532, + 0.010723407380282879, + 0.015200569294393063, + 0.015968082472682, + 0.014789024367928505, + 0.007233510259538889, + 0.00896878819912672, + 0.010513336397707462, + 0.010514847002923489, + 0.02009483240544796, + 0.010227970778942108, + 0.018035700544714928, + 0.006969277281314135, + 0.00644432520493865, + 0.006012432277202606, + 0.004225172568112612, + 0.006721023004502058, + 0.008464117534458637, + 0.005259823985397816, + 0.005066941492259502, + 0.004645828623324633, + 0.0061178551986813545, + 0.01872711442410946, + 0.006082239095121622, + 0.009962367825210094, + 0.006265698466449976, + 0.008926169015467167, + 0.005594278685748577, + 0.008476567454636097, + 0.0051447865553200245, + 0.03065907210111618, + 0.00479486957192421, + 0.013677600771188736, + 0.012661896646022797, + 0.01104073878377676, + 0.003215071512386203, + 0.007841629907488823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.029802024364471436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072737164795399 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072737164795399, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072737164795399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09457673877477646, + 0.04570050165057182, + 0.1709383726119995, + 0.13225415349006653, + 0.1816525012254715, + 0.1464819312095642, + 0.04270026460289955, + 0.05015445873141289, + 0.09823284298181534, + 0.037308257073163986 + ], + "credit_map": [ + 0.011719949543476105, + 0.00798810925334692, + 0.013583758845925331, + 0.011749972589313984, + 0.008748553693294525, + 0.005144880153238773, + 0.013361177407205105, + 0.022280342876911163, + 0.005073260981589556, + 0.004785891156643629, + 0.005973659455776215, + 0.007672215346246958, + 0.006597454659640789, + 0.0039608110673725605, + 0.003633045358583331, + 0.00800416525453329, + 0.0070871165953576565, + 0.005830142181366682, + 0.05501658469438553, + 0.017418012022972107, + 0.006193892098963261, + 0.021005943417549133, + 0.016332970932126045, + 0.0420537106692791, + 0.011260326951742172, + 0.008394718170166016, + 0.01721288450062275, + 0.016384344547986984, + 0.02170223370194435, + 0.020849935710430145, + 0.018576808273792267, + 0.017872905358672142, + 0.012838317081332207, + 0.020124059170484543, + 0.028575323522090912, + 0.025665245950222015, + 0.046560630202293396, + 0.013380871154367924, + 0.019083309918642044, + 0.015424739569425583, + 0.009802555665373802, + 0.0131378173828125, + 0.02176203951239586, + 0.0343567356467247, + 0.009987560100853443, + 0.018696552142500877, + 0.020401401445269585, + 0.018337270244956017, + 0.004417521879076958, + 0.006408528424799442, + 0.0053982799872756, + 0.005437476094812155, + 0.004552596714347601, + 0.005293963942676783, + 0.0062157087959349155, + 0.004976189695298672, + 0.0058364770375192165, + 0.006831460632383823, + 0.006651806179434061, + 0.005804494488984346, + 0.006965090520679951, + 0.005106217693537474, + 0.006566123571246862, + 0.006392787676304579, + 0.007483727764338255, + 0.04098479077219963, + 0.004601051565259695, + 0.012698579579591751, + 0.008479228243231773, + 0.011173182167112827, + 0.004337921738624573, + 0.008474349975585938, + 0.0036913822405040264, + 0.006924882996827364, + 0.0030711302533745766, + 0.003969954326748848, + 0.002831466728821397, + 0.010916607454419136, + 0.002096845768392086, + 0.003805987536907196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.011072737164795399 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028886049985886 + ], + "opposite_logprob_tokens": [ + -2.3502888679504395 + ], + "logprob_reward": -0.10028886049985886, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028886049985886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048011619597673416, + 0.0403665229678154, + 0.0520031675696373, + 0.034861985594034195, + 0.20130057632923126, + 0.1603250354528427, + 0.14212995767593384, + 0.1500099003314972, + 0.07506227493286133, + 0.09592901170253754 + ], + "credit_map": [ + 0.0050596813671290874, + 0.0043549188412725925, + 0.005900161806493998, + 0.009310484863817692, + 0.00454057939350605, + 0.0022248891182243824, + 0.007544780150055885, + 0.009076124057173729, + 0.00419211620464921, + 0.0034898039884865284, + 0.0065485392697155476, + 0.010630249977111816, + 0.0018122013425454497, + 0.0018032541265711188, + 0.004611059091985226, + 0.007279296405613422, + 0.005202603992074728, + 0.004336080048233271, + 0.01019201334565878, + 0.00884236115962267, + 0.0023645104374736547, + 0.007166747469455004, + 0.007838803343474865, + 0.006060048472136259, + 0.0028820757288485765, + 0.002098250901326537, + 0.005884386133402586, + 0.006578151136636734, + 0.004568415693938732, + 0.003274948336184025, + 0.006353205069899559, + 0.003222550731152296, + 0.012400157749652863, + 0.018839066848158836, + 0.0330197811126709, + 0.028580736368894577, + 0.05177421495318413, + 0.018038176000118256, + 0.021065346896648407, + 0.017583098262548447, + 0.012083301320672035, + 0.01704988069832325, + 0.025841984897851944, + 0.03695937618613243, + 0.015802035108208656, + 0.019178763031959534, + 0.016473980620503426, + 0.016935698688030243, + 0.010830978862941265, + 0.015150206163525581, + 0.02645968459546566, + 0.017030656337738037, + 0.016411757096648216, + 0.019993241876363754, + 0.022811362519860268, + 0.013442066498100758, + 0.01841951720416546, + 0.029275257140398026, + 0.016060447320342064, + 0.01780327782034874, + 0.02171795256435871, + 0.015450955368578434, + 0.014884961768984795, + 0.016397519037127495, + 0.005059093702584505, + 0.021471958607435226, + 0.006581708323210478, + 0.010131055489182472, + 0.007594651076942682, + 0.01061537116765976, + 0.005627121310681105, + 0.007981309667229652, + 0.005393347702920437, + 0.033278707414865494, + 0.005899990443140268, + 0.010944659821689129, + 0.01660171151161194, + 0.012352772988379002, + 0.0036237440072000027, + 0.007834082469344139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.10028886049985886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a slightly elevated angle and panning slowly to the right, captures a friendly children's soccer game in full swing in a sunny suburban park. On the neatly manicured, fresh-smelling emerald green lawn, two teams of children, approximately eight to ten years old, dressed in vibrant jerseys—one a dazzling lemon yellow, the other a brilliant royal blue—energize and chase a black and white soccer ball. The warm, golden afternoon sun streams across the field, elongating the soft shadows cast by the children as they run. The camera pans to the sidelines, focusing on a crowd of parents and siblings seated on folding canvas chairs or plaid picnic mats. A father in a gray hoodie excitedly rises from his chair, fists clenched, and cheers on his son on the field. Nearby, a smiling mother gently pats the back of her young daughter as she records the video on her phone. When a young player in a yellow jersey deftly dribbled the ball past the defense, the crowd erupted in cheers mixed with laughter and encouragement. The whole scene was filled with family warmth, healthy vitality, and pure joy, creating a community atmosphere of unity and friendship.", + "seed": 12594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_1950_seed_12594_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025737684033811093, + "vlm_logprob": -0.025737684033811093, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.025737684033811093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037399932742118835 + ], + "opposite_logprob_tokens": [ + -3.787400007247925 + ], + "logprob_reward": -0.037399932742118835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.037399932742118835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21664515137672424, + 0.13070400059223175, + 0.0577889159321785, + 0.10959859192371368, + 0.11813931912183762, + 0.07411015778779984, + 0.08087383955717087, + 0.06797878444194794, + 0.038362957537174225, + 0.10579823702573776 + ], + "credit_map": [ + 0.0436314158141613, + 0.028504515066742897, + 0.028607472777366638, + 0.026683131232857704, + 0.0147707499563694, + 0.017570694908499718, + 0.027645226567983627, + 0.02923196367919445, + 0.022419221699237823, + 0.02062617801129818, + 0.017463406547904015, + 0.02027539350092411, + 0.013636423274874687, + 0.01358374860137701, + 0.01291697658598423, + 0.009782659821212292, + 0.009120296686887741, + 0.01058212760835886, + 0.0069604539312422276, + 0.0075554740615189075, + 0.006095077842473984, + 0.004195722751319408, + 0.007292157504707575, + 0.0059876046143472195, + 0.009526967070996761, + 0.01929999701678753, + 0.01666359417140484, + 0.018502244725823402, + 0.010241543874144554, + 0.009406530298292637, + 0.016239341348409653, + 0.009718373417854309, + 0.016940869390964508, + 0.018258705735206604, + 0.01328772772103548, + 0.014207116328179836, + 0.016116352751851082, + 0.015407510101795197, + 0.01459425501525402, + 0.009326777420938015, + 0.009389964863657951, + 0.00545347249135375, + 0.00848856195807457, + 0.009923434816300869, + 0.010965781286358833, + 0.009579145349562168, + 0.010388154536485672, + 0.009921636432409286, + 0.00746139045804739, + 0.010361665859818459, + 0.01525065116584301, + 0.009689009748399258, + 0.008882544934749603, + 0.006641217973083258, + 0.010632657445967197, + 0.011954696848988533, + 0.007217569276690483, + 0.008113809861242771, + 0.00818889494985342, + 0.011038700118660927, + 0.007974515669047832, + 0.011893405579030514, + 0.006725209765136242, + 0.006826680153608322, + 0.0038764739874750376, + 0.005499656777828932, + 0.004068530630320311, + 0.005046788137406111, + 0.006210636347532272, + 0.004235933069139719, + 0.0057472651824355125, + 0.0036776731722056866, + 0.014640484005212784, + 0.020077578723430634, + 0.014237377792596817, + 0.010156694799661636, + 0.009461267851293087, + 0.013161255978047848, + 0.014545165933668613, + 0.009518411941826344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.037399932742118835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658486418426037 + ], + "opposite_logprob_tokens": [ + -4.758658409118652 + ], + "logprob_reward": -0.008658486418426037, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658486418426037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09394536167383194, + 0.08526432514190674, + 0.1747148185968399, + 0.06041616573929787, + 0.17695698142051697, + 0.052014391869306564, + 0.06713235378265381, + 0.06803162395954132, + 0.09761868417263031, + 0.12390527874231339 + ], + "credit_map": [ + 0.015571173280477524, + 0.010840797796845436, + 0.013397233560681343, + 0.014215917326509953, + 0.009619973599910736, + 0.008215793408453465, + 0.01252950169146061, + 0.009554973803460598, + 0.007678097579628229, + 0.008359442465007305, + 0.01186329685151577, + 0.010547909885644913, + 0.010048233903944492, + 0.009933867491781712, + 0.011102075688540936, + 0.01573139987885952, + 0.01967170089483261, + 0.024356121197342873, + 0.02291717566549778, + 0.02114187553524971, + 0.036464303731918335, + 0.017535926774144173, + 0.014724919572472572, + 0.017902802675962448, + 0.006396151147782803, + 0.00812511146068573, + 0.010155804455280304, + 0.00927025731652975, + 0.00561754172667861, + 0.0064783948473632336, + 0.008376042358577251, + 0.005996866151690483, + 0.014475172385573387, + 0.010033808648586273, + 0.02806488610804081, + 0.02693471871316433, + 0.016590438783168793, + 0.020011547952890396, + 0.027184370905160904, + 0.033662036061286926, + 0.005021070595830679, + 0.008324199356138706, + 0.005449374206364155, + 0.006526316050440073, + 0.009704183787107468, + 0.004520324990153313, + 0.006335277575999498, + 0.006133645307272673, + 0.007997282780706882, + 0.007855914533138275, + 0.007927591912448406, + 0.008748923428356647, + 0.011314158327877522, + 0.008044491522014141, + 0.00795016996562481, + 0.007293822709470987, + 0.007863882929086685, + 0.007670377381145954, + 0.009952239692211151, + 0.01105041615664959, + 0.006218849681317806, + 0.012238025665283203, + 0.005243061576038599, + 0.0077947708778083324, + 0.015099060721695423, + 0.013898191042244434, + 0.011111217550933361, + 0.010689680464565754, + 0.014052964746952057, + 0.011147650890052319, + 0.012787927873432636, + 0.008831996470689774, + 0.010668010450899601, + 0.010892937891185284, + 0.02642376348376274, + 0.0134894335642457, + 0.017209351062774658, + 0.014086838811635971, + 0.01734965480864048, + 0.013785285875201225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.008658486418426037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803412035107613 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803412035107613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803412035107613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09732664376497269, + 0.058715250343084335, + 0.09471315890550613, + 0.09354221820831299, + 0.12778240442276, + 0.3300296366214752, + 0.03922179341316223, + 0.04276397079229355, + 0.058269865810871124, + 0.057635094970464706 + ], + "credit_map": [ + 0.010499934665858746, + 0.01153965201228857, + 0.011640641838312149, + 0.018644122406840324, + 0.008456910029053688, + 0.008221544325351715, + 0.013877395540475845, + 0.014446437358856201, + 0.006555347703397274, + 0.006199320778250694, + 0.008757533505558968, + 0.011367712169885635, + 0.0075973523780703545, + 0.0057799480855464935, + 0.00655162101611495, + 0.005906418431550264, + 0.007841861806809902, + 0.01536483969539404, + 0.01181519590318203, + 0.01417594589293003, + 0.013302604667842388, + 0.008495117537677288, + 0.01293819397687912, + 0.010779396630823612, + 0.01078339759260416, + 0.01302266400307417, + 0.011235939338803291, + 0.012737770564854145, + 0.007665818557143211, + 0.011250246316194534, + 0.01008350495249033, + 0.016762878745794296, + 0.0095713771879673, + 0.008870698511600494, + 0.020344320684671402, + 0.01997886598110199, + 0.012819301337003708, + 0.017577888444066048, + 0.02364310994744301, + 0.014976851642131805, + 0.03408021479845047, + 0.03983183950185776, + 0.03593197092413902, + 0.059038471430540085, + 0.0653253123164177, + 0.025293678045272827, + 0.04320800304412842, + 0.027320129796862602, + 0.00522335572168231, + 0.00464413920417428, + 0.0046496945433318615, + 0.0047056241892278194, + 0.006441053468734026, + 0.005455761216580868, + 0.00405952287837863, + 0.0040426417253911495, + 0.0042474414221942425, + 0.004561767913401127, + 0.005514349788427353, + 0.0050406064838171005, + 0.005606598220765591, + 0.008531585335731506, + 0.004289599135518074, + 0.0049720206297934055, + 0.00988216232508421, + 0.009653275832533836, + 0.006638555787503719, + 0.007019161246716976, + 0.007148461882025003, + 0.0061654942110180855, + 0.006741633173078299, + 0.005021124612540007, + 0.005207017529755831, + 0.004690084606409073, + 0.009621280245482922, + 0.006307096220552921, + 0.009565855376422405, + 0.008306087926030159, + 0.008967969566583633, + 0.004969704896211624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.029803412035107613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779518470168114 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779518470168114, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779518470168114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17626182734966278, + 0.12532785534858704, + 0.1536400020122528, + 0.05761512368917465, + 0.14427490532398224, + 0.0877138301730156, + 0.08329419791698456, + 0.05534404516220093, + 0.06350698322057724, + 0.053021159023046494 + ], + "credit_map": [ + 0.019282184541225433, + 0.015320788137614727, + 0.027853580191731453, + 0.032944925129413605, + 0.011371467262506485, + 0.011407243087887764, + 0.034257903695106506, + 0.023823752999305725, + 0.012066137045621872, + 0.013772266916930676, + 0.016519872471690178, + 0.01778622344136238, + 0.019663633778691292, + 0.013113495893776417, + 0.018202703446149826, + 0.014203530736267567, + 0.012267669662833214, + 0.024880481883883476, + 0.022921940311789513, + 0.01773799955844879, + 0.01927519403398037, + 0.014120513573288918, + 0.019505858421325684, + 0.02293034829199314, + 0.005432994570583105, + 0.005327663850039244, + 0.00553300604224205, + 0.00960121676325798, + 0.007257257588207722, + 0.0069308774545788765, + 0.006099392659962177, + 0.011432723142206669, + 0.010497111827135086, + 0.011677547357976437, + 0.01953006349503994, + 0.019375398755073547, + 0.014418888837099075, + 0.0295726228505373, + 0.02018970437347889, + 0.01901358924806118, + 0.011121219955384731, + 0.006780148018151522, + 0.009987781755626202, + 0.011653253808617592, + 0.009379756636917591, + 0.013173434883356094, + 0.012241511605679989, + 0.013376729562878609, + 0.007669907528907061, + 0.009564649313688278, + 0.012082734145224094, + 0.009878871031105518, + 0.006942586973309517, + 0.012794743292033672, + 0.010108704678714275, + 0.014252008870244026, + 0.006175292190164328, + 0.010735257528722286, + 0.007281939964741468, + 0.007142562884837389, + 0.006346043199300766, + 0.0073610651306807995, + 0.005291950888931751, + 0.005009937100112438, + 0.012272309511899948, + 0.00866031739860773, + 0.007444078102707863, + 0.006766066886484623, + 0.0073103089816868305, + 0.006917642429471016, + 0.007600076030939817, + 0.0065361857414245605, + 0.005927450489252806, + 0.004126004874706268, + 0.009619660675525665, + 0.005171339493244886, + 0.006228555925190449, + 0.008028145879507065, + 0.00823252834379673, + 0.005687476135790348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.009779518470168114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304707050323486 + ], + "opposite_logprob_tokens": [ + -3.1680469512939453 + ], + "logprob_reward": -0.04304707050323486, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304707050323486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06337711960077286, + 0.039143405854701996, + 0.04276541247963905, + 0.038115713745355606, + 0.14658895134925842, + 0.3883412480354309, + 0.057917121797800064, + 0.04091387614607811, + 0.0866442397236824, + 0.0961928740143776 + ], + "credit_map": [ + 0.007245325017720461, + 0.007266930770128965, + 0.00997963733971119, + 0.007638240698724985, + 0.006559242960065603, + 0.006589697673916817, + 0.009427816607058048, + 0.008670227602124214, + 0.004190625622868538, + 0.0039825658313930035, + 0.005737727042287588, + 0.006557030603289604, + 0.003785899607464671, + 0.003524517174810171, + 0.005760736297816038, + 0.005604300182312727, + 0.00399497477337718, + 0.0047654761001467705, + 0.006949593313038349, + 0.004459526389837265, + 0.005030737724155188, + 0.003960644826292992, + 0.006590079050511122, + 0.007014378439635038, + 0.0032909137662500143, + 0.006552404724061489, + 0.004893770441412926, + 0.005193037446588278, + 0.003623483469709754, + 0.004035914316773415, + 0.006098671816289425, + 0.004427517764270306, + 0.008313199505209923, + 0.008851354941725731, + 0.01915065012872219, + 0.01834416575729847, + 0.02054906263947487, + 0.03188274800777435, + 0.0206031184643507, + 0.018894657492637634, + 0.04279179871082306, + 0.05384761467576027, + 0.04185718670487404, + 0.05335063859820366, + 0.07637940347194672, + 0.03416696935892105, + 0.04676783084869385, + 0.039179783314466476, + 0.005329914391040802, + 0.009831959381699562, + 0.005050723906606436, + 0.00697475578635931, + 0.008922753855586052, + 0.009940153919160366, + 0.005798256956040859, + 0.006068604998290539, + 0.0047886064276099205, + 0.003325055353343487, + 0.006187780760228634, + 0.007678509224206209, + 0.003977685235440731, + 0.006450660992413759, + 0.004017484374344349, + 0.004488096106797457, + 0.013052084483206272, + 0.01007118634879589, + 0.009608202613890171, + 0.009318150579929352, + 0.013239344581961632, + 0.010933536104857922, + 0.011184421367943287, + 0.009237312711775303, + 0.011756965890526772, + 0.0070129092782735825, + 0.01699264906346798, + 0.007367186713963747, + 0.012405939400196075, + 0.01081195380538702, + 0.012325519695878029, + 0.017519749701023102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.04304707050323486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a slightly elevated angle and panning slowly to the right, captures a friendly children's soccer game in full swing in a sunny suburban park. On the neatly manicured, fresh-smelling emerald green lawn, two teams of children, approximately eight to ten years old, dressed in vibrant jerseys—one a dazzling lemon yellow, the other a brilliant royal blue—energize and chase a black and white soccer ball. The warm, golden afternoon sun streams across the field, elongating the soft shadows cast by the children as they run. The camera pans to the sidelines, focusing on a crowd of parents and siblings seated on folding canvas chairs or plaid picnic mats. A father in a gray hoodie excitedly rises from his chair, fists clenched, and cheers on his son on the field. Nearby, a smiling mother gently pats the back of her young daughter as she records the video on her phone. When a young player in a yellow jersey deftly dribbled the ball past the defense, the crowd erupted in cheers mixed with laughter and encouragement. The whole scene was filled with family warmth, healthy vitality, and pure joy, creating a community atmosphere of unity and friendship.", + "seed": 12595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_1950_seed_12595_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02919252421706915, + "vlm_logprob": -0.02919252421706915, + "vlm_yes_no_margin": 3.8749999523162844, + "vlm_token_credit": -0.02919252421706915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03869670629501343 + ], + "opposite_logprob_tokens": [ + -4.038696765899658 + ], + "logprob_reward": -0.03869670629501343, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03869670629501343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1129312589764595, + 0.0583459809422493, + 0.07584402710199356, + 0.03513010963797569, + 0.026388688012957573, + 0.030180614441633224, + 0.16028696298599243, + 0.08787821233272552, + 0.11586245894432068, + 0.29715168476104736 + ], + "credit_map": [ + 0.008451892994344234, + 0.012752426788210869, + 0.014547478407621384, + 0.022222639992833138, + 0.009836248122155666, + 0.010320735163986683, + 0.019929787144064903, + 0.014870049431920052, + 0.0041448334231972694, + 0.007450997829437256, + 0.007208413444459438, + 0.010150699876248837, + 0.00993425864726305, + 0.003806361462920904, + 0.010571595281362534, + 0.005078824702650309, + 0.008500042371451855, + 0.008298887871205807, + 0.011392940767109394, + 0.010059447027742863, + 0.00909550953656435, + 0.009125172160565853, + 0.011476077139377594, + 0.007895959541201591, + 0.004800694528967142, + 0.004415543749928474, + 0.0046631651930511, + 0.004319582600146532, + 0.003110797144472599, + 0.004271283280104399, + 0.004001760855317116, + 0.0055472832173109055, + 0.002965430961921811, + 0.0036867791786789894, + 0.003586462466046214, + 0.004474797751754522, + 0.0025485707446932793, + 0.0024961333256214857, + 0.004012736491858959, + 0.002617777558043599, + 0.0023166383616626263, + 0.003937732428312302, + 0.0025438203010708094, + 0.004285240545868874, + 0.0027668580878525972, + 0.004803040064871311, + 0.0064534470438957214, + 0.003073839470744133, + 0.017709173262119293, + 0.014387383125722408, + 0.03912155702710152, + 0.009771736338734627, + 0.04593218117952347, + 0.010531575419008732, + 0.013452018611133099, + 0.009381352923810482, + 0.014569849707186222, + 0.00984082743525505, + 0.007977968081831932, + 0.008377146907150745, + 0.011207538656890392, + 0.019378844648599625, + 0.0089988699182868, + 0.007527172099798918, + 0.007526347879320383, + 0.004379756283015013, + 0.0376652330160141, + 0.0044781919568777084, + 0.00612108176574111, + 0.04444278031587601, + 0.004344466608017683, + 0.00690460903570056, + 0.0794852077960968, + 0.029904497787356377, + 0.0995122566819191, + 0.020241957157850266, + 0.013507883064448833, + 0.018702764064073563, + 0.013635831885039806, + 0.02216131053864956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.03869670629501343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006755608133971691 + ], + "opposite_logprob_tokens": [ + -5.006755828857422 + ], + "logprob_reward": -0.006755608133971691, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006755608133971691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09343818575143814, + 0.11293058842420578, + 0.06917641311883926, + 0.07287861406803131, + 0.02789115346968174, + 0.05222257226705551, + 0.057172611355781555, + 0.10963956266641617, + 0.09508388489484787, + 0.30956634879112244 + ], + "credit_map": [ + 0.004900977946817875, + 0.010453930124640465, + 0.014405625872313976, + 0.015899335965514183, + 0.0089937224984169, + 0.007628786843270063, + 0.02002795971930027, + 0.011127857491374016, + 0.009478856809437275, + 0.01472173910588026, + 0.01322849839925766, + 0.020402677357196808, + 0.019467171281576157, + 0.007664740085601807, + 0.017225708812475204, + 0.01074121706187725, + 0.010480150580406189, + 0.006054863333702087, + 0.01070986595004797, + 0.011281284503638744, + 0.004218715708702803, + 0.010976305231451988, + 0.006395447533577681, + 0.009059793315827847, + 0.007588670123368502, + 0.0068785520270466805, + 0.009943414479494095, + 0.010818361304700375, + 0.00716144684702158, + 0.008335083723068237, + 0.012488334439694881, + 0.00966475997120142, + 0.004606761038303375, + 0.0034830232616513968, + 0.003714826190844178, + 0.003331632586196065, + 0.0028888003434985876, + 0.0035149159375578165, + 0.0024308243300765753, + 0.003920374903827906, + 0.0030034349765628576, + 0.005435104016214609, + 0.0038873020093888044, + 0.010847028344869614, + 0.004729618318378925, + 0.007207274436950684, + 0.012939756736159325, + 0.004173062741756439, + 0.007817337289452553, + 0.005316573660820723, + 0.011099250055849552, + 0.005051832180470228, + 0.00947948731482029, + 0.00725584989413619, + 0.007022674661129713, + 0.0041296170093119144, + 0.014608118683099747, + 0.01088083814829588, + 0.011289612390100956, + 0.013465971685945988, + 0.016948815435171127, + 0.02423004060983658, + 0.009866909123957157, + 0.008349276147782803, + 0.008110559545457363, + 0.005838214885443449, + 0.017054442316293716, + 0.0033890842460095882, + 0.009065228514373302, + 0.036025840789079666, + 0.004792284686118364, + 0.010808248072862625, + 0.08329758048057556, + 0.030103305354714394, + 0.1043233647942543, + 0.019454892724752426, + 0.01648956537246704, + 0.01869378425180912, + 0.013983355835080147, + 0.023220527917146683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.006755608133971691 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695679157972336 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.033695679157972336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695679157972336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05471646413207054, + 0.27755406498908997, + 0.2553752362728119, + 0.08994243294000626, + 0.06309810280799866, + 0.05046635493636131, + 0.02445986494421959, + 0.06941719353199005, + 0.06193963438272476, + 0.05303065851330757 + ], + "credit_map": [ + 0.004605868365615606, + 0.007533839903771877, + 0.006843201350420713, + 0.008541867136955261, + 0.0061239684000611305, + 0.006855969317257404, + 0.007634144742041826, + 0.006577601656317711, + 0.021212995052337646, + 0.03537632152438164, + 0.03335391357541084, + 0.06521224230527878, + 0.03434593603014946, + 0.01924503594636917, + 0.037897512316703796, + 0.03091011568903923, + 0.021252771839499474, + 0.0314042754471302, + 0.0425170436501503, + 0.05614922195672989, + 0.016625402495265007, + 0.02814640663564205, + 0.03625401481986046, + 0.023026084527373314, + 0.006674643140286207, + 0.010527295991778374, + 0.015785304829478264, + 0.01833084411919117, + 0.00892766285687685, + 0.009266878478229046, + 0.012172664515674114, + 0.008257140405476093, + 0.0053182034753263, + 0.006754610687494278, + 0.007715854328125715, + 0.0066481479443609715, + 0.008068658411502838, + 0.010085785761475563, + 0.009235597215592861, + 0.009271242655813694, + 0.0030038252007216215, + 0.006996842101216316, + 0.004092564340680838, + 0.006689182948321104, + 0.005736346356570721, + 0.008459442295134068, + 0.010403740219771862, + 0.005084413569420576, + 0.003886217949911952, + 0.0029933685436844826, + 0.0047892858274281025, + 0.0024342117831110954, + 0.002565717324614525, + 0.0018186400411650538, + 0.004262691829353571, + 0.001709732343442738, + 0.010366248898208141, + 0.005727134644985199, + 0.006554493214935064, + 0.007225412875413895, + 0.01005926076322794, + 0.016652263700962067, + 0.0066598583944141865, + 0.006172523368149996, + 0.004594517406076193, + 0.0026337839663028717, + 0.014546811580657959, + 0.002583053894340992, + 0.003846526611596346, + 0.024821151047945023, + 0.002893444849178195, + 0.006020341534167528, + 0.01330022793263197, + 0.006246797274798155, + 0.01512874010950327, + 0.004067325498908758, + 0.0034755594097077847, + 0.0039179762825369835, + 0.0028128663543611765, + 0.004081160295754671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.033695679157972336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180515617132187 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180515617132187, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018180515617132187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13935981690883636, + 0.12653084099292755, + 0.0897308737039566, + 0.08561574667692184, + 0.08685244619846344, + 0.0677238255739212, + 0.04949311912059784, + 0.1953352391719818, + 0.113335981965065, + 0.04602210223674774 + ], + "credit_map": [ + 0.009354914538562298, + 0.01258858758956194, + 0.018587863072752953, + 0.0324416384100914, + 0.017637118697166443, + 0.011138570494949818, + 0.020422954112291336, + 0.0171881765127182, + 0.0114297429099679, + 0.015818458050489426, + 0.014355434104800224, + 0.026512479409575462, + 0.01866833120584488, + 0.009855085983872414, + 0.015027206391096115, + 0.01486410852521658, + 0.014639693312346935, + 0.007187001872807741, + 0.01640484854578972, + 0.017407234758138657, + 0.006305057089775801, + 0.008882906287908554, + 0.01072560902684927, + 0.008178523741662502, + 0.010252372361719608, + 0.009415067732334137, + 0.012948326766490936, + 0.017564255744218826, + 0.006406404543668032, + 0.008118662983179092, + 0.01164862047880888, + 0.009262035600841045, + 0.0071994890458881855, + 0.006643592845648527, + 0.011107539758086205, + 0.010647091083228588, + 0.006101483479142189, + 0.012584765441715717, + 0.015008553862571716, + 0.017559925094246864, + 0.005060361232608557, + 0.008752120658755302, + 0.00492670014500618, + 0.0082417456433177, + 0.008290549740195274, + 0.01272561401128769, + 0.01319451630115509, + 0.006532211787998676, + 0.0047603813000023365, + 0.006119130179286003, + 0.008700292557477951, + 0.005408280994743109, + 0.010236652567982674, + 0.0038473664317280054, + 0.005917178001254797, + 0.004503836389631033, + 0.029932724311947823, + 0.017386676743626595, + 0.016378164291381836, + 0.021095674484968185, + 0.024201396852731705, + 0.05081702023744583, + 0.016973713412880898, + 0.018549876287579536, + 0.008509380742907524, + 0.004990437533706427, + 0.030524427071213722, + 0.0051904101856052876, + 0.006313912104815245, + 0.04348566755652428, + 0.005349100101739168, + 0.008972643874585629, + 0.00949518196284771, + 0.006864197552204132, + 0.0089030209928751, + 0.004378628451377153, + 0.00321030686609447, + 0.004028310067951679, + 0.004368865862488747, + 0.0047735911794006824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.018180515617132187 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048634111881256104 + ], + "opposite_logprob_tokens": [ + -3.0486340522766113 + ], + "logprob_reward": -0.048634111881256104, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048634111881256104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03118940442800522, + 0.03490886464715004, + 0.024957768619060516, + 0.01925249584019184, + 0.033209752291440964, + 0.06723818928003311, + 0.0851307287812233, + 0.4613122045993805, + 0.18512363731861115, + 0.0576770156621933 + ], + "credit_map": [ + 0.002792950486764312, + 0.0036059000995010138, + 0.003883042838424444, + 0.004299211781471968, + 0.004499972797930241, + 0.0030778739601373672, + 0.005734904203563929, + 0.0032955463975667953, + 0.0031294028740376234, + 0.004871274344623089, + 0.004302060231566429, + 0.006373500917106867, + 0.005138315260410309, + 0.002718674950301647, + 0.004043439868837595, + 0.004332195967435837, + 0.0035869767889380455, + 0.0033453030046075583, + 0.003685935866087675, + 0.0029320160392671824, + 0.002270621480420232, + 0.0027630957774817944, + 0.003305296879261732, + 0.003068521386012435, + 0.0012966099893674254, + 0.002407850231975317, + 0.0026270966045558453, + 0.003038580296561122, + 0.0026564844883978367, + 0.002185387071222067, + 0.0030116341076791286, + 0.002028852701187134, + 0.003455950180068612, + 0.0038062925450503826, + 0.004391541238874197, + 0.0031718311365693808, + 0.003694360377267003, + 0.0051663825288414955, + 0.0030408231541514397, + 0.006482567172497511, + 0.004332696553319693, + 0.00969730969518423, + 0.0056133572943508625, + 0.012161735445261002, + 0.008242770098149776, + 0.011291882023215294, + 0.0067629823461174965, + 0.009135455824434757, + 0.008815404959022999, + 0.008180186152458191, + 0.008636567741632462, + 0.01362865511327982, + 0.021135276183485985, + 0.006286615040153265, + 0.012208398431539536, + 0.006239612586796284, + 0.06725757569074631, + 0.0428122915327549, + 0.04214411973953247, + 0.05211615562438965, + 0.0636456236243248, + 0.11365673691034317, + 0.03951868414878845, + 0.04016100987792015, + 0.011312819086015224, + 0.006168043706566095, + 0.05455705523490906, + 0.007796808145940304, + 0.010628978721797466, + 0.07353173196315765, + 0.007533331401646137, + 0.013594849966466427, + 0.016147561371326447, + 0.005310246720910072, + 0.02140512317419052, + 0.0030432529747486115, + 0.0023347411770373583, + 0.002520880429074168, + 0.002447139471769333, + 0.0044680698774755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.048634111881256104 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a slightly elevated angle and panning slowly to the right, captures a friendly children's soccer game in full swing in a sunny suburban park. On the neatly manicured, fresh-smelling emerald green lawn, two teams of children, approximately eight to ten years old, dressed in vibrant jerseys—one a dazzling lemon yellow, the other a brilliant royal blue—energize and chase a black and white soccer ball. The warm, golden afternoon sun streams across the field, elongating the soft shadows cast by the children as they run. The camera pans to the sidelines, focusing on a crowd of parents and siblings seated on folding canvas chairs or plaid picnic mats. A father in a gray hoodie excitedly rises from his chair, fists clenched, and cheers on his son on the field. Nearby, a smiling mother gently pats the back of her young daughter as she records the video on her phone. When a young player in a yellow jersey deftly dribbled the ball past the defense, the crowd erupted in cheers mixed with laughter and encouragement. The whole scene was filled with family warmth, healthy vitality, and pure joy, creating a community atmosphere of unity and friendship.", + "seed": 12596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_15/index_1950_seed_12596_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03071870831772685, + "vlm_logprob": -0.03071870831772685, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03071870831772685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03472764045000076 + ], + "opposite_logprob_tokens": [ + -4.284727573394775 + ], + "logprob_reward": -0.03472764045000076, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03472764045000076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06836975365877151, + 0.07636835426092148, + 0.03418079391121864, + 0.06438892334699631, + 0.05055263265967369, + 0.4480748772621155, + 0.1253560483455658, + 0.04301253333687782, + 0.061052411794662476, + 0.02864363230764866 + ], + "credit_map": [ + 0.0070149218663573265, + 0.007055358961224556, + 0.008815811946988106, + 0.009705747477710247, + 0.005964300595223904, + 0.005762872286140919, + 0.014768346212804317, + 0.009282397106289864, + 0.006163025740534067, + 0.006362866144627333, + 0.011115767993032932, + 0.015659989789128304, + 0.011281092651188374, + 0.007062332704663277, + 0.012250649742782116, + 0.006472633685916662, + 0.0020287607330828905, + 0.002578203333541751, + 0.0052466001361608505, + 0.006481599994003773, + 0.004133609589189291, + 0.0028283216524869204, + 0.005635756999254227, + 0.005247939378023148, + 0.005767939146608114, + 0.010011500678956509, + 0.008236506953835487, + 0.00843362882733345, + 0.01356099545955658, + 0.009484470821917057, + 0.005258341319859028, + 0.003635534318163991, + 0.0038654496893286705, + 0.008494564332067966, + 0.007218413054943085, + 0.008407936431467533, + 0.006858513690531254, + 0.0073575288988649845, + 0.0036574488040059805, + 0.004692775662988424, + 0.10636209696531296, + 0.05268504470586777, + 0.0635557547211647, + 0.024319427087903023, + 0.04945257306098938, + 0.07827049493789673, + 0.03773023188114166, + 0.035699278116226196, + 0.011299120262265205, + 0.006419999059289694, + 0.007119240239262581, + 0.011060466058552265, + 0.007130153477191925, + 0.027991732582449913, + 0.048076946288347244, + 0.006258388515561819, + 0.005987971555441618, + 0.004514702595770359, + 0.004984327591955662, + 0.0036649522371590137, + 0.007493306882679462, + 0.004793132655322552, + 0.007977212779223919, + 0.0035969268064945936, + 0.005344188306480646, + 0.015034974552690983, + 0.0065231104381382465, + 0.008348495699465275, + 0.0048191482201218605, + 0.004088253248482943, + 0.007125980220735073, + 0.009768260642886162, + 0.002647505374625325, + 0.003884863341227174, + 0.003583259880542755, + 0.003154946956783533, + 0.003077120054513216, + 0.0064096711575984955, + 0.002880414482206106, + 0.0030058519914746284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.03472764045000076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007649888750165701 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007649888750165701, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649888750165701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08225136250257492, + 0.06579093635082245, + 0.044212911278009415, + 0.046555645763874054, + 0.03871752694249153, + 0.17155666649341583, + 0.08035148680210114, + 0.06365357339382172, + 0.3884509205818176, + 0.018459001556038857 + ], + "credit_map": [ + 0.00833215657621622, + 0.007814440876245499, + 0.010780326090753078, + 0.011750606819987297, + 0.008570213802158833, + 0.007726348005235195, + 0.016998952254652977, + 0.010278326459228992, + 0.005422349087893963, + 0.0052591669373214245, + 0.008654097095131874, + 0.012494715861976147, + 0.007950842380523682, + 0.006025396753102541, + 0.012831737287342548, + 0.007152633275836706, + 0.003036890411749482, + 0.0036563414614647627, + 0.006621351465582848, + 0.007105263415724039, + 0.006598366890102625, + 0.005463340785354376, + 0.006840348243713379, + 0.004891012795269489, + 0.0031562396325170994, + 0.004093653056770563, + 0.008574229665100574, + 0.007570840418338776, + 0.0030875818338245153, + 0.0036465914454311132, + 0.008417487144470215, + 0.008009028621017933, + 0.0028122616931796074, + 0.0073657166212797165, + 0.004460161551833153, + 0.0055875713005661964, + 0.005422626622021198, + 0.0037198741920292377, + 0.005201233550906181, + 0.00414808327332139, + 0.035635028034448624, + 0.020772164687514305, + 0.024466700851917267, + 0.011210852302610874, + 0.02013644203543663, + 0.03119600936770439, + 0.013549558818340302, + 0.014589923433959484, + 0.009140516631305218, + 0.004612543620169163, + 0.005225918721407652, + 0.007845351472496986, + 0.006423671264201403, + 0.013084790669381618, + 0.029226943850517273, + 0.004791761748492718, + 0.009921587072312832, + 0.00756849953904748, + 0.012272964231669903, + 0.00879701878875494, + 0.0069008441641926765, + 0.00478404713794589, + 0.008092714473605156, + 0.00531590124592185, + 0.026597248390316963, + 0.08442968130111694, + 0.026569508016109467, + 0.1616635024547577, + 0.013184587471187115, + 0.021464018151164055, + 0.020341919735074043, + 0.03420045226812363, + 0.0023778066970407963, + 0.0020525006111711264, + 0.0019904046785086393, + 0.003572230227291584, + 0.0019259139662608504, + 0.002509724348783493, + 0.0022345525212585926, + 0.0017958712996914983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.007649888750165701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048651598393917084 + ], + "opposite_logprob_tokens": [ + -3.048651695251465 + ], + "logprob_reward": -0.048651598393917084, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048651598393917084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12290219217538834, + 0.38432031869888306, + 0.07803226262331009, + 0.10929550230503082, + 0.04846831038594246, + 0.05671953037381172, + 0.04115328565239906, + 0.12416985630989075, + 0.019639471545815468, + 0.015299202874302864 + ], + "credit_map": [ + 0.012730714865028858, + 0.01482432521879673, + 0.012681463733315468, + 0.02178763970732689, + 0.013091535307466984, + 0.013237220235168934, + 0.02061692252755165, + 0.013932377099990845, + 0.026267722249031067, + 0.031234515830874443, + 0.04667395353317261, + 0.0808955505490303, + 0.0448225736618042, + 0.0223261546343565, + 0.0980713739991188, + 0.03402850404381752, + 0.004264922346919775, + 0.00558409933000803, + 0.01298406533896923, + 0.015028362162411213, + 0.011493980884552002, + 0.010873653925955296, + 0.00904902908951044, + 0.008754147216677666, + 0.0069759744219481945, + 0.011537178419530392, + 0.018760574981570244, + 0.017667293548583984, + 0.011206395924091339, + 0.009135397151112556, + 0.01801900379359722, + 0.015993690118193626, + 0.003917306195944548, + 0.01090322807431221, + 0.005410641431808472, + 0.007133534178137779, + 0.00743440818041563, + 0.004778452683240175, + 0.005036697257310152, + 0.0038540400564670563, + 0.012045120820403099, + 0.005534101277589798, + 0.008095289580523968, + 0.0043767886236310005, + 0.006709444802254438, + 0.009022407233715057, + 0.005640799645334482, + 0.005295580253005028, + 0.010747887194156647, + 0.003230480710044503, + 0.0040892320685088634, + 0.0037857445422559977, + 0.005560684949159622, + 0.0035006594844162464, + 0.007210345007479191, + 0.003028252860531211, + 0.014355247840285301, + 0.01041898038238287, + 0.015079802833497524, + 0.03531253710389137, + 0.009040127508342266, + 0.004849366378039122, + 0.025685841217637062, + 0.009427955374121666, + 0.0024195730220526457, + 0.0026479416992515326, + 0.002869329648092389, + 0.002957521239295602, + 0.0027380238752812147, + 0.0021112102549523115, + 0.0021373212803155184, + 0.0017585493624210358, + 0.0015165491495281458, + 0.0018287986749783158, + 0.0020866598933935165, + 0.0021991324611008167, + 0.001441990491002798, + 0.0031469492241740227, + 0.0015074026305228472, + 0.001571720466017723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.048651598393917084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059960074722766876, + 0.0829220712184906, + 0.05848804861307144, + 0.0952831283211708, + 0.115007683634758, + 0.11259403824806213, + 0.08619751036167145, + 0.18924550712108612, + 0.18368396162986755, + 0.016617894172668457 + ], + "credit_map": [ + 0.009967618621885777, + 0.006385055370628834, + 0.006460650824010372, + 0.0074211610481143, + 0.008028320968151093, + 0.007824213244020939, + 0.006627541501075029, + 0.007245518267154694, + 0.00405882578343153, + 0.007242721505463123, + 0.012230315245687962, + 0.019010569900274277, + 0.012885363772511482, + 0.006424013525247574, + 0.014880252070724964, + 0.006190010812133551, + 0.0029830278363078833, + 0.0050264159217476845, + 0.009922748431563377, + 0.010821998119354248, + 0.010177917778491974, + 0.006749697029590607, + 0.006504247896373272, + 0.006302004214376211, + 0.006342125590890646, + 0.009098703041672707, + 0.01245956216007471, + 0.014441344887018204, + 0.008567072451114655, + 0.011798013001680374, + 0.017733976244926453, + 0.014842340722680092, + 0.010243472643196583, + 0.021301599219441414, + 0.014431921765208244, + 0.020234471186995506, + 0.014638662338256836, + 0.015759525820612907, + 0.009701911360025406, + 0.00869612954556942, + 0.02200605720281601, + 0.011895817704498768, + 0.015255289152264595, + 0.007132064551115036, + 0.010999221354722977, + 0.019172459840774536, + 0.0123603455722332, + 0.013772787526249886, + 0.019188232719898224, + 0.00710368063300848, + 0.009538275189697742, + 0.006766339763998985, + 0.007261720485985279, + 0.01468871533870697, + 0.015871018171310425, + 0.005779540631920099, + 0.023254316300153732, + 0.017353761941194534, + 0.02270858734846115, + 0.04959437996149063, + 0.016248835250735283, + 0.01196412742137909, + 0.03263076767325401, + 0.015490737743675709, + 0.012643998488783836, + 0.044042881578207016, + 0.015068870969116688, + 0.0732322633266449, + 0.007378377020359039, + 0.00988061260432005, + 0.012262333184480667, + 0.00917463656514883, + 0.0016345084877684712, + 0.0013704891316592693, + 0.002186461817473173, + 0.0022872230038046837, + 0.0024354232009500265, + 0.0026439803186804056, + 0.0019997546914964914, + 0.0020600545685738325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.007629185914993286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a children's soccer game taking place in a sunny suburban park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two teams of children visible, one wearing lemon yellow jerseys and the other wearing royal blue jerseys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black and white soccer ball visible being chased or dribbled by the children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a slightly elevated angle and panning slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a father in a gray hoodie visible rising from a chair to cheer on his son?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493522807955742 + ], + "opposite_logprob_tokens": [ + -2.9299352169036865 + ], + "logprob_reward": -0.05493522807955742, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493522807955742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10991116613149643, + 0.04354742169380188, + 0.04137970879673958, + 0.10205056518316269, + 0.16200606524944305, + 0.10127654671669006, + 0.18884773552417755, + 0.10027530044317245, + 0.12276521325111389, + 0.02794032357633114 + ], + "credit_map": [ + 0.01111286785453558, + 0.012227505445480347, + 0.012853961437940598, + 0.01530844159424305, + 0.013535557314753532, + 0.013986779376864433, + 0.019254466518759727, + 0.01163158193230629, + 0.002687560860067606, + 0.0029198681004345417, + 0.0066772159188985825, + 0.010033221915364265, + 0.005375283770263195, + 0.0031875590793788433, + 0.009059268981218338, + 0.003607445862144232, + 0.0020154528319835663, + 0.002397253643721342, + 0.0099626574665308, + 0.009738403372466564, + 0.0035757410805672407, + 0.005284366197884083, + 0.004547472111880779, + 0.003858362790197134, + 0.005932510364800692, + 0.012588773854076862, + 0.011405473574995995, + 0.012398475781083107, + 0.014931567013263702, + 0.01061940100044012, + 0.019524596631526947, + 0.01464975904673338, + 0.012678872793912888, + 0.022888552397489548, + 0.020860187709331512, + 0.023689262568950653, + 0.02957059256732464, + 0.024411514401435852, + 0.014167615212500095, + 0.013739475980401039, + 0.008218836970627308, + 0.008427257649600506, + 0.013944552280008793, + 0.012585218995809555, + 0.013980395160615444, + 0.006641735322773457, + 0.020772187039256096, + 0.01670636050403118, + 0.03061852790415287, + 0.013494077138602734, + 0.013348951004445553, + 0.019629159942269325, + 0.01570110023021698, + 0.018321970477700233, + 0.06665817648172379, + 0.01107577234506607, + 0.011978231370449066, + 0.011805243790149689, + 0.014340830035507679, + 0.02093389444053173, + 0.01006417628377676, + 0.007999014109373093, + 0.012926995754241943, + 0.010226909071207047, + 0.01177777349948883, + 0.029110047966241837, + 0.008897810243070126, + 0.04487340897321701, + 0.00541211711242795, + 0.006561354734003544, + 0.006540075875818729, + 0.009592633694410324, + 0.0032307221554219723, + 0.003046952420845628, + 0.004138868767768145, + 0.0029352945275604725, + 0.002999653806909919, + 0.004967134445905685, + 0.004241585731506348, + 0.0023801110219210386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.05493522807955742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_tensor([4817, 4817])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_tensor([4817, 4817])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c6bed17337d7f159b3e3437a1c6e5f5579d10e6f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_tensor([4817, 4817])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, tracking shot, close to the ground, follows two young men as they effortlessly glide across a futuristic city square at dusk on their electric unicycles. The young man on the left, with shaggy black hair and a tech-fabric black hoodie with reflective silver stripes on the sleeves, wears a broad smile. The young woman next to him, with her long, bright pink hair tied in a high ponytail, throws her head back in a sleek white bomber jacket and laughs heartily. The unicycles they ride are matte black, their futuristic LED strips emitting brilliant cyan and magenta light on either side, casting vibrant reflections on the polished dark stone floor, leaving long trails of light as the wheels spin at breakneck speed. The camera switches focus swiftly and precisely between close-ups of their vivid, joyful facial expressions and the wheels spinning so rapidly that they appear almost animate, against a backdrop of towering skyscrapers, their windows lit up, and flashing holographic billboards. The entire picture has a cold tone, mainly dark blue and black, but is illuminated by the neon lights of the unicycle and the warm lights in the distance, interspersed with slight, stylized digital screen flashes and lens flares that represent a sense of technology, creating a futuristic atmosphere that is free, cheerful and full of urban dynamism.", + "seed": 22843, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_4817_seed_22843_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03459826149046421, + "vlm_logprob": -0.03459826149046421, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03459826149046421, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035734958946704865 + ], + "opposite_logprob_tokens": [ + -3.7857348918914795 + ], + "logprob_reward": -0.035734958946704865, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.035734958946704865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15783900022506714, + 0.06757008284330368, + 0.12477196007966995, + 0.040495071560144424, + 0.13036096096038818, + 0.10584847629070282, + 0.10815523564815521, + 0.032827768474817276, + 0.06670238077640533, + 0.16542915999889374 + ], + "credit_map": [ + 0.013173858635127544, + 0.01393570750951767, + 0.02272973395884037, + 0.028626414015889168, + 0.01601451449096203, + 0.01843845285475254, + 0.021789932623505592, + 0.023130353540182114, + 0.009428860619664192, + 0.007720572408288717, + 0.006197446491569281, + 0.007997809909284115, + 0.008946442976593971, + 0.006931927055120468, + 0.010913549922406673, + 0.00943346694111824, + 0.008036650717258453, + 0.010265683755278587, + 0.007676088251173496, + 0.020648948848247528, + 0.012535464018583298, + 0.030093196779489517, + 0.020588383078575134, + 0.014927529729902744, + 0.0045621683821082115, + 0.0035392497666180134, + 0.004365328699350357, + 0.005470436066389084, + 0.0057856072671711445, + 0.00500286091119051, + 0.006560527253895998, + 0.005208886694163084, + 0.011909243650734425, + 0.012863731943070889, + 0.01986618898808956, + 0.013191677629947662, + 0.012088069692254066, + 0.02835911139845848, + 0.01674896664917469, + 0.015333963558077812, + 0.01661936566233635, + 0.01197793334722519, + 0.011080224066972733, + 0.00992164108902216, + 0.013380919583141804, + 0.018362021073698997, + 0.010515166446566582, + 0.013991194777190685, + 0.012170006521046162, + 0.016074884682893753, + 0.01292747724801302, + 0.013927306979894638, + 0.013422952964901924, + 0.01350532565265894, + 0.017104951664805412, + 0.009022322483360767, + 0.004914541263133287, + 0.003484841203317046, + 0.00504379952326417, + 0.0036666044034063816, + 0.002828501397743821, + 0.005647765472531319, + 0.0031538307666778564, + 0.0040878779254853725, + 0.005810327362269163, + 0.005965758115053177, + 0.006986408494412899, + 0.007428973913192749, + 0.005891832523047924, + 0.017394937574863434, + 0.00718046585097909, + 0.010043667629361153, + 0.006316795479506254, + 0.005928208120167255, + 0.01077644806355238, + 0.009785821661353111, + 0.011332343332469463, + 0.05505967512726784, + 0.010427819564938545, + 0.05580202117562294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.035734958946704865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088299565017223 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088299565017223, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088299565017223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06552039086818695, + 0.016814397647976875, + 0.09696543961763382, + 0.0284198597073555, + 0.03659731522202492, + 0.03591718152165413, + 0.1668611466884613, + 0.033021070063114166, + 0.052487559616565704, + 0.4673956036567688 + ], + "credit_map": [ + 0.004067267291247845, + 0.006625707261264324, + 0.009086203761398792, + 0.007433625403791666, + 0.009546120651066303, + 0.0122371232137084, + 0.00714343274012208, + 0.009380904957652092, + 0.001601053518243134, + 0.0019658247474581003, + 0.0019219386158511043, + 0.002141195349395275, + 0.001865818747319281, + 0.0018101936439052224, + 0.003110058605670929, + 0.002398315817117691, + 0.0039472137577831745, + 0.0061722407117486, + 0.011216329410672188, + 0.0279054157435894, + 0.010136323049664497, + 0.010394112206995487, + 0.013337220065295696, + 0.013856583274900913, + 0.0019045426743105054, + 0.0020857478957623243, + 0.004219017922878265, + 0.002812577411532402, + 0.005267420783638954, + 0.0026400303468108177, + 0.006017505656927824, + 0.0034730175975710154, + 0.0036814052145928144, + 0.006022075656801462, + 0.0018592124106362462, + 0.005798637866973877, + 0.0031111103016883135, + 0.005923047196120024, + 0.005583252292126417, + 0.004618574399501085, + 0.004231607541441917, + 0.007374508306384087, + 0.0033042277209460735, + 0.005063782911747694, + 0.0032748831436038017, + 0.003418281441554427, + 0.0055360873229801655, + 0.0037138022016733885, + 0.01855778880417347, + 0.023605287075042725, + 0.019427867606282234, + 0.019288256764411926, + 0.0215262733399868, + 0.023969944566488266, + 0.025430895388126373, + 0.015054828487336636, + 0.0029086389113217592, + 0.0037227615248411894, + 0.002983465790748596, + 0.004985273350030184, + 0.0031771769281476736, + 0.006856620777398348, + 0.0030223382636904716, + 0.0053647966124117374, + 0.003430579323321581, + 0.0042111133225262165, + 0.00766793405637145, + 0.005392452236264944, + 0.0034730019979178905, + 0.0077792382799088955, + 0.014345129020512104, + 0.00618811184540391, + 0.019578615203499794, + 0.038072988390922546, + 0.01960260048508644, + 0.03585676848888397, + 0.04772234335541725, + 0.1424209475517273, + 0.02416842058300972, + 0.13997291028499603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.011088299565017223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08421187102794647, + 0.014627971686422825, + 0.062287408858537674, + 0.029643015936017036, + 0.016006026417016983, + 0.025195898488163948, + 0.05140643194317818, + 0.017516430467367172, + 0.049259528517723083, + 0.6498454809188843 + ], + "credit_map": [ + 0.004605158697813749, + 0.006346643902361393, + 0.011887652799487114, + 0.005440984852612019, + 0.011396288871765137, + 0.014356723055243492, + 0.00880407728254795, + 0.02137434110045433, + 0.0016599927330389619, + 0.0016520139761269093, + 0.0016633769264444709, + 0.0016335950931534171, + 0.001853786176070571, + 0.0016738784033805132, + 0.002484657568857074, + 0.0020066711585968733, + 0.0022400482557713985, + 0.004233540501445532, + 0.006643877364695072, + 0.017880506813526154, + 0.00786405336111784, + 0.006026152055710554, + 0.008361263200640678, + 0.009037972427904606, + 0.0018725197296589613, + 0.0022729465272277594, + 0.003957684617489576, + 0.003090521087870002, + 0.0053230952471494675, + 0.002776992740109563, + 0.006519349757581949, + 0.003829908324405551, + 0.0015659458003938198, + 0.0026897285133600235, + 0.0013420794857665896, + 0.0021126242354512215, + 0.0015218735206872225, + 0.0016588644357398152, + 0.0031776195392012596, + 0.0019372891401872039, + 0.002937351819127798, + 0.005114221479743719, + 0.0023821708746254444, + 0.0038719819858670235, + 0.003240149701014161, + 0.002208438701927662, + 0.0027400259859859943, + 0.00270155631005764, + 0.00577893853187561, + 0.006339280866086483, + 0.006084623280912638, + 0.005172760225832462, + 0.00801683496683836, + 0.006214907392859459, + 0.00857977382838726, + 0.005219312850385904, + 0.0014735813019797206, + 0.0017514190403744578, + 0.0017919610254466534, + 0.003360254690051079, + 0.0015659488271921873, + 0.002793662017211318, + 0.0016026466619223356, + 0.0031769576016813517, + 0.003457802813500166, + 0.004105817526578903, + 0.008379665203392506, + 0.004907791502773762, + 0.003526512999087572, + 0.006238813046365976, + 0.012071983888745308, + 0.006571140140295029, + 0.027736062183976173, + 0.051656898111104965, + 0.030827168375253677, + 0.05717581883072853, + 0.061630845069885254, + 0.1930203139781952, + 0.03502584993839264, + 0.19277246296405792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.004639456048607826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863425899296999 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.00863425899296999, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863425899296999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17325812578201294, + 0.07918756455183029, + 0.31612443923950195, + 0.02759668044745922, + 0.052237384021282196, + 0.03597423434257507, + 0.04453660547733307, + 0.040307480841875076, + 0.06819729506969452, + 0.162580206990242 + ], + "credit_map": [ + 0.010929223150014877, + 0.018891409039497375, + 0.017573613673448563, + 0.018320217728614807, + 0.018785400316119194, + 0.02176959067583084, + 0.03286334127187729, + 0.034125328063964844, + 0.007075774949043989, + 0.006293796468526125, + 0.008185457438230515, + 0.010402671061456203, + 0.012181206606328487, + 0.010370373725891113, + 0.012976376339793205, + 0.01170190330594778, + 0.019899414852261543, + 0.019589629024267197, + 0.02772141806781292, + 0.07083797454833984, + 0.039702367037534714, + 0.040178507566452026, + 0.0586521252989769, + 0.039542991667985916, + 0.0026692592073231936, + 0.0028601372614502907, + 0.0037222958635538816, + 0.0034355968236923218, + 0.004535543266683817, + 0.0037347893230617046, + 0.0031257998198270798, + 0.00351326004602015, + 0.00651916628703475, + 0.007786526344716549, + 0.004243460949510336, + 0.006757339928299189, + 0.0046416642144322395, + 0.009207486175000668, + 0.006344306748360395, + 0.006737437564879656, + 0.004729622043669224, + 0.004277700092643499, + 0.0035292143002152443, + 0.004999236203730106, + 0.004291203338652849, + 0.005583150777965784, + 0.005315208341926336, + 0.003248899709433317, + 0.005903364159166813, + 0.005022325087338686, + 0.005938265006989241, + 0.005661475472152233, + 0.0069817774929106236, + 0.00416332995519042, + 0.006306415889412165, + 0.004559651482850313, + 0.003901999443769455, + 0.004479559138417244, + 0.0070803202688694, + 0.005048920400440693, + 0.004304137080907822, + 0.005691776983439922, + 0.004439977928996086, + 0.005360787734389305, + 0.005684571340680122, + 0.005268268287181854, + 0.011477425694465637, + 0.006919249426573515, + 0.005200529936701059, + 0.011829953640699387, + 0.01371274795383215, + 0.00810454785823822, + 0.01005546748638153, + 0.02246467024087906, + 0.012780207209289074, + 0.03094860352575779, + 0.017708150669932365, + 0.027069900184869766, + 0.01659368723630905, + 0.02495952509343624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.00863425899296999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289433389902115 + ], + "opposite_logprob_tokens": [ + -2.237894296646118 + ], + "logprob_reward": -0.11289433389902115, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289433389902115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13519662618637085, + 0.06327898055315018, + 0.09377662092447281, + 0.047557730227708817, + 0.03917841613292694, + 0.08375751972198486, + 0.3625480830669403, + 0.030772002413868904, + 0.07915935665369034, + 0.0647747665643692 + ], + "credit_map": [ + 0.00738014979287982, + 0.009141338057816029, + 0.03175267577171326, + 0.010775300674140453, + 0.016434092074632645, + 0.0186766404658556, + 0.011364178732037544, + 0.02967223711311817, + 0.006749317049980164, + 0.009681456722319126, + 0.008371560834348202, + 0.008951584808528423, + 0.005144152324646711, + 0.006374758202582598, + 0.008162586949765682, + 0.009843552485108376, + 0.0062041436322033405, + 0.007141090929508209, + 0.009581387974321842, + 0.021276172250509262, + 0.010343984700739384, + 0.013200740329921246, + 0.015318010933697224, + 0.010711084119975567, + 0.0032998777460306883, + 0.004476366098970175, + 0.007240667007863522, + 0.009686813689768314, + 0.005202482920140028, + 0.004822563845664263, + 0.007145332638174295, + 0.005683620925992727, + 0.004413147456943989, + 0.00495752552524209, + 0.0033630249090492725, + 0.006437226198613644, + 0.003963161259889603, + 0.005500786937773228, + 0.006434840150177479, + 0.004108701832592487, + 0.010755886323750019, + 0.01015706080943346, + 0.007735655177384615, + 0.012977402657270432, + 0.010411351919174194, + 0.008949757553637028, + 0.014624182134866714, + 0.008146215230226517, + 0.04270152002573013, + 0.04906248673796654, + 0.04585651680827141, + 0.043093811720609665, + 0.04412495344877243, + 0.05243339762091637, + 0.05102564021945, + 0.034249719232320786, + 0.0033785176929086447, + 0.003635415341705084, + 0.005111314356327057, + 0.00441919406875968, + 0.0038972797337919474, + 0.003301301971077919, + 0.0038967090658843517, + 0.003132268786430359, + 0.0064425659365952015, + 0.004982872866094112, + 0.013965999707579613, + 0.007464317139238119, + 0.007154840510338545, + 0.011083907447755337, + 0.015402616932988167, + 0.01266222633421421, + 0.004797691013664007, + 0.006947360001504421, + 0.01120743341743946, + 0.0063103423453867435, + 0.007735798601061106, + 0.010367057286202908, + 0.00573303597047925, + 0.011676045134663582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.11289433389902115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, tracking shot, close to the ground, follows two young men as they effortlessly glide across a futuristic city square at dusk on their electric unicycles. The young man on the left, with shaggy black hair and a tech-fabric black hoodie with reflective silver stripes on the sleeves, wears a broad smile. The young woman next to him, with her long, bright pink hair tied in a high ponytail, throws her head back in a sleek white bomber jacket and laughs heartily. The unicycles they ride are matte black, their futuristic LED strips emitting brilliant cyan and magenta light on either side, casting vibrant reflections on the polished dark stone floor, leaving long trails of light as the wheels spin at breakneck speed. The camera switches focus swiftly and precisely between close-ups of their vivid, joyful facial expressions and the wheels spinning so rapidly that they appear almost animate, against a backdrop of towering skyscrapers, their windows lit up, and flashing holographic billboards. The entire picture has a cold tone, mainly dark blue and black, but is illuminated by the neon lights of the unicycle and the warm lights in the distance, interspersed with slight, stylized digital screen flashes and lens flares that represent a sense of technology, creating a futuristic atmosphere that is free, cheerful and full of urban dynamism.", + "seed": 22844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_4817_seed_22844_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02481126682832837, + "vlm_logprob": -0.02481126682832837, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02481126682832837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0688156932592392 + ], + "opposite_logprob_tokens": [ + -3.0688157081604004 + ], + "logprob_reward": -0.0688156932592392, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0688156932592392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18786965310573578, + 0.10228925943374634, + 0.4038439095020294, + 0.03904120996594429, + 0.04467171058058739, + 0.040213268250226974, + 0.05219157785177231, + 0.041390810161828995, + 0.03455545753240585, + 0.05393316224217415 + ], + "credit_map": [ + 0.013694051653146744, + 0.016701621934771538, + 0.018903501331806183, + 0.01682766154408455, + 0.015662483870983124, + 0.042131055146455765, + 0.01628517545759678, + 0.04766407981514931, + 0.012456534430384636, + 0.008980530314147472, + 0.012872118502855301, + 0.012018841691315174, + 0.008303852751851082, + 0.014660842716693878, + 0.017930587753653526, + 0.015065940096974373, + 0.01183080580085516, + 0.006820610258728266, + 0.03853226080536842, + 0.012623446062207222, + 0.009232054464519024, + 0.15999169647693634, + 0.023364106193184853, + 0.14144888520240784, + 0.003946474753320217, + 0.003998544067144394, + 0.0053365291096270084, + 0.005293093156069517, + 0.005281435325741768, + 0.005387419369071722, + 0.004795982036739588, + 0.005001729354262352, + 0.005712225567549467, + 0.004335844423621893, + 0.008067307993769646, + 0.004742007236927748, + 0.004393396899104118, + 0.0031390332151204348, + 0.006120864301919937, + 0.008161026984453201, + 0.006228263024240732, + 0.00436922162771225, + 0.004038581624627113, + 0.006274264771491289, + 0.004719962365925312, + 0.0037774601951241493, + 0.0069243828766047955, + 0.00388112710788846, + 0.009050972759723663, + 0.006895408499985933, + 0.007316334638744593, + 0.0055638388730585575, + 0.007202341686934233, + 0.008019935339689255, + 0.0033982149325311184, + 0.004744524136185646, + 0.007290459703654051, + 0.006258520297706127, + 0.00666071567684412, + 0.005498239304870367, + 0.0035478712525218725, + 0.0034275532234460115, + 0.0034865771885961294, + 0.005220865830779076, + 0.004113734234124422, + 0.0033219216857105494, + 0.0045092846266925335, + 0.00464264303445816, + 0.00708311190828681, + 0.0037578882183879614, + 0.004281224682927132, + 0.0028456468135118484, + 0.004535497631877661, + 0.005514797754585743, + 0.007291657850146294, + 0.007988429628312588, + 0.009431659244000912, + 0.006716207135468721, + 0.00712922727689147, + 0.005325681529939175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0688156932592392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646339945495129 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646339945495129, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646339945495129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09304425865411758, + 0.09010715037584305, + 0.1839018017053604, + 0.08632409572601318, + 0.10065826773643494, + 0.04477505013346672, + 0.15630584955215454, + 0.06527262926101685, + 0.14645498991012573, + 0.033155933022499084 + ], + "credit_map": [ + 0.007762862369418144, + 0.00646204873919487, + 0.007676407229155302, + 0.006284734234213829, + 0.012491941452026367, + 0.023572802543640137, + 0.007958335801959038, + 0.020835118368268013, + 0.009631277993321419, + 0.006975193042308092, + 0.01588280312716961, + 0.010057337582111359, + 0.005705596879124641, + 0.014963170513510704, + 0.010165310464799404, + 0.016726454719901085, + 0.00687920767813921, + 0.005455105099827051, + 0.015012440271675587, + 0.00811087153851986, + 0.008398742415010929, + 0.0643235594034195, + 0.01751231960952282, + 0.058209557086229324, + 0.009172987192869186, + 0.010360073298215866, + 0.011621900834143162, + 0.011109726503491402, + 0.012164250947535038, + 0.011278596706688404, + 0.008344696834683418, + 0.012271854095160961, + 0.014064433053135872, + 0.009547492489218712, + 0.013206392526626587, + 0.01465483009815216, + 0.00980707723647356, + 0.01087978295981884, + 0.015115699730813503, + 0.013382556848227978, + 0.008237221278250217, + 0.004888234194368124, + 0.004948446061462164, + 0.007940595969557762, + 0.0029788396786898375, + 0.004902305081486702, + 0.006099533289670944, + 0.004779869224876165, + 0.0254774559289217, + 0.0189700685441494, + 0.0197641272097826, + 0.02339814603328705, + 0.01598937064409256, + 0.02193286083638668, + 0.013170684687793255, + 0.017603131011128426, + 0.006010038778185844, + 0.006931796669960022, + 0.0056090219877660275, + 0.006522654555737972, + 0.006849917117506266, + 0.009060895070433617, + 0.009780511260032654, + 0.014507780782878399, + 0.014866843819618225, + 0.013252072967588902, + 0.014312364161014557, + 0.01756991818547249, + 0.033374492079019547, + 0.019573520869016647, + 0.02126907929778099, + 0.012236694805324078, + 0.003361708950251341, + 0.004647436086088419, + 0.003170135896652937, + 0.005161082372069359, + 0.004519292153418064, + 0.003897556336596608, + 0.004554563667625189, + 0.0038441536016762257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.007646339945495129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036183618940413 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036183618940413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036183618940413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12719976902008057, + 0.14245373010635376, + 0.12361927330493927, + 0.0815475732088089, + 0.07264147698879242, + 0.06869008392095566, + 0.14546507596969604, + 0.09810960292816162, + 0.05119919404387474, + 0.08907429873943329 + ], + "credit_map": [ + 0.009749424643814564, + 0.01029615942388773, + 0.01007821410894394, + 0.010195116512477398, + 0.022870704531669617, + 0.028485817834734917, + 0.013585121370851994, + 0.02193920686841011, + 0.012368730269372463, + 0.012034676969051361, + 0.030066778883337975, + 0.012467600405216217, + 0.010304407216608524, + 0.02031436376273632, + 0.011865762062370777, + 0.033031389117240906, + 0.005449293646961451, + 0.005594258662313223, + 0.0143038434907794, + 0.009167800657451153, + 0.006210243795067072, + 0.04010894522070885, + 0.011515996418893337, + 0.0312688834965229, + 0.007397128269076347, + 0.008665680885314941, + 0.007945749908685684, + 0.009408478625118732, + 0.012080347165465355, + 0.01064152829349041, + 0.013755988329648972, + 0.011652661487460136, + 0.01202194020152092, + 0.0069846585392951965, + 0.010055615566670895, + 0.0080493139103055, + 0.00929155945777893, + 0.007529046852141619, + 0.009306112304329872, + 0.009403222240507603, + 0.008326913230121136, + 0.008202461525797844, + 0.007279688958078623, + 0.012369193136692047, + 0.006853337399661541, + 0.008204315789043903, + 0.010593006387352943, + 0.0068611628375947475, + 0.020146792754530907, + 0.01937846653163433, + 0.018212007358670235, + 0.016317861154675484, + 0.01653353124856949, + 0.029126742854714394, + 0.011223895475268364, + 0.01452576369047165, + 0.01184349600225687, + 0.013027693144977093, + 0.013301873579621315, + 0.010695435106754303, + 0.008817114867269993, + 0.012367510236799717, + 0.012163771316409111, + 0.015892697498202324, + 0.005567642394453287, + 0.004334606695920229, + 0.00794411264359951, + 0.006914264056831598, + 0.011765698902308941, + 0.004291456658393145, + 0.006356252823024988, + 0.004025150090456009, + 0.005831481888890266, + 0.012124346569180489, + 0.006318202242255211, + 0.012876659631729126, + 0.016633164137601852, + 0.009298738092184067, + 0.013621374033391476, + 0.012370328418910503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0036183618940413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189188368618488 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189188368618488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189188368618488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06908311694860458, + 0.17196141183376312, + 0.22121326625347137, + 0.16117747128009796, + 0.08080202341079712, + 0.04599924013018608, + 0.0782625675201416, + 0.07176277041435242, + 0.05127092823386192, + 0.04846722632646561 + ], + "credit_map": [ + 0.0070757572539150715, + 0.005405017640441656, + 0.009301450103521347, + 0.006037569139152765, + 0.009510988369584084, + 0.012985628098249435, + 0.00836676824837923, + 0.010399926453828812, + 0.014909244142472744, + 0.018978556618094444, + 0.02941334806382656, + 0.016216309741139412, + 0.014809658750891685, + 0.026498811319470406, + 0.01858440227806568, + 0.03255106881260872, + 0.011288135312497616, + 0.006515694782137871, + 0.021380223333835602, + 0.012606047093868256, + 0.008622475899755955, + 0.07787566632032394, + 0.017513733357191086, + 0.0654112845659256, + 0.01548051368445158, + 0.013531682081520557, + 0.01637793704867363, + 0.01771743781864643, + 0.017864663153886795, + 0.026014111936092377, + 0.017981644719839096, + 0.03620946779847145, + 0.01054392196238041, + 0.008313436061143875, + 0.011378192342817783, + 0.009677985683083534, + 0.00836026668548584, + 0.0077056786976754665, + 0.014437878504395485, + 0.010384656488895416, + 0.004900163505226374, + 0.004740789067000151, + 0.004940493497997522, + 0.009005137719213963, + 0.00400230847299099, + 0.00555731076747179, + 0.006873250938951969, + 0.005979782901704311, + 0.010636999271810055, + 0.010021446272730827, + 0.010607264004647732, + 0.006884023547172546, + 0.011434617452323437, + 0.01390225999057293, + 0.0050092171877622604, + 0.009766729548573494, + 0.011372607201337814, + 0.009777206927537918, + 0.012685252353549004, + 0.009214353747665882, + 0.006336749531328678, + 0.006414438132196665, + 0.00886506401002407, + 0.007097095251083374, + 0.00528968358412385, + 0.004393823444843292, + 0.005125639960169792, + 0.007034195587038994, + 0.010396492667496204, + 0.006202505901455879, + 0.007968044839799404, + 0.004860542248934507, + 0.004437233787029982, + 0.005618661642074585, + 0.004314879421144724, + 0.011146657168865204, + 0.008563674986362457, + 0.004870113916695118, + 0.005084118340164423, + 0.004431882873177528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.014189188368618488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978675067424774 + ], + "opposite_logprob_tokens": [ + -3.5297868251800537 + ], + "logprob_reward": -0.02978675067424774, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978675067424774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0716637670993805, + 0.28236886858940125, + 0.2916313111782074, + 0.03900574520230293, + 0.0849996954202652, + 0.03781279921531677, + 0.03066280670464039, + 0.06159193813800812, + 0.02953638695180416, + 0.07072668522596359 + ], + "credit_map": [ + 0.0065614632330834866, + 0.0046144393272697926, + 0.007656926289200783, + 0.006229712162166834, + 0.005405592266470194, + 0.017075300216674805, + 0.007996778003871441, + 0.016123555600643158, + 0.02473609149456024, + 0.03588704764842987, + 0.04474906623363495, + 0.03154388815164566, + 0.02648627944290638, + 0.0340193472802639, + 0.039648544043302536, + 0.04529861360788345, + 0.010648434981703758, + 0.008568699471652508, + 0.02669878490269184, + 0.014832882210612297, + 0.010000539012253284, + 0.11031059175729752, + 0.017551343888044357, + 0.09302002191543579, + 0.0033568302169442177, + 0.003174643265083432, + 0.0038581793196499348, + 0.004175885580480099, + 0.0051011089235544205, + 0.005940054543316364, + 0.004493173211812973, + 0.008905869908630848, + 0.012451198883354664, + 0.008561547845602036, + 0.011506161652505398, + 0.011550524272024632, + 0.008796053938567638, + 0.008456672541797161, + 0.01432831957936287, + 0.009349213913083076, + 0.004628276452422142, + 0.0037028540391474962, + 0.003584206337109208, + 0.004567979369312525, + 0.006409374065697193, + 0.0052413977682590485, + 0.0063053700141608715, + 0.003373337909579277, + 0.004397353623062372, + 0.004436796996742487, + 0.0036592388059943914, + 0.004065190441906452, + 0.0030541277956217527, + 0.004319378174841404, + 0.0025406810455024242, + 0.004190040752291679, + 0.011534163728356361, + 0.006209405139088631, + 0.010007338598370552, + 0.00788917113095522, + 0.0049658832140266895, + 0.00679227989166975, + 0.006501148920506239, + 0.007692547515034676, + 0.002962671685963869, + 0.002894971054047346, + 0.004151416942477226, + 0.003731631441041827, + 0.0034932096023112535, + 0.004286317154765129, + 0.0042776563204824924, + 0.0037385120522230864, + 0.005536434706300497, + 0.006553116720169783, + 0.004204362630844116, + 0.0069127208553254604, + 0.013434752821922302, + 0.006000509485602379, + 0.0180154237896204, + 0.010069360956549644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.02978675067424774 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, tracking shot, close to the ground, follows two young men as they effortlessly glide across a futuristic city square at dusk on their electric unicycles. The young man on the left, with shaggy black hair and a tech-fabric black hoodie with reflective silver stripes on the sleeves, wears a broad smile. The young woman next to him, with her long, bright pink hair tied in a high ponytail, throws her head back in a sleek white bomber jacket and laughs heartily. The unicycles they ride are matte black, their futuristic LED strips emitting brilliant cyan and magenta light on either side, casting vibrant reflections on the polished dark stone floor, leaving long trails of light as the wheels spin at breakneck speed. The camera switches focus swiftly and precisely between close-ups of their vivid, joyful facial expressions and the wheels spinning so rapidly that they appear almost animate, against a backdrop of towering skyscrapers, their windows lit up, and flashing holographic billboards. The entire picture has a cold tone, mainly dark blue and black, but is illuminated by the neon lights of the unicycle and the warm lights in the distance, interspersed with slight, stylized digital screen flashes and lens flares that represent a sense of technology, creating a futuristic atmosphere that is free, cheerful and full of urban dynamism.", + "seed": 22845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_4817_seed_22845_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018579246336594225, + "vlm_logprob": -0.018579246336594225, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.018579246336594225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029195524752140045 + ], + "opposite_logprob_tokens": [ + -3.904195547103882 + ], + "logprob_reward": -0.029195524752140045, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029195524752140045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06479363888502121, + 0.05983076989650726, + 0.030300604179501534, + 0.49725663661956787, + 0.04625367000699043, + 0.11946805566549301, + 0.09663001447916031, + 0.020021159201860428, + 0.041642796248197556, + 0.023802585899829865 + ], + "credit_map": [ + 0.007947725243866444, + 0.007283431477844715, + 0.007337597664445639, + 0.007993925362825394, + 0.010274720378220081, + 0.010129643604159355, + 0.004984848201274872, + 0.008841749280691147, + 0.010969337075948715, + 0.01058648806065321, + 0.007703977636992931, + 0.006641557905822992, + 0.004186921752989292, + 0.0062090386636555195, + 0.004339990206062794, + 0.009193460457026958, + 0.002957310061901808, + 0.002767870668321848, + 0.003553861752152443, + 0.004693667870014906, + 0.0028728623874485493, + 0.004724026191979647, + 0.004658421967178583, + 0.004072586074471474, + 0.0367412269115448, + 0.011352655477821827, + 0.16693967580795288, + 0.037737950682640076, + 0.012544794008135796, + 0.03821461647748947, + 0.018622728064656258, + 0.17510302364826202, + 0.006178345065563917, + 0.004242801107466221, + 0.006577275227755308, + 0.004813255276530981, + 0.004324394278228283, + 0.008047160692512989, + 0.004343456123024225, + 0.007726981770247221, + 0.04419731721282005, + 0.007852950133383274, + 0.03862697258591652, + 0.005955650005489588, + 0.006835498847067356, + 0.00569259375333786, + 0.004811340942978859, + 0.005495740566402674, + 0.03292933106422424, + 0.0044358582235872746, + 0.0058580199256539345, + 0.004105646628886461, + 0.0028068942483514547, + 0.0053504775278270245, + 0.005157608073204756, + 0.03598618134856224, + 0.0023374103475362062, + 0.002484778640791774, + 0.00254815979860723, + 0.0020982392597943544, + 0.0039494833908975124, + 0.0019637050572782755, + 0.0026445353869348764, + 0.0019948487170040607, + 0.00651771854609251, + 0.004396566189825535, + 0.004875588696449995, + 0.0031689750030636787, + 0.0038556386716663837, + 0.008040981367230415, + 0.005201808642596006, + 0.005585520062595606, + 0.003035841276869178, + 0.004266722127795219, + 0.002612336305901408, + 0.002566130366176367, + 0.0020659330766648054, + 0.003681342350319028, + 0.002748481696471572, + 0.0028257989324629307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.029195524752140045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254948042333126 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.01254948042333126, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254948042333126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06934335082769394, + 0.03725353628396988, + 0.028793450444936752, + 0.3185124099254608, + 0.0638139545917511, + 0.2686269283294678, + 0.072567418217659, + 0.06964724510908127, + 0.049794603139162064, + 0.02164716273546219 + ], + "credit_map": [ + 0.004569624084979296, + 0.006748332176357508, + 0.009742495603859425, + 0.013724682852625847, + 0.005828664638102055, + 0.007743020076304674, + 0.00816409382969141, + 0.012822428718209267, + 0.003918985370546579, + 0.002607928588986397, + 0.00677125807851553, + 0.005974584724754095, + 0.0028693510685116053, + 0.005902555771172047, + 0.0032111702021211386, + 0.005997699685394764, + 0.0025893556885421276, + 0.003194336546584964, + 0.003197197802364826, + 0.004176360089331865, + 0.0032046071719378233, + 0.005353492219001055, + 0.003543598810210824, + 0.0035345014184713364, + 0.0252256877720356, + 0.01582307741045952, + 0.0924561470746994, + 0.02469506859779358, + 0.014868604950606823, + 0.030894655734300613, + 0.013842035084962845, + 0.10070708394050598, + 0.005912836641073227, + 0.009661072865128517, + 0.007721561938524246, + 0.009748123586177826, + 0.009451989084482193, + 0.004678911529481411, + 0.010174951516091824, + 0.006464498117566109, + 0.0955469012260437, + 0.016088617965579033, + 0.09858536720275879, + 0.012101572938263416, + 0.012018129229545593, + 0.008826160803437233, + 0.008766219019889832, + 0.016693923622369766, + 0.020728595554828644, + 0.005004821345210075, + 0.006065381225198507, + 0.003954729065299034, + 0.0031076206360012293, + 0.00455264188349247, + 0.004389075096696615, + 0.02476455084979534, + 0.008715878240764141, + 0.007410555146634579, + 0.007101262919604778, + 0.008361706510186195, + 0.014442315325140953, + 0.007257678080350161, + 0.009109444916248322, + 0.007248396519571543, + 0.0050615849904716015, + 0.00403420627117157, + 0.006777316331863403, + 0.00440613180398941, + 0.006902019958943129, + 0.00885891355574131, + 0.007345184218138456, + 0.006409239489585161, + 0.0021440251730382442, + 0.003159071784466505, + 0.0027690385468304157, + 0.0026943355333060026, + 0.002089964458718896, + 0.0020270568784326315, + 0.004215147811919451, + 0.0025485209189355373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.01254948042333126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04967411234974861, + 0.020716233178973198, + 0.03664766252040863, + 0.3559477925300598, + 0.02754473127424717, + 0.23504534363746643, + 0.18409816920757294, + 0.05908024311065674, + 0.019512439146637917, + 0.011733231134712696 + ], + "credit_map": [ + 0.0037380754947662354, + 0.004460293799638748, + 0.007915990427136421, + 0.010628921911120415, + 0.003127944888547063, + 0.005760821979492903, + 0.005735244136303663, + 0.008306823670864105, + 0.0018143578199669719, + 0.0013448274694383144, + 0.002519940957427025, + 0.002451683161780238, + 0.0019304790766909719, + 0.004195003304630518, + 0.0021152414847165346, + 0.004344702232629061, + 0.002967148320749402, + 0.004544632509350777, + 0.004430667031556368, + 0.0051105390302836895, + 0.0038759217131882906, + 0.006396143231540918, + 0.004714274778962135, + 0.004608337767422199, + 0.027455810457468033, + 0.018433764576911926, + 0.10198426246643066, + 0.027980895712971687, + 0.016885844990611076, + 0.0362970307469368, + 0.01574215106666088, + 0.11116807907819748, + 0.0029248602222651243, + 0.00304000498726964, + 0.0032263994216918945, + 0.003002976765856147, + 0.0051847659051418304, + 0.0030950047075748444, + 0.003750514006242156, + 0.0033202078193426132, + 0.08726106584072113, + 0.013765716925263405, + 0.09117398411035538, + 0.01054949127137661, + 0.005762972868978977, + 0.005538693629205227, + 0.005574411246925592, + 0.015419050119817257, + 0.06344370543956757, + 0.008555224165320396, + 0.010892234742641449, + 0.00800698809325695, + 0.0049179778434336185, + 0.009505701251327991, + 0.010222526267170906, + 0.06855382025241852, + 0.005697532556951046, + 0.004235975444316864, + 0.008345126174390316, + 0.007849616929888725, + 0.01283178385347128, + 0.007206670939922333, + 0.0074274116195738316, + 0.005486132111400366, + 0.0025570366997271776, + 0.0022591478191316128, + 0.001986471237614751, + 0.00250623794272542, + 0.0026825794484466314, + 0.0020139180123806, + 0.003090217709541321, + 0.0024168319068849087, + 0.001497552846558392, + 0.0020114288199692965, + 0.0016938396729528904, + 0.00137683458160609, + 0.0010380768217146397, + 0.0010205412982031703, + 0.0018728921422734857, + 0.0012220664648339152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176183104515076 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176183104515076, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176183104515076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03848927095532417, + 0.021847276017069817, + 0.02350878342986107, + 0.4411102831363678, + 0.01650906912982464, + 0.34762153029441833, + 0.05726025998592377, + 0.021771134808659554, + 0.023647276684641838, + 0.008235102519392967 + ], + "credit_map": [ + 0.0031810577493160963, + 0.004015503916889429, + 0.006332517601549625, + 0.0077277920208871365, + 0.004147286526858807, + 0.004588984418660402, + 0.003449903102591634, + 0.005046226549893618, + 0.0023771850392222404, + 0.0017125274753198028, + 0.003957964479923248, + 0.003757026046514511, + 0.0016903990181162953, + 0.0029344861395657063, + 0.0018032199004665017, + 0.0036144680343568325, + 0.0016106270486488938, + 0.0019304535817354918, + 0.002956779208034277, + 0.0033202855847775936, + 0.00266603147611022, + 0.0031967631075531244, + 0.003591131651774049, + 0.004236710257828236, + 0.033708978444337845, + 0.017436916008591652, + 0.13249027729034424, + 0.03410365805029869, + 0.017435776069760323, + 0.042912691831588745, + 0.01955249533057213, + 0.14346948266029358, + 0.0017758863978087902, + 0.002218563575297594, + 0.0020138374529778957, + 0.002639602869749069, + 0.001578207709826529, + 0.002371532376855612, + 0.0024201988708227873, + 0.0014912403421476483, + 0.1255650669336319, + 0.020374594256281853, + 0.12819212675094604, + 0.01577068120241165, + 0.015878554433584213, + 0.0098475506529212, + 0.009733643382787704, + 0.022259339690208435, + 0.019048625603318214, + 0.002786488737910986, + 0.003461033571511507, + 0.0026053334586322308, + 0.0017860663356259465, + 0.003371089929714799, + 0.003123608184978366, + 0.021078011021018028, + 0.002434731228277087, + 0.001954844919964671, + 0.003895791247487068, + 0.002311728661879897, + 0.0036440789699554443, + 0.003056779270991683, + 0.0027145626954734325, + 0.0017586175817996264, + 0.004738479387015104, + 0.002703302074223757, + 0.0032107592560350895, + 0.0018709051655605435, + 0.002307828050106764, + 0.0035696839913725853, + 0.0022346938494592905, + 0.003011624561622739, + 0.001291781198233366, + 0.0012553819688037038, + 0.0011278935708105564, + 0.0009074988774955273, + 0.0006722915568388999, + 0.0009606632520444691, + 0.0008766287937760353, + 0.0011429626028984785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.018176183104515076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029782237485051155 + ], + "opposite_logprob_tokens": [ + -3.529782295227051 + ], + "logprob_reward": -0.029782237485051155, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029782237485051155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034625548869371414, + 0.02388612926006317, + 0.02151055447757244, + 0.4486177861690521, + 0.04792354255914688, + 0.10682464390993118, + 0.06200641021132469, + 0.11347667872905731, + 0.12379095703363419, + 0.01733773574233055 + ], + "credit_map": [ + 0.002374435542151332, + 0.002752127591520548, + 0.005506892688572407, + 0.004865845199674368, + 0.0032897016499191523, + 0.003799853613600135, + 0.0038522337563335896, + 0.008184458129107952, + 0.003043570788577199, + 0.003018639050424099, + 0.0023732921108603477, + 0.0032451541628688574, + 0.002031471114605665, + 0.00340553792193532, + 0.0021030306816101074, + 0.004665435757488012, + 0.0016872950363904238, + 0.002051824238151312, + 0.002547535113990307, + 0.002927499357610941, + 0.0030780513770878315, + 0.0024713645689189434, + 0.0036110118962824345, + 0.0031359742861241102, + 0.033059388399124146, + 0.01033025048673153, + 0.14867214858531952, + 0.03428271412849426, + 0.011111586354672909, + 0.03686732053756714, + 0.018055668100714684, + 0.15623874962329865, + 0.005405560601502657, + 0.005656259134411812, + 0.005743828136473894, + 0.007372354157269001, + 0.006057083606719971, + 0.005349801387637854, + 0.00642803218215704, + 0.005910628940910101, + 0.011173631995916367, + 0.006726094055920839, + 0.03068859502673149, + 0.007949918508529663, + 0.007533450145274401, + 0.009655439294874668, + 0.006178323179483414, + 0.026919201016426086, + 0.01666550152003765, + 0.0047921244986355305, + 0.005966654978692532, + 0.004277786239981651, + 0.0037041623145341873, + 0.007038530893623829, + 0.0041510737501084805, + 0.01541057787835598, + 0.012247394770383835, + 0.010592467151582241, + 0.018151279538869858, + 0.013665168546140194, + 0.018839741125702858, + 0.018323760479688644, + 0.011063494719564915, + 0.010593375191092491, + 0.023730270564556122, + 0.011792720295488834, + 0.01779811829328537, + 0.010677063837647438, + 0.015431485138833523, + 0.014842666685581207, + 0.013785431161522865, + 0.015733202919363976, + 0.0019964869134128094, + 0.0023815506137907505, + 0.0024580496829003096, + 0.0024930972140282393, + 0.0016517831245437264, + 0.0013911162968724966, + 0.0030190087854862213, + 0.0019466449739411473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.029782237485051155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, tracking shot, close to the ground, follows two young men as they effortlessly glide across a futuristic city square at dusk on their electric unicycles. The young man on the left, with shaggy black hair and a tech-fabric black hoodie with reflective silver stripes on the sleeves, wears a broad smile. The young woman next to him, with her long, bright pink hair tied in a high ponytail, throws her head back in a sleek white bomber jacket and laughs heartily. The unicycles they ride are matte black, their futuristic LED strips emitting brilliant cyan and magenta light on either side, casting vibrant reflections on the polished dark stone floor, leaving long trails of light as the wheels spin at breakneck speed. The camera switches focus swiftly and precisely between close-ups of their vivid, joyful facial expressions and the wheels spinning so rapidly that they appear almost animate, against a backdrop of towering skyscrapers, their windows lit up, and flashing holographic billboards. The entire picture has a cold tone, mainly dark blue and black, but is illuminated by the neon lights of the unicycle and the warm lights in the distance, interspersed with slight, stylized digital screen flashes and lens flares that represent a sense of technology, creating a futuristic atmosphere that is free, cheerful and full of urban dynamism.", + "seed": 22846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_2/index_4817_seed_22846_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025624214159324765, + "vlm_logprob": -0.025624214159324765, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.025624214159324765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08071012794971466 + ], + "opposite_logprob_tokens": [ + -2.9557101726531982 + ], + "logprob_reward": -0.08071012794971466, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08071012794971466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17635215818881989, + 0.13958585262298584, + 0.04681435972452164, + 0.11890709400177002, + 0.09527308493852615, + 0.03043476678431034, + 0.07697906345129013, + 0.07026571780443192, + 0.21965353190898895, + 0.02573438361287117 + ], + "credit_map": [ + 0.0199422650039196, + 0.023757413029670715, + 0.023386264219880104, + 0.028993841260671616, + 0.020091041922569275, + 0.011921665631234646, + 0.03183956444263458, + 0.016420096158981323, + 0.012032746337354183, + 0.015421957708895206, + 0.021231522783637047, + 0.02403983660042286, + 0.019950391724705696, + 0.013946819119155407, + 0.019309520721435547, + 0.01365305483341217, + 0.004861666355282068, + 0.006508154794573784, + 0.004549617879092693, + 0.0072973622009158134, + 0.007448549382388592, + 0.005728171207010746, + 0.006350706797093153, + 0.004070131108164787, + 0.012251446954905987, + 0.015315425582230091, + 0.02638429030776024, + 0.009551632218062878, + 0.015760673210024834, + 0.013825718313455582, + 0.012850780040025711, + 0.012967121787369251, + 0.012996161356568336, + 0.012882805429399014, + 0.022140618413686752, + 0.005831887945532799, + 0.008167670108377934, + 0.010233686305582523, + 0.011783716268837452, + 0.011236538179218769, + 0.0030288533307611942, + 0.0038762737531214952, + 0.0032661068253219128, + 0.002912999829277396, + 0.0036301841028034687, + 0.004194381646811962, + 0.005450142547488213, + 0.004075823351740837, + 0.013260546140372753, + 0.010431675240397453, + 0.010677346959710121, + 0.009436774998903275, + 0.007963103242218494, + 0.008229510858654976, + 0.008339786902070045, + 0.008640320971608162, + 0.009634154848754406, + 0.009779764339327812, + 0.009330793283879757, + 0.00968458317220211, + 0.006926214788109064, + 0.007463324349373579, + 0.009043009020388126, + 0.008403876796364784, + 0.02313380315899849, + 0.027456561103463173, + 0.020013781264424324, + 0.019273322075605392, + 0.027558406814932823, + 0.04475785419344902, + 0.023410474881529808, + 0.03404933214187622, + 0.0030172381084412336, + 0.003217470133677125, + 0.003215524833649397, + 0.003251594491302967, + 0.0035085149575024843, + 0.003076486987993121, + 0.0029960740357637405, + 0.0034514819271862507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.08071012794971466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750871893018484 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.006750871893018484, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750871893018484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14546877145767212, + 0.0967932641506195, + 0.055096860975027084, + 0.048704542219638824, + 0.0647372156381607, + 0.05277349799871445, + 0.03720446303486824, + 0.11923582851886749, + 0.11092076450586319, + 0.26906469464302063 + ], + "credit_map": [ + 0.0114135155454278, + 0.018921520560979843, + 0.012945406138896942, + 0.028432460501790047, + 0.01673094369471073, + 0.01976913958787918, + 0.023314842954277992, + 0.013940954580903053, + 0.00910191796720028, + 0.009946564212441444, + 0.014968452975153923, + 0.014690429903566837, + 0.014472899958491325, + 0.012916652485728264, + 0.012176844291388988, + 0.00851951539516449, + 0.004586209077388048, + 0.007969447411596775, + 0.0075528547167778015, + 0.008095496334135532, + 0.006179886870086193, + 0.007045109290629625, + 0.006911353208124638, + 0.006756506394594908, + 0.004313341807574034, + 0.00651518115773797, + 0.006464918609708548, + 0.004148019943386316, + 0.009498395957052708, + 0.005502423737198114, + 0.00614584144204855, + 0.0061164237558841705, + 0.006678787525743246, + 0.01041396427899599, + 0.0077798557467758656, + 0.007348331622779369, + 0.00668380968272686, + 0.008708481676876545, + 0.0071763102896511555, + 0.009947678074240685, + 0.005863357335329056, + 0.008953670971095562, + 0.007387321908026934, + 0.0067266677506268024, + 0.006502893753349781, + 0.005641172174364328, + 0.00757110258564353, + 0.00412731571123004, + 0.005751537159085274, + 0.003422683337703347, + 0.00439803721383214, + 0.003814226482063532, + 0.004070598166435957, + 0.004929179325699806, + 0.0044021280482411385, + 0.006416076794266701, + 0.020144827663898468, + 0.021465325728058815, + 0.016619591042399406, + 0.01828126423060894, + 0.011558791622519493, + 0.00875544361770153, + 0.011762329377233982, + 0.010648275725543499, + 0.010844642296433449, + 0.0203881673514843, + 0.016969773918390274, + 0.018092002719640732, + 0.011843828484416008, + 0.01248631440103054, + 0.011538294143974781, + 0.008757744915783405, + 0.030546125024557114, + 0.04361095279455185, + 0.024803640320897102, + 0.03264743834733963, + 0.05529426783323288, + 0.022954080253839493, + 0.04033639654517174, + 0.01887180656194687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.006750871893018484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031932813581079245 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031932813581079245, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031932813581079245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20022085309028625, + 0.1993912160396576, + 0.0929815024137497, + 0.055032841861248016, + 0.0602944977581501, + 0.04147657752037048, + 0.09789606928825378, + 0.1233922690153122, + 0.07069984823465347, + 0.05861435458064079 + ], + "credit_map": [ + 0.01510432455688715, + 0.018159478902816772, + 0.014779706485569477, + 0.03707333281636238, + 0.0197067279368639, + 0.03505965694785118, + 0.032399170100688934, + 0.02793845161795616, + 0.01890162006020546, + 0.018231980502605438, + 0.03204551339149475, + 0.02346864715218544, + 0.031696029007434845, + 0.031070388853549957, + 0.02446114271879196, + 0.019515885040163994, + 0.006581056397408247, + 0.011390323750674725, + 0.011593800038099289, + 0.007638043724000454, + 0.012174591422080994, + 0.018345637246966362, + 0.010133672505617142, + 0.015124373137950897, + 0.004925192799419165, + 0.008213918656110764, + 0.008795016445219517, + 0.0040504904463887215, + 0.009373851120471954, + 0.007371081970632076, + 0.005940335337072611, + 0.006362956948578358, + 0.0059577422216534615, + 0.00816426333039999, + 0.011528984643518925, + 0.004744760226458311, + 0.005833944771438837, + 0.007350131869316101, + 0.009301170706748962, + 0.007413502316921949, + 0.003764979774132371, + 0.0062664709985256195, + 0.003331777174025774, + 0.004865729715675116, + 0.006713278125971556, + 0.006842560600489378, + 0.00607722532004118, + 0.00361455837264657, + 0.013304663822054863, + 0.008115722797811031, + 0.011243830434978008, + 0.013140090741217136, + 0.009456750005483627, + 0.01258259080350399, + 0.013295387849211693, + 0.016757037490606308, + 0.018692675977945328, + 0.01592952571809292, + 0.014744622632861137, + 0.01614847406744957, + 0.015158622525632381, + 0.013171295635402203, + 0.01288757286965847, + 0.01665947027504444, + 0.009007656946778297, + 0.008128288201987743, + 0.008100851438939571, + 0.00856642797589302, + 0.009416190907359123, + 0.013336622156202793, + 0.005909602157771587, + 0.008234205655753613, + 0.006437196861952543, + 0.007365695666521788, + 0.005068470258265734, + 0.008930088952183723, + 0.012582332827150822, + 0.005607915110886097, + 0.007996215485036373, + 0.0046264431439340115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0031932813581079245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190246351063251 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190246351063251, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190246351063251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13853032886981964, + 0.1341497153043747, + 0.07708192616701126, + 0.07850752025842667, + 0.1526193916797638, + 0.1331794559955597, + 0.05456315353512764, + 0.06763031333684921, + 0.09094405174255371, + 0.07279408723115921 + ], + "credit_map": [ + 0.014159015379846096, + 0.017336269840598106, + 0.019038110971450806, + 0.026297645643353462, + 0.015650924295186996, + 0.012902762740850449, + 0.018119165673851967, + 0.015026452951133251, + 0.018666837364435196, + 0.013719933107495308, + 0.028966156765818596, + 0.019553622230887413, + 0.013083770871162415, + 0.012866940349340439, + 0.01660303957760334, + 0.010689418762922287, + 0.006456131115555763, + 0.011121156625449657, + 0.007612911053001881, + 0.010709875263273716, + 0.010425278916954994, + 0.011568539775907993, + 0.009175054728984833, + 0.010012982413172722, + 0.009183619171380997, + 0.009464338421821594, + 0.014764519408345222, + 0.006712811999022961, + 0.009323135018348694, + 0.007025876082479954, + 0.011881181038916111, + 0.010152047500014305, + 0.016779672354459763, + 0.023067962378263474, + 0.032589659094810486, + 0.015241408720612526, + 0.013819555751979351, + 0.014645724557340145, + 0.021831678226590157, + 0.014643731527030468, + 0.010308454744517803, + 0.017314938828349113, + 0.0120161771774292, + 0.017015134915709496, + 0.022852590307593346, + 0.019434848800301552, + 0.022726207971572876, + 0.011511108838021755, + 0.009359783492982388, + 0.006179900839924812, + 0.00689254654571414, + 0.010487384162843227, + 0.00488731823861599, + 0.004400491714477539, + 0.00725756399333477, + 0.005098169203847647, + 0.010536925867199898, + 0.009193241596221924, + 0.009390060789883137, + 0.010623031295835972, + 0.0062117199413478374, + 0.006075207144021988, + 0.00840127281844616, + 0.0071988580748438835, + 0.008817481808364391, + 0.018354663625359535, + 0.009445100091397762, + 0.012438696809113026, + 0.011476692743599415, + 0.012910110875964165, + 0.008657896891236305, + 0.008843411691486835, + 0.011058344505727291, + 0.010891745798289776, + 0.008676293306052685, + 0.007466016337275505, + 0.011543902568519115, + 0.006495837587863207, + 0.010029119439423084, + 0.006632825825363398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.014190246351063251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young people riding electric unicycles across a futuristic city square at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a black hoodie with reflective silver stripes and has shaggy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman next to him wearing a white bomber jacket and has long bright pink hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the matte black unicycles feature LED strips emitting cyan and magenta light that cast reflections on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus swiftly between close-ups of their faces and the spinning wheels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023276543244719505 + ], + "opposite_logprob_tokens": [ + -3.7732765674591064 + ], + "logprob_reward": -0.023276543244719505, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023276543244719505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10334397852420807, + 0.12238747626543045, + 0.12612688541412354, + 0.11628684401512146, + 0.14976908266544342, + 0.08256117254495621, + 0.05802575871348381, + 0.10184743255376816, + 0.0804569274187088, + 0.05919446796178818 + ], + "credit_map": [ + 0.014716921374201775, + 0.010238075628876686, + 0.014562332071363926, + 0.017819300293922424, + 0.010047809220850468, + 0.009537073783576488, + 0.018596695736050606, + 0.007825776934623718, + 0.014511393383145332, + 0.010799730196595192, + 0.015093326568603516, + 0.025643818080425262, + 0.015591043047606945, + 0.01295617874711752, + 0.01840122975409031, + 0.009390764869749546, + 0.010923069901764393, + 0.016440540552139282, + 0.012924210168421268, + 0.014048060402274132, + 0.021175259724259377, + 0.018876124173402786, + 0.013966161757707596, + 0.017773469910025597, + 0.012110709212720394, + 0.02002030424773693, + 0.016980791464447975, + 0.008058101870119572, + 0.021722743287682533, + 0.010981129482388496, + 0.015252704732120037, + 0.011160369962453842, + 0.02202288806438446, + 0.018070600926876068, + 0.039901748299598694, + 0.010090035386383533, + 0.013864366337656975, + 0.012702172622084618, + 0.01902208849787712, + 0.014095188118517399, + 0.008860423229634762, + 0.014244530349969864, + 0.008938956074416637, + 0.007762083783745766, + 0.016001731157302856, + 0.011274033226072788, + 0.010010877624154091, + 0.005468552000820637, + 0.008767809718847275, + 0.0048126112669706345, + 0.008758846670389175, + 0.008291474543511868, + 0.006159410811960697, + 0.007405993528664112, + 0.0073786466382443905, + 0.006450972985476255, + 0.012572051957249641, + 0.012411832809448242, + 0.014053992927074432, + 0.015904773026704788, + 0.011483936570584774, + 0.010555812157690525, + 0.01458412129431963, + 0.010280920192599297, + 0.01056590210646391, + 0.013197196647524834, + 0.0106633510440588, + 0.009865716099739075, + 0.010696892626583576, + 0.012041795998811722, + 0.006787512451410294, + 0.00663856603205204, + 0.00870770774781704, + 0.010126371867954731, + 0.007055342197418213, + 0.006367396097630262, + 0.009483976289629936, + 0.004711943678557873, + 0.008154411800205708, + 0.004587325267493725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.023276543244719505 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_tensor([449, 449])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_tensor([449, 449])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6e3093e04a5add6af1a2c4f120215540c60e87c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_tensor([449, 449])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion shot, shot from a low angle and tilted slightly upward, captures the decisive moment of a young man in a dark gray quick-drying T-shirt and khaki cargo shorts completing a breathtaking leap. As his dusty hiking boots landed precisely and powerfully on the massive, jagged summit of gray granite, his knees bent and he squatted, deftly absorbing the impact, sending the tiny gravel beneath his feet bouncing slightly. Then, steadily and deliberately, he straightened his body, arms extended to his sides like bird wings, fingers slightly spread apart, carefully adjusting his balance to maintain perfect balance. His short, sweat-soaked brown hair blew tousled in the breeze, and his focused, resolute face, muscles taut and defined, betrayed a fierce concentration. The golden afterglow of the setting sun, streaming from behind and sideways, cast a dazzling golden rim across his figure, while his front was cast in soft shadow, enhancing the three-dimensionality of the image. The background is the silhouette of rolling mountains, dyed a deep purple by the setting sun. The whole scene is full of the sense of accomplishment of conquering nature and the magnificent tranquility of sunset.", + "seed": 5213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_449_seed_5213_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12831572461873292, + "vlm_logprob": -0.12831572461873292, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.12831572461873292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5173847675323486 + ], + "opposite_logprob_tokens": [ + -1.0173847675323486 + ], + "logprob_reward": -0.5173847675323486, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5173847675323486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18756552040576935, + 0.1688377559185028, + 0.19341571629047394, + 0.04985873028635979, + 0.06774909049272537, + 0.08027268201112747, + 0.047860898077487946, + 0.048168327659368515, + 0.0796954482793808, + 0.07657591253519058 + ], + "credit_map": [ + 0.025235461071133614, + 0.030330361798405647, + 0.016128724440932274, + 0.02289295382797718, + 0.024934250861406326, + 0.024737874045968056, + 0.02461087517440319, + 0.018695032224059105, + 0.022482335567474365, + 0.021112671121954918, + 0.01992620714008808, + 0.017411300912499428, + 0.026924822479486465, + 0.030984770506620407, + 0.013681743294000626, + 0.016313893720507622, + 0.020999472588300705, + 0.022460494190454483, + 0.029098717495799065, + 0.023112310096621513, + 0.020474832504987717, + 0.021832773461937904, + 0.034673795104026794, + 0.02076331526041031, + 0.006407454144209623, + 0.006405585445463657, + 0.005632187705487013, + 0.004768720827996731, + 0.006989207584410906, + 0.01068735122680664, + 0.00447473768144846, + 0.004493489861488342, + 0.008535257540643215, + 0.010633851401507854, + 0.005995587445795536, + 0.006199462804943323, + 0.01199402753263712, + 0.008712134324014187, + 0.008542224764823914, + 0.0071365367621183395, + 0.010570812039077282, + 0.007324906066060066, + 0.010409216396510601, + 0.013738387264311314, + 0.012919764034450054, + 0.009081937372684479, + 0.008563145063817501, + 0.007664512377232313, + 0.010473125614225864, + 0.00445619085803628, + 0.0060792043805122375, + 0.003657920053228736, + 0.005931065417826176, + 0.008449791930615902, + 0.0034152271691709757, + 0.005398372653871775, + 0.01218027900904417, + 0.007161893416196108, + 0.006757319439202547, + 0.0034410844091326, + 0.003470171010121703, + 0.006490429397672415, + 0.0039818743243813515, + 0.004685273393988609, + 0.010636553168296814, + 0.012150387279689312, + 0.01836385205388069, + 0.008139188401401043, + 0.007956079207360744, + 0.007632568012923002, + 0.006405304651707411, + 0.008411518298089504, + 0.012884018942713737, + 0.006229862105101347, + 0.006998025346547365, + 0.00813518837094307, + 0.013096733018755913, + 0.013066916726529598, + 0.0055126016959548, + 0.010652566328644753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.5173847675323486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983557991683483 + ], + "opposite_logprob_tokens": [ + -3.5298354625701904 + ], + "logprob_reward": -0.02983557991683483, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983557991683483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1706310659646988, + 0.2783746123313904, + 0.09477771073579788, + 0.04821646213531494, + 0.0467459037899971, + 0.05063987523317337, + 0.12604977190494537, + 0.0441109724342823, + 0.10540652275085449, + 0.035047128796577454 + ], + "credit_map": [ + 0.026551121845841408, + 0.02335813269019127, + 0.020741216838359833, + 0.019494125619530678, + 0.030481845140457153, + 0.012760180048644543, + 0.021810825914144516, + 0.015433629043400288, + 0.043022520840168, + 0.0356350876390934, + 0.042078956961631775, + 0.019827978685498238, + 0.04519251361489296, + 0.041084062308073044, + 0.02390197105705738, + 0.027631500735878944, + 0.008813826367259026, + 0.012070161290466785, + 0.01147747877985239, + 0.015393786132335663, + 0.012443154118955135, + 0.01121965516358614, + 0.01289451029151678, + 0.010465131141245365, + 0.00804887991398573, + 0.007300708908587694, + 0.005340175237506628, + 0.004211222752928734, + 0.007182063534855843, + 0.007107674144208431, + 0.004526953212916851, + 0.004498786758631468, + 0.006244323216378689, + 0.007165181916207075, + 0.0046561239287257195, + 0.004396883305162191, + 0.008113578893244267, + 0.006069157738238573, + 0.0058877370320260525, + 0.004212924279272556, + 0.0072328499518334866, + 0.004799802787601948, + 0.006325533613562584, + 0.0068394881673157215, + 0.007582308724522591, + 0.006566360592842102, + 0.006207192782312632, + 0.005086340010166168, + 0.028507964685559273, + 0.016157841309905052, + 0.01929335668683052, + 0.008282718248665333, + 0.010057546198368073, + 0.022736776620149612, + 0.009503272362053394, + 0.011510306037962437, + 0.006123939994722605, + 0.008544227108359337, + 0.004442594014108181, + 0.002880141604691744, + 0.0035723689943552017, + 0.010804225690662861, + 0.0040404945611953735, + 0.0037029804661870003, + 0.013582412153482437, + 0.019475243985652924, + 0.029910653829574585, + 0.008873901329934597, + 0.004845427814871073, + 0.013040429912507534, + 0.004366847220808268, + 0.011311609297990799, + 0.005199076607823372, + 0.0029864057432860136, + 0.004865296185016632, + 0.003941753879189491, + 0.00427315104752779, + 0.005438595544546843, + 0.0038697656709700823, + 0.00447308411821723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.02983557991683483 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809638321399689 + ], + "opposite_logprob_tokens": [ + -3.2880964279174805 + ], + "logprob_reward": -0.03809638321399689, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809638321399689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11499588191509247, + 0.09400160610675812, + 0.05643310025334358, + 0.15765736997127533, + 0.05877894163131714, + 0.03933308273553848, + 0.054671067744493484, + 0.1351507008075714, + 0.07923122495412827, + 0.2097470760345459 + ], + "credit_map": [ + 0.01584572158753872, + 0.01849316991865635, + 0.01159632857888937, + 0.012938405387103558, + 0.015928147360682487, + 0.01561222318559885, + 0.01391540002077818, + 0.010666475631296635, + 0.010647134855389595, + 0.010758256539702415, + 0.011165211908519268, + 0.011935632675886154, + 0.013328938744962215, + 0.01935228519141674, + 0.00824750866740942, + 0.008566626347601414, + 0.008622454479336739, + 0.005663673393428326, + 0.0056640878319740295, + 0.01104777492582798, + 0.005415431223809719, + 0.010950944386422634, + 0.004912379197776318, + 0.004156345501542091, + 0.01952190138399601, + 0.023455776274204254, + 0.010938093066215515, + 0.012941302731633186, + 0.03702504560351372, + 0.025175761431455612, + 0.016622044146060944, + 0.011977414600551128, + 0.0071342443116009235, + 0.006538359913975, + 0.0058055236004292965, + 0.005621528252959251, + 0.009937469847500324, + 0.010257914662361145, + 0.009134817868471146, + 0.004349077586084604, + 0.005700955633074045, + 0.003735159756615758, + 0.0051543754525482655, + 0.005573675036430359, + 0.005086402874439955, + 0.006572392303496599, + 0.0031197830103337765, + 0.004390335641801357, + 0.0069862729869782925, + 0.008569454774260521, + 0.0073822722770273685, + 0.0049443054012954235, + 0.005208011716604233, + 0.009146751835942268, + 0.004409682471305132, + 0.008024308830499649, + 0.024134058505296707, + 0.019478769972920418, + 0.015465443953871727, + 0.008737118914723396, + 0.01434533391147852, + 0.032461490482091904, + 0.009176905266940594, + 0.011351564899086952, + 0.0130178052932024, + 0.016092777252197266, + 0.015482880175113678, + 0.007133426144719124, + 0.005610348656773567, + 0.008553839288651943, + 0.005432399921119213, + 0.007907734252512455, + 0.03367415815591812, + 0.0165049247443676, + 0.029785171151161194, + 0.021388331428170204, + 0.023837611079216003, + 0.03461482748389244, + 0.019872628152370453, + 0.030069399625062943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.03809638321399689 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864228889346123 + ], + "opposite_logprob_tokens": [ + -3.04864239692688 + ], + "logprob_reward": -0.04864228889346123, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864228889346123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2735787332057953, + 0.33046022057533264, + 0.0943111777305603, + 0.05120081454515457, + 0.0820799395442009, + 0.05618658661842346, + 0.025117985904216766, + 0.027334855869412422, + 0.029840022325515747, + 0.029889674857258797 + ], + "credit_map": [ + 0.03282282128930092, + 0.024977413937449455, + 0.02186966873705387, + 0.033317964524030685, + 0.037706632167100906, + 0.05154133960604668, + 0.047303877770900726, + 0.024039005860686302, + 0.03346557915210724, + 0.027305953204631805, + 0.03329472988843918, + 0.03422592207789421, + 0.04068370908498764, + 0.0792836919426918, + 0.03742410987615585, + 0.04477651044726372, + 0.013125264085829258, + 0.0076744272373616695, + 0.008892770856618881, + 0.02871246077120304, + 0.00848880223929882, + 0.010984357446432114, + 0.008500585332512856, + 0.007932511158287525, + 0.009958948008716106, + 0.0044580139219760895, + 0.0076182689517736435, + 0.0048830839805305, + 0.006341400556266308, + 0.008090582676231861, + 0.004237503744661808, + 0.005613009445369244, + 0.010981359519064426, + 0.010791568085551262, + 0.008395053446292877, + 0.00923203770071268, + 0.009623252786695957, + 0.017887583002448082, + 0.006715869065374136, + 0.008453216403722763, + 0.009389886632561684, + 0.01005687564611435, + 0.005940691102296114, + 0.005392471328377724, + 0.008176806382834911, + 0.007003686390817165, + 0.004447642248123884, + 0.0057785287499427795, + 0.003950592130422592, + 0.0031282417476177216, + 0.002664344385266304, + 0.0028995254542678595, + 0.0036306725814938545, + 0.0035972241312265396, + 0.0026866113767027855, + 0.0025607759598642588, + 0.004237906076014042, + 0.0038890535943210125, + 0.004101860336959362, + 0.00199095718562603, + 0.002595286350697279, + 0.004979322664439678, + 0.002315916819497943, + 0.0032245537731796503, + 0.004888137802481651, + 0.004008040763437748, + 0.0060654254630208015, + 0.0023837334010750055, + 0.0025946651585400105, + 0.004333821590989828, + 0.0016741916770115495, + 0.003892007516697049, + 0.00424991175532341, + 0.002153594046831131, + 0.004632941912859678, + 0.0029849340207874775, + 0.003818850265815854, + 0.0037794490344822407, + 0.00174062664154917, + 0.006529365200549364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04864228889346123 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076196035370230675 + ], + "opposite_logprob_tokens": [ + -4.882619380950928 + ], + "logprob_reward": -0.0076196035370230675, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076196035370230675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14404797554016113, + 0.07993271946907043, + 0.05447541922330856, + 0.07007849216461182, + 0.03120487369596958, + 0.08569309860467911, + 0.09703311324119568, + 0.11072540283203125, + 0.15787722170352936, + 0.16893166303634644 + ], + "credit_map": [ + 0.02666683867573738, + 0.02055378630757332, + 0.010684877634048462, + 0.011739429086446762, + 0.023070840165019035, + 0.023882780224084854, + 0.01427918579429388, + 0.013170233927667141, + 0.012352610938251019, + 0.014652308076620102, + 0.012523120269179344, + 0.006007207557559013, + 0.007836895994842052, + 0.010974464938044548, + 0.006657785270363092, + 0.008928325027227402, + 0.0071027749218046665, + 0.004186237696558237, + 0.0057226670905947685, + 0.006857977248728275, + 0.0049438150599598885, + 0.014206789433956146, + 0.0058519565500319, + 0.005603204946964979, + 0.010837338864803314, + 0.00730076152831316, + 0.011294250376522541, + 0.0070087723433971405, + 0.007007113192230463, + 0.011746276170015335, + 0.008760370314121246, + 0.006123613100498915, + 0.003849827917292714, + 0.005443331319838762, + 0.0033528634812682867, + 0.0035317661240696907, + 0.003910550847649574, + 0.004856486804783344, + 0.0029107395093888044, + 0.0033493083901703358, + 0.01024721097201109, + 0.008112219162285328, + 0.01274104043841362, + 0.019433045759797096, + 0.008675923570990562, + 0.013667336665093899, + 0.005198880564421415, + 0.007617440540343523, + 0.01574232615530491, + 0.013865180313587189, + 0.01008868683129549, + 0.005344083532691002, + 0.00726244505494833, + 0.026212448254227638, + 0.005866555031388998, + 0.012651391327381134, + 0.016730962321162224, + 0.01912679709494114, + 0.01819271221756935, + 0.008016400970518589, + 0.008245856501162052, + 0.021305831149220467, + 0.007634001784026623, + 0.011472844518721104, + 0.021949997171759605, + 0.019194742664694786, + 0.05061894282698631, + 0.01258769165724516, + 0.008938395418226719, + 0.017639977857470512, + 0.007119165267795324, + 0.0198283102363348, + 0.021905308589339256, + 0.013935228809714317, + 0.029470354318618774, + 0.014685871079564095, + 0.02630736492574215, + 0.024897772818803787, + 0.00872570089995861, + 0.02900405414402485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0076196035370230675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low angle and tilted slightly upward, captures the decisive moment of a young man in a dark gray quick-drying T-shirt and khaki cargo shorts completing a breathtaking leap. As his dusty hiking boots landed precisely and powerfully on the massive, jagged summit of gray granite, his knees bent and he squatted, deftly absorbing the impact, sending the tiny gravel beneath his feet bouncing slightly. Then, steadily and deliberately, he straightened his body, arms extended to his sides like bird wings, fingers slightly spread apart, carefully adjusting his balance to maintain perfect balance. His short, sweat-soaked brown hair blew tousled in the breeze, and his focused, resolute face, muscles taut and defined, betrayed a fierce concentration. The golden afterglow of the setting sun, streaming from behind and sideways, cast a dazzling golden rim across his figure, while his front was cast in soft shadow, enhancing the three-dimensionality of the image. The background is the silhouette of rolling mountains, dyed a deep purple by the setting sun. The whole scene is full of the sense of accomplishment of conquering nature and the magnificent tranquility of sunset.", + "seed": 5214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_449_seed_5214_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07807598561048508, + "vlm_logprob": -0.07807598561048508, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.07807598561048508, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2001344859600067 + ], + "opposite_logprob_tokens": [ + -1.950134515762329 + ], + "logprob_reward": -0.2001344859600067, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.2001344859600067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14553897082805634, + 0.08735237270593643, + 0.14303471148014069, + 0.1610371321439743, + 0.03750915080308914, + 0.04950432851910591, + 0.09183873236179352, + 0.11154370754957199, + 0.034456901252269745, + 0.13818396627902985 + ], + "credit_map": [ + 0.015443171374499798, + 0.021273372694849968, + 0.011386311613023281, + 0.011244176886975765, + 0.03220590204000473, + 0.0248127281665802, + 0.017300067469477654, + 0.011873232200741768, + 0.012848683632910252, + 0.01281125470995903, + 0.011361640878021717, + 0.013594531454145908, + 0.009139674715697765, + 0.012742120772600174, + 0.007676648907363415, + 0.007177817635238171, + 0.015964465215802193, + 0.017454611137509346, + 0.019893428310751915, + 0.013614494353532791, + 0.020686283707618713, + 0.02188965305685997, + 0.018119052052497864, + 0.01541272085160017, + 0.017060041427612305, + 0.011090646497905254, + 0.026406805962324142, + 0.01069513987749815, + 0.017317865043878555, + 0.046907033771276474, + 0.014057233929634094, + 0.017502380535006523, + 0.0037354666274040937, + 0.004593651741743088, + 0.0056657991372048855, + 0.003892673412337899, + 0.00573824904859066, + 0.0035019293427467346, + 0.005427972413599491, + 0.004953410942107439, + 0.006927235051989555, + 0.004719277378171682, + 0.007647699676454067, + 0.004137777257710695, + 0.004713584203273058, + 0.00938386470079422, + 0.004738573916256428, + 0.00723631726577878, + 0.024631157517433167, + 0.008588013239204884, + 0.01582638919353485, + 0.009338178671896458, + 0.005984270945191383, + 0.010732362978160381, + 0.008649484254419804, + 0.00808887742459774, + 0.020407436415553093, + 0.010635909624397755, + 0.014437571167945862, + 0.015002106316387653, + 0.014648042619228363, + 0.010687482543289661, + 0.01395732443779707, + 0.011767834424972534, + 0.0046745482832193375, + 0.004645369481295347, + 0.005655364599078894, + 0.003833872266113758, + 0.003515595104545355, + 0.004531779792159796, + 0.002802340080961585, + 0.00479803467169404, + 0.03384201228618622, + 0.0077860006131231785, + 0.014625728130340576, + 0.006976442877203226, + 0.009664805606007576, + 0.034304894506931305, + 0.00851528625935316, + 0.022468795999884605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.2001344859600067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026386484503746033 + ], + "opposite_logprob_tokens": [ + -3.6513864994049072 + ], + "logprob_reward": -0.026386484503746033, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026386484503746033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38669154047966003, + 0.09119653701782227, + 0.08662742376327515, + 0.09918222576379776, + 0.05674062669277191, + 0.10251785069704056, + 0.029508676379919052, + 0.04250436648726463, + 0.03972078487277031, + 0.06531000137329102 + ], + "credit_map": [ + 0.040419451892375946, + 0.06052495166659355, + 0.026921899989247322, + 0.030141346156597137, + 0.08065953105688095, + 0.06508693099021912, + 0.04926921799778938, + 0.033668190240859985, + 0.010411983355879784, + 0.011842155829071999, + 0.010800369083881378, + 0.01189188752323389, + 0.011439974419772625, + 0.013401951640844345, + 0.011618330143392086, + 0.00978988129645586, + 0.01003967598080635, + 0.012264973483979702, + 0.012160840444266796, + 0.007739880122244358, + 0.01322045736014843, + 0.014883515425026417, + 0.00881663616746664, + 0.007501444313675165, + 0.009589271619915962, + 0.016282610595226288, + 0.015185658819973469, + 0.013391339220106602, + 0.011615701951086521, + 0.00951165147125721, + 0.014090836979448795, + 0.009515156969428062, + 0.00574309378862381, + 0.005637357011437416, + 0.00895777065306902, + 0.006714068818837404, + 0.0060285357758402824, + 0.007429267279803753, + 0.008685116656124592, + 0.007545414846390486, + 0.027186816558241844, + 0.006245205644518137, + 0.021617354825139046, + 0.008516745641827583, + 0.0065780458971858025, + 0.015581264160573483, + 0.008114764466881752, + 0.008677654899656773, + 0.004481468815356493, + 0.004175465553998947, + 0.004037666134536266, + 0.004046525806188583, + 0.003015755908563733, + 0.003897740738466382, + 0.0031495976727455854, + 0.0027044538874179125, + 0.008518538437783718, + 0.005729110911488533, + 0.006123127415776253, + 0.005012480542063713, + 0.003504587337374687, + 0.004883936140686274, + 0.003842931007966399, + 0.004889653995633125, + 0.008404879830777645, + 0.004253404680639505, + 0.007006032858043909, + 0.003451636526733637, + 0.0034434301778674126, + 0.0050029749982059, + 0.004701132420450449, + 0.0034572912845760584, + 0.012519832700490952, + 0.0043396963737905025, + 0.008114862255752087, + 0.00499389972537756, + 0.00528451893478632, + 0.010875262320041656, + 0.0063399202190339565, + 0.012842009775340557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.026386484503746033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054949332028627396 + ], + "opposite_logprob_tokens": [ + -2.9299492835998535 + ], + "logprob_reward": -0.054949332028627396, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054949332028627396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09999728947877884, + 0.06865062564611435, + 0.14212524890899658, + 0.07616593688726425, + 0.22495274245738983, + 0.08997725695371628, + 0.07837484776973724, + 0.05705269053578377, + 0.03085697628557682, + 0.1318463832139969 + ], + "credit_map": [ + 0.012022112496197224, + 0.01508728601038456, + 0.009639899246394634, + 0.009020671248435974, + 0.012797783128917217, + 0.017004510387778282, + 0.014799217693507671, + 0.009625805541872978, + 0.008724219165742397, + 0.005930375307798386, + 0.009933462366461754, + 0.007466218899935484, + 0.006795999128371477, + 0.015878895297646523, + 0.007022834848612547, + 0.0068986183032393456, + 0.017557479441165924, + 0.020458711311221123, + 0.018422039225697517, + 0.01572498306632042, + 0.021369529888033867, + 0.015982137992978096, + 0.01544105913490057, + 0.017169304192066193, + 0.0076083652675151825, + 0.00813316646963358, + 0.011272541247308254, + 0.0075506106950342655, + 0.009566687978804111, + 0.012399821542203426, + 0.009193764999508858, + 0.010440981946885586, + 0.02770424634218216, + 0.030615510419011116, + 0.028990663588047028, + 0.021101368591189384, + 0.0356270857155323, + 0.02165752835571766, + 0.029766904190182686, + 0.029489435255527496, + 0.01297595351934433, + 0.009063716046512127, + 0.01513854693621397, + 0.013772109523415565, + 0.007227740250527859, + 0.007356889545917511, + 0.013883114792406559, + 0.010559184476733208, + 0.009400973096489906, + 0.008588774129748344, + 0.010357418097555637, + 0.015749601647257805, + 0.011081756092607975, + 0.005796547047793865, + 0.01108875498175621, + 0.006311021279543638, + 0.009424952790141106, + 0.010794461704790592, + 0.007082416210323572, + 0.0059384312480688095, + 0.005819345358759165, + 0.00824570469558239, + 0.004787273705005646, + 0.004960106685757637, + 0.0038365290965884924, + 0.002311058109626174, + 0.004815172869712114, + 0.0031252566259354353, + 0.0031715405639261007, + 0.004967875778675079, + 0.003571250941604376, + 0.00505829369649291, + 0.01404496468603611, + 0.010748968459665775, + 0.016118379309773445, + 0.01233131717890501, + 0.013376736082136631, + 0.027444541454315186, + 0.012951977550983429, + 0.024829505011439323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.054949332028627396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028044879436493 + ], + "opposite_logprob_tokens": [ + -2.350280523300171 + ], + "logprob_reward": -0.10028044879436493, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028044879436493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09229206293821335, + 0.08732707798480988, + 0.19517894089221954, + 0.07773596793413162, + 0.16841398179531097, + 0.179823100566864, + 0.03528731316328049, + 0.03140588849782944, + 0.07069700211286545, + 0.061838679015636444 + ], + "credit_map": [ + 0.009278260171413422, + 0.01599886082112789, + 0.01049769762903452, + 0.00794010516256094, + 0.0190671868622303, + 0.010454936884343624, + 0.0093530910089612, + 0.00970191415399313, + 0.010740582831203938, + 0.00951087474822998, + 0.011984182521700859, + 0.010233860462903976, + 0.01130885910242796, + 0.01766055077314377, + 0.007158468011766672, + 0.008729701861739159, + 0.023237036541104317, + 0.02429424226284027, + 0.026155611500144005, + 0.021215397864580154, + 0.031247595325112343, + 0.02844528667628765, + 0.021305007860064507, + 0.01927875727415085, + 0.009420447051525116, + 0.006319078151136637, + 0.016220422461628914, + 0.007935280911624432, + 0.008846989832818508, + 0.013494250364601612, + 0.0073592206463217735, + 0.008140280842781067, + 0.016640983521938324, + 0.01829506643116474, + 0.026160454377532005, + 0.01168047171086073, + 0.031206738203763962, + 0.0262202937155962, + 0.01662025973200798, + 0.021589722484350204, + 0.04447558522224426, + 0.010586508549749851, + 0.039101146161556244, + 0.016199709847569466, + 0.010682609863579273, + 0.024458035826683044, + 0.01580432988703251, + 0.018515173345804214, + 0.0067090135999023914, + 0.0034800933208316565, + 0.0045105875469744205, + 0.0045353746972978115, + 0.003356068627908826, + 0.005602945573627949, + 0.003349416656419635, + 0.0037438126746565104, + 0.004845903720706701, + 0.0031155755277723074, + 0.005105299409478903, + 0.0021891468204557896, + 0.002338622696697712, + 0.004587624687701464, + 0.0025067850947380066, + 0.00671693030744791, + 0.014070184901356697, + 0.009966274723410606, + 0.01081934105604887, + 0.005873335059732199, + 0.005410204641520977, + 0.009965472854673862, + 0.00539789954200387, + 0.00919429212808609, + 0.012047924101352692, + 0.0035471725277602673, + 0.008863195776939392, + 0.005527998320758343, + 0.003780792700126767, + 0.017486531287431717, + 0.0057260626927018166, + 0.004859003704041243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.10028044879436493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008629176765680313 + ], + "opposite_logprob_tokens": [ + -4.758629322052002 + ], + "logprob_reward": -0.008629176765680313, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008629176765680313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18158318102359772, + 0.08306914567947388, + 0.06401068717241287, + 0.036640748381614685, + 0.07461367547512054, + 0.15175677835941315, + 0.06910614669322968, + 0.11011583358049393, + 0.053746722638607025, + 0.17535711824893951 + ], + "credit_map": [ + 0.019762640818953514, + 0.0356837660074234, + 0.011994498781859875, + 0.01520491298288107, + 0.031007688492536545, + 0.022896872833371162, + 0.024865686893463135, + 0.020167110487818718, + 0.01219034381210804, + 0.012495397590100765, + 0.010707532986998558, + 0.013513082638382912, + 0.00904778391122818, + 0.012627177871763706, + 0.006758388131856918, + 0.005729437805712223, + 0.0070855868980288506, + 0.009147640317678452, + 0.008647141978144646, + 0.007545550353825092, + 0.009932151064276695, + 0.0061409263871610165, + 0.00879198033362627, + 0.006719715427607298, + 0.00442527886480093, + 0.004094282630831003, + 0.005300127901136875, + 0.0025096843019127846, + 0.003668632125481963, + 0.0075701754540205, + 0.004358963575214148, + 0.004713600035756826, + 0.00874335877597332, + 0.008478505536913872, + 0.009931392967700958, + 0.006734011694788933, + 0.013503257185220718, + 0.007935342378914356, + 0.009749755263328552, + 0.009538055397570133, + 0.042399197816848755, + 0.008883628994226456, + 0.0356818325817585, + 0.012996416538953781, + 0.007867124862968922, + 0.02072874642908573, + 0.010637345723807812, + 0.012562493793666363, + 0.012386852875351906, + 0.012297110632061958, + 0.010311267338693142, + 0.0062352074310183525, + 0.005629359744489193, + 0.010673776268959045, + 0.004998007323592901, + 0.006574563682079315, + 0.023313798010349274, + 0.009343189187347889, + 0.022204848006367683, + 0.006775728426873684, + 0.009307960979640484, + 0.01661909930408001, + 0.00891893357038498, + 0.013632284477353096, + 0.01102164015173912, + 0.004369061440229416, + 0.012896017171442509, + 0.004963345360010862, + 0.0033348589204251766, + 0.0056778304278850555, + 0.004603775683790445, + 0.006880191620439291, + 0.0529358945786953, + 0.007534360513091087, + 0.0165726188570261, + 0.006426657550036907, + 0.007685358636081219, + 0.06478297710418701, + 0.006810450926423073, + 0.012608795426785946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.008629176765680313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low angle and tilted slightly upward, captures the decisive moment of a young man in a dark gray quick-drying T-shirt and khaki cargo shorts completing a breathtaking leap. As his dusty hiking boots landed precisely and powerfully on the massive, jagged summit of gray granite, his knees bent and he squatted, deftly absorbing the impact, sending the tiny gravel beneath his feet bouncing slightly. Then, steadily and deliberately, he straightened his body, arms extended to his sides like bird wings, fingers slightly spread apart, carefully adjusting his balance to maintain perfect balance. His short, sweat-soaked brown hair blew tousled in the breeze, and his focused, resolute face, muscles taut and defined, betrayed a fierce concentration. The golden afterglow of the setting sun, streaming from behind and sideways, cast a dazzling golden rim across his figure, while his front was cast in soft shadow, enhancing the three-dimensionality of the image. The background is the silhouette of rolling mountains, dyed a deep purple by the setting sun. The whole scene is full of the sense of accomplishment of conquering nature and the magnificent tranquility of sunset.", + "seed": 5215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_449_seed_5215_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07405919507145882, + "vlm_logprob": -0.07405919507145882, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.07405919507145882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24056802690029144 + ], + "opposite_logprob_tokens": [ + -1.7405680418014526 + ], + "logprob_reward": -0.24056802690029144, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.24056802690029144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.122765451669693, + 0.12489687651395798, + 0.11300202459096909, + 0.07567206025123596, + 0.04678662121295929, + 0.03400208055973053, + 0.171620175242424, + 0.0402543731033802, + 0.18689647316932678, + 0.0841037854552269 + ], + "credit_map": [ + 0.011914657428860664, + 0.009131020866334438, + 0.0073677292093634605, + 0.008085637353360653, + 0.036782119423151016, + 0.03605048730969429, + 0.007677019573748112, + 0.005756792146712542, + 0.0220088679343462, + 0.02121911011636257, + 0.014471367932856083, + 0.010178419761359692, + 0.021056896075606346, + 0.017549410462379456, + 0.00927743036299944, + 0.00913538970053196, + 0.015877483412623405, + 0.012990185059607029, + 0.011774037033319473, + 0.011417589150369167, + 0.020974913612008095, + 0.01726384460926056, + 0.010446883738040924, + 0.01225709542632103, + 0.015692606568336487, + 0.00954663846641779, + 0.014323470182716846, + 0.005613573826849461, + 0.008975645527243614, + 0.010032995603978634, + 0.005309603177011013, + 0.006177529226988554, + 0.005546363536268473, + 0.007014986127614975, + 0.004274401348084211, + 0.004527000244706869, + 0.0055643972009420395, + 0.00811888836324215, + 0.004634265787899494, + 0.007106324192136526, + 0.0039390679448843, + 0.004601232707500458, + 0.0031141513027250767, + 0.004353806376457214, + 0.006380555685609579, + 0.0038768122904002666, + 0.004356598947197199, + 0.003379856003448367, + 0.031225822865962982, + 0.02944333292543888, + 0.0617336705327034, + 0.019321279600262642, + 0.005608087871223688, + 0.010697264224290848, + 0.0038790584076195955, + 0.0097116744145751, + 0.004603526555001736, + 0.007536790799349546, + 0.004529059398919344, + 0.0038460807409137487, + 0.004331137053668499, + 0.00540539575740695, + 0.00496818358078599, + 0.005034205969423056, + 0.02345356158912182, + 0.06317628175020218, + 0.03600896894931793, + 0.013833613134920597, + 0.008820602670311928, + 0.02497350238263607, + 0.006090349052101374, + 0.01053962204605341, + 0.01179156731814146, + 0.01417881902307272, + 0.007811674848198891, + 0.009409436024725437, + 0.010218597948551178, + 0.009842544794082642, + 0.011817575432360172, + 0.009033583104610443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.24056802690029144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026375453919172287 + ], + "opposite_logprob_tokens": [ + -3.6513755321502686 + ], + "logprob_reward": -0.026375453919172287, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026375453919172287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15382324159145355, + 0.07100628316402435, + 0.05583018809556961, + 0.07947424054145813, + 0.10191857069730759, + 0.04364258423447609, + 0.2829141616821289, + 0.04661251977086067, + 0.12071742862462997, + 0.04406077787280083 + ], + "credit_map": [ + 0.017359772697091103, + 0.012464907951653004, + 0.010608634911477566, + 0.008596234023571014, + 0.04548802599310875, + 0.04406654089689255, + 0.006768868770450354, + 0.008470260538160801, + 0.010250299237668514, + 0.011451471596956253, + 0.010452629067003727, + 0.008405981585383415, + 0.010559561662375927, + 0.009391608648002148, + 0.005334046203643084, + 0.005160687491297722, + 0.009251506999135017, + 0.00665467232465744, + 0.006375497207045555, + 0.0048975758254528046, + 0.009065333753824234, + 0.008312810212373734, + 0.005113739985972643, + 0.006159048527479172, + 0.02001783438026905, + 0.008789179846644402, + 0.01729024574160576, + 0.007794772740453482, + 0.007067675702273846, + 0.0070403520949184895, + 0.004505504388362169, + 0.006968676578253508, + 0.01582423783838749, + 0.02128223329782486, + 0.011974591761827469, + 0.009510808624327183, + 0.009415226057171822, + 0.01886378601193428, + 0.005332274828106165, + 0.009715406224131584, + 0.010385093279182911, + 0.007288509514182806, + 0.0053707328625023365, + 0.004733906127512455, + 0.0031681545078754425, + 0.005531274247914553, + 0.0029695648699998856, + 0.004195348359644413, + 0.034415967762470245, + 0.07193920761346817, + 0.09728384763002396, + 0.028869012370705605, + 0.009540185332298279, + 0.0197585616260767, + 0.00546218641102314, + 0.01564519852399826, + 0.006655498407781124, + 0.013842512853443623, + 0.003978393971920013, + 0.004294689744710922, + 0.004399574361741543, + 0.0062790196388959885, + 0.0035466060508042574, + 0.0036162245087325573, + 0.016766371205449104, + 0.0407540500164032, + 0.0156357754021883, + 0.008115914650261402, + 0.006560183595865965, + 0.016319850459694862, + 0.004832303151488304, + 0.011732985265552998, + 0.00717527698725462, + 0.006713953800499439, + 0.005612936336547136, + 0.005062239244580269, + 0.004113249946385622, + 0.005582654848694801, + 0.003721166169270873, + 0.006079302169382572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.026375453919172287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204133853316307 + ], + "opposite_logprob_tokens": [ + -2.8120412826538086 + ], + "logprob_reward": -0.06204133853316307, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204133853316307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03487642481923103, + 0.04409022256731987, + 0.03114892542362213, + 0.16338197886943817, + 0.055269043892621994, + 0.03423905000090599, + 0.17330747842788696, + 0.06261509656906128, + 0.2653995752334595, + 0.13567213714122772 + ], + "credit_map": [ + 0.0034194181207567453, + 0.004061669111251831, + 0.0049318838864564896, + 0.004232872277498245, + 0.00524663832038641, + 0.006758333649486303, + 0.00315670995041728, + 0.0030688976403325796, + 0.0070478301495313644, + 0.005489001050591469, + 0.007891937159001827, + 0.004245780408382416, + 0.005265089683234692, + 0.006645426154136658, + 0.003565221792086959, + 0.003939937334507704, + 0.004839947447180748, + 0.004679857287555933, + 0.0028356921393424273, + 0.0034966764505952597, + 0.005246847867965698, + 0.005346792750060558, + 0.0019385857740417123, + 0.0027645237278193235, + 0.04976530000567436, + 0.015597852878272533, + 0.04421557858586311, + 0.013255643658339977, + 0.007581609301269054, + 0.012519778683781624, + 0.007348878774791956, + 0.013097330927848816, + 0.008784003555774689, + 0.009858239442110062, + 0.005763388704508543, + 0.004264757968485355, + 0.007031363435089588, + 0.009779427200555801, + 0.004357369150966406, + 0.005430493038147688, + 0.00688402634114027, + 0.008090130984783173, + 0.0030572290997952223, + 0.002651839517056942, + 0.004442539997398853, + 0.003860100172460079, + 0.0025267358869314194, + 0.002726449631154537, + 0.04272719472646713, + 0.0704454556107521, + 0.023012584075331688, + 0.006569946650415659, + 0.00859623309224844, + 0.010427955538034439, + 0.004224180709570646, + 0.007303929887712002, + 0.008756550960242748, + 0.015195581130683422, + 0.006250578444451094, + 0.006556775886565447, + 0.0075613828375935555, + 0.006191764492541552, + 0.006555023603141308, + 0.005547439679503441, + 0.04627089947462082, + 0.09883641451597214, + 0.02928672730922699, + 0.016431273892521858, + 0.01428318116813898, + 0.03538721427321434, + 0.008771638385951519, + 0.016132201999425888, + 0.01628175377845764, + 0.016379082575440407, + 0.013800599612295628, + 0.014842708595097065, + 0.014832155779004097, + 0.019153010100126266, + 0.022633254528045654, + 0.017749574035406113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.06204133853316307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369072452187538 + ], + "opposite_logprob_tokens": [ + -3.4086906909942627 + ], + "logprob_reward": -0.03369072452187538, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369072452187538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08235783874988556, + 0.0911412462592125, + 0.048481471836566925, + 0.08439766615629196, + 0.032429039478302, + 0.032689500600099564, + 0.1849520057439804, + 0.01305741723626852, + 0.41129860281944275, + 0.019195156171917915 + ], + "credit_map": [ + 0.006585911847651005, + 0.006531245075166225, + 0.0051015447825193405, + 0.005155438557267189, + 0.023008257150650024, + 0.025561517104506493, + 0.004757985472679138, + 0.005655938759446144, + 0.014033420011401176, + 0.017369994893670082, + 0.013758309185504913, + 0.008222649805247784, + 0.011328923515975475, + 0.013749225065112114, + 0.00605673436075449, + 0.006621987093240023, + 0.010430068708956242, + 0.0066862222738564014, + 0.004533412866294384, + 0.0046658688224852085, + 0.0057581947185099125, + 0.008294453844428062, + 0.0028786901384592056, + 0.005234561860561371, + 0.007973522879183292, + 0.018820174038410187, + 0.009370958432555199, + 0.012818478047847748, + 0.008111618459224701, + 0.010527017526328564, + 0.007230096962302923, + 0.009545805864036083, + 0.006969565059989691, + 0.006432554218918085, + 0.004099978134036064, + 0.0035059174988418818, + 0.0021574466954916716, + 0.0046449038200080395, + 0.0019194846972823143, + 0.002699189120903611, + 0.004249462857842445, + 0.006448206026107073, + 0.003418674459680915, + 0.002972960937768221, + 0.004065496381372213, + 0.004753620829433203, + 0.0029684847686439753, + 0.0038125920109450817, + 0.06024986132979393, + 0.015970483422279358, + 0.05980096384882927, + 0.020321661606431007, + 0.005347293335944414, + 0.008064951747655869, + 0.004749703221023083, + 0.010447097942233086, + 0.0017914179479703307, + 0.002464222488924861, + 0.0018246974796056747, + 0.0016353075625374913, + 0.0012287562713027, + 0.001446871436201036, + 0.0012130640679970384, + 0.0014530797488987446, + 0.07856637984514236, + 0.17805318534374237, + 0.05165295675396919, + 0.01586269773542881, + 0.011829245835542679, + 0.059756822884082794, + 0.0046582999639213085, + 0.010919011197984219, + 0.0023769650142639875, + 0.0021675056777894497, + 0.002937598852440715, + 0.0024618846364319324, + 0.002269042655825615, + 0.0017185963224619627, + 0.0030307318083941936, + 0.0022328307386487722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.03369072452187538 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007620431482791901 + ], + "opposite_logprob_tokens": [ + -4.882620334625244 + ], + "logprob_reward": -0.007620431482791901, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007620431482791901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12362109869718552, + 0.05003713071346283, + 0.04482489451766014, + 0.061031635850667953, + 0.048627398908138275, + 0.058475248515605927, + 0.08829333633184433, + 0.1055498868227005, + 0.35066699981689453, + 0.0688723772764206 + ], + "credit_map": [ + 0.007238362450152636, + 0.008015146479010582, + 0.0071982876397669315, + 0.006417317781597376, + 0.04303126782178879, + 0.0411578007042408, + 0.005500794854015112, + 0.00506211631000042, + 0.005694981664419174, + 0.00925231073051691, + 0.007545815780758858, + 0.004385924898087978, + 0.007566452492028475, + 0.00712073128670454, + 0.0032750864047557116, + 0.0051958272233605385, + 0.008619656786322594, + 0.006046251393854618, + 0.004951367620378733, + 0.004383898805826902, + 0.006095421500504017, + 0.008068033494055271, + 0.00235991133376956, + 0.004300354979932308, + 0.014915629290044308, + 0.007123279385268688, + 0.014799132011830807, + 0.005220131482928991, + 0.004730490036308765, + 0.004498444963246584, + 0.004634643904864788, + 0.005109886638820171, + 0.010490800254046917, + 0.010956084355711937, + 0.005347034428268671, + 0.00564546650275588, + 0.003328067949041724, + 0.00619424507021904, + 0.003085261443629861, + 0.003580436808988452, + 0.013993805274367332, + 0.014217297546565533, + 0.006277256645262241, + 0.0037113684229552746, + 0.00576753169298172, + 0.007558662444353104, + 0.003236947348341346, + 0.0037123782094568014, + 0.009192144498229027, + 0.030906205996870995, + 0.02323049306869507, + 0.0076025426387786865, + 0.0033357685897499323, + 0.0070426734164357185, + 0.0023457063362002373, + 0.004637803416699171, + 0.016044676303863525, + 0.02444017119705677, + 0.01559191383421421, + 0.013802298344671726, + 0.006330434698611498, + 0.015345500782132149, + 0.004450530279427767, + 0.00954436231404543, + 0.06963198632001877, + 0.15317198634147644, + 0.0390855148434639, + 0.013541506603360176, + 0.010264493525028229, + 0.051229462027549744, + 0.004361310973763466, + 0.009380729869008064, + 0.008669877424836159, + 0.008709012530744076, + 0.015174699015915394, + 0.010617674328386784, + 0.005238483659923077, + 0.004519811365753412, + 0.004388394299894571, + 0.011554427444934845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.007620431482791901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low angle and tilted slightly upward, captures the decisive moment of a young man in a dark gray quick-drying T-shirt and khaki cargo shorts completing a breathtaking leap. As his dusty hiking boots landed precisely and powerfully on the massive, jagged summit of gray granite, his knees bent and he squatted, deftly absorbing the impact, sending the tiny gravel beneath his feet bouncing slightly. Then, steadily and deliberately, he straightened his body, arms extended to his sides like bird wings, fingers slightly spread apart, carefully adjusting his balance to maintain perfect balance. His short, sweat-soaked brown hair blew tousled in the breeze, and his focused, resolute face, muscles taut and defined, betrayed a fierce concentration. The golden afterglow of the setting sun, streaming from behind and sideways, cast a dazzling golden rim across his figure, while his front was cast in soft shadow, enhancing the three-dimensionality of the image. The background is the silhouette of rolling mountains, dyed a deep purple by the setting sun. The whole scene is full of the sense of accomplishment of conquering nature and the magnificent tranquility of sunset.", + "seed": 5216, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_3/index_449_seed_5216_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0930071578361094, + "vlm_logprob": -0.0930071578361094, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.0930071578361094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.274244487285614 + ], + "opposite_logprob_tokens": [ + -1.6492445468902588 + ], + "logprob_reward": -0.274244487285614, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.274244487285614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1696372777223587, + 0.040661994367837906, + 0.021488351747393608, + 0.05780233442783356, + 0.03602462261915207, + 0.11112554371356964, + 0.10988045483827591, + 0.318258672952652, + 0.045895326882600784, + 0.08922545611858368 + ], + "credit_map": [ + 0.028548700734972954, + 0.021333973854780197, + 0.016461387276649475, + 0.008912323974072933, + 0.04223635792732239, + 0.03058508038520813, + 0.008607948198914528, + 0.012951504439115524, + 0.007029522676020861, + 0.007229405455291271, + 0.003190357703715563, + 0.0018071436788886786, + 0.007972384802997112, + 0.004983928054571152, + 0.0037457707803696394, + 0.004703477956354618, + 0.005067827645689249, + 0.0034812830854207277, + 0.0018568418454378843, + 0.002003822475671768, + 0.0024787455331534147, + 0.0022103236988186836, + 0.0017045753775164485, + 0.0026849317364394665, + 0.010705162771046162, + 0.00909693818539381, + 0.004364685155451298, + 0.005552819464355707, + 0.006768996827304363, + 0.008098750375211239, + 0.00640019541606307, + 0.006814786698669195, + 0.008842973038554192, + 0.006140556186437607, + 0.003452595556154847, + 0.00325932283885777, + 0.003281261306256056, + 0.004263376351445913, + 0.002964874031022191, + 0.003819664940237999, + 0.015342182479798794, + 0.018932409584522247, + 0.007982579059898853, + 0.011219690553843975, + 0.010776735842227936, + 0.02074841782450676, + 0.008300975896418095, + 0.017822550609707832, + 0.01714777573943138, + 0.022877303883433342, + 0.01420428417623043, + 0.01329051610082388, + 0.006323943845927715, + 0.01808425597846508, + 0.005793274845927954, + 0.012159101665019989, + 0.04135278984904289, + 0.02190886065363884, + 0.11663392186164856, + 0.0776490643620491, + 0.007901372388005257, + 0.018569575622677803, + 0.009534899145364761, + 0.024708179756999016, + 0.002970885019749403, + 0.003984053619205952, + 0.01312554907053709, + 0.007989340461790562, + 0.0037643606774508953, + 0.002649840898811817, + 0.005334592889994383, + 0.006076703779399395, + 0.025453703477978706, + 0.02298898622393608, + 0.007036079186946154, + 0.005813946947455406, + 0.00497085228562355, + 0.004878791980445385, + 0.010171300731599331, + 0.007911798544228077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.274244487285614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02333407662808895 + ], + "opposite_logprob_tokens": [ + -3.77333402633667 + ], + "logprob_reward": -0.02333407662808895, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333407662808895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09013336151838303, + 0.039443548768758774, + 0.020071865990757942, + 0.04310759902000427, + 0.027481840923428535, + 0.02087528072297573, + 0.15395919978618622, + 0.5197461843490601, + 0.05154326558113098, + 0.03363780304789543 + ], + "credit_map": [ + 0.027566134929656982, + 0.008304058574140072, + 0.012560388073325157, + 0.004613549914211035, + 0.008936239406466484, + 0.017951637506484985, + 0.004910040646791458, + 0.005291308742016554, + 0.00677351513877511, + 0.007394385524094105, + 0.003534342860803008, + 0.0017902720719575882, + 0.006772344466298819, + 0.005393481347709894, + 0.0031748914625495672, + 0.004610313568264246, + 0.003637261688709259, + 0.0034089551772922277, + 0.0013880804181098938, + 0.0012407510075718164, + 0.003470475785434246, + 0.00322199915535748, + 0.0018530577654018998, + 0.0018512839451432228, + 0.007467296905815601, + 0.008453772403299809, + 0.003527967957779765, + 0.003307982347905636, + 0.006579203996807337, + 0.006558617576956749, + 0.003084799973294139, + 0.004127957858145237, + 0.00701460475102067, + 0.005845905281603336, + 0.0020010957960039377, + 0.0022548744454979897, + 0.00224921852350235, + 0.003063101088628173, + 0.0020949533209204674, + 0.002958086784929037, + 0.004707674961537123, + 0.0033737740013748407, + 0.001986893592402339, + 0.0025112212169915438, + 0.0019179051741957664, + 0.0020055770874023438, + 0.0015286182751879096, + 0.0028436146676540375, + 0.02377181127667427, + 0.024707160890102386, + 0.03474482521414757, + 0.027622489258646965, + 0.005120653193444014, + 0.017055943608283997, + 0.0054488168098032475, + 0.015487500466406345, + 0.05025485157966614, + 0.1218012347817421, + 0.17865507304668427, + 0.10669375211000443, + 0.011031017638742924, + 0.02176273986697197, + 0.008235061541199684, + 0.02131248638033867, + 0.004654628690332174, + 0.005194346886128187, + 0.012560312636196613, + 0.010247373953461647, + 0.002891070442274213, + 0.003664460964500904, + 0.004820340778678656, + 0.007510732859373093, + 0.0065394798293709755, + 0.007036282680928707, + 0.004539028741419315, + 0.0013953505549579859, + 0.0023741736076772213, + 0.004084308631718159, + 0.0029980686958879232, + 0.004671111702919006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.02333407662808895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698905259370804 + ], + "opposite_logprob_tokens": [ + -3.408698797225952 + ], + "logprob_reward": -0.033698905259370804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698905259370804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03818262740969658, + 0.034960515797138214, + 0.016647635027766228, + 0.033531032502651215, + 0.034774478524923325, + 0.06154797598719597, + 0.07735575735569, + 0.4224162995815277, + 0.055091675370931625, + 0.22549207508563995 + ], + "credit_map": [ + 0.007229804061353207, + 0.00365451630204916, + 0.004145028535276651, + 0.0018938970752060413, + 0.008120812475681305, + 0.0061589921824634075, + 0.002810397418215871, + 0.004169179126620293, + 0.005731984972953796, + 0.008305949158966541, + 0.002596877748146653, + 0.0016278091352432966, + 0.004821746610105038, + 0.0063653248362243176, + 0.0022091660648584366, + 0.0033016602974385023, + 0.0036974167451262474, + 0.002636742778122425, + 0.0015532727120444179, + 0.0017030254239216447, + 0.002043677493929863, + 0.0018821714911609888, + 0.0013012884883210063, + 0.0018300402443856, + 0.008120640181005001, + 0.007386055774986744, + 0.0020520996768027544, + 0.0030014009680598974, + 0.0032946262508630753, + 0.0037763873115181923, + 0.002557906787842512, + 0.0033419146202504635, + 0.007538266479969025, + 0.00610565859824419, + 0.00270771118812263, + 0.0037633837200701237, + 0.002973963739350438, + 0.003955453168600798, + 0.00335294590331614, + 0.004377095494419336, + 0.008335785008966923, + 0.011562994681298733, + 0.004965115338563919, + 0.007315691560506821, + 0.006736611481755972, + 0.006892178673297167, + 0.004431473556905985, + 0.011308128945529461, + 0.012861482799053192, + 0.020948506891727448, + 0.008402934297919273, + 0.007776143029332161, + 0.004280310124158859, + 0.011242023669183254, + 0.003679751418530941, + 0.008164603263139725, + 0.027865251526236534, + 0.050044890493154526, + 0.1591484397649765, + 0.09941565245389938, + 0.01353700365871191, + 0.01648818328976631, + 0.016994725912809372, + 0.038922134786844254, + 0.003389237681403756, + 0.0035735205747187138, + 0.01667999103665352, + 0.010458061471581459, + 0.0027433631476014853, + 0.00407714257016778, + 0.006281028967350721, + 0.00788932852447033, + 0.05723005160689354, + 0.05194132402539253, + 0.027098625898361206, + 0.015067669562995434, + 0.015062453225255013, + 0.012023327872157097, + 0.025978228077292442, + 0.021090393885970116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.033698905259370804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702852487564087 + ], + "opposite_logprob_tokens": [ + -2.127028465270996 + ], + "logprob_reward": -0.12702852487564087, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702852487564087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05345480144023895, + 0.12578755617141724, + 0.08797414600849152, + 0.20388124883174896, + 0.033342283219099045, + 0.026369059458374977, + 0.18327702581882477, + 0.14504601061344147, + 0.04611913859844208, + 0.09474869072437286 + ], + "credit_map": [ + 0.011395283974707127, + 0.00586387375369668, + 0.005616148468106985, + 0.0034182313829660416, + 0.010387859307229519, + 0.009754308499395847, + 0.0029640779830515385, + 0.0040550194680690765, + 0.01918100379407406, + 0.01991754025220871, + 0.008632348850369453, + 0.004931616131216288, + 0.03714015707373619, + 0.016031140461564064, + 0.008931498043239117, + 0.01102224551141262, + 0.024310030043125153, + 0.016191313043236732, + 0.0071639856323599815, + 0.00783140305429697, + 0.009718003682792187, + 0.009142335504293442, + 0.006907620467245579, + 0.006709453649818897, + 0.032166432589292526, + 0.035672303289175034, + 0.018083956092596054, + 0.02178817056119442, + 0.030748965218663216, + 0.02708384580910206, + 0.018123267218470573, + 0.020214296877384186, + 0.0069995722733438015, + 0.005431052763015032, + 0.0025806385092437267, + 0.003661140101030469, + 0.0030145011842250824, + 0.0046566869132220745, + 0.0030682755168527365, + 0.00393041642382741, + 0.005049798637628555, + 0.004510201513767242, + 0.0024245386011898518, + 0.002883798209950328, + 0.0023295104037970304, + 0.0036236755549907684, + 0.0017943565035238862, + 0.00375318038277328, + 0.03271643817424774, + 0.041869860142469406, + 0.0322500579059124, + 0.031653162091970444, + 0.006574500352144241, + 0.014608759433031082, + 0.0059263333678245544, + 0.01767790876328945, + 0.03564046695828438, + 0.05750463157892227, + 0.0202051792293787, + 0.010018479079008102, + 0.003394597675651312, + 0.01128595694899559, + 0.0022430962417274714, + 0.004753618035465479, + 0.004130132962018251, + 0.005801001098006964, + 0.013733656145632267, + 0.007504611276090145, + 0.0029183803126215935, + 0.003555453382432461, + 0.0036399441305547953, + 0.004835961852222681, + 0.025829749181866646, + 0.02582252398133278, + 0.010769911110401154, + 0.0040304833091795444, + 0.004333718679845333, + 0.009543615393340588, + 0.005972321145236492, + 0.008446374908089638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.12702852487564087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle tilted slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young man wearing a dark gray quick-drying T-shirt and khaki cargo shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the man landing on a massive, jagged summit of gray granite with gravel bouncing beneath his feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling mountains silhouetted in deep purple against a golden setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006729795131832361 + ], + "opposite_logprob_tokens": [ + -5.006729602813721 + ], + "logprob_reward": -0.006729795131832361, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006729795131832361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0474475622177124, + 0.03933335095643997, + 0.020794274285435677, + 0.03942383453249931, + 0.016523249447345734, + 0.01799081824719906, + 0.04727223888039589, + 0.6780530214309692, + 0.05041978508234024, + 0.04274177551269531 + ], + "credit_map": [ + 0.012540529482066631, + 0.00456614512950182, + 0.004664685111492872, + 0.0032769443932920694, + 0.006611300632357597, + 0.008177531883120537, + 0.0032677713315933943, + 0.004342658445239067, + 0.006948762107640505, + 0.009321513585746288, + 0.0032823681831359863, + 0.0021964958868920803, + 0.004604400601238012, + 0.006426401436328888, + 0.00288212182931602, + 0.0036712910514324903, + 0.007180949207395315, + 0.004478925839066505, + 0.0012498999712988734, + 0.0012621245114132762, + 0.002247406402602792, + 0.002055385382845998, + 0.0011572045041248202, + 0.001162379514425993, + 0.007608573418110609, + 0.00782685074955225, + 0.003206638852134347, + 0.0037547890096902847, + 0.00547168729826808, + 0.0053625996224582195, + 0.002808661898598075, + 0.003384036011993885, + 0.003974996507167816, + 0.0028847840148955584, + 0.0011570702772587538, + 0.0014434391632676125, + 0.0011164796305820346, + 0.002179659204557538, + 0.001555111724883318, + 0.002211711835116148, + 0.003991066012531519, + 0.004069982096552849, + 0.0014426031848415732, + 0.0011155109386891127, + 0.0017852201126515865, + 0.002475196961313486, + 0.0011200339067727327, + 0.001991206780076027, + 0.008744106628000736, + 0.011921047233045101, + 0.004603562410920858, + 0.005234976764768362, + 0.001872593304142356, + 0.008684451691806316, + 0.0018157020676881075, + 0.004395803902298212, + 0.10964649170637131, + 0.21479398012161255, + 0.1811373382806778, + 0.10142737627029419, + 0.012465821579098701, + 0.03810034692287445, + 0.007422838360071182, + 0.013058931566774845, + 0.006428035907447338, + 0.009151481091976166, + 0.011588801629841328, + 0.006875963881611824, + 0.00237066438421607, + 0.006798791699111462, + 0.0026456215418875217, + 0.00456042867153883, + 0.012715058401226997, + 0.011142505332827568, + 0.004867997486144304, + 0.00279175559990108, + 0.0017391776200383902, + 0.0019892374984920025, + 0.0035856980830430984, + 0.003910350147634745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.006729795131832361 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_tensor([8194, 8194])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_tensor([8194, 8194])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1089267585f63f0436e437f0922dbe932c837875 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_tensor([8194, 8194])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot at an extremely high frame rate, the high-speed macro camera footage begins with static, piercing focus, capturing the inner workings of a highly competitive gearbox against a deep, non-reflective black background. In profound slow motion, a massive, mirror-polished steel gear, its surface gleaming with brilliant mirror highlights, slams into position with silent, impactful force. Its perfectly machined teeth mesh with another with utter precision, the perfect interlock sending a powerful, resonant tremor through the entire assembly, causing it to shake and sway as if it were a ship pitching and tumbling on a rough sea. The impact violently launches droplets of golden, viscous oil into the air; these globular beads, suspended in the void, arc in elegant arcs, catching the harsh lights of a movie studio like tiny, glowing amber gems. Then, the camera's focus subtly tilts to reveal the intricate dance of the selector and fork, their matte gunmetal finish contrasting sharply with the polished gears as they move with swift, deliberate purpose. The entire scene is a surreal and primal spectacle, with the stark contrast between cool, gleaming steel, warm, glistening oil, and the absolute black background creating a captivating and tense atmosphere of controlled mechanical rage.", + "seed": 29676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_8194_seed_29676_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.039970542304217815, + "vlm_logprob": -0.039970542304217815, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.039970542304217815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04832668602466583 + ], + "opposite_logprob_tokens": [ + -3.1733267307281494 + ], + "logprob_reward": -0.04832668602466583, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04832668602466583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27571651339530945, + 0.02948339655995369, + 0.05698029324412346, + 0.03558764234185219, + 0.14899486303329468, + 0.023823140189051628, + 0.23324698209762573, + 0.0397392101585865, + 0.07108406722545624, + 0.08534398674964905 + ], + "credit_map": [ + 0.013136832974851131, + 0.014135030098259449, + 0.017076438292860985, + 0.01674506813287735, + 0.02172429859638214, + 0.015869630500674248, + 0.08765918016433716, + 0.089369997382164, + 0.0029951264150440693, + 0.0027121035382151604, + 0.003927955869585276, + 0.003962312359362841, + 0.002899525221437216, + 0.003790416521951556, + 0.003851866815239191, + 0.005344087723642588, + 0.008838271722197533, + 0.008063504472374916, + 0.006675571668893099, + 0.00868451502174139, + 0.005707792472094297, + 0.007757879793643951, + 0.00628844927996397, + 0.004964303690940142, + 0.004414185881614685, + 0.004142211750149727, + 0.0033482294529676437, + 0.004922065883874893, + 0.005072969477623701, + 0.004861549474298954, + 0.005008049309253693, + 0.003818378783762455, + 0.02175806649029255, + 0.018501242622733116, + 0.012356920167803764, + 0.02322397381067276, + 0.013637684285640717, + 0.023218471556901932, + 0.020773984491825104, + 0.015524503774940968, + 0.0037440794985741377, + 0.0061182198114693165, + 0.0024177883751690388, + 0.0018336904468014836, + 0.002088174456730485, + 0.0022350994404405355, + 0.0034667192958295345, + 0.0019193680491298437, + 0.033702120184898376, + 0.026730945333838463, + 0.06136532500386238, + 0.018044203519821167, + 0.027373675256967545, + 0.021698562428355217, + 0.015543836168944836, + 0.028788277879357338, + 0.0052018193528056145, + 0.00565089238807559, + 0.007608912885189056, + 0.0030477959662675858, + 0.006147376261651516, + 0.0045236810110509396, + 0.003071973565965891, + 0.004486752673983574, + 0.006608989089727402, + 0.01049291156232357, + 0.013516229577362537, + 0.012138968333601952, + 0.013290757313370705, + 0.004641551990061998, + 0.004486915655434132, + 0.005907732527703047, + 0.008488316088914871, + 0.007484536152333021, + 0.007142280228435993, + 0.020571893081068993, + 0.007675030268728733, + 0.015055187977850437, + 0.01074136234819889, + 0.008185362443327904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.04832668602466583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651158772408962 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651158772408962, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651158772408962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44377967715263367, + 0.06537435203790665, + 0.05647667497396469, + 0.041069645434617996, + 0.09653183817863464, + 0.04903938248753548, + 0.02136938087642193, + 0.08889134973287582, + 0.08795685321092606, + 0.04951082915067673 + ], + "credit_map": [ + 0.02132653258740902, + 0.016767600551247597, + 0.03840591758489609, + 0.026996083557605743, + 0.030176324769854546, + 0.02493448555469513, + 0.1745128184556961, + 0.11065991967916489, + 0.00570938503369689, + 0.005082044284790754, + 0.008471657522022724, + 0.007414421532303095, + 0.007639849092811346, + 0.007251125294715166, + 0.009349861182272434, + 0.0144560057669878, + 0.009944483637809753, + 0.006311564706265926, + 0.006428560707718134, + 0.007563375402241945, + 0.004885382484644651, + 0.007121846545487642, + 0.008283056318759918, + 0.005938402842730284, + 0.004085545893758535, + 0.003952786326408386, + 0.0059641385450959206, + 0.008671768009662628, + 0.003707777475938201, + 0.004330743569880724, + 0.004204413387924433, + 0.006152472924441099, + 0.01422204077243805, + 0.011984081007540226, + 0.007387117948383093, + 0.01780710555613041, + 0.008224662393331528, + 0.012841461226344109, + 0.013351562432944775, + 0.01071380265057087, + 0.007894108071923256, + 0.007847641594707966, + 0.004451895132660866, + 0.004151544533669949, + 0.003645590739324689, + 0.005619209259748459, + 0.009807540103793144, + 0.005621850956231356, + 0.004526639822870493, + 0.002934650518000126, + 0.003301674732938409, + 0.002337861806154251, + 0.0020651943050324917, + 0.002750928280875087, + 0.0014231721870601177, + 0.0020292613189667463, + 0.02212788537144661, + 0.017736883834004402, + 0.011704111471772194, + 0.0039261444471776485, + 0.018747709691524506, + 0.005006145685911179, + 0.0037191829178482294, + 0.0059232888743281364, + 0.009882153943181038, + 0.012157639488577843, + 0.015036500990390778, + 0.014850719831883907, + 0.01657346449792385, + 0.005658510606735945, + 0.006808844860643148, + 0.0069890208542346954, + 0.007507792674005032, + 0.005625430028885603, + 0.004594518803060055, + 0.008625052869319916, + 0.006120615638792515, + 0.0067733582109212875, + 0.005796895828098059, + 0.004467163234949112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.008651158772408962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800057411193848 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029800057411193848, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800057411193848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28092533349990845, + 0.05838320031762123, + 0.046550773084163666, + 0.03334962576627731, + 0.05424784868955612, + 0.0590050183236599, + 0.0891534611582756, + 0.2243773341178894, + 0.07470410317182541, + 0.07930328696966171 + ], + "credit_map": [ + 0.017189156264066696, + 0.021333180367946625, + 0.05354698374867439, + 0.03600779548287392, + 0.029641056433320045, + 0.01680528186261654, + 0.08748459815979004, + 0.018917283043265343, + 0.005004156846553087, + 0.004707054700702429, + 0.0064930482767522335, + 0.005456261802464724, + 0.00734669016674161, + 0.007360529154539108, + 0.010992906987667084, + 0.011022556573152542, + 0.010871161706745625, + 0.004712740890681744, + 0.0042601218447089195, + 0.004403477068990469, + 0.0028280233964323997, + 0.006786466576159, + 0.005625635385513306, + 0.007063145749270916, + 0.004101655911654234, + 0.0036348525900393724, + 0.003677795873954892, + 0.004933160729706287, + 0.004269177559763193, + 0.004088600166141987, + 0.004712356720119715, + 0.00393202668055892, + 0.009181196801364422, + 0.013511649332940578, + 0.005266765132546425, + 0.01027669943869114, + 0.0037820376455783844, + 0.004253927152603865, + 0.0046847546473145485, + 0.003290816443040967, + 0.011405092664062977, + 0.019991787150502205, + 0.005837402306497097, + 0.004600991494953632, + 0.0031423300970345736, + 0.004023080226033926, + 0.005548011977225542, + 0.004456325434148312, + 0.013622645288705826, + 0.010670618154108524, + 0.01473036129027605, + 0.008114520460367203, + 0.011109634302556515, + 0.010628430172801018, + 0.007690149825066328, + 0.012587109580636024, + 0.05058663710951805, + 0.04283398762345314, + 0.031214270740747452, + 0.012739044614136219, + 0.050227873027324677, + 0.010218260809779167, + 0.009686884470283985, + 0.016870371997356415, + 0.00722214300185442, + 0.01185136754065752, + 0.010977555997669697, + 0.00912072416394949, + 0.01688506081700325, + 0.004905201494693756, + 0.005279053468257189, + 0.0084629962220788, + 0.008075849153101444, + 0.006814553868025541, + 0.005614054389297962, + 0.019033141434192657, + 0.008163213729858398, + 0.011651184409856796, + 0.01075752917677164, + 0.009193760342895985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.029800057411193848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043060772120952606 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.043060772120952606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043060772120952606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1696251481771469, + 0.055858198553323746, + 0.07390756160020828, + 0.038716934621334076, + 0.12054453790187836, + 0.08713600784540176, + 0.11246532946825027, + 0.20540638267993927, + 0.06561609357595444, + 0.07072378695011139 + ], + "credit_map": [ + 0.009192518889904022, + 0.010971471667289734, + 0.015438064932823181, + 0.014556470327079296, + 0.014766190201044083, + 0.01237940602004528, + 0.043184392154216766, + 0.04913663491606712, + 0.0054246471263468266, + 0.004640423227101564, + 0.006792123895138502, + 0.005203971639275551, + 0.008398517966270447, + 0.00913652777671814, + 0.007593567483127117, + 0.008668421767652035, + 0.012484030798077583, + 0.007490368559956551, + 0.00924849882721901, + 0.006970139220356941, + 0.007837285287678242, + 0.007150197401642799, + 0.009849894791841507, + 0.01287714671343565, + 0.004467568825930357, + 0.0066041951067745686, + 0.0050911628641188145, + 0.006119999568909407, + 0.003935558255761862, + 0.00444169994443655, + 0.003926494624465704, + 0.004130254965275526, + 0.018221579492092133, + 0.015197088941931725, + 0.010981312021613121, + 0.023040998727083206, + 0.013549529947340488, + 0.013570082373917103, + 0.01527879573404789, + 0.010705159977078438, + 0.018200144171714783, + 0.01770063303411007, + 0.008535643108189106, + 0.007163615431636572, + 0.007669826038181782, + 0.008311218582093716, + 0.009973868727684021, + 0.009581057354807854, + 0.01650230586528778, + 0.013356748037040234, + 0.014441000297665596, + 0.01394171454012394, + 0.016494901850819588, + 0.014121259562671185, + 0.009739923290908337, + 0.013867469504475594, + 0.028511177748441696, + 0.027081143110990524, + 0.01866123452782631, + 0.016355527564883232, + 0.03343243896961212, + 0.024003438651561737, + 0.02051013708114624, + 0.03685129061341286, + 0.008264188654720783, + 0.008884039707481861, + 0.010088731534779072, + 0.014161590486764908, + 0.009813934564590454, + 0.004700096789747477, + 0.0033911296632140875, + 0.006312383338809013, + 0.006303143687546253, + 0.00690833292901516, + 0.004982237704098225, + 0.01469371933490038, + 0.007639871444553137, + 0.013332145288586617, + 0.00883301068097353, + 0.008031328208744526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.043060772120952606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001403719186783 + ], + "opposite_logprob_tokens": [ + -2.695013999938965 + ], + "logprob_reward": -0.07001403719186783, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001403719186783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20269489288330078, + 0.057303495705127716, + 0.05977831408381462, + 0.04268048703670502, + 0.14333952963352203, + 0.08963111788034439, + 0.0742260292172432, + 0.12877243757247925, + 0.07313386350870132, + 0.12843982875347137 + ], + "credit_map": [ + 0.010466726496815681, + 0.00730609567835927, + 0.012703297659754753, + 0.012871054001152515, + 0.012841681018471718, + 0.011871084570884705, + 0.0797654390335083, + 0.05486953631043434, + 0.005365004763007164, + 0.004995578434318304, + 0.00816385354846716, + 0.007306603714823723, + 0.007493774406611919, + 0.007357374764978886, + 0.005633150227367878, + 0.01098815631121397, + 0.006695355288684368, + 0.009695177897810936, + 0.009103153832256794, + 0.009028870612382889, + 0.005311697255820036, + 0.008538933470845222, + 0.006413309834897518, + 0.004991818685084581, + 0.004133499693125486, + 0.005263135302811861, + 0.005602181423455477, + 0.007247165311127901, + 0.003661239054054022, + 0.004926909692585468, + 0.005272472742944956, + 0.006573881953954697, + 0.02296551689505577, + 0.01854165643453598, + 0.013070384040474892, + 0.03368283063173294, + 0.012316183187067509, + 0.01763349398970604, + 0.014395087026059628, + 0.010734380222856998, + 0.015462051145732403, + 0.021457703784108162, + 0.010638617910444736, + 0.005966163706034422, + 0.008665555156767368, + 0.00654745614156127, + 0.012301534414291382, + 0.008592033758759499, + 0.01392337679862976, + 0.010355777107179165, + 0.009411098435521126, + 0.008534181863069534, + 0.008333675563335419, + 0.010554337874054909, + 0.005571612622588873, + 0.007541969418525696, + 0.019816963002085686, + 0.016981789842247963, + 0.01335239503532648, + 0.010357053019106388, + 0.02029750868678093, + 0.012199106626212597, + 0.014626005664467812, + 0.02114160545170307, + 0.009332854300737381, + 0.009279758669435978, + 0.010940593667328358, + 0.01576382666826248, + 0.009588532149791718, + 0.005269790533930063, + 0.004513869062066078, + 0.008444631472229958, + 0.009724347852170467, + 0.009196290746331215, + 0.008724365383386612, + 0.03301052376627922, + 0.014705603942275047, + 0.028989888727664948, + 0.015162917785346508, + 0.008925894275307655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.07001403719186783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot at an extremely high frame rate, the high-speed macro camera footage begins with static, piercing focus, capturing the inner workings of a highly competitive gearbox against a deep, non-reflective black background. In profound slow motion, a massive, mirror-polished steel gear, its surface gleaming with brilliant mirror highlights, slams into position with silent, impactful force. Its perfectly machined teeth mesh with another with utter precision, the perfect interlock sending a powerful, resonant tremor through the entire assembly, causing it to shake and sway as if it were a ship pitching and tumbling on a rough sea. The impact violently launches droplets of golden, viscous oil into the air; these globular beads, suspended in the void, arc in elegant arcs, catching the harsh lights of a movie studio like tiny, glowing amber gems. Then, the camera's focus subtly tilts to reveal the intricate dance of the selector and fork, their matte gunmetal finish contrasting sharply with the polished gears as they move with swift, deliberate purpose. The entire scene is a surreal and primal spectacle, with the stark contrast between cool, gleaming steel, warm, glistening oil, and the absolute black background creating a captivating and tense atmosphere of controlled mechanical rage.", + "seed": 29677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_8194_seed_29677_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03626501634716987, + "vlm_logprob": -0.03626501634716987, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.03626501634716987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010844580829143524 + ], + "opposite_logprob_tokens": [ + -4.885844707489014 + ], + "logprob_reward": -0.010844580829143524, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010844580829143524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05384108051657677, + 0.05464458465576172, + 0.025108719244599342, + 0.16359907388687134, + 0.04839261248707771, + 0.03191962465643883, + 0.025146452710032463, + 0.07973981648683548, + 0.18598847091197968, + 0.3316195607185364 + ], + "credit_map": [ + 0.0075437831692397594, + 0.006032607983797789, + 0.009265468455851078, + 0.006491557694971561, + 0.0055067106150090694, + 0.006783984135836363, + 0.006704165134578943, + 0.005512802395969629, + 0.00889701209962368, + 0.0073428889736533165, + 0.008219307288527489, + 0.009365176782011986, + 0.0047409855760633945, + 0.005281259771436453, + 0.005181694403290749, + 0.005616258829832077, + 0.0028799118008464575, + 0.0033045304007828236, + 0.00273341778665781, + 0.0038634352385997772, + 0.0038757396396249533, + 0.003540457459166646, + 0.002746938494965434, + 0.0021642884239554405, + 0.012963403947651386, + 0.011452089063823223, + 0.007404679898172617, + 0.010514277033507824, + 0.01786610297858715, + 0.038107164204120636, + 0.022170547395944595, + 0.04312080889940262, + 0.004362940322607756, + 0.0026452899910509586, + 0.0021683387458324432, + 0.00465579004958272, + 0.005472568795084953, + 0.012645267881453037, + 0.007429976481944323, + 0.009012441150844097, + 0.002535484032705426, + 0.0016360563458874822, + 0.0024960448499768972, + 0.0029749267268925905, + 0.002906845649704337, + 0.006695693824440241, + 0.0060060094110667706, + 0.006668565794825554, + 0.002106440020725131, + 0.002067058812826872, + 0.0028712074272334576, + 0.0019913052674382925, + 0.0026428522542119026, + 0.005544617306441069, + 0.0034109337721019983, + 0.0045120371505618095, + 0.006724694278091192, + 0.007947766222059727, + 0.008683247491717339, + 0.006657129153609276, + 0.0048728338442742825, + 0.016805967316031456, + 0.011879202909767628, + 0.016168976202607155, + 0.01217617280781269, + 0.00897988025099039, + 0.018587490543723106, + 0.028674649074673653, + 0.00676442077383399, + 0.036547355353832245, + 0.04233572259545326, + 0.031922776252031326, + 0.015479515306651592, + 0.02440430037677288, + 0.01890210434794426, + 0.01877998374402523, + 0.04536108300089836, + 0.06614838540554047, + 0.026396332308650017, + 0.11614787578582764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.010844580829143524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052642300724983215 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.0052642300724983215, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052642300724983215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09086106717586517, + 0.0866813063621521, + 0.058504581451416016, + 0.2889026701450348, + 0.0809207633137703, + 0.071895070374012, + 0.048247311264276505, + 0.022588172927498817, + 0.10675915330648422, + 0.14463986456394196 + ], + "credit_map": [ + 0.012828468345105648, + 0.009483910165727139, + 0.015900088474154472, + 0.011029425077140331, + 0.008155104704201221, + 0.009814824908971786, + 0.014364604838192463, + 0.00928464811295271, + 0.02080831304192543, + 0.021057119593024254, + 0.007318506948649883, + 0.012745189480483532, + 0.006123960018157959, + 0.005413093604147434, + 0.005384915042668581, + 0.007830202579498291, + 0.007066249847412109, + 0.007329699117690325, + 0.005801226478070021, + 0.007822792045772076, + 0.011925455182790756, + 0.007013259921222925, + 0.007426319178193808, + 0.004119575954973698, + 0.03268604725599289, + 0.02118823677301407, + 0.014128770679235458, + 0.024328647181391716, + 0.029438002035021782, + 0.05933219939470291, + 0.04594910517334938, + 0.06185168772935867, + 0.00947596039623022, + 0.006309662479907274, + 0.0038149969186633825, + 0.005572712980210781, + 0.0193999782204628, + 0.01250870805233717, + 0.010280507616698742, + 0.013558235950767994, + 0.006339694373309612, + 0.0053061372600495815, + 0.004534154664725065, + 0.0064020911231637, + 0.0145852230489254, + 0.013499505817890167, + 0.0091179721057415, + 0.012110303156077862, + 0.0027810093015432358, + 0.0032791919074952602, + 0.004271504934877157, + 0.0064608571119606495, + 0.003204177599400282, + 0.012739376164972782, + 0.006617221049964428, + 0.008893972262740135, + 0.002271731849759817, + 0.001470708753913641, + 0.0024313898757100105, + 0.001858031959272921, + 0.0014173772651702166, + 0.004310841206461191, + 0.0038229278288781643, + 0.005005164071917534, + 0.007048991043120623, + 0.005621692631393671, + 0.009589049965143204, + 0.014922930859029293, + 0.0051542310975492, + 0.02147511951625347, + 0.027662670239806175, + 0.015284462831914425, + 0.012762409634888172, + 0.013405953533947468, + 0.016886282712221146, + 0.01471256185323, + 0.01893630623817444, + 0.022656027227640152, + 0.020265331491827965, + 0.025014985352754593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.0052642300724983215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809535130858421 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809535130858421, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809535130858421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05882958695292473, + 0.03707381337881088, + 0.049875225871801376, + 0.3802013695240021, + 0.29348060488700867, + 0.028694864362478256, + 0.01953166536986828, + 0.02957089990377426, + 0.041022781282663345, + 0.061719123274087906 + ], + "credit_map": [ + 0.007237654644995928, + 0.0070123374462127686, + 0.008567425422370434, + 0.007842104882001877, + 0.005978261586278677, + 0.007326025515794754, + 0.008968900889158249, + 0.005896875634789467, + 0.0076276808977127075, + 0.007591376081109047, + 0.004664954263716936, + 0.0038234719540923834, + 0.0035797914024442434, + 0.0032373550347983837, + 0.0031948976684361696, + 0.00335428724065423, + 0.006763183977454901, + 0.005141769535839558, + 0.004408192355185747, + 0.005183760076761246, + 0.009316781535744667, + 0.006503232289105654, + 0.008043467998504639, + 0.00451483903452754, + 0.0342884361743927, + 0.023618856444954872, + 0.01668092980980873, + 0.021722085773944855, + 0.033828284591436386, + 0.08167799562215805, + 0.051553111523389816, + 0.11683167517185211, + 0.034274760633707047, + 0.014906507916748524, + 0.012647600844502449, + 0.03101600520312786, + 0.0291603934019804, + 0.06569221615791321, + 0.048767779022455215, + 0.05701534077525139, + 0.002725247060880065, + 0.0018153260461986065, + 0.0020379042252898216, + 0.0020768234971910715, + 0.005354580469429493, + 0.006108728237450123, + 0.003736161394044757, + 0.004840095527470112, + 0.0011687369551509619, + 0.0013321663718670607, + 0.0015078906435519457, + 0.0023733351845294237, + 0.001346612349152565, + 0.004712611436843872, + 0.002812158316373825, + 0.004278155509382486, + 0.00258674961514771, + 0.002829632256180048, + 0.0032504971604794264, + 0.0024857069365680218, + 0.0015823309076949954, + 0.00556733226403594, + 0.004773246124386787, + 0.0064954026602208614, + 0.0031942478381097317, + 0.0019273344660177827, + 0.004164192825555801, + 0.007214065175503492, + 0.0017821226501837373, + 0.007516141980886459, + 0.009768961928784847, + 0.005455716513097286, + 0.007355157285928726, + 0.007405459880828857, + 0.006098144222050905, + 0.004151027649641037, + 0.00885741040110588, + 0.010238763876259327, + 0.0055277906358242035, + 0.012085366994142532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.03809535130858421 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809443116188049 + ], + "opposite_logprob_tokens": [ + -3.2880945205688477 + ], + "logprob_reward": -0.03809443116188049, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809443116188049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18545711040496826, + 0.08079448342323303, + 0.07103519886732101, + 0.19751876592636108, + 0.060929346829652786, + 0.030769001692533493, + 0.04039982706308365, + 0.048178110271692276, + 0.09106608480215073, + 0.19385209679603577 + ], + "credit_map": [ + 0.023684674873948097, + 0.022608567029237747, + 0.0279732346534729, + 0.019676119089126587, + 0.02264990285038948, + 0.025007691234350204, + 0.02493022382259369, + 0.01892668940126896, + 0.016172394156455994, + 0.015135948546230793, + 0.00749568734318018, + 0.013273186050355434, + 0.006284971255809069, + 0.006630375515669584, + 0.007240840699523687, + 0.008561080321669579, + 0.009711893275380135, + 0.008556092157959938, + 0.008925363421440125, + 0.009655009023845196, + 0.010944082401692867, + 0.00868728756904602, + 0.008483326993882656, + 0.006072140298783779, + 0.015795260667800903, + 0.01785343512892723, + 0.010550334118306637, + 0.01301167905330658, + 0.019903378561139107, + 0.046464063227176666, + 0.026190219447016716, + 0.04775039851665497, + 0.0043031638488173485, + 0.0033761155791580677, + 0.0023275662679225206, + 0.006262839771807194, + 0.008758830837905407, + 0.014117544516921043, + 0.009386247023940086, + 0.01239703968167305, + 0.0044279806315898895, + 0.0023738404270261526, + 0.0025943380314856768, + 0.003210440743714571, + 0.0027773764450103045, + 0.005873135756701231, + 0.00479471730068326, + 0.004717171657830477, + 0.0019477307796478271, + 0.0018412560457363725, + 0.002959810197353363, + 0.005674449261277914, + 0.002767226891592145, + 0.010426260530948639, + 0.005860657896846533, + 0.008922438137233257, + 0.004527413286268711, + 0.004280879162251949, + 0.006107111927121878, + 0.00493237841874361, + 0.0023070406168699265, + 0.00930362194776535, + 0.006987220142036676, + 0.009732445701956749, + 0.008029122836887836, + 0.005112889222800732, + 0.0093314740806818, + 0.019757483154535294, + 0.005157100036740303, + 0.013359421864151955, + 0.015428553335368633, + 0.014890038408339024, + 0.01905432716012001, + 0.01932765357196331, + 0.024682387709617615, + 0.015414930880069733, + 0.023080289363861084, + 0.035127509385347366, + 0.019161591306328773, + 0.038003403693437576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.03809443116188049 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902648836374283 + ], + "opposite_logprob_tokens": [ + -2.46402645111084 + ], + "logprob_reward": -0.08902648836374283, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902648836374283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0601465217769146, + 0.08289897441864014, + 0.03589963540434837, + 0.5551424622535706, + 0.02008424699306488, + 0.02864188328385353, + 0.02361760288476944, + 0.01854913868010044, + 0.10400483757257462, + 0.0710146427154541 + ], + "credit_map": [ + 0.009658550843596458, + 0.009572659619152546, + 0.012204003520309925, + 0.0069483621045947075, + 0.005190221592783928, + 0.004749974701553583, + 0.00810373667627573, + 0.0037190115544945, + 0.01773362047970295, + 0.01744951494038105, + 0.00691318279132247, + 0.020098449662327766, + 0.004156041424721479, + 0.004077027086168528, + 0.005010178312659264, + 0.007460959255695343, + 0.004687775392085314, + 0.004582767374813557, + 0.0036908667534589767, + 0.00444002402946353, + 0.006254435516893864, + 0.005100239533931017, + 0.003941190429031849, + 0.0032023389358073473, + 0.10082044452428818, + 0.029129283502697945, + 0.020199313759803772, + 0.027580445632338524, + 0.041065990924835205, + 0.0877121314406395, + 0.1159520223736763, + 0.13268287479877472, + 0.002555632498115301, + 0.0012004541931673884, + 0.0008572670631110668, + 0.0012230491265654564, + 0.002711506560444832, + 0.00447643268853426, + 0.002864948008209467, + 0.004194958135485649, + 0.0026904880069196224, + 0.0018851453205570579, + 0.00237826956436038, + 0.0023190125357359648, + 0.006138043012470007, + 0.005218956619501114, + 0.0033514227252453566, + 0.004660547245293856, + 0.0014458625810220838, + 0.0019703705329447985, + 0.001405893242917955, + 0.002370368456467986, + 0.0013498347252607346, + 0.005995910614728928, + 0.0038498525973409414, + 0.00522951316088438, + 0.0011551403440535069, + 0.0015767901204526424, + 0.0019807578064501286, + 0.0016011272091418505, + 0.000996665097773075, + 0.002899145008996129, + 0.004382278770208359, + 0.003957235254347324, + 0.008921891450881958, + 0.0069837565533816814, + 0.009789667092263699, + 0.016766157001256943, + 0.0074152108281850815, + 0.015762561932206154, + 0.013228490948677063, + 0.025137105956673622, + 0.004538603592664003, + 0.006898616906255484, + 0.008185124956071377, + 0.006148026790469885, + 0.008241736330091953, + 0.013667144812643528, + 0.006794836837798357, + 0.016540559008717537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.08902648836374283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot at an extremely high frame rate, the high-speed macro camera footage begins with static, piercing focus, capturing the inner workings of a highly competitive gearbox against a deep, non-reflective black background. In profound slow motion, a massive, mirror-polished steel gear, its surface gleaming with brilliant mirror highlights, slams into position with silent, impactful force. Its perfectly machined teeth mesh with another with utter precision, the perfect interlock sending a powerful, resonant tremor through the entire assembly, causing it to shake and sway as if it were a ship pitching and tumbling on a rough sea. The impact violently launches droplets of golden, viscous oil into the air; these globular beads, suspended in the void, arc in elegant arcs, catching the harsh lights of a movie studio like tiny, glowing amber gems. Then, the camera's focus subtly tilts to reveal the intricate dance of the selector and fork, their matte gunmetal finish contrasting sharply with the polished gears as they move with swift, deliberate purpose. The entire scene is a surreal and primal spectacle, with the stark contrast between cool, gleaming steel, warm, glistening oil, and the absolute black background creating a captivating and tense atmosphere of controlled mechanical rage.", + "seed": 29678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_8194_seed_29678_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029360903333872555, + "vlm_logprob": -0.029360903333872555, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.029360903333872555, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015298171900212765 + ], + "opposite_logprob_tokens": [ + -4.890298366546631 + ], + "logprob_reward": -0.015298171900212765, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015298171900212765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0855337455868721, + 0.05655430629849434, + 0.45970529317855835, + 0.03926195204257965, + 0.06597133725881577, + 0.09047505259513855, + 0.06890451908111572, + 0.039006445556879044, + 0.039613813161849976, + 0.05497358739376068 + ], + "credit_map": [ + 0.0053597139194607735, + 0.013307287357747555, + 0.005786558613181114, + 0.012587441131472588, + 0.009116162545979023, + 0.011136768385767937, + 0.01619846187531948, + 0.012041348032653332, + 0.00446995347738266, + 0.005667457822710276, + 0.007359336595982313, + 0.004789749626070261, + 0.015777207911014557, + 0.007180882152169943, + 0.0039994786493480206, + 0.0073102423921227455, + 0.02183244749903679, + 0.009732694365084171, + 0.009891609661281109, + 0.00840364582836628, + 0.02018522284924984, + 0.19522739946842194, + 0.011725079268217087, + 0.18270717561244965, + 0.003471566829830408, + 0.0049506365321576595, + 0.00419225450605154, + 0.004770171828567982, + 0.005241790320724249, + 0.005072710104286671, + 0.005334577057510614, + 0.006228242069482803, + 0.014471483416855335, + 0.006781155243515968, + 0.005440734326839447, + 0.005983096547424793, + 0.010345766320824623, + 0.009598296135663986, + 0.006407049018889666, + 0.006943753454834223, + 0.014460770413279533, + 0.011303392238914967, + 0.00957088079303503, + 0.0060998848639428616, + 0.015534686855971813, + 0.012205179780721664, + 0.010444926097989082, + 0.010855334810912609, + 0.006938283797353506, + 0.00547023955732584, + 0.004147239029407501, + 0.006065412424504757, + 0.018060237169265747, + 0.004008214455097914, + 0.020458083599805832, + 0.00375680741854012, + 0.003546870080754161, + 0.004749959334731102, + 0.006423576734960079, + 0.0052738795056939125, + 0.004423336125910282, + 0.0035812268033623695, + 0.004856116604059935, + 0.006151482462882996, + 0.004785122349858284, + 0.002743886085227132, + 0.006182502023875713, + 0.003985222429037094, + 0.004708766471594572, + 0.005025964695960283, + 0.007550437934696674, + 0.004631910938769579, + 0.0036072349175810814, + 0.0034619062207639217, + 0.013673421926796436, + 0.008429566398262978, + 0.002653727540746331, + 0.005817139521241188, + 0.01449732854962349, + 0.002833261387422681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.015298171900212765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961731541901827 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961731541901827, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961731541901827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16834917664527893, + 0.08344874531030655, + 0.11527308076620102, + 0.08281327039003372, + 0.05349849537014961, + 0.22374355792999268, + 0.13257654011249542, + 0.04440655559301376, + 0.03127986565232277, + 0.0646107867360115 + ], + "credit_map": [ + 0.01393235195428133, + 0.027296191081404686, + 0.012918103486299515, + 0.029070094227790833, + 0.021564291790127754, + 0.019396262243390083, + 0.020754646509885788, + 0.023417219519615173, + 0.007843981496989727, + 0.008908743970096111, + 0.012460566125810146, + 0.015131453052163124, + 0.011011986993253231, + 0.010896404273808002, + 0.009115376509726048, + 0.008080226369202137, + 0.006990616675466299, + 0.008785146288573742, + 0.003996571525931358, + 0.0056868079118430614, + 0.006506210193037987, + 0.03662126138806343, + 0.007785257417708635, + 0.03890120983123779, + 0.007431633770465851, + 0.009748771786689758, + 0.008587168529629707, + 0.012542693875730038, + 0.011140861548483372, + 0.012650908902287483, + 0.009201315231621265, + 0.011509904637932777, + 0.006811727304011583, + 0.004359818529337645, + 0.004891771357506514, + 0.0041931793093681335, + 0.012794713489711285, + 0.006730412598699331, + 0.007658068090677261, + 0.0060588037595152855, + 0.03545157238841057, + 0.029936768114566803, + 0.03210343047976494, + 0.040775272995233536, + 0.018231328576803207, + 0.024373149499297142, + 0.022865476086735725, + 0.020006537437438965, + 0.012032878585159779, + 0.007214698474854231, + 0.006813953630626202, + 0.008199600502848625, + 0.038210224360227585, + 0.006127568893134594, + 0.046643443405628204, + 0.007334162015467882, + 0.005069178529083729, + 0.00803708378225565, + 0.004969283472746611, + 0.00616002082824707, + 0.005363155622035265, + 0.003834361210465431, + 0.0056516206823289394, + 0.005321844946593046, + 0.0048021008260548115, + 0.003283692756667733, + 0.004317468032240868, + 0.004436481278389692, + 0.0031435727141797543, + 0.0028225122950971127, + 0.004820397589355707, + 0.003653638530522585, + 0.004592653829604387, + 0.005206684581935406, + 0.010893573984503746, + 0.008891791105270386, + 0.005816514138132334, + 0.010321301408112049, + 0.01014018151909113, + 0.008748081512749195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.005961731541901827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290053009986877 + ], + "opposite_logprob_tokens": [ + -3.7732901573181152 + ], + "logprob_reward": -0.023290053009986877, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290053009986877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1513247936964035, + 0.15695500373840332, + 0.12695187330245972, + 0.04672440513968468, + 0.03230859339237213, + 0.05196285620331764, + 0.13298307359218597, + 0.05577666312456131, + 0.02647269517183304, + 0.21854010224342346 + ], + "credit_map": [ + 0.010066255927085876, + 0.02869197353720665, + 0.007383297197520733, + 0.018850022926926613, + 0.019657501950860023, + 0.024269860237836838, + 0.021273763850331306, + 0.02113211713731289, + 0.014601869508624077, + 0.01853214018046856, + 0.014031895436346531, + 0.01997014693915844, + 0.015402253717184067, + 0.03502305969595909, + 0.022213462740182877, + 0.01718016341328621, + 0.006620559375733137, + 0.008403538726270199, + 0.00592383136972785, + 0.006554892752319574, + 0.007308679167181253, + 0.04133421927690506, + 0.008239845745265484, + 0.04256631061434746, + 0.003983317408710718, + 0.0052960943430662155, + 0.006130855064839125, + 0.004862667061388493, + 0.005061920266598463, + 0.006398769095540047, + 0.0072725145146250725, + 0.007718272041529417, + 0.005768186878412962, + 0.0033105360344052315, + 0.0024292143061757088, + 0.005125297233462334, + 0.003194752149283886, + 0.004130937624722719, + 0.0036172624677419662, + 0.004732409492135048, + 0.007352150045335293, + 0.005314959213137627, + 0.005668148398399353, + 0.005590932909399271, + 0.008681127801537514, + 0.006770467851310968, + 0.005451873876154423, + 0.007133199367672205, + 0.010313538834452629, + 0.009373057633638382, + 0.007362876087427139, + 0.008554349653422832, + 0.04238629341125488, + 0.006592575926333666, + 0.0424911305308342, + 0.005909239873290062, + 0.005562532227486372, + 0.008493512868881226, + 0.006620369385927916, + 0.007074511609971523, + 0.007427260745316744, + 0.005893327295780182, + 0.006117241457104683, + 0.008587905205786228, + 0.003050470259040594, + 0.0030744269024580717, + 0.004643320571631193, + 0.0038524523843079805, + 0.002318593440577388, + 0.0018146486254408956, + 0.0050199273973703384, + 0.002698854310438037, + 0.012674808502197266, + 0.010451342910528183, + 0.05697823315858841, + 0.018688799813389778, + 0.021456623449921608, + 0.027584057301282883, + 0.05441741272807121, + 0.016288813203573227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.023290053009986877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283181712031364 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283181712031364, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283181712031364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13807463645935059, + 0.08057809621095657, + 0.13180653750896454, + 0.05521462857723236, + 0.06634388864040375, + 0.07716928422451019, + 0.10629484802484512, + 0.02534380555152893, + 0.03949751332402229, + 0.27967679500579834 + ], + "credit_map": [ + 0.01165688131004572, + 0.024825094267725945, + 0.008115026168525219, + 0.013807660900056362, + 0.01587556302547455, + 0.02259376458823681, + 0.02150830812752247, + 0.019692320376634598, + 0.007156569976359606, + 0.008783443830907345, + 0.010008234530687332, + 0.008113500662147999, + 0.021100347861647606, + 0.011822054162621498, + 0.005693699698895216, + 0.007900241762399673, + 0.006224071606993675, + 0.006391395349055529, + 0.006363371387124062, + 0.00616955803707242, + 0.006915200036019087, + 0.045519497245550156, + 0.007398056332021952, + 0.04682536795735359, + 0.005673418287187815, + 0.005908271297812462, + 0.006375435274094343, + 0.006413637660443783, + 0.007632056716829538, + 0.00825534202158451, + 0.006924637593328953, + 0.00803182739764452, + 0.007756588514894247, + 0.006598787847906351, + 0.0059874095022678375, + 0.007692794315516949, + 0.011107386089861393, + 0.007034469395875931, + 0.0122756352648139, + 0.007890811190009117, + 0.01251404918730259, + 0.010008725337684155, + 0.009527819231152534, + 0.01054614968597889, + 0.010356293059885502, + 0.009599341079592705, + 0.005109586287289858, + 0.00950730498880148, + 0.010220430791378021, + 0.0062695774249732494, + 0.005872852634638548, + 0.007179249078035355, + 0.030410699546337128, + 0.004937225021421909, + 0.0358709916472435, + 0.0055338130332529545, + 0.0031744337175041437, + 0.004859857261180878, + 0.0029163165017962456, + 0.004228842444717884, + 0.002774212509393692, + 0.002215719548985362, + 0.0024313030298799276, + 0.0027431189082562923, + 0.005058967042714357, + 0.003926598932594061, + 0.008795770816504955, + 0.005542169790714979, + 0.003233537543565035, + 0.003919337410479784, + 0.004553842358291149, + 0.00446728290989995, + 0.012671846896409988, + 0.010582209564745426, + 0.09103943407535553, + 0.028508245944976807, + 0.01587308570742607, + 0.019457470625638962, + 0.08766045421361923, + 0.013884040527045727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.023283181712031364 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897137850522995 + ], + "opposite_logprob_tokens": [ + -2.5789713859558105 + ], + "logprob_reward": -0.07897137850522995, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897137850522995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1748439222574234, + 0.04879032075405121, + 0.11562135070562363, + 0.1031603068113327, + 0.07271061837673187, + 0.08514628559350967, + 0.08261819183826447, + 0.03628023341298103, + 0.03978990018367767, + 0.24103885889053345 + ], + "credit_map": [ + 0.018553294241428375, + 0.029764942824840546, + 0.012393246404826641, + 0.020327499136328697, + 0.017314469441771507, + 0.022686492651700974, + 0.028230881318449974, + 0.025573087856173515, + 0.006052056327462196, + 0.00681948009878397, + 0.0063439300283789635, + 0.008488921448588371, + 0.00627492181956768, + 0.006475261878222227, + 0.004378489684313536, + 0.0039572580717504025, + 0.007606406696140766, + 0.012342363595962524, + 0.0062993373721838, + 0.0113743981346488, + 0.010898001492023468, + 0.03124321810901165, + 0.011565121822059155, + 0.024292510002851486, + 0.008838661946356297, + 0.010999730788171291, + 0.011897653341293335, + 0.011444831266999245, + 0.013944551348686218, + 0.013187520205974579, + 0.016412923112511635, + 0.01643444038927555, + 0.011594103649258614, + 0.004810987040400505, + 0.003952027298510075, + 0.004792514257133007, + 0.008496524766087532, + 0.010928014293313026, + 0.010398822836577892, + 0.017737625166773796, + 0.012294894084334373, + 0.01237531565129757, + 0.010152137838304043, + 0.00994900707155466, + 0.008792515844106674, + 0.011803454719483852, + 0.007512654177844524, + 0.012266308069229126, + 0.008836052380502224, + 0.006615990772843361, + 0.0046369777992367744, + 0.006464366801083088, + 0.022600499913096428, + 0.004114815965294838, + 0.024337617680430412, + 0.00501187052577734, + 0.0031426860950887203, + 0.0037964689545333385, + 0.005881559103727341, + 0.00462215393781662, + 0.004328271374106407, + 0.003042067401111126, + 0.004808133468031883, + 0.006658889353275299, + 0.0033922402653843164, + 0.003681469475850463, + 0.0061141992919147015, + 0.006175058428198099, + 0.003989869263023138, + 0.00455467076972127, + 0.007463438436388969, + 0.004418953321874142, + 0.011349983513355255, + 0.006627976894378662, + 0.0774245485663414, + 0.01989368535578251, + 0.013611949980258942, + 0.01817062869668007, + 0.07921451330184937, + 0.01474559586495161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07897137850522995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot at an extremely high frame rate, the high-speed macro camera footage begins with static, piercing focus, capturing the inner workings of a highly competitive gearbox against a deep, non-reflective black background. In profound slow motion, a massive, mirror-polished steel gear, its surface gleaming with brilliant mirror highlights, slams into position with silent, impactful force. Its perfectly machined teeth mesh with another with utter precision, the perfect interlock sending a powerful, resonant tremor through the entire assembly, causing it to shake and sway as if it were a ship pitching and tumbling on a rough sea. The impact violently launches droplets of golden, viscous oil into the air; these globular beads, suspended in the void, arc in elegant arcs, catching the harsh lights of a movie studio like tiny, glowing amber gems. Then, the camera's focus subtly tilts to reveal the intricate dance of the selector and fork, their matte gunmetal finish contrasting sharply with the polished gears as they move with swift, deliberate purpose. The entire scene is a surreal and primal spectacle, with the stark contrast between cool, gleaming steel, warm, glistening oil, and the absolute black background creating a captivating and tense atmosphere of controlled mechanical rage.", + "seed": 29679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_4/index_8194_seed_29679_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0330324157141149, + "vlm_logprob": -0.0330324157141149, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.0330324157141149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015198964625597 + ], + "opposite_logprob_tokens": [ + -4.515199184417725 + ], + "logprob_reward": -0.015198964625597, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015198964625597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2827991843223572, + 0.06623518466949463, + 0.16473031044006348, + 0.06001051515340805, + 0.05866147205233574, + 0.09389284998178482, + 0.11711230129003525, + 0.0999990701675415, + 0.026711730286478996, + 0.029847417026758194 + ], + "credit_map": [ + 0.02889139950275421, + 0.07032529264688492, + 0.07268644869327545, + 0.023288091644644737, + 0.01984766311943531, + 0.022865410894155502, + 0.022476589307188988, + 0.02241828851401806, + 0.009474419988691807, + 0.006717653013765812, + 0.007481148466467857, + 0.006409135181456804, + 0.009330379776656628, + 0.00872792862355709, + 0.007970348000526428, + 0.01012417022138834, + 0.026107540354132652, + 0.01586950197815895, + 0.023359594866633415, + 0.018662743270397186, + 0.016442440450191498, + 0.023771433159708977, + 0.025492748245596886, + 0.015024316497147083, + 0.018539922311902046, + 0.0030256356112658978, + 0.005163935944437981, + 0.0036447884049266577, + 0.004244272597134113, + 0.014150830917060375, + 0.0035418507177382708, + 0.007699276320636272, + 0.006929359398782253, + 0.004713356960564852, + 0.004934999626129866, + 0.010492897592484951, + 0.003788638161495328, + 0.005365491844713688, + 0.01537736039608717, + 0.00705936923623085, + 0.010819991119205952, + 0.006902307737618685, + 0.008078676648437977, + 0.013490027748048306, + 0.008192292414605618, + 0.014705678448081017, + 0.013332408852875233, + 0.01837146282196045, + 0.007685039658099413, + 0.00698377750813961, + 0.00867496058344841, + 0.009072380140423775, + 0.019083796069025993, + 0.023973653092980385, + 0.021297236904501915, + 0.020341448485851288, + 0.008412141352891922, + 0.005369268823415041, + 0.007830671966075897, + 0.009887265041470528, + 0.006686946377158165, + 0.024764925241470337, + 0.007866020314395428, + 0.029181839898228645, + 0.0034936454612761736, + 0.0037945774383842945, + 0.004886149428784847, + 0.002224474446848035, + 0.003781314706429839, + 0.0027307362761348486, + 0.002291938289999962, + 0.003508895169943571, + 0.004184623248875141, + 0.003800756298005581, + 0.0036411674227565527, + 0.0032151970081031322, + 0.0035812559071928263, + 0.0029120366089046, + 0.004049512092024088, + 0.004462868440896273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.015198964625597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003202668856829405 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.003202668856829405, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003202668856829405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27482250332832336, + 0.10021325200796127, + 0.25462260842323303, + 0.0513274259865284, + 0.04570930823683739, + 0.06703422218561172, + 0.06363396346569061, + 0.03881519287824631, + 0.04544343054294586, + 0.05837809666991234 + ], + "credit_map": [ + 0.03422534838318825, + 0.06662026047706604, + 0.0753161683678627, + 0.020263532176613808, + 0.022901320829987526, + 0.022141695022583008, + 0.0203225277364254, + 0.013031619600951672, + 0.010451355949044228, + 0.008485158905386925, + 0.020502954721450806, + 0.007445340510457754, + 0.00952134095132351, + 0.012766299769282341, + 0.021465091034770012, + 0.009575718082487583, + 0.040372587740421295, + 0.02395135909318924, + 0.02891317382454872, + 0.0314406119287014, + 0.028209945186972618, + 0.02809927985072136, + 0.047683898359537125, + 0.025951769202947617, + 0.009666850790381432, + 0.0037426839116960764, + 0.005650609731674194, + 0.006785516627132893, + 0.0041660768911242485, + 0.012158037163317204, + 0.0029704063199460506, + 0.006187244318425655, + 0.006851868703961372, + 0.003217513905838132, + 0.00442570960149169, + 0.00794361624866724, + 0.003678356297314167, + 0.00439961813390255, + 0.009177804924547672, + 0.006014818325638771, + 0.012294656597077847, + 0.0040975227020680904, + 0.006174567621201277, + 0.00874343328177929, + 0.008880484849214554, + 0.00840374082326889, + 0.009637169539928436, + 0.008802646771073341, + 0.012102882377803326, + 0.0053987568244338036, + 0.003950732294470072, + 0.004898624029010534, + 0.005467880982905626, + 0.011187825351953506, + 0.010158213786780834, + 0.010469047352671623, + 0.0029195700772106647, + 0.0030217936728149652, + 0.006513915490359068, + 0.004868717398494482, + 0.0019974864553660154, + 0.008097236044704914, + 0.0028043093625456095, + 0.008592162281274796, + 0.0055931140668690205, + 0.005871410947293043, + 0.009125773794949055, + 0.004221704322844744, + 0.005651105660945177, + 0.004622795153409243, + 0.005188601091504097, + 0.0051689245738089085, + 0.00968372542411089, + 0.006010361947119236, + 0.007597160059958696, + 0.006135488394647837, + 0.0059852697886526585, + 0.00564139150083065, + 0.005601848941296339, + 0.011722849681973457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.003202668856829405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043060317635536194 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.043060317635536194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043060317635536194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35768571496009827, + 0.09107954800128937, + 0.24697166681289673, + 0.04689931496977806, + 0.020560353994369507, + 0.05920877307653427, + 0.05999275669455528, + 0.03151152282953262, + 0.03136444464325905, + 0.05472591146826744 + ], + "credit_map": [ + 0.028850773349404335, + 0.0996304526925087, + 0.08220001310110092, + 0.029779013246297836, + 0.03270753473043442, + 0.028688792139291763, + 0.03033445030450821, + 0.025494668632745743, + 0.0049421945586800575, + 0.006604993715882301, + 0.015670912340283394, + 0.007119004148989916, + 0.0074867126531898975, + 0.012906946241855621, + 0.02103946916759014, + 0.015309317037463188, + 0.03163059055805206, + 0.01970529556274414, + 0.03276273235678673, + 0.029489563778042793, + 0.026163585484027863, + 0.02726925164461136, + 0.058833491057157516, + 0.021117165684700012, + 0.012099171057343483, + 0.00281926104798913, + 0.006951052695512772, + 0.0028631906025111675, + 0.003819370176643133, + 0.011998257599771023, + 0.002466165227815509, + 0.0038828460965305567, + 0.0020578240510076284, + 0.00179184228181839, + 0.0028828936628997326, + 0.005263229366391897, + 0.001505562337115407, + 0.0019380067242309451, + 0.0027371342293918133, + 0.002383859595283866, + 0.00396376010030508, + 0.00591270811855793, + 0.004726157058030367, + 0.010031997226178646, + 0.008120662532746792, + 0.007791003677994013, + 0.008996043354272842, + 0.009666439145803452, + 0.004869991913437843, + 0.004832250531762838, + 0.005897860042750835, + 0.006167794577777386, + 0.006373301614075899, + 0.011520405299961567, + 0.009451242163777351, + 0.010879912413656712, + 0.004134883172810078, + 0.0020092143677175045, + 0.0026012412272393703, + 0.0030438320245593786, + 0.0044466182589530945, + 0.005317592993378639, + 0.00277705118060112, + 0.007181089837104082, + 0.00431026192381978, + 0.0037880470044910908, + 0.007121448405086994, + 0.0032543421257287264, + 0.00386746390722692, + 0.0024451902136206627, + 0.0030139281880110502, + 0.0035637603141367435, + 0.009032747708261013, + 0.0057396311312913895, + 0.006830190308392048, + 0.0059704966843128204, + 0.0056159901432693005, + 0.0052019087597727776, + 0.004655920434743166, + 0.01167902909219265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.043060317635536194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033691760152578354 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.033691760152578354, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033691760152578354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0766790360212326, + 0.04734308645129204, + 0.3396165072917938, + 0.06607300788164139, + 0.06658373773097992, + 0.04632188379764557, + 0.050173286348581314, + 0.05929334834218025, + 0.1077142283320427, + 0.14020183682441711 + ], + "credit_map": [ + 0.008294600993394852, + 0.01928231492638588, + 0.018527476117014885, + 0.007072461768984795, + 0.0056737554259598255, + 0.006514830980449915, + 0.0063974191434681416, + 0.0049161831848323345, + 0.004735765978693962, + 0.005618702154606581, + 0.005193660967051983, + 0.004181487485766411, + 0.006190114188939333, + 0.005434915889054537, + 0.005601591896265745, + 0.010386854410171509, + 0.04110393673181534, + 0.02614641562104225, + 0.04425380378961563, + 0.03961365669965744, + 0.03731909394264221, + 0.04103836789727211, + 0.07983733713626862, + 0.030303923413157463, + 0.017135687172412872, + 0.004212178755551577, + 0.004182109143584967, + 0.005188754294067621, + 0.006651866249740124, + 0.020483776926994324, + 0.0032732165418565273, + 0.004945424851030111, + 0.010311998426914215, + 0.006374143064022064, + 0.007636218797415495, + 0.007719428278505802, + 0.005028735846281052, + 0.006081456784158945, + 0.012218098156154156, + 0.011213664896786213, + 0.0069730691611766815, + 0.0034188355784863234, + 0.0029759611934423447, + 0.006588297430425882, + 0.006488846614956856, + 0.007000561337918043, + 0.006971622351557016, + 0.005904694087803364, + 0.005627778824418783, + 0.0036653857678174973, + 0.0053272489458322525, + 0.0047140480019152164, + 0.0037428787909448147, + 0.0085708973929286, + 0.009703307412564754, + 0.008821745403110981, + 0.006029979325830936, + 0.004345266614109278, + 0.005277174524962902, + 0.0035780451726168394, + 0.008247526362538338, + 0.011077825911343098, + 0.0074329497292637825, + 0.013304583728313446, + 0.015893051400780678, + 0.014209320768713951, + 0.01766422763466835, + 0.011937873438000679, + 0.011600538156926632, + 0.009338044561445713, + 0.012836924754083157, + 0.014234263449907303, + 0.020520400255918503, + 0.017676418647170067, + 0.021497534587979317, + 0.015321153216063976, + 0.014361623674631119, + 0.011807537637650967, + 0.013995328918099403, + 0.02502184733748436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.033691760152578354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro close-up of a gearbox assembly with a deep, non-reflective black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, mirror-polished steel gear visible with brilliant mirror highlights and perfectly machined teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden, viscous oil droplets appear as globular beads arcing in the air after the gear impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a selector and fork visible with a matte gunmetal finish contrasting with the polished gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus subtly tilt to reveal the selector and fork during the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000836730003357 + ], + "opposite_logprob_tokens": [ + -2.6950082778930664 + ], + "logprob_reward": -0.07000836730003357, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000836730003357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05089382827281952, + 0.10187801718711853, + 0.04079756885766983, + 0.08187674731016159, + 0.08694440126419067, + 0.17857258021831512, + 0.1381513625383377, + 0.07404110580682755, + 0.09315172582864761, + 0.15369266271591187 + ], + "credit_map": [ + 0.0052474383264780045, + 0.007801673840731382, + 0.004729490727186203, + 0.006887913215905428, + 0.006595219019800425, + 0.006754640955477953, + 0.005778241902589798, + 0.007099210284650326, + 0.00981686171144247, + 0.0073013389483094215, + 0.02322133630514145, + 0.008267956785857677, + 0.008432266302406788, + 0.012635567225515842, + 0.01970340684056282, + 0.012499283999204636, + 0.005894937086850405, + 0.004209475591778755, + 0.0043522207997739315, + 0.00458542350679636, + 0.005877918563783169, + 0.004495073109865189, + 0.0072538480162620544, + 0.004128671251237392, + 0.02069244533777237, + 0.0041224220767617226, + 0.011622472666203976, + 0.006214808207005262, + 0.005011874716728926, + 0.01283013541251421, + 0.008136848919093609, + 0.013245745562016964, + 0.010215111076831818, + 0.008127009496092796, + 0.008082527667284012, + 0.01209288090467453, + 0.006727113388478756, + 0.010521139949560165, + 0.019310802221298218, + 0.011867811903357506, + 0.019778858870267868, + 0.010887600481510162, + 0.013343448750674725, + 0.03643650934100151, + 0.020767610520124435, + 0.026942968368530273, + 0.023293273523449898, + 0.02712230384349823, + 0.01324159000068903, + 0.014089591801166534, + 0.010688128881156445, + 0.01235126331448555, + 0.015114104375243187, + 0.02562727779150009, + 0.02355671487748623, + 0.02348269149661064, + 0.009658490307629108, + 0.0055795880034565926, + 0.010394199751317501, + 0.00958480779081583, + 0.010545812547206879, + 0.010393446311354637, + 0.0076520745642483234, + 0.010232685133814812, + 0.013580475002527237, + 0.012256857939064503, + 0.013311211951076984, + 0.009720348753035069, + 0.008973713964223862, + 0.010606189258396626, + 0.011922262609004974, + 0.012780665419995785, + 0.023900650441646576, + 0.02027999609708786, + 0.020176252350211143, + 0.017220990732312202, + 0.016447922214865685, + 0.015963038429617882, + 0.013008852489292622, + 0.026694968342781067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.07000836730003357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_tensor([4557, 4557])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_tensor([4557, 4557])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f841d546a3dd7d42c2bedf45aee3c0d7b12800c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_tensor([4557, 4557])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up shot from a slightly overhead angle ushers the viewer into a scene imbued with melancholic chiaroscuro. A strong sidelight cuts through the gloom from the left side of the frame, precisely sculpting the chiseled figure of a secret agent bent over a large wooden table. His strong jawline and prominent nose are accentuated by the light, while the other side of his face is completely lost in deep shadow. He wears a matte, dark, high-collared tactical shirt, leaning forward in a posture that betrays intense concentration. A fine-tipped pen, carefully and steadily tracing a new route on an intricate, yellowed map, the deep black ink leaving a clear mark on the paper. Within this piercing beam of light, countless dust particles, captured by slow-motion particle effects, dance and flicker like suspended stars, adding a sense of movement and poetry to the otherwise tranquil scene. The background is mired in deep shadows, with only vague outlines visible. All visual focus is on him, his focused movements, and the crucial map, creating an extremely tense, lonely, and contemplative atmosphere.", + "seed": 14274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_4557_seed_14274_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03736529555171728, + "vlm_logprob": -0.03736529555171728, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.03736529555171728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04620629921555519 + ], + "opposite_logprob_tokens": [ + -4.421206474304199 + ], + "logprob_reward": -0.04620629921555519, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.04620629921555519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16018585860729218, + 0.06751405447721481, + 0.10627036541700363, + 0.18853606283664703, + 0.049400538206100464, + 0.07204601168632507, + 0.08906084299087524, + 0.1393308788537979, + 0.0716041699051857, + 0.05605122819542885 + ], + "credit_map": [ + 0.0181008018553257, + 0.023148972541093826, + 0.012555813416838646, + 0.011759602464735508, + 0.027905991300940514, + 0.024975823238492012, + 0.02012399025261402, + 0.02161485329270363, + 0.007519465871155262, + 0.00863443873822689, + 0.00588120287284255, + 0.005831640679389238, + 0.011642402969300747, + 0.008569516241550446, + 0.012120435014367104, + 0.007314941380172968, + 0.01591743342578411, + 0.016821371391415596, + 0.01168927550315857, + 0.008145279251039028, + 0.013303844258189201, + 0.015568528324365616, + 0.012622365728020668, + 0.012202262878417969, + 0.026951413601636887, + 0.04447992146015167, + 0.016141219064593315, + 0.017254289239645004, + 0.017701271921396255, + 0.027670584619045258, + 0.017958691343665123, + 0.02037867158651352, + 0.006333955563604832, + 0.007181144319474697, + 0.007622808218002319, + 0.005682350602000952, + 0.005402123089879751, + 0.006139622535556555, + 0.00537699181586504, + 0.005661538802087307, + 0.009223122149705887, + 0.009702063165605068, + 0.011917270720005035, + 0.00807673018425703, + 0.00870566163212061, + 0.00981433317065239, + 0.006335627753287554, + 0.008271208964288235, + 0.012313309125602245, + 0.016136271879076958, + 0.007786015048623085, + 0.010598591528832912, + 0.012071887031197548, + 0.012276386842131615, + 0.009019081480801105, + 0.008859299123287201, + 0.013327976688742638, + 0.029817404225468636, + 0.013202237896621227, + 0.017557788640260696, + 0.01651601865887642, + 0.023677978664636612, + 0.007569546345621347, + 0.017661921679973602, + 0.010019189678132534, + 0.010950482450425625, + 0.0069410065189003944, + 0.009159683249890804, + 0.010169613175094128, + 0.010587155818939209, + 0.007140395697206259, + 0.006636640522629023, + 0.005236565135419369, + 0.006524553522467613, + 0.0075491187162697315, + 0.007298629730939865, + 0.005781836342066526, + 0.00853564403951168, + 0.008960427716374397, + 0.006164450664073229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.04620629921555519 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006838917732239 + ], + "opposite_logprob_tokens": [ + -2.695068359375 + ], + "logprob_reward": -0.07006838917732239, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006838917732239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26360687613487244, + 0.06683797389268875, + 0.04018651321530342, + 0.1887979656457901, + 0.22206008434295654, + 0.058851733803749084, + 0.05072643607854843, + 0.048824846744537354, + 0.03322840481996536, + 0.026879196986556053 + ], + "credit_map": [ + 0.02858969010412693, + 0.03530765697360039, + 0.022685298696160316, + 0.019398784264922142, + 0.05261792615056038, + 0.04291246831417084, + 0.03351358696818352, + 0.02858147770166397, + 0.0077646407298743725, + 0.00927199050784111, + 0.005944405682384968, + 0.0047427029348909855, + 0.01488745678216219, + 0.010493859648704529, + 0.007302473299205303, + 0.006430438719689846, + 0.005071607418358326, + 0.005949868820607662, + 0.004887870047241449, + 0.0028512931894510984, + 0.006147995591163635, + 0.005461009684950113, + 0.005836874712258577, + 0.003979993984103203, + 0.016120951622724533, + 0.03724455460906029, + 0.012250966392457485, + 0.018398467451334, + 0.031137879937887192, + 0.028351543471217155, + 0.02402001991868019, + 0.02127358689904213, + 0.01481832005083561, + 0.023308727890253067, + 0.019706396386027336, + 0.019217900931835175, + 0.05107872560620308, + 0.03316293656826019, + 0.03336010128259659, + 0.02740699239075184, + 0.008052819408476353, + 0.006862059701234102, + 0.005513983778655529, + 0.006929484196007252, + 0.008387509733438492, + 0.008995242416858673, + 0.006481225602328777, + 0.007629408501088619, + 0.0047057573683559895, + 0.00819335039705038, + 0.0035253800451755524, + 0.004583082161843777, + 0.011473173275589943, + 0.007548144087195396, + 0.0056900326162576675, + 0.005007514264434576, + 0.0028201057575643063, + 0.007555629592388868, + 0.004289850126951933, + 0.005218941252678633, + 0.01045931875705719, + 0.007409124169498682, + 0.005980545189231634, + 0.005091328639537096, + 0.0043877591378986835, + 0.005677719600498676, + 0.002700410084798932, + 0.003146170638501644, + 0.00533349858596921, + 0.00413126964122057, + 0.004334178287535906, + 0.0035174007061868906, + 0.002454963745549321, + 0.003368722042068839, + 0.0033922751899808645, + 0.002673305105417967, + 0.0038738586008548737, + 0.0036357506178319454, + 0.004854022990912199, + 0.0026262986939400434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.07006838917732239 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814607113599777 + ], + "opposite_logprob_tokens": [ + -3.2881460189819336 + ], + "logprob_reward": -0.03814607113599777, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814607113599777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1707397848367691, + 0.09913235157728195, + 0.0447116419672966, + 0.2550536096096039, + 0.2626985013484955, + 0.03741907328367233, + 0.04752524197101593, + 0.0287657268345356, + 0.0235469788312912, + 0.03040706366300583 + ], + "credit_map": [ + 0.03885927051305771, + 0.028275761753320694, + 0.01221716869622469, + 0.011774387210607529, + 0.01744116097688675, + 0.023026684299111366, + 0.019519463181495667, + 0.019625891000032425, + 0.011491152457892895, + 0.013969660736620426, + 0.007116318214684725, + 0.006970459129661322, + 0.016315853223204613, + 0.015099622309207916, + 0.015331997536122799, + 0.012837288901209831, + 0.00499410554766655, + 0.006569467484951019, + 0.004228765610605478, + 0.0034310556948184967, + 0.007145848590880632, + 0.006382194813340902, + 0.00754007650539279, + 0.004420125856995583, + 0.03842467814683914, + 0.06017113849520683, + 0.021324649453163147, + 0.024014804512262344, + 0.021326016634702682, + 0.03779067471623421, + 0.02432304434478283, + 0.027678607031702995, + 0.017967872321605682, + 0.02537042833864689, + 0.029599100351333618, + 0.024913093075156212, + 0.058304984122514725, + 0.03825769200921059, + 0.03905053436756134, + 0.029234785586595535, + 0.003624935168772936, + 0.005466419737786055, + 0.0031115557067096233, + 0.0037290723994374275, + 0.007463890593498945, + 0.0049958983436226845, + 0.005502446088939905, + 0.003524856874719262, + 0.004227231256663799, + 0.005539820063859224, + 0.003980747889727354, + 0.004336917772889137, + 0.012065495364367962, + 0.0071174209006130695, + 0.005506024230271578, + 0.004751585889607668, + 0.003493889234960079, + 0.0042473990470170975, + 0.0028532049618661404, + 0.003089179517701268, + 0.004420636221766472, + 0.004255901090800762, + 0.0032742777839303017, + 0.0031312373466789722, + 0.0027941670268774033, + 0.0028787069022655487, + 0.001936569926328957, + 0.002849607029929757, + 0.0035053964238613844, + 0.0031168730929493904, + 0.00397717859596014, + 0.0024884797167032957, + 0.00256377924233675, + 0.004624160006642342, + 0.003959181718528271, + 0.003075903980061412, + 0.003788143629208207, + 0.004353791009634733, + 0.004905493929982185, + 0.003136611543595791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03814607113599777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420846302062273 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.01420846302062273, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420846302062273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19464023411273956, + 0.11547186225652695, + 0.07079194486141205, + 0.09999118745326996, + 0.0812033861875534, + 0.09633450210094452, + 0.09817814826965332, + 0.12602952122688293, + 0.061505142599344254, + 0.05585404112935066 + ], + "credit_map": [ + 0.01897307299077511, + 0.025109270587563515, + 0.020757421851158142, + 0.01416328176856041, + 0.03287186473608017, + 0.028075311332941055, + 0.029224209487438202, + 0.02546580322086811, + 0.02102673426270485, + 0.015023392625153065, + 0.0114490557461977, + 0.007917063310742378, + 0.018868006765842438, + 0.015486021526157856, + 0.013686860911548138, + 0.012014728970825672, + 0.010832400992512703, + 0.010561911389231682, + 0.008936543017625809, + 0.005567073822021484, + 0.008600953966379166, + 0.008853867650032043, + 0.009108619764447212, + 0.008330577053129673, + 0.014248497784137726, + 0.02324477769434452, + 0.006855110637843609, + 0.010108371265232563, + 0.010228498838841915, + 0.01630866527557373, + 0.009329098276793957, + 0.00966816395521164, + 0.011368360370397568, + 0.009767827577888966, + 0.016815748065710068, + 0.010883602313697338, + 0.006021210923790932, + 0.010262515395879745, + 0.007814566604793072, + 0.008269554935395718, + 0.00761071452870965, + 0.012825037352740765, + 0.010828816331923008, + 0.006963606458157301, + 0.02050912193953991, + 0.0134784746915102, + 0.014140529558062553, + 0.009978203102946281, + 0.009801525622606277, + 0.019988149404525757, + 0.008134854957461357, + 0.01195797324180603, + 0.016817212104797363, + 0.013069575652480125, + 0.010075535625219345, + 0.008333326317369938, + 0.010754208080470562, + 0.024145670235157013, + 0.013554531149566174, + 0.013807961717247963, + 0.017025841400027275, + 0.021038105711340904, + 0.01078032050281763, + 0.014922878704965115, + 0.007525450550019741, + 0.01078127883374691, + 0.007424261420965195, + 0.010099001228809357, + 0.0059655336663126945, + 0.008234784938395023, + 0.006091509014368057, + 0.005383324343711138, + 0.006754376459866762, + 0.01109023205935955, + 0.006533487234264612, + 0.006285327486693859, + 0.005979208275675774, + 0.007920583710074425, + 0.005723288282752037, + 0.005567539948970079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01420846302062273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197255209088326 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197255209088326, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197255209088326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2902356684207916, + 0.11891219019889832, + 0.05887560173869133, + 0.11185073852539062, + 0.15858173370361328, + 0.05455326661467552, + 0.0643107071518898, + 0.071855328977108, + 0.03790546581149101, + 0.03291924297809601 + ], + "credit_map": [ + 0.030510667711496353, + 0.04749194160103798, + 0.030258305370807648, + 0.021255793049931526, + 0.035049919039011, + 0.05097781866788864, + 0.033089764416217804, + 0.0416015088558197, + 0.017173901200294495, + 0.02236950770020485, + 0.011544629000127316, + 0.006960202008485794, + 0.014170941896736622, + 0.02082984149456024, + 0.01182890497148037, + 0.014034278690814972, + 0.007768016308546066, + 0.012246277183294296, + 0.005094029009342194, + 0.0043946499936282635, + 0.006540891248732805, + 0.008457825519144535, + 0.007578829303383827, + 0.006795089691877365, + 0.010945837944746017, + 0.013062756508588791, + 0.006346175447106361, + 0.010491857305169106, + 0.018910681828856468, + 0.021316012367606163, + 0.015201471745967865, + 0.0155759546905756, + 0.01827569119632244, + 0.015570223331451416, + 0.029285728931427002, + 0.015409968793392181, + 0.02864333800971508, + 0.019147340208292007, + 0.01661369763314724, + 0.01563575305044651, + 0.004576893523335457, + 0.008052929304540157, + 0.006317080929875374, + 0.008047632873058319, + 0.006444739177823067, + 0.00784470234066248, + 0.006134852766990662, + 0.007134437095373869, + 0.010671825148165226, + 0.015957506373524666, + 0.0057259793393313885, + 0.008012015372514725, + 0.004289610777050257, + 0.008700251579284668, + 0.005001273937523365, + 0.005952250212430954, + 0.006134222727268934, + 0.013096483424305916, + 0.005140575580298901, + 0.008902573958039284, + 0.013659365475177765, + 0.009762187488377094, + 0.007549504283815622, + 0.0076104262843728065, + 0.0038966990541666746, + 0.004674633964896202, + 0.003774108598008752, + 0.004163058008998632, + 0.007165452465415001, + 0.005274904891848564, + 0.005200248211622238, + 0.003756366204470396, + 0.0036679324693977833, + 0.005938101559877396, + 0.003527278546243906, + 0.0031008878722786903, + 0.0058511244133114815, + 0.004172279965132475, + 0.0035291630774736404, + 0.0031324781011790037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.018197255209088326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot from a slightly overhead angle ushers the viewer into a scene imbued with melancholic chiaroscuro. A strong sidelight cuts through the gloom from the left side of the frame, precisely sculpting the chiseled figure of a secret agent bent over a large wooden table. His strong jawline and prominent nose are accentuated by the light, while the other side of his face is completely lost in deep shadow. He wears a matte, dark, high-collared tactical shirt, leaning forward in a posture that betrays intense concentration. A fine-tipped pen, carefully and steadily tracing a new route on an intricate, yellowed map, the deep black ink leaving a clear mark on the paper. Within this piercing beam of light, countless dust particles, captured by slow-motion particle effects, dance and flicker like suspended stars, adding a sense of movement and poetry to the otherwise tranquil scene. The background is mired in deep shadows, with only vague outlines visible. All visual focus is on him, his focused movements, and the crucial map, creating an extremely tense, lonely, and contemplative atmosphere.", + "seed": 14275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_4557_seed_14275_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06623426768928767, + "vlm_logprob": -0.06623426768928767, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06623426768928767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07992136478424072 + ], + "opposite_logprob_tokens": [ + -2.829921245574951 + ], + "logprob_reward": -0.07992136478424072, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07992136478424072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11308106034994125, + 0.06417699158191681, + 0.10550545901060104, + 0.23395568132400513, + 0.16158059239387512, + 0.07334132492542267, + 0.0682111531496048, + 0.08854137361049652, + 0.02666347660124302, + 0.06494292616844177 + ], + "credit_map": [ + 0.013830321840941906, + 0.011678643524646759, + 0.01215195283293724, + 0.01540529727935791, + 0.014466162770986557, + 0.015310132876038551, + 0.01288793794810772, + 0.01735059730708599, + 0.005789791699498892, + 0.0047293491661548615, + 0.007492391858249903, + 0.0058564734645187855, + 0.008550722151994705, + 0.013064512982964516, + 0.006672198418527842, + 0.012021541595458984, + 0.006218609865754843, + 0.005719826556742191, + 0.0069610136561095715, + 0.006517954170703888, + 0.007125684525817633, + 0.0333857461810112, + 0.007841253653168678, + 0.03173535689711571, + 0.005969037301838398, + 0.004643207415938377, + 0.006535931024700403, + 0.005314132198691368, + 0.007817494682967663, + 0.09845536947250366, + 0.008555606938898563, + 0.09666486084461212, + 0.003810918191447854, + 0.0031905199866741896, + 0.004737141076475382, + 0.002659580437466502, + 0.00449998676776886, + 0.07008425891399384, + 0.004397224169224501, + 0.0682009607553482, + 0.005995181854814291, + 0.005144214723259211, + 0.004117135889828205, + 0.003630773164331913, + 0.004777336493134499, + 0.022103222087025642, + 0.0062775276601314545, + 0.021295925602316856, + 0.006146223749965429, + 0.004935919307172298, + 0.00547462934628129, + 0.005611298140138388, + 0.00595397362485528, + 0.01780012622475624, + 0.004536490421742201, + 0.01775248907506466, + 0.007153310347348452, + 0.006176271475851536, + 0.006203142926096916, + 0.006129808723926544, + 0.007480614352971315, + 0.02434336207807064, + 0.007056352682411671, + 0.023998495191335678, + 0.002297190483659506, + 0.0024268648121505976, + 0.001929630758240819, + 0.00243941112421453, + 0.002504371339455247, + 0.006171162240207195, + 0.0026372019201517105, + 0.006257642526179552, + 0.0064541045576334, + 0.006645700894296169, + 0.007026114501059055, + 0.005927353631705046, + 0.006843322888016701, + 0.011743701994419098, + 0.007405312266200781, + 0.012897316366434097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.07992136478424072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994850873947144 + ], + "opposite_logprob_tokens": [ + -1.8049485683441162 + ], + "logprob_reward": -0.17994850873947144, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994850873947144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1190374568104744, + 0.07911436259746552, + 0.22596728801727295, + 0.3407032787799835, + 0.03887981176376343, + 0.09970571845769882, + 0.009161676280200481, + 0.03975387290120125, + 0.006847199983894825, + 0.04082939773797989 + ], + "credit_map": [ + 0.010537274181842804, + 0.01420508325099945, + 0.009759319014847279, + 0.012734219431877136, + 0.011556652374565601, + 0.0246003195643425, + 0.011808122508227825, + 0.02383645996451378, + 0.00406706053763628, + 0.0037648340221494436, + 0.00375976855866611, + 0.003913120366632938, + 0.005024858750402927, + 0.027854060754179955, + 0.00414624810218811, + 0.026584399864077568, + 0.008950985036790371, + 0.012399012222886086, + 0.013892772607505322, + 0.007728223223239183, + 0.010818480513989925, + 0.08210468292236328, + 0.012041048146784306, + 0.07803205400705338, + 0.008402802981436253, + 0.011206093244254589, + 0.011358138173818588, + 0.010296552442014217, + 0.01045663096010685, + 0.14116667211055756, + 0.00852342788130045, + 0.13929294049739838, + 0.0031230002641677856, + 0.002597292186692357, + 0.004262606613337994, + 0.002996820257976651, + 0.0029107246082276106, + 0.010151737369596958, + 0.0028528349939733744, + 0.00998479314148426, + 0.00483312364667654, + 0.004703798331320286, + 0.004430138971656561, + 0.004877607338130474, + 0.00442791823297739, + 0.0351426936686039, + 0.005971395410597324, + 0.03531903028488159, + 0.0009359767427667975, + 0.0009179366170428693, + 0.0009760019020177424, + 0.0011169776553288102, + 0.0011724168434739113, + 0.0015636790776625276, + 0.0009105896460823715, + 0.0015680971555411816, + 0.00327687687240541, + 0.0025688910391181707, + 0.003145018359646201, + 0.00271528004668653, + 0.0034903883934020996, + 0.01060960628092289, + 0.003317308146506548, + 0.010630500502884388, + 0.0005195006378926337, + 0.0006541142938658595, + 0.0006383721483871341, + 0.0005182183813303709, + 0.000791964412201196, + 0.001414065482094884, + 0.000739561568479985, + 0.0015714026521891356, + 0.0034336999524384737, + 0.0030649309046566486, + 0.0036723283119499683, + 0.003075912594795227, + 0.0031894349958747625, + 0.01034505758434534, + 0.0034486076328903437, + 0.01059942040592432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.17994850873947144 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815203532576561 + ], + "opposite_logprob_tokens": [ + -3.288151979446411 + ], + "logprob_reward": -0.03815203532576561, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815203532576561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14509464800357819, + 0.08152133971452713, + 0.12841516733169556, + 0.2896532714366913, + 0.07142889499664307, + 0.037165913730859756, + 0.014777418226003647, + 0.0413101464509964, + 0.0328315831720829, + 0.15780159831047058 + ], + "credit_map": [ + 0.012232491746544838, + 0.01796180196106434, + 0.017427289858460426, + 0.016151387244462967, + 0.014830393716692924, + 0.020926615223288536, + 0.021436631679534912, + 0.024128025397658348, + 0.007310078479349613, + 0.00840379111468792, + 0.00657670758664608, + 0.005580184515565634, + 0.007079756818711758, + 0.01703532412648201, + 0.010822483338415623, + 0.01871301233768463, + 0.009781081229448318, + 0.014565936289727688, + 0.013459041714668274, + 0.010521948337554932, + 0.014230438508093357, + 0.023990845307707787, + 0.01832718960940838, + 0.023538686335086823, + 0.007046045269817114, + 0.0065494440495967865, + 0.008840614929795265, + 0.0061454023234546185, + 0.006026162300258875, + 0.12605439126491547, + 0.005643743090331554, + 0.12334748357534409, + 0.002653785515576601, + 0.0028416914865374565, + 0.0023777580354362726, + 0.002484129508957267, + 0.002642096485942602, + 0.02777833119034767, + 0.003337452420964837, + 0.02731364406645298, + 0.0023164667654782534, + 0.0017299106111750007, + 0.0018803277052938938, + 0.0019760499708354473, + 0.0018765508430078626, + 0.012905428186058998, + 0.0013902670470997691, + 0.01309091504663229, + 0.0012351329205557704, + 0.001971073215827346, + 0.0018681843066588044, + 0.0019568626303225756, + 0.0015980794560164213, + 0.002138108015060425, + 0.0019030612893402576, + 0.002106916392222047, + 0.003948923200368881, + 0.004214507527649403, + 0.003673762548714876, + 0.003421712201088667, + 0.004985272418707609, + 0.008580464869737625, + 0.004652382805943489, + 0.007833119481801987, + 0.0023914927151054144, + 0.0023351882118731737, + 0.0022151644807308912, + 0.0021568748634308577, + 0.003444952191784978, + 0.008825726807117462, + 0.002404833910986781, + 0.009057353250682354, + 0.016950255259871483, + 0.015789788216352463, + 0.016786988824605942, + 0.014149519614875317, + 0.014719877392053604, + 0.0290980264544487, + 0.01818564534187317, + 0.03212149441242218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.03815203532576561 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599007606506348 + ], + "opposite_logprob_tokens": [ + -3.895598888397217 + ], + "logprob_reward": -0.020599007606506348, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599007606506348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19548986852169037, + 0.047056641429662704, + 0.301033616065979, + 0.167864590883255, + 0.052383072674274445, + 0.031109852716326714, + 0.014940724708139896, + 0.13556525111198425, + 0.02178296260535717, + 0.03277336806058884 + ], + "credit_map": [ + 0.01591322012245655, + 0.014366637915372849, + 0.013687810860574245, + 0.018908819183707237, + 0.014518406242132187, + 0.051943086087703705, + 0.01673799194395542, + 0.049413904547691345, + 0.0034580202773213387, + 0.003632640466094017, + 0.0024509946815669537, + 0.0026214004028588533, + 0.00671660341322422, + 0.012477549724280834, + 0.004321272950619459, + 0.011378161609172821, + 0.012640377506613731, + 0.016160404309630394, + 0.0115973437204957, + 0.013670998625457287, + 0.014109903015196323, + 0.10755757242441177, + 0.020796585828065872, + 0.1045004352927208, + 0.010167860426008701, + 0.009654026478528976, + 0.0068207704462111, + 0.00653741555288434, + 0.011705949902534485, + 0.05832284688949585, + 0.008310324512422085, + 0.05634540691971779, + 0.003941991366446018, + 0.005412290338426828, + 0.00415407121181488, + 0.003799173515290022, + 0.003507563378661871, + 0.014157626777887344, + 0.004100201651453972, + 0.013310160487890244, + 0.002632207004353404, + 0.0024025733582675457, + 0.0028991622384637594, + 0.002591117052361369, + 0.0029211125802248716, + 0.007185312919318676, + 0.003481913125142455, + 0.006996457930654287, + 0.0016720205312594771, + 0.001770113711245358, + 0.001280511962249875, + 0.0013176088687032461, + 0.001750629162415862, + 0.0029492126777768135, + 0.0010118496138602495, + 0.0031887798104435205, + 0.011115621775388718, + 0.01246163435280323, + 0.011480998247861862, + 0.010260169394314289, + 0.011424691416323185, + 0.03416617214679718, + 0.011324475519359112, + 0.033331498503685, + 0.0017901374958455563, + 0.0016267845639958978, + 0.0015212955186143517, + 0.0013670880580320954, + 0.001823684317059815, + 0.005794718861579895, + 0.0019452619599178433, + 0.005913993343710899, + 0.002325234701856971, + 0.002852173987776041, + 0.0024146006908267736, + 0.0023990601766854525, + 0.0025991990696638823, + 0.008556867018342018, + 0.0031593050807714462, + 0.00846693105995655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.020599007606506348 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550421990454197 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550421990454197, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550421990454197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11314578354358673, + 0.045247480273246765, + 0.07880149781703949, + 0.10664117336273193, + 0.29367199540138245, + 0.2518725097179413, + 0.039076272398233414, + 0.029163936153054237, + 0.008232863619923592, + 0.034146521240472794 + ], + "credit_map": [ + 0.01169463898986578, + 0.010530381463468075, + 0.011096775531768799, + 0.012414018623530865, + 0.009995647706091404, + 0.021891187876462936, + 0.01196413766592741, + 0.02355898916721344, + 0.0033844683784991503, + 0.002803449984639883, + 0.0029149630572646856, + 0.002094127470627427, + 0.004024025984108448, + 0.013946202583611012, + 0.0031581150833517313, + 0.012922130525112152, + 0.006259537301957607, + 0.0064310310408473015, + 0.0044526816345751286, + 0.006840741727501154, + 0.008158497512340546, + 0.01959860324859619, + 0.006847221404314041, + 0.020213183015584946, + 0.0032751711551100016, + 0.002920065075159073, + 0.0035485378466546535, + 0.0028286499436944723, + 0.0030619727913290262, + 0.044715434312820435, + 0.002685361076146364, + 0.0436059832572937, + 0.016077524051070213, + 0.016535822302103043, + 0.016294557601213455, + 0.01347307674586773, + 0.014847125858068466, + 0.10183454304933548, + 0.016188113018870354, + 0.09842123091220856, + 0.009681344032287598, + 0.005693034268915653, + 0.008136231452226639, + 0.006731242872774601, + 0.00734757399186492, + 0.10554441064596176, + 0.006920472718775272, + 0.10181819647550583, + 0.0031179841607809067, + 0.003100659465417266, + 0.003338288050144911, + 0.003582892706617713, + 0.003742908826097846, + 0.009858980774879456, + 0.0028664779383689165, + 0.009468082338571548, + 0.003984129521995783, + 0.0036096712574362755, + 0.0036118256393820047, + 0.0034228642471134663, + 0.0028408814687281847, + 0.004127116408199072, + 0.00285161891952157, + 0.004715828690677881, + 0.0012046884512528777, + 0.0008412178140133619, + 0.0009496753918938339, + 0.0006529124802909791, + 0.0007465234957635403, + 0.0015773489139974117, + 0.0007458031177520752, + 0.0015146943042054772, + 0.002099714009091258, + 0.0019083484075963497, + 0.0021216801833361387, + 0.0017024658154696226, + 0.0019869995303452015, + 0.011094451881945133, + 0.0023136092349886894, + 0.01091925147920847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.012550421990454197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot from a slightly overhead angle ushers the viewer into a scene imbued with melancholic chiaroscuro. A strong sidelight cuts through the gloom from the left side of the frame, precisely sculpting the chiseled figure of a secret agent bent over a large wooden table. His strong jawline and prominent nose are accentuated by the light, while the other side of his face is completely lost in deep shadow. He wears a matte, dark, high-collared tactical shirt, leaning forward in a posture that betrays intense concentration. A fine-tipped pen, carefully and steadily tracing a new route on an intricate, yellowed map, the deep black ink leaving a clear mark on the paper. Within this piercing beam of light, countless dust particles, captured by slow-motion particle effects, dance and flicker like suspended stars, adding a sense of movement and poetry to the otherwise tranquil scene. The background is mired in deep shadows, with only vague outlines visible. All visual focus is on him, his focused movements, and the crucial map, creating an extremely tense, lonely, and contemplative atmosphere.", + "seed": 14276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_4557_seed_14276_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03949498198926449, + "vlm_logprob": -0.03949498198926449, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.03949498198926449, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04793735221028328 + ], + "opposite_logprob_tokens": [ + -3.5479373931884766 + ], + "logprob_reward": -0.04793735221028328, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04793735221028328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20887760818004608, + 0.08146262168884277, + 0.06744755804538727, + 0.06020507961511612, + 0.04803259298205376, + 0.10505159944295883, + 0.049575746059417725, + 0.04404476657509804, + 0.05392186716198921, + 0.2813805639743805 + ], + "credit_map": [ + 0.016126485541462898, + 0.014005765318870544, + 0.03936709836125374, + 0.03375351056456566, + 0.02982316166162491, + 0.029352987185120583, + 0.025458596646785736, + 0.02099001593887806, + 0.00886489823460579, + 0.006057782098650932, + 0.011262351647019386, + 0.011787769384682178, + 0.012586221098899841, + 0.010461066849529743, + 0.011123137548565865, + 0.009319395758211613, + 0.004730287007987499, + 0.00451199384406209, + 0.010332279838621616, + 0.010834837332367897, + 0.00851193442940712, + 0.00998410489410162, + 0.008220439776778221, + 0.010321682319045067, + 0.004284171853214502, + 0.0032001750078052282, + 0.012611578218638897, + 0.010865437798202038, + 0.006113150622695684, + 0.013053659349679947, + 0.005251978058367968, + 0.004824926145374775, + 0.004062149208039045, + 0.004163051955401897, + 0.008199910633265972, + 0.006553074344992638, + 0.007106701843440533, + 0.008215019479393959, + 0.004706056322902441, + 0.005026630125939846, + 0.007809062488377094, + 0.006508032791316509, + 0.02327357605099678, + 0.016699640080332756, + 0.012513036839663982, + 0.021885255351662636, + 0.009867552667856216, + 0.0064954436384141445, + 0.0034581120125949383, + 0.003916493616998196, + 0.008478322997689247, + 0.007555490825325251, + 0.009845609776675701, + 0.011346418410539627, + 0.0027540037408471107, + 0.002221293980255723, + 0.0043372493237257, + 0.0032031729351729155, + 0.007166103459894657, + 0.006129661574959755, + 0.007882697507739067, + 0.00871890876442194, + 0.002753890585154295, + 0.003853084985166788, + 0.003778004553169012, + 0.002609868301078677, + 0.014760125428438187, + 0.005529478657990694, + 0.007221046835184097, + 0.013958471827208996, + 0.003419896587729454, + 0.002644971013069153, + 0.01176860835403204, + 0.008329911157488823, + 0.09688173979520798, + 0.02029373124241829, + 0.016845349222421646, + 0.10469594597816467, + 0.010474126785993576, + 0.01209114771336317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.04793735221028328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005671411752701 + ], + "opposite_logprob_tokens": [ + -2.695056676864624 + ], + "logprob_reward": -0.07005671411752701, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005671411752701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32646963000297546, + 0.11152257770299911, + 0.08933917433023453, + 0.04599637910723686, + 0.12197232991456985, + 0.08138716965913773, + 0.02743390202522278, + 0.04032241925597191, + 0.021084584295749664, + 0.1344718188047409 + ], + "credit_map": [ + 0.027416471391916275, + 0.020372910425066948, + 0.05502249300479889, + 0.04195399582386017, + 0.04459843412041664, + 0.044850174337625504, + 0.04637744650244713, + 0.045877713710069656, + 0.010442452505230904, + 0.008495437912642956, + 0.019613027572631836, + 0.016164643689990044, + 0.01515931822359562, + 0.014714838936924934, + 0.013906893320381641, + 0.01302596740424633, + 0.007295897230505943, + 0.009286279790103436, + 0.014755596406757832, + 0.012694560922682285, + 0.016305699944496155, + 0.015668408945202827, + 0.006038694176822901, + 0.007294042967259884, + 0.0031952352728694677, + 0.003088306402787566, + 0.008353170938789845, + 0.006694576237350702, + 0.005417650565505028, + 0.009143744595348835, + 0.004677222575992346, + 0.005426471587270498, + 0.008420796133577824, + 0.011182003654539585, + 0.020317010581493378, + 0.01741340197622776, + 0.015271943062543869, + 0.024516671895980835, + 0.010894968174397945, + 0.013955540023744106, + 0.005695325788110495, + 0.00605763727799058, + 0.01758783496916294, + 0.013032578863203526, + 0.008835677988827229, + 0.0171451885253191, + 0.00621101725846529, + 0.006821910385042429, + 0.0021193986758589745, + 0.00203260057605803, + 0.004436240065842867, + 0.0032938215881586075, + 0.005426246672868729, + 0.005696231499314308, + 0.0018989446107298136, + 0.0025304195005446672, + 0.00462960172444582, + 0.003747164737433195, + 0.007054267916828394, + 0.004227815195918083, + 0.006850576493889093, + 0.007720353547483683, + 0.0025959608610719442, + 0.0034966804087162018, + 0.0011401640949770808, + 0.000941504433285445, + 0.006375840399414301, + 0.0020187394693493843, + 0.0019458528840914369, + 0.0061742039397358894, + 0.0014273881679400802, + 0.0010608875891193748, + 0.004898684564977884, + 0.003458683378994465, + 0.04872467368841171, + 0.008530999533832073, + 0.006800817791372538, + 0.05255325883626938, + 0.004413509741425514, + 0.005091184284538031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.07005671411752701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822273179888725 + ], + "opposite_logprob_tokens": [ + -3.52982234954834 + ], + "logprob_reward": -0.029822273179888725, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822273179888725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2384326159954071, + 0.14858277142047882, + 0.13064005970954895, + 0.06836409866809845, + 0.0357765294611454, + 0.1823415905237198, + 0.0746692642569542, + 0.04583146795630455, + 0.021638615056872368, + 0.053723014891147614 + ], + "credit_map": [ + 0.023860415443778038, + 0.017643628641963005, + 0.037776388227939606, + 0.03383740410208702, + 0.0336863249540329, + 0.03732926771044731, + 0.021872935816645622, + 0.03242626041173935, + 0.01629733480513096, + 0.010953646153211594, + 0.021264810115098953, + 0.020032783970236778, + 0.020354410633444786, + 0.019504489377141, + 0.020990250632166862, + 0.019185034558176994, + 0.009464165195822716, + 0.006959886290132999, + 0.01956634223461151, + 0.01940985396504402, + 0.014703203924000263, + 0.01815837249159813, + 0.017048705369234085, + 0.025329528376460075, + 0.005003586411476135, + 0.005708174780011177, + 0.012256047688424587, + 0.011390209197998047, + 0.010479713790118694, + 0.012092864140868187, + 0.005126163829118013, + 0.006307334639132023, + 0.0026378557085990906, + 0.0023353740107268095, + 0.006104145664721727, + 0.004694514442235231, + 0.006268686149269342, + 0.006692490074783564, + 0.0031083186622709036, + 0.003935143817216158, + 0.013751916587352753, + 0.013921594247221947, + 0.0380219928920269, + 0.028763126581907272, + 0.019667712971568108, + 0.03835369274020195, + 0.014100815169513226, + 0.015760721638798714, + 0.005254839546978474, + 0.006389893125742674, + 0.01258073654025793, + 0.010172500275075436, + 0.014769394882023335, + 0.015633605420589447, + 0.004846404772251844, + 0.00502188503742218, + 0.003353297244757414, + 0.004451372195035219, + 0.007257129065692425, + 0.006761061027646065, + 0.008289716206490993, + 0.006174370180815458, + 0.004543623421341181, + 0.005000896751880646, + 0.001737502054311335, + 0.0016862187767401338, + 0.004253139719367027, + 0.002049266127869487, + 0.002885733963921666, + 0.005164640489965677, + 0.0018324138363823295, + 0.0020297011360526085, + 0.0026709011290222406, + 0.0015458345878869295, + 0.018434766680002213, + 0.0032233467791229486, + 0.004667150788009167, + 0.01945490390062332, + 0.0016038190806284547, + 0.0021222911309450865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029822273179888725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359083130955696 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026359083130955696, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359083130955696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26726239919662476, + 0.20626790821552277, + 0.08516155183315277, + 0.13520726561546326, + 0.02929314784705639, + 0.054964929819107056, + 0.03126833215355873, + 0.038918450474739075, + 0.02796526998281479, + 0.12369077652692795 + ], + "credit_map": [ + 0.01837141253054142, + 0.013019281439483166, + 0.05205903947353363, + 0.04597994312644005, + 0.03076670505106449, + 0.031206253916025162, + 0.04314109683036804, + 0.03271866962313652, + 0.018741121515631676, + 0.009941982105374336, + 0.029823243618011475, + 0.031119586899876595, + 0.025927666574716568, + 0.027593020349740982, + 0.03541654720902443, + 0.027704723179340363, + 0.005922675132751465, + 0.006091110408306122, + 0.011307850480079651, + 0.010089089162647724, + 0.01175349485129118, + 0.010467207990586758, + 0.01049030851572752, + 0.019039807841181755, + 0.01058485358953476, + 0.008781581185758114, + 0.01968577690422535, + 0.020833628252148628, + 0.017131948843598366, + 0.026859693229198456, + 0.01504520047456026, + 0.01628456637263298, + 0.0025585240218788385, + 0.0022208276204764843, + 0.004997221287339926, + 0.005105871241539717, + 0.0037854029797017574, + 0.004791140556335449, + 0.0029525854624807835, + 0.002881573513150215, + 0.004547757096588612, + 0.004093723371624947, + 0.011529718525707722, + 0.0084793446585536, + 0.006950851529836655, + 0.01014814805239439, + 0.005225513130426407, + 0.003989872056990862, + 0.0024168454110622406, + 0.0021552983671426773, + 0.0055114468559622765, + 0.004312430042773485, + 0.005661431234329939, + 0.006391502451151609, + 0.002437649294734001, + 0.002381725236773491, + 0.0029808308463543653, + 0.0032799136824905872, + 0.006615881808102131, + 0.005456697661429644, + 0.006582633126527071, + 0.0057927826419472694, + 0.003711099037900567, + 0.004498612601310015, + 0.002194046974182129, + 0.0022415241692215204, + 0.0059870160184800625, + 0.003748965682461858, + 0.003933651372790337, + 0.005584916565567255, + 0.0019229972967877984, + 0.002352149225771427, + 0.004057806916534901, + 0.002722936449572444, + 0.045909054577350616, + 0.0075128329917788506, + 0.005845983047038317, + 0.04926362261176109, + 0.0040183174423873425, + 0.004360207822173834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.026359083130955696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299487307667732 + ], + "opposite_logprob_tokens": [ + -3.7732994556427 + ], + "logprob_reward": -0.023299487307667732, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299487307667732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23877763748168945, + 0.17305268347263336, + 0.10083826631307602, + 0.08464143425226212, + 0.06713330000638962, + 0.19642485678195953, + 0.02699347399175167, + 0.027206579223275185, + 0.036934543401002884, + 0.04799724370241165 + ], + "credit_map": [ + 0.024155547842383385, + 0.014946093782782555, + 0.044482965022325516, + 0.041326723992824554, + 0.023303793743252754, + 0.0305339302867651, + 0.03392314165830612, + 0.026105448603630066, + 0.014317560940980911, + 0.010241548530757427, + 0.02432861737906933, + 0.024270467460155487, + 0.02420014888048172, + 0.023861486464738846, + 0.021308807656168938, + 0.03052404895424843, + 0.009124012663960457, + 0.008028106763958931, + 0.015822838991880417, + 0.0168477650731802, + 0.013356945477426052, + 0.014480029232800007, + 0.011528920382261276, + 0.011649653315544128, + 0.006468418054282665, + 0.0060745724476873875, + 0.013391518034040928, + 0.013055791147053242, + 0.011881367303431034, + 0.015125521458685398, + 0.008998140692710876, + 0.009646103717386723, + 0.00485382042825222, + 0.005460331682115793, + 0.01090353261679411, + 0.010809206403791904, + 0.010336863808333874, + 0.01140984334051609, + 0.005976036190986633, + 0.00738366786390543, + 0.013199897482991219, + 0.01384138036519289, + 0.04309260845184326, + 0.032117269933223724, + 0.021278705447912216, + 0.04341203346848488, + 0.014638669788837433, + 0.014844287186861038, + 0.0024488167837262154, + 0.0030169435776770115, + 0.004853785503655672, + 0.005459499545395374, + 0.003737516701221466, + 0.004840774927288294, + 0.0012538579758256674, + 0.0013822794426232576, + 0.003587186336517334, + 0.0022550439462065697, + 0.004351743031293154, + 0.004625814501196146, + 0.003123414469882846, + 0.004670070018619299, + 0.002412008587270975, + 0.0021812948398292065, + 0.0025894197169691324, + 0.001895489520393312, + 0.010328950360417366, + 0.004128753673285246, + 0.005101880989968777, + 0.008948557078838348, + 0.0022924894001334906, + 0.0016490034759044647, + 0.0030772890895605087, + 0.0017203770112246275, + 0.015945078805088997, + 0.0029897133354097605, + 0.004490641877055168, + 0.015580225735902786, + 0.0019426748622208834, + 0.002251239726319909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.023299487307667732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot from a slightly overhead angle ushers the viewer into a scene imbued with melancholic chiaroscuro. A strong sidelight cuts through the gloom from the left side of the frame, precisely sculpting the chiseled figure of a secret agent bent over a large wooden table. His strong jawline and prominent nose are accentuated by the light, while the other side of his face is completely lost in deep shadow. He wears a matte, dark, high-collared tactical shirt, leaning forward in a posture that betrays intense concentration. A fine-tipped pen, carefully and steadily tracing a new route on an intricate, yellowed map, the deep black ink leaving a clear mark on the paper. Within this piercing beam of light, countless dust particles, captured by slow-motion particle effects, dance and flicker like suspended stars, adding a sense of movement and poetry to the otherwise tranquil scene. The background is mired in deep shadows, with only vague outlines visible. All visual focus is on him, his focused movements, and the crucial map, creating an extremely tense, lonely, and contemplative atmosphere.", + "seed": 14277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_5/index_4557_seed_14277_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04435851909220219, + "vlm_logprob": -0.04435851909220219, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04435851909220219, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05979020148515701 + ], + "opposite_logprob_tokens": [ + -3.4347901344299316 + ], + "logprob_reward": -0.05979020148515701, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.05979020148515701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2889726758003235, + 0.1797475963830948, + 0.11328091472387314, + 0.06402575224637985, + 0.0990375429391861, + 0.04495001584291458, + 0.07175908237695694, + 0.06339140981435776, + 0.040648482739925385, + 0.034186579287052155 + ], + "credit_map": [ + 0.021306801587343216, + 0.0367618203163147, + 0.015330596826970577, + 0.021359635517001152, + 0.05372392386198044, + 0.05928150191903114, + 0.04829847440123558, + 0.032909881323575974, + 0.01201728917658329, + 0.020215466618537903, + 0.012352445162832737, + 0.01612284407019615, + 0.032612789422273636, + 0.03204897791147232, + 0.025960396975278854, + 0.028417358174920082, + 0.00634269742295146, + 0.01129705086350441, + 0.00805922131985426, + 0.008456770330667496, + 0.022688254714012146, + 0.022306205704808235, + 0.01701091229915619, + 0.017119789496064186, + 0.005338361952453852, + 0.006902068853378296, + 0.003365796525031328, + 0.004670608788728714, + 0.013731390237808228, + 0.014703012071549892, + 0.007540701422840357, + 0.007773802615702152, + 0.007532298099249601, + 0.009444798342883587, + 0.0045108976773917675, + 0.006445286329835653, + 0.020295416936278343, + 0.023460932075977325, + 0.013826022855937481, + 0.013521886430680752, + 0.005261405371129513, + 0.004262483213096857, + 0.003218653379008174, + 0.004629270173609257, + 0.007972655817866325, + 0.007372220512479544, + 0.0061662690714001656, + 0.006067054346203804, + 0.006717946380376816, + 0.0077242678962647915, + 0.005101601127535105, + 0.005969710182398558, + 0.013196366839110851, + 0.01694674603641033, + 0.007211802992969751, + 0.008890626952052116, + 0.004894263111054897, + 0.00745387701317668, + 0.003741326043382287, + 0.0040036765858531, + 0.014826704747974873, + 0.012568757869303226, + 0.008176608942449093, + 0.007726193871349096, + 0.004794846288859844, + 0.004836866166442633, + 0.0024398064706474543, + 0.0032863314263522625, + 0.008984928019344807, + 0.007251559291034937, + 0.004409885499626398, + 0.004644255619496107, + 0.0028010664973407984, + 0.00423899432644248, + 0.003168676979839802, + 0.0022986719850450754, + 0.007237398996949196, + 0.005947575438767672, + 0.004580792970955372, + 0.003913398366421461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05979020148515701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890696719288826 + ], + "opposite_logprob_tokens": [ + -2.4640696048736572 + ], + "logprob_reward": -0.0890696719288826, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890696719288826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3301480710506439, + 0.1112554669380188, + 0.24351051449775696, + 0.030728386715054512, + 0.0745629072189331, + 0.038286276161670685, + 0.0697387233376503, + 0.02253442443907261, + 0.06174096465110779, + 0.01749430224299431 + ], + "credit_map": [ + 0.026629574596881866, + 0.03746943920850754, + 0.01700774021446705, + 0.026715610176324844, + 0.07348684966564178, + 0.06542448699474335, + 0.05319199711084366, + 0.03022236004471779, + 0.008092663250863552, + 0.013489638455212116, + 0.00983983650803566, + 0.010184667073190212, + 0.021457673981785774, + 0.02004706673324108, + 0.015400872565805912, + 0.012743032537400723, + 0.017010938376188278, + 0.03476447984576225, + 0.016241412609815598, + 0.02322063408792019, + 0.05206100642681122, + 0.04241349548101425, + 0.03024236299097538, + 0.027556154876947403, + 0.0027961968444287777, + 0.00448957271873951, + 0.002829846693202853, + 0.0027008051984012127, + 0.0053848205134272575, + 0.005592505447566509, + 0.0033223459031432867, + 0.0036122880410403013, + 0.006448543630540371, + 0.01353215891867876, + 0.005258864723145962, + 0.005390759091824293, + 0.014077694155275822, + 0.01361637469381094, + 0.008331161923706532, + 0.007907342165708542, + 0.0027832386549562216, + 0.004551491234451532, + 0.0022449521347880363, + 0.002571702003479004, + 0.008412701077759266, + 0.008466708473861217, + 0.004320724867284298, + 0.0049347528256475925, + 0.0048615154810249805, + 0.005162259563803673, + 0.004282223526388407, + 0.00570337288081646, + 0.012225832790136337, + 0.014737756922841072, + 0.013754901476204395, + 0.009010854177176952, + 0.0017832060111686587, + 0.002853174228221178, + 0.0016154360491782427, + 0.002067988272756338, + 0.003445687238126993, + 0.004414096940308809, + 0.003302306402474642, + 0.0030525277834385633, + 0.004850463476032019, + 0.007676674984395504, + 0.00522486912086606, + 0.004427796229720116, + 0.013701888732612133, + 0.012832858599722385, + 0.007086619734764099, + 0.0059397853910923, + 0.0018545216880738735, + 0.001629668870009482, + 0.0012043040478602052, + 0.0013201335677877069, + 0.003845109837129712, + 0.003222147235646844, + 0.0019294054945930839, + 0.0024890101049095392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0890696719288826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038132645189762115 + ], + "opposite_logprob_tokens": [ + -3.288132667541504 + ], + "logprob_reward": -0.038132645189762115, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038132645189762115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36075764894485474, + 0.18384255468845367, + 0.13330799341201782, + 0.03896369785070419, + 0.06468579173088074, + 0.08788453042507172, + 0.027658939361572266, + 0.04081357643008232, + 0.037124428898096085, + 0.024960745126008987 + ], + "credit_map": [ + 0.02554362267255783, + 0.053495556116104126, + 0.02274966426193714, + 0.03112194500863552, + 0.07443039119243622, + 0.07137073576450348, + 0.04319552704691887, + 0.03885022923350334, + 0.013278246857225895, + 0.015675725415349007, + 0.010777839459478855, + 0.016313383355736732, + 0.03668885678052902, + 0.0382629856467247, + 0.022853342816233635, + 0.02999218925833702, + 0.00908083189278841, + 0.013632482849061489, + 0.0073339082300662994, + 0.010881447233259678, + 0.02891058661043644, + 0.026993699371814728, + 0.015988031402230263, + 0.020487023517489433, + 0.002897548256441951, + 0.0036521859001368284, + 0.003367674071341753, + 0.003566938918083906, + 0.008225961588323116, + 0.009602461941540241, + 0.003769812872633338, + 0.0038811180274933577, + 0.005165382754057646, + 0.008304591290652752, + 0.0046201725490391254, + 0.0048314849846065044, + 0.013160514645278454, + 0.013864242471754551, + 0.008224408142268658, + 0.0065150014124810696, + 0.006078551057726145, + 0.006249609403312206, + 0.004655471071600914, + 0.007773230783641338, + 0.02121495082974434, + 0.02030714973807335, + 0.009265908971428871, + 0.012339667417109013, + 0.0013499943306669593, + 0.002142959972843528, + 0.0018005043966695666, + 0.00248331343755126, + 0.004165024030953646, + 0.0060288794338703156, + 0.0062441155314445496, + 0.0034441519528627396, + 0.0032465681433677673, + 0.005007528699934483, + 0.002913526725023985, + 0.00318766082637012, + 0.006452209781855345, + 0.007385958917438984, + 0.007704840507358313, + 0.00491529144346714, + 0.003167601302266121, + 0.005518182180821896, + 0.0033654526341706514, + 0.0026877636555582285, + 0.007920569740235806, + 0.006405894178897142, + 0.004629543982446194, + 0.0034294261131435633, + 0.0024044448509812355, + 0.0025718596298247576, + 0.002192461397498846, + 0.0024241511709988117, + 0.004788629710674286, + 0.004403791390359402, + 0.0024950450751930475, + 0.003680364927276969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.038132645189762115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597372204065323 + ], + "opposite_logprob_tokens": [ + -3.895597457885742 + ], + "logprob_reward": -0.020597372204065323, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597372204065323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26465556025505066, + 0.24466310441493988, + 0.06609761714935303, + 0.0591595284640789, + 0.1271268129348755, + 0.041322607547044754, + 0.03156716749072075, + 0.04743218049407005, + 0.06254152953624725, + 0.055433906614780426 + ], + "credit_map": [ + 0.027593152597546577, + 0.0367949903011322, + 0.018394574522972107, + 0.02804388850927353, + 0.03816148638725281, + 0.053563788533210754, + 0.02466406300663948, + 0.037439607083797455, + 0.02008611522614956, + 0.028352031484246254, + 0.018248068168759346, + 0.02703552320599556, + 0.04848999157547951, + 0.04474342241883278, + 0.03247275948524475, + 0.02523517794907093, + 0.004312296863645315, + 0.008251912891864777, + 0.007476343773305416, + 0.006211509462445974, + 0.012943500652909279, + 0.011535289697349072, + 0.007574362680315971, + 0.007792395073920488, + 0.0035457145422697067, + 0.009073702618479729, + 0.004755231551826, + 0.006035160273313522, + 0.010539921931922436, + 0.010725043714046478, + 0.007036635652184486, + 0.007448114920407534, + 0.013735492713749409, + 0.012723427265882492, + 0.010460460558533669, + 0.012157252989709377, + 0.02192899025976658, + 0.023152420297265053, + 0.02041103132069111, + 0.012557720765471458, + 0.003023359226062894, + 0.0059815687127411366, + 0.0035090085584670305, + 0.003210955997928977, + 0.007717235945165157, + 0.008364581502974033, + 0.00493358401581645, + 0.004582309164106846, + 0.0023227334022521973, + 0.0026639210991561413, + 0.002202533883973956, + 0.002470464911311865, + 0.004921365529298782, + 0.006753284018486738, + 0.006557561457157135, + 0.0036753013264387846, + 0.00556154502555728, + 0.007260520942509174, + 0.004478468094021082, + 0.004797553177922964, + 0.007284671999514103, + 0.00747890118509531, + 0.006011053919792175, + 0.00455946009606123, + 0.005971728824079037, + 0.007076244801282883, + 0.005050092935562134, + 0.0052792406640946865, + 0.014079844579100609, + 0.010732177644968033, + 0.007754875347018242, + 0.006597323343157768, + 0.0054345992393791676, + 0.004746891092509031, + 0.0046867188066244125, + 0.0050583346746861935, + 0.012763317674398422, + 0.008761673234403133, + 0.00757667375728488, + 0.006405695807188749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.020597372204065323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a secret agent with a strong jawline and prominent nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the agent wearing a matte, dark, high-collared tactical shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine-tipped pen tracing a route on an intricate, yellowed map with black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a strong sidelight from the left creating deep shadows on the right side of the face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202704653143883 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202704653143883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202704653143883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.325903981924057, + 0.18683241307735443, + 0.08281875401735306, + 0.11977997422218323, + 0.06732665002346039, + 0.07312148064374924, + 0.032798729836940765, + 0.05041491612792015, + 0.03266308456659317, + 0.028340071439743042 + ], + "credit_map": [ + 0.030925847589969635, + 0.04542125016450882, + 0.023495949804782867, + 0.02666269987821579, + 0.05929016321897507, + 0.057591553777456284, + 0.039255961775779724, + 0.04326056316494942, + 0.008861495181918144, + 0.02012927271425724, + 0.007101371884346008, + 0.013756479136645794, + 0.03848717734217644, + 0.034242935478687286, + 0.027568742632865906, + 0.036684922873973846, + 0.006185835227370262, + 0.010518905706703663, + 0.006357358768582344, + 0.009412799030542374, + 0.01701345108449459, + 0.014490287750959396, + 0.009651754051446915, + 0.009188351221382618, + 0.009860964491963387, + 0.017060447484254837, + 0.005760215688496828, + 0.009517032653093338, + 0.022675588726997375, + 0.02294762432575226, + 0.013996992260217667, + 0.01796109974384308, + 0.006154167465865612, + 0.009741040877997875, + 0.0043621547520160675, + 0.005057988688349724, + 0.011482433415949345, + 0.012911398895084858, + 0.007942060939967632, + 0.009675395675003529, + 0.006388136185705662, + 0.011641076765954494, + 0.004127847030758858, + 0.005473596043884754, + 0.014036889187991619, + 0.013321705162525177, + 0.008594202809035778, + 0.009538021869957447, + 0.0030248917173594236, + 0.003629253478720784, + 0.0022758985869586468, + 0.0025666342116892338, + 0.005298970732837915, + 0.006017081439495087, + 0.0054442379623651505, + 0.004541758447885513, + 0.005534588824957609, + 0.005860509816557169, + 0.004181451629847288, + 0.004316122271120548, + 0.009236111305654049, + 0.009163862094283104, + 0.006845930591225624, + 0.005276333075016737, + 0.002918617334216833, + 0.005548074375838041, + 0.002337573328986764, + 0.0023616550024598837, + 0.006672095041722059, + 0.005892659537494183, + 0.0036463874857872725, + 0.0032860219944268465, + 0.002194484230130911, + 0.003320341231301427, + 0.002001754008233547, + 0.0023358804173767567, + 0.004661672282963991, + 0.005114263854920864, + 0.004529598169028759, + 0.004182074218988419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014202704653143883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_tensor([5372, 5372])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_tensor([5372, 5372])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b402c16086258aef42a716b3ba2134e63dad7e6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_tensor([5372, 5372])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a low angle, the camera slowly tilting upward, begins with the image of a hand clad in a soft black leather glove. Fingers grip the ornate crossguard of a longsword, crafted from a dull silver material and intricately carved with arabesques. The hand and hilt are ensconced in the deep shadows of the ancient library, illuminated only by the warm, flickering light of a candle from outside the frame, casting soft golden highlights on the metal and leather. With a remarkably steady and deliberate movement, the hand pulls upward, slowly and gently, the sword is drawn from its dark wooden scabbard, inlaid with polished metalwork, inch by inch, accompanied only by the subtle hiss of steel against the lining. As the razor-sharp, mirror-polished blade gleamed with a cold light, it perfectly captured and distorted the reflection of its surroundings: rows of magnificent mahogany bookshelves reaching the ceiling, filled with thick, leather-bound ancient books, the faded gold-stamped text on the spines flowing as the blade moved. The entire scene was filled with a solemn, suspenseful, and historically tranquil atmosphere.", + "seed": 21861, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_5372_seed_21861_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030759974475950004, + "vlm_logprob": -0.030759974475950004, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.030759974475950004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06916327774524689 + ], + "opposite_logprob_tokens": [ + -3.1941633224487305 + ], + "logprob_reward": -0.06916327774524689, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06916327774524689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20789580047130585, + 0.21733397245407104, + 0.09507320076227188, + 0.18598031997680664, + 0.040373750030994415, + 0.0905236005783081, + 0.025832099840044975, + 0.04577193036675453, + 0.053475599735975266, + 0.03773971274495125 + ], + "credit_map": [ + 0.02081655152142048, + 0.024970417842268944, + 0.02450665645301342, + 0.03161068633198738, + 0.025964993983507156, + 0.024405445903539658, + 0.028492528945207596, + 0.02712852507829666, + 0.01818668842315674, + 0.02148793637752533, + 0.029073268175125122, + 0.04791515693068504, + 0.026406140998005867, + 0.021553387865424156, + 0.03209661319851875, + 0.020614787936210632, + 0.0076170205138623714, + 0.016414673998951912, + 0.01123830396682024, + 0.016977673396468163, + 0.010033945553004742, + 0.011735658161342144, + 0.010564344003796577, + 0.010491584427654743, + 0.015557351522147655, + 0.013152383267879486, + 0.0439477264881134, + 0.017204366624355316, + 0.016073236241936684, + 0.024262314662337303, + 0.013174885883927345, + 0.04260806366801262, + 0.006566089112311602, + 0.005968940909951925, + 0.005949562415480614, + 0.004472068510949612, + 0.004816030152142048, + 0.005956389009952545, + 0.0030526176560670137, + 0.0035920541267842054, + 0.007924236357212067, + 0.004211689345538616, + 0.020637119188904762, + 0.006506173871457577, + 0.0047096265479922295, + 0.01447264850139618, + 0.02370493672788143, + 0.008357174694538116, + 0.004344778135418892, + 0.004719928838312626, + 0.0031793012749403715, + 0.0028659424278885126, + 0.002918054349720478, + 0.00305826379917562, + 0.0021287391427904367, + 0.0026170925702899694, + 0.012290522456169128, + 0.004988755099475384, + 0.005992915015667677, + 0.004544122144579887, + 0.0030934622045606375, + 0.002691349945962429, + 0.0027232200372964144, + 0.009447584860026836, + 0.0075539033859968185, + 0.00533142639324069, + 0.007018181961029768, + 0.010913483798503876, + 0.005263727158308029, + 0.004953634925186634, + 0.006666174624115229, + 0.00577506935223937, + 0.005317025352269411, + 0.0037411400116980076, + 0.006036767736077309, + 0.0036532508675009012, + 0.004230198450386524, + 0.003697715001180768, + 0.00456992955878377, + 0.006493687629699707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.06916327774524689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636372670531273 + ], + "opposite_logprob_tokens": [ + -3.6513636112213135 + ], + "logprob_reward": -0.02636372670531273, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636372670531273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053286243230104446, + 0.08773516863584518, + 0.07076063007116318, + 0.2999548614025116, + 0.03490534424781799, + 0.2958580255508423, + 0.01596160978078842, + 0.058588992804288864, + 0.03924112766981125, + 0.04370800405740738 + ], + "credit_map": [ + 0.004725737497210503, + 0.004671535454690456, + 0.00711465161293745, + 0.00786236859858036, + 0.007042082492262125, + 0.008057277649641037, + 0.006775099318474531, + 0.007037491537630558, + 0.008702489547431469, + 0.012165145948529243, + 0.012389659881591797, + 0.012429744005203247, + 0.010428003966808319, + 0.010581842623651028, + 0.01255766674876213, + 0.008480612188577652, + 0.005894801579415798, + 0.011510055512189865, + 0.008588769473135471, + 0.008058547042310238, + 0.007301501464098692, + 0.011636738665401936, + 0.005796915851533413, + 0.011973296292126179, + 0.027130261063575745, + 0.019016796723008156, + 0.07414918392896652, + 0.024613523855805397, + 0.02078019641339779, + 0.04293468967080116, + 0.016235150396823883, + 0.07509506493806839, + 0.004792630672454834, + 0.00656100083142519, + 0.005135038401931524, + 0.0058117094449698925, + 0.0025806918274611235, + 0.004229307174682617, + 0.0025261184200644493, + 0.0032688486389815807, + 0.02643871307373047, + 0.011262119747698307, + 0.11599470674991608, + 0.01912664994597435, + 0.010720889084041119, + 0.050189755856990814, + 0.04647236317396164, + 0.01565282605588436, + 0.0023332468699663877, + 0.0019678245298564434, + 0.0015572266420349479, + 0.0025954314041882753, + 0.0017902367981150746, + 0.0020301430486142635, + 0.0018730873707681894, + 0.001814411603845656, + 0.014878466725349426, + 0.006097853183746338, + 0.005922101438045502, + 0.007046446204185486, + 0.007376379799097776, + 0.0050959233194589615, + 0.004178475588560104, + 0.007993345148861408, + 0.004651346709579229, + 0.0031002224422991276, + 0.006082560401409864, + 0.003412070218473673, + 0.0039840564131736755, + 0.006856505759060383, + 0.0038708180654793978, + 0.007283544167876244, + 0.006407434586435556, + 0.005591634660959244, + 0.0064870696514844894, + 0.004623045213520527, + 0.0054290988482534885, + 0.00444742850959301, + 0.004609058145433664, + 0.006113233044743538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.02636372670531273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038112103939056396 + ], + "opposite_logprob_tokens": [ + -3.288112163543701 + ], + "logprob_reward": -0.038112103939056396, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038112103939056396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1422508955001831, + 0.09864460676908493, + 0.19359718263149261, + 0.26172497868537903, + 0.13053566217422485, + 0.051122035831213, + 0.02948545664548874, + 0.027432098984718323, + 0.02613232471048832, + 0.03907478228211403 + ], + "credit_map": [ + 0.012417747639119625, + 0.01753913424909115, + 0.020870262756943703, + 0.020999116823077202, + 0.018129926174879074, + 0.019037550315260887, + 0.019444601610302925, + 0.013812556862831116, + 0.008104066364467144, + 0.011793606914579868, + 0.013303527608513832, + 0.019172227010130882, + 0.014190444722771645, + 0.01139255240559578, + 0.011841085739433765, + 0.008847087621688843, + 0.017572885379195213, + 0.02536076121032238, + 0.023596737533807755, + 0.03479555621743202, + 0.02018226683139801, + 0.022609787061810493, + 0.023888496682047844, + 0.025590676814317703, + 0.025962049141526222, + 0.020891910418868065, + 0.0607406347990036, + 0.023554448038339615, + 0.017654981464147568, + 0.04078167676925659, + 0.013713393360376358, + 0.058425866067409515, + 0.016130931675434113, + 0.029905453324317932, + 0.01839636079967022, + 0.025478200986981392, + 0.00988290086388588, + 0.009839986450970173, + 0.009572597220540047, + 0.011329226195812225, + 0.004701352212578058, + 0.0030119314324110746, + 0.014194579795002937, + 0.0038678841665387154, + 0.005409372970461845, + 0.006378184072673321, + 0.00918078888207674, + 0.004377941135317087, + 0.0037702617701143026, + 0.004262072965502739, + 0.0031688406597822905, + 0.005331615451723337, + 0.004150823689997196, + 0.0025029596872627735, + 0.0035150584299117327, + 0.0027838218957185745, + 0.00455051613971591, + 0.002294725738465786, + 0.0023128101602196693, + 0.0039576138369739056, + 0.004561152309179306, + 0.004886710550636053, + 0.00217242818325758, + 0.002696139505133033, + 0.0032022972591221333, + 0.002586982911452651, + 0.0030294284224510193, + 0.004758840426802635, + 0.0027403905987739563, + 0.002237253822386265, + 0.0035298564471304417, + 0.004047272261232138, + 0.006136713083833456, + 0.004143706988543272, + 0.004705467261373997, + 0.004148782230913639, + 0.006398821249604225, + 0.00408887816593051, + 0.006044641137123108, + 0.0034077675081789494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.038112103939056396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627647835761309 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627647835761309, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627647835761309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15640753507614136, + 0.1491497904062271, + 0.07539469748735428, + 0.28236058354377747, + 0.08398042619228363, + 0.05687214061617851, + 0.07139143347740173, + 0.041121967136859894, + 0.0504889115691185, + 0.032832566648721695 + ], + "credit_map": [ + 0.016266491264104843, + 0.008275829255580902, + 0.029419541358947754, + 0.014776220545172691, + 0.013824548572301865, + 0.021025855094194412, + 0.014674250967800617, + 0.0381447970867157, + 0.01236022263765335, + 0.013086706399917603, + 0.021653451025485992, + 0.02066132426261902, + 0.013509204611182213, + 0.03556985408067703, + 0.014166810549795628, + 0.01814221404492855, + 0.0058603896759450436, + 0.010370503179728985, + 0.00945983175188303, + 0.008050888776779175, + 0.011428792029619217, + 0.011407890357077122, + 0.00836468581110239, + 0.010451714508235455, + 0.02236238494515419, + 0.02217581681907177, + 0.07811512053012848, + 0.023188065737485886, + 0.016265783458948135, + 0.03615902364253998, + 0.013249939307570457, + 0.07084444165229797, + 0.007219020742923021, + 0.01356489397585392, + 0.011408631689846516, + 0.015535562299191952, + 0.007464653812348843, + 0.010769200511276722, + 0.008323975838720798, + 0.009694484062492847, + 0.003837974276393652, + 0.0038017239421606064, + 0.0053932941518723965, + 0.004418729338794947, + 0.004339827224612236, + 0.012421404011547565, + 0.015546352602541447, + 0.007112833671271801, + 0.010607603937387466, + 0.007163699716329575, + 0.00882168859243393, + 0.014852416701614857, + 0.00642394507303834, + 0.010393147356808186, + 0.007832149975001812, + 0.005296786315739155, + 0.01161815132945776, + 0.003351812483742833, + 0.003907820209860802, + 0.004308230243623257, + 0.003640673588961363, + 0.005345460027456284, + 0.002894432283937931, + 0.006055388133972883, + 0.006535346154123545, + 0.004464001394808292, + 0.005859717261046171, + 0.007009330205619335, + 0.005195013713091612, + 0.007389083504676819, + 0.00574449822306633, + 0.008291923440992832, + 0.006274597253650427, + 0.0030393812339752913, + 0.004619984421879053, + 0.002825923962518573, + 0.004141737706959248, + 0.0034301099367439747, + 0.0055866241455078125, + 0.0029142077546566725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.007627647835761309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533116154372692 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533116154372692, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533116154372692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06329193711280823, + 0.07107905298471451, + 0.057537395507097244, + 0.22847476601600647, + 0.0713031217455864, + 0.18477775156497955, + 0.015477718785405159, + 0.04271957650780678, + 0.17572084069252014, + 0.08961784839630127 + ], + "credit_map": [ + 0.006452181842178106, + 0.004942421801388264, + 0.00860120914876461, + 0.005876548588275909, + 0.0074170553125441074, + 0.011156820692121983, + 0.004696289077401161, + 0.014149407856166363, + 0.005362901370972395, + 0.00620540976524353, + 0.008482737466692924, + 0.01150793582201004, + 0.006807577330619097, + 0.012063221074640751, + 0.009137870743870735, + 0.011511393822729588, + 0.005008626729249954, + 0.006820688955485821, + 0.007111651822924614, + 0.010113178752362728, + 0.006822324823588133, + 0.00787089392542839, + 0.005312318913638592, + 0.008477709256112576, + 0.02280990593135357, + 0.018764302134513855, + 0.04409976676106453, + 0.02118847332894802, + 0.020546305924654007, + 0.03941696137189865, + 0.015610282309353352, + 0.04603877291083336, + 0.008803686127066612, + 0.010927241295576096, + 0.009549572132527828, + 0.010497964918613434, + 0.007092961575835943, + 0.010247908532619476, + 0.005163330584764481, + 0.009020456112921238, + 0.01543122436851263, + 0.007153155282139778, + 0.07439731061458588, + 0.010077591985464096, + 0.005268316250294447, + 0.040554046630859375, + 0.022588852792978287, + 0.009307265281677246, + 0.002148882718756795, + 0.002397160977125168, + 0.0019508287077769637, + 0.0014070061733946204, + 0.0019609006121754646, + 0.0020023782271891832, + 0.0013418737798929214, + 0.0022686864249408245, + 0.011341910809278488, + 0.005589702632278204, + 0.003769531147554517, + 0.0041147381998598576, + 0.005476579535752535, + 0.004827397875487804, + 0.0025078284088522196, + 0.005091889761388302, + 0.01844608038663864, + 0.013018141500651836, + 0.025222664698958397, + 0.01662720926105976, + 0.016569631174206734, + 0.034779053181409836, + 0.014216357842087746, + 0.03684170916676521, + 0.010629874654114246, + 0.006889660842716694, + 0.014371340163052082, + 0.010731668211519718, + 0.00934872217476368, + 0.008843421004712582, + 0.012527532875537872, + 0.016275623813271523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.012533116154372692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a low angle, the camera slowly tilting upward, begins with the image of a hand clad in a soft black leather glove. Fingers grip the ornate crossguard of a longsword, crafted from a dull silver material and intricately carved with arabesques. The hand and hilt are ensconced in the deep shadows of the ancient library, illuminated only by the warm, flickering light of a candle from outside the frame, casting soft golden highlights on the metal and leather. With a remarkably steady and deliberate movement, the hand pulls upward, slowly and gently, the sword is drawn from its dark wooden scabbard, inlaid with polished metalwork, inch by inch, accompanied only by the subtle hiss of steel against the lining. As the razor-sharp, mirror-polished blade gleamed with a cold light, it perfectly captured and distorted the reflection of its surroundings: rows of magnificent mahogany bookshelves reaching the ceiling, filled with thick, leather-bound ancient books, the faded gold-stamped text on the spines flowing as the blade moved. The entire scene was filled with a solemn, suspenseful, and historically tranquil atmosphere.", + "seed": 21862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_5372_seed_21862_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0327662032097578, + "vlm_logprob": -0.0327662032097578, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.0327662032097578, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07451551407575607 + ], + "opposite_logprob_tokens": [ + -3.6995155811309814 + ], + "logprob_reward": -0.07451551407575607, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.07451551407575607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07386061549186707, + 0.0531391017138958, + 0.14102070033550262, + 0.1192089170217514, + 0.05894576758146286, + 0.07909354567527771, + 0.11173371225595474, + 0.1520427018404007, + 0.16339917480945587, + 0.04755571484565735 + ], + "credit_map": [ + 0.009846709668636322, + 0.006955415941774845, + 0.00971599668264389, + 0.008667616173624992, + 0.011201182380318642, + 0.008077851496636868, + 0.010636302642524242, + 0.008759544230997562, + 0.003958660643547773, + 0.0048359171487390995, + 0.0054236287251114845, + 0.005517593119293451, + 0.003432769328355789, + 0.012580621987581253, + 0.0056101055815815926, + 0.011779810301959515, + 0.020209236070513725, + 0.011922366917133331, + 0.017795884981751442, + 0.014487545005977154, + 0.018093900755047798, + 0.021620193496346474, + 0.011664509773254395, + 0.02522706426680088, + 0.013396762311458588, + 0.015160888433456421, + 0.013810386881232262, + 0.01244616974145174, + 0.01837443932890892, + 0.01745534874498844, + 0.01175319030880928, + 0.016811739653348923, + 0.009972670115530491, + 0.005838467739522457, + 0.008470405824482441, + 0.003918067552149296, + 0.009533803910017014, + 0.008436447940766811, + 0.006111524533480406, + 0.006664388347417116, + 0.01357820350676775, + 0.005415833089500666, + 0.015598173253238201, + 0.008324469439685345, + 0.007785255089402199, + 0.011324525810778141, + 0.0065791490487754345, + 0.010487940162420273, + 0.01612231135368347, + 0.0063149710185825825, + 0.0356968492269516, + 0.0064472900703549385, + 0.006117070093750954, + 0.011770148761570454, + 0.0075463405810296535, + 0.021718738600611687, + 0.013301225379109383, + 0.018384844064712524, + 0.03519327938556671, + 0.012242446653544903, + 0.016762733459472656, + 0.025399180129170418, + 0.013437775894999504, + 0.017321227118372917, + 0.020555315539240837, + 0.0157381109893322, + 0.034816887229681015, + 0.013146612793207169, + 0.015066774562001228, + 0.02224966511130333, + 0.019830381497740746, + 0.02199544943869114, + 0.006383925676345825, + 0.005582048092037439, + 0.006366152316331863, + 0.003546057501807809, + 0.004738944116979837, + 0.007394719868898392, + 0.007343691773712635, + 0.006200180388987064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.07451551407575607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820380985736847 + ], + "opposite_logprob_tokens": [ + -4.0182037353515625 + ], + "logprob_reward": -0.01820380985736847, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820380985736847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05665580928325653, + 0.04334442317485809, + 0.2777348458766937, + 0.040351614356040955, + 0.04793839901685715, + 0.07647500932216644, + 0.07671293616294861, + 0.07546689361333847, + 0.2321850210428238, + 0.07313503324985504 + ], + "credit_map": [ + 0.007121015805751085, + 0.0037290852051228285, + 0.005631887819617987, + 0.004355954006314278, + 0.005110830068588257, + 0.013863017782568932, + 0.0044892942532896996, + 0.012354722246527672, + 0.00470215268433094, + 0.006043798290193081, + 0.005563237238675356, + 0.007251190021634102, + 0.005116323009133339, + 0.005055263172835112, + 0.004337833262979984, + 0.005274628754705191, + 0.03552844375371933, + 0.022624509409070015, + 0.036817535758018494, + 0.028389466926455498, + 0.032913293689489365, + 0.041469138115644455, + 0.025603927671909332, + 0.05438854172825813, + 0.004405589774250984, + 0.003865417093038559, + 0.005833851173520088, + 0.004098472651094198, + 0.004119924735277891, + 0.006565459538251162, + 0.00360321206972003, + 0.007859691046178341, + 0.004935606848448515, + 0.0045987083576619625, + 0.004441882483661175, + 0.004587501287460327, + 0.008484615944325924, + 0.006268166936933994, + 0.008222280070185661, + 0.0063996389508247375, + 0.014942292124032974, + 0.0069047315046191216, + 0.011418869718909264, + 0.00877844076603651, + 0.00825601164251566, + 0.010511879809200764, + 0.006567878182977438, + 0.009094908833503723, + 0.00910439994186163, + 0.005597424227744341, + 0.019435971975326538, + 0.005674112122505903, + 0.005958658177405596, + 0.011168868280947208, + 0.005798731930553913, + 0.013974769972264767, + 0.012075359001755714, + 0.006711524911224842, + 0.011526343412697315, + 0.007524934131652117, + 0.008319252170622349, + 0.010110720060765743, + 0.009229219518601894, + 0.009969537146389484, + 0.0384662002325058, + 0.026024870574474335, + 0.044949233531951904, + 0.023325225338339806, + 0.017846262082457542, + 0.031104840338230133, + 0.023419398814439774, + 0.02704901434481144, + 0.007100902497768402, + 0.007400244940072298, + 0.008164090104401112, + 0.006530934013426304, + 0.010165956802666187, + 0.015181371942162514, + 0.00881999358534813, + 0.00977153517305851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.01820380985736847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635246515274048 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.02635246515274048, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635246515274048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.196660578250885, + 0.09791512787342072, + 0.1359054148197174, + 0.04295859858393669, + 0.07258149236440659, + 0.06597928702831268, + 0.12558680772781372, + 0.053246017545461655, + 0.1488359570503235, + 0.06033071503043175 + ], + "credit_map": [ + 0.020588833838701248, + 0.01776433177292347, + 0.01917208917438984, + 0.025641189888119698, + 0.025665832683444023, + 0.03397468850016594, + 0.018504051491618156, + 0.03534955531358719, + 0.0077401017770171165, + 0.00831614900380373, + 0.008442741818726063, + 0.010153214447200298, + 0.00940892193466425, + 0.02235737442970276, + 0.005877309944480658, + 0.02561931312084198, + 0.020209163427352905, + 0.013126907870173454, + 0.016095582395792007, + 0.017109032720327377, + 0.016159888356924057, + 0.0219647828489542, + 0.012110424228012562, + 0.01912962645292282, + 0.005801710300147533, + 0.00384205998852849, + 0.005169704090803862, + 0.0032078444492071867, + 0.004932418931275606, + 0.006440306082367897, + 0.004436344839632511, + 0.009128211997449398, + 0.012034421786665916, + 0.006473680958151817, + 0.008026165887713432, + 0.005403887014836073, + 0.012299424037337303, + 0.010719847865402699, + 0.00767440814524889, + 0.009949657134711742, + 0.01050016563385725, + 0.005787941627204418, + 0.0071897441521286964, + 0.0061493245884776115, + 0.008118522353470325, + 0.010008899495005608, + 0.008612707257270813, + 0.009611982852220535, + 0.017075074836611748, + 0.0073966835625469685, + 0.03943011164665222, + 0.008677593432366848, + 0.00620518671348691, + 0.01407882571220398, + 0.007890182547271252, + 0.02483314834535122, + 0.010005908086895943, + 0.0050867474637925625, + 0.008269699290394783, + 0.004524769727140665, + 0.006018845364451408, + 0.006754795555025339, + 0.006831783801317215, + 0.00575347151607275, + 0.021933099254965782, + 0.016248928382992744, + 0.028529852628707886, + 0.01209220103919506, + 0.014942027628421783, + 0.0202987901866436, + 0.01737554557621479, + 0.017415523529052734, + 0.009217415936291218, + 0.006013896781951189, + 0.008908615447580814, + 0.0032667077612131834, + 0.005621568765491247, + 0.00934507604688406, + 0.008953538723289967, + 0.009003893472254276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.02635246515274048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09829270094633102, + 0.12180622667074203, + 0.09237337112426758, + 0.07611853629350662, + 0.1127421036362648, + 0.07732651382684708, + 0.08451689034700394, + 0.05923926457762718, + 0.12258540093898773, + 0.15499894320964813 + ], + "credit_map": [ + 0.012594526633620262, + 0.007836371660232544, + 0.016605066135525703, + 0.01136520504951477, + 0.008806035853922367, + 0.01649271883070469, + 0.00722153764218092, + 0.017371252179145813, + 0.006563820410519838, + 0.005459227599203587, + 0.010141829960048199, + 0.011350143700838089, + 0.006596902385354042, + 0.03814561292529106, + 0.007834293879568577, + 0.03571441024541855, + 0.015524274669587612, + 0.007237681653350592, + 0.01078106090426445, + 0.010275796987116337, + 0.012151180766522884, + 0.013990814797580242, + 0.008653485216200352, + 0.013759082183241844, + 0.00823754072189331, + 0.007254971656948328, + 0.007411289028823376, + 0.004921054467558861, + 0.012446647509932518, + 0.011716246604919434, + 0.005782450549304485, + 0.01834835112094879, + 0.013736407272517681, + 0.009302156046032906, + 0.017549175769090652, + 0.010305798612535, + 0.01640372723340988, + 0.014542246237397194, + 0.013973919674754143, + 0.01692868396639824, + 0.01293215248733759, + 0.004485929384827614, + 0.012162002734839916, + 0.010686736553907394, + 0.005834150593727827, + 0.011334124021232128, + 0.007402236573398113, + 0.012489196844398975, + 0.015136364847421646, + 0.006343984045088291, + 0.02335377410054207, + 0.005280130077153444, + 0.004848330747336149, + 0.010259066708385944, + 0.004799854941666126, + 0.014495392329990864, + 0.007276915013790131, + 0.006214706692844629, + 0.007782194763422012, + 0.005168319679796696, + 0.00847830343991518, + 0.0056909616105258465, + 0.009690983220934868, + 0.008936885744333267, + 0.014146605506539345, + 0.01363762840628624, + 0.01713615469634533, + 0.009996440261602402, + 0.022357342764735222, + 0.018173692747950554, + 0.01286512240767479, + 0.014272432774305344, + 0.021447112783789635, + 0.01505863107740879, + 0.02586638554930687, + 0.009243584237992764, + 0.017589008435606956, + 0.02996671013534069, + 0.015129403211176395, + 0.02069813944399357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.011074034497141838 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685192465782166 + ], + "opposite_logprob_tokens": [ + -3.4086852073669434 + ], + "logprob_reward": -0.033685192465782166, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685192465782166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06111430749297142, + 0.05385496839880943, + 0.03103105165064335, + 0.055401284247636795, + 0.09311763942241669, + 0.0626608282327652, + 0.2519438862800598, + 0.04392712935805321, + 0.2594367265701294, + 0.08751218020915985 + ], + "credit_map": [ + 0.009078684262931347, + 0.004479529801756144, + 0.007541856728494167, + 0.005846103187650442, + 0.004320311360061169, + 0.014057382941246033, + 0.004344284534454346, + 0.011446150951087475, + 0.002627328271046281, + 0.003386987140402198, + 0.0040169828571379185, + 0.006781405303627253, + 0.0026243191678076982, + 0.014547942206263542, + 0.004278541076928377, + 0.015591464005410671, + 0.0033974051475524902, + 0.004091266542673111, + 0.0037032028194516897, + 0.005019127856940031, + 0.0024999654851853848, + 0.004865084309130907, + 0.002695763483643532, + 0.00475923391059041, + 0.0072751110419631, + 0.005596734583377838, + 0.006698119919747114, + 0.0057890466414391994, + 0.006978342309594154, + 0.008333591744303703, + 0.005724933929741383, + 0.009005406871438026, + 0.015161419287323952, + 0.006818105466663837, + 0.013309769332408905, + 0.007514152210205793, + 0.014694992452859879, + 0.013825921341776848, + 0.008657007478177547, + 0.013136274181306362, + 0.014374936930835247, + 0.0031681291293352842, + 0.011497050523757935, + 0.006981355138123035, + 0.003172158496454358, + 0.00928895827382803, + 0.005659565329551697, + 0.008518680930137634, + 0.04412543773651123, + 0.017244579270482063, + 0.0697389543056488, + 0.01567612774670124, + 0.012325771152973175, + 0.03313811868429184, + 0.013923062942922115, + 0.04577184095978737, + 0.006031188648194075, + 0.005355945322662592, + 0.007477758917957544, + 0.0037907373625785112, + 0.005050391890108585, + 0.006823776289820671, + 0.004019524436444044, + 0.005377807654440403, + 0.02968660742044449, + 0.020250171422958374, + 0.05966973677277565, + 0.016690002754330635, + 0.0385967381298542, + 0.040195565670728683, + 0.026486100628972054, + 0.027861813083291054, + 0.009628920815885067, + 0.010476733557879925, + 0.012914917431771755, + 0.006617103237658739, + 0.010602318681776524, + 0.016699757426977158, + 0.009401081129908562, + 0.011171339079737663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.033685192465782166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a low angle, the camera slowly tilting upward, begins with the image of a hand clad in a soft black leather glove. Fingers grip the ornate crossguard of a longsword, crafted from a dull silver material and intricately carved with arabesques. The hand and hilt are ensconced in the deep shadows of the ancient library, illuminated only by the warm, flickering light of a candle from outside the frame, casting soft golden highlights on the metal and leather. With a remarkably steady and deliberate movement, the hand pulls upward, slowly and gently, the sword is drawn from its dark wooden scabbard, inlaid with polished metalwork, inch by inch, accompanied only by the subtle hiss of steel against the lining. As the razor-sharp, mirror-polished blade gleamed with a cold light, it perfectly captured and distorted the reflection of its surroundings: rows of magnificent mahogany bookshelves reaching the ceiling, filled with thick, leather-bound ancient books, the faded gold-stamped text on the spines flowing as the blade moved. The entire scene was filled with a solemn, suspenseful, and historically tranquil atmosphere.", + "seed": 21863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_5372_seed_21863_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03128667008131743, + "vlm_logprob": -0.03128667008131743, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03128667008131743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09077528864145279 + ], + "opposite_logprob_tokens": [ + -3.09077525138855 + ], + "logprob_reward": -0.09077528864145279, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.09077528864145279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10368528962135315, + 0.058889616280794144, + 0.07536346465349197, + 0.22886165976524353, + 0.1326218545436859, + 0.03646773099899292, + 0.08909667283296585, + 0.052539147436618805, + 0.12631714344024658, + 0.09615743905305862 + ], + "credit_map": [ + 0.00987579207867384, + 0.009304003790020943, + 0.022280633449554443, + 0.022573597729206085, + 0.009719174355268478, + 0.008796105161309242, + 0.013748171739280224, + 0.00738780340179801, + 0.009720593690872192, + 0.009993916377425194, + 0.00963509175926447, + 0.010226842015981674, + 0.006089417263865471, + 0.0032293335534632206, + 0.005366015713661909, + 0.004628399387001991, + 0.00433791009709239, + 0.006346812937408686, + 0.017030758783221245, + 0.020283350721001625, + 0.008824375458061695, + 0.0032965675927698612, + 0.011653618887066841, + 0.003590063191950321, + 0.016847804188728333, + 0.03851881995797157, + 0.02700127847492695, + 0.019679337739944458, + 0.03880547359585762, + 0.016333237290382385, + 0.05637900531291962, + 0.015296682715415955, + 0.015510682947933674, + 0.024315720424056053, + 0.01658039353787899, + 0.016425523906946182, + 0.01573328673839569, + 0.014085778966546059, + 0.01260132435709238, + 0.017369138076901436, + 0.004402208141982555, + 0.0033953513484448195, + 0.005264087114483118, + 0.004584910813719034, + 0.003497835248708725, + 0.004002786241471767, + 0.007939962670207024, + 0.0033805868588387966, + 0.007247012574225664, + 0.008139540441334248, + 0.02105735056102276, + 0.008820103481411934, + 0.009157346561551094, + 0.008956066332757473, + 0.007687659934163094, + 0.018031593412160873, + 0.006228820886462927, + 0.008384325541555882, + 0.005286238621920347, + 0.005801370833069086, + 0.0038332196418195963, + 0.01135655865073204, + 0.0027834251523017883, + 0.00886518880724907, + 0.012670234777033329, + 0.016092833131551743, + 0.02963261865079403, + 0.009539959952235222, + 0.008429920300841331, + 0.013126038014888763, + 0.005814742762595415, + 0.031010789796710014, + 0.00911751575767994, + 0.01827624998986721, + 0.010372328571975231, + 0.012451468966901302, + 0.011876252479851246, + 0.014525805599987507, + 0.010126650333404541, + 0.009411165490746498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.09077528864145279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223976992070675 + ], + "opposite_logprob_tokens": [ + -4.264224052429199 + ], + "logprob_reward": -0.014223976992070675, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223976992070675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05841045081615448, + 0.10982176661491394, + 0.09882901608943939, + 0.0779079794883728, + 0.07409003376960754, + 0.04452056065201759, + 0.2176937609910965, + 0.06243673339486122, + 0.19175435602664948, + 0.06453528255224228 + ], + "credit_map": [ + 0.006683737505227327, + 0.004756968934088945, + 0.011905000545084476, + 0.012022077105939388, + 0.005781345069408417, + 0.006295688915997744, + 0.00588632095605135, + 0.005079311318695545, + 0.01132255233824253, + 0.009353432804346085, + 0.02503378689289093, + 0.025162670761346817, + 0.01231585256755352, + 0.005669401027262211, + 0.014861255884170532, + 0.006102810148149729, + 0.011281473562121391, + 0.007238679565489292, + 0.02214459516108036, + 0.025023413822054863, + 0.009184151887893677, + 0.005665272008627653, + 0.011246692389249802, + 0.007044734433293343, + 0.006422147620469332, + 0.006676401011645794, + 0.009512636810541153, + 0.00855420995503664, + 0.011188933625817299, + 0.008954107761383057, + 0.014906883239746094, + 0.011692661792039871, + 0.007860948331654072, + 0.010975376702845097, + 0.007885443978011608, + 0.008071577176451683, + 0.011511795222759247, + 0.007941027171909809, + 0.011083048768341541, + 0.008760815486311913, + 0.006444770377129316, + 0.004557522013783455, + 0.00610427837818861, + 0.004206201992928982, + 0.004516540560871363, + 0.0063965460285544395, + 0.007521944120526314, + 0.004772757645696402, + 0.01812107302248478, + 0.019758284091949463, + 0.05152888223528862, + 0.027462447062134743, + 0.01754382997751236, + 0.024835780262947083, + 0.021077679470181465, + 0.037365786731243134, + 0.009507521986961365, + 0.01098127942532301, + 0.007554817013442516, + 0.006376239005476236, + 0.004897952079772949, + 0.010909016244113445, + 0.0037214842159301043, + 0.008488423191010952, + 0.019608935341238976, + 0.023883791640400887, + 0.04039435088634491, + 0.012369138188660145, + 0.01645846851170063, + 0.01938926987349987, + 0.011340707540512085, + 0.04830969497561455, + 0.004586281254887581, + 0.008245089091360569, + 0.006319183390587568, + 0.01194695383310318, + 0.0053335027769207954, + 0.009344839490950108, + 0.008051836863160133, + 0.010707595385611057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.014223976992070675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363495737314224 + ], + "opposite_logprob_tokens": [ + -3.6513636112213135 + ], + "logprob_reward": -0.026363495737314224, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363495737314224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17689834535121918, + 0.0791475847363472, + 0.12534095346927643, + 0.13413245975971222, + 0.06444885581731796, + 0.04005512595176697, + 0.10416273027658463, + 0.029336582869291306, + 0.17258375883102417, + 0.07389361411333084 + ], + "credit_map": [ + 0.015703648328781128, + 0.01745583489537239, + 0.036854036152362823, + 0.04315958917140961, + 0.015709541738033295, + 0.011618336662650108, + 0.023550478741526604, + 0.012846878729760647, + 0.00869720708578825, + 0.010485244914889336, + 0.018425755202770233, + 0.017436666414141655, + 0.008706700056791306, + 0.003128022886812687, + 0.008874267339706421, + 0.003393723163753748, + 0.007775710429996252, + 0.015742355957627296, + 0.02523786388337612, + 0.01989605464041233, + 0.01960054598748684, + 0.009225612506270409, + 0.017530294135212898, + 0.01033251266926527, + 0.01231966633349657, + 0.022696293890476227, + 0.020425129681825638, + 0.013301314786076546, + 0.021564384922385216, + 0.015492169186472893, + 0.013527494855225086, + 0.014805988408625126, + 0.007430946454405785, + 0.010887865908443928, + 0.007152384612709284, + 0.007017852272838354, + 0.008189193904399872, + 0.007723155431449413, + 0.006890835706144571, + 0.009156622923910618, + 0.004489942919462919, + 0.004284628666937351, + 0.005879030097275972, + 0.0035706914495676756, + 0.003370379563421011, + 0.005772327072918415, + 0.005495200399309397, + 0.007192922756075859, + 0.010139038786292076, + 0.009705454111099243, + 0.02458370104432106, + 0.01317722350358963, + 0.008080961182713509, + 0.010307242162525654, + 0.010341922752559185, + 0.017827186733484268, + 0.005245952866971493, + 0.0049438984133303165, + 0.00499779311940074, + 0.004065331537276506, + 0.0023319905158132315, + 0.0028715927619487047, + 0.0015104101039469242, + 0.003369614016264677, + 0.01700209453701973, + 0.022331153973937035, + 0.037646468728780746, + 0.013385646045207977, + 0.014093614183366299, + 0.01755746454000473, + 0.00956137664616108, + 0.04100595414638519, + 0.005588902626186609, + 0.013780275359749794, + 0.005774042569100857, + 0.012707484886050224, + 0.006365374196320772, + 0.01117745228111744, + 0.008514264598488808, + 0.009985814802348614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.026363495737314224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06864354759454727, + 0.09665842354297638, + 0.13206766545772552, + 0.1780349463224411, + 0.052520502358675, + 0.07169703394174576, + 0.1199677437543869, + 0.07759054005146027, + 0.07127216458320618, + 0.1315474659204483 + ], + "credit_map": [ + 0.005504146218299866, + 0.007298130542039871, + 0.015325404703617096, + 0.01680438034236431, + 0.0073709795251488686, + 0.004891761112958193, + 0.007620100397616625, + 0.0038286400958895683, + 0.009978833608329296, + 0.006758757866919041, + 0.02510884776711464, + 0.024869820103049278, + 0.01293911598622799, + 0.003675509011372924, + 0.00856173224747181, + 0.004765805322676897, + 0.00971589982509613, + 0.009552350267767906, + 0.024639848619699478, + 0.02813924103975296, + 0.013830193318426609, + 0.007653762586414814, + 0.02564847096800804, + 0.012887904420495033, + 0.013987868092954159, + 0.02499011717736721, + 0.026907971128821373, + 0.01851700246334076, + 0.031081590801477432, + 0.02266724780201912, + 0.016913095489144325, + 0.022970061749219894, + 0.007318656425923109, + 0.009200630709528923, + 0.006457929965108633, + 0.006944525986909866, + 0.004644858185201883, + 0.004126998130232096, + 0.008195901289582253, + 0.005630998406559229, + 0.008498228155076504, + 0.008565627038478851, + 0.009118431247770786, + 0.00885761808604002, + 0.007609536871314049, + 0.011289077810943127, + 0.01024620234966278, + 0.007512309588491917, + 0.011324322782456875, + 0.010782326571643353, + 0.02766985446214676, + 0.019221363589167595, + 0.009276936762034893, + 0.010636267252266407, + 0.012326144613325596, + 0.0187305249273777, + 0.009357859380543232, + 0.015176026150584221, + 0.008291402831673622, + 0.013160496950149536, + 0.0052813077345490456, + 0.011690870858728886, + 0.00303712603636086, + 0.011595446616411209, + 0.007113991770893335, + 0.010044974274933338, + 0.01573609933257103, + 0.005502729211002588, + 0.008975940756499767, + 0.009072710759937763, + 0.005802057217806578, + 0.009023657999932766, + 0.00996861606836319, + 0.015188281424343586, + 0.014693531207740307, + 0.01905140094459057, + 0.01830066554248333, + 0.019390428438782692, + 0.017497539520263672, + 0.017457006499171257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.012539473362267017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531115673482418 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531115673482418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531115673482418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20012956857681274, + 0.06737400591373444, + 0.07219741493463516, + 0.05564456060528755, + 0.05112674459815025, + 0.04176906496286392, + 0.04970753937959671, + 0.08366981148719788, + 0.14940865337848663, + 0.22897270321846008 + ], + "credit_map": [ + 0.02099611796438694, + 0.01835131272673607, + 0.040240783244371414, + 0.04594868794083595, + 0.019660357385873795, + 0.018462015315890312, + 0.020781978964805603, + 0.015688277781009674, + 0.007587790489196777, + 0.009074036031961441, + 0.012846535071730614, + 0.014347204938530922, + 0.007434039376676083, + 0.004031653981655836, + 0.008008119650185108, + 0.0040446193888783455, + 0.004448696039617062, + 0.005349651910364628, + 0.014512085355818272, + 0.017623163759708405, + 0.007068312261253595, + 0.00453026732429862, + 0.014459076337516308, + 0.004206155892461538, + 0.004167122300714254, + 0.009466947056353092, + 0.006830854807049036, + 0.004422138445079327, + 0.00955109391361475, + 0.006071402225643396, + 0.008550839498639107, + 0.006584159564226866, + 0.006074602250009775, + 0.009815847501158714, + 0.00636851042509079, + 0.006351180374622345, + 0.005983359646052122, + 0.005057099740952253, + 0.00487179309129715, + 0.006604344584047794, + 0.005924389231950045, + 0.006028120871633291, + 0.00629116827622056, + 0.004195932764559984, + 0.003559606382623315, + 0.006312861572951078, + 0.0059484499506652355, + 0.0035085303243249655, + 0.00559775996953249, + 0.0037811892107129097, + 0.011390741914510727, + 0.008387935347855091, + 0.0034074210561811924, + 0.005393032915890217, + 0.004161069169640541, + 0.007588385604321957, + 0.010546100325882435, + 0.01313397753983736, + 0.01664232648909092, + 0.010498019866645336, + 0.007161846850067377, + 0.009298124350607395, + 0.005330378655344248, + 0.011059027165174484, + 0.015154453925788403, + 0.02133307233452797, + 0.03135158121585846, + 0.014345376752316952, + 0.015374093316495419, + 0.016480734571814537, + 0.011576112359762192, + 0.023793216794729233, + 0.021314676851034164, + 0.03648396208882332, + 0.022068144753575325, + 0.041700638830661774, + 0.018185297027230263, + 0.03469224274158478, + 0.02517741359770298, + 0.029350310564041138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012531115673482418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a low angle, the camera slowly tilting upward, begins with the image of a hand clad in a soft black leather glove. Fingers grip the ornate crossguard of a longsword, crafted from a dull silver material and intricately carved with arabesques. The hand and hilt are ensconced in the deep shadows of the ancient library, illuminated only by the warm, flickering light of a candle from outside the frame, casting soft golden highlights on the metal and leather. With a remarkably steady and deliberate movement, the hand pulls upward, slowly and gently, the sword is drawn from its dark wooden scabbard, inlaid with polished metalwork, inch by inch, accompanied only by the subtle hiss of steel against the lining. As the razor-sharp, mirror-polished blade gleamed with a cold light, it perfectly captured and distorted the reflection of its surroundings: rows of magnificent mahogany bookshelves reaching the ceiling, filled with thick, leather-bound ancient books, the faded gold-stamped text on the spines flowing as the blade moved. The entire scene was filled with a solemn, suspenseful, and historically tranquil atmosphere.", + "seed": 21864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_6/index_5372_seed_21864_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1102805756032467, + "vlm_logprob": -0.1102805756032467, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.1102805756032467, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35046061873435974 + ], + "opposite_logprob_tokens": [ + -1.3504606485366821 + ], + "logprob_reward": -0.35046061873435974, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.35046061873435974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0652698278427124, + 0.043161679059267044, + 0.036564599722623825, + 0.10073868185281754, + 0.06924797594547272, + 0.10891690850257874, + 0.0884750485420227, + 0.023764362558722496, + 0.10326799750328064, + 0.36059290170669556 + ], + "credit_map": [ + 0.0054551768116652966, + 0.003399781882762909, + 0.012940012849867344, + 0.013815859332680702, + 0.0069939871318638325, + 0.003930911887437105, + 0.01141535583883524, + 0.00731874443590641, + 0.002679739845916629, + 0.0017190850339829922, + 0.008369243703782558, + 0.005219112150371075, + 0.005012220237404108, + 0.0033305881079286337, + 0.007798734121024609, + 0.009032958187162876, + 0.0026431491132825613, + 0.0021694933529943228, + 0.006576350890100002, + 0.0039023070130497217, + 0.00333206495270133, + 0.004292350262403488, + 0.00877113826572895, + 0.004877745173871517, + 0.004204887896776199, + 0.006057333201169968, + 0.013681677170097828, + 0.016191933304071426, + 0.01773321069777012, + 0.007396282162517309, + 0.02480732463300228, + 0.010666035115718842, + 0.011866441927850246, + 0.009558837860822678, + 0.006615335121750832, + 0.010181020945310593, + 0.0053586214780807495, + 0.009313889779150486, + 0.008564633317291737, + 0.007789198774844408, + 0.011117142625153065, + 0.00911979004740715, + 0.01725238934159279, + 0.012085548602044582, + 0.015282019972801208, + 0.021894069388508797, + 0.010257856920361519, + 0.011908088810741901, + 0.007492772303521633, + 0.006247048731893301, + 0.006146733649075031, + 0.026261886581778526, + 0.007976374588906765, + 0.005371383856981993, + 0.008615783415734768, + 0.020363064482808113, + 0.0035110434982925653, + 0.004538388457149267, + 0.002025925787165761, + 0.0032513707410544157, + 0.0028947354294359684, + 0.003323029726743698, + 0.002428201260045171, + 0.0017916702199727297, + 0.01078253798186779, + 0.014691194519400597, + 0.009612507186830044, + 0.022949527949094772, + 0.006443919613957405, + 0.012399416416883469, + 0.008411353453993797, + 0.017977537587285042, + 0.03988908603787422, + 0.11220855265855789, + 0.13624995946884155, + 0.019500620663166046, + 0.012323490343987942, + 0.010781305842101574, + 0.011484676040709019, + 0.018155228346586227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.35046061873435974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374074399471283 + ], + "opposite_logprob_tokens": [ + -3.408740758895874 + ], + "logprob_reward": -0.03374074399471283, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374074399471283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1381257176399231, + 0.13233400881290436, + 0.06898532807826996, + 0.07346092909574509, + 0.1126597598195076, + 0.06474220007658005, + 0.09342988580465317, + 0.1256808042526245, + 0.08775901794433594, + 0.10282231867313385 + ], + "credit_map": [ + 0.011736656539142132, + 0.008648273535072803, + 0.025161851197481155, + 0.02360120601952076, + 0.010969801805913448, + 0.009735614992678165, + 0.02039974182844162, + 0.02787257917225361, + 0.008115287870168686, + 0.007841764949262142, + 0.027884379029273987, + 0.024243483319878578, + 0.013057786040008068, + 0.009044285863637924, + 0.021924646571278572, + 0.020222362130880356, + 0.004523451905697584, + 0.0032474875915795565, + 0.017215313389897346, + 0.0057411519810557365, + 0.0062331329099833965, + 0.004000506363809109, + 0.021665435284376144, + 0.006358851678669453, + 0.005913203116506338, + 0.0074665904976427555, + 0.01358647458255291, + 0.00804887618869543, + 0.012338795699179173, + 0.006366891786456108, + 0.012862072326242924, + 0.006878019310534, + 0.014570916071534157, + 0.019526056945323944, + 0.010506330989301205, + 0.008772041648626328, + 0.006901516113430262, + 0.01647094637155533, + 0.009060117416083813, + 0.026851831004023552, + 0.009743420407176018, + 0.008575573563575745, + 0.010762150399386883, + 0.005373626947402954, + 0.0075609334744513035, + 0.006788903381675482, + 0.005577920004725456, + 0.010359671898186207, + 0.008809377439320087, + 0.006567973643541336, + 0.007849989458918571, + 0.028877729550004005, + 0.00500946119427681, + 0.006001720204949379, + 0.00724980840459466, + 0.02306382544338703, + 0.01842983439564705, + 0.01472193282097578, + 0.013642347417771816, + 0.014101563021540642, + 0.011403419077396393, + 0.01851492002606392, + 0.012142992578446865, + 0.02272379770874977, + 0.010910640470683575, + 0.015245893970131874, + 0.009000353515148163, + 0.013497436419129372, + 0.007635337766259909, + 0.00889397319406271, + 0.008399699814617634, + 0.01417567953467369, + 0.011248094029724598, + 0.02060778997838497, + 0.022079629823565483, + 0.01480475626885891, + 0.0065009091049432755, + 0.006312447600066662, + 0.007124654483050108, + 0.01414403971284628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.03374074399471283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899759709835052 + ], + "opposite_logprob_tokens": [ + -2.5789976119995117 + ], + "logprob_reward": -0.07899759709835052, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899759709835052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1311378926038742, + 0.2064419388771057, + 0.025708744302392006, + 0.1340475082397461, + 0.05541238561272621, + 0.037256333976984024, + 0.10978255420923233, + 0.04032684862613678, + 0.08443548530340195, + 0.17545032501220703 + ], + "credit_map": [ + 0.00973600521683693, + 0.008250507526099682, + 0.023548655211925507, + 0.0272528063505888, + 0.014962108805775642, + 0.00722867576405406, + 0.02157589979469776, + 0.018583251163363457, + 0.01215163990855217, + 0.013705762103199959, + 0.038228049874305725, + 0.03794574737548828, + 0.029489904642105103, + 0.016898907721042633, + 0.04036908969283104, + 0.017652859911322594, + 0.0020598587580025196, + 0.001808853354305029, + 0.003791431663557887, + 0.004651067312806845, + 0.0027745896950364113, + 0.0026415702886879444, + 0.004614040721207857, + 0.0033673341386020184, + 0.005821137223392725, + 0.007496001664549112, + 0.019567497074604034, + 0.018083009868860245, + 0.022162921726703644, + 0.012239786796271801, + 0.03338586911559105, + 0.015291295945644379, + 0.00788567028939724, + 0.009450506418943405, + 0.004823255352675915, + 0.006604487542062998, + 0.003097382141277194, + 0.008625313639640808, + 0.00483891973271966, + 0.010086852125823498, + 0.0059007517993450165, + 0.004423405509442091, + 0.00557055464014411, + 0.003573246067389846, + 0.004922283813357353, + 0.005089948885142803, + 0.0030758106149733067, + 0.004700337536633015, + 0.009845527820289135, + 0.012751802802085876, + 0.010099677368998528, + 0.022169630974531174, + 0.012443003244698048, + 0.008170686662197113, + 0.008661288768053055, + 0.02564094401896, + 0.005357359070330858, + 0.005285829305648804, + 0.004925926215946674, + 0.005739297717809677, + 0.0036203900817781687, + 0.005291300360113382, + 0.004393938463181257, + 0.005712810903787613, + 0.009049979969859123, + 0.01094699278473854, + 0.009060012176632881, + 0.017983464524149895, + 0.005261354148387909, + 0.008507279679179192, + 0.005783687345683575, + 0.017842721194028854, + 0.019496463239192963, + 0.05329924076795578, + 0.06364760547876358, + 0.008995244279503822, + 0.007757158018648624, + 0.005610760301351547, + 0.00608490314334631, + 0.01055894698947668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.07899759709835052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002504169940948 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002504169940948, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002504169940948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17754784226417542, + 0.09111534804105759, + 0.040481872856616974, + 0.10572607815265656, + 0.07521213591098785, + 0.027626143768429756, + 0.1289319396018982, + 0.051492881029844284, + 0.05328095331788063, + 0.2485847771167755 + ], + "credit_map": [ + 0.012991918250918388, + 0.01084317546337843, + 0.03086988441646099, + 0.03720388934016228, + 0.022444402799010277, + 0.010785802267491817, + 0.03213274106383324, + 0.020276028662919998, + 0.005089553073048592, + 0.004132323898375034, + 0.018565304577350616, + 0.016350841149687767, + 0.011254757642745972, + 0.006417117081582546, + 0.017813095822930336, + 0.011492354795336723, + 0.0023677696008235216, + 0.002680912148207426, + 0.006946175824850798, + 0.004898402374237776, + 0.0067274309694767, + 0.005730029661208391, + 0.005909004248678684, + 0.005222151521593332, + 0.0064490619115531445, + 0.012200708501040936, + 0.01414456032216549, + 0.009821380488574505, + 0.01724153757095337, + 0.011951994150876999, + 0.02463129349052906, + 0.009285558946430683, + 0.013341541402041912, + 0.010225260630249977, + 0.005916392430663109, + 0.007926811464130878, + 0.004583558067679405, + 0.010095421224832535, + 0.006456803530454636, + 0.016666356474161148, + 0.0029941953253000975, + 0.0023169650230556726, + 0.00517465453594923, + 0.0031687442678958178, + 0.0029931501485407352, + 0.004012182354927063, + 0.0029827405232936144, + 0.003983513452112675, + 0.01302420161664486, + 0.012378731742501259, + 0.010318410582840443, + 0.02726840227842331, + 0.020339172333478928, + 0.008434442803263664, + 0.018293922767043114, + 0.01887466572225094, + 0.007012334652245045, + 0.008000210858881474, + 0.006247298792004585, + 0.008049080148339272, + 0.004444357473403215, + 0.005952800624072552, + 0.0057882750406861305, + 0.005998526234179735, + 0.006672735791653395, + 0.0072463881224393845, + 0.005531503818929195, + 0.011306241154670715, + 0.004161672666668892, + 0.005786378402262926, + 0.003817789489403367, + 0.008758246898651123, + 0.027286384254693985, + 0.0756717100739479, + 0.08928185701370239, + 0.013748914003372192, + 0.009721040725708008, + 0.008767353370785713, + 0.011501068249344826, + 0.012606477364897728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.07002504169940948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand wearing a soft black leather glove gripping an ornate crossguard of a longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sword drawn slowly and deliberately from a dark wooden scabbard inlaid with polished metalwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rows of mahogany bookshelves filled with thick leather-bound books visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05715705826878548, + 0.061036378145217896, + 0.028114991262555122, + 0.039588265120983124, + 0.05880606919527054, + 0.02596130222082138, + 0.11715521663427353, + 0.04937318339943886, + 0.13713675737380981, + 0.4256707727909088 + ], + "credit_map": [ + 0.004664186388254166, + 0.0037773032672703266, + 0.0096198208630085, + 0.01123584434390068, + 0.007118172012269497, + 0.0031831900123506784, + 0.00958021730184555, + 0.00797832477837801, + 0.002963049104437232, + 0.0051018474623560905, + 0.011400778777897358, + 0.014843101613223553, + 0.008080492727458477, + 0.003868039697408676, + 0.009686704725027084, + 0.005092362407594919, + 0.002074912656098604, + 0.0018071795348078012, + 0.005087864585220814, + 0.004158752970397472, + 0.003388711716979742, + 0.0035979480016976595, + 0.004504334647208452, + 0.0034952862188220024, + 0.0026333739515393972, + 0.005308056250214577, + 0.0049773817881941795, + 0.004481959622353315, + 0.0070461309514939785, + 0.00492847990244627, + 0.00480430806055665, + 0.005408576689660549, + 0.007395456545054913, + 0.0074479905888438225, + 0.005120856687426567, + 0.007986113429069519, + 0.00570994708687067, + 0.008859487250447273, + 0.007013284135609865, + 0.009272933006286621, + 0.005464952904731035, + 0.00332573801279068, + 0.0037806807085871696, + 0.0023908675648272038, + 0.0024671258870512247, + 0.0031934946309775114, + 0.001819174038246274, + 0.0035192666109651327, + 0.013779284432530403, + 0.015198185108602047, + 0.012554549612104893, + 0.014896837994456291, + 0.009093326516449451, + 0.011976893991231918, + 0.009784545749425888, + 0.029871592298150063, + 0.007279932964593172, + 0.006567166652530432, + 0.0045052082277834415, + 0.005295159760862589, + 0.009125404991209507, + 0.004868451971560717, + 0.005436235573142767, + 0.006295621860772371, + 0.015549979172647, + 0.0170292966067791, + 0.013496100902557373, + 0.022148678079247475, + 0.009545690380036831, + 0.017351102083921432, + 0.008587280288338661, + 0.0334286130964756, + 0.0454326830804348, + 0.12347935140132904, + 0.15037275850772858, + 0.02832077257335186, + 0.015492104925215244, + 0.016865722835063934, + 0.019150642678141594, + 0.026556706055998802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.018178876489400864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_tensor([6237, 6237])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_tensor([6237, 6237])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a79515ad23aaaa75e74295d6daecc7a9c8aa1f3d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_tensor([6237, 6237])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, surreal macro shot begins with a static frame and slowly tilts downward throughout the sequence, following the journey of water through the pour-over coffee process. The scene begins with a pristine bed of coarse dark roasted coffee grounds in a white, ribbed paper filter. A single drop of shimmering, scalding, clear water breaks free from the unseen spout of the gooseneck kettle and falls, its descent captured in delicate slow motion. Upon impact, it creates a satisfyingly small crater, and the amber grounds instantly darken as it is absorbed. The camera carefully follows the water as it steadily and deliberately dehydrates downward, soaking the paper and causing the grounds to slightly swell. The downward tilt ends at the very tip of the now-saturated, dark brown filter cone, where a viscous drop of rich, opaque black coffee forms, elongating under the pull of gravity before finally separating. It falls into the clear glass bottle below, creating subtle, expanding ripples on the surface of the black liquid. The entire photorealistic scene is bathed in soft, directional morning light, creating brilliant mirrored highlights on the water and glass, enhancing textures and evoking an atmosphere of tranquility, concentration, and contemplation.", + "seed": 3221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3221_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10927394283935428, + "vlm_logprob": -0.10927394283935428, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.10927394283935428, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01843579299747944 + ], + "opposite_logprob_tokens": [ + -4.768435955047607 + ], + "logprob_reward": -0.01843579299747944, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01843579299747944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08812690526247025, + 0.034197814762592316, + 0.06336036324501038, + 0.0768226683139801, + 0.3165781497955322, + 0.1277039647102356, + 0.03913575038313866, + 0.06850606948137283, + 0.09340853989124298, + 0.09215976297855377 + ], + "credit_map": [ + 0.013820546679198742, + 0.014674202539026737, + 0.010067759081721306, + 0.008960873819887638, + 0.015268439427018166, + 0.007962584495544434, + 0.008571697399020195, + 0.008800803683698177, + 0.004640977364033461, + 0.007999352179467678, + 0.0022157817147672176, + 0.0037255454808473587, + 0.004584842827171087, + 0.002977418014779687, + 0.005125813186168671, + 0.0029280816670507193, + 0.0061916783452034, + 0.006757569499313831, + 0.0057233055122196674, + 0.009193791076540947, + 0.009581947699189186, + 0.006338433362543583, + 0.012453571893274784, + 0.007120066788047552, + 0.004401001147925854, + 0.012960991822183132, + 0.006403176113963127, + 0.010838937014341354, + 0.01729751192033291, + 0.00912222359329462, + 0.01007460243999958, + 0.005724222864955664, + 0.0227227583527565, + 0.026433279737830162, + 0.02120695821940899, + 0.03786270692944527, + 0.026456719264388084, + 0.06397058069705963, + 0.04893110692501068, + 0.06899406015872955, + 0.011112634092569351, + 0.019163981080055237, + 0.009831078350543976, + 0.0120781809091568, + 0.028952648863196373, + 0.014401169493794441, + 0.02050403691828251, + 0.011660252697765827, + 0.003829645924270153, + 0.00668437359854579, + 0.002244194271042943, + 0.005039839539676905, + 0.0060824244283139706, + 0.0076723224483430386, + 0.004484446253627539, + 0.003098504152148962, + 0.005271759815514088, + 0.018915284425020218, + 0.004754203837364912, + 0.005166233982890844, + 0.015170237049460411, + 0.008148450404405594, + 0.005760902538895607, + 0.0053189960308372974, + 0.011346669867634773, + 0.01277210470288992, + 0.007299782708287239, + 0.01161256991326809, + 0.01560247503221035, + 0.01362406462430954, + 0.011840276420116425, + 0.00931059755384922, + 0.011174528859555721, + 0.009142047725617886, + 0.00792348850518465, + 0.012057694606482983, + 0.017814835533499718, + 0.009040125645697117, + 0.017119212076067924, + 0.00788783561438322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.01843579299747944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742666482925415 + ], + "opposite_logprob_tokens": [ + -0.9742666482925415 + ], + "logprob_reward": -0.4742666482925415, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742666482925415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09198575466871262, + 0.048613276332616806, + 0.08420303463935852, + 0.2620755732059479, + 0.10162319988012314, + 0.054255351424217224, + 0.09132779389619827, + 0.12879784405231476, + 0.08179113268852234, + 0.05532700568437576 + ], + "credit_map": [ + 0.013385352678596973, + 0.01447110902518034, + 0.010616165585815907, + 0.00963178277015686, + 0.010298171080648899, + 0.0076646581292152405, + 0.018649496138095856, + 0.007269022520631552, + 0.0061901649460196495, + 0.009244651533663273, + 0.004264844581484795, + 0.0058210548013448715, + 0.006235126405954361, + 0.006269468925893307, + 0.006641005631536245, + 0.003946959041059017, + 0.006375271826982498, + 0.007798981852829456, + 0.01047018263489008, + 0.01439510378986597, + 0.010949465446174145, + 0.009193702600896358, + 0.016512714326381683, + 0.00850761029869318, + 0.012960250489413738, + 0.05180954188108444, + 0.013977116905152798, + 0.03207894414663315, + 0.07437995821237564, + 0.01889783889055252, + 0.040233638137578964, + 0.017738284543156624, + 0.009095955640077591, + 0.01383956428617239, + 0.006764900404959917, + 0.011628558859229088, + 0.010969886556267738, + 0.01977449096739292, + 0.014139163307845592, + 0.015410677529871464, + 0.0055434792302548885, + 0.00910341925919056, + 0.004382118582725525, + 0.0067216139286756516, + 0.008682162500917912, + 0.0064667812548577785, + 0.0082322396337986, + 0.00512353889644146, + 0.007934764958918095, + 0.013833161443471909, + 0.007847566157579422, + 0.012486572377383709, + 0.015807760879397392, + 0.017330477014183998, + 0.008261253125965595, + 0.007826240733265877, + 0.014554928988218307, + 0.023025551810860634, + 0.011922599747776985, + 0.011473516933619976, + 0.023069828748703003, + 0.01941259577870369, + 0.012998722493648529, + 0.012340113520622253, + 0.006802545394748449, + 0.01237842720001936, + 0.006986974738538265, + 0.01260281540453434, + 0.015119008719921112, + 0.010673854500055313, + 0.009970010258257389, + 0.0072574918158352375, + 0.0057175038382411, + 0.00522176269441843, + 0.0051646181382238865, + 0.006669916212558746, + 0.009465598501265049, + 0.010012775659561157, + 0.007876148447394371, + 0.005198682192713022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.4742666482925415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648592788726091 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648592788726091, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648592788726091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14644257724285126, + 0.10226306319236755, + 0.18399977684020996, + 0.059610284864902496, + 0.051466040313243866, + 0.24069181084632874, + 0.04482317715883255, + 0.050139449536800385, + 0.04591180384159088, + 0.07465194910764694 + ], + "credit_map": [ + 0.020463982596993446, + 0.024887559935450554, + 0.01136772707104683, + 0.01530888769775629, + 0.021388404071331024, + 0.01670854724943638, + 0.022592909634113312, + 0.013724572025239468, + 0.010105745866894722, + 0.021167421713471413, + 0.0071610151790082455, + 0.013243474997580051, + 0.017114214599132538, + 0.010550279170274734, + 0.015656065195798874, + 0.007264848332852125, + 0.01813105121254921, + 0.029629886150360107, + 0.01632106490433216, + 0.022288164123892784, + 0.025069136172533035, + 0.02272816188633442, + 0.02639896795153618, + 0.02343336120247841, + 0.003805895335972309, + 0.011457235552370548, + 0.003893461776897311, + 0.007605839055031538, + 0.013200799934566021, + 0.006406708154827356, + 0.007694755680859089, + 0.005545596592128277, + 0.005400811787694693, + 0.010166007094085217, + 0.003763089422136545, + 0.006367288995534182, + 0.007839098572731018, + 0.00600092951208353, + 0.006506996229290962, + 0.005421821493655443, + 0.015574402175843716, + 0.045438122004270554, + 0.01602061465382576, + 0.02306504175066948, + 0.06023896113038063, + 0.02421855553984642, + 0.04180295765399933, + 0.014333188533782959, + 0.004791564308106899, + 0.007814702577888966, + 0.0031093975994735956, + 0.005921013653278351, + 0.006857068277895451, + 0.00791255384683609, + 0.004635421093553305, + 0.00378145813010633, + 0.003925708122551441, + 0.012013393454253674, + 0.003504229011014104, + 0.004500643815845251, + 0.010843902826309204, + 0.006215198896825314, + 0.004412385635077953, + 0.004723990801721811, + 0.005861073732376099, + 0.0070459856651723385, + 0.004017310682684183, + 0.006405384745448828, + 0.007456521969288588, + 0.004214941989630461, + 0.007361576892435551, + 0.0035490121226757765, + 0.008723444305360317, + 0.007520284503698349, + 0.006768918130546808, + 0.009547906927764416, + 0.013410805724561214, + 0.005600541364401579, + 0.017334500327706337, + 0.005745554342865944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.004648592788726091 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956991575658321 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.005956991575658321, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956991575658321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24971801042556763, + 0.0793585330247879, + 0.050003327429294586, + 0.035154763609170914, + 0.048393383622169495, + 0.21369318664073944, + 0.033718302845954895, + 0.03855738416314125, + 0.05046343430876732, + 0.20093971490859985 + ], + "credit_map": [ + 0.03560679033398628, + 0.04851904138922691, + 0.016893265768885612, + 0.028525415807962418, + 0.03366175293922424, + 0.0255393385887146, + 0.03474908322095871, + 0.02622331865131855, + 0.011632353067398071, + 0.016392314806580544, + 0.007287006825208664, + 0.00859733298420906, + 0.00963542889803648, + 0.009831741452217102, + 0.009669718332588673, + 0.00631262781098485, + 0.004423164296895266, + 0.0069360327906906605, + 0.005218335893005133, + 0.006853007711470127, + 0.008390570059418678, + 0.005709419492632151, + 0.0077765630558133125, + 0.004696229472756386, + 0.0028730989433825016, + 0.005064474884420633, + 0.004082784987986088, + 0.0058592380955815315, + 0.005100801587104797, + 0.0048799836076796055, + 0.004188929218798876, + 0.0031054469291120768, + 0.004353493917733431, + 0.0067072222009301186, + 0.0033551405649632215, + 0.005961624905467033, + 0.005453948862850666, + 0.005583472549915314, + 0.007437099236994982, + 0.009541374631226063, + 0.01860773004591465, + 0.04003240168094635, + 0.019324608147144318, + 0.02710748463869095, + 0.03577735275030136, + 0.02558503858745098, + 0.030585721135139465, + 0.016672829166054726, + 0.002488735131919384, + 0.006068931892514229, + 0.0018748085713014007, + 0.0032984197605401278, + 0.00555466627702117, + 0.0081843426451087, + 0.0034216311760246754, + 0.0028267623856663704, + 0.002047407440841198, + 0.013324858620762825, + 0.0016811358509585261, + 0.002233463106676936, + 0.010792144574224949, + 0.0032734833657741547, + 0.0027781652752310038, + 0.002426720457151532, + 0.009532079100608826, + 0.006907281465828419, + 0.004578270483762026, + 0.006234502885490656, + 0.007842396385967731, + 0.00454974127933383, + 0.0065726484172046185, + 0.004246507305651903, + 0.020120101049542427, + 0.017933489754796028, + 0.019736962392926216, + 0.027568012475967407, + 0.03586779534816742, + 0.02713046409189701, + 0.035948172211647034, + 0.016634704545140266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.005956991575658321 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306168854236603 + ], + "opposite_logprob_tokens": [ + -3.1680617332458496 + ], + "logprob_reward": -0.04306168854236603, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306168854236603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09499673545360565, + 0.05525924637913704, + 0.06886204332113266, + 0.046653863042593, + 0.05930320918560028, + 0.1664218008518219, + 0.06181187927722931, + 0.11517798900604248, + 0.25188833475112915, + 0.07962489873170853 + ], + "credit_map": [ + 0.011584759689867496, + 0.013428324833512306, + 0.008251266553997993, + 0.009893076494336128, + 0.01230533979833126, + 0.009263857267796993, + 0.01957760937511921, + 0.010692507028579712, + 0.007474600803107023, + 0.010061230510473251, + 0.004166371654719114, + 0.006538920104503632, + 0.008752553723752499, + 0.006825156044214964, + 0.00658431276679039, + 0.004856106359511614, + 0.00702108908444643, + 0.007561733014881611, + 0.0056371307000517845, + 0.010551505722105503, + 0.008870275691151619, + 0.00856553204357624, + 0.011757674627006054, + 0.008897113613784313, + 0.003259244142100215, + 0.007996603846549988, + 0.0037139037158340216, + 0.007688585668802261, + 0.008788923732936382, + 0.005186084657907486, + 0.005818459205329418, + 0.004202059470117092, + 0.003969918005168438, + 0.007808417081832886, + 0.005871180910617113, + 0.0074642617255449295, + 0.006753518246114254, + 0.010881680063903332, + 0.008164049126207829, + 0.008390188217163086, + 0.011339242570102215, + 0.027425481006503105, + 0.012624071910977364, + 0.018218105658888817, + 0.039127498865127563, + 0.020274832844734192, + 0.0252445749938488, + 0.012167994864284992, + 0.0057325162924826145, + 0.008185524493455887, + 0.005830422043800354, + 0.008317505940794945, + 0.010425953194499016, + 0.011130849830806255, + 0.006946586538106203, + 0.005242526531219482, + 0.00827360711991787, + 0.013825591653585434, + 0.0095992311835289, + 0.009805877692997456, + 0.019805312156677246, + 0.022856127470731735, + 0.01663077436387539, + 0.0143814692273736, + 0.028490405529737473, + 0.03168485313653946, + 0.0200465340167284, + 0.032693877816200256, + 0.04219362512230873, + 0.03431105986237526, + 0.03862687200307846, + 0.023841116577386856, + 0.00799538567662239, + 0.0070947082713246346, + 0.007949845865368843, + 0.00902166310697794, + 0.017847441136837006, + 0.008798173628747463, + 0.013470793142914772, + 0.007446897216141224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.04306168854236603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, surreal macro shot begins with a static frame and slowly tilts downward throughout the sequence, following the journey of water through the pour-over coffee process. The scene begins with a pristine bed of coarse dark roasted coffee grounds in a white, ribbed paper filter. A single drop of shimmering, scalding, clear water breaks free from the unseen spout of the gooseneck kettle and falls, its descent captured in delicate slow motion. Upon impact, it creates a satisfyingly small crater, and the amber grounds instantly darken as it is absorbed. The camera carefully follows the water as it steadily and deliberately dehydrates downward, soaking the paper and causing the grounds to slightly swell. The downward tilt ends at the very tip of the now-saturated, dark brown filter cone, where a viscous drop of rich, opaque black coffee forms, elongating under the pull of gravity before finally separating. It falls into the clear glass bottle below, creating subtle, expanding ripples on the surface of the black liquid. The entire photorealistic scene is bathed in soft, directional morning light, creating brilliant mirrored highlights on the water and glass, enhancing textures and evoking an atmosphere of tranquility, concentration, and contemplation.", + "seed": 3222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3222_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.048184650391340254, + "vlm_logprob": -0.048184650391340254, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.048184650391340254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015052071772515774 + ], + "opposite_logprob_tokens": [ + -5.26505184173584 + ], + "logprob_reward": -0.015052071772515774, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.015052071772515774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.404563844203949, + 0.06678099185228348, + 0.07888932526111603, + 0.053752221167087555, + 0.1574573814868927, + 0.05388486757874489, + 0.028150545433163643, + 0.05704457312822342, + 0.04336448013782501, + 0.05611180141568184 + ], + "credit_map": [ + 0.0345502607524395, + 0.0472215861082077, + 0.02292472869157791, + 0.06108812242746353, + 0.05004023015499115, + 0.034496892243623734, + 0.055948272347450256, + 0.09829375892877579, + 0.004320402629673481, + 0.013371595181524754, + 0.006630888674408197, + 0.006560266483575106, + 0.0157739520072937, + 0.0077859205193817616, + 0.005762697197496891, + 0.00657526683062315, + 0.004658462479710579, + 0.016046134755015373, + 0.007131448946893215, + 0.011096736416220665, + 0.015405047684907913, + 0.009509810246527195, + 0.0076922085136175156, + 0.007349479012191296, + 0.004312196746468544, + 0.009441189467906952, + 0.005056878551840782, + 0.008782525546848774, + 0.006638091988861561, + 0.0056086149998009205, + 0.00824692565947771, + 0.00566580006852746, + 0.011275879107415676, + 0.017421305179595947, + 0.03308859094977379, + 0.009518261067569256, + 0.05098118260502815, + 0.011584888212382793, + 0.01583704724907875, + 0.007750237826257944, + 0.00401161378249526, + 0.010777086019515991, + 0.00299649010412395, + 0.004734470043331385, + 0.014560496434569359, + 0.007100994233042002, + 0.005233084782958031, + 0.004470628220587969, + 0.0026477400679141283, + 0.005297607742249966, + 0.0029544627759605646, + 0.004778285510838032, + 0.003827332751825452, + 0.002397349337115884, + 0.002920419443398714, + 0.0033273512963205576, + 0.010522004216909409, + 0.007873849011957645, + 0.0056459978222846985, + 0.0074764396995306015, + 0.006927747745066881, + 0.0058212680742144585, + 0.005904700141400099, + 0.0068725645542144775, + 0.003557745600119233, + 0.005016202572733164, + 0.004215453285723925, + 0.005177467130124569, + 0.005383903626352549, + 0.0035045105032622814, + 0.00838727317750454, + 0.008121926337480545, + 0.00489156786352396, + 0.008705471642315388, + 0.004593654070049524, + 0.006614990066736937, + 0.007719157263636589, + 0.009000848978757858, + 0.009423921816051006, + 0.005162192974239588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.015052071772515774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988866567611694 + ], + "opposite_logprob_tokens": [ + -1.8048887252807617 + ], + "logprob_reward": -0.17988866567611694, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988866567611694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40964239835739136, + 0.052141834050416946, + 0.047917913645505905, + 0.1894516944885254, + 0.13361190259456635, + 0.04970202222466469, + 0.040901873260736465, + 0.026522045955061913, + 0.03288911655545235, + 0.017219178378582 + ], + "credit_map": [ + 0.03413308411836624, + 0.04864807426929474, + 0.022681545466184616, + 0.06302196532487869, + 0.0534144751727581, + 0.03452495113015175, + 0.05676082521677017, + 0.09645748138427734, + 0.004986680578440428, + 0.00790340080857277, + 0.006738652475178242, + 0.007942033931612968, + 0.006484805140644312, + 0.00484914006665349, + 0.007642305456101894, + 0.005594812333583832, + 0.0025993376038968563, + 0.006014981772750616, + 0.004361960105597973, + 0.005238011479377747, + 0.008328249678015709, + 0.007256787270307541, + 0.008567515760660172, + 0.005551070906221867, + 0.009687039069831371, + 0.013515076600015163, + 0.01487073116004467, + 0.044629015028476715, + 0.011570894159376621, + 0.021184705197811127, + 0.06008785590529442, + 0.013906382955610752, + 0.007669038604944944, + 0.01231041643768549, + 0.032881539314985275, + 0.008246788755059242, + 0.04277385026216507, + 0.01075014378875494, + 0.010748323984444141, + 0.00823180191218853, + 0.0035044923424720764, + 0.010533411055803299, + 0.0033874446526169777, + 0.0039769248105585575, + 0.013380927965044975, + 0.004552803933620453, + 0.0047685676254332066, + 0.005597446579486132, + 0.0044109332375228405, + 0.007486009504646063, + 0.004406217020004988, + 0.006828599609434605, + 0.006981963291764259, + 0.0033420559484511614, + 0.0033892043866217136, + 0.004056889563798904, + 0.0029293515253812075, + 0.0033796741627156734, + 0.0024120465386658907, + 0.004545372445136309, + 0.004082078114151955, + 0.0023587248288094997, + 0.00333101162686944, + 0.003483787877485156, + 0.004095201846212149, + 0.005361819639801979, + 0.0035146900918334723, + 0.003667729441076517, + 0.005111140664666891, + 0.00262061576358974, + 0.005199596751481295, + 0.003318323055282235, + 0.0019802351016551256, + 0.0020523008424788713, + 0.002013644902035594, + 0.0019341191509738564, + 0.0025085455272346735, + 0.00253704353235662, + 0.0024422237183898687, + 0.001751066418364644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17988866567611694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646219778805971 + ], + "opposite_logprob_tokens": [ + -5.379646301269531 + ], + "logprob_reward": -0.004646219778805971, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646219778805971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15219125151634216, + 0.061777275055646896, + 0.06329000741243362, + 0.11606202274560928, + 0.14465129375457764, + 0.3032771646976471, + 0.03659031540155411, + 0.0492887906730175, + 0.034386783838272095, + 0.038485076278448105 + ], + "credit_map": [ + 0.021875077858567238, + 0.01769477128982544, + 0.008555270731449127, + 0.020308036357164383, + 0.021518370136618614, + 0.014569243416190147, + 0.019542740657925606, + 0.028127754107117653, + 0.004376648459583521, + 0.012664246372878551, + 0.0054631163366138935, + 0.006093870382755995, + 0.013523011468350887, + 0.008267234079539776, + 0.005512714385986328, + 0.005876434035599232, + 0.004031735938042402, + 0.014065065421164036, + 0.004923434462398291, + 0.006741137709468603, + 0.014669610187411308, + 0.008369199000298977, + 0.005078399553894997, + 0.005411428865045309, + 0.009967428632080555, + 0.01587478630244732, + 0.0189813245087862, + 0.01733933947980404, + 0.01385729480534792, + 0.013975038193166256, + 0.016989776864647865, + 0.009077031165361404, + 0.013666540384292603, + 0.021714847534894943, + 0.031496062874794006, + 0.015143051743507385, + 0.016578275710344315, + 0.013597941026091576, + 0.023001329973340034, + 0.00945323333144188, + 0.022838717326521873, + 0.055692240595817566, + 0.009101311676204205, + 0.012241465970873833, + 0.10985224694013596, + 0.027018576860427856, + 0.014484104700386524, + 0.052048515528440475, + 0.002818017965182662, + 0.007086531724780798, + 0.003001659642904997, + 0.0069444142282009125, + 0.005129783879965544, + 0.0023707093205302954, + 0.005422530695796013, + 0.0038166700396686792, + 0.005518798250705004, + 0.006449139676988125, + 0.0053659905679523945, + 0.007526891306042671, + 0.00816280022263527, + 0.004621561616659164, + 0.005854678340256214, + 0.005788928363472223, + 0.004196548834443092, + 0.006771965883672237, + 0.0032495956402271986, + 0.004357024095952511, + 0.0045695314183831215, + 0.002643946325406432, + 0.004316734150052071, + 0.0042814346961677074, + 0.0036112647503614426, + 0.004506140481680632, + 0.003559170989319682, + 0.006045532878488302, + 0.005338487681001425, + 0.005621518008410931, + 0.00598911615088582, + 0.003813845105469227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004646219778805971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638887036591768 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007638887036591768, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638887036591768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28037264943122864, + 0.08300796151161194, + 0.06345271319150925, + 0.04166519641876221, + 0.03377842903137207, + 0.03537948057055473, + 0.031523898243904114, + 0.36167457699775696, + 0.016069503501057625, + 0.05307559669017792 + ], + "credit_map": [ + 0.028707977384328842, + 0.03503795713186264, + 0.02157733589410782, + 0.040963008999824524, + 0.02864646166563034, + 0.02278415858745575, + 0.047333069145679474, + 0.055322661995887756, + 0.006593654863536358, + 0.013036154210567474, + 0.00921791885048151, + 0.009032823145389557, + 0.014174623414874077, + 0.010774733498692513, + 0.011590423993766308, + 0.008587619289755821, + 0.004328825511038303, + 0.013585356064140797, + 0.004811782855540514, + 0.005571193993091583, + 0.009684967808425426, + 0.014187864027917385, + 0.005469136405736208, + 0.005813580472022295, + 0.003005373990163207, + 0.003970067948102951, + 0.005812766496092081, + 0.007636433467268944, + 0.004075393080711365, + 0.005301392171531916, + 0.008470557630062103, + 0.003393209306523204, + 0.0022420277819037437, + 0.003753915661945939, + 0.002372672315686941, + 0.0038777620065957308, + 0.0035854752641171217, + 0.0033857307862490416, + 0.00881713442504406, + 0.005743708461523056, + 0.0033412363845854998, + 0.005868350155651569, + 0.0027140469755977392, + 0.004792987369000912, + 0.006712134927511215, + 0.003727594157680869, + 0.0037819468416273594, + 0.004441182129085064, + 0.0031455480493605137, + 0.005825944244861603, + 0.002757162554189563, + 0.005308898631483316, + 0.005213514436036348, + 0.002880767220631242, + 0.0039402577094733715, + 0.0024518033023923635, + 0.03607933223247528, + 0.04516899585723877, + 0.03952357545495033, + 0.0582166388630867, + 0.046308405697345734, + 0.03301610052585602, + 0.06358885020017624, + 0.03977266699075699, + 0.002274795901030302, + 0.002826640848070383, + 0.0017550886841490865, + 0.0023473354522138834, + 0.0021099871955811977, + 0.0012805816950276494, + 0.0019784625619649887, + 0.0014966102316975594, + 0.0064522079192101955, + 0.008731802925467491, + 0.00479478295892477, + 0.006571026984602213, + 0.009109687991440296, + 0.004278212320059538, + 0.008280483074486256, + 0.00485738692805171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.007638887036591768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369740769267082 + ], + "opposite_logprob_tokens": [ + -3.4086973667144775 + ], + "logprob_reward": -0.03369740769267082, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369740769267082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26013100147247314, + 0.1082911491394043, + 0.050952132791280746, + 0.04117342084646225, + 0.15074311196804047, + 0.08539289981126785, + 0.03460412472486496, + 0.12101641297340393, + 0.054841961711645126, + 0.09285381436347961 + ], + "credit_map": [ + 0.02706259675323963, + 0.039710044860839844, + 0.017526574432849884, + 0.03311034291982651, + 0.04414790868759155, + 0.02431240864098072, + 0.03496294096112251, + 0.03929818794131279, + 0.01106418576091528, + 0.01539033092558384, + 0.012024009600281715, + 0.011290493421256542, + 0.021597102284431458, + 0.010428227484226227, + 0.013572612777352333, + 0.012924187816679478, + 0.002978773321956396, + 0.005253252107650042, + 0.004218165297061205, + 0.007309694308787584, + 0.010383808985352516, + 0.01051347516477108, + 0.005901247262954712, + 0.004393719602376223, + 0.0025817472487688065, + 0.0035712793469429016, + 0.004178117960691452, + 0.008581375703215599, + 0.004160238429903984, + 0.005042640957981348, + 0.009547977708280087, + 0.00351004209369421, + 0.009517496451735497, + 0.011237076483666897, + 0.03997368738055229, + 0.012840806506574154, + 0.024606866762042046, + 0.009856113232672215, + 0.0348876491189003, + 0.007823417894542217, + 0.005737849045544863, + 0.0135792912915349, + 0.004879573360085487, + 0.00838247500360012, + 0.024649523198604584, + 0.008106227032840252, + 0.00729736452922225, + 0.012760600075125694, + 0.0022166988346725702, + 0.007049998734146357, + 0.0035254640970379114, + 0.0072454363107681274, + 0.004639987833797932, + 0.0022534874733537436, + 0.004625400993973017, + 0.003047648584470153, + 0.014928190968930721, + 0.016466662287712097, + 0.014916366897523403, + 0.016313297674059868, + 0.014137400314211845, + 0.015156052075326443, + 0.014558680355548859, + 0.014539760537445545, + 0.00510377436876297, + 0.007998389191925526, + 0.0058598509058356285, + 0.0071696205995976925, + 0.009154598228633404, + 0.0035717450082302094, + 0.010388997383415699, + 0.005594990216195583, + 0.009721637703478336, + 0.010916604660451412, + 0.009146088734269142, + 0.016035718843340874, + 0.014953050762414932, + 0.006695957388728857, + 0.015318408608436584, + 0.010066349059343338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03369740769267082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, surreal macro shot begins with a static frame and slowly tilts downward throughout the sequence, following the journey of water through the pour-over coffee process. The scene begins with a pristine bed of coarse dark roasted coffee grounds in a white, ribbed paper filter. A single drop of shimmering, scalding, clear water breaks free from the unseen spout of the gooseneck kettle and falls, its descent captured in delicate slow motion. Upon impact, it creates a satisfyingly small crater, and the amber grounds instantly darken as it is absorbed. The camera carefully follows the water as it steadily and deliberately dehydrates downward, soaking the paper and causing the grounds to slightly swell. The downward tilt ends at the very tip of the now-saturated, dark brown filter cone, where a viscous drop of rich, opaque black coffee forms, elongating under the pull of gravity before finally separating. It falls into the clear glass bottle below, creating subtle, expanding ripples on the surface of the black liquid. The entire photorealistic scene is bathed in soft, directional morning light, creating brilliant mirrored highlights on the water and glass, enhancing textures and evoking an atmosphere of tranquility, concentration, and contemplation.", + "seed": 3223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3223_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09242932666093111, + "vlm_logprob": -0.09242932666093111, + "vlm_yes_no_margin": 3.450000011920929, + "vlm_token_credit": -0.09242932666093111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0240266565233469 + ], + "opposite_logprob_tokens": [ + -4.274026870727539 + ], + "logprob_reward": -0.0240266565233469, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0240266565233469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10081573575735092, + 0.0390017069876194, + 0.03378612920641899, + 0.1939772218465805, + 0.38972899317741394, + 0.05628220736980438, + 0.040681254118680954, + 0.031007390469312668, + 0.05554290488362312, + 0.05917635187506676 + ], + "credit_map": [ + 0.010439906269311905, + 0.01329093612730503, + 0.01317804679274559, + 0.014033397659659386, + 0.012155106291174889, + 0.012068032287061214, + 0.011922130361199379, + 0.0137281883507967, + 0.0034495405852794647, + 0.00506498571485281, + 0.004936445504426956, + 0.006484080106019974, + 0.004430774599313736, + 0.006326804868876934, + 0.0042677088640630245, + 0.004041370470076799, + 0.005560461897403002, + 0.003812821116298437, + 0.003750670002773404, + 0.004420539829879999, + 0.006038798484951258, + 0.004120089579373598, + 0.003000798402354121, + 0.003081951756030321, + 0.017832394689321518, + 0.007299795746803284, + 0.03471275791525841, + 0.030404677614569664, + 0.009789466857910156, + 0.008844003081321716, + 0.07202127575874329, + 0.013072880916297436, + 0.016844088211655617, + 0.019679538905620575, + 0.01755395159125328, + 0.12361956387758255, + 0.03514865040779114, + 0.02329828590154648, + 0.13479334115982056, + 0.01879158429801464, + 0.007021340541541576, + 0.0075255706906318665, + 0.006528572179377079, + 0.007109606638550758, + 0.009120367467403412, + 0.00615535257384181, + 0.007841281592845917, + 0.004980124533176422, + 0.005290214903652668, + 0.008471122942864895, + 0.0031503376085311174, + 0.0043683238327503204, + 0.009528634138405323, + 0.0030416464433073997, + 0.0039213234558701515, + 0.002909656148403883, + 0.0033484913874417543, + 0.004756075795739889, + 0.0037303566932678223, + 0.004519912879914045, + 0.004469903185963631, + 0.003134692320600152, + 0.003713794983923435, + 0.0033341660164296627, + 0.0044064950197935104, + 0.010397974401712418, + 0.010802666656672955, + 0.006808751728385687, + 0.006071151699870825, + 0.005385586526244879, + 0.006312389392405748, + 0.0053578936494886875, + 0.008476022630929947, + 0.006376619450747967, + 0.007467273622751236, + 0.007709653116762638, + 0.007850999012589455, + 0.004684217739850283, + 0.0105371605604887, + 0.006074409931898117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0240266565233469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348625123500824 + ], + "opposite_logprob_tokens": [ + -1.2236251831054688 + ], + "logprob_reward": -0.348625123500824, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.348625123500824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07969354838132858, + 0.04891693592071533, + 0.029920604079961777, + 0.31985393166542053, + 0.366080641746521, + 0.07156994193792343, + 0.023242324590682983, + 0.02600257284939289, + 0.023504815995693207, + 0.011214611120522022 + ], + "credit_map": [ + 0.011821883730590343, + 0.013367779552936554, + 0.010224692523479462, + 0.010270822793245316, + 0.007710837293416262, + 0.008541370742022991, + 0.007760415785014629, + 0.009995760396122932, + 0.004466489423066378, + 0.005393789615482092, + 0.005205832421779633, + 0.007675665896385908, + 0.0067185466177761555, + 0.00609245989471674, + 0.008808724582195282, + 0.0045554363168776035, + 0.003948607016354799, + 0.0038113617338240147, + 0.0024039251729846, + 0.0029135141521692276, + 0.005516148637980223, + 0.004077569581568241, + 0.004144466947764158, + 0.003105015493929386, + 0.02539639361202717, + 0.012005940079689026, + 0.060161348432302475, + 0.0494939386844635, + 0.016778089106082916, + 0.013475283980369568, + 0.12154167890548706, + 0.021001292392611504, + 0.021315161138772964, + 0.01490091159939766, + 0.013785375282168388, + 0.1267378330230713, + 0.020333411172032356, + 0.012537793256342411, + 0.13965076208114624, + 0.01681944541633129, + 0.008313173428177834, + 0.00882613006979227, + 0.009036985225975513, + 0.009931709617376328, + 0.01242283545434475, + 0.00956256315112114, + 0.008869455195963383, + 0.004607104230672121, + 0.0028515940066426992, + 0.003913714550435543, + 0.001655276515521109, + 0.00354958837851882, + 0.0043823858723044395, + 0.002492418047040701, + 0.002108389977365732, + 0.0022889620158821344, + 0.0034746939782053232, + 0.005317474249750376, + 0.0030049115885049105, + 0.0033464364241808653, + 0.0035644331946969032, + 0.0026486932765692472, + 0.002148100407794118, + 0.0024978353176265955, + 0.0027514081448316574, + 0.0027664219960570335, + 0.003045904217287898, + 0.004141767043620348, + 0.003281260374933481, + 0.0023408338893204927, + 0.0027969982475042343, + 0.00238022580742836, + 0.0011188172502443194, + 0.0016149971634149551, + 0.0013682374265044928, + 0.0017362476792186499, + 0.0019702862482517958, + 0.0007287233020178974, + 0.0018289051949977875, + 0.0008483989513479173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.348625123500824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259960889816284 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005259960889816284, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259960889816284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057414710521698, + 0.18545790016651154, + 0.06679436564445496, + 0.5155977606773376, + 0.034174609929323196, + 0.07747697085142136, + 0.010776063427329063, + 0.013575932942330837, + 0.022293325513601303, + 0.01643836311995983 + ], + "credit_map": [ + 0.0058104232884943485, + 0.009691890329122543, + 0.007069526240229607, + 0.009927581995725632, + 0.005593393929302692, + 0.005814670119434595, + 0.005449631251394749, + 0.00805759523063898, + 0.017312893643975258, + 0.024492783471941948, + 0.01700383611023426, + 0.026327328756451607, + 0.02942419797182083, + 0.019999075680971146, + 0.02891012281179428, + 0.021987663581967354, + 0.007379158865660429, + 0.007660465314984322, + 0.005601614248007536, + 0.007145041599869728, + 0.010655410587787628, + 0.011569837108254433, + 0.009812045842409134, + 0.006970790680497885, + 0.03854959458112717, + 0.019092455506324768, + 0.09236988425254822, + 0.08091390132904053, + 0.02669767290353775, + 0.021152133122086525, + 0.20072412490844727, + 0.03609795868396759, + 0.001900933450087905, + 0.0012840243289247155, + 0.0018834314541891217, + 0.011443090625107288, + 0.0019006803631782532, + 0.001318783382885158, + 0.012748142704367638, + 0.0016955252503976226, + 0.010957188904285431, + 0.00879060197621584, + 0.00746543612331152, + 0.009581219404935837, + 0.012404958717525005, + 0.010355440899729729, + 0.011018331162631512, + 0.006903789471834898, + 0.0016417171573266387, + 0.0017104133730754256, + 0.0010866033844649792, + 0.0013236739905551076, + 0.0017663893522694707, + 0.001057033659890294, + 0.0014077319065108895, + 0.0007825004868209362, + 0.0013192477636039257, + 0.0027020133566111326, + 0.0018247737316414714, + 0.0015658922493457794, + 0.0021288027055561543, + 0.0014535450609400868, + 0.0013746644835919142, + 0.001206993474625051, + 0.0022082726936787367, + 0.0035784761421382427, + 0.004004118498414755, + 0.002747626043856144, + 0.0026626880280673504, + 0.002374087693169713, + 0.0022458485327661037, + 0.0024722071830183268, + 0.001767070614732802, + 0.0018132301047444344, + 0.0017956800293177366, + 0.0025171437300741673, + 0.0024086928460747004, + 0.001216130331158638, + 0.0031885476782917976, + 0.0017318690661340952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.005259960889816284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209403656423092 + ], + "opposite_logprob_tokens": [ + -4.264209270477295 + ], + "logprob_reward": -0.014209403656423092, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209403656423092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06405273824930191, + 0.05385909602046013, + 0.07883262634277344, + 0.26058855652809143, + 0.16607625782489777, + 0.050181757658720016, + 0.04629702493548393, + 0.05538933724164963, + 0.09938889741897583, + 0.12533362209796906 + ], + "credit_map": [ + 0.006932953838258982, + 0.00991786364465952, + 0.007085800636559725, + 0.010200193151831627, + 0.007079739589244127, + 0.005946909077465534, + 0.007025728467851877, + 0.009863554500043392, + 0.005300534889101982, + 0.008913126774132252, + 0.0052863736636936665, + 0.007730401586741209, + 0.0076996935531497, + 0.008027032017707825, + 0.004706001374870539, + 0.006195933558046818, + 0.009207568131387234, + 0.008666637353599072, + 0.0065093799494206905, + 0.009975146502256393, + 0.012044391594827175, + 0.01150048989802599, + 0.013600616715848446, + 0.007328395266085863, + 0.020533403381705284, + 0.01020857784897089, + 0.046507179737091064, + 0.04081006720662117, + 0.013255223631858826, + 0.011085347272455692, + 0.10006566345691681, + 0.018123099580407143, + 0.012135391123592854, + 0.011457362212240696, + 0.007983783259987831, + 0.05201726406812668, + 0.01038870308548212, + 0.006085656117647886, + 0.057896047830581665, + 0.008112072944641113, + 0.006075453478842974, + 0.006623482797294855, + 0.006584171671420336, + 0.0072183916345238686, + 0.00890500470995903, + 0.004713722039014101, + 0.0065457639284431934, + 0.003515774616971612, + 0.0053297774866223335, + 0.010287631303071976, + 0.0037526648957282305, + 0.005467693321406841, + 0.0105455182492733, + 0.0032810214906930923, + 0.004403627477586269, + 0.0032290935050696135, + 0.005977054592221975, + 0.011347854509949684, + 0.007238996215164661, + 0.007006197236478329, + 0.008367128670215607, + 0.006128297187387943, + 0.004366636276245117, + 0.004957173485308886, + 0.011134404689073563, + 0.023074105381965637, + 0.02015763707458973, + 0.009433796629309654, + 0.012416534125804901, + 0.00808386318385601, + 0.0075949570164084435, + 0.007493611425161362, + 0.016641581431031227, + 0.015270445495843887, + 0.011318130418658257, + 0.016156382858753204, + 0.022699024528265, + 0.012065676040947437, + 0.018701152876019478, + 0.012481230311095715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.014209403656423092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700254887342453 + ], + "opposite_logprob_tokens": [ + -2.6950254440307617 + ], + "logprob_reward": -0.0700254887342453, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700254887342453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25817734003067017, + 0.05055488273501396, + 0.06268865615129471, + 0.12818403542041779, + 0.1970759481191635, + 0.06018437445163727, + 0.057227496057748795, + 0.03816729784011841, + 0.0711570680141449, + 0.07658293098211288 + ], + "credit_map": [ + 0.023378070443868637, + 0.03927656635642052, + 0.02677113562822342, + 0.041103821247816086, + 0.03807079419493675, + 0.024191100150346756, + 0.03127887472510338, + 0.03410695493221283, + 0.00450351694598794, + 0.006263775750994682, + 0.005869375541806221, + 0.005630739964544773, + 0.008377603255212307, + 0.007516505196690559, + 0.006572891492396593, + 0.005820470862090588, + 0.007269543129950762, + 0.009048127569258213, + 0.00520097091794014, + 0.009892905130982399, + 0.007924123667180538, + 0.009340597316622734, + 0.006870415061712265, + 0.007141966372728348, + 0.0189112089574337, + 0.010643724352121353, + 0.018063189461827278, + 0.016884060576558113, + 0.007727230433374643, + 0.009901388548314571, + 0.01442144624888897, + 0.03163177892565727, + 0.017487946897745132, + 0.02818623185157776, + 0.01612076722085476, + 0.029274756088852882, + 0.04955272376537323, + 0.017125161364674568, + 0.029069552198052406, + 0.010258774273097515, + 0.00664674723520875, + 0.008122052997350693, + 0.008091606199741364, + 0.008382080122828484, + 0.010375935584306717, + 0.006027831230312586, + 0.0075135366059839725, + 0.005024573300033808, + 0.00766784930601716, + 0.008463859558105469, + 0.006305119022727013, + 0.007718222215771675, + 0.00899697095155716, + 0.006944287568330765, + 0.006400052923709154, + 0.004731128457933664, + 0.003546415362507105, + 0.00824227649718523, + 0.0031751201022416353, + 0.004558985587209463, + 0.005796423181891441, + 0.004330563824623823, + 0.005672246217727661, + 0.002845264272764325, + 0.009479080326855183, + 0.010114895179867744, + 0.008660311810672283, + 0.006889448966830969, + 0.011264968663454056, + 0.009565090760588646, + 0.007727956399321556, + 0.0074553107842803, + 0.00758521631360054, + 0.011745743453502655, + 0.009999296627938747, + 0.010324062779545784, + 0.011429890058934689, + 0.007960012182593346, + 0.011437659151852131, + 0.006101035047322512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0700254887342453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, surreal macro shot begins with a static frame and slowly tilts downward throughout the sequence, following the journey of water through the pour-over coffee process. The scene begins with a pristine bed of coarse dark roasted coffee grounds in a white, ribbed paper filter. A single drop of shimmering, scalding, clear water breaks free from the unseen spout of the gooseneck kettle and falls, its descent captured in delicate slow motion. Upon impact, it creates a satisfyingly small crater, and the amber grounds instantly darken as it is absorbed. The camera carefully follows the water as it steadily and deliberately dehydrates downward, soaking the paper and causing the grounds to slightly swell. The downward tilt ends at the very tip of the now-saturated, dark brown filter cone, where a viscous drop of rich, opaque black coffee forms, elongating under the pull of gravity before finally separating. It falls into the clear glass bottle below, creating subtle, expanding ripples on the surface of the black liquid. The entire photorealistic scene is bathed in soft, directional morning light, creating brilliant mirrored highlights on the water and glass, enhancing textures and evoking an atmosphere of tranquility, concentration, and contemplation.", + "seed": 3224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_7/index_6237_seed_3224_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06195530034601689, + "vlm_logprob": -0.06195530034601689, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.06195530034601689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015174307860434055 + ], + "opposite_logprob_tokens": [ + -5.515174388885498 + ], + "logprob_reward": -0.015174307860434055, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.015174307860434055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40456685423851013, + 0.05021532624959946, + 0.017211612313985825, + 0.021822115406394005, + 0.029012802988290787, + 0.05219854786992073, + 0.051718246191740036, + 0.15346889197826385, + 0.13171470165252686, + 0.08807092159986496 + ], + "credit_map": [ + 0.09971031546592712, + 0.1597532331943512, + 0.015945514664053917, + 0.021817296743392944, + 0.039887238293886185, + 0.01532728224992752, + 0.020860804244875908, + 0.03126515820622444, + 0.006027555093169212, + 0.006721199955791235, + 0.009445009753108025, + 0.006406622938811779, + 0.004825798328965902, + 0.00469494191929698, + 0.005858596879988909, + 0.006235601846128702, + 0.0020892044994980097, + 0.0014575942186638713, + 0.0022768941707909107, + 0.0022699160035699606, + 0.0020842445082962513, + 0.0016470192931592464, + 0.002604489913210273, + 0.0027822500560432673, + 0.0014688292285427451, + 0.0033610337413847446, + 0.001885613426566124, + 0.003286341205239296, + 0.003591769840568304, + 0.0025514017324894667, + 0.003047018079087138, + 0.0026301078032702208, + 0.002087211236357689, + 0.004098455421626568, + 0.0038622019346803427, + 0.004660357255488634, + 0.004149597603827715, + 0.002496568253263831, + 0.004259355831891298, + 0.003399054054170847, + 0.004469984211027622, + 0.00874408334493637, + 0.007771487347781658, + 0.00572653952986002, + 0.007736982777714729, + 0.005408782511949539, + 0.007855406031012535, + 0.004485281649976969, + 0.006011251825839281, + 0.00616089953109622, + 0.007460303604602814, + 0.005363417323678732, + 0.007404125761240721, + 0.0075074429623782635, + 0.00663907453417778, + 0.00517172971740365, + 0.028368540108203888, + 0.020443953573703766, + 0.02723648026585579, + 0.012109870091080666, + 0.01754840463399887, + 0.02749190665781498, + 0.008768770843744278, + 0.011500957421958447, + 0.012441386468708515, + 0.013837402686476707, + 0.014087854884564877, + 0.021955519914627075, + 0.020411474630236626, + 0.01643848977982998, + 0.020751798525452614, + 0.011790777556598186, + 0.013424421660602093, + 0.01083388738334179, + 0.007873611524701118, + 0.010022745467722416, + 0.010829899460077286, + 0.02214490808546543, + 0.007379187270998955, + 0.005562261212617159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.015174307860434055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015514224767685 + ], + "opposite_logprob_tokens": [ + -1.7015514373779297 + ], + "logprob_reward": -0.2015514224767685, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015514224767685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2866753041744232, + 0.14615751802921295, + 0.03486895188689232, + 0.028868315741419792, + 0.07116156071424484, + 0.027061015367507935, + 0.10792194306850433, + 0.060024458914995193, + 0.14882756769657135, + 0.08843342214822769 + ], + "credit_map": [ + 0.06612726300954819, + 0.10431859642267227, + 0.012067335657775402, + 0.01873789168894291, + 0.029140502214431763, + 0.013914882205426693, + 0.017263635993003845, + 0.02510516345500946, + 0.012291989289224148, + 0.014311471953988075, + 0.02541303262114525, + 0.014962772838771343, + 0.016423488035798073, + 0.01360036339610815, + 0.01658068224787712, + 0.03257370740175247, + 0.0035865167155861855, + 0.0031783615704625845, + 0.0036692111752927303, + 0.004536224529147148, + 0.004706566222012043, + 0.004627950955182314, + 0.005897803697735071, + 0.004666314925998449, + 0.0030432133935391903, + 0.004693645052611828, + 0.0030732806771993637, + 0.0036193637643009424, + 0.0041804127395153046, + 0.0030884789302945137, + 0.0034992569126188755, + 0.003670662408694625, + 0.005331953056156635, + 0.00817873328924179, + 0.008527901954948902, + 0.01187028456479311, + 0.008359991014003754, + 0.006315057631582022, + 0.013567076064646244, + 0.0090105552226305, + 0.0032999643590301275, + 0.003325584577396512, + 0.003220431972295046, + 0.003498109057545662, + 0.003454165533185005, + 0.003105569863691926, + 0.0038983733393251896, + 0.0032588148023933172, + 0.007933972403407097, + 0.017590411007404327, + 0.013829132542014122, + 0.013072056695818901, + 0.01497675385326147, + 0.01553603820502758, + 0.012478976510465145, + 0.012504602782428265, + 0.010923925787210464, + 0.007484614849090576, + 0.009186545386910439, + 0.005519379395991564, + 0.006550136487931013, + 0.011249097064137459, + 0.0044270968064665794, + 0.004683659411966801, + 0.01556034293025732, + 0.021540442481637, + 0.020586496219038963, + 0.018556291237473488, + 0.029749322682619095, + 0.022732315585017204, + 0.011855137534439564, + 0.008247209712862968, + 0.013601460494101048, + 0.01095557864755392, + 0.007128109224140644, + 0.012717179954051971, + 0.01124513614922762, + 0.017436280846595764, + 0.00833834521472454, + 0.0070113311521708965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.2015514224767685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001518169417977333 + ], + "opposite_logprob_tokens": [ + -6.501518249511719 + ], + "logprob_reward": -0.001518169417977333, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001518169417977333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2927265763282776, + 0.07177277654409409, + 0.12110329419374466, + 0.04481426998972893, + 0.04517684131860733, + 0.04870370402932167, + 0.030073154717683792, + 0.11074623465538025, + 0.09239301830530167, + 0.14249016344547272 + ], + "credit_map": [ + 0.06870561093091965, + 0.1079745963215828, + 0.011664249934256077, + 0.016552666202187538, + 0.030987288802862167, + 0.012857827357947826, + 0.015201902016997337, + 0.028782406821846962, + 0.0068741366267204285, + 0.007492246571928263, + 0.005949190352112055, + 0.006357917096465826, + 0.010961596854031086, + 0.005503844935446978, + 0.008031157776713371, + 0.020602688193321228, + 0.009000666439533234, + 0.018799947574734688, + 0.013787908479571342, + 0.01628454215824604, + 0.01630893535912037, + 0.01962262950837612, + 0.013836939819157124, + 0.01346172858029604, + 0.0033328912686556578, + 0.004013533238321543, + 0.004590602125972509, + 0.0078310277312994, + 0.005507058463990688, + 0.007026800885796547, + 0.007124018389731646, + 0.005388337653130293, + 0.0032405827660113573, + 0.004627535585314035, + 0.00645476346835494, + 0.006428901571780443, + 0.0067182378843426704, + 0.004526571370661259, + 0.007088986691087484, + 0.006091260351240635, + 0.005533246789127588, + 0.005827560555189848, + 0.006408208049833775, + 0.006344884634017944, + 0.0061231511645019054, + 0.005156686529517174, + 0.007545231841504574, + 0.005764733534306288, + 0.003060397459194064, + 0.002968721091747284, + 0.0036519940476864576, + 0.0037202159874141216, + 0.003743474604561925, + 0.003938933368772268, + 0.005338076502084732, + 0.0036513437516987324, + 0.01932760328054428, + 0.01497905608266592, + 0.019629642367362976, + 0.009470625780522823, + 0.011883242055773735, + 0.01917922869324684, + 0.007472465746104717, + 0.008804366923868656, + 0.008263426832854748, + 0.014433109201490879, + 0.008019204251468182, + 0.01269867829978466, + 0.018831636756658554, + 0.011620978824794292, + 0.011980190873146057, + 0.0065457881428301334, + 0.022873371839523315, + 0.019326334819197655, + 0.011968536302447319, + 0.02016901597380638, + 0.01699947938323021, + 0.031002460047602654, + 0.010587967000901699, + 0.009562990628182888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.001518169417977333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255230512470007 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.01255230512470007, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255230512470007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2365993857383728, + 0.05055578425526619, + 0.03273535519838333, + 0.017607133835554123, + 0.06185733526945114, + 0.034842588007450104, + 0.04043853282928467, + 0.35707923769950867, + 0.06502067297697067, + 0.10326395183801651 + ], + "credit_map": [ + 0.05786467343568802, + 0.09228721261024475, + 0.00885303970426321, + 0.013305366039276123, + 0.02334897220134735, + 0.009714020416140556, + 0.01116690319031477, + 0.02005922794342041, + 0.006425119005143642, + 0.004658197518438101, + 0.004547895863652229, + 0.007882579229772091, + 0.004922301042824984, + 0.004557631444185972, + 0.007200649008154869, + 0.010361415334045887, + 0.0030231988057494164, + 0.0024680106434971094, + 0.0037981225177645683, + 0.005158094689249992, + 0.0031532212160527706, + 0.00444070715457201, + 0.005174373276531696, + 0.005519629456102848, + 0.0014002626994624734, + 0.00206857081502676, + 0.0017923411214724183, + 0.0024597826413810253, + 0.0035670683719217777, + 0.002061007544398308, + 0.0021619705948978662, + 0.002096131443977356, + 0.004379012621939182, + 0.006886366289108992, + 0.007388656958937645, + 0.011104071512818336, + 0.007609732914716005, + 0.006844660267233849, + 0.011311221867799759, + 0.006333615630865097, + 0.004303023684769869, + 0.004687963053584099, + 0.004946735687553883, + 0.003872223198413849, + 0.005078213755041361, + 0.004209162667393684, + 0.004580870736390352, + 0.003164395922794938, + 0.0035349626559764147, + 0.004240204580128193, + 0.005957323592156172, + 0.004936734680086374, + 0.0051099397242069244, + 0.005092711187899113, + 0.0061041805893182755, + 0.005462479777634144, + 0.05687788128852844, + 0.0438508503139019, + 0.06668272614479065, + 0.03070744127035141, + 0.03948014974594116, + 0.06869589537382126, + 0.019817430526018143, + 0.030966894701123238, + 0.007336687296628952, + 0.008325657807290554, + 0.008535468019545078, + 0.010671272873878479, + 0.010416783392429352, + 0.008574694395065308, + 0.0060367342084646225, + 0.005123375449329615, + 0.015638038516044617, + 0.014829736202955246, + 0.007828492671251297, + 0.013265617191791534, + 0.013628087006509304, + 0.022320475429296494, + 0.007938052527606487, + 0.007815462537109852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.01255230512470007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a pour-over coffee process?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a static frame and slowly tilt downward throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are coarse dark roasted coffee grounds visible in a white ribbed paper filter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft directional morning light creating mirrored highlights on water and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a drop of water fall from a gooseneck kettle and create ripples in a clear glass bottle below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898029685020447 + ], + "opposite_logprob_tokens": [ + -2.5789802074432373 + ], + "logprob_reward": -0.07898029685020447, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898029685020447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09707099199295044, + 0.0357503667473793, + 0.03340558335185051, + 0.053378451615571976, + 0.057892587035894394, + 0.04260416701436043, + 0.05803908035159111, + 0.17065757513046265, + 0.11502580344676971, + 0.33617544174194336 + ], + "credit_map": [ + 0.012219283729791641, + 0.01515903789550066, + 0.00790059007704258, + 0.008922199718654156, + 0.018684733659029007, + 0.007936512120068073, + 0.010485735721886158, + 0.015762897208333015, + 0.00410145940259099, + 0.004132860340178013, + 0.0031310010235756636, + 0.003927956335246563, + 0.005575475282967091, + 0.003274098038673401, + 0.004675576929003, + 0.006931938696652651, + 0.0019344390602782369, + 0.005153259262442589, + 0.0036052933428436518, + 0.003990925382822752, + 0.004642830230295658, + 0.0051150843501091, + 0.0042169769294559956, + 0.0047467732802033424, + 0.00323423626832664, + 0.00750407949090004, + 0.005372975487262011, + 0.007338446564972401, + 0.007409052923321724, + 0.008320610038936138, + 0.0070828828029334545, + 0.007116170134395361, + 0.003533452283591032, + 0.005080406554043293, + 0.007405752781778574, + 0.009418647736310959, + 0.007825879380106926, + 0.0048965224996209145, + 0.013300245627760887, + 0.006431681104004383, + 0.0048518008552491665, + 0.006320123560726643, + 0.004615985322743654, + 0.00625892449170351, + 0.005427312105894089, + 0.005082315299659967, + 0.006302546709775925, + 0.00374516099691391, + 0.005859856493771076, + 0.007286480627954006, + 0.008124210871756077, + 0.006523854099214077, + 0.009431357495486736, + 0.00817045010626316, + 0.007738265674561262, + 0.004904603585600853, + 0.026430286467075348, + 0.022948144003748894, + 0.03365330025553703, + 0.012436645105481148, + 0.019616978242993355, + 0.029863005504012108, + 0.010694034397602081, + 0.015015175566077232, + 0.008639016188681126, + 0.019622277468442917, + 0.010699629783630371, + 0.01952393352985382, + 0.02005430869758129, + 0.013289843685925007, + 0.015316428616642952, + 0.007880372926592827, + 0.043876294046640396, + 0.045573700219392776, + 0.03664645180106163, + 0.06475003808736801, + 0.04294230416417122, + 0.04730704426765442, + 0.03507057949900627, + 0.02000902220606804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.07898029685020447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_tensor([9462, 9462])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_tensor([9462, 9462])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9eed706821a6f99f7b47552b3e8a9ad830702607 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_tensor([9462, 9462])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward-tilting wide-angle shot reveals the interior of a massive, cylindrical, high-tech greenhouse tower, its polished steel structure and glass walls stretching upward as far as the eye can see. Rows of vertical hydroponic racks spiral upward, densely packed with lush leafy greens—primarily crisp, dew-drenched lettuce and deep, veined kale—a deep emerald hue. The space is bathed in vibrant purple and soft pink glow cast by rows of LED lights, a surreal illumination that makes the greens of the plants appear strikingly vivid. The camera then pans smoothly to the left, following a small, white quadcopter drone as it silently glides along a pre-set magnetic track, displaying perfect stability as it hovers above a row of plants. Then, carefully extending its precision nozzle, it evenly sprays a fine, nutrient-rich mist of water onto the plants' roots. In the foreground walkway, a researcher in a pristine white lab coat stood absorbed in his work, his fingers rapidly sliding and tapping across a translucent blue holographic tablet suspended in mid-air. The tablet displayed a real-time display of glowing green growth graphs and complex data streams. The entire scene blended the tranquility of organic life with the efficiency of sterile technology, creating an atmosphere of focus, cleanliness, and a futuristic feel.", + "seed": 28705, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_9462_seed_28705_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1800164205022156, + "vlm_logprob": -0.1800164205022156, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.1800164205022156, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2557835578918457 + ], + "opposite_logprob_tokens": [ + -1.5057835578918457 + ], + "logprob_reward": -0.2557835578918457, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2557835578918457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06232643872499466, + 0.0899692252278328, + 0.05869963765144348, + 0.38755106925964355, + 0.13847099244594574, + 0.06667018681764603, + 0.08934362232685089, + 0.04800599440932274, + 0.03402075916528702, + 0.024942057207226753 + ], + "credit_map": [ + 0.007129085715860128, + 0.005485232919454575, + 0.005967525765299797, + 0.004101249855011702, + 0.01564795710146427, + 0.01389558706432581, + 0.0038542195688933134, + 0.006245577707886696, + 0.007746429182589054, + 0.006428021006286144, + 0.009931469336152077, + 0.007569981273263693, + 0.022782212123274803, + 0.020777029916644096, + 0.00951844360679388, + 0.005215639714151621, + 0.008027424104511738, + 0.006513338070362806, + 0.006914590485394001, + 0.006920580752193928, + 0.006708499044179916, + 0.009165366180241108, + 0.008158450946211815, + 0.006291386671364307, + 0.10177435725927353, + 0.14810453355312347, + 0.017659684643149376, + 0.02895405702292919, + 0.01903252676129341, + 0.02366625890135765, + 0.015554307959973812, + 0.03280537948012352, + 0.021658800542354584, + 0.02135532908141613, + 0.00995660200715065, + 0.010025021620094776, + 0.011921683326363564, + 0.018319591879844666, + 0.02360113151371479, + 0.0216328464448452, + 0.010922886431217194, + 0.013163421303033829, + 0.007023379672318697, + 0.0060342540964484215, + 0.005888534244149923, + 0.004918805789202452, + 0.0075430311262607574, + 0.011175872758030891, + 0.014730654656887054, + 0.012165810912847519, + 0.01650502160191536, + 0.007078051567077637, + 0.014153266325592995, + 0.008269124664366245, + 0.007368748541921377, + 0.009072939865291119, + 0.0065328278578817844, + 0.005564216990023851, + 0.00573083758354187, + 0.01092342659831047, + 0.004679656121879816, + 0.004714137874543667, + 0.0049781836569309235, + 0.004882706329226494, + 0.003461959771811962, + 0.003935576882213354, + 0.0040002125315368176, + 0.0036945249885320663, + 0.004504234995692968, + 0.005395715590566397, + 0.0037857878487557173, + 0.005242747720330954, + 0.0048081218264997005, + 0.004778696224093437, + 0.0031838833820074797, + 0.0030711141880601645, + 0.0026123991701751947, + 0.0019266011659055948, + 0.001985752023756504, + 0.002575488295406103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.2557835578918457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706653773784637 + ], + "opposite_logprob_tokens": [ + -2.1270666122436523 + ], + "logprob_reward": -0.12706653773784637, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706653773784637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11781210452318192, + 0.0762823298573494, + 0.1169329360127449, + 0.26464778184890747, + 0.11353456228971481, + 0.19517336785793304, + 0.05078663304448128, + 0.02025436982512474, + 0.031365327537059784, + 0.013210594654083252 + ], + "credit_map": [ + 0.011992004700005054, + 0.012295330874621868, + 0.00793512724339962, + 0.008959068916738033, + 0.030655624344944954, + 0.026800762861967087, + 0.008700731210410595, + 0.010473450645804405, + 0.009457038715481758, + 0.007039392367005348, + 0.009248645044863224, + 0.00803526584059, + 0.013225065544247627, + 0.012371637858450413, + 0.009485679678618908, + 0.007419613189995289, + 0.014027315191924572, + 0.011573580093681812, + 0.012660237960517406, + 0.010708795860409737, + 0.010827370919287205, + 0.020133888348937035, + 0.022830883041024208, + 0.01417087484151125, + 0.053963541984558105, + 0.09303855150938034, + 0.016307057812809944, + 0.029203025624155998, + 0.016559500247240067, + 0.022021837532520294, + 0.012326101772487164, + 0.021228153258562088, + 0.02157856896519661, + 0.019071198999881744, + 0.00982798170298338, + 0.00858708843588829, + 0.008468907326459885, + 0.015328475274145603, + 0.012608659453690052, + 0.018063683062791824, + 0.025442255660891533, + 0.03764898702502251, + 0.017451198771595955, + 0.016755258664488792, + 0.0188997033983469, + 0.016379419714212418, + 0.02812846377491951, + 0.034468069672584534, + 0.008460132405161858, + 0.005443960893899202, + 0.01307242177426815, + 0.003729145275428891, + 0.007087723817676306, + 0.004344161134213209, + 0.004647247493267059, + 0.004001840017735958, + 0.0034567515831440687, + 0.0015535046113654971, + 0.002461676951497793, + 0.0023255024570971727, + 0.0019819983281195164, + 0.0037491191178560257, + 0.00245348340831697, + 0.0022723341826349497, + 0.00497035589069128, + 0.003517858451232314, + 0.006460146047174931, + 0.002687464002519846, + 0.003366737859323621, + 0.003955581691116095, + 0.003054628148674965, + 0.003352558007463813, + 0.001208335510455072, + 0.001563151367008686, + 0.002147726248949766, + 0.001684390939772129, + 0.001266295206733048, + 0.0016699270345270634, + 0.0016736382385715842, + 0.0019971297588199377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.12706653773784637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022004456259310246 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.0022004456259310246, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022004456259310246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0937444195151329, + 0.13358858227729797, + 0.10393461585044861, + 0.29183098673820496, + 0.16558098793029785, + 0.08968176692724228, + 0.03455418720841408, + 0.0345357283949852, + 0.015071650967001915, + 0.03747709468007088 + ], + "credit_map": [ + 0.00929864589124918, + 0.013923166319727898, + 0.007134575862437487, + 0.009927374310791492, + 0.01124463602900505, + 0.01596551574766636, + 0.011169977486133575, + 0.015080526471138, + 0.014844309538602829, + 0.006718358490616083, + 0.006970901042222977, + 0.005931641440838575, + 0.04424511641263962, + 0.04096178337931633, + 0.009402504190802574, + 0.004513976164162159, + 0.012647735886275768, + 0.013538765721023083, + 0.013569812290370464, + 0.009469812735915184, + 0.00955659244209528, + 0.01752299815416336, + 0.013045924715697765, + 0.014582982286810875, + 0.08188517391681671, + 0.11481893807649612, + 0.01218919362872839, + 0.018885068595409393, + 0.01620219275355339, + 0.0176585353910923, + 0.00982594396919012, + 0.020365910604596138, + 0.02673315815627575, + 0.02328774891793728, + 0.011790073476731777, + 0.011290078051388264, + 0.013222401961684227, + 0.0198047012090683, + 0.03403552621603012, + 0.025417307391762733, + 0.020888840779662132, + 0.007459848653525114, + 0.00837691780179739, + 0.005896030459553003, + 0.005647899582982063, + 0.007687433157116175, + 0.007282540202140808, + 0.026442257687449455, + 0.006397818215191364, + 0.003394986502826214, + 0.007533811964094639, + 0.002702168421819806, + 0.0037150259595364332, + 0.0032414833549410105, + 0.004221717361360788, + 0.0033471754286438227, + 0.006274097599089146, + 0.0029089204035699368, + 0.004269417840987444, + 0.0035901193041354418, + 0.003535793162882328, + 0.004813470412045717, + 0.005566516891121864, + 0.003577391617000103, + 0.0015661140205338597, + 0.0017894067568704486, + 0.0011815535835921764, + 0.001388302887789905, + 0.0021360348910093307, + 0.002357356483116746, + 0.002287345938384533, + 0.002365535357967019, + 0.00578144658356905, + 0.005023511126637459, + 0.005254487972706556, + 0.003979806322604418, + 0.003495382145047188, + 0.0038806081283837557, + 0.004339843522757292, + 0.005722008179873228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.0022004456259310246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015780359506607 + ], + "opposite_logprob_tokens": [ + -1.7015780210494995 + ], + "logprob_reward": -0.2015780359506607, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015780359506607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04027014970779419, + 0.20066098868846893, + 0.03323156014084816, + 0.30034276843070984, + 0.08111149072647095, + 0.14431221783161163, + 0.03902508318424225, + 0.03996530920267105, + 0.030103642493486404, + 0.09097680449485779 + ], + "credit_map": [ + 0.005185677204281092, + 0.003986187744885683, + 0.003375606844201684, + 0.0034536775201559067, + 0.009088665246963501, + 0.006646876689046621, + 0.004420341923832893, + 0.0041131190955638885, + 0.023333365097641945, + 0.01467628963291645, + 0.01651499792933464, + 0.013279937207698822, + 0.047346331179142, + 0.037725355476140976, + 0.018729323521256447, + 0.029055388644337654, + 0.0042733801528811455, + 0.0036201488692313433, + 0.0039033587090671062, + 0.00398392416536808, + 0.0029569482430815697, + 0.004254826810210943, + 0.006417933851480484, + 0.003821038408204913, + 0.08654008060693741, + 0.12207325547933578, + 0.011789040639996529, + 0.022224975749850273, + 0.014408158138394356, + 0.017812205478549004, + 0.011665549129247665, + 0.01382953580468893, + 0.010265501216053963, + 0.012144157662987709, + 0.007653673645108938, + 0.007767579983919859, + 0.008051462471485138, + 0.012353558093309402, + 0.008695719763636589, + 0.014179835096001625, + 0.017696563154459, + 0.0398591011762619, + 0.011385302059352398, + 0.01115813571959734, + 0.0138902822509408, + 0.01356460340321064, + 0.02095942385494709, + 0.015798790380358696, + 0.004669981077313423, + 0.004960712511092424, + 0.0069548217579722404, + 0.003824834479019046, + 0.005306163337081671, + 0.004781557712703943, + 0.0037470185197889805, + 0.004779990762472153, + 0.0060459948144853115, + 0.0043508498929440975, + 0.004324205685406923, + 0.0057905009016394615, + 0.003789676120504737, + 0.00785667635500431, + 0.004284119699150324, + 0.0035232845693826675, + 0.004357846919447184, + 0.0032642430160194635, + 0.003663663985207677, + 0.0029862443916499615, + 0.0035722381435334682, + 0.003625913057476282, + 0.004162508994340897, + 0.004470983054488897, + 0.022241750732064247, + 0.014801111072301865, + 0.012958202511072159, + 0.008727948181331158, + 0.007916681468486786, + 0.008246900513768196, + 0.006967408582568169, + 0.009116807952523232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.2015780359506607 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134535253047943 + ], + "opposite_logprob_tokens": [ + -1.3134535551071167 + ], + "logprob_reward": -0.3134535253047943, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134535253047943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11166943609714508, + 0.12084552645683289, + 0.07738301903009415, + 0.11439923197031021, + 0.1347048580646515, + 0.10309262573719025, + 0.06821970641613007, + 0.1371791809797287, + 0.029439691454172134, + 0.10306676477193832 + ], + "credit_map": [ + 0.011309736408293247, + 0.012367773801088333, + 0.011693621054291725, + 0.009003719314932823, + 0.02639901638031006, + 0.02063433639705181, + 0.01006295531988144, + 0.010198280215263367, + 0.013649272732436657, + 0.010573930107057095, + 0.01592724584043026, + 0.01275466289371252, + 0.01945815049111843, + 0.0152014484629035, + 0.016074931249022484, + 0.017205888405442238, + 0.010242943651974201, + 0.011428670957684517, + 0.008856459520757198, + 0.007269121240824461, + 0.008158300071954727, + 0.008418124169111252, + 0.010042463429272175, + 0.012966935522854328, + 0.020939035341143608, + 0.03625664860010147, + 0.008424703031778336, + 0.0130847729742527, + 0.008199634961783886, + 0.010192056186497211, + 0.007687197998166084, + 0.00961516983807087, + 0.018191929906606674, + 0.022897589951753616, + 0.01181121077388525, + 0.010944925248622894, + 0.014185498468577862, + 0.020108632743358612, + 0.013722818344831467, + 0.022842243313789368, + 0.015813367441296577, + 0.02162317931652069, + 0.009507856331765652, + 0.008257907815277576, + 0.009635142050683498, + 0.010532643646001816, + 0.016692936420440674, + 0.011029591783881187, + 0.011253288947045803, + 0.008223929442465305, + 0.008547988720238209, + 0.00540931336581707, + 0.012690079398453236, + 0.007510407827794552, + 0.007139355409890413, + 0.007445336785167456, + 0.02607855387032032, + 0.015252342447638512, + 0.01713225618004799, + 0.02242494933307171, + 0.012348589487373829, + 0.01473939884454012, + 0.017472442239522934, + 0.011730652302503586, + 0.00437987269833684, + 0.003426286391913891, + 0.0036098561249673367, + 0.0032034108880907297, + 0.004108577035367489, + 0.005067440681159496, + 0.002491441322490573, + 0.00315280444920063, + 0.02130180224776268, + 0.01500372402369976, + 0.015754254534840584, + 0.008814644999802113, + 0.010403089225292206, + 0.012036359868943691, + 0.008317478001117706, + 0.011435410939157009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.3134535253047943 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-tilting wide-angle shot reveals the interior of a massive, cylindrical, high-tech greenhouse tower, its polished steel structure and glass walls stretching upward as far as the eye can see. Rows of vertical hydroponic racks spiral upward, densely packed with lush leafy greens—primarily crisp, dew-drenched lettuce and deep, veined kale—a deep emerald hue. The space is bathed in vibrant purple and soft pink glow cast by rows of LED lights, a surreal illumination that makes the greens of the plants appear strikingly vivid. The camera then pans smoothly to the left, following a small, white quadcopter drone as it silently glides along a pre-set magnetic track, displaying perfect stability as it hovers above a row of plants. Then, carefully extending its precision nozzle, it evenly sprays a fine, nutrient-rich mist of water onto the plants' roots. In the foreground walkway, a researcher in a pristine white lab coat stood absorbed in his work, his fingers rapidly sliding and tapping across a translucent blue holographic tablet suspended in mid-air. The tablet displayed a real-time display of glowing green growth graphs and complex data streams. The entire scene blended the tranquility of organic life with the efficiency of sterile technology, creating an atmosphere of focus, cleanliness, and a futuristic feel.", + "seed": 28706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_9462_seed_28706_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11922292539384216, + "vlm_logprob": -0.11922292539384216, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.11922292539384216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1816985309123993 + ], + "opposite_logprob_tokens": [ + -1.8066985607147217 + ], + "logprob_reward": -0.1816985309123993, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1816985309123993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17870491743087769, + 0.08366016298532486, + 0.136351078748703, + 0.06999262422323227, + 0.02931070886552334, + 0.0474131740629673, + 0.028306826949119568, + 0.03192286938428879, + 0.2792679965496063, + 0.11506971716880798 + ], + "credit_map": [ + 0.02565988153219223, + 0.02721431851387024, + 0.023711223155260086, + 0.01908821240067482, + 0.023330308496952057, + 0.022816423326730728, + 0.015929756686091423, + 0.020954765379428864, + 0.006448240485042334, + 0.01211515162140131, + 0.018752846866846085, + 0.022884303703904152, + 0.006223189178854227, + 0.006194401998072863, + 0.007549682166427374, + 0.0034923339262604713, + 0.008260335773229599, + 0.011742748320102692, + 0.023920530453324318, + 0.007171839941293001, + 0.04279891401529312, + 0.025607287883758545, + 0.0103691928088665, + 0.006480207201093435, + 0.009176966734230518, + 0.00872284546494484, + 0.013164934702217579, + 0.0064599174074828625, + 0.009380815550684929, + 0.010039051994681358, + 0.005674788728356361, + 0.007373289205133915, + 0.002946765162050724, + 0.004974160343408585, + 0.0038781706243753433, + 0.0041954838670790195, + 0.003705532057210803, + 0.0053446367383003235, + 0.0022539934143424034, + 0.0020119633991271257, + 0.005422002170234919, + 0.00589203042909503, + 0.005932214204221964, + 0.005427380092442036, + 0.00828953180462122, + 0.0079194325953722, + 0.004659807309508324, + 0.0038707698695361614, + 0.0031461897306144238, + 0.004598409403115511, + 0.004068025387823582, + 0.005818269681185484, + 0.0028904045466333628, + 0.0029936544597148895, + 0.002426605671644211, + 0.0023652624804526567, + 0.005801382940262556, + 0.0036594278644770384, + 0.006456412374973297, + 0.0028470216784626245, + 0.005254876334220171, + 0.0031769839115440845, + 0.002347977366298437, + 0.0023787834215909243, + 0.010220583528280258, + 0.018052974715828896, + 0.07495370507240295, + 0.011552880518138409, + 0.11258427798748016, + 0.036793794482946396, + 0.009192991070449352, + 0.005916748195886612, + 0.009940430521965027, + 0.006692236755043268, + 0.04180941730737686, + 0.02937708981335163, + 0.010609623044729233, + 0.006271719932556152, + 0.005477928090840578, + 0.004891251213848591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.1816985309123993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201233819127083 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201233819127083, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201233819127083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0603160485625267, + 0.00860085804015398, + 0.10568589717149734, + 0.005445378366857767, + 0.011192010715603828, + 0.026371633633971214, + 0.010614010505378246, + 0.005498518235981464, + 0.7577722668647766, + 0.00850333459675312 + ], + "credit_map": [ + 0.00900511909276247, + 0.005641164258122444, + 0.006178163457661867, + 0.005903498735278845, + 0.007266317494213581, + 0.008297420106828213, + 0.009794317185878754, + 0.0082300566136837, + 0.0009676378103904426, + 0.0010440170299261808, + 0.0021786955185234547, + 0.000991038978099823, + 0.0009521167376078665, + 0.0009312195470556617, + 0.0008220907184295356, + 0.0007140428060665727, + 0.007213980425149202, + 0.00841023214161396, + 0.026971464976668358, + 0.004862741567194462, + 0.03277458995580673, + 0.015682563185691833, + 0.0051547931507229805, + 0.004615538753569126, + 0.0007393616251647472, + 0.0006131399422883987, + 0.0010524244280532002, + 0.0005735931918025017, + 0.0009259165963158011, + 0.000738740258384496, + 0.0004134744522161782, + 0.0003887283382937312, + 0.0010455885203555226, + 0.0021398102398961782, + 0.001073274528607726, + 0.0012197495670989156, + 0.0013426992809399962, + 0.0020197229459881783, + 0.0011403841199353337, + 0.0012107820948585868, + 0.001977570354938507, + 0.0023310501128435135, + 0.0021017282269895077, + 0.0037955520674586296, + 0.0026954086497426033, + 0.003613709006458521, + 0.005164986941963434, + 0.004691629204899073, + 0.0012020363938063383, + 0.0012157848104834557, + 0.0017017606878653169, + 0.0011303180363029242, + 0.00111247634049505, + 0.001305644866079092, + 0.0015595974400639534, + 0.001386392512358725, + 0.00038080537342466414, + 0.0007049827836453915, + 0.000852564291562885, + 0.0007607984589412808, + 0.0011305732186883688, + 0.000617319077719003, + 0.0005108665791340172, + 0.0005406087730079889, + 0.025561919435858727, + 0.03811464458703995, + 0.1948910355567932, + 0.024158790707588196, + 0.32517778873443604, + 0.11891467869281769, + 0.018878264352679253, + 0.012075157836079597, + 0.0006872822996228933, + 0.0006037693819962442, + 0.001823194557800889, + 0.0013891073176637292, + 0.001183283282443881, + 0.000748730672057718, + 0.001087468583136797, + 0.0009804994333535433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.018201233819127083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019425348145887256 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019425348145887256, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019425348145887256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06981988251209259, + 0.18109363317489624, + 0.08929520845413208, + 0.04889750853180885, + 0.03654770180583, + 0.05654469504952431, + 0.03313066437840462, + 0.01672045886516571, + 0.37138599157333374, + 0.09656427800655365 + ], + "credit_map": [ + 0.00664122449234128, + 0.0069181229919195175, + 0.00782079715281725, + 0.007799011655151844, + 0.0060892789624631405, + 0.009037387557327747, + 0.014945484697818756, + 0.010568567551672459, + 0.011552688665688038, + 0.054410625249147415, + 0.01723889820277691, + 0.04652145504951477, + 0.01871057227253914, + 0.012154074385762215, + 0.012087133713066578, + 0.008418170735239983, + 0.005878453142940998, + 0.008646620437502861, + 0.01766100339591503, + 0.0055402666330337524, + 0.02855944260954857, + 0.013136343099176884, + 0.0057191625237464905, + 0.004153906833380461, + 0.008073519915342331, + 0.006879816763103008, + 0.007699035108089447, + 0.003237234428524971, + 0.007697647903114557, + 0.006952417548745871, + 0.0032619419507682323, + 0.0050958930514752865, + 0.003384743584319949, + 0.006638945546001196, + 0.004703917540609837, + 0.0032432819716632366, + 0.004862423054873943, + 0.007493589539080858, + 0.0031942962668836117, + 0.0030265026725828648, + 0.004439907614141703, + 0.005800499580800533, + 0.006430306937545538, + 0.007957476191222668, + 0.006018847227096558, + 0.008529177866876125, + 0.010214748792350292, + 0.00715372571721673, + 0.003167165443301201, + 0.004228610545396805, + 0.004410127177834511, + 0.005813660100102425, + 0.004222137853503227, + 0.0041634440422058105, + 0.00400075688958168, + 0.003124760929495096, + 0.0014407128328457475, + 0.0016722818836569786, + 0.002334882505238056, + 0.0018928971840068698, + 0.0033290015999227762, + 0.00215554959140718, + 0.0022880344185978174, + 0.001607098150998354, + 0.011306151747703552, + 0.020545784384012222, + 0.0797358825802803, + 0.010762438178062439, + 0.16155490279197693, + 0.07135779410600662, + 0.009364066645503044, + 0.006758917588740587, + 0.006956739351153374, + 0.005124976858496666, + 0.037100035697221756, + 0.020170768722891808, + 0.012451909482479095, + 0.005686569958925247, + 0.004407140426337719, + 0.004666130989789963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.0019425348145887256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294522881507874 + ], + "opposite_logprob_tokens": [ + -2.237945318222046 + ], + "logprob_reward": -0.11294522881507874, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294522881507874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02012951485812664, + 0.08468659967184067, + 0.030997702851891518, + 0.08635250478982925, + 0.018209874629974365, + 0.0403735488653183, + 0.00817916914820671, + 0.010264834389090538, + 0.662167489528656, + 0.038638826459646225 + ], + "credit_map": [ + 0.0034155675675719976, + 0.00255110883153975, + 0.001675209030508995, + 0.001606741570867598, + 0.002410764340311289, + 0.003151391167193651, + 0.0035197206307202578, + 0.0017990098567679524, + 0.004586980678141117, + 0.020782068371772766, + 0.01227810699492693, + 0.024485813453793526, + 0.007215097080916166, + 0.004925081972032785, + 0.006523732095956802, + 0.003889714600518346, + 0.003184604225680232, + 0.002133877482265234, + 0.012812787666916847, + 0.002711578970775008, + 0.0039321044459939, + 0.0034150388091802597, + 0.0015821567503735423, + 0.0012255527544766665, + 0.013716952875256538, + 0.011711830273270607, + 0.014559343457221985, + 0.006663909647613764, + 0.01068895310163498, + 0.012381179258227348, + 0.006912008859217167, + 0.009718324989080429, + 0.0017314276192337275, + 0.004438997246325016, + 0.0019295456586405635, + 0.0020080609247088432, + 0.002522655064240098, + 0.0031890501268208027, + 0.0011996919056400657, + 0.0011904437560588121, + 0.004499136470258236, + 0.005775651894509792, + 0.006511235609650612, + 0.005229228641837835, + 0.0041946652345359325, + 0.003894747933372855, + 0.006322021596133709, + 0.003946860320866108, + 0.0008759046322666109, + 0.001216341508552432, + 0.0010390861425548792, + 0.0014725885121151805, + 0.0010119403013959527, + 0.0009996236767619848, + 0.0008708015084266663, + 0.000692882516887039, + 0.0018045061733573675, + 0.0010697228135541081, + 0.002604261040687561, + 0.0009552433621138334, + 0.0013959187781438231, + 0.0010779408039525151, + 0.0007291833753697574, + 0.0006280571105889976, + 0.022437511011958122, + 0.033291660249233246, + 0.17174439132213593, + 0.020719796419143677, + 0.28310492634773254, + 0.1035609245300293, + 0.01660856045782566, + 0.010699611157178879, + 0.005596231669187546, + 0.0028377152048051357, + 0.010388359427452087, + 0.009782045148313046, + 0.0030816292855888605, + 0.002222829032689333, + 0.0020832717418670654, + 0.002646745415404439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.11294522881507874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132709860801697 + ], + "opposite_logprob_tokens": [ + -1.4063271284103394 + ], + "logprob_reward": -0.28132709860801697, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132709860801697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04426748305559158, + 0.17042899131774902, + 0.019467726349830627, + 0.016835590824484825, + 0.026089098304510117, + 0.05937936529517174, + 0.017483141273260117, + 0.057081613689661026, + 0.05717974528670311, + 0.5317872166633606 + ], + "credit_map": [ + 0.0038435396272689104, + 0.0031548275146633387, + 0.003681500442326069, + 0.006730700843036175, + 0.004425146616995335, + 0.0034065567888319492, + 0.01175494771450758, + 0.0072702690958976746, + 0.009649360552430153, + 0.04646174609661102, + 0.018597841262817383, + 0.04904315620660782, + 0.01619678921997547, + 0.008979453705251217, + 0.012402410618960857, + 0.009098230861127377, + 0.0014700543833896518, + 0.0013087581610307097, + 0.003869080450385809, + 0.0013886978849768639, + 0.00622042128816247, + 0.003364631673321128, + 0.0009258875506930053, + 0.0009201968205161393, + 0.002731012413278222, + 0.002051415154710412, + 0.0032444079406559467, + 0.002353772521018982, + 0.0018355946522206068, + 0.0022623203694820404, + 0.001174631412141025, + 0.0011824375251308084, + 0.0027673335280269384, + 0.005790651775896549, + 0.00284174014814198, + 0.0025317880790680647, + 0.003667006967589259, + 0.004615153651684523, + 0.0020671123638749123, + 0.001808315166272223, + 0.00842999666929245, + 0.01015380211174488, + 0.010258168913424015, + 0.006341388914734125, + 0.006357856094837189, + 0.0055289617739617825, + 0.006145070306956768, + 0.006164129823446274, + 0.00231615686789155, + 0.002679660217836499, + 0.002871118253096938, + 0.0025029575917869806, + 0.00178195396438241, + 0.002148143481463194, + 0.0017971106572076678, + 0.0013860419858247042, + 0.011114740744233131, + 0.005889739841222763, + 0.013631986454129219, + 0.004839926026761532, + 0.007774557918310165, + 0.005504203960299492, + 0.004692675545811653, + 0.003633788786828518, + 0.0033192462287843227, + 0.004051907453685999, + 0.018247190862894058, + 0.004896210972219706, + 0.017105620354413986, + 0.004229606129229069, + 0.002941251266747713, + 0.002388722961768508, + 0.04298361390829086, + 0.02939320169389248, + 0.19953881204128265, + 0.13266269862651825, + 0.053235121071338654, + 0.028258899226784706, + 0.02489808015525341, + 0.020816853269934654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.28132709860801697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-tilting wide-angle shot reveals the interior of a massive, cylindrical, high-tech greenhouse tower, its polished steel structure and glass walls stretching upward as far as the eye can see. Rows of vertical hydroponic racks spiral upward, densely packed with lush leafy greens—primarily crisp, dew-drenched lettuce and deep, veined kale—a deep emerald hue. The space is bathed in vibrant purple and soft pink glow cast by rows of LED lights, a surreal illumination that makes the greens of the plants appear strikingly vivid. The camera then pans smoothly to the left, following a small, white quadcopter drone as it silently glides along a pre-set magnetic track, displaying perfect stability as it hovers above a row of plants. Then, carefully extending its precision nozzle, it evenly sprays a fine, nutrient-rich mist of water onto the plants' roots. In the foreground walkway, a researcher in a pristine white lab coat stood absorbed in his work, his fingers rapidly sliding and tapping across a translucent blue holographic tablet suspended in mid-air. The tablet displayed a real-time display of glowing green growth graphs and complex data streams. The entire scene blended the tranquility of organic life with the efficiency of sterile technology, creating an atmosphere of focus, cleanliness, and a futuristic feel.", + "seed": 28707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_9462_seed_28707_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21472444655373693, + "vlm_logprob": -0.21472444655373693, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.21472444655373693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47527071833610535 + ], + "opposite_logprob_tokens": [ + -0.9752707481384277 + ], + "logprob_reward": -0.47527071833610535, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47527071833610535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5446710586547852, + 0.053958918899297714, + 0.05099596828222275, + 0.03432454168796539, + 0.10243632644414902, + 0.04731130599975586, + 0.05403042957186699, + 0.045140694826841354, + 0.029260994866490364, + 0.03786976635456085 + ], + "credit_map": [ + 0.018571002408862114, + 0.02154500037431717, + 0.025513039901852608, + 0.027717648074030876, + 0.03900996223092079, + 0.17632167041301727, + 0.047111883759498596, + 0.18888084590435028, + 0.006352636963129044, + 0.004641853738576174, + 0.005832301918417215, + 0.0070285978727042675, + 0.007062291726469994, + 0.005326590500771999, + 0.009024808183312416, + 0.00868984218686819, + 0.005067278165370226, + 0.007148949429392815, + 0.008068103343248367, + 0.0060194144025444984, + 0.005782618187367916, + 0.004627472721040249, + 0.008328176103532314, + 0.005953953601419926, + 0.002852948382496834, + 0.003711677622050047, + 0.0027343451511114836, + 0.003100969595834613, + 0.005307073704898357, + 0.003647074569016695, + 0.007561592850834131, + 0.005408862605690956, + 0.004986002575606108, + 0.0031053570564836264, + 0.0025867680087685585, + 0.004875851795077324, + 0.007235744968056679, + 0.03370478376746178, + 0.0077719977125525475, + 0.03816981613636017, + 0.004339251201599836, + 0.004839262925088406, + 0.007423904724419117, + 0.004574006889015436, + 0.009746446274220943, + 0.004595424514263868, + 0.005065524019300938, + 0.006727486848831177, + 0.004966722801327705, + 0.003570504952222109, + 0.009068742394447327, + 0.022793537005782127, + 0.0024578615557402372, + 0.0024190284311771393, + 0.0062170191667973995, + 0.002537012565881014, + 0.006073432508856058, + 0.008131470531225204, + 0.005480317864567041, + 0.006865766830742359, + 0.004561678972095251, + 0.004149006213992834, + 0.005266475025564432, + 0.004612546414136887, + 0.004601189401000738, + 0.005076558329164982, + 0.0034240628592669964, + 0.005099277012050152, + 0.002749783918261528, + 0.002510065445676446, + 0.0030523359309881926, + 0.002747722202911973, + 0.009275959804654121, + 0.0034723635762929916, + 0.005016795825213194, + 0.0067624980583786964, + 0.0035277067217975855, + 0.002403503516688943, + 0.004512816667556763, + 0.002898119855672121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.47527071833610535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990808188915253 + ], + "opposite_logprob_tokens": [ + -1.804908037185669 + ], + "logprob_reward": -0.17990808188915253, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990808188915253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3392183780670166, + 0.03639186546206474, + 0.06706397980451584, + 0.03929165005683899, + 0.10463891923427582, + 0.2077248990535736, + 0.02633734792470932, + 0.10223334282636642, + 0.06702297925949097, + 0.010076618753373623 + ], + "credit_map": [ + 0.023148655891418457, + 0.022943787276744843, + 0.026511967182159424, + 0.02788848988711834, + 0.060552507638931274, + 0.06023737043142319, + 0.055884964764118195, + 0.06205061823129654, + 0.004529709927737713, + 0.0030295292381197214, + 0.005068765487521887, + 0.004571925383061171, + 0.004279619082808495, + 0.004308846313506365, + 0.005866332445293665, + 0.004737141076475382, + 0.006798821967095137, + 0.011433562263846397, + 0.013565459288656712, + 0.00887245498597622, + 0.0063208225183188915, + 0.0053233494982123375, + 0.007032802794128656, + 0.007716700900346041, + 0.00354023277759552, + 0.0042032767087221146, + 0.005502872634679079, + 0.004896083381026983, + 0.004400840029120445, + 0.005118813831359148, + 0.005979475565254688, + 0.005650052800774574, + 0.005479814484715462, + 0.005084977950900793, + 0.005455642007291317, + 0.01382393017411232, + 0.005174707155674696, + 0.027885690331459045, + 0.009257219731807709, + 0.03247693181037903, + 0.011566303670406342, + 0.06633130460977554, + 0.014848926104605198, + 0.010247107595205307, + 0.06942827254533768, + 0.01393609307706356, + 0.01145016960799694, + 0.009916734881699085, + 0.002686435356736183, + 0.0019407439976930618, + 0.0040879566222429276, + 0.010391884483397007, + 0.0012686453992500901, + 0.0016316752880811691, + 0.0026987267192453146, + 0.0016312821535393596, + 0.01381458155810833, + 0.014767976477742195, + 0.013404026627540588, + 0.01735566183924675, + 0.0089181549847126, + 0.010400095023214817, + 0.012013607658445835, + 0.011559242382645607, + 0.00799595657736063, + 0.014852996915578842, + 0.007141525391489267, + 0.00588010111823678, + 0.010355685837566853, + 0.010911768302321434, + 0.004789245314896107, + 0.005095700733363628, + 0.0018015444511547685, + 0.0017275206046178937, + 0.0013571518938988447, + 0.00215464411303401, + 0.0008315052255056798, + 0.00082671974087134, + 0.0007551123853772879, + 0.0006224204553291202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.17990808188915253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955806467682123 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.005955806467682123, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955806467682123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4418148994445801, + 0.046312812715768814, + 0.02204015664756298, + 0.02404787577688694, + 0.1489444077014923, + 0.04599151760339737, + 0.05908357352018356, + 0.029963001608848572, + 0.1615806221961975, + 0.02022111415863037 + ], + "credit_map": [ + 0.019742537289857864, + 0.020458318293094635, + 0.02594108320772648, + 0.02591124176979065, + 0.05078381299972534, + 0.11939464509487152, + 0.05367746204137802, + 0.12590579688549042, + 0.004893168341368437, + 0.0038921909872442484, + 0.005340879317373037, + 0.005305254366248846, + 0.006449347361922264, + 0.005276952404528856, + 0.008019102737307549, + 0.007135914638638496, + 0.0018607054371386766, + 0.0026923425029963255, + 0.0029632621444761753, + 0.00244248379021883, + 0.002678696298971772, + 0.0037041353061795235, + 0.0034203510731458664, + 0.0022781784646213055, + 0.0021142223849892616, + 0.0029470252338796854, + 0.002024667104706168, + 0.002345244400203228, + 0.003508401568979025, + 0.0028184938710182905, + 0.004593192599713802, + 0.0036966288462281227, + 0.0071806893683969975, + 0.005894219037145376, + 0.004935766104608774, + 0.013052502647042274, + 0.010592642240226269, + 0.04407479986548424, + 0.011902601458132267, + 0.0513111911714077, + 0.0026956035289913416, + 0.012432685121893883, + 0.0046800123527646065, + 0.003647660603746772, + 0.012367083691060543, + 0.0034911464899778366, + 0.003884766483679414, + 0.002792560262605548, + 0.005596108268946409, + 0.0038300552405416965, + 0.01240574847906828, + 0.02238084375858307, + 0.002620300045236945, + 0.0030425298027694225, + 0.006315924227237701, + 0.0028920641634613276, + 0.003775863442569971, + 0.006211507134139538, + 0.0027495615649968386, + 0.004564283881336451, + 0.0030599927995353937, + 0.0025202755350619555, + 0.004489194601774216, + 0.002592321950942278, + 0.018142761662602425, + 0.03270626813173294, + 0.01501508429646492, + 0.01235287543386221, + 0.02847360074520111, + 0.033116504549980164, + 0.010382343083620071, + 0.01139118243008852, + 0.0036332830786705017, + 0.001921404036693275, + 0.003548978129401803, + 0.003696325235068798, + 0.0017529474571347237, + 0.0015074082184582949, + 0.0025582686066627502, + 0.0016024996293708682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.005955806467682123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037845611572266 + ], + "opposite_logprob_tokens": [ + -1.9103784561157227 + ], + "logprob_reward": -0.16037845611572266, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037845611572266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09727878123521805, + 0.05877688527107239, + 0.06345407664775848, + 0.06127208471298218, + 0.26780810952186584, + 0.12420038133859634, + 0.07491511851549149, + 0.128060981631279, + 0.040978770703077316, + 0.08325478434562683 + ], + "credit_map": [ + 0.009571004658937454, + 0.010029253549873829, + 0.012623106129467487, + 0.012797268107533455, + 0.012112751603126526, + 0.010075709782540798, + 0.019615380093455315, + 0.010454299859702587, + 0.006277464330196381, + 0.009007606655359268, + 0.006677502766251564, + 0.007524399086833, + 0.006413707509636879, + 0.006526962388306856, + 0.009144223295152187, + 0.007205019239336252, + 0.005097085144370794, + 0.004131485242396593, + 0.008425632491707802, + 0.006199222058057785, + 0.005682718940079212, + 0.016286293044686317, + 0.008626566268503666, + 0.009005068801343441, + 0.005765678361058235, + 0.008152550086379051, + 0.0064390879124403, + 0.0055562574416399, + 0.008501388132572174, + 0.010325754061341286, + 0.008741885423660278, + 0.007789485156536102, + 0.013219615444540977, + 0.00952149461954832, + 0.007102674804627895, + 0.013757049106061459, + 0.025670653209090233, + 0.08219397068023682, + 0.021274248138070107, + 0.09506841003894806, + 0.008397176861763, + 0.01805010624229908, + 0.014511591754853725, + 0.007687306497246027, + 0.034657422453165054, + 0.01097792573273182, + 0.011841081082820892, + 0.01807776652276516, + 0.00818695966154337, + 0.005342594347894192, + 0.012683035805821419, + 0.028030114248394966, + 0.0027177059091627598, + 0.00493284547701478, + 0.008983801119029522, + 0.004038064740598202, + 0.01796221174299717, + 0.017663568258285522, + 0.017247848212718964, + 0.023988831788301468, + 0.010038095526397228, + 0.012141978368163109, + 0.01531470101326704, + 0.013703743927180767, + 0.004585299640893936, + 0.008271065540611744, + 0.004287953954190016, + 0.00411008158698678, + 0.006709345616400242, + 0.007269596215337515, + 0.002830458339303732, + 0.002914970275014639, + 0.016535447910428047, + 0.012034555897116661, + 0.010833367705345154, + 0.013287538662552834, + 0.009123732335865498, + 0.005309108179062605, + 0.010202985256910324, + 0.005928047467023134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.16037845611572266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252109169960022 + ], + "opposite_logprob_tokens": [ + -1.502109169960022 + ], + "logprob_reward": -0.252109169960022, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252109169960022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25889068841934204, + 0.05918694660067558, + 0.04562078416347504, + 0.047234490513801575, + 0.03464747592806816, + 0.12643764913082123, + 0.06809879839420319, + 0.0801427811384201, + 0.09303686022758484, + 0.18670354783535004 + ], + "credit_map": [ + 0.016638385131955147, + 0.016578659415245056, + 0.023092105984687805, + 0.022750403732061386, + 0.041907571256160736, + 0.04408770054578781, + 0.04379412904381752, + 0.050041716545820236, + 0.005554663483053446, + 0.008318763226270676, + 0.005550011061131954, + 0.009377194568514824, + 0.006776227615773678, + 0.005829004105180502, + 0.010160358622670174, + 0.007620722055435181, + 0.006778168957680464, + 0.008458111435174942, + 0.005906736943870783, + 0.004614660516381264, + 0.004539391491562128, + 0.004530575592070818, + 0.00590029451996088, + 0.004892843775451183, + 0.004009948577731848, + 0.005342242773622274, + 0.005158307962119579, + 0.0049623362720012665, + 0.009618418291211128, + 0.006421597674489021, + 0.007070313673466444, + 0.00465132761746645, + 0.002349250251427293, + 0.002085366053506732, + 0.0030260910280048847, + 0.007281710393726826, + 0.004699293989688158, + 0.006267169956117868, + 0.0038465289399027824, + 0.005092065781354904, + 0.010217025876045227, + 0.013134191744029522, + 0.015746809542179108, + 0.008597622625529766, + 0.038336753845214844, + 0.01000988855957985, + 0.012196322903037071, + 0.018199020996689796, + 0.00693895760923624, + 0.004206213168799877, + 0.017306309193372726, + 0.02227206714451313, + 0.003604051424190402, + 0.0035781534388661385, + 0.005901101510971785, + 0.00429194699972868, + 0.009723804891109467, + 0.011616485193371773, + 0.009701509028673172, + 0.01720346510410309, + 0.005683626048266888, + 0.007608957588672638, + 0.010020369663834572, + 0.008584561757743359, + 0.009548447094857693, + 0.020202364772558212, + 0.009458008222281933, + 0.00787625927478075, + 0.013155529275536537, + 0.019165456295013428, + 0.006672152318060398, + 0.006958637852221727, + 0.04364114999771118, + 0.019757915288209915, + 0.03548557683825493, + 0.04226627200841904, + 0.011809284798800945, + 0.009615382179617882, + 0.014404378831386566, + 0.009723584167659283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.252109169960022 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-tilting wide-angle shot reveals the interior of a massive, cylindrical, high-tech greenhouse tower, its polished steel structure and glass walls stretching upward as far as the eye can see. Rows of vertical hydroponic racks spiral upward, densely packed with lush leafy greens—primarily crisp, dew-drenched lettuce and deep, veined kale—a deep emerald hue. The space is bathed in vibrant purple and soft pink glow cast by rows of LED lights, a surreal illumination that makes the greens of the plants appear strikingly vivid. The camera then pans smoothly to the left, following a small, white quadcopter drone as it silently glides along a pre-set magnetic track, displaying perfect stability as it hovers above a row of plants. Then, carefully extending its precision nozzle, it evenly sprays a fine, nutrient-rich mist of water onto the plants' roots. In the foreground walkway, a researcher in a pristine white lab coat stood absorbed in his work, his fingers rapidly sliding and tapping across a translucent blue holographic tablet suspended in mid-air. The tablet displayed a real-time display of glowing green growth graphs and complex data streams. The entire scene blended the tranquility of organic life with the efficiency of sterile technology, creating an atmosphere of focus, cleanliness, and a futuristic feel.", + "seed": 28708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_8/index_9462_seed_28708_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06735904077067971, + "vlm_logprob": -0.06735904077067971, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.06735904077067971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10338085144758224 + ], + "opposite_logprob_tokens": [ + -2.3533809185028076 + ], + "logprob_reward": -0.10338085144758224, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10338085144758224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08782340586185455, + 0.0391412079334259, + 0.3674321174621582, + 0.049375515431165695, + 0.10114132612943649, + 0.08413385599851608, + 0.039441321045160294, + 0.07936786115169525, + 0.02951662242412567, + 0.12262675166130066 + ], + "credit_map": [ + 0.013441307470202446, + 0.007631022948771715, + 0.011099333874881268, + 0.009065673686563969, + 0.013820623978972435, + 0.012856666930019855, + 0.009499265812337399, + 0.010409506969153881, + 0.005051683634519577, + 0.00391279486939311, + 0.004595309961587191, + 0.0034105402883142233, + 0.00538631621748209, + 0.008457762189209461, + 0.0041030170395970345, + 0.004223784431815147, + 0.0255753044039011, + 0.011105245910584927, + 0.019061530008912086, + 0.0164633821696043, + 0.025227822363376617, + 0.10583879053592682, + 0.119040347635746, + 0.045119691640138626, + 0.007934005931019783, + 0.004411811474710703, + 0.007124932948499918, + 0.004659587051719427, + 0.006345896050333977, + 0.010613192804157734, + 0.004220978822559118, + 0.0040651108138263226, + 0.018063070252537727, + 0.010403844527900219, + 0.029356934130191803, + 0.00813008937984705, + 0.011871123686432838, + 0.006868941709399223, + 0.005764164496213198, + 0.010683154687285423, + 0.01581593044102192, + 0.013396594673395157, + 0.012857534922659397, + 0.014503099024295807, + 0.00737879378721118, + 0.008756445720791817, + 0.0047279102727770805, + 0.0066975378431379795, + 0.007757992018014193, + 0.006837876513600349, + 0.009217655286192894, + 0.004332813434302807, + 0.0033655345905572176, + 0.0024457245599478483, + 0.0027029027696698904, + 0.002780826995149255, + 0.00917287077754736, + 0.010473580099642277, + 0.011460697278380394, + 0.00965179968625307, + 0.007289881817996502, + 0.010831385850906372, + 0.012846427969634533, + 0.00764121487736702, + 0.004383700899779797, + 0.006624352186918259, + 0.004557162523269653, + 0.0038984878920018673, + 0.0035612499341368675, + 0.0015064718900248408, + 0.0030441253911703825, + 0.0019410692621022463, + 0.009787704795598984, + 0.006623043213039637, + 0.013726716861128807, + 0.008220250718295574, + 0.04222492501139641, + 0.0043664006516337395, + 0.03228354454040527, + 0.005394166801124811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.10338085144758224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292500525712967 + ], + "opposite_logprob_tokens": [ + -3.7732925415039062 + ], + "logprob_reward": -0.023292500525712967, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292500525712967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1189480572938919, + 0.16696321964263916, + 0.10863350331783295, + 0.06388302147388458, + 0.10487134009599686, + 0.05675949528813362, + 0.13941912353038788, + 0.1017528772354126, + 0.047611039131879807, + 0.09115833044052124 + ], + "credit_map": [ + 0.013183848932385445, + 0.013074299320578575, + 0.017797384411096573, + 0.015639908611774445, + 0.015831010416150093, + 0.015323814935982227, + 0.013673869892954826, + 0.014423927292227745, + 0.019333241507411003, + 0.019036171957850456, + 0.019485699012875557, + 0.02211681567132473, + 0.014885722659528255, + 0.022055363282561302, + 0.029591573402285576, + 0.020458633080124855, + 0.012666508555412292, + 0.010014928877353668, + 0.018178926780819893, + 0.013556077145040035, + 0.01183296274393797, + 0.01480928622186184, + 0.015881521627306938, + 0.01169330719858408, + 0.00670502008870244, + 0.0053908186964690685, + 0.008068270049989223, + 0.007815101183950901, + 0.011048911139369011, + 0.014050236903131008, + 0.005737010855227709, + 0.005067647900432348, + 0.020694125443696976, + 0.00837384071201086, + 0.030152132734656334, + 0.01084759272634983, + 0.009621510282158852, + 0.007379016373306513, + 0.008109682239592075, + 0.009693440981209278, + 0.008863234892487526, + 0.00819097738713026, + 0.014799674972891808, + 0.006895716767758131, + 0.004729531705379486, + 0.003891262225806713, + 0.0045790900476276875, + 0.004810010548681021, + 0.03159720078110695, + 0.036165352910757065, + 0.02338942512869835, + 0.01934516802430153, + 0.006960938684642315, + 0.004953967407345772, + 0.010084237903356552, + 0.006922848057001829, + 0.013604024425148964, + 0.014764244668185711, + 0.012609004974365234, + 0.010192224755883217, + 0.011650899425148964, + 0.01653597317636013, + 0.012025905773043633, + 0.010370602831244469, + 0.005842859856784344, + 0.0064720530062913895, + 0.006471989676356316, + 0.009350140579044819, + 0.005061679054051638, + 0.003735074307769537, + 0.006769443862140179, + 0.003907802980393171, + 0.009555787779390812, + 0.013327100314199924, + 0.014743918552994728, + 0.024572981521487236, + 0.011708966456353664, + 0.004959520418196917, + 0.008007248863577843, + 0.004282813053578138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.023292500525712967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07411515712738037, + 0.04534529149532318, + 0.05473607778549194, + 0.15286684036254883, + 0.2395666390657425, + 0.06883223354816437, + 0.09238727390766144, + 0.06951205432415009, + 0.05659784749150276, + 0.14604054391384125 + ], + "credit_map": [ + 0.01309448853135109, + 0.007052087690681219, + 0.008015562780201435, + 0.007997212000191212, + 0.009684623219072819, + 0.011631150729954243, + 0.006928850896656513, + 0.009711183607578278, + 0.005930020473897457, + 0.00566945830360055, + 0.005222078878432512, + 0.004681480582803488, + 0.0063190157525241375, + 0.007077739108353853, + 0.005501436069607735, + 0.004944059532135725, + 0.00477364519611001, + 0.003717275569215417, + 0.004765281453728676, + 0.0037327513564378023, + 0.0075616417452692986, + 0.012149106711149216, + 0.011172140948474407, + 0.006864232011139393, + 0.025681782513856888, + 0.012139608152210712, + 0.019153360277414322, + 0.011671620421111584, + 0.021750204265117645, + 0.03158790245652199, + 0.01376165822148323, + 0.01712070032954216, + 0.04598067328333855, + 0.02223827689886093, + 0.07791180163621902, + 0.021294748410582542, + 0.015180576592683792, + 0.01335715502500534, + 0.012967127375304699, + 0.030636275187134743, + 0.012171239592134953, + 0.010217323899269104, + 0.012446437031030655, + 0.014056137762963772, + 0.0043217213824391365, + 0.005642454605549574, + 0.004683769773691893, + 0.005293145775794983, + 0.02199431322515011, + 0.015296895988285542, + 0.01927395910024643, + 0.01131477952003479, + 0.007720184978097677, + 0.004291878547519445, + 0.006626483518630266, + 0.005868778098374605, + 0.008687916211783886, + 0.007257275283336639, + 0.010442882776260376, + 0.008219562470912933, + 0.006538868881762028, + 0.009580666199326515, + 0.010126377455890179, + 0.008658510632812977, + 0.008084024302661419, + 0.005216139834374189, + 0.005876224488019943, + 0.0086975684389472, + 0.007656473200768232, + 0.005954646971076727, + 0.00949054304510355, + 0.005622225347906351, + 0.017915211617946625, + 0.022215111181139946, + 0.020512327551841736, + 0.02695242315530777, + 0.02151176892220974, + 0.006280490662902594, + 0.02247525192797184, + 0.008177954703569412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.004097756464034319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898327708244324 + ], + "opposite_logprob_tokens": [ + -2.5789833068847656 + ], + "logprob_reward": -0.07898327708244324, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898327708244324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044475190341472626, + 0.022972039878368378, + 0.25977838039398193, + 0.025681015104055405, + 0.09873121976852417, + 0.05074742063879967, + 0.06133563071489334, + 0.03564033657312393, + 0.08708957582712173, + 0.3135492205619812 + ], + "credit_map": [ + 0.006948562804609537, + 0.0036368577275425196, + 0.004452403634786606, + 0.004261333495378494, + 0.006215675733983517, + 0.008267260156571865, + 0.0036466168239712715, + 0.007046480197459459, + 0.002526860684156418, + 0.0024233139120042324, + 0.0038746390491724014, + 0.0021085224580019712, + 0.003428621916100383, + 0.004086836241185665, + 0.002489547710865736, + 0.0020336988382041454, + 0.01789059303700924, + 0.008239289745688438, + 0.012769048102200031, + 0.011649874038994312, + 0.018283234909176826, + 0.07581968605518341, + 0.08393575251102448, + 0.031190888956189156, + 0.007324654143303633, + 0.002270087832584977, + 0.0037803896702826023, + 0.0026909050066024065, + 0.0018122210167348385, + 0.0028686460573226213, + 0.0019283801084384322, + 0.003005731152370572, + 0.0250316821038723, + 0.007134578190743923, + 0.033626459538936615, + 0.011224635876715183, + 0.004940826445817947, + 0.00423338683322072, + 0.004938192665576935, + 0.007601460441946983, + 0.007375698070973158, + 0.007028896827250719, + 0.013756892643868923, + 0.006061029154807329, + 0.004285715986043215, + 0.004340803250670433, + 0.003985850606113672, + 0.003912534564733505, + 0.011377050541341305, + 0.012655451893806458, + 0.009945632889866829, + 0.007980660535395145, + 0.006513836327940226, + 0.004229712300002575, + 0.004663003608584404, + 0.003970287274569273, + 0.004829391837120056, + 0.003619532799348235, + 0.004676789976656437, + 0.004894344136118889, + 0.0030579054728150368, + 0.005042603705078363, + 0.004357359372079372, + 0.005162411369383335, + 0.009569675661623478, + 0.028263794258236885, + 0.005495840683579445, + 0.017914511263370514, + 0.00817902572453022, + 0.004758116789162159, + 0.008098882623016834, + 0.004809732083231211, + 0.031531915068626404, + 0.044908832758665085, + 0.0353776253759861, + 0.052165016531944275, + 0.06601615250110626, + 0.012993528507649899, + 0.05555374547839165, + 0.015002395026385784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.07898327708244324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive cylindrical high-tech greenhouse tower with polished steel and glass walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plants primarily lush leafy greens like crisp lettuce and deep veined kale with a deep emerald hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the space illuminated by vibrant purple and soft pink glow from rows of LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a small white quadcopter drone glide along a magnetic track and spray a fine mist of water onto the plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a researcher in a white lab coat visible interacting with a translucent blue holographic tablet displaying green growth graphs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270408183336258 + ], + "opposite_logprob_tokens": [ + -2.1270408630371094 + ], + "logprob_reward": -0.1270408183336258, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270408183336258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044833701103925705, + 0.030664728954434395, + 0.05435030162334442, + 0.025355828925967216, + 0.03066265769302845, + 0.04485466331243515, + 0.30014628171920776, + 0.036493536084890366, + 0.07198011130094528, + 0.3606581687927246 + ], + "credit_map": [ + 0.004884014837443829, + 0.0035860484931617975, + 0.007586266845464706, + 0.007035580463707447, + 0.005100008100271225, + 0.005359316244721413, + 0.004635789431631565, + 0.006646677851676941, + 0.004249849356710911, + 0.003352655563503504, + 0.003971795551478863, + 0.0032268238719552755, + 0.003401672700420022, + 0.004459394607692957, + 0.004161722492426634, + 0.003840815741568804, + 0.006073371507227421, + 0.004135113209486008, + 0.00633515790104866, + 0.0038755731657147408, + 0.004192101303488016, + 0.011122525669634342, + 0.012796949595212936, + 0.0058195097371935844, + 0.008049244992434978, + 0.003015318652614951, + 0.0035784943029284477, + 0.0020709289237856865, + 0.0018784524872899055, + 0.0023398578632622957, + 0.0020134574733674526, + 0.0024100723676383495, + 0.004683635663241148, + 0.0030454599764198065, + 0.004966801963746548, + 0.004575233906507492, + 0.0029220441356301308, + 0.003537382697686553, + 0.0032845763489603996, + 0.003647521836683154, + 0.007581913378089666, + 0.007929296232759953, + 0.009963818825781345, + 0.005445444490760565, + 0.0031636301428079605, + 0.003223598236218095, + 0.004040571860969067, + 0.0035063927061855793, + 0.0637354701757431, + 0.07130943983793259, + 0.05250304192304611, + 0.04105807840824127, + 0.017550025135278702, + 0.010563171468675137, + 0.02504667639732361, + 0.018380386754870415, + 0.006441624369472265, + 0.004161642398685217, + 0.0063971043564379215, + 0.005704006180167198, + 0.0034328538458794355, + 0.0036456144880503416, + 0.004009605851024389, + 0.0027010864578187466, + 0.00845230370759964, + 0.0077825491316616535, + 0.013492767699062824, + 0.011270270682871342, + 0.007915319874882698, + 0.006446956656873226, + 0.009682826697826385, + 0.006937121972441673, + 0.036777060478925705, + 0.052558910101652145, + 0.04582573100924492, + 0.09043645113706589, + 0.059619154781103134, + 0.01674281619489193, + 0.046338483691215515, + 0.012359557673335075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.1270408183336258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_tensor([7936, 7936])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_tensor([7936, 7936])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b9fab81bc091512c87e432fa3c1eb114dfbbbb76 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_tensor([7936, 7936])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera sweeps smoothly from left to right across a brightly lit, modern office, at eye level. It first captures a woman with dark brown hair tied in a neat ponytail, wearing a vibrant red, long-sleeved silk shirt. She gently rocks on her feet as she carefully flips through a stack of white documents. As the saucepan continues to simmer, she leans forward and places the pile on the polished white surface of a large conference table. Then, a second woman with short blond hair and a sophisticated charcoal-gray blazer appears, seated at the table. She reaches out with her right hand and steadily and deliberately picks up the documents. The camera finally settles on a third woman, standing at the head of the table, dressed in a professional, navy-blue, sheath dress. As the camera settles on her, she begins speaking with a confident expression, raising her right hand with an open palm to emphasize a point, holding the gesture for a moment, and then slowly and gracefully lowering it to her side. All this is set in a scene bathed in soft, natural sunlight, with large windows creating an atmosphere of professional focus and collaboration.", + "seed": 18352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_7936_seed_18352_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06482247877866029, + "vlm_logprob": -0.06482247877866029, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.06482247877866029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22142909467220306 + ], + "opposite_logprob_tokens": [ + -1.8464291095733643 + ], + "logprob_reward": -0.22142909467220306, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.22142909467220306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10400209575891495, + 0.03792974725365639, + 0.06919965893030167, + 0.07350672036409378, + 0.2729296088218689, + 0.13206666707992554, + 0.07996346801519394, + 0.05672479793429375, + 0.13335737586021423, + 0.04031979665160179 + ], + "credit_map": [ + 0.020782532170414925, + 0.0113638024777174, + 0.015472136437892914, + 0.01220397837460041, + 0.012048836797475815, + 0.010207475163042545, + 0.010759202763438225, + 0.011164124123752117, + 0.004763682838529348, + 0.004609744995832443, + 0.0055702305398881435, + 0.005597605369985104, + 0.004851138684898615, + 0.005161406937986612, + 0.004447817802429199, + 0.0029281226452440023, + 0.01137311290949583, + 0.009838464669883251, + 0.007832794450223446, + 0.009539862163364887, + 0.0054728807881474495, + 0.009562687017023563, + 0.007864569313824177, + 0.0077152862213552, + 0.010539132170379162, + 0.008540185168385506, + 0.007191444281488657, + 0.00915118120610714, + 0.011205350048840046, + 0.012561480514705181, + 0.008155575022101402, + 0.006162365898489952, + 0.04652688279747963, + 0.05645834282040596, + 0.05094124749302864, + 0.0231610257178545, + 0.025775592774152756, + 0.03900342807173729, + 0.012533379718661308, + 0.018529728055000305, + 0.011156906373798847, + 0.021121766418218613, + 0.012339604087173939, + 0.013067884370684624, + 0.030295925214886665, + 0.01115155778825283, + 0.022760750725865364, + 0.010172282345592976, + 0.007582986261695623, + 0.0163604486733675, + 0.006344436202198267, + 0.009401694871485233, + 0.01567271538078785, + 0.008998643606901169, + 0.008471234701573849, + 0.007131303660571575, + 0.010248664766550064, + 0.00945685151964426, + 0.0068822805769741535, + 0.006536815315485001, + 0.0065344530157744884, + 0.006317604333162308, + 0.0054773068986833096, + 0.0052708168514072895, + 0.01652420125901699, + 0.021502986550331116, + 0.012876025401055813, + 0.01932418905198574, + 0.022715644910931587, + 0.015939682722091675, + 0.015279165469110012, + 0.009195481427013874, + 0.007099439390003681, + 0.008024976588785648, + 0.002978584496304393, + 0.004633471369743347, + 0.005180605687201023, + 0.004704299848526716, + 0.004491139203310013, + 0.003207281231880188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.22142909467220306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086315736174583 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086315736174583, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086315736174583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22366555035114288, + 0.09503989666700363, + 0.2842795252799988, + 0.04534802585840225, + 0.03449839726090431, + 0.03501126170158386, + 0.052090443670749664, + 0.0523630790412426, + 0.062068454921245575, + 0.11563540995121002 + ], + "credit_map": [ + 0.035178862512111664, + 0.023398352786898613, + 0.043416835367679596, + 0.03763221204280853, + 0.019021209329366684, + 0.03045077435672283, + 0.018155241385102272, + 0.01641206070780754, + 0.017499137669801712, + 0.014108805917203426, + 0.014081866480410099, + 0.012044726870954037, + 0.007842703722417355, + 0.015755491331219673, + 0.00651774974539876, + 0.0071894144639372826, + 0.04685864970088005, + 0.038850896060466766, + 0.04493437334895134, + 0.056880638003349304, + 0.021571313962340355, + 0.03828014060854912, + 0.01850597746670246, + 0.018397506326436996, + 0.006448713596910238, + 0.005775907076895237, + 0.007418530061841011, + 0.00781780295073986, + 0.005037324037402868, + 0.00507716741412878, + 0.004116799682378769, + 0.003655780106782913, + 0.005964776035398245, + 0.004977858159691095, + 0.005404178984463215, + 0.003927732352167368, + 0.00419579865410924, + 0.005186822731047869, + 0.0023166523315012455, + 0.0025245766155421734, + 0.0029325666837394238, + 0.004674005787819624, + 0.0034127559047192335, + 0.003293182235211134, + 0.007887275889515877, + 0.00372550031170249, + 0.006236664485186338, + 0.002849309705197811, + 0.006304589565843344, + 0.012102962471544743, + 0.004457908216863871, + 0.007080052513629198, + 0.007304431404918432, + 0.005705338437110186, + 0.004513659980148077, + 0.004621499218046665, + 0.007688141893595457, + 0.007910766638815403, + 0.006745082791894674, + 0.0059250472113490105, + 0.005713402759283781, + 0.005924050230532885, + 0.00742732547223568, + 0.0050292592495679855, + 0.008977657184004784, + 0.008784237317740917, + 0.0062164245173335075, + 0.00839785486459732, + 0.008588694967329502, + 0.007279167417436838, + 0.00785001553595066, + 0.005974400322884321, + 0.014925549738109112, + 0.017368439584970474, + 0.010015846230089664, + 0.01445548515766859, + 0.017323030158877373, + 0.012523185461759567, + 0.019863298162817955, + 0.009160569868981838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.016086315736174583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076628230512142 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076628230512142, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076628230512142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17029009759426117, + 0.11898022145032883, + 0.09237653017044067, + 0.05690672993659973, + 0.06248258054256439, + 0.051802653819322586, + 0.09703120589256287, + 0.10666797310113907, + 0.09358430653810501, + 0.14987774193286896 + ], + "credit_map": [ + 0.018605034798383713, + 0.01509478222578764, + 0.02587801404297352, + 0.02456786297261715, + 0.032092802226543427, + 0.017969077453017235, + 0.023304589092731476, + 0.012777939438819885, + 0.013528620824217796, + 0.012594531290233135, + 0.01323763933032751, + 0.015068800188601017, + 0.02786513976752758, + 0.013203369453549385, + 0.014944300055503845, + 0.00853782519698143, + 0.01194687094539404, + 0.013460824266076088, + 0.012698586098849773, + 0.02053200826048851, + 0.010357421822845936, + 0.01079679373651743, + 0.00719775864854455, + 0.0053862608037889, + 0.008455771952867508, + 0.007311693858355284, + 0.004754793364554644, + 0.007058064453303814, + 0.010922458954155445, + 0.007622010540217161, + 0.006355448625981808, + 0.004426493309438229, + 0.012787160463631153, + 0.009394215419888496, + 0.009459398686885834, + 0.007333720102906227, + 0.006910284981131554, + 0.008676506578922272, + 0.003907482605427504, + 0.004013807978481054, + 0.0053819394670426846, + 0.009266825392842293, + 0.005241727456450462, + 0.0048865447752177715, + 0.008743347600102425, + 0.007601500954478979, + 0.0067719523794949055, + 0.003908815328031778, + 0.010391565971076488, + 0.018778042867779732, + 0.006625025067478418, + 0.01158501673489809, + 0.024134662002325058, + 0.010757969692349434, + 0.007704293821007013, + 0.007054617628455162, + 0.01235655415803194, + 0.02261364832520485, + 0.009084893390536308, + 0.016668375581502914, + 0.012987491674721241, + 0.011252411641180515, + 0.013029414229094982, + 0.008675182238221169, + 0.012213821522891521, + 0.011061911471188068, + 0.006282353308051825, + 0.00920067448168993, + 0.02201153337955475, + 0.011505628935992718, + 0.013383309356868267, + 0.007925075478851795, + 0.020245159044861794, + 0.02516467496752739, + 0.013010382652282715, + 0.01632334291934967, + 0.02309040166437626, + 0.016035357490181923, + 0.022703571245074272, + 0.013304849155247211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.011076628230512142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057892084121704 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057892084121704, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057892084121704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0730467215180397, + 0.03433666005730629, + 0.046382468193769455, + 0.04636209085583687, + 0.11171482503414154, + 0.22417007386684418, + 0.0606568418443203, + 0.19398990273475647, + 0.11592592298984528, + 0.09341452270746231 + ], + "credit_map": [ + 0.011360046453773975, + 0.008273608051240444, + 0.009956509806215763, + 0.01429315097630024, + 0.008084448985755444, + 0.006678076460957527, + 0.007335646077990532, + 0.007065235171467066, + 0.0042120651341974735, + 0.0040473006665706635, + 0.004664112348109484, + 0.0038130658213049173, + 0.004223326221108437, + 0.004758934490382671, + 0.005350357387214899, + 0.0032674979884177446, + 0.006119688041508198, + 0.005512709729373455, + 0.0070320307277143, + 0.006724154576659203, + 0.0060736690647900105, + 0.004719430580735207, + 0.005898611154407263, + 0.004302177578210831, + 0.007775037549436092, + 0.005348884034901857, + 0.004104417748749256, + 0.0042876992374658585, + 0.007699752692133188, + 0.006963586900383234, + 0.005818672012537718, + 0.004364039748907089, + 0.012840302661061287, + 0.017188409343361855, + 0.012416938319802284, + 0.008291434496641159, + 0.015497368760406971, + 0.03001478686928749, + 0.006421108730137348, + 0.0090444665402174, + 0.023167692124843597, + 0.0329180546104908, + 0.023211805149912834, + 0.03018495813012123, + 0.04095058888196945, + 0.022987883538007736, + 0.02981756068766117, + 0.020931538194417953, + 0.008027056232094765, + 0.008480018936097622, + 0.006415426731109619, + 0.007572402711957693, + 0.008349242620170116, + 0.007000713609158993, + 0.009487169794738293, + 0.0053248112089931965, + 0.02172406017780304, + 0.05030186474323273, + 0.0178538765758276, + 0.03608466312289238, + 0.019982902333140373, + 0.020004086196422577, + 0.017648808658123016, + 0.010389652103185654, + 0.013234872370958328, + 0.01440820749849081, + 0.007694311439990997, + 0.011228296905755997, + 0.031156662851572037, + 0.01454467885196209, + 0.014373470097780228, + 0.00928543135523796, + 0.013149437494575977, + 0.017897766083478928, + 0.008883904665708542, + 0.01126895472407341, + 0.01224473025649786, + 0.009577637538313866, + 0.013207859359681606, + 0.007184223737567663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.02057892084121704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941434413194656 + ], + "opposite_logprob_tokens": [ + -2.929941415786743 + ], + "logprob_reward": -0.054941434413194656, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941434413194656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06959950178861618, + 0.028277907520532608, + 0.04332384839653969, + 0.02551833726465702, + 0.07302636653184891, + 0.06873737275600433, + 0.17849470674991608, + 0.11530416458845139, + 0.22696349024772644, + 0.17075428366661072 + ], + "credit_map": [ + 0.008740689605474472, + 0.006784350611269474, + 0.01121017336845398, + 0.014851829037070274, + 0.00778354424983263, + 0.007450990378856659, + 0.008042976260185242, + 0.004734945949167013, + 0.003971761092543602, + 0.003491893643513322, + 0.0036194317508488894, + 0.0038807580713182688, + 0.0033535694237798452, + 0.004214073531329632, + 0.003073092084378004, + 0.0026733260601758957, + 0.006422691512852907, + 0.005531754344701767, + 0.006174631882458925, + 0.009637395851314068, + 0.003262864425778389, + 0.005714734550565481, + 0.0031385847833007574, + 0.003441188018769026, + 0.004139029420912266, + 0.003075061831623316, + 0.0036291165743023157, + 0.00299826986156404, + 0.00250619906000793, + 0.0046647838316857815, + 0.002115317853167653, + 0.002390559995546937, + 0.007728950120508671, + 0.009899153374135494, + 0.007202324457466602, + 0.007701530586928129, + 0.009319337084889412, + 0.020401587709784508, + 0.00391768105328083, + 0.006855804473161697, + 0.005663436371833086, + 0.009982743300497532, + 0.006891040131449699, + 0.007281618658453226, + 0.014842002652585506, + 0.0059472485445439816, + 0.011434156447649002, + 0.006695129442960024, + 0.01328276190906763, + 0.03183066472411156, + 0.015923671424388885, + 0.021793482825160027, + 0.036234401166439056, + 0.015942638739943504, + 0.028875773772597313, + 0.014611311256885529, + 0.018759600818157196, + 0.016474252566695213, + 0.013801383785903454, + 0.015279985032975674, + 0.014949539676308632, + 0.012375928461551666, + 0.013311948627233505, + 0.010351531207561493, + 0.025932062417268753, + 0.024179678410291672, + 0.015519714914262295, + 0.02158360555768013, + 0.06078990176320076, + 0.028066370636224747, + 0.030798358842730522, + 0.020093781873583794, + 0.02172129414975643, + 0.02272857166826725, + 0.01376333087682724, + 0.017320483922958374, + 0.030361464247107506, + 0.018684295937418938, + 0.026824744418263435, + 0.019350092858076096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.054941434413194656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera sweeps smoothly from left to right across a brightly lit, modern office, at eye level. It first captures a woman with dark brown hair tied in a neat ponytail, wearing a vibrant red, long-sleeved silk shirt. She gently rocks on her feet as she carefully flips through a stack of white documents. As the saucepan continues to simmer, she leans forward and places the pile on the polished white surface of a large conference table. Then, a second woman with short blond hair and a sophisticated charcoal-gray blazer appears, seated at the table. She reaches out with her right hand and steadily and deliberately picks up the documents. The camera finally settles on a third woman, standing at the head of the table, dressed in a professional, navy-blue, sheath dress. As the camera settles on her, she begins speaking with a confident expression, raising her right hand with an open palm to emphasize a point, holding the gesture for a moment, and then slowly and gracefully lowering it to her side. All this is set in a scene bathed in soft, natural sunlight, with large windows creating an atmosphere of professional focus and collaboration.", + "seed": 18353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_7936_seed_18353_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12937616351991893, + "vlm_logprob": -0.12937616351991893, + "vlm_yes_no_margin": 2.7250000119209288, + "vlm_token_credit": -0.12937616351991893, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42137736082077026 + ], + "opposite_logprob_tokens": [ + -1.171377420425415 + ], + "logprob_reward": -0.42137736082077026, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.42137736082077026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11846709251403809, + 0.06181662529706955, + 0.08945930004119873, + 0.08756260573863983, + 0.12702161073684692, + 0.19904252886772156, + 0.132572740316391, + 0.07150129228830338, + 0.0770348533987999, + 0.03552134707570076 + ], + "credit_map": [ + 0.018824944272637367, + 0.01900414749979973, + 0.013914293609559536, + 0.009084390476346016, + 0.009406818076968193, + 0.01971159502863884, + 0.006648984272032976, + 0.02187192067503929, + 0.006947338581085205, + 0.007794522680342197, + 0.011448720470070839, + 0.008165220730006695, + 0.00819371733814478, + 0.007128957659006119, + 0.00554787740111351, + 0.006590273231267929, + 0.01779867894947529, + 0.014438729733228683, + 0.015580806881189346, + 0.007509998045861721, + 0.010729536414146423, + 0.008095687255263329, + 0.006266802549362183, + 0.009039059281349182, + 0.01817462407052517, + 0.01754315197467804, + 0.018879985436797142, + 0.008455690927803516, + 0.006429088301956654, + 0.0055533237755298615, + 0.005981156136840582, + 0.00654558464884758, + 0.023580016568303108, + 0.019925542175769806, + 0.021990645676851273, + 0.02134878747165203, + 0.010004588402807713, + 0.011180935427546501, + 0.010549650527536869, + 0.008441440761089325, + 0.04976480081677437, + 0.030426332727074623, + 0.03484974429011345, + 0.021584011614322662, + 0.017973437905311584, + 0.016817543655633926, + 0.0150239123031497, + 0.012602749280631542, + 0.017797958105802536, + 0.0176963172852993, + 0.012231270782649517, + 0.01375709380954504, + 0.016885586082935333, + 0.016293544322252274, + 0.02497454732656479, + 0.012936411425471306, + 0.008405935950577259, + 0.011567210778594017, + 0.007518075406551361, + 0.013947089202702045, + 0.008048122748732567, + 0.008683303371071815, + 0.007689417339861393, + 0.005642138887196779, + 0.010366277769207954, + 0.01185239665210247, + 0.011164071969687939, + 0.010097583755850792, + 0.010635009966790676, + 0.006864210590720177, + 0.0067751542665064335, + 0.009280153550207615, + 0.003265179693698883, + 0.004832310602068901, + 0.0034861117601394653, + 0.0049000694416463375, + 0.003824882674962282, + 0.0054403068497776985, + 0.003835400566458702, + 0.005937085021287203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.42137736082077026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906563371419907 + ], + "opposite_logprob_tokens": [ + -2.4640655517578125 + ], + "logprob_reward": -0.08906563371419907, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906563371419907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23863039910793304, + 0.517076849937439, + 0.07469931244850159, + 0.04019377380609512, + 0.019811321049928665, + 0.022880280390381813, + 0.019521920010447502, + 0.011837814003229141, + 0.028942875564098358, + 0.026405425742268562 + ], + "credit_map": [ + 0.038862958550453186, + 0.02158336713910103, + 0.056561436504125595, + 0.020627429708838463, + 0.014538672752678394, + 0.03626801818609238, + 0.012998553924262524, + 0.03718997910618782, + 0.07364224642515182, + 0.08240047097206116, + 0.11304831504821777, + 0.060743946582078934, + 0.0341787189245224, + 0.06721780449151993, + 0.028791366145014763, + 0.05705396085977554, + 0.017005978152155876, + 0.015760432928800583, + 0.014236544258892536, + 0.006574796978384256, + 0.005171123426407576, + 0.005475210025906563, + 0.0030753216706216335, + 0.007399903144687414, + 0.007904485799372196, + 0.0049505895003676414, + 0.007329065352678299, + 0.005692342761904001, + 0.004208295606076717, + 0.003124527633190155, + 0.0034480111207813025, + 0.0035364553332328796, + 0.003179658204317093, + 0.002822587266564369, + 0.003854120383039117, + 0.0028450314421206713, + 0.0017490210011601448, + 0.0019224719144403934, + 0.0014513981295749545, + 0.0019870344549417496, + 0.0036005026195198298, + 0.0026730282697826624, + 0.0028665398713201284, + 0.002196453744545579, + 0.001968301134184003, + 0.005250993184745312, + 0.0017222230089828372, + 0.002602237043902278, + 0.0020651021040976048, + 0.0028530750423669815, + 0.0021423539146780968, + 0.0025219956878572702, + 0.002408177126199007, + 0.0022396654821932316, + 0.002916680183261633, + 0.0023748704697936773, + 0.0019349603680893779, + 0.0019458226161077619, + 0.0016752445371821523, + 0.002223012037575245, + 0.0011946505401283503, + 0.0012161597842350602, + 0.0007821621256880462, + 0.0008658014121465385, + 0.0039112442173063755, + 0.004349015187472105, + 0.00422300212085247, + 0.004569303244352341, + 0.0035362469498068094, + 0.003682304173707962, + 0.0023840523790568113, + 0.002287707757204771, + 0.004622948355972767, + 0.004582122899591923, + 0.003179320367053151, + 0.004217194393277168, + 0.0026005571708083153, + 0.0032882618252187967, + 0.001989379059523344, + 0.0019256423693150282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.08906563371419907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978801865130663 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.00978801865130663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978801865130663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16787296533584595, + 0.26586371660232544, + 0.1106899306178093, + 0.08198146522045135, + 0.09872175008058548, + 0.06775528192520142, + 0.06434448063373566, + 0.06077621504664421, + 0.05802260711789131, + 0.023971570655703545 + ], + "credit_map": [ + 0.019272465258836746, + 0.030456526204943657, + 0.01369958370923996, + 0.01343626156449318, + 0.028299605473876, + 0.024458639323711395, + 0.011385442689061165, + 0.026864441111683846, + 0.035459574311971664, + 0.038029663264751434, + 0.0513136200606823, + 0.0262632817029953, + 0.03269459679722786, + 0.03320595249533653, + 0.02205595001578331, + 0.02684108540415764, + 0.02048531360924244, + 0.019624359905719757, + 0.021251684054732323, + 0.008387298323214054, + 0.013159448280930519, + 0.011415406130254269, + 0.007899417541921139, + 0.008467007428407669, + 0.018714798614382744, + 0.015500745736062527, + 0.008618702180683613, + 0.00658043147996068, + 0.010536870919167995, + 0.00780075415968895, + 0.006113534327596426, + 0.008115620352327824, + 0.013223852962255478, + 0.013100403361022472, + 0.014934033155441284, + 0.014419656246900558, + 0.008868584409356117, + 0.014716117642819881, + 0.007950843311846256, + 0.01150825247168541, + 0.0075059859082102776, + 0.008604338392615318, + 0.007884244434535503, + 0.007657809183001518, + 0.009617812000215054, + 0.013744515366852283, + 0.006131114438176155, + 0.0066094608046114445, + 0.009158763103187084, + 0.008214554749429226, + 0.011735630221664906, + 0.006912207696586847, + 0.007260110694915056, + 0.006455838680267334, + 0.009471134282648563, + 0.005136237479746342, + 0.006965182721614838, + 0.009949298575520515, + 0.005562308710068464, + 0.01346668228507042, + 0.006689473986625671, + 0.009882711805403233, + 0.0041529168374836445, + 0.004107635002583265, + 0.007703989278525114, + 0.008735309354960918, + 0.0076293302699923515, + 0.007873504422605038, + 0.00792551226913929, + 0.006115280091762543, + 0.004765670746564865, + 0.007274008821696043, + 0.002228304510936141, + 0.003820827230811119, + 0.0027187836822122335, + 0.004367005545645952, + 0.0025449569802731276, + 0.0022328393533825874, + 0.0032101720571517944, + 0.0028486817609518766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.00978801865130663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003054678440094 + ], + "opposite_logprob_tokens": [ + -2.3503055572509766 + ], + "logprob_reward": -0.1003054678440094, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003054678440094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1126350462436676, + 0.11132539808750153, + 0.06816939264535904, + 0.08792489767074585, + 0.10276615619659424, + 0.060909170657396317, + 0.19399598240852356, + 0.10729790478944778, + 0.06567971408367157, + 0.08929634839296341 + ], + "credit_map": [ + 0.011937640607357025, + 0.016893206164240837, + 0.01384011097252369, + 0.010902751237154007, + 0.010060608386993408, + 0.02034207247197628, + 0.007664763368666172, + 0.020993893966078758, + 0.012433948926627636, + 0.013026082888245583, + 0.01596411131322384, + 0.011516759172081947, + 0.022749943658709526, + 0.013791454024612904, + 0.011507256887853146, + 0.010335845872759819, + 0.012639195658266544, + 0.012892225757241249, + 0.01224104780703783, + 0.007243984378874302, + 0.00800318643450737, + 0.005244496278464794, + 0.004611042328178883, + 0.005294214002788067, + 0.013701156713068485, + 0.009826426394283772, + 0.010763063095510006, + 0.00893525779247284, + 0.014786510728299618, + 0.010036741383373737, + 0.011522670276463032, + 0.008353063836693764, + 0.01807958073914051, + 0.01330301072448492, + 0.01713910512626171, + 0.013741528615355492, + 0.009315872564911842, + 0.015111738815903664, + 0.009195612743496895, + 0.006879706401377916, + 0.006331631913781166, + 0.007436273619532585, + 0.006106743589043617, + 0.007146923802793026, + 0.008683418855071068, + 0.012018156237900257, + 0.006057138554751873, + 0.007128883618861437, + 0.016999680548906326, + 0.023742275312542915, + 0.01746371202170849, + 0.015458906069397926, + 0.027238324284553528, + 0.0415814071893692, + 0.026003912091255188, + 0.025507766753435135, + 0.011820986866950989, + 0.015302526764571667, + 0.014662831090390682, + 0.01264896523207426, + 0.014795237220823765, + 0.017885388806462288, + 0.012760570272803307, + 0.007421394810080528, + 0.007542951963841915, + 0.009960077702999115, + 0.006588211748749018, + 0.009947371669113636, + 0.0073685855604708195, + 0.009820426814258099, + 0.00716626551002264, + 0.007285820785909891, + 0.009115314111113548, + 0.01431227196007967, + 0.011036008596420288, + 0.013565793633460999, + 0.008724953979253769, + 0.011906436644494534, + 0.010939065366983414, + 0.009696504101157188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.1003054678440094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344336569309235 + ], + "opposite_logprob_tokens": [ + -3.6513442993164062 + ], + "logprob_reward": -0.026344336569309235, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344336569309235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1265931874513626, + 0.09727830439805984, + 0.10102865844964981, + 0.05525148659944534, + 0.07818886637687683, + 0.08916360139846802, + 0.09008818119764328, + 0.15315493941307068, + 0.13230465352535248, + 0.0769481211900711 + ], + "credit_map": [ + 0.014503171667456627, + 0.013114385306835175, + 0.02390334941446781, + 0.008689545094966888, + 0.01038429606705904, + 0.024559155106544495, + 0.0073968213982880116, + 0.02404245361685753, + 0.011519737541675568, + 0.01488479319959879, + 0.015896791592240334, + 0.014913615770637989, + 0.014761819504201412, + 0.011050534434616566, + 0.006865530274808407, + 0.0073854802176356316, + 0.018948879092931747, + 0.014649607241153717, + 0.027799982577562332, + 0.007157639600336552, + 0.008193754591047764, + 0.009382596239447594, + 0.005249278619885445, + 0.009646919555962086, + 0.00989879947155714, + 0.00838150642812252, + 0.00813443772494793, + 0.00793214701116085, + 0.005774345714598894, + 0.004616522695869207, + 0.004947618581354618, + 0.00556610943749547, + 0.009776911698281765, + 0.009244757704436779, + 0.014508492313325405, + 0.01209560688585043, + 0.006643085274845362, + 0.010665190406143665, + 0.006307974457740784, + 0.008946849033236504, + 0.016731025651097298, + 0.01268812082707882, + 0.012099510990083218, + 0.011509167961776257, + 0.009258609265089035, + 0.010534903965890408, + 0.009316989220678806, + 0.007025271654129028, + 0.00872130785137415, + 0.010174558497965336, + 0.009074741043150425, + 0.007253666408360004, + 0.016038082540035248, + 0.015537422150373459, + 0.013357141986489296, + 0.009931260719895363, + 0.018454160541296005, + 0.02381512150168419, + 0.020036879926919937, + 0.024628527462482452, + 0.012333360500633717, + 0.017511608079075813, + 0.0190499946475029, + 0.017325295135378838, + 0.018202368170022964, + 0.019833065569400787, + 0.016068244352936745, + 0.014090260490775108, + 0.01804967224597931, + 0.01585165224969387, + 0.01615033484995365, + 0.014059057459235191, + 0.010201610624790192, + 0.012816665694117546, + 0.00930743757635355, + 0.010683929547667503, + 0.00794565211981535, + 0.007655196823179722, + 0.007535669952630997, + 0.010801960714161396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.026344336569309235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera sweeps smoothly from left to right across a brightly lit, modern office, at eye level. It first captures a woman with dark brown hair tied in a neat ponytail, wearing a vibrant red, long-sleeved silk shirt. She gently rocks on her feet as she carefully flips through a stack of white documents. As the saucepan continues to simmer, she leans forward and places the pile on the polished white surface of a large conference table. Then, a second woman with short blond hair and a sophisticated charcoal-gray blazer appears, seated at the table. She reaches out with her right hand and steadily and deliberately picks up the documents. The camera finally settles on a third woman, standing at the head of the table, dressed in a professional, navy-blue, sheath dress. As the camera settles on her, she begins speaking with a confident expression, raising her right hand with an open palm to emphasize a point, holding the gesture for a moment, and then slowly and gracefully lowering it to her side. All this is set in a scene bathed in soft, natural sunlight, with large windows creating an atmosphere of professional focus and collaboration.", + "seed": 18354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_7936_seed_18354_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08329907478764653, + "vlm_logprob": -0.08329907478764653, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.08329907478764653, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22077487409114838 + ], + "opposite_logprob_tokens": [ + -1.8457748889923096 + ], + "logprob_reward": -0.22077487409114838, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.22077487409114838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13524512946605682, + 0.12441688030958176, + 0.06894499808549881, + 0.07463761419057846, + 0.10843545198440552, + 0.24668464064598083, + 0.07577449083328247, + 0.06617268919944763, + 0.05854557082056999, + 0.04114251583814621 + ], + "credit_map": [ + 0.012528609484434128, + 0.016667351126670837, + 0.012169108726084232, + 0.012476007454097271, + 0.023144302889704704, + 0.017956716939806938, + 0.02582884021103382, + 0.014474190771579742, + 0.013874786905944347, + 0.009856189601123333, + 0.010392931289970875, + 0.016072655096650124, + 0.02164646051824093, + 0.018240973353385925, + 0.020539594814181328, + 0.013793288730084896, + 0.006655063014477491, + 0.00535224424675107, + 0.008707376196980476, + 0.008458588272333145, + 0.010585000738501549, + 0.010450858622789383, + 0.008889589458703995, + 0.009846271015703678, + 0.00729807373136282, + 0.005508108995854855, + 0.009214990772306919, + 0.014466388151049614, + 0.012220405973494053, + 0.011380931362509727, + 0.007624720688909292, + 0.006923995912075043, + 0.008260371163487434, + 0.012217655777931213, + 0.01159138884395361, + 0.01087142899632454, + 0.017402147874236107, + 0.015286928974092007, + 0.01927630603313446, + 0.013529232703149319, + 0.025160284712910652, + 0.08758772164583206, + 0.03009680286049843, + 0.023468373343348503, + 0.013656506314873695, + 0.020763646811246872, + 0.01024822611361742, + 0.03570307046175003, + 0.009729931131005287, + 0.011696483939886093, + 0.012969371862709522, + 0.00807436928153038, + 0.006617330946028233, + 0.01334870420396328, + 0.006043531931936741, + 0.007294770795851946, + 0.009267756715416908, + 0.007347797974944115, + 0.012954977340996265, + 0.00976064894348383, + 0.0063719479367136955, + 0.005258764140307903, + 0.007812505587935448, + 0.007398285903036594, + 0.010826971381902695, + 0.009770178236067295, + 0.007322140969336033, + 0.005787404719740152, + 0.005143098067492247, + 0.009922309778630733, + 0.004021530505269766, + 0.005751938093453646, + 0.0033532604575157166, + 0.00545528857037425, + 0.004503733944147825, + 0.006194733548909426, + 0.005761913489550352, + 0.00555590121075511, + 0.00590702285990119, + 0.0044106608256697655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.22077487409114838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372080624103546 + ], + "opposite_logprob_tokens": [ + -3.4087207317352295 + ], + "logprob_reward": -0.03372080624103546, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372080624103546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25667598843574524, + 0.15733440220355988, + 0.08998148143291473, + 0.060830116271972656, + 0.023755814880132675, + 0.1351677030324936, + 0.06285229325294495, + 0.04962972551584244, + 0.04658305272459984, + 0.11718940734863281 + ], + "credit_map": [ + 0.029459955170750618, + 0.03280887007713318, + 0.03528591990470886, + 0.03059687465429306, + 0.03360921889543533, + 0.04240048676729202, + 0.02778218686580658, + 0.024732477962970734, + 0.028769303113222122, + 0.020865626633167267, + 0.013896016404032707, + 0.017274152487516403, + 0.021450936794281006, + 0.019864648580551147, + 0.01695401966571808, + 0.018259694799780846, + 0.008860945701599121, + 0.005368008278310299, + 0.007711718790233135, + 0.008013858459889889, + 0.008814756758511066, + 0.026382915675640106, + 0.011246117763221264, + 0.013583164662122726, + 0.0056768436916172504, + 0.005383060779422522, + 0.005689200013875961, + 0.011089934036135674, + 0.008001839742064476, + 0.008871673606336117, + 0.007584637030959129, + 0.008532927371561527, + 0.0019799829460680485, + 0.0030012805946171284, + 0.002992692170664668, + 0.002555510262027383, + 0.0033047054894268513, + 0.0038177967071533203, + 0.0030589704401791096, + 0.0030448753386735916, + 0.014981864020228386, + 0.049889545887708664, + 0.017661742866039276, + 0.013437764719128609, + 0.007915189489722252, + 0.010128713212907314, + 0.004129091743379831, + 0.01702379435300827, + 0.009657987393438816, + 0.01154699083417654, + 0.005650409962981939, + 0.003252316964790225, + 0.004973842762410641, + 0.01916753128170967, + 0.004691144451498985, + 0.00391206843778491, + 0.0038044233806431293, + 0.005531147122383118, + 0.006854064762592316, + 0.003817061660811305, + 0.008422808721661568, + 0.006840542424470186, + 0.008060436695814133, + 0.006299241911619902, + 0.0038509357254952192, + 0.006481162738054991, + 0.009260260500013828, + 0.004322494380176067, + 0.0044537815265357494, + 0.00868740864098072, + 0.0034737405367195606, + 0.006053270306438208, + 0.01096008438616991, + 0.014608852565288544, + 0.011645970866084099, + 0.015112318098545074, + 0.01858007349073887, + 0.013706812635064125, + 0.01500757411122322, + 0.017567727714776993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03372080624103546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09825217723846436, + 0.0792379379272461, + 0.056661207228899, + 0.4202221632003784, + 0.09862066805362701, + 0.06179715692996979, + 0.10263140499591827, + 0.023064039647579193, + 0.027885928750038147, + 0.03162733465433121 + ], + "credit_map": [ + 0.007594980765134096, + 0.00832296535372734, + 0.009250693023204803, + 0.009229307994246483, + 0.011700499802827835, + 0.022880196571350098, + 0.015252957120537758, + 0.014020571485161781, + 0.007607182487845421, + 0.008701004087924957, + 0.007245848886668682, + 0.00755118764936924, + 0.01871001347899437, + 0.009974515996873379, + 0.011173932813107967, + 0.008274258114397526, + 0.003238022793084383, + 0.004295151215046644, + 0.00739341089501977, + 0.006277693435549736, + 0.007727364543825388, + 0.009622333571314812, + 0.008564917370676994, + 0.009542316198348999, + 0.05002376809716225, + 0.04645957797765732, + 0.05043714493513107, + 0.0758737102150917, + 0.04073302075266838, + 0.050387781113386154, + 0.06336447596549988, + 0.04294271394610405, + 0.0074917771853506565, + 0.01468114834278822, + 0.015242825262248516, + 0.012563706375658512, + 0.01721806637942791, + 0.0073156338185071945, + 0.015254771336913109, + 0.008852731436491013, + 0.005623951088637114, + 0.018584677949547768, + 0.008359691128134727, + 0.005747515242546797, + 0.005217486992478371, + 0.005594744812697172, + 0.0039694905281066895, + 0.008699599653482437, + 0.018678098917007446, + 0.02012341096997261, + 0.009797291830182076, + 0.004993916023522615, + 0.004941857419908047, + 0.03208767622709274, + 0.004556499887257814, + 0.007452651858329773, + 0.0025583039969205856, + 0.0023478863295167685, + 0.004157563205808401, + 0.0012101586908102036, + 0.004046217072755098, + 0.002514258725568652, + 0.003553207265213132, + 0.0026764445938169956, + 0.0034066797234117985, + 0.003539796220138669, + 0.0075010620057582855, + 0.0037490669637918472, + 0.002598375780507922, + 0.0028868101071566343, + 0.0020295034628361464, + 0.0021746333222836256, + 0.0035920129157602787, + 0.003967630676925182, + 0.00406121788546443, + 0.0043964325450360775, + 0.005186424124985933, + 0.004700826946645975, + 0.0033043292351067066, + 0.0024184593930840492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.006745188031345606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031431168317795 + ], + "opposite_logprob_tokens": [ + -2.3503143787384033 + ], + "logprob_reward": -0.10031431168317795, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031431168317795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11999063193798065, + 0.08327408879995346, + 0.08072417974472046, + 0.06338471174240112, + 0.14993685483932495, + 0.11984812468290329, + 0.06166251003742218, + 0.13297419250011444, + 0.07371708005666733, + 0.11448763310909271 + ], + "credit_map": [ + 0.00884804967790842, + 0.007288347464054823, + 0.010246645659208298, + 0.010226440615952015, + 0.017463989555835724, + 0.028640814125537872, + 0.019886277616024017, + 0.017390063032507896, + 0.008421617560088634, + 0.010384364053606987, + 0.007978237234055996, + 0.009385526180267334, + 0.015093273483216763, + 0.010721804574131966, + 0.011898906901478767, + 0.009390359744429588, + 0.006150220986455679, + 0.0049100276082754135, + 0.005715105216950178, + 0.007727531716227531, + 0.017504803836345673, + 0.014104163274168968, + 0.014830670319497585, + 0.009781653992831707, + 0.007053546141833067, + 0.005425571464002132, + 0.009035283699631691, + 0.011151469312608242, + 0.007961285300552845, + 0.008472318761050701, + 0.008828211575746536, + 0.005457027815282345, + 0.011224946938455105, + 0.015158550813794136, + 0.01655563712120056, + 0.016380716115236282, + 0.028628576546907425, + 0.02141617052257061, + 0.024218443781137466, + 0.016353799030184746, + 0.015394984744489193, + 0.010310055688023567, + 0.012088939547538757, + 0.011029123328626156, + 0.015879224985837936, + 0.014529788866639137, + 0.012735208496451378, + 0.027880799025297165, + 0.01027277298271656, + 0.011046206578612328, + 0.006091536488384008, + 0.0034119777847081423, + 0.004939942620694637, + 0.013436597771942616, + 0.005633531138300896, + 0.006829944904893637, + 0.010557078756392002, + 0.017496317625045776, + 0.008008957840502262, + 0.009267487563192844, + 0.026152461767196655, + 0.01584569737315178, + 0.02381829358637333, + 0.02182788960635662, + 0.010168834589421749, + 0.009478619322180748, + 0.009887326508760452, + 0.006790100131183863, + 0.012638176791369915, + 0.009025057777762413, + 0.006718610413372517, + 0.00901035126298666, + 0.009279616177082062, + 0.01613987609744072, + 0.01744471862912178, + 0.009098340757191181, + 0.017673036083579063, + 0.0196550153195858, + 0.015155667439103127, + 0.010041358880698681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.10031431168317795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494019389152527 + ], + "opposite_logprob_tokens": [ + -2.9299402236938477 + ], + "logprob_reward": -0.05494019389152527, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494019389152527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05943647772073746, + 0.04561605677008629, + 0.09883920103311539, + 0.05208636075258255, + 0.04698089137673378, + 0.14788110554218292, + 0.21024970710277557, + 0.05941814184188843, + 0.11793666332960129, + 0.1615554243326187 + ], + "credit_map": [ + 0.003932660445570946, + 0.006322769448161125, + 0.005872874520719051, + 0.006669892929494381, + 0.009578736498951912, + 0.012122021988034248, + 0.007632440887391567, + 0.007305081468075514, + 0.0053697493858635426, + 0.0051877666264772415, + 0.003628441598266363, + 0.0055693150497972965, + 0.008300983346998692, + 0.006210241932421923, + 0.006380036938935518, + 0.004969519563019276, + 0.008073804900050163, + 0.006134075112640858, + 0.007435083854943514, + 0.007246064953505993, + 0.01085554901510477, + 0.03167765587568283, + 0.013592777773737907, + 0.013824193738400936, + 0.005970440339297056, + 0.0057959239929914474, + 0.006489387713372707, + 0.010185382328927517, + 0.004927075933665037, + 0.005679371301084757, + 0.008114896714687347, + 0.004923883359879255, + 0.004108033608645201, + 0.00519188866019249, + 0.004030704498291016, + 0.0063638933934271336, + 0.008332259021699429, + 0.0077934409491717815, + 0.006171539891511202, + 0.004989132285118103, + 0.011708357371389866, + 0.04681245610117912, + 0.016010798513889313, + 0.0122069688513875, + 0.010238268412649632, + 0.014525136910378933, + 0.007622464094310999, + 0.028756648302078247, + 0.03761134296655655, + 0.03623313084244728, + 0.019786924123764038, + 0.010111146606504917, + 0.020262224599719048, + 0.049594733864068985, + 0.01882011629641056, + 0.017830077558755875, + 0.007303330115973949, + 0.006384368985891342, + 0.009173765778541565, + 0.0046094078570604324, + 0.008293695747852325, + 0.00542127899825573, + 0.012380258180201054, + 0.005852037575095892, + 0.014534862712025642, + 0.015340331010520458, + 0.02636547200381756, + 0.012082666158676147, + 0.011115677654743195, + 0.016013039276003838, + 0.011608239263296127, + 0.01087637897580862, + 0.010133759118616581, + 0.022904515266418457, + 0.015051167458295822, + 0.015862036496400833, + 0.028224749490618706, + 0.028547653928399086, + 0.025769520550966263, + 0.015062031336128712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.05494019389152527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera sweeps smoothly from left to right across a brightly lit, modern office, at eye level. It first captures a woman with dark brown hair tied in a neat ponytail, wearing a vibrant red, long-sleeved silk shirt. She gently rocks on her feet as she carefully flips through a stack of white documents. As the saucepan continues to simmer, she leans forward and places the pile on the polished white surface of a large conference table. Then, a second woman with short blond hair and a sophisticated charcoal-gray blazer appears, seated at the table. She reaches out with her right hand and steadily and deliberately picks up the documents. The camera finally settles on a third woman, standing at the head of the table, dressed in a professional, navy-blue, sheath dress. As the camera settles on her, she begins speaking with a confident expression, raising her right hand with an open palm to emphasize a point, holding the gesture for a moment, and then slowly and gracefully lowering it to her side. All this is set in a scene bathed in soft, natural sunlight, with large windows creating an atmosphere of professional focus and collaboration.", + "seed": 18355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000112/group_9/index_7936_seed_18355_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0635952316224575, + "vlm_logprob": -0.0635952316224575, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.0635952316224575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19298551976680756 + ], + "opposite_logprob_tokens": [ + -1.9429855346679688 + ], + "logprob_reward": -0.19298551976680756, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19298551976680756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1351058930158615, + 0.08692679554224014, + 0.0900820940732956, + 0.053693365305662155, + 0.13231737911701202, + 0.35721302032470703, + 0.05104673281311989, + 0.013392592780292034, + 0.0671272724866867, + 0.01309486199170351 + ], + "credit_map": [ + 0.026243578642606735, + 0.024672118946909904, + 0.023059075698256493, + 0.01789022423326969, + 0.015119017101824284, + 0.010712712071835995, + 0.008820747025310993, + 0.008588419295847416, + 0.012138647958636284, + 0.009216872975230217, + 0.012345460243523121, + 0.011424744501709938, + 0.008390089496970177, + 0.01773451454937458, + 0.005732601974159479, + 0.009943860583007336, + 0.010145852342247963, + 0.010122212581336498, + 0.020298117771744728, + 0.009405655786395073, + 0.00842412281781435, + 0.01432445365935564, + 0.006058874074369669, + 0.011302811093628407, + 0.006904028356075287, + 0.0052267154678702354, + 0.008323444984853268, + 0.008815720677375793, + 0.007380640134215355, + 0.006754973437637091, + 0.005247431341558695, + 0.005040409043431282, + 0.015388024039566517, + 0.01224270649254322, + 0.021001700311899185, + 0.023934917524456978, + 0.013472714461386204, + 0.023633327335119247, + 0.011388364247977734, + 0.011255632154643536, + 0.031838323920965195, + 0.06708674877882004, + 0.03717906400561333, + 0.11412229388952255, + 0.02597351372241974, + 0.02263418398797512, + 0.022428246214985847, + 0.03595063090324402, + 0.009698123671114445, + 0.005522941239178181, + 0.007911840453743935, + 0.0056513212621212006, + 0.004631310235708952, + 0.0038357810117304325, + 0.0051976521499454975, + 0.008597765117883682, + 0.001968821743503213, + 0.001242646249011159, + 0.0015424685552716255, + 0.0013174426276236773, + 0.0019135577604174614, + 0.0015257636550813913, + 0.0021945415064692497, + 0.0016873519634827971, + 0.00891287624835968, + 0.0027034927625209093, + 0.0026277131401002407, + 0.006679217331111431, + 0.00511707179248333, + 0.006020789500325918, + 0.007710759993642569, + 0.02735535427927971, + 0.0013710753992199898, + 0.0011860303347930312, + 0.001419308828189969, + 0.001909457496367395, + 0.0022756829857826233, + 0.0010817400179803371, + 0.0021411566995084286, + 0.001710411161184311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.19298551976680756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308921605348587 + ], + "opposite_logprob_tokens": [ + -3.7733089923858643 + ], + "logprob_reward": -0.023308921605348587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308921605348587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3541751801967621, + 0.0761333703994751, + 0.06624796241521835, + 0.12651298940181732, + 0.07484354078769684, + 0.10657185316085815, + 0.04052341729402542, + 0.02045642025768757, + 0.11323966085910797, + 0.02129562385380268 + ], + "credit_map": [ + 0.0490090511739254, + 0.05872964486479759, + 0.050783466547727585, + 0.062306102365255356, + 0.037569839507341385, + 0.03855489194393158, + 0.031978532671928406, + 0.025243647396564484, + 0.006285171955823898, + 0.010011117905378342, + 0.007754845544695854, + 0.009596330113708973, + 0.008047056384384632, + 0.01813441514968872, + 0.006117240991443396, + 0.010187197476625443, + 0.006460567004978657, + 0.006126982159912586, + 0.015754161402583122, + 0.009781711734831333, + 0.005479886662214994, + 0.010548701509833336, + 0.0043376749381423, + 0.007758286781609058, + 0.0128821711987257, + 0.018542639911174774, + 0.014384725131094456, + 0.017845947295427322, + 0.018852436915040016, + 0.018439488485455513, + 0.012054478749632835, + 0.013511092402040958, + 0.00891850795596838, + 0.012386571615934372, + 0.009755171835422516, + 0.014031589031219482, + 0.006497368682175875, + 0.010828623548150063, + 0.005582449026405811, + 0.00684325722977519, + 0.008965162560343742, + 0.0186797883361578, + 0.013087323866784573, + 0.034992240369319916, + 0.005775123834609985, + 0.007236816454678774, + 0.006521909032016993, + 0.01131348591297865, + 0.008702464401721954, + 0.005166533403098583, + 0.005018843337893486, + 0.004088744521141052, + 0.006682177539914846, + 0.0032939149532467127, + 0.004005517344921827, + 0.003565223189070821, + 0.0024204805959016085, + 0.001750453026033938, + 0.00252172420732677, + 0.0024375112261623144, + 0.0020600389689207077, + 0.002386081963777542, + 0.0020243842154741287, + 0.004855746403336525, + 0.016174664720892906, + 0.004013726022094488, + 0.0033458261750638485, + 0.011226257309317589, + 0.0052847410552203655, + 0.010992921888828278, + 0.012502791360020638, + 0.04969872906804085, + 0.0024284624960273504, + 0.002670422662049532, + 0.0026744224596768618, + 0.0031353647354990244, + 0.0035094835329800844, + 0.0024596164003014565, + 0.0022345534525811672, + 0.002183298347517848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.023308921605348587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16322484612464905, + 0.08803126215934753, + 0.061770983040332794, + 0.052253223955631256, + 0.15413255989551544, + 0.12350521981716156, + 0.08604959398508072, + 0.03597415238618851, + 0.1831168383359909, + 0.05194127932190895 + ], + "credit_map": [ + 0.021878143772482872, + 0.022793259471654892, + 0.031683843582868576, + 0.015543089248239994, + 0.018020058050751686, + 0.020281648263335228, + 0.014065009541809559, + 0.018959783017635345, + 0.00713181309401989, + 0.01269841194152832, + 0.006485906895250082, + 0.009331152774393559, + 0.012178315781056881, + 0.022808630019426346, + 0.007379413582384586, + 0.010017617605626583, + 0.005692463833838701, + 0.00520306034013629, + 0.013386668637394905, + 0.007484756410121918, + 0.007437955588102341, + 0.007761766668409109, + 0.009117821231484413, + 0.005686488468199968, + 0.004350648727267981, + 0.004161303397268057, + 0.00907878763973713, + 0.01097968127578497, + 0.006135653704404831, + 0.007519126404076815, + 0.0047494955360889435, + 0.0052785249426960945, + 0.01909492537379265, + 0.03083720989525318, + 0.01815936341881752, + 0.020262662321329117, + 0.014214585535228252, + 0.028216568753123283, + 0.009907931089401245, + 0.013439326547086239, + 0.011002114973962307, + 0.02140706218779087, + 0.014788446947932243, + 0.037610575556755066, + 0.008705979213118553, + 0.008542598225176334, + 0.008645712397992611, + 0.012802745215594769, + 0.017194734886288643, + 0.00981070939451456, + 0.015449215658009052, + 0.011777501553297043, + 0.008447918109595776, + 0.005958135239779949, + 0.008376570418477058, + 0.009034807793796062, + 0.005032353103160858, + 0.0026257915887981653, + 0.0072194961830973625, + 0.004238212946802378, + 0.005030250176787376, + 0.003139002248644829, + 0.0049139061011374, + 0.0037751400377601385, + 0.025608636438846588, + 0.006328654941171408, + 0.006273822393268347, + 0.017680376768112183, + 0.0073846252635121346, + 0.018972603604197502, + 0.019892429932951927, + 0.08097569644451141, + 0.005845296662300825, + 0.003101970301941037, + 0.004560357891023159, + 0.007358152884989977, + 0.009791289456188679, + 0.0059257736429572105, + 0.010174814611673355, + 0.005183625500649214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809649869799614 + ], + "opposite_logprob_tokens": [ + -3.2880964279174805 + ], + "logprob_reward": -0.03809649869799614, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809649869799614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10193219780921936, + 0.05367358401417732, + 0.06240228936076164, + 0.02949213795363903, + 0.08287106454372406, + 0.14599360525608063, + 0.12316897511482239, + 0.024291042238473892, + 0.3109881281852722, + 0.0651869848370552 + ], + "credit_map": [ + 0.016251999884843826, + 0.01622178591787815, + 0.019654855132102966, + 0.010382267646491528, + 0.012539496645331383, + 0.01038071047514677, + 0.007592718582600355, + 0.008908362127840519, + 0.005130822770297527, + 0.0069922832772135735, + 0.004191163927316666, + 0.0077402014285326, + 0.005357667338103056, + 0.014075168408453465, + 0.004090479109436274, + 0.006095795426517725, + 0.006597544066607952, + 0.005709396675229073, + 0.015157824382185936, + 0.006905827671289444, + 0.006808009464293718, + 0.009277697652578354, + 0.0055480170994997025, + 0.006397973746061325, + 0.0022932561114430428, + 0.0025668079033493996, + 0.0053722751326859, + 0.0061568161472678185, + 0.0031236920040100813, + 0.004385304171591997, + 0.002799409907311201, + 0.0027945751789957285, + 0.01022808812558651, + 0.015272707678377628, + 0.009746798314154148, + 0.014163308776915073, + 0.005572907160967588, + 0.013403855264186859, + 0.005708768032491207, + 0.008774633519351482, + 0.013323469087481499, + 0.026020236313343048, + 0.015898054465651512, + 0.04529130086302757, + 0.011929985135793686, + 0.009570552036166191, + 0.009938693605363369, + 0.014021316543221474, + 0.01944163627922535, + 0.014901355840265751, + 0.020982669666409492, + 0.019442671909928322, + 0.014117428101599216, + 0.009482666850090027, + 0.013122983276844025, + 0.011677569709718227, + 0.0028743313159793615, + 0.0020064623095095158, + 0.004294566810131073, + 0.0024024026934057474, + 0.002829672070220113, + 0.0020966597367078066, + 0.0029800680931657553, + 0.004806878976523876, + 0.04384072497487068, + 0.010793422348797321, + 0.00960273016244173, + 0.030652549117803574, + 0.011787046678364277, + 0.03229415789246559, + 0.03366611525416374, + 0.1383514255285263, + 0.007818921469151974, + 0.0061348117887973785, + 0.00764111801981926, + 0.008798387832939625, + 0.011056453920900822, + 0.005633690394461155, + 0.01053061243146658, + 0.007572993170469999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.03809649869799614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera sweep smoothly from left to right across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit, modern office with large windows and soft natural sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the first woman wearing a vibrant red, long-sleeved silk shirt and has dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second woman have short blond hair and wear a charcoal-gray blazer while picking up documents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the third woman standing at the head of the table wearing a navy-blue sheath dress and raising her right hand to speak?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494493246078491 + ], + "opposite_logprob_tokens": [ + -2.9299449920654297 + ], + "logprob_reward": -0.05494493246078491, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494493246078491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06678478419780731, + 0.055776484310626984, + 0.07132653146982193, + 0.03524992614984512, + 0.04911455139517784, + 0.1778370440006256, + 0.08167025446891785, + 0.05275067314505577, + 0.2072775959968567, + 0.20221218466758728 + ], + "credit_map": [ + 0.00840030051767826, + 0.010492891073226929, + 0.008679560385644436, + 0.0077469684183597565, + 0.008142075501382351, + 0.006474115885794163, + 0.008220277726650238, + 0.00862859282642603, + 0.006124056875705719, + 0.005279960110783577, + 0.006674282252788544, + 0.00891480129212141, + 0.008547700941562653, + 0.00984163861721754, + 0.005169492214918137, + 0.005224548280239105, + 0.008364715613424778, + 0.0070036305114626884, + 0.014819271862506866, + 0.007916617207229137, + 0.0070640575140714645, + 0.010038722306489944, + 0.007634749636054039, + 0.008484771475195885, + 0.00478522339835763, + 0.0030240232590585947, + 0.004862749949097633, + 0.005068525206297636, + 0.005004173144698143, + 0.0053767547942698, + 0.003244910156354308, + 0.00388357019983232, + 0.0062765683978796005, + 0.006311265751719475, + 0.0061324783600866795, + 0.008196594193577766, + 0.005032673478126526, + 0.007147262338548899, + 0.004633413627743721, + 0.0053842943161726, + 0.015699069947004318, + 0.034509994089603424, + 0.020262517035007477, + 0.053973566740751266, + 0.012255478650331497, + 0.012867639772593975, + 0.011728261597454548, + 0.016540508717298508, + 0.012685995548963547, + 0.00788934901356697, + 0.010825837962329388, + 0.012011829763650894, + 0.010548428632318974, + 0.007430317346006632, + 0.01091244351118803, + 0.009366055019199848, + 0.004288531374186277, + 0.003301982767879963, + 0.006579069420695305, + 0.004416703712195158, + 0.008200122974812984, + 0.005337637849152088, + 0.008404317311942577, + 0.012222304940223694, + 0.026415476575493813, + 0.0076491134241223335, + 0.013291388750076294, + 0.018709201365709305, + 0.015809373930096626, + 0.020974164828658104, + 0.023492641746997833, + 0.0809362530708313, + 0.02194102853536606, + 0.012727348133921623, + 0.01387128047645092, + 0.023443907499313354, + 0.04142962023615837, + 0.01675419881939888, + 0.04262024536728859, + 0.029424550011754036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.05494493246078491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_tensor([5670, 5670])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_tensor([5670, 5670])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8e617c9b731bf76f084770c887dc474457c6c350 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_tensor([5670, 5670])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up shot from a slightly upward angle, the camera remains still, focused with photographic realism on a pair of calloused hands, their knuckles thick and stained with white marble dust. Steady and powerful, they grip a wooden-handled steel chisel, polished and gleaming with ageless use. Carefully, they aim its sharp tip at the face of an unfinished Carrara marble statue. Within an artistic studio, a strong, directional light streams in from the side, illuminating the work area and revealing countless tiny particles of dust floating in the air. As the craftsman's mallet strikes with rhythmic precision, each impact lifts a tiny fleck of fine white marble shards and dust from the smooth stone surface. These fragments shimmer briefly in the light before slowly drifting away. The light casts deep shadows on the newly carved eyelids and the curve of the nose, highlighting every subtle undulation and texture, making the cold stone seem to be breathed with life. The entire picture has a cold tone, full of a high sense of concentration and the quiet beauty of the creative process, perfectly capturing the texture and reality of this moment.", + "seed": 16870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_5670_seed_16870_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10190853457897901, + "vlm_logprob": -0.10190853457897901, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.10190853457897901, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028085164725780487 + ], + "opposite_logprob_tokens": [ + -3.653085231781006 + ], + "logprob_reward": -0.028085164725780487, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028085164725780487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5858397483825684, + 0.03482237085700035, + 0.04937971755862236, + 0.018949609249830246, + 0.035059209913015366, + 0.04546220228075981, + 0.02279110997915268, + 0.10050582885742188, + 0.0765627771615982, + 0.03062744066119194 + ], + "credit_map": [ + 0.025461876764893532, + 0.027331527322530746, + 0.11554915457963943, + 0.24694201350212097, + 0.011311412788927555, + 0.01917392946779728, + 0.0318945050239563, + 0.10817532986402512, + 0.004834705963730812, + 0.00331652513705194, + 0.004294554237276316, + 0.0054609403014183044, + 0.002254677703604102, + 0.004955002572387457, + 0.0038322526961565018, + 0.005873709451407194, + 0.0037259464152157307, + 0.004867666866630316, + 0.007618901319801807, + 0.007900819182395935, + 0.005007558967918158, + 0.008878381922841072, + 0.004315672442317009, + 0.007064769510179758, + 0.0014364778762683272, + 0.0024793827906250954, + 0.004365585744380951, + 0.0018605986842885613, + 0.002035359386354685, + 0.0023131780326366425, + 0.0018592489650472999, + 0.002599778352305293, + 0.004786234814673662, + 0.0032077652867883444, + 0.0039021463599056005, + 0.007014639675617218, + 0.002357762772589922, + 0.0046834261156618595, + 0.003048982471227646, + 0.006058251019567251, + 0.005766996182501316, + 0.004007090348750353, + 0.004323992878198624, + 0.004722896032035351, + 0.00545223755761981, + 0.007006341591477394, + 0.004520064685493708, + 0.009662582539021969, + 0.0025009505916386843, + 0.0019614927005022764, + 0.004427362233400345, + 0.0051488266326487064, + 0.0018581115873530507, + 0.002148569095879793, + 0.0018459452548995614, + 0.002899852115660906, + 0.005107003264129162, + 0.00473977392539382, + 0.008596304804086685, + 0.03853795677423477, + 0.0051097446121275425, + 0.009282943792641163, + 0.020649610087275505, + 0.008482499979436398, + 0.0067237382754683495, + 0.007619817741215229, + 0.012660977430641651, + 0.011241992004215717, + 0.0066116852685809135, + 0.0128725441172719, + 0.010904788039624691, + 0.00792723149061203, + 0.0035652737133204937, + 0.0029716596473008394, + 0.004164980724453926, + 0.002747707301750779, + 0.0029845326207578182, + 0.006466311402618885, + 0.0037473132833838463, + 0.003979662898927927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.028085164725780487 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206889823079109 + ], + "opposite_logprob_tokens": [ + -2.8120689392089844 + ], + "logprob_reward": -0.06206889823079109, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206889823079109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42907822132110596, + 0.04054407775402069, + 0.08065911382436752, + 0.13139964640140533, + 0.03213741257786751, + 0.054618995636701584, + 0.04116634652018547, + 0.1631692796945572, + 0.014442631974816322, + 0.012784313410520554 + ], + "credit_map": [ + 0.021088628098368645, + 0.01741308532655239, + 0.09163018316030502, + 0.15230168402194977, + 0.011969313956797123, + 0.021730201318860054, + 0.030635923147201538, + 0.08230917155742645, + 0.00604659179225564, + 0.0037231033202260733, + 0.005504608619958162, + 0.006559378933161497, + 0.002552468329668045, + 0.005614525638520718, + 0.005114543717354536, + 0.0054288520477712154, + 0.006162664387375116, + 0.008518664166331291, + 0.013064435683190823, + 0.014452705159783363, + 0.007753349840641022, + 0.012454715557396412, + 0.006575645413249731, + 0.011676936410367489, + 0.007966711185872555, + 0.006955741439014673, + 0.035472482442855835, + 0.016956057399511337, + 0.0057146367616951466, + 0.008428889326751232, + 0.028852984309196472, + 0.02105214074254036, + 0.004164539743214846, + 0.004159705713391304, + 0.005711446050554514, + 0.0041982391849160194, + 0.00239786715246737, + 0.003949353471398354, + 0.0029053473845124245, + 0.004650909453630447, + 0.006523084826767445, + 0.0056968750432133675, + 0.005206258036196232, + 0.00802055187523365, + 0.005872241221368313, + 0.007546372711658478, + 0.0043029156513512135, + 0.011450693011283875, + 0.0037980754859745502, + 0.003518591867759824, + 0.007554607465863228, + 0.011747992597520351, + 0.0033972952514886856, + 0.0033249047119170427, + 0.003691745223477483, + 0.004133131820708513, + 0.009117720648646355, + 0.00767305726185441, + 0.01447399239987135, + 0.05568200349807739, + 0.009715670719742775, + 0.02040671557188034, + 0.030006732791662216, + 0.016093382611870766, + 0.001549076521769166, + 0.001849773689173162, + 0.0026355553418397903, + 0.0020667146891355515, + 0.0013352315872907639, + 0.0016773419920355082, + 0.0016832640394568443, + 0.0016456728335469961, + 0.0012893492821604013, + 0.0018654644954949617, + 0.0017943407874554396, + 0.0021883186418563128, + 0.0017577933613210917, + 0.0013416411820799112, + 0.0011622578604146838, + 0.0013851469848304987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.06206889823079109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2783712148666382, + 0.0947393923997879, + 0.09072386473417282, + 0.13175247609615326, + 0.11075177043676376, + 0.09070082008838654, + 0.03713003173470497, + 0.042285285890102386, + 0.09460662305355072, + 0.028938572853803635 + ], + "credit_map": [ + 0.025739513337612152, + 0.014926676638424397, + 0.0350058414041996, + 0.10788699239492416, + 0.007941498421132565, + 0.015195408836007118, + 0.02043105848133564, + 0.05124422162771225, + 0.011778459884226322, + 0.008270258083939552, + 0.010932017117738724, + 0.019854076206684113, + 0.008625815622508526, + 0.010189628228545189, + 0.009653531946241856, + 0.015435608103871346, + 0.006992465350776911, + 0.009036545641720295, + 0.01590059883892536, + 0.01969827711582184, + 0.007864591665565968, + 0.011342420242726803, + 0.00923614390194416, + 0.010652812197804451, + 0.008195583708584309, + 0.014071501791477203, + 0.04797907546162605, + 0.021570907905697823, + 0.0054874783381819725, + 0.008701234124600887, + 0.007598182186484337, + 0.018148507922887802, + 0.014407016336917877, + 0.00796546507626772, + 0.012720105238258839, + 0.01930476725101471, + 0.008268109522759914, + 0.01770559698343277, + 0.010081090964376926, + 0.02029961347579956, + 0.008978007361292839, + 0.01239454559981823, + 0.01434473879635334, + 0.01144441869109869, + 0.010889925993978977, + 0.015147613361477852, + 0.00816479790955782, + 0.009336763992905617, + 0.0038223068695515394, + 0.004633395932614803, + 0.006527676712721586, + 0.007116320542991161, + 0.0029391616117209196, + 0.004738010466098785, + 0.002782238181680441, + 0.0045709190890192986, + 0.003222758648917079, + 0.0032015719916671515, + 0.005844449158757925, + 0.01302388310432434, + 0.0024837483651936054, + 0.003724830225110054, + 0.0073537263087928295, + 0.0034303180873394012, + 0.007389343809336424, + 0.007830001413822174, + 0.01655762642621994, + 0.015298301354050636, + 0.0074323005974292755, + 0.015217486768960953, + 0.013925980776548386, + 0.010955573990941048, + 0.0026378012262284756, + 0.004424792714416981, + 0.00470359344035387, + 0.006018312182277441, + 0.003216315060853958, + 0.0022377048153430223, + 0.003785692621022463, + 0.0019143602112308145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005949762649834156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38707825541496277 + ], + "opposite_logprob_tokens": [ + -1.1370782852172852 + ], + "logprob_reward": -0.38707825541496277, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38707825541496277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26262620091438293, + 0.04885457456111908, + 0.13463228940963745, + 0.1855606734752655, + 0.08180602639913559, + 0.057134728878736496, + 0.02820652164518833, + 0.11482787132263184, + 0.06898470222949982, + 0.017366398125886917 + ], + "credit_map": [ + 0.021830732002854347, + 0.01089349016547203, + 0.03409871086478233, + 0.0930243656039238, + 0.008400967344641685, + 0.013071888126432896, + 0.027926281094551086, + 0.05337977409362793, + 0.008568739518523216, + 0.005189584102481604, + 0.0073312618769705296, + 0.009668783284723759, + 0.002767541678622365, + 0.004300009924918413, + 0.00440648477524519, + 0.0066221654415130615, + 0.009574943222105503, + 0.014491989277303219, + 0.02242353931069374, + 0.027872219681739807, + 0.012546541169285774, + 0.020375346764922142, + 0.011771551333367825, + 0.015576153062283993, + 0.01130695454776287, + 0.01832093857228756, + 0.059739965945482254, + 0.030050449073314667, + 0.007191683631390333, + 0.016387462615966797, + 0.027083653956651688, + 0.01547956746071577, + 0.011308797635138035, + 0.007282048463821411, + 0.008703689090907574, + 0.018718140199780464, + 0.007317740935832262, + 0.010605287738144398, + 0.006522513460367918, + 0.011347812600433826, + 0.005215732380747795, + 0.006821339949965477, + 0.006723592057824135, + 0.01155218668282032, + 0.007150827441364527, + 0.008510954678058624, + 0.004250105936080217, + 0.0069099911488592625, + 0.0030500460416078568, + 0.002343878149986267, + 0.0050812154076993465, + 0.006449595093727112, + 0.002231860999017954, + 0.0033192855771631002, + 0.002313723787665367, + 0.003416916821151972, + 0.0077702696435153484, + 0.006868033669888973, + 0.013370666652917862, + 0.02910717763006687, + 0.007666465826332569, + 0.01928916573524475, + 0.016393577679991722, + 0.014362518675625324, + 0.006856908090412617, + 0.007134252227842808, + 0.01276029460132122, + 0.011613238602876663, + 0.005668737459927797, + 0.00967487134039402, + 0.009138497523963451, + 0.006137902848422527, + 0.0017117284005507827, + 0.0020140144042670727, + 0.0028939933981746435, + 0.0024366576690226793, + 0.0017818922642618418, + 0.00255099986679852, + 0.0026926931459456682, + 0.0012844182783737779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.38707825541496277 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360591873526573 + ], + "opposite_logprob_tokens": [ + -3.651360511779785 + ], + "logprob_reward": -0.026360591873526573, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360591873526573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.305990993976593, + 0.07296904921531677, + 0.056395139545202255, + 0.08242347836494446, + 0.043742306530475616, + 0.10031529515981674, + 0.09568584710359573, + 0.18945030868053436, + 0.03914159536361694, + 0.013885914348065853 + ], + "credit_map": [ + 0.018810918554663658, + 0.015506615862250328, + 0.02176765725016594, + 0.11051306128501892, + 0.010122302919626236, + 0.015963979065418243, + 0.021056929603219032, + 0.0922495499253273, + 0.011530979536473751, + 0.00688254926353693, + 0.008568620309233665, + 0.013227986171841621, + 0.00426768371835351, + 0.011201794259250164, + 0.0077204215340316296, + 0.009569020941853523, + 0.004386635031551123, + 0.005953922402113676, + 0.009411372244358063, + 0.012386132031679153, + 0.005065422970801592, + 0.006807084195315838, + 0.0054401527158916, + 0.006944421213120222, + 0.006247867830097675, + 0.007266979664564133, + 0.012427138164639473, + 0.00920278299599886, + 0.0040238699875772, + 0.00809535477310419, + 0.018209682777523994, + 0.01694980449974537, + 0.004884833004325628, + 0.004634429235011339, + 0.007535669486969709, + 0.00989596825093031, + 0.003193451324477792, + 0.005632713902741671, + 0.00304476753808558, + 0.004920471925288439, + 0.012303578667342663, + 0.005814737640321255, + 0.006721485406160355, + 0.01613539643585682, + 0.011542974039912224, + 0.018592894077301025, + 0.009554448537528515, + 0.01964978501200676, + 0.010375005193054676, + 0.014454432763159275, + 0.0243437010794878, + 0.01709049567580223, + 0.007570631802082062, + 0.007572802249342203, + 0.006722530350089073, + 0.007556252181529999, + 0.011156536638736725, + 0.010242177173495293, + 0.018144063651561737, + 0.06306910514831543, + 0.012409202754497528, + 0.021570323035120964, + 0.03377426415681839, + 0.019084637984633446, + 0.0032449993304908276, + 0.0034928887616842985, + 0.006250835955142975, + 0.006466267630457878, + 0.0034078527241945267, + 0.00683661550283432, + 0.004983673337846994, + 0.004458460956811905, + 0.0012960718013346195, + 0.0018449524650350213, + 0.0017684345366433263, + 0.002297961385920644, + 0.0015959731535986066, + 0.002008677925914526, + 0.0016303532756865025, + 0.0014434894546866417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.026360591873526573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot from a slightly upward angle, the camera remains still, focused with photographic realism on a pair of calloused hands, their knuckles thick and stained with white marble dust. Steady and powerful, they grip a wooden-handled steel chisel, polished and gleaming with ageless use. Carefully, they aim its sharp tip at the face of an unfinished Carrara marble statue. Within an artistic studio, a strong, directional light streams in from the side, illuminating the work area and revealing countless tiny particles of dust floating in the air. As the craftsman's mallet strikes with rhythmic precision, each impact lifts a tiny fleck of fine white marble shards and dust from the smooth stone surface. These fragments shimmer briefly in the light before slowly drifting away. The light casts deep shadows on the newly carved eyelids and the curve of the nose, highlighting every subtle undulation and texture, making the cold stone seem to be breathed with life. The entire picture has a cold tone, full of a high sense of concentration and the quiet beauty of the creative process, perfectly capturing the texture and reality of this moment.", + "seed": 16871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_5670_seed_16871_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11688407808542252, + "vlm_logprob": -0.11688407808542252, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11688407808542252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03903886675834656 + ], + "opposite_logprob_tokens": [ + -3.289038896560669 + ], + "logprob_reward": -0.03903886675834656, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03903886675834656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09690579771995544, + 0.08734376728534698, + 0.10117340832948685, + 0.17276710271835327, + 0.027224961668252945, + 0.03690292686223984, + 0.08695320039987564, + 0.04865891486406326, + 0.29332634806632996, + 0.048743583261966705 + ], + "credit_map": [ + 0.010295715183019638, + 0.010979161597788334, + 0.009339691139757633, + 0.012634806334972382, + 0.010709023103117943, + 0.02189049683511257, + 0.008290423080325127, + 0.01276647299528122, + 0.009206010960042477, + 0.008266308344900608, + 0.008636096492409706, + 0.014266219921410084, + 0.005960769020020962, + 0.014982216991484165, + 0.012770970351994038, + 0.013255174271762371, + 0.017445718869566917, + 0.005556778516620398, + 0.0056336671113967896, + 0.009270455688238144, + 0.007253967225551605, + 0.019180817529559135, + 0.00865184050053358, + 0.028180159628391266, + 0.03083064965903759, + 0.011159579269587994, + 0.006225664634257555, + 0.005787746515125036, + 0.007177943829447031, + 0.03365347534418106, + 0.01782924309372902, + 0.060102809220552444, + 0.003177602542564273, + 0.001594036933965981, + 0.001944165676832199, + 0.003622627118602395, + 0.0021516047418117523, + 0.00563050527125597, + 0.0033239189069718122, + 0.005780501291155815, + 0.003750458126887679, + 0.003116186009719968, + 0.005057820584625006, + 0.009496796876192093, + 0.003755947109311819, + 0.003937647212296724, + 0.0031001372262835503, + 0.004687932785600424, + 0.018674740567803383, + 0.007386059034615755, + 0.00957497674971819, + 0.005233650095760822, + 0.0063224430195987225, + 0.01327305007725954, + 0.005732634570449591, + 0.020755641162395477, + 0.00498475506901741, + 0.0034806858748197556, + 0.005490628536790609, + 0.007015714887529612, + 0.004137515556067228, + 0.006901988293975592, + 0.008026892319321632, + 0.008620732463896275, + 0.05714457854628563, + 0.012864447198808193, + 0.01367514580488205, + 0.015554234385490417, + 0.014756000600755215, + 0.04846584051847458, + 0.02977777272462845, + 0.10108833014965057, + 0.006013895850628614, + 0.0029673289973288774, + 0.004096216522157192, + 0.0029929738957434893, + 0.004130592104047537, + 0.007686650846153498, + 0.01020725816488266, + 0.010648667812347412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.03903886675834656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430985689163208 + ], + "opposite_logprob_tokens": [ + -3.1680984497070312 + ], + "logprob_reward": -0.0430985689163208, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430985689163208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12001979351043701, + 0.07796484231948853, + 0.20705872774124146, + 0.20148517191410065, + 0.054896555840969086, + 0.04280434921383858, + 0.0920632928609848, + 0.05644090473651886, + 0.03997300565242767, + 0.10729333758354187 + ], + "credit_map": [ + 0.010248911567032337, + 0.01145560760051012, + 0.01398745458573103, + 0.021750301122665405, + 0.012757189571857452, + 0.02126653678715229, + 0.01067415438592434, + 0.017879636958241463, + 0.009666035883128643, + 0.006479516159743071, + 0.007041028700768948, + 0.010253875516355038, + 0.007178273983299732, + 0.015690091997385025, + 0.011503283865749836, + 0.010152735747396946, + 0.04984714463353157, + 0.008588744327425957, + 0.0038011984433978796, + 0.0048014577478170395, + 0.005354105029255152, + 0.03931594640016556, + 0.005772063974291086, + 0.08957807719707489, + 0.03449048101902008, + 0.009501109831035137, + 0.007052321918308735, + 0.008471081033349037, + 0.008275792002677917, + 0.04306197538971901, + 0.020026208832859993, + 0.07060620933771133, + 0.006984819192439318, + 0.0035731797106564045, + 0.002715095179155469, + 0.006436750292778015, + 0.004159805364906788, + 0.013203178532421589, + 0.006036125123500824, + 0.01178760640323162, + 0.004509972874075174, + 0.0028195178601890802, + 0.0040276250801980495, + 0.006486539263278246, + 0.0029773630667477846, + 0.008382510393857956, + 0.007012063171714544, + 0.006588757503777742, + 0.012879638932645321, + 0.00922096986323595, + 0.010830502957105637, + 0.009673893451690674, + 0.006184274330735207, + 0.013581644743680954, + 0.014813782647252083, + 0.014878589659929276, + 0.008564256131649017, + 0.005071636289358139, + 0.00352403218857944, + 0.006477270275354385, + 0.0037261606194078922, + 0.009611534886062145, + 0.006984235718846321, + 0.012481783516705036, + 0.004323609173297882, + 0.0031975230667740107, + 0.0041960980743169785, + 0.003621733281761408, + 0.0029580029658973217, + 0.009271683171391487, + 0.007319203112274408, + 0.005085153039544821, + 0.01995052956044674, + 0.005234377458691597, + 0.008233188651502132, + 0.006141971331089735, + 0.005451112054288387, + 0.01893436908721924, + 0.008096951991319656, + 0.03525085002183914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0430985689163208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22352327406406403, + 0.05950731039047241, + 0.10756861418485641, + 0.03622335195541382, + 0.041591934859752655, + 0.03145331144332886, + 0.07832761108875275, + 0.05084722489118576, + 0.24375846982002258, + 0.1271989345550537 + ], + "credit_map": [ + 0.019856557250022888, + 0.024085262790322304, + 0.025058655068278313, + 0.04031285271048546, + 0.021686574444174767, + 0.035875383764505386, + 0.02080199122428894, + 0.03584596887230873, + 0.007794857490807772, + 0.003156780730932951, + 0.003453881246969104, + 0.006234615109860897, + 0.005518235266208649, + 0.012913767248392105, + 0.012583384290337563, + 0.007851785980165005, + 0.020218703895807266, + 0.004280045162886381, + 0.002873929450288415, + 0.004731610883027315, + 0.004236430395394564, + 0.023016341030597687, + 0.009005237370729446, + 0.03920630365610123, + 0.004275768529623747, + 0.005660302471369505, + 0.0023437063209712505, + 0.0029214995447546244, + 0.0031134902965277433, + 0.006424741353839636, + 0.005037042777985334, + 0.006446796003729105, + 0.005139871034771204, + 0.001763752312399447, + 0.0013319933786988258, + 0.0032140002585947514, + 0.004313385579735041, + 0.010911407880485058, + 0.008237817324697971, + 0.006679704412817955, + 0.005172878038138151, + 0.001859996817074716, + 0.0027885022573173046, + 0.002918841317296028, + 0.002171044936403632, + 0.005854730494320393, + 0.0029360309708863497, + 0.007751286495476961, + 0.011572549119591713, + 0.007139086257666349, + 0.008835073560476303, + 0.006900454871356487, + 0.004515225999057293, + 0.01191852055490017, + 0.012625576928257942, + 0.014821117743849754, + 0.004901117645204067, + 0.003075523069128394, + 0.0038353954441845417, + 0.007207509595900774, + 0.004055009223520756, + 0.00739314965903759, + 0.012760026380419731, + 0.007619485259056091, + 0.05024384334683418, + 0.00864621065557003, + 0.00929240696132183, + 0.010805843397974968, + 0.011065680533647537, + 0.04180384427309036, + 0.022051000967621803, + 0.08984960615634918, + 0.02257474884390831, + 0.005215012934058905, + 0.006937037687748671, + 0.009058820083737373, + 0.007071387022733688, + 0.02020970918238163, + 0.019823390990495682, + 0.03630882129073143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.009783769026398659 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47429853677749634 + ], + "opposite_logprob_tokens": [ + -0.9742985367774963 + ], + "logprob_reward": -0.47429853677749634, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47429853677749634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052152425050735474, + 0.06769726425409317, + 0.2769443988800049, + 0.2739301919937134, + 0.04289918765425682, + 0.020887693390250206, + 0.09928630292415619, + 0.024270106106996536, + 0.11404910683631897, + 0.027883322909474373 + ], + "credit_map": [ + 0.004718045704066753, + 0.004220569506287575, + 0.004907920025289059, + 0.006126998923718929, + 0.0050536151975393295, + 0.010688015259802341, + 0.005887729115784168, + 0.010549529455602169, + 0.008081821724772453, + 0.003543446073308587, + 0.0041242241859436035, + 0.006609948351979256, + 0.005300621967762709, + 0.01379972230643034, + 0.017030399292707443, + 0.009207077324390411, + 0.06460320949554443, + 0.010514774359762669, + 0.005761817563325167, + 0.006996049080044031, + 0.008070995099842548, + 0.05231484770774841, + 0.010144156403839588, + 0.11853853613138199, + 0.05417206883430481, + 0.010842825286090374, + 0.005439314991235733, + 0.007910030893981457, + 0.008707345463335514, + 0.05648665130138397, + 0.020675502717494965, + 0.10969644039869308, + 0.004281283356249332, + 0.002961420686915517, + 0.003969090990722179, + 0.007241961546242237, + 0.003531987778842449, + 0.007667787838727236, + 0.007074661552906036, + 0.006170993205159903, + 0.0028467851225286722, + 0.0012997877784073353, + 0.0018057756824418902, + 0.0036823104601353407, + 0.0016175853088498116, + 0.003415935207158327, + 0.0020655663684010506, + 0.004153948277235031, + 0.016699861735105515, + 0.008177468553185463, + 0.008703264407813549, + 0.00796804204583168, + 0.006026669405400753, + 0.013219601474702358, + 0.01596272923052311, + 0.022528672590851784, + 0.0030725623946636915, + 0.0026701067108660936, + 0.0028648534789681435, + 0.004078431520611048, + 0.002141456352546811, + 0.004144477657973766, + 0.002374277450144291, + 0.002923938911408186, + 0.02268531732261181, + 0.0056142727844417095, + 0.004656651057302952, + 0.006494478322565556, + 0.004358440637588501, + 0.020003998652100563, + 0.008096047677099705, + 0.042139895260334015, + 0.0032181679271161556, + 0.002655073069036007, + 0.003946109674870968, + 0.004272849764674902, + 0.0023744769860059023, + 0.003884910373017192, + 0.0037003965117037296, + 0.0038313374388962984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.47429853677749634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200648948550224 + ], + "opposite_logprob_tokens": [ + -4.018200874328613 + ], + "logprob_reward": -0.018200648948550224, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200648948550224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13388583064079285, + 0.08741433918476105, + 0.08695312589406967, + 0.1137453243136406, + 0.050149232149124146, + 0.0626978725194931, + 0.048079751431941986, + 0.09937955439090729, + 0.14124809205532074, + 0.1764468550682068 + ], + "credit_map": [ + 0.013390351086854935, + 0.01289539597928524, + 0.015428940765559673, + 0.019566399976611137, + 0.014427212998270988, + 0.0248830895870924, + 0.01175112184137106, + 0.02154332958161831, + 0.012439875863492489, + 0.0066991811618208885, + 0.005690513178706169, + 0.009039290249347687, + 0.006301886402070522, + 0.015906348824501038, + 0.01572909764945507, + 0.015608156099915504, + 0.013837003149092197, + 0.007424559909850359, + 0.003036330919712782, + 0.00665950495749712, + 0.005987255368381739, + 0.02215762808918953, + 0.008286471478641033, + 0.019564377143979073, + 0.013409309089183807, + 0.006996688898652792, + 0.011034974828362465, + 0.010994503274559975, + 0.006431235931813717, + 0.026184115558862686, + 0.01756976917386055, + 0.021124735474586487, + 0.0063732764683663845, + 0.003953579347580671, + 0.0030424590222537518, + 0.004437226336449385, + 0.0045916736125946045, + 0.008942849934101105, + 0.010533851571381092, + 0.00827431958168745, + 0.009306256659328938, + 0.004534049890935421, + 0.007953048683702946, + 0.008420407772064209, + 0.0050088344141840935, + 0.008721002377569675, + 0.0055236793123185635, + 0.013230600394308567, + 0.007665213197469711, + 0.004027525428682566, + 0.004411135800182819, + 0.0030694894958287477, + 0.0026944405399262905, + 0.006943153217434883, + 0.008414504118263721, + 0.010854294523596764, + 0.012088587507605553, + 0.008629752323031425, + 0.01279743667691946, + 0.013797041960060596, + 0.005943526513874531, + 0.014280619099736214, + 0.013080526143312454, + 0.018762074410915375, + 0.020968329161405563, + 0.010811083018779755, + 0.013432461768388748, + 0.013523733243346214, + 0.011766644194722176, + 0.020739596337080002, + 0.020048972219228745, + 0.029957275837659836, + 0.03636704012751579, + 0.0074536618776619434, + 0.005185320973396301, + 0.010083423927426338, + 0.008841253817081451, + 0.029547270387411118, + 0.018603049218654633, + 0.06036585941910744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018200648948550224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot from a slightly upward angle, the camera remains still, focused with photographic realism on a pair of calloused hands, their knuckles thick and stained with white marble dust. Steady and powerful, they grip a wooden-handled steel chisel, polished and gleaming with ageless use. Carefully, they aim its sharp tip at the face of an unfinished Carrara marble statue. Within an artistic studio, a strong, directional light streams in from the side, illuminating the work area and revealing countless tiny particles of dust floating in the air. As the craftsman's mallet strikes with rhythmic precision, each impact lifts a tiny fleck of fine white marble shards and dust from the smooth stone surface. These fragments shimmer briefly in the light before slowly drifting away. The light casts deep shadows on the newly carved eyelids and the curve of the nose, highlighting every subtle undulation and texture, making the cold stone seem to be breathed with life. The entire picture has a cold tone, full of a high sense of concentration and the quiet beauty of the creative process, perfectly capturing the texture and reality of this moment.", + "seed": 16872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_5670_seed_16872_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03823702298104763, + "vlm_logprob": -0.03823702298104763, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03823702298104763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034523360431194305 + ], + "opposite_logprob_tokens": [ + -3.4095232486724854 + ], + "logprob_reward": -0.034523360431194305, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.034523360431194305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07421407848596573, + 0.3630746901035309, + 0.0843520313501358, + 0.028181057423353195, + 0.17647725343704224, + 0.05062124505639076, + 0.03543420881032944, + 0.0402567982673645, + 0.031950920820236206, + 0.11543767899274826 + ], + "credit_map": [ + 0.006178615149110556, + 0.008564907126128674, + 0.010700205340981483, + 0.013316500931978226, + 0.011789832264184952, + 0.006098790559917688, + 0.010733331553637981, + 0.00683189881965518, + 0.024850057438015938, + 0.037016186863183975, + 0.041312456130981445, + 0.043363891541957855, + 0.049172382801771164, + 0.07305493205785751, + 0.05160294100642204, + 0.04270186647772789, + 0.006705715786665678, + 0.011521661654114723, + 0.016316091641783714, + 0.013746627606451511, + 0.010574281215667725, + 0.009129020385444164, + 0.010122991167008877, + 0.006235649809241295, + 0.0023657383862882853, + 0.002933545969426632, + 0.005645001772791147, + 0.0017324896762147546, + 0.003950957208871841, + 0.0034585657995194197, + 0.00484486622735858, + 0.0032498943619430065, + 0.014129500836133957, + 0.020119650289416313, + 0.03813597932457924, + 0.026275886222720146, + 0.018599215894937515, + 0.018643824383616447, + 0.021484188735485077, + 0.019089026376605034, + 0.0027852063067257404, + 0.004883190616965294, + 0.011119232513010502, + 0.009012876078486443, + 0.003126387484371662, + 0.005890083499252796, + 0.007076634559780359, + 0.0067276377230882645, + 0.0027403247077018023, + 0.004964733496308327, + 0.003997999243438244, + 0.006401195656508207, + 0.0027792106848210096, + 0.005163159687072039, + 0.0050591290928423405, + 0.004328458569943905, + 0.0020158193074166775, + 0.004037928301841021, + 0.008287048898637295, + 0.0075096976943314075, + 0.003714586142450571, + 0.0024315824266523123, + 0.00567357474938035, + 0.006586562842130661, + 0.0026140855625271797, + 0.0035435601603239775, + 0.0023903290275484324, + 0.00584031455218792, + 0.004167641047388315, + 0.004770715720951557, + 0.0043172589503228664, + 0.00430701719596982, + 0.008833496831357479, + 0.014913724735379219, + 0.010468208231031895, + 0.01670236513018608, + 0.015584939159452915, + 0.01811154931783676, + 0.015056179836392403, + 0.01576722413301468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.034523360431194305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003960013389587 + ], + "opposite_logprob_tokens": [ + -2.6950395107269287 + ], + "logprob_reward": -0.07003960013389587, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003960013389587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1194123923778534, + 0.2238435000181198, + 0.13796289265155792, + 0.09459208697080612, + 0.09634824097156525, + 0.03321877121925354, + 0.061496589332818985, + 0.13222606480121613, + 0.02732895500957966, + 0.07357057183980942 + ], + "credit_map": [ + 0.00886745098978281, + 0.013033865951001644, + 0.016432445496320724, + 0.018101980909705162, + 0.019701363518834114, + 0.010887155309319496, + 0.019769536331295967, + 0.012618577107787132, + 0.01399306207895279, + 0.02116210013628006, + 0.02330928109586239, + 0.0253370963037014, + 0.029356615617871284, + 0.048162732273340225, + 0.03382406383752823, + 0.028698524460196495, + 0.006193746346980333, + 0.017400994896888733, + 0.010918484069406986, + 0.018496252596378326, + 0.02970403991639614, + 0.016088368371129036, + 0.023941943421959877, + 0.015219049528241158, + 0.0057843467220664024, + 0.007768458221107721, + 0.021777398884296417, + 0.004754409659653902, + 0.014160025864839554, + 0.013592411763966084, + 0.01534784771502018, + 0.011407176032662392, + 0.0063918777741491795, + 0.011518280021846294, + 0.01873520202934742, + 0.015019034966826439, + 0.00801398977637291, + 0.011010498739778996, + 0.015091921202838421, + 0.01056742575019598, + 0.0023714497219771147, + 0.004295083694159985, + 0.0064193932339549065, + 0.004982746206223965, + 0.004025376867502928, + 0.003309886669740081, + 0.004416708368808031, + 0.0033981234300881624, + 0.0045136925764381886, + 0.007581373676657677, + 0.00975447241216898, + 0.008221862837672234, + 0.005367509555071592, + 0.008140458725392818, + 0.007921732030808926, + 0.009995478205382824, + 0.008493229746818542, + 0.0094713494181633, + 0.043062955141067505, + 0.01998542621731758, + 0.008231216110289097, + 0.007347275968641043, + 0.01645096018910408, + 0.019183626398444176, + 0.002350075403228402, + 0.003735578153282404, + 0.002495699329301715, + 0.00498585170134902, + 0.0034867573995143175, + 0.002838175743818283, + 0.0038124159909784794, + 0.00362439826130867, + 0.004664028994739056, + 0.010136713273823261, + 0.00620768778026104, + 0.00995171070098877, + 0.012997780926525593, + 0.011183417402207851, + 0.008891758508980274, + 0.00953747145831585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.07003960013389587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078985407948494 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078985407948494, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078985407948494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0762544721364975, + 0.41513487696647644, + 0.04822646453976631, + 0.14826691150665283, + 0.12142743170261383, + 0.028688739985227585, + 0.05994803458452225, + 0.02985917404294014, + 0.02054496668279171, + 0.05164884403347969 + ], + "credit_map": [ + 0.005643222946673632, + 0.009326930157840252, + 0.010342596098780632, + 0.011058786883950233, + 0.011322716251015663, + 0.010514822788536549, + 0.010626097209751606, + 0.007419304456561804, + 0.025786345824599266, + 0.04233403876423836, + 0.04460421949625015, + 0.04347962141036987, + 0.05628403276205063, + 0.08501983433961868, + 0.06424177438020706, + 0.053385041654109955, + 0.0029402917716652155, + 0.007254748605191708, + 0.005758417770266533, + 0.0056077539920806885, + 0.010254289023578167, + 0.004764121491461992, + 0.007141270209103823, + 0.004505574237555265, + 0.008985218591988087, + 0.010653029195964336, + 0.030413001775741577, + 0.010085799731314182, + 0.022726276889443398, + 0.021680928766727448, + 0.025498514994978905, + 0.018224159255623817, + 0.006766644772142172, + 0.016916166990995407, + 0.01651659980416298, + 0.018073853105306625, + 0.014101595617830753, + 0.01405141782015562, + 0.022947043180465698, + 0.012054120190441608, + 0.001686382805928588, + 0.0036948241759091616, + 0.005792253650724888, + 0.005012746900320053, + 0.0017276933649554849, + 0.00296558509580791, + 0.003954485524445772, + 0.00385476928204298, + 0.004226525779813528, + 0.007413154002279043, + 0.012773902155458927, + 0.007834271527826786, + 0.005060885101556778, + 0.008392740041017532, + 0.0065476675517857075, + 0.007698891218751669, + 0.0014582125004380941, + 0.003217932768166065, + 0.005262704566121101, + 0.006189639680087566, + 0.0024040834978222847, + 0.0018845597514882684, + 0.00429515540599823, + 0.005146891809999943, + 0.001601635362021625, + 0.0022329487837851048, + 0.0018617131281644106, + 0.0029625827446579933, + 0.0023259336594492197, + 0.002921173581853509, + 0.0027902587316930294, + 0.003848724067211151, + 0.0040035732090473175, + 0.004487332887947559, + 0.004220929928123951, + 0.006758322473615408, + 0.009641882963478565, + 0.00958090741187334, + 0.0064791180193424225, + 0.006476786453276873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.011078985407948494 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054956890642642975 + ], + "opposite_logprob_tokens": [ + -2.9299569129943848 + ], + "logprob_reward": -0.054956890642642975, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054956890642642975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1169387474656105, + 0.4176689088344574, + 0.07478111237287521, + 0.07744477689266205, + 0.04745004326105118, + 0.05219384655356407, + 0.053752049803733826, + 0.04870446026325226, + 0.022523464635014534, + 0.08854252845048904 + ], + "credit_map": [ + 0.0088560301810503, + 0.014816312119364738, + 0.018568435683846474, + 0.017696740105748177, + 0.016935566440224648, + 0.014632993377745152, + 0.014973861165344715, + 0.010458812117576599, + 0.02647305093705654, + 0.03988637775182724, + 0.04606107249855995, + 0.05066075921058655, + 0.05378087982535362, + 0.08187483996152878, + 0.06230108439922333, + 0.05663083866238594, + 0.005470252595841885, + 0.008861555717885494, + 0.009419601410627365, + 0.008533800020813942, + 0.016043897718191147, + 0.007303708698600531, + 0.012370775453746319, + 0.006777525879442692, + 0.005413472652435303, + 0.008030789904296398, + 0.013208172284066677, + 0.006954600568860769, + 0.014339901506900787, + 0.009146366268396378, + 0.012608058750629425, + 0.007743413560092449, + 0.0043362947180867195, + 0.005526737309992313, + 0.009420296177268028, + 0.0075144837610423565, + 0.004780882969498634, + 0.005314547568559647, + 0.005977477878332138, + 0.0045793200843036175, + 0.0027625723741948605, + 0.006870117969810963, + 0.008709964342415333, + 0.008986824192106724, + 0.005083298776298761, + 0.004510336555540562, + 0.006843640934675932, + 0.00842709094285965, + 0.0032328341621905565, + 0.007322225254029036, + 0.00605744170024991, + 0.011655351147055626, + 0.0038748844526708126, + 0.008001330308616161, + 0.00791641790419817, + 0.005691565573215485, + 0.003437012666836381, + 0.003491382347419858, + 0.008290847763419151, + 0.008469062857329845, + 0.004299814347177744, + 0.004245582036674023, + 0.007397127337753773, + 0.009073629043996334, + 0.001339800306595862, + 0.002192904707044363, + 0.002068049507215619, + 0.004704065155237913, + 0.002063014544546604, + 0.0037406282499432564, + 0.003409996861591935, + 0.0030050065834075212, + 0.007506380323320627, + 0.010329262353479862, + 0.010093296878039837, + 0.01433470007032156, + 0.009850378148257732, + 0.010942986235022545, + 0.011744596064090729, + 0.013740934431552887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.054956890642642975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586278289556503 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586278289556503, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586278289556503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056083012372255325, + 0.6534925699234009, + 0.031758759170770645, + 0.022370874881744385, + 0.045454591512680054, + 0.028065495193004608, + 0.047043778002262115, + 0.038038309663534164, + 0.052023522555828094, + 0.025669043883681297 + ], + "credit_map": [ + 0.004526182077825069, + 0.005572199355810881, + 0.009121989831328392, + 0.01046750508248806, + 0.0066335806623101234, + 0.005722550209611654, + 0.008343277499079704, + 0.005695726256817579, + 0.04277613013982773, + 0.06542447954416275, + 0.07354969531297684, + 0.07365065068006516, + 0.08755888789892197, + 0.13197018206119537, + 0.09651339054107666, + 0.0820491835474968, + 0.0017270524986088276, + 0.003180488245561719, + 0.005115044768899679, + 0.005856816191226244, + 0.0045779915526509285, + 0.002843711990863085, + 0.0051744235679507256, + 0.0032832359429448843, + 0.0015397367533296347, + 0.002009882126003504, + 0.003450007876381278, + 0.002150057116523385, + 0.004549331497400999, + 0.0024508393835276365, + 0.004025982227176428, + 0.0021950404625386, + 0.0022239205427467823, + 0.005205369088798761, + 0.01030950527638197, + 0.007051687687635422, + 0.0036249086260795593, + 0.005801324266940355, + 0.0074363937601447105, + 0.0038014831952750683, + 0.0018604492070153356, + 0.0029221107251942158, + 0.006864879745990038, + 0.004128646105527878, + 0.0034132644068449736, + 0.0029322407208383083, + 0.0033178203739225864, + 0.0026260856539011, + 0.0039663733914494514, + 0.004113431554287672, + 0.008079583756625652, + 0.0049487510696053505, + 0.003794383956119418, + 0.010998492129147053, + 0.005795127246528864, + 0.005347634665668011, + 0.0018476078985258937, + 0.0032721359748393297, + 0.006030753720551729, + 0.004743479657918215, + 0.004032629542052746, + 0.003266077721491456, + 0.006958319805562496, + 0.007887309417128563, + 0.003434140468016267, + 0.004785810597240925, + 0.007113107014447451, + 0.005683020688593388, + 0.005286604631692171, + 0.008782798424363136, + 0.006334364879876375, + 0.010603678412735462, + 0.001797475852072239, + 0.004018793348222971, + 0.002818961162120104, + 0.004100310616195202, + 0.0030825273133814335, + 0.003249057801440358, + 0.0031542808283120394, + 0.003447639988735318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.020586278289556503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot from a slightly upward angle, the camera remains still, focused with photographic realism on a pair of calloused hands, their knuckles thick and stained with white marble dust. Steady and powerful, they grip a wooden-handled steel chisel, polished and gleaming with ageless use. Carefully, they aim its sharp tip at the face of an unfinished Carrara marble statue. Within an artistic studio, a strong, directional light streams in from the side, illuminating the work area and revealing countless tiny particles of dust floating in the air. As the craftsman's mallet strikes with rhythmic precision, each impact lifts a tiny fleck of fine white marble shards and dust from the smooth stone surface. These fragments shimmer briefly in the light before slowly drifting away. The light casts deep shadows on the newly carved eyelids and the curve of the nose, highlighting every subtle undulation and texture, making the cold stone seem to be breathed with life. The entire picture has a cold tone, full of a high sense of concentration and the quiet beauty of the creative process, perfectly capturing the texture and reality of this moment.", + "seed": 16873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_0/index_5670_seed_16873_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08854927895590663, + "vlm_logprob": -0.08854927895590663, + "vlm_yes_no_margin": 3.374999988079071, + "vlm_token_credit": -0.08854927895590663, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03075258433818817 + ], + "opposite_logprob_tokens": [ + -3.530752658843994 + ], + "logprob_reward": -0.03075258433818817, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03075258433818817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.642608106136322, + 0.14078494906425476, + 0.02137693576514721, + 0.02867322601377964, + 0.023801937699317932, + 0.022546064108610153, + 0.03169720619916916, + 0.010970803909003735, + 0.03457468003034592, + 0.04296606406569481 + ], + "credit_map": [ + 0.039718419313430786, + 0.015070304274559021, + 0.019932707771658897, + 0.04694059118628502, + 0.059046778827905655, + 0.2751915752887726, + 0.030633267015218735, + 0.15607455372810364, + 0.006790273357182741, + 0.00551643455401063, + 0.029540348798036575, + 0.05789821222424507, + 0.005426385439932346, + 0.008344165980815887, + 0.008489909581840038, + 0.018779229372739792, + 0.0016958910273388028, + 0.0014659659937024117, + 0.0023680313024669886, + 0.0029453563038259745, + 0.0026462215464562178, + 0.0048649851232767105, + 0.0030999493319541216, + 0.002290537115186453, + 0.003242525504902005, + 0.0020052390173077583, + 0.003228710498660803, + 0.003398190950974822, + 0.0027696345932781696, + 0.004000285640358925, + 0.0038557234220206738, + 0.006172916851937771, + 0.0016595068154856563, + 0.001980698900297284, + 0.001641003997065127, + 0.003458926221355796, + 0.002987787127494812, + 0.0030632440466433764, + 0.005863219033926725, + 0.003147554351016879, + 0.0018997339066118002, + 0.0016094779130071402, + 0.002768856706097722, + 0.0028736870735883713, + 0.0024405128788203, + 0.004070810507982969, + 0.003957800567150116, + 0.002925184555351734, + 0.002023120643571019, + 0.0020759054459631443, + 0.001614209613762796, + 0.0021829940378665924, + 0.009200268425047398, + 0.005874984432011843, + 0.004777353722602129, + 0.003948370926082134, + 0.0011196460109204054, + 0.0011196606792509556, + 0.0014272785047069192, + 0.0013257665559649467, + 0.0012953768018633127, + 0.0017044964479282498, + 0.0013637866359204054, + 0.001614793436601758, + 0.002832012018188834, + 0.0028970043640583754, + 0.0041422126814723015, + 0.004142689984291792, + 0.005297431722283363, + 0.003174504032358527, + 0.007422589231282473, + 0.0046662394888699055, + 0.004637465812265873, + 0.0026939688250422478, + 0.003890948137268424, + 0.006226120516657829, + 0.00451803021132946, + 0.004534436855465174, + 0.011397658847272396, + 0.005067433230578899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03075258433818817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811795637011528 + ], + "opposite_logprob_tokens": [ + -3.2881178855895996 + ], + "logprob_reward": -0.03811795637011528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811795637011528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6170143485069275, + 0.1448533833026886, + 0.07567846029996872, + 0.03677191957831383, + 0.02964344248175621, + 0.016753699630498886, + 0.019782889634370804, + 0.01198457833379507, + 0.020597156137228012, + 0.026920070871710777 + ], + "credit_map": [ + 0.036970000714063644, + 0.01569087989628315, + 0.020318808034062386, + 0.04605672135949135, + 0.0558454729616642, + 0.2617138624191284, + 0.030198074877262115, + 0.1502205729484558, + 0.005939687602221966, + 0.006900802720338106, + 0.023358605802059174, + 0.06079316884279251, + 0.005490550305694342, + 0.008864868432283401, + 0.009314065799117088, + 0.024191638454794884, + 0.006467954255640507, + 0.003997304476797581, + 0.004488557111471891, + 0.006666518747806549, + 0.006773562170565128, + 0.02304244227707386, + 0.009294485673308372, + 0.014947639778256416, + 0.0045373449102044106, + 0.0034895569551736116, + 0.005635768640786409, + 0.005856332369148731, + 0.0031881718896329403, + 0.006603818386793137, + 0.003663742681965232, + 0.003797185141593218, + 0.0020631104707717896, + 0.001937651657499373, + 0.002833784092217684, + 0.003939719405025244, + 0.0036602215841412544, + 0.004630308132618666, + 0.006676320917904377, + 0.0039023286662995815, + 0.0013663249555975199, + 0.001556035946123302, + 0.00160906370729208, + 0.002588175004348159, + 0.0015311649767681956, + 0.0028855300042778254, + 0.0030680783092975616, + 0.0021493288222700357, + 0.0015413963701575994, + 0.001070495811291039, + 0.0010676842648535967, + 0.0011123678414151073, + 0.006695302203297615, + 0.004129886161535978, + 0.0021284043323248625, + 0.0020373540464788675, + 0.0010244430741295218, + 0.0011715407017618418, + 0.0015014687087386847, + 0.0019125938415527344, + 0.001469156821258366, + 0.0022698843386024237, + 0.0014401827938854694, + 0.0011953088687732816, + 0.0018628736725077033, + 0.0019331821240484715, + 0.0025672907941043377, + 0.0028662453405559063, + 0.002801415743306279, + 0.0023988522589206696, + 0.003404798684641719, + 0.0027624983340501785, + 0.0029744512867182493, + 0.001926262746565044, + 0.002729599131271243, + 0.00501174945384264, + 0.0027292603626847267, + 0.0028180917724967003, + 0.004651975817978382, + 0.004078681115061045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03811795637011528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642422776669264 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642422776669264, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642422776669264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6030351519584656, + 0.058447424322366714, + 0.06185375154018402, + 0.04636331647634506, + 0.050320591777563095, + 0.020724669098854065, + 0.017321765422821045, + 0.07208473980426788, + 0.04728652164340019, + 0.022562021389603615 + ], + "credit_map": [ + 0.03922462835907936, + 0.017217520624399185, + 0.02242118865251541, + 0.048253290355205536, + 0.05578048154711723, + 0.25511300563812256, + 0.029216673225164413, + 0.13580840826034546, + 0.0030455016531050205, + 0.00329015520401299, + 0.005474227946251631, + 0.023158974945545197, + 0.002767080906778574, + 0.003742585889995098, + 0.0044996896758675575, + 0.012469210661947727, + 0.006969590671360493, + 0.00481685996055603, + 0.007947963662445545, + 0.014132268726825714, + 0.005125338677316904, + 0.01050200592726469, + 0.005977584980428219, + 0.00638214498758316, + 0.00563395768404007, + 0.004158128518611193, + 0.006745180115103722, + 0.008309359662234783, + 0.004272435326129198, + 0.007173618767410517, + 0.005518651567399502, + 0.004551985301077366, + 0.004061524290591478, + 0.0035693934187293053, + 0.004433886613696814, + 0.006516238208860159, + 0.006644947454333305, + 0.0073142945766448975, + 0.011414574459195137, + 0.006365736946463585, + 0.002309226430952549, + 0.0014718082966282964, + 0.002067884895950556, + 0.0035375840961933136, + 0.0020903595723211765, + 0.0038404150400310755, + 0.0025522157084196806, + 0.0028551758732646704, + 0.0015103333862498403, + 0.0012844266602769494, + 0.0013957396149635315, + 0.00179462437517941, + 0.0022797966375947, + 0.0039706300012767315, + 0.002762486692517996, + 0.0023237315472215414, + 0.00512735778465867, + 0.0031636864878237247, + 0.002330115530639887, + 0.002054035197943449, + 0.027712881565093994, + 0.01046422217041254, + 0.003987638745456934, + 0.017244815826416016, + 0.006578175351023674, + 0.004716257099062204, + 0.005083233118057251, + 0.006009784527122974, + 0.005355494562536478, + 0.006140433251857758, + 0.006745257414877415, + 0.006657888181507587, + 0.0027589863166213036, + 0.0017682084580883384, + 0.0022429998498409986, + 0.0021168242674320936, + 0.002625192515552044, + 0.0029204203747212887, + 0.004951378796249628, + 0.0031780111603438854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.004642422776669264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34864598512649536 + ], + "opposite_logprob_tokens": [ + -1.2236459255218506 + ], + "logprob_reward": -0.34864598512649536, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34864598512649536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.700717031955719, + 0.0325135812163353, + 0.057489074766635895, + 0.05511142686009407, + 0.011954235844314098, + 0.025608502328395844, + 0.015768656507134438, + 0.04609907045960426, + 0.026682116091251373, + 0.028056275099515915 + ], + "credit_map": [ + 0.042284995317459106, + 0.01492836233228445, + 0.01975385472178459, + 0.04664255306124687, + 0.06297691911458969, + 0.29916292428970337, + 0.03231021761894226, + 0.18265721201896667, + 0.0036807262804359198, + 0.0019401194294914603, + 0.005656650755554438, + 0.007567196153104305, + 0.003204549429938197, + 0.003333640517666936, + 0.0025011212565004826, + 0.004629576578736305, + 0.006870281882584095, + 0.005754244513809681, + 0.004978439304977655, + 0.008683538064360619, + 0.007850277237594128, + 0.007943255826830864, + 0.00962929055094719, + 0.005779746454209089, + 0.005317778326570988, + 0.00443671178072691, + 0.005300982855260372, + 0.009987971745431423, + 0.004872380290180445, + 0.0069845556281507015, + 0.008474497124552727, + 0.009736552834510803, + 0.0011425005504861474, + 0.0009945309720933437, + 0.0012364359572529793, + 0.0016030077822506428, + 0.002312305150553584, + 0.0015559622552245855, + 0.001632125349715352, + 0.0014773677103221416, + 0.0033477414399385452, + 0.0018695686012506485, + 0.0026147523894906044, + 0.003030359512194991, + 0.0028268173336982727, + 0.003931496292352676, + 0.00475198682397604, + 0.003235779469832778, + 0.0013970636064186692, + 0.0008767092949710786, + 0.0010720014106482267, + 0.001220806734636426, + 0.0021886234171688557, + 0.0035255346447229385, + 0.0031816933769732714, + 0.0023062238469719887, + 0.003302659373730421, + 0.0029607194010168314, + 0.002535604638978839, + 0.0031373328529298306, + 0.015991196036338806, + 0.009261311031877995, + 0.002346523804590106, + 0.0065637268126010895, + 0.0034734930377453566, + 0.0030560821760445833, + 0.004484171513468027, + 0.0036238848697394133, + 0.0025884381029754877, + 0.0031296161469072104, + 0.0031828628852963448, + 0.0031435678247362375, + 0.003126061288639903, + 0.0020865180995315313, + 0.00252578966319561, + 0.0034258426167070866, + 0.0030977572314441204, + 0.0034927683882415295, + 0.006989344023168087, + 0.0033121949527412653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.34864598512649536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of calloused hands with thick knuckles stained with white marble dust visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hands gripping a wooden-handled steel chisel that appears polished and gleaming?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chisel aimed at the face of an unfinished Carrara marble statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a mallet striking the statue with rhythmic precision, lifting tiny flecks of white marble shards and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an artistic studio with strong directional side lighting and visible floating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058744616806507 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058744616806507, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058744616806507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7148926258087158, + 0.05129111558198929, + 0.017170356586575508, + 0.02304266393184662, + 0.03069440834224224, + 0.03264934942126274, + 0.011699256487190723, + 0.05410895496606827, + 0.028126049786806107, + 0.03632528707385063 + ], + "credit_map": [ + 0.04427299275994301, + 0.016551444306969643, + 0.02247781865298748, + 0.0488521046936512, + 0.05943290516734123, + 0.29481977224349976, + 0.03494991362094879, + 0.19353561103343964, + 0.0030214323196560144, + 0.002567728515714407, + 0.011710935272276402, + 0.02084466442465782, + 0.0023670915979892015, + 0.00283029000274837, + 0.0030925979372113943, + 0.004856366664171219, + 0.0014525370206683874, + 0.0011086371960118413, + 0.0013491151621565223, + 0.002493705367669463, + 0.0015501833986490965, + 0.004521532915532589, + 0.0018873487133532763, + 0.002807294251397252, + 0.0023788001853972673, + 0.0019297348335385323, + 0.00240657408721745, + 0.005710231605917215, + 0.0018869920168071985, + 0.003967834170907736, + 0.0021357035730034113, + 0.002626791363582015, + 0.0022625974379479885, + 0.0021316870115697384, + 0.0019698352552950382, + 0.0033715281169861555, + 0.003690961515530944, + 0.004379497841000557, + 0.008472070097923279, + 0.004416226409375668, + 0.003318900940939784, + 0.001975306076928973, + 0.002653356408700347, + 0.004421450663357973, + 0.0032436514738947153, + 0.0059438953176140785, + 0.006038948893547058, + 0.005053835455328226, + 0.0013497686013579369, + 0.0009653680608607829, + 0.001091267797164619, + 0.0012791112530976534, + 0.0015459107235074043, + 0.002112844493240118, + 0.0017465229611843824, + 0.0016084611415863037, + 0.003510825801640749, + 0.0020708274096250534, + 0.001932007959112525, + 0.0019744408782571554, + 0.020906707271933556, + 0.010548626072704792, + 0.0028688148595392704, + 0.010296697728335857, + 0.0036995478440076113, + 0.002920715603977442, + 0.003321694443002343, + 0.003588947234675288, + 0.0030720483046025038, + 0.003513336181640625, + 0.004283688962459564, + 0.003726068651303649, + 0.002877461491152644, + 0.002829574281349778, + 0.0040254145860672, + 0.008052722550928593, + 0.003643142059445381, + 0.0036253822036087513, + 0.006541315466165543, + 0.004730265121906996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02058744616806507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_tensor([4202, 4202])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_tensor([4202, 4202])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..96cac9fb7f16a8eda88de7698008f86f2a6678da --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_tensor([4202, 4202])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, low-angle, static shot meticulously depicts a slender Amanita mushroom as it steadily creeps through the cracks of a dark, cracked sidewalk. The asphalt, slick and reflective from a recent downpour, reflects the gloomy sky. The mushroom's cap is a smooth, smoky-gray dome, its fragile stem displaying immense strength in its slow, deliberate upward thrust. Suddenly, a bright, silent flash of lightning violently exposes the scene for a moment, painting the frame pure white. When the intense light momentarily fades, the mushroom reveals intricate, glowing blue circuits that run like veins through its now translucent stem and delicate gills. This internal electric blue light gently pulsates, casting a captivating cyberpunk neon glow across the damp, textured ground below, transforming the mundane urban decay into an eerie biotechnological marvel.", + "seed": 9104, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_4202_seed_9104_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07318966267630458, + "vlm_logprob": -0.07318966267630458, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.07318966267630458, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12849299609661102 + ], + "opposite_logprob_tokens": [ + -2.253493070602417 + ], + "logprob_reward": -0.12849299609661102, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12849299609661102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37554627656936646, + 0.11148135364055634, + 0.1480368822813034, + 0.05016433075070381, + 0.04400099441409111, + 0.07080940157175064, + 0.04906461015343666, + 0.05117730423808098, + 0.07536367326974869, + 0.02435518056154251 + ], + "credit_map": [ + 0.028179071843624115, + 0.03889639675617218, + 0.025932442396879196, + 0.024972012266516685, + 0.07082519680261612, + 0.05305875837802887, + 0.07325053960084915, + 0.0604318305850029, + 0.007639440707862377, + 0.011296113021671772, + 0.0094009879976511, + 0.010860259644687176, + 0.024476852267980576, + 0.014243967831134796, + 0.01696464605629444, + 0.016599096357822418, + 0.011463250033557415, + 0.013321089558303356, + 0.01716645248234272, + 0.012886433862149715, + 0.030088301748037338, + 0.017528098076581955, + 0.02611021138727665, + 0.019473036751151085, + 0.0033758303616195917, + 0.0046953437849879265, + 0.004177586175501347, + 0.00451489444822073, + 0.009593631140887737, + 0.007178134750574827, + 0.009211614727973938, + 0.00741729699075222, + 0.0049204896204173565, + 0.004839809611439705, + 0.0035935193300247192, + 0.0034663791302591562, + 0.008277804590761662, + 0.006066707894206047, + 0.005818880628794432, + 0.0070174019783735275, + 0.011350944638252258, + 0.010593661107122898, + 0.007017231546342373, + 0.006046749651432037, + 0.007900371216237545, + 0.006343570537865162, + 0.012317533604800701, + 0.00923934392631054, + 0.010635674931108952, + 0.007262285798788071, + 0.005759926978498697, + 0.004281711298972368, + 0.004740484990179539, + 0.0036271752323955297, + 0.005210197996348143, + 0.007547152228653431, + 0.006496103946119547, + 0.006144106388092041, + 0.006903947796672583, + 0.005328044295310974, + 0.00501984590664506, + 0.00879957340657711, + 0.00613821716979146, + 0.006347467191517353, + 0.01016488391906023, + 0.009395884349942207, + 0.011249118484556675, + 0.00642814114689827, + 0.0052892123349010944, + 0.015310912393033504, + 0.007334361784160137, + 0.010191160254180431, + 0.0022601536475121975, + 0.0027083877939730883, + 0.002914316253736615, + 0.0031361037399619818, + 0.00345090520568192, + 0.003118563210591674, + 0.003963744733482599, + 0.0028030069079250097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.12849299609661102 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040049493312836 + ], + "opposite_logprob_tokens": [ + -1.9104005098342896 + ], + "logprob_reward": -0.16040049493312836, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040049493312836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16088755428791046, + 0.047845467925071716, + 0.09163860976696014, + 0.10635443776845932, + 0.05043380707502365, + 0.10509206354618073, + 0.069992296397686, + 0.09795750677585602, + 0.24624107778072357, + 0.023557186126708984 + ], + "credit_map": [ + 0.013901574537158012, + 0.011404691264033318, + 0.0201082993298769, + 0.012459414079785347, + 0.02612590789794922, + 0.024038419127464294, + 0.025010770186781883, + 0.027838479727506638, + 0.006387172266840935, + 0.005081060342490673, + 0.006439997348934412, + 0.004635023884475231, + 0.006763841025531292, + 0.006119886878877878, + 0.0062360637821257114, + 0.00618241960182786, + 0.007634475827217102, + 0.008470402099192142, + 0.011458350345492363, + 0.008385621011257172, + 0.013199608772993088, + 0.01342232245951891, + 0.017387228086590767, + 0.011680607683956623, + 0.008154294453561306, + 0.01344191376119852, + 0.010749559849500656, + 0.010533206164836884, + 0.02132163941860199, + 0.01515858806669712, + 0.01564740762114525, + 0.011347822844982147, + 0.005279280245304108, + 0.004956508055329323, + 0.005659529473632574, + 0.004939976613968611, + 0.009322736412286758, + 0.005888207349926233, + 0.00734095461666584, + 0.007046610116958618, + 0.021944623440504074, + 0.01811353676021099, + 0.012557200156152248, + 0.010203525424003601, + 0.009285829961299896, + 0.009930913336575031, + 0.008485246449708939, + 0.01457119733095169, + 0.010391206480562687, + 0.007798351813107729, + 0.010881305672228336, + 0.008890807628631592, + 0.007153677754104137, + 0.006085229571908712, + 0.008378778584301472, + 0.01041293703019619, + 0.019101817160844803, + 0.011995984241366386, + 0.016897788271307945, + 0.010901334695518017, + 0.004963799845427275, + 0.012747852131724358, + 0.0075983586721122265, + 0.013750575482845306, + 0.025677228346467018, + 0.023619739338755608, + 0.04113016277551651, + 0.025057172402739525, + 0.017533456906676292, + 0.06333085894584656, + 0.016987748444080353, + 0.03290468454360962, + 0.0023244437761604786, + 0.0016245911829173565, + 0.002924442756921053, + 0.0029274208936840296, + 0.0025148442946374416, + 0.00349180749617517, + 0.003976786509156227, + 0.0037728503812104464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.16040049493312836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265416111797094 + ], + "opposite_logprob_tokens": [ + -5.255265235900879 + ], + "logprob_reward": -0.005265416111797094, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265416111797094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13452041149139404, + 0.049168650060892105, + 0.0432538241147995, + 0.1225360706448555, + 0.11341123282909393, + 0.09302331507205963, + 0.12176959216594696, + 0.05636702477931976, + 0.0747215524315834, + 0.19122830033302307 + ], + "credit_map": [ + 0.011530397459864616, + 0.014135592617094517, + 0.008043879643082619, + 0.009124666452407837, + 0.026675190776586533, + 0.01671854965388775, + 0.02326819859445095, + 0.0250239335000515, + 0.004034235142171383, + 0.0033754962496459484, + 0.004848548676818609, + 0.00450384384021163, + 0.008319251239299774, + 0.007602088153362274, + 0.008624873124063015, + 0.007860313169658184, + 0.0035914438776671886, + 0.003283064579591155, + 0.005264591425657272, + 0.0031251213513314724, + 0.008162358775734901, + 0.00635613314807415, + 0.007867061533033848, + 0.005604048725217581, + 0.008812122978270054, + 0.014497343450784683, + 0.012424108572304249, + 0.01578816957771778, + 0.01978464238345623, + 0.019173195585608482, + 0.01743755117058754, + 0.014618935063481331, + 0.009922655299305916, + 0.011190840974450111, + 0.010263162665069103, + 0.009277581237256527, + 0.027963899075984955, + 0.014443637803196907, + 0.014135042205452919, + 0.016214413568377495, + 0.010138463228940964, + 0.009532993659377098, + 0.010511541739106178, + 0.011677218601107597, + 0.017077932134270668, + 0.010029911063611507, + 0.011293794959783554, + 0.012761462479829788, + 0.020781226456165314, + 0.01889495737850666, + 0.01318663265556097, + 0.01242520660161972, + 0.013564222492277622, + 0.011526157148182392, + 0.013709141872823238, + 0.017682045698165894, + 0.00830281712114811, + 0.006708801724016666, + 0.006830207537859678, + 0.00663356576114893, + 0.004701865371316671, + 0.009530418552458286, + 0.006793168839067221, + 0.006866182200610638, + 0.00560800451785326, + 0.00641001807525754, + 0.007668538950383663, + 0.006251177750527859, + 0.01000962033867836, + 0.015375901013612747, + 0.01179603673517704, + 0.011602255515754223, + 0.012230970896780491, + 0.01228458620607853, + 0.0189793910831213, + 0.016427168622612953, + 0.0292416550219059, + 0.038716137409210205, + 0.036280177533626556, + 0.027068205177783966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.005265416111797094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809741511940956 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809741511940956, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809741511940956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11172625422477722, + 0.057432737201452255, + 0.05298929288983345, + 0.09182985872030258, + 0.08599694073200226, + 0.2086828202009201, + 0.13639871776103973, + 0.039361659437417984, + 0.06899749487638474, + 0.14658425748348236 + ], + "credit_map": [ + 0.009356214664876461, + 0.008696800097823143, + 0.011877849698066711, + 0.010001681745052338, + 0.021314002573490143, + 0.014092047698795795, + 0.02096061035990715, + 0.015427043661475182, + 0.003748612944036722, + 0.004623823799192905, + 0.0037896204739809036, + 0.0048457863740623, + 0.015686387196183205, + 0.0084418635815382, + 0.007370325271040201, + 0.008926311507821083, + 0.003779098391532898, + 0.004342503380030394, + 0.0038103905972093344, + 0.004511070437729359, + 0.013680522330105305, + 0.00909380242228508, + 0.00643963273614645, + 0.0073322718963027, + 0.005397734697908163, + 0.005941367242485285, + 0.00920569896697998, + 0.007056381553411484, + 0.02261943556368351, + 0.0116501459851861, + 0.017453931272029877, + 0.012505159713327885, + 0.008012011647224426, + 0.00851176306605339, + 0.009076506830751896, + 0.008668223395943642, + 0.01524562668055296, + 0.011743182316422462, + 0.011771820485591888, + 0.01296780165284872, + 0.05280512198805809, + 0.03261497616767883, + 0.0243923868983984, + 0.015104642137885094, + 0.016613399609923363, + 0.01905668154358864, + 0.02290089800953865, + 0.025194713845849037, + 0.016172071918845177, + 0.017909815534949303, + 0.014545712620019913, + 0.01656177081167698, + 0.01687062717974186, + 0.014087124727666378, + 0.021623967215418816, + 0.018627623096108437, + 0.0031958690378814936, + 0.0035884191747754812, + 0.004182267468422651, + 0.0038765789940953255, + 0.005093068815767765, + 0.005943916272372007, + 0.007908778265118599, + 0.00557276327162981, + 0.004314648453146219, + 0.00617271987721324, + 0.007662052754312754, + 0.006895488128066063, + 0.009700828231871128, + 0.010492721572518349, + 0.013665223494172096, + 0.010093812830746174, + 0.009469777345657349, + 0.014118483290076256, + 0.012412482872605324, + 0.013701885007321835, + 0.03669169917702675, + 0.017928471788764, + 0.022279750555753708, + 0.01998170092701912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.03809741511940956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369199112057686 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.03369199112057686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369199112057686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2935080826282501, + 0.1137973740696907, + 0.07990042865276337, + 0.08623437583446503, + 0.12147105485200882, + 0.08766433596611023, + 0.05011473596096039, + 0.04846193641424179, + 0.08031763881444931, + 0.03853004798293114 + ], + "credit_map": [ + 0.018597908318042755, + 0.018062284216284752, + 0.033277057111263275, + 0.03384646028280258, + 0.05555811896920204, + 0.026935214176774025, + 0.06439650058746338, + 0.04283454641699791, + 0.0065695745870471, + 0.008182771503925323, + 0.014195083640515804, + 0.012872186489403248, + 0.0250502098351717, + 0.012527968734502792, + 0.017890211194753647, + 0.01650938019156456, + 0.005822617560625076, + 0.00681056035682559, + 0.00864574033766985, + 0.0069857374764978886, + 0.015002654865384102, + 0.011358610354363918, + 0.014129182323813438, + 0.011145327240228653, + 0.007832646369934082, + 0.007444456685334444, + 0.008752665482461452, + 0.010333712212741375, + 0.014901842921972275, + 0.012784581631422043, + 0.011189539916813374, + 0.012994935736060143, + 0.008478980511426926, + 0.010869454592466354, + 0.013542823493480682, + 0.013130960054695606, + 0.024683356285095215, + 0.016103284433484077, + 0.01708633452653885, + 0.017575863748788834, + 0.00986697431653738, + 0.012021702714264393, + 0.010214447975158691, + 0.009985861368477345, + 0.012007974088191986, + 0.00891542062163353, + 0.013447490520775318, + 0.011204475536942482, + 0.004273931495845318, + 0.0058162035420536995, + 0.005818753503262997, + 0.005708918906748295, + 0.007819706574082375, + 0.0054264971986413, + 0.008679922670125961, + 0.006570805795490742, + 0.004344609100371599, + 0.0038396732416003942, + 0.005227150861173868, + 0.004405064973980188, + 0.007987499237060547, + 0.006448988802731037, + 0.00878159236162901, + 0.007427362259477377, + 0.004849500488489866, + 0.006046930328011513, + 0.007042940240353346, + 0.006690420676022768, + 0.010207954794168472, + 0.01816416345536709, + 0.014732815325260162, + 0.012582915835082531, + 0.0023369062691926956, + 0.003559964243322611, + 0.003635021159425378, + 0.0034093696158379316, + 0.008035201579332352, + 0.005343790166079998, + 0.0065870219841599464, + 0.005622774828225374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03369199112057686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot meticulously depicts a slender Amanita mushroom as it steadily creeps through the cracks of a dark, cracked sidewalk. The asphalt, slick and reflective from a recent downpour, reflects the gloomy sky. The mushroom's cap is a smooth, smoky-gray dome, its fragile stem displaying immense strength in its slow, deliberate upward thrust. Suddenly, a bright, silent flash of lightning violently exposes the scene for a moment, painting the frame pure white. When the intense light momentarily fades, the mushroom reveals intricate, glowing blue circuits that run like veins through its now translucent stem and delicate gills. This internal electric blue light gently pulsates, casting a captivating cyberpunk neon glow across the damp, textured ground below, transforming the mundane urban decay into an eerie biotechnological marvel.", + "seed": 9105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_4202_seed_9105_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1279950648546219, + "vlm_logprob": -0.1279950648546219, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.1279950648546219, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.307293176651001 + ], + "opposite_logprob_tokens": [ + -1.432293176651001 + ], + "logprob_reward": -0.307293176651001, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.307293176651001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25263506174087524, + 0.1661464124917984, + 0.15210629999637604, + 0.11214284598827362, + 0.05562923103570938, + 0.04998074844479561, + 0.04146708548069, + 0.05203485116362572, + 0.04771055653691292, + 0.070146843791008 + ], + "credit_map": [ + 0.056672051548957825, + 0.01598939672112465, + 0.04834238812327385, + 0.022181201726198196, + 0.025591522455215454, + 0.037204399704933167, + 0.01634097844362259, + 0.030313139781355858, + 0.029964668676257133, + 0.009237726218998432, + 0.02520477958023548, + 0.015295852907001972, + 0.01934157870709896, + 0.03209087252616882, + 0.011897100135684013, + 0.023113833740353584, + 0.022802095860242844, + 0.011367796920239925, + 0.013140634633600712, + 0.012858071364462376, + 0.025216247886419296, + 0.02802194282412529, + 0.022647593170404434, + 0.016051936894655228, + 0.01260241400450468, + 0.01772061549127102, + 0.012735279276967049, + 0.010285369120538235, + 0.014122682623565197, + 0.02024698071181774, + 0.011178149841725826, + 0.013251355849206448, + 0.005918421316891909, + 0.004926742520183325, + 0.009939218871295452, + 0.006139002740383148, + 0.006026123184710741, + 0.009424334391951561, + 0.006658842787146568, + 0.006596545223146677, + 0.00527389207854867, + 0.0058139231987297535, + 0.00470224441960454, + 0.0051184059120714664, + 0.007715169806033373, + 0.008358695544302464, + 0.007156596519052982, + 0.005841825623065233, + 0.004194364417344332, + 0.0053626056760549545, + 0.00619446812197566, + 0.0048932963982224464, + 0.005665133241564035, + 0.005395127460360527, + 0.005257639102637768, + 0.004504446405917406, + 0.006401906721293926, + 0.008685392327606678, + 0.005731227342039347, + 0.007812316995114088, + 0.005689649377018213, + 0.004875404294580221, + 0.005267215892672539, + 0.007571740075945854, + 0.005172477103769779, + 0.006689988076686859, + 0.00488253915682435, + 0.006506005767732859, + 0.0059529878199100494, + 0.0057995994575321674, + 0.007435851730406284, + 0.005271109752357006, + 0.0050055901519954205, + 0.013268663547933102, + 0.005753208417445421, + 0.007885496132075787, + 0.009969623759388924, + 0.009407521225512028, + 0.011223513633012772, + 0.0076332297176122665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.307293176651001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271173506975174 + ], + "opposite_logprob_tokens": [ + -2.127117395401001 + ], + "logprob_reward": -0.1271173506975174, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1271173506975174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0707077905535698, + 0.09026777744293213, + 0.05089987814426422, + 0.07246695458889008, + 0.2277124673128128, + 0.19924549758434296, + 0.0909101739525795, + 0.04858864098787308, + 0.09762823581695557, + 0.05157254636287689 + ], + "credit_map": [ + 0.006838475354015827, + 0.004641497507691383, + 0.00962426234036684, + 0.008871853351593018, + 0.009462796151638031, + 0.013006014749407768, + 0.008748597465455532, + 0.009514294564723969, + 0.018901778385043144, + 0.00499693350866437, + 0.010306976735591888, + 0.009198217652738094, + 0.00990210473537445, + 0.018077190965414047, + 0.006808767560869455, + 0.012075807899236679, + 0.006729611195623875, + 0.0029585491865873337, + 0.005666681099683046, + 0.005229217000305653, + 0.008566929027438164, + 0.010175147093832493, + 0.005584881640970707, + 0.005988863296806812, + 0.006734985858201981, + 0.007775720674544573, + 0.00794128142297268, + 0.009586403146386147, + 0.011654031462967396, + 0.008467333391308784, + 0.012487934902310371, + 0.007819263264536858, + 0.01798158511519432, + 0.02127937786281109, + 0.02419513277709484, + 0.02134758047759533, + 0.03955679386854172, + 0.045570988208055496, + 0.03028610534965992, + 0.027494901791214943, + 0.024658044800162315, + 0.02686980366706848, + 0.02436658926308155, + 0.019732994958758354, + 0.023314321413636208, + 0.04574599862098694, + 0.0108741894364357, + 0.023683568462729454, + 0.01040282379835844, + 0.015195905230939388, + 0.013722386211156845, + 0.012078351341187954, + 0.008206415921449661, + 0.015551049262285233, + 0.0058900583535432816, + 0.00986318476498127, + 0.0054382216185331345, + 0.010989181697368622, + 0.004051529336720705, + 0.005911797285079956, + 0.005552377551794052, + 0.006008324678987265, + 0.006045760121196508, + 0.004591451492160559, + 0.008130754344165325, + 0.013600648380815983, + 0.014117144979536533, + 0.01590854302048683, + 0.009982509538531303, + 0.01454963069409132, + 0.00949171558022499, + 0.011847290210425854, + 0.004706406034529209, + 0.007179965730756521, + 0.007641978561878204, + 0.0064565446227788925, + 0.006920453626662493, + 0.006544324103742838, + 0.005746428854763508, + 0.006376449018716812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.1271173506975174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036292895674705505 + ], + "opposite_logprob_tokens": [ + -5.628629207611084 + ], + "logprob_reward": -0.0036292895674705505, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036292895674705505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1497415155172348, + 0.08329363167285919, + 0.04042327031493187, + 0.0836857259273529, + 0.1763683408498764, + 0.105447918176651, + 0.04279197007417679, + 0.07757056504487991, + 0.09803815186023712, + 0.14263886213302612 + ], + "credit_map": [ + 0.011862370185554028, + 0.01187961082905531, + 0.021032750606536865, + 0.015031707473099232, + 0.02385447546839714, + 0.026527510955929756, + 0.018972693011164665, + 0.020580416545271873, + 0.016263054683804512, + 0.005127543583512306, + 0.013421514071524143, + 0.007747406139969826, + 0.011457230895757675, + 0.014053464867174625, + 0.0045610154047608376, + 0.010662409476935863, + 0.003854482201859355, + 0.0040334067307412624, + 0.004387943539768457, + 0.0042882091365754604, + 0.006742785219103098, + 0.005776004865765572, + 0.006625880487263203, + 0.004714564420282841, + 0.009503666311502457, + 0.01201654877513647, + 0.011545774526894093, + 0.011465679854154587, + 0.012317066080868244, + 0.008779660798609257, + 0.009512810036540031, + 0.00854452233761549, + 0.015145059674978256, + 0.014319602400064468, + 0.020664134994149208, + 0.020579904317855835, + 0.03191499039530754, + 0.0257475096732378, + 0.028387581929564476, + 0.019609574228525162, + 0.013575215823948383, + 0.013896954245865345, + 0.013110269792377949, + 0.009764275513589382, + 0.014339013025164604, + 0.015011095441877842, + 0.013336390256881714, + 0.012414716184139252, + 0.0051306975074112415, + 0.005235198885202408, + 0.0042954133823513985, + 0.004443752579391003, + 0.007266617845743895, + 0.005169979762285948, + 0.00660623237490654, + 0.0046440837904810905, + 0.006263322662562132, + 0.009756434708833694, + 0.0069524808786809444, + 0.011554776690900326, + 0.0147560341283679, + 0.008131914772093296, + 0.011661401949822903, + 0.00849420391023159, + 0.013739694841206074, + 0.01601121947169304, + 0.012554164975881577, + 0.01103190891444683, + 0.009828622452914715, + 0.013701106421649456, + 0.0070636337623000145, + 0.014107812196016312, + 0.015534271486103535, + 0.016049617901444435, + 0.014609308913350105, + 0.01640784740447998, + 0.020825903862714767, + 0.014849657192826271, + 0.028457291424274445, + 0.015904981642961502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0036292895674705505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292265355587006 + ], + "opposite_logprob_tokens": [ + -2.2379226684570312 + ], + "logprob_reward": -0.11292265355587006, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292265355587006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11849302798509598, + 0.06069594994187355, + 0.04498671367764473, + 0.25128477811813354, + 0.20087505877017975, + 0.06077155843377113, + 0.016380518674850464, + 0.058422576636075974, + 0.08053214848041534, + 0.10755763202905655 + ], + "credit_map": [ + 0.01180764939635992, + 0.010541358962655067, + 0.01791001483798027, + 0.017904620617628098, + 0.013883906416594982, + 0.014231768436729908, + 0.014981027692556381, + 0.017232680693268776, + 0.00654634740203619, + 0.005774557124823332, + 0.008722818456590176, + 0.006167090963572264, + 0.009282959625124931, + 0.009733780287206173, + 0.007099458947777748, + 0.007368939463049173, + 0.00746507104486227, + 0.003449545009061694, + 0.006025543436408043, + 0.005757422652095556, + 0.007406752556562424, + 0.005036457907408476, + 0.005357732065021992, + 0.004488189704716206, + 0.02584809996187687, + 0.03791624680161476, + 0.02657340280711651, + 0.0324748195707798, + 0.04429321363568306, + 0.029695207253098488, + 0.025521164759993553, + 0.0289626382291317, + 0.01642470434308052, + 0.016076957806944847, + 0.02780730277299881, + 0.016094770282506943, + 0.035778723657131195, + 0.03568840026855469, + 0.028795570135116577, + 0.024208636954426765, + 0.007535210810601711, + 0.009225272573530674, + 0.005575947929173708, + 0.006462832912802696, + 0.007839729078114033, + 0.008787195198237896, + 0.008639645762741566, + 0.0067057241685688496, + 0.0017262991750612855, + 0.0024321232922375202, + 0.0022461649496108294, + 0.0024837919045239687, + 0.0019328949274495244, + 0.0015718378126621246, + 0.00194394844584167, + 0.002043458865955472, + 0.0045865243300795555, + 0.007116587366908789, + 0.006028740666806698, + 0.007257758639752865, + 0.010021920315921307, + 0.006362833082675934, + 0.010478422045707703, + 0.006569792050868273, + 0.00768496748059988, + 0.0108694639056921, + 0.008146407082676888, + 0.010868746787309647, + 0.01129175815731287, + 0.010981891304254532, + 0.009970147162675858, + 0.010718770325183868, + 0.012829416431486607, + 0.012599315494298935, + 0.009778184816241264, + 0.011517288163304329, + 0.01642869971692562, + 0.011487668380141258, + 0.02182961441576481, + 0.011087432503700256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.11292265355587006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901285380125046 + ], + "opposite_logprob_tokens": [ + -2.464012861251831 + ], + "logprob_reward": -0.08901285380125046, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901285380125046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250331699848175, + 0.08960920572280884, + 0.0508464016020298, + 0.11033453792333603, + 0.10949365049600601, + 0.07274039834737778, + 0.0770743265748024, + 0.09323914349079132, + 0.1093134954571724, + 0.16231563687324524 + ], + "credit_map": [ + 0.01469892542809248, + 0.00883827917277813, + 0.016699479892849922, + 0.014371485449373722, + 0.019475877285003662, + 0.02084686979651451, + 0.013231380842626095, + 0.016870880499482155, + 0.016526831313967705, + 0.0047488161362707615, + 0.013851560652256012, + 0.00916000735014677, + 0.010973606258630753, + 0.017003804445266724, + 0.006040217354893684, + 0.011304368264973164, + 0.00857111532241106, + 0.0036270844284445047, + 0.0063791233114898205, + 0.0058465441688895226, + 0.006632678210735321, + 0.008496358059346676, + 0.004845994058996439, + 0.006447506137192249, + 0.009350406005978584, + 0.010505386628210545, + 0.0129556180909276, + 0.014853235334157944, + 0.020062336698174477, + 0.01566988043487072, + 0.015486801974475384, + 0.011450882069766521, + 0.006935277488082647, + 0.01250798162072897, + 0.011898599565029144, + 0.008804166689515114, + 0.01975296251475811, + 0.020866699516773224, + 0.015713518485426903, + 0.013014450669288635, + 0.007155290339142084, + 0.008800634182989597, + 0.00795983150601387, + 0.007613539695739746, + 0.00925182830542326, + 0.015543889254331589, + 0.00651300884783268, + 0.009902384132146835, + 0.011617456562817097, + 0.011919023469090462, + 0.008949861861765385, + 0.010652121156454086, + 0.007256220560520887, + 0.010264342650771141, + 0.008867346681654453, + 0.0075479596853256226, + 0.008515022695064545, + 0.012128086760640144, + 0.00751951988786459, + 0.011219058185815811, + 0.014827175065875053, + 0.014199025928974152, + 0.014754743315279484, + 0.010076516307890415, + 0.008128918707370758, + 0.015053655952215195, + 0.01106936950236559, + 0.012561815790832043, + 0.020758790895342827, + 0.0112559013068676, + 0.019282376393675804, + 0.011202670633792877, + 0.012294908985495567, + 0.023394983261823654, + 0.014630251564085484, + 0.018740441650152206, + 0.023715075105428696, + 0.021716076880693436, + 0.029397839680314064, + 0.01842607744038105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.08901285380125046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot meticulously depicts a slender Amanita mushroom as it steadily creeps through the cracks of a dark, cracked sidewalk. The asphalt, slick and reflective from a recent downpour, reflects the gloomy sky. The mushroom's cap is a smooth, smoky-gray dome, its fragile stem displaying immense strength in its slow, deliberate upward thrust. Suddenly, a bright, silent flash of lightning violently exposes the scene for a moment, painting the frame pure white. When the intense light momentarily fades, the mushroom reveals intricate, glowing blue circuits that run like veins through its now translucent stem and delicate gills. This internal electric blue light gently pulsates, casting a captivating cyberpunk neon glow across the damp, textured ground below, transforming the mundane urban decay into an eerie biotechnological marvel.", + "seed": 9106, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_4202_seed_9106_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10393291246145964, + "vlm_logprob": -0.10393291246145964, + "vlm_yes_no_margin": 2.925000023841858, + "vlm_token_credit": -0.10393291246145964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1580464094877243 + ], + "opposite_logprob_tokens": [ + -2.0330464839935303 + ], + "logprob_reward": -0.1580464094877243, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1580464094877243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49769940972328186, + 0.10328848659992218, + 0.029875632375478745, + 0.057942986488342285, + 0.06034253165125847, + 0.02944371849298477, + 0.048411354422569275, + 0.05356823652982712, + 0.032353874295949936, + 0.08707372844219208 + ], + "credit_map": [ + 0.0397292822599411, + 0.05606772005558014, + 0.04180940240621567, + 0.06819114089012146, + 0.06016234681010246, + 0.0735718160867691, + 0.08254575729370117, + 0.07562196254730225, + 0.008156306110322475, + 0.010256621055305004, + 0.010247958824038506, + 0.014918630942702293, + 0.01033896766602993, + 0.016532521694898605, + 0.01586655154824257, + 0.016970936208963394, + 0.003171321237459779, + 0.00265646749176085, + 0.002774375956505537, + 0.003956303931772709, + 0.003367582568898797, + 0.004637680482119322, + 0.004484867677092552, + 0.004827036522328854, + 0.00560346432030201, + 0.005939822178333998, + 0.008500452153384686, + 0.009622064419090748, + 0.004588668700307608, + 0.008946986868977547, + 0.006717970594763756, + 0.008023558184504509, + 0.00426325760781765, + 0.0055975099094212055, + 0.006323631387203932, + 0.008826158009469509, + 0.0062366388738155365, + 0.0074679856188595295, + 0.012448970228433609, + 0.009178382344543934, + 0.0024164430797100067, + 0.0020553453359752893, + 0.005057250615209341, + 0.004571556579321623, + 0.0022578020580112934, + 0.003001568606123328, + 0.00528793316334486, + 0.00479581905528903, + 0.004668628331273794, + 0.004467999562621117, + 0.005391014274209738, + 0.004589826799929142, + 0.005091613624244928, + 0.011640148237347603, + 0.005409848410636187, + 0.007152279373258352, + 0.004928174428641796, + 0.009377120062708855, + 0.004630893934518099, + 0.005702659487724304, + 0.003602012060582638, + 0.005452677607536316, + 0.005026514641940594, + 0.014848185703158379, + 0.003561069956049323, + 0.004025748465210199, + 0.003853571368381381, + 0.004013787489384413, + 0.0036117550916969776, + 0.004557155538350344, + 0.005493887234479189, + 0.003236902877688408, + 0.009829619899392128, + 0.01655638962984085, + 0.0075539518147706985, + 0.008308586664497852, + 0.01758355088531971, + 0.009606871753931046, + 0.008763445541262627, + 0.008871319703757763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.1580464094877243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995329201221466 + ], + "opposite_logprob_tokens": [ + -1.8049533367156982 + ], + "logprob_reward": -0.17995329201221466, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995329201221466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43271830677986145, + 0.07377704977989197, + 0.02926916815340519, + 0.0535920187830925, + 0.023166242986917496, + 0.05669840797781944, + 0.04399769380688667, + 0.19137179851531982, + 0.05944054201245308, + 0.03596873581409454 + ], + "credit_map": [ + 0.03347140923142433, + 0.053353846073150635, + 0.032075244933366776, + 0.05866484344005585, + 0.05415823683142662, + 0.06263754516839981, + 0.07137923687696457, + 0.06697797030210495, + 0.00651058554649353, + 0.007725083269178867, + 0.006239557638764381, + 0.009754965081810951, + 0.00789804756641388, + 0.012059318833053112, + 0.012181457132101059, + 0.011408040300011635, + 0.0019805612973868847, + 0.002570286625996232, + 0.0028199166990816593, + 0.004467683844268322, + 0.003722491906955838, + 0.0032090849708765745, + 0.00642642006278038, + 0.0040727234445512295, + 0.004199143499135971, + 0.005659585818648338, + 0.0034977798350155354, + 0.008372820913791656, + 0.007007007487118244, + 0.005901285447180271, + 0.011826911009848118, + 0.00712748384103179, + 0.002060441765934229, + 0.001921942108310759, + 0.003036676673218608, + 0.0029979730024933815, + 0.003031406318768859, + 0.0026054929476231337, + 0.004221837967634201, + 0.0032904702238738537, + 0.0043380833230912685, + 0.00661905761808157, + 0.004472715780138969, + 0.006210722960531712, + 0.01017611101269722, + 0.004721799865365028, + 0.011924800463020802, + 0.008235118351876736, + 0.004036430735141039, + 0.003938318695873022, + 0.0036865295842289925, + 0.004663502331823111, + 0.004811374004930258, + 0.008758116513490677, + 0.007215926889330149, + 0.006887495983392, + 0.0140844052657485, + 0.04287374019622803, + 0.013484067283570766, + 0.0216461643576622, + 0.01000783871859312, + 0.011778934858739376, + 0.006847424432635307, + 0.07064923644065857, + 0.005161626264452934, + 0.005247701425105333, + 0.0058425478637218475, + 0.005311787128448486, + 0.005044101737439632, + 0.014755991287529469, + 0.004833913873881102, + 0.013242870569229126, + 0.004392968025058508, + 0.00395227549597621, + 0.00732786301523447, + 0.004315582104027271, + 0.004839200526475906, + 0.003201397368684411, + 0.0039541469886898994, + 0.003985305782407522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.17995329201221466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268261767923832 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005268261767923832, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268261767923832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5129250884056091, + 0.07788775861263275, + 0.02913796715438366, + 0.022841686382889748, + 0.019616704434156418, + 0.03375488892197609, + 0.07213590294122696, + 0.08510982990264893, + 0.06228503957390785, + 0.08430507034063339 + ], + "credit_map": [ + 0.039555083960294724, + 0.06345254927873611, + 0.03802337870001793, + 0.0697367787361145, + 0.06087661162018776, + 0.07693550735712051, + 0.08445433527231216, + 0.07989086955785751, + 0.0059325844049453735, + 0.008603682741522789, + 0.006912978366017342, + 0.011829102411866188, + 0.009815679863095284, + 0.011051452718675137, + 0.013914493843913078, + 0.00982778798788786, + 0.002291339449584484, + 0.002611500211060047, + 0.0027395503129810095, + 0.004277926869690418, + 0.004142988473176956, + 0.0029851444996893406, + 0.004853368736803532, + 0.0052361502312123775, + 0.0015644310042262077, + 0.002134929411113262, + 0.002569768112152815, + 0.003219315316528082, + 0.002329879440367222, + 0.003322425065562129, + 0.004139784257858992, + 0.0035611565690487623, + 0.0014093121280893683, + 0.0020154290832579136, + 0.0021465697791427374, + 0.00274886772967875, + 0.0022835293784737587, + 0.0019920042250305414, + 0.004167850129306316, + 0.0028531423304229975, + 0.0022475854493677616, + 0.0033251065760850906, + 0.003129095770418644, + 0.004898985847830772, + 0.003967299126088619, + 0.004122772254049778, + 0.00785451103001833, + 0.004209533333778381, + 0.0060424646362662315, + 0.008146104402840137, + 0.005414292216300964, + 0.012610715813934803, + 0.008430462330579758, + 0.008222414180636406, + 0.013847357593476772, + 0.009422102011740208, + 0.007555334363132715, + 0.011489158496260643, + 0.006435403600335121, + 0.009030221030116081, + 0.003877492155879736, + 0.012585372664034367, + 0.004869639873504639, + 0.0292672086507082, + 0.006285133771598339, + 0.006130986846983433, + 0.0067544919438660145, + 0.0074856700375676155, + 0.005212853197008371, + 0.008747375570237637, + 0.008513469249010086, + 0.01315506361424923, + 0.009195312857627869, + 0.014368773438036442, + 0.009107548743486404, + 0.008926342241466045, + 0.017727576196193695, + 0.006928177550435066, + 0.009088770486414433, + 0.00896257534623146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.005268261767923832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4565121531486511, + 0.05751090869307518, + 0.018326474353671074, + 0.016199108213186264, + 0.015966380015015602, + 0.02107630856335163, + 0.06856603920459747, + 0.07635657489299774, + 0.2370162308216095, + 0.03246978297829628 + ], + "credit_map": [ + 0.03552363067865372, + 0.05620061233639717, + 0.03282659873366356, + 0.06207916513085365, + 0.056196440011262894, + 0.06750272959470749, + 0.07609548419713974, + 0.07008752226829529, + 0.00465295510366559, + 0.0076569924131035805, + 0.004878389183431864, + 0.008073811419308186, + 0.0073069310747087, + 0.008429791778326035, + 0.008892939426004887, + 0.00761910667642951, + 0.001614870736375451, + 0.0016615379136055708, + 0.002278510481119156, + 0.002694103866815567, + 0.0013914393493905663, + 0.003570863278582692, + 0.0022454618010669947, + 0.002869687508791685, + 0.001351004233583808, + 0.0017081585247069597, + 0.002043388318270445, + 0.0030413127969950438, + 0.0016959039494395256, + 0.001729375682771206, + 0.0024903840385377407, + 0.0021395832300186157, + 0.0012397188693284988, + 0.0013697546673938632, + 0.0014016464119777083, + 0.0026710834354162216, + 0.001595054636709392, + 0.0021523060277104378, + 0.0031141985673457384, + 0.0024226191453635693, + 0.0016908750403672457, + 0.0015299733495339751, + 0.0025788801722228527, + 0.002581915818154812, + 0.0020763834472745657, + 0.0034760339185595512, + 0.004115914460271597, + 0.0030263348016887903, + 0.008069873787462711, + 0.005831936839967966, + 0.008362958207726479, + 0.006824410054832697, + 0.006229623220860958, + 0.014965558424592018, + 0.00888274796307087, + 0.009398937225341797, + 0.006210720632225275, + 0.01381805818527937, + 0.006260327063500881, + 0.009193526580929756, + 0.005099515430629253, + 0.00678902817890048, + 0.009007494896650314, + 0.019977910444140434, + 0.019615652039647102, + 0.015162477269768715, + 0.017840558663010597, + 0.024717286229133606, + 0.013731510378420353, + 0.0575302354991436, + 0.021992696449160576, + 0.06642584502696991, + 0.004146214108914137, + 0.004977615550160408, + 0.0037347853649407625, + 0.0032658306881785393, + 0.004411920439451933, + 0.005385281052440405, + 0.0035540671087801456, + 0.0029940700624138117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.016067897900938988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603287011384964 + ], + "opposite_logprob_tokens": [ + -1.91032874584198 + ], + "logprob_reward": -0.1603287011384964, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603287011384964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4535236954689026, + 0.019971929490566254, + 0.026806768029928207, + 0.035706374794244766, + 0.017298735678195953, + 0.10142297297716141, + 0.06267639994621277, + 0.0332673043012619, + 0.15341611206531525, + 0.09590969979763031 + ], + "credit_map": [ + 0.03594392165541649, + 0.05489441752433777, + 0.032618630677461624, + 0.061980459839105606, + 0.05651122331619263, + 0.06647007167339325, + 0.07483651489019394, + 0.07026843726634979, + 0.0021273253951221704, + 0.0018251353176310658, + 0.002312862081453204, + 0.0028880969621241093, + 0.002129443921148777, + 0.0024735217448323965, + 0.0026411612052470446, + 0.003574382746592164, + 0.0020893218461424112, + 0.002643618267029524, + 0.0032012751325964928, + 0.004079555161297321, + 0.0032630518544465303, + 0.0023928924929350615, + 0.004424997605383396, + 0.004712054971605539, + 0.002608145819976926, + 0.003178686834871769, + 0.003817754564806819, + 0.004973647650331259, + 0.004540182184427977, + 0.003719872795045376, + 0.007239704951643944, + 0.005628378130495548, + 0.0015168049139901996, + 0.00137356820050627, + 0.0022331394720822573, + 0.0023246132768690586, + 0.0018491504015401006, + 0.002001581946387887, + 0.0032207053154706955, + 0.0027791736647486687, + 0.007494912017136812, + 0.012524035759270191, + 0.008110543712973595, + 0.012548678554594517, + 0.013951948843896389, + 0.008623147383332253, + 0.023852907121181488, + 0.014316797256469727, + 0.0053073023445904255, + 0.00613224133849144, + 0.005789614282548428, + 0.008100859820842743, + 0.007678230758756399, + 0.0106954425573349, + 0.01178276352584362, + 0.007189942058175802, + 0.0044618952088057995, + 0.003797100856900215, + 0.0033931154757738113, + 0.0032971107866615057, + 0.0022178427316248417, + 0.006812357343733311, + 0.003113760845735669, + 0.006174121052026749, + 0.010403379797935486, + 0.008052150718867779, + 0.009154180996119976, + 0.01356114074587822, + 0.00905598234385252, + 0.04908629506826401, + 0.012103652581572533, + 0.04199932888150215, + 0.01163563970476389, + 0.016186969354748726, + 0.010604341514408588, + 0.010920418426394463, + 0.018399303779006004, + 0.008936993777751923, + 0.010580088011920452, + 0.008645943365991116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.1603287011384964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot meticulously depicts a slender Amanita mushroom as it steadily creeps through the cracks of a dark, cracked sidewalk. The asphalt, slick and reflective from a recent downpour, reflects the gloomy sky. The mushroom's cap is a smooth, smoky-gray dome, its fragile stem displaying immense strength in its slow, deliberate upward thrust. Suddenly, a bright, silent flash of lightning violently exposes the scene for a moment, painting the frame pure white. When the intense light momentarily fades, the mushroom reveals intricate, glowing blue circuits that run like veins through its now translucent stem and delicate gills. This internal electric blue light gently pulsates, casting a captivating cyberpunk neon glow across the damp, textured ground below, transforming the mundane urban decay into an eerie biotechnological marvel.", + "seed": 9107, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_1/index_4202_seed_9107_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09854980185627937, + "vlm_logprob": -0.09854980185627937, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.09854980185627937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10894368588924408 + ], + "opposite_logprob_tokens": [ + -2.4839437007904053 + ], + "logprob_reward": -0.10894368588924408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10894368588924408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.118870809674263, + 0.032466862350702286, + 0.022154657170176506, + 0.020302869379520416, + 0.03713180124759674, + 0.058616649359464645, + 0.09774085879325867, + 0.009046242572367191, + 0.48721861839294434, + 0.11645065993070602 + ], + "credit_map": [ + 0.010615563951432705, + 0.00921321939677, + 0.010487028397619724, + 0.008615762926638126, + 0.025389602407813072, + 0.026521552354097366, + 0.011360779404640198, + 0.016667302697896957, + 0.003152864519506693, + 0.004144354723393917, + 0.0025621908716857433, + 0.0023983167484402657, + 0.005687321536242962, + 0.005816243588924408, + 0.004498227033764124, + 0.0042073423974215984, + 0.002319569233804941, + 0.0023578659165650606, + 0.0021872164215892553, + 0.0027027723845094442, + 0.003371619153767824, + 0.003477760823443532, + 0.002778078429400921, + 0.002959775971248746, + 0.0018693535821512341, + 0.0032050746958702803, + 0.001977452775463462, + 0.0024105519987642765, + 0.0036550972145050764, + 0.0020587246399372816, + 0.003022921970114112, + 0.002103692851960659, + 0.006250141188502312, + 0.004213384818285704, + 0.0023935320787131786, + 0.0015969047090038657, + 0.008082324638962746, + 0.0054892380721867085, + 0.0022061257623136044, + 0.006900153122842312, + 0.02169015072286129, + 0.016584770753979683, + 0.00232116156257689, + 0.0016621678369119763, + 0.006021200213581324, + 0.005500193685293198, + 0.0018726274138316512, + 0.0029643746092915535, + 0.010906504467129707, + 0.01513700745999813, + 0.006349658593535423, + 0.008592412807047367, + 0.025798164308071136, + 0.015441226772964, + 0.007522742263972759, + 0.007993148639798164, + 0.0012035187100991607, + 0.0011974034132435918, + 0.0007579347002319992, + 0.000895938603207469, + 0.0019173126202076674, + 0.0012642752844840288, + 0.0009764048154465854, + 0.0008334534941241145, + 0.02526707947254181, + 0.012034254148602486, + 0.010544911026954651, + 0.008963928557932377, + 0.01460526417940855, + 0.22223825752735138, + 0.007795844227075577, + 0.18576909601688385, + 0.005365521181374788, + 0.0069546508602797985, + 0.004953163675963879, + 0.003007227089256048, + 0.007142236456274986, + 0.04379302263259888, + 0.0043825204484164715, + 0.040852323174476624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.10894368588924408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016109675168991 + ], + "opposite_logprob_tokens": [ + -1.7016109228134155 + ], + "logprob_reward": -0.2016109675168991, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016109675168991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03726613521575928, + 0.02103245072066784, + 0.02305058017373085, + 0.020081425085663795, + 0.036626774817705154, + 0.21998965740203857, + 0.042900532484054565, + 0.041218556463718414, + 0.5067041516304016, + 0.05112973600625992 + ], + "credit_map": [ + 0.005213430151343346, + 0.00259519019164145, + 0.005467810668051243, + 0.0032878208439797163, + 0.00729389488697052, + 0.005137067753821611, + 0.00351341487839818, + 0.004757509101182222, + 0.003633978310972452, + 0.0016380691668018699, + 0.003937747795134783, + 0.00184211740270257, + 0.0029647997580468655, + 0.003036009380593896, + 0.0013981793308630586, + 0.002581549808382988, + 0.0038663316518068314, + 0.0017469300655648112, + 0.004725389648228884, + 0.002297372091561556, + 0.0034344790037721395, + 0.002521539106965065, + 0.0017398333875462413, + 0.002718706149607897, + 0.0031250440515577793, + 0.0016966930124908686, + 0.0036594346165657043, + 0.002345754997804761, + 0.0026750427205115557, + 0.002356857294216752, + 0.0016888114623725414, + 0.0025337873958051205, + 0.004420422483235598, + 0.0022989134304225445, + 0.0031101920176297426, + 0.0014808306004852057, + 0.004415431991219521, + 0.009557600133121014, + 0.001751552103087306, + 0.00959183182567358, + 0.09122925996780396, + 0.07306133955717087, + 0.008416234515607357, + 0.005976511165499687, + 0.008491292595863342, + 0.017166761681437492, + 0.003793630050495267, + 0.011854637414216995, + 0.004482590593397617, + 0.0035933395847678185, + 0.0050858622416853905, + 0.00811120867729187, + 0.006960769183933735, + 0.005325223784893751, + 0.004504328593611717, + 0.0048372093588113785, + 0.007171331904828548, + 0.004740475676953793, + 0.005125050898641348, + 0.0046440488658845425, + 0.004736654926091433, + 0.006905985530465841, + 0.003197185229510069, + 0.00469782343134284, + 0.02545875683426857, + 0.0126736955717206, + 0.012131205759942532, + 0.009097311645746231, + 0.014898587949573994, + 0.23054568469524384, + 0.008125891909003258, + 0.1937730312347412, + 0.003774490440264344, + 0.0024118423461914062, + 0.003131423844024539, + 0.0016584205441176891, + 0.002523398259654641, + 0.0185270793735981, + 0.0022640423849225044, + 0.01683903858065605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.2016109675168991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204349368810654 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204349368810654, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204349368810654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03532536327838898, + 0.009198324754834175, + 0.008375962264835835, + 0.006751325447112322, + 0.014189847745001316, + 0.21383805572986603, + 0.03041333705186844, + 0.020619485527276993, + 0.6257985234260559, + 0.03548981994390488 + ], + "credit_map": [ + 0.00510107446461916, + 0.0024193325079977512, + 0.005568754393607378, + 0.0026112154591828585, + 0.0053345258347690105, + 0.005671299993991852, + 0.003542267717421055, + 0.005076896864920855, + 0.0013454033760353923, + 0.0007238407270051539, + 0.0012366451555863023, + 0.0007940282812342048, + 0.0011870925081893802, + 0.0016456091543659568, + 0.0010651672491803765, + 0.0012005384778603911, + 0.0012211190769448876, + 0.000870495627168566, + 0.0011592407245188951, + 0.0009205598034895957, + 0.0011685307836160064, + 0.0010364892659708858, + 0.0009095208952203393, + 0.0010900060879066586, + 0.0008634647238068283, + 0.0008082788554020226, + 0.0005811384180560708, + 0.0005208830698393285, + 0.0012701209634542465, + 0.000877508835401386, + 0.0011044364655390382, + 0.0007254946394823492, + 0.0014676348073408008, + 0.001062125200405717, + 0.0014921629335731268, + 0.0008806525147520006, + 0.0025825046468526125, + 0.0032624234445393085, + 0.00087404326768592, + 0.002568301046267152, + 0.08855991065502167, + 0.06835746765136719, + 0.008389968425035477, + 0.006048808805644512, + 0.008597072213888168, + 0.01890600100159645, + 0.0037341758143156767, + 0.011244652792811394, + 0.006123875733464956, + 0.003510294249281287, + 0.004353872034698725, + 0.002927783178165555, + 0.00405774125829339, + 0.0037181347142904997, + 0.0022824013140052557, + 0.003439234336838126, + 0.002519630128517747, + 0.0019051198614761233, + 0.0016613831976428628, + 0.0021457839757204056, + 0.0031386390328407288, + 0.0037756930105388165, + 0.0027550687082111835, + 0.0027181676123291254, + 0.03377001732587814, + 0.014860745519399643, + 0.014221147634088993, + 0.011825810186564922, + 0.019078824669122696, + 0.2851865887641907, + 0.010115253739058971, + 0.2367401272058487, + 0.0019689123146235943, + 0.00233368668705225, + 0.0017546536400914192, + 0.0010606736177578568, + 0.0030187729280442, + 0.011982133612036705, + 0.0019929222762584686, + 0.01137806661427021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.014204349368810654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902997523546219 + ], + "opposite_logprob_tokens": [ + -2.4640300273895264 + ], + "logprob_reward": -0.08902997523546219, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902997523546219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04009955748915672, + 0.013363208621740341, + 0.01671699993312359, + 0.022064408287405968, + 0.018690502271056175, + 0.2039729803800583, + 0.01336982287466526, + 0.024001069366931915, + 0.4134157598018646, + 0.23430560529232025 + ], + "credit_map": [ + 0.005266728810966015, + 0.0032087918370962143, + 0.005188031122088432, + 0.002360054524615407, + 0.0094566335901618, + 0.005961557384580374, + 0.004058671183884144, + 0.004599094390869141, + 0.0013419832102954388, + 0.0016013074200600386, + 0.0014055910287424922, + 0.0011956084053963423, + 0.002350615570321679, + 0.002163685392588377, + 0.001803309889510274, + 0.0015011084033176303, + 0.0016702701104804873, + 0.002038746839389205, + 0.0021426742896437645, + 0.001570848049595952, + 0.0032938621006906033, + 0.0024177017621695995, + 0.001838739961385727, + 0.0017441578675061464, + 0.0027978175785392523, + 0.0021955869160592556, + 0.0027618652675300837, + 0.0023724515922367573, + 0.0034138215705752373, + 0.0034817249979823828, + 0.0025622716639190912, + 0.0024788700975477695, + 0.003098456421867013, + 0.001372124650515616, + 0.002176386769860983, + 0.0011785741662606597, + 0.0031499334145337343, + 0.00326893525198102, + 0.0015881792642176151, + 0.002857914427295327, + 0.08079779893159866, + 0.05885094776749611, + 0.007988868281245232, + 0.006075050216168165, + 0.012325935065746307, + 0.020481601357460022, + 0.00625568674877286, + 0.011197106912732124, + 0.001777050201781094, + 0.0012402195716276765, + 0.001016063615679741, + 0.0009784640278667212, + 0.0028802237939089537, + 0.002696672687307, + 0.0010447690729051828, + 0.001736361300572753, + 0.00413651205599308, + 0.0020100492984056473, + 0.0031950953416526318, + 0.002284812508150935, + 0.003661079565063119, + 0.0038192381616681814, + 0.0018109080847352743, + 0.003083376679569483, + 0.02272201143205166, + 0.009876642376184464, + 0.009546922519803047, + 0.007788673974573612, + 0.012498322874307632, + 0.18827012181282043, + 0.006757680792361498, + 0.15595541894435883, + 0.009811045601963997, + 0.007640046067535877, + 0.004678595811128616, + 0.004177797585725784, + 0.006754401605576277, + 0.10405358672142029, + 0.003872272092849016, + 0.09331788867712021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.08902997523546219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a slender Amanita mushroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom display glowing blue circuits within its translucent stem and gills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bright flash of lightning occur that momentarily paints the frame pure white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mushroom's stem shown performing a slow, deliberate upward thrust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896003127098083 + ], + "opposite_logprob_tokens": [ + -2.5789599418640137 + ], + "logprob_reward": -0.07896003127098083, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896003127098083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09580659121274948, + 0.020353589206933975, + 0.026572132483124733, + 0.03181521221995354, + 0.10826519131660461, + 0.06261546164751053, + 0.09025760740041733, + 0.06681197136640549, + 0.3575802147388458, + 0.13992208242416382 + ], + "credit_map": [ + 0.008197558112442493, + 0.009259091690182686, + 0.006295936647802591, + 0.011863465420901775, + 0.018651215359568596, + 0.014022879302501678, + 0.016323324292898178, + 0.011193113401532173, + 0.0019752311054617167, + 0.002226815093308687, + 0.0019442730117589235, + 0.0014627623604610562, + 0.004440263845026493, + 0.002600393956527114, + 0.0033631587866693735, + 0.002340689767152071, + 0.0030292843002825975, + 0.002946848515421152, + 0.002037328202277422, + 0.0022346244659274817, + 0.005130233243107796, + 0.003964680712670088, + 0.003903722856193781, + 0.003325409023091197, + 0.0046602897346019745, + 0.0035181844141334295, + 0.0032953687477856874, + 0.003483384381979704, + 0.004390927031636238, + 0.003836976829916239, + 0.004770740866661072, + 0.003859336953610182, + 0.00921480730175972, + 0.007954517379403114, + 0.004717385862022638, + 0.0055083297193050385, + 0.02164049819111824, + 0.030989404767751694, + 0.006657748017460108, + 0.02158248983323574, + 0.020072253420948982, + 0.01634308695793152, + 0.0049425046890974045, + 0.002535121515393257, + 0.005623070057481527, + 0.006312841083854437, + 0.0024751515593379736, + 0.004311432596296072, + 0.010122311301529408, + 0.012381674721837044, + 0.005572683177888393, + 0.006007941905409098, + 0.026877928525209427, + 0.013421915471553802, + 0.008081521838903427, + 0.007791630458086729, + 0.01272611878812313, + 0.008687159977853298, + 0.008621455170214176, + 0.007528594695031643, + 0.008222989737987518, + 0.008424835279583931, + 0.005658068228513002, + 0.00694273691624403, + 0.022457722574472427, + 0.009407497011125088, + 0.011684026569128036, + 0.0070485337637364864, + 0.01089826412498951, + 0.1588258594274521, + 0.007543432060629129, + 0.12971484661102295, + 0.005821747239679098, + 0.005352174397557974, + 0.003987742122262716, + 0.0028028511442244053, + 0.0058412738144397736, + 0.059240322560071945, + 0.003083290997892618, + 0.053792670369148254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.07896003127098083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_tensor([8637, 8637])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_tensor([8637, 8637])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..44c4f7be9d5e3f46f6ceab39a33d89d7016b55aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_tensor([8637, 8637])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "full-body shot, shot from a slightly elevated perspective, the camera slowly pans rightward, following the subject in a solemn and steady rhythm, capturing the empress's majestic demeanor. Her hair, crowned by a towering, magnificent bun, is adorned with an intricately carved golden phoenix hairpin, which sways gently with her movements, its tassels cascading down her temples. She wears an intricately crafted crimson silk phoenix robe. Its wide sleeves and long, floor-length skirt are embroidered with countless strands of gold thread, creating a vivid image of a phoenix ready to take flight. Each feather gleams with a metallic sheen. She walks solemnly down a grand and profound palace corridor, flanked by massive vermilion columns adorned with carved and painted beams. The floor is paved with a mirror-like, black marble that reflects a hazy figure. Her steps were steady and deliberate, each step sending her voluminous skirt undulating gracefully across the ground like a crimson wave. The delicate phoenix embroidery on her robe shone brilliantly in the dappled light filtering through a carved window pane on one side, while the gemstones and pearls adorned it with shimmering glints. Motes of dust floating in the air were clearly visible in the slanting light, adding texture to the scene. The entire scene, meticulously rendered with cinematic animation, exuded an exquisite magnificence and solemnity, perfectly showcasing her undeniable imperial demeanor.", + "seed": 6125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_8637_seed_6125_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01914897784590721, + "vlm_logprob": -0.01914897784590721, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01914897784590721, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018851052969694138 + ], + "opposite_logprob_tokens": [ + -4.268851280212402 + ], + "logprob_reward": -0.018851052969694138, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018851052969694138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26508304476737976, + 0.07284730672836304, + 0.07914827018976212, + 0.09724213182926178, + 0.15990474820137024, + 0.05867166072130203, + 0.06030197814106941, + 0.050255678594112396, + 0.07579497247934341, + 0.08075013756752014 + ], + "credit_map": [ + 0.027495456859469414, + 0.021062789484858513, + 0.05502568930387497, + 0.03511759638786316, + 0.027531221508979797, + 0.026546532288193703, + 0.04646146669983864, + 0.025842303410172462, + 0.007980987429618835, + 0.006088640540838242, + 0.014407048001885414, + 0.009308628737926483, + 0.007616315968334675, + 0.008758356794714928, + 0.00712078670039773, + 0.011566540226340294, + 0.0072690327651798725, + 0.008839240297675133, + 0.009586002677679062, + 0.009557734243571758, + 0.006537212524563074, + 0.009071738459169865, + 0.00625316146761179, + 0.022034145891666412, + 0.008137647993862629, + 0.019753441214561462, + 0.013854644261300564, + 0.019088150933384895, + 0.009088819846510887, + 0.01065001729875803, + 0.008151520043611526, + 0.008517892099916935, + 0.02371782436966896, + 0.02599966898560524, + 0.023307504132390022, + 0.01930003985762596, + 0.014647198840975761, + 0.012315595522522926, + 0.025325793772935867, + 0.015291138552129269, + 0.0061910212971270084, + 0.006141112186014652, + 0.007299813907593489, + 0.00919173751026392, + 0.007764913607388735, + 0.006341451313346624, + 0.010114781558513641, + 0.0056268335320055485, + 0.005524323787540197, + 0.008183623664081097, + 0.00862706545740366, + 0.011704071424901485, + 0.005071864929050207, + 0.004251838196069002, + 0.011882200837135315, + 0.0050569926388561726, + 0.004566750023514032, + 0.008229829370975494, + 0.008314100094139576, + 0.0056920843198895454, + 0.005784564185887575, + 0.005415687337517738, + 0.003964797127991915, + 0.008287861943244934, + 0.00791436992585659, + 0.009745700284838676, + 0.008877616375684738, + 0.009100792929530144, + 0.009586993604898453, + 0.010807156562805176, + 0.008942839689552784, + 0.010819503106176853, + 0.012879626825451851, + 0.008555635809898376, + 0.010676449164748192, + 0.010713196359574795, + 0.008371293544769287, + 0.009718586690723896, + 0.011909308843314648, + 0.007926044054329395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.018851052969694138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088770814239979 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088770814239979, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088770814239979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.251045823097229, + 0.08947048336267471, + 0.08589088171720505, + 0.10909803211688995, + 0.14844609797000885, + 0.0718778446316719, + 0.03927949070930481, + 0.0305902399122715, + 0.06473171710968018, + 0.10956931114196777 + ], + "credit_map": [ + 0.01911240629851818, + 0.018625307828187943, + 0.03414616361260414, + 0.03330674767494202, + 0.034020788967609406, + 0.03431199491024017, + 0.04927188530564308, + 0.028250571340322495, + 0.006871730554848909, + 0.012568727135658264, + 0.01618840917944908, + 0.011717027053236961, + 0.010297175496816635, + 0.0096210278570652, + 0.011814087629318237, + 0.010392313823103905, + 0.006650631781667471, + 0.011176001280546188, + 0.011459626257419586, + 0.010584416799247265, + 0.009178857319056988, + 0.010516026988625526, + 0.009139393456280231, + 0.01718594692647457, + 0.0079238535836339, + 0.01675643026828766, + 0.016075480729341507, + 0.013866007328033447, + 0.011401779018342495, + 0.012083965353667736, + 0.02063707634806633, + 0.010353460907936096, + 0.019236551597714424, + 0.013441386632621288, + 0.02345016784965992, + 0.01587231270968914, + 0.016937563195824623, + 0.013587737455964088, + 0.0314149409532547, + 0.014505472965538502, + 0.008431271649897099, + 0.00659040454775095, + 0.009286821819841862, + 0.010954098775982857, + 0.010773973539471626, + 0.006897635292261839, + 0.011827466078102589, + 0.00711618410423398, + 0.007072107866406441, + 0.006657711695879698, + 0.005215820856392384, + 0.005631425883620977, + 0.0029511358588933945, + 0.0026924158446490765, + 0.0058412104845047, + 0.0032176710665225983, + 0.002525629475712776, + 0.0040481239557266235, + 0.005923958495259285, + 0.004005872178822756, + 0.002888060174882412, + 0.00451418524608016, + 0.003061346709728241, + 0.00362306903116405, + 0.005954813677817583, + 0.00706688966602087, + 0.00751387607306242, + 0.006422680336982012, + 0.010081390850245953, + 0.009330499917268753, + 0.0072305649518966675, + 0.011131019331514835, + 0.022497134283185005, + 0.01233023963868618, + 0.015437478199601173, + 0.00613027298822999, + 0.011796511709690094, + 0.014489376917481422, + 0.008024392649531364, + 0.01886393129825592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.011088770814239979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533704750239849 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533704750239849, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533704750239849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16851067543029785, + 0.07519547641277313, + 0.09415365755558014, + 0.07646319270133972, + 0.1699400246143341, + 0.060398202389478683, + 0.07490991801023483, + 0.1739942878484726, + 0.06843889504671097, + 0.03799569979310036 + ], + "credit_map": [ + 0.020649712532758713, + 0.026877859607338905, + 0.025635957717895508, + 0.021619660779833794, + 0.014069621451199055, + 0.019716760143637657, + 0.02081175707280636, + 0.019129354506731033, + 0.010700873099267483, + 0.014730044640600681, + 0.01081999484449625, + 0.010480216704308987, + 0.005850807763636112, + 0.0068096984177827835, + 0.0063878558576107025, + 0.009415989741683006, + 0.005901204887777567, + 0.012887537479400635, + 0.016020920127630234, + 0.013816739432513714, + 0.009360425174236298, + 0.013456640765070915, + 0.009283582679927349, + 0.013426602818071842, + 0.006729993969202042, + 0.014339444227516651, + 0.011048880405724049, + 0.01560061052441597, + 0.007994157262146473, + 0.005018685944378376, + 0.009489088319242, + 0.006242331117391586, + 0.0348932147026062, + 0.02232118509709835, + 0.026412220671772957, + 0.017250724136829376, + 0.014478127472102642, + 0.013879228383302689, + 0.02612672746181488, + 0.014578594826161861, + 0.007743956055492163, + 0.004865563008934259, + 0.010240321047604084, + 0.006033674348145723, + 0.006158806849271059, + 0.006258622743189335, + 0.00890630204230547, + 0.010190953500568867, + 0.013130243867635727, + 0.015828518196940422, + 0.012565432116389275, + 0.010531075298786163, + 0.006353270728141069, + 0.0062010278925299644, + 0.004984825849533081, + 0.005315522663295269, + 0.016422579064965248, + 0.02049860917031765, + 0.026768658310174942, + 0.020021749660372734, + 0.02120152860879898, + 0.022575750946998596, + 0.01739981211721897, + 0.02910560742020607, + 0.005861197132617235, + 0.008931610733270645, + 0.007652857806533575, + 0.007034201640635729, + 0.012068629264831543, + 0.0095566026866436, + 0.006935339421033859, + 0.01039845496416092, + 0.006096506025642157, + 0.0035652057267725468, + 0.006032832432538271, + 0.003148101270198822, + 0.004994161427021027, + 0.004970789887011051, + 0.004668075125664473, + 0.004520031623542309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.012533704750239849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048633091151714325 + ], + "opposite_logprob_tokens": [ + -3.048633098602295 + ], + "logprob_reward": -0.048633091151714325, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048633091151714325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05571284890174866, + 0.0747487023472786, + 0.05629887804389, + 0.09756182134151459, + 0.1510481834411621, + 0.07337578386068344, + 0.08278241753578186, + 0.21955876052379608, + 0.12993890047073364, + 0.05897372588515282 + ], + "credit_map": [ + 0.006253052037209272, + 0.005496102850884199, + 0.011544850654900074, + 0.007104015909135342, + 0.004710179753601551, + 0.00839901715517044, + 0.0053815096616744995, + 0.006824116222560406, + 0.0079274270683527, + 0.010797153227031231, + 0.014626233838498592, + 0.0113737303763628, + 0.006979618687182665, + 0.00789016205817461, + 0.007872404530644417, + 0.007281970698386431, + 0.005464318208396435, + 0.008204983547329903, + 0.008554106578230858, + 0.0063799587078392506, + 0.0061012692749500275, + 0.006724770646542311, + 0.0056453305296599865, + 0.00922413356602192, + 0.012266930192708969, + 0.016012001782655716, + 0.014936106279492378, + 0.012408890761435032, + 0.007982744835317135, + 0.008623432368040085, + 0.014036073349416256, + 0.011295631527900696, + 0.03822816163301468, + 0.021087899804115295, + 0.024867508560419083, + 0.014751658774912357, + 0.011912589892745018, + 0.01553995069116354, + 0.013442333787679672, + 0.0112180570140481, + 0.009432266466319561, + 0.007309487089514732, + 0.009468227624893188, + 0.007590714376419783, + 0.010491089895367622, + 0.00586330471560359, + 0.01215703971683979, + 0.011063642799854279, + 0.008332578465342522, + 0.014151622541248798, + 0.01041404902935028, + 0.011301126331090927, + 0.010116716846823692, + 0.006689024623483419, + 0.011354885995388031, + 0.01042240671813488, + 0.020847629755735397, + 0.030972789973020554, + 0.034360844641923904, + 0.03161592781543732, + 0.020596256479620934, + 0.025903776288032532, + 0.027870584279298782, + 0.027390915900468826, + 0.013260804116725922, + 0.014731676317751408, + 0.015285857953131199, + 0.011228618212044239, + 0.01728879287838936, + 0.015071979723870754, + 0.013071409426629543, + 0.029999740421772003, + 0.01049607153981924, + 0.005925382487475872, + 0.008638286963105202, + 0.0033074291422963142, + 0.008036416955292225, + 0.007419185247272253, + 0.006553569808602333, + 0.008597375825047493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.048633091151714325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638269543647766 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638269543647766, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638269543647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10556437820196152, + 0.059048257768154144, + 0.16925682127475739, + 0.09933967143297195, + 0.13051050901412964, + 0.20363688468933105, + 0.03666916489601135, + 0.05235638841986656, + 0.05252891033887863, + 0.09108901768922806 + ], + "credit_map": [ + 0.010164576582610607, + 0.009907142259180546, + 0.016315268352627754, + 0.012143180705606937, + 0.015361838042736053, + 0.011724110692739487, + 0.019505620002746582, + 0.010442632250487804, + 0.0040609161369502544, + 0.0074821156449615955, + 0.008768469095230103, + 0.006764610297977924, + 0.006693268660455942, + 0.0068146358244121075, + 0.00952771957963705, + 0.008936512283980846, + 0.007361245341598988, + 0.020012835040688515, + 0.021210549399256706, + 0.029191583395004272, + 0.018003197386860847, + 0.016731588169932365, + 0.02846662700176239, + 0.028279203921556473, + 0.011578058823943138, + 0.011237208731472492, + 0.012792852707207203, + 0.007541916333138943, + 0.008115662261843681, + 0.015788421034812927, + 0.010418005287647247, + 0.021867536008358, + 0.017954709008336067, + 0.01219917368143797, + 0.021360531449317932, + 0.014494325965642929, + 0.014915098436176777, + 0.01363817136734724, + 0.024858132004737854, + 0.01109036523848772, + 0.020832428708672523, + 0.018957413733005524, + 0.01922389678657055, + 0.0345267727971077, + 0.028654463589191437, + 0.02066841535270214, + 0.03606603667140007, + 0.024707462638616562, + 0.003940064925700426, + 0.00537953432649374, + 0.004196743946522474, + 0.005326465237885714, + 0.0035876897163689137, + 0.0025918364990502596, + 0.007297903764992952, + 0.004348925314843655, + 0.0046755475923419, + 0.007534853182733059, + 0.006939044687896967, + 0.008330787532031536, + 0.006832458544522524, + 0.006047227419912815, + 0.006869762670248747, + 0.005126700270920992, + 0.00530024990439415, + 0.00457300478592515, + 0.0077684116549789906, + 0.005224315915256739, + 0.006534876301884651, + 0.006169171072542667, + 0.005843570921570063, + 0.011115306988358498, + 0.015645289793610573, + 0.0090775853022933, + 0.01391078531742096, + 0.007425948977470398, + 0.006617587525397539, + 0.01463960949331522, + 0.010961893014609814, + 0.012810306623578072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.004638269543647766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "full-body shot, shot from a slightly elevated perspective, the camera slowly pans rightward, following the subject in a solemn and steady rhythm, capturing the empress's majestic demeanor. Her hair, crowned by a towering, magnificent bun, is adorned with an intricately carved golden phoenix hairpin, which sways gently with her movements, its tassels cascading down her temples. She wears an intricately crafted crimson silk phoenix robe. Its wide sleeves and long, floor-length skirt are embroidered with countless strands of gold thread, creating a vivid image of a phoenix ready to take flight. Each feather gleams with a metallic sheen. She walks solemnly down a grand and profound palace corridor, flanked by massive vermilion columns adorned with carved and painted beams. The floor is paved with a mirror-like, black marble that reflects a hazy figure. Her steps were steady and deliberate, each step sending her voluminous skirt undulating gracefully across the ground like a crimson wave. The delicate phoenix embroidery on her robe shone brilliantly in the dappled light filtering through a carved window pane on one side, while the gemstones and pearls adorned it with shimmering glints. Motes of dust floating in the air were clearly visible in the slanting light, adding texture to the scene. The entire scene, meticulously rendered with cinematic animation, exuded an exquisite magnificence and solemnity, perfectly showcasing her undeniable imperial demeanor.", + "seed": 6126, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_8637_seed_6126_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0172818741761148, + "vlm_logprob": -0.0172818741761148, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.0172818741761148, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010355550795793533 + ], + "opposite_logprob_tokens": [ + -4.760355472564697 + ], + "logprob_reward": -0.010355550795793533, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010355550795793533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09694885462522507, + 0.03527668118476868, + 0.11726392805576324, + 0.06148125231266022, + 0.08116120100021362, + 0.28498414158821106, + 0.056848134845495224, + 0.09284799546003342, + 0.03105652891099453, + 0.1421312838792801 + ], + "credit_map": [ + 0.013485065661370754, + 0.014754209667444229, + 0.012780120596289635, + 0.016529113054275513, + 0.010140543803572655, + 0.00957863125950098, + 0.011943245306611061, + 0.007737911771982908, + 0.004728348460048437, + 0.004594322293996811, + 0.0035473022144287825, + 0.005296847317367792, + 0.003246811218559742, + 0.004845284856855869, + 0.005031817592680454, + 0.003985944669693708, + 0.02927008830010891, + 0.003437537932768464, + 0.00843803957104683, + 0.010115193203091621, + 0.0036172077525407076, + 0.03969155624508858, + 0.003933988511562347, + 0.018760308623313904, + 0.010112442076206207, + 0.0054898010566830635, + 0.007558355573564768, + 0.0065122125670313835, + 0.007740189787000418, + 0.00759776821359992, + 0.008302275091409683, + 0.008168202824890614, + 0.004380004480481148, + 0.022345395758748055, + 0.005069495644420385, + 0.00930819008499384, + 0.0066678388975560665, + 0.02156599424779415, + 0.007339105475693941, + 0.004485169891268015, + 0.030308565124869347, + 0.047199442982673645, + 0.04078184440732002, + 0.05179128423333168, + 0.02800552174448967, + 0.020660720765590668, + 0.043749235570430756, + 0.02248750999569893, + 0.009219822473824024, + 0.010726802051067352, + 0.006672754883766174, + 0.00851354282349348, + 0.006798182614147663, + 0.0036240036133676767, + 0.007212347351014614, + 0.004080669954419136, + 0.028757311403751373, + 0.004968379624187946, + 0.009829243645071983, + 0.005530242342501879, + 0.003216368844732642, + 0.028699535876512527, + 0.005679239518940449, + 0.006167664658278227, + 0.003728628158569336, + 0.0034283355344086885, + 0.0038042354863137007, + 0.004520528484135866, + 0.006127931177616119, + 0.0027151829563081264, + 0.0036245081573724747, + 0.003107176162302494, + 0.010501752607524395, + 0.05299022048711777, + 0.03919127956032753, + 0.009156069718301296, + 0.006441122852265835, + 0.013828117400407791, + 0.0037633441388607025, + 0.006259358488023281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.010355550795793533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101080819964409 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101080819964409, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101080819964409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09306159615516663, + 0.057424936443567276, + 0.050684891641139984, + 0.05410539358854294, + 0.09411070495843887, + 0.05011202022433281, + 0.05574607104063034, + 0.4098633825778961, + 0.04801531136035919, + 0.08687570691108704 + ], + "credit_map": [ + 0.010605553165078163, + 0.014091146178543568, + 0.008979293517768383, + 0.01504146121442318, + 0.014779268763959408, + 0.007189156021922827, + 0.014761232770979404, + 0.0076144696213305, + 0.007594319060444832, + 0.00664288317784667, + 0.006169695872813463, + 0.007756283972412348, + 0.00607990100979805, + 0.007614954840391874, + 0.009285538457334042, + 0.0062813605181872845, + 0.00967336818575859, + 0.0029471286106854677, + 0.006043440196663141, + 0.008966702036559582, + 0.006261799018830061, + 0.006969343405216932, + 0.004605024587363005, + 0.005218077916651964, + 0.008666045032441616, + 0.005447381641715765, + 0.007229895330965519, + 0.0075328839011490345, + 0.006617728155106306, + 0.0075955139473080635, + 0.00570666091516614, + 0.005309276282787323, + 0.005606836173683405, + 0.016423895955085754, + 0.00591433048248291, + 0.008629970252513885, + 0.02090064063668251, + 0.019034158438444138, + 0.010820737108588219, + 0.00678013451397419, + 0.0041218409314751625, + 0.008890524506568909, + 0.006641737185418606, + 0.00784394796937704, + 0.006553346291184425, + 0.0024316529743373394, + 0.00921802967786789, + 0.004410936497151852, + 0.006206281017512083, + 0.009779680520296097, + 0.006133911199867725, + 0.007731813006103039, + 0.0058993143029510975, + 0.006517816800624132, + 0.005831248126924038, + 0.007646000944077969, + 0.13494576513767242, + 0.026585334911942482, + 0.03048478439450264, + 0.0221449863165617, + 0.016198555007576942, + 0.13762320578098297, + 0.02291812188923359, + 0.01896260865032673, + 0.005655589047819376, + 0.00717903720214963, + 0.004762706346809864, + 0.005429769400507212, + 0.006994687020778656, + 0.00638580834493041, + 0.006363455671817064, + 0.005244249477982521, + 0.006764180492609739, + 0.024971896782517433, + 0.019302982836961746, + 0.007066374644637108, + 0.009290821850299835, + 0.0092741958796978, + 0.005129961296916008, + 0.005075288470834494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.004101080819964409 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029781080782413483 + ], + "opposite_logprob_tokens": [ + -3.5297811031341553 + ], + "logprob_reward": -0.029781080782413483, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029781080782413483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16403816640377045, + 0.038431569933891296, + 0.12841780483722687, + 0.06349547207355499, + 0.043435994535684586, + 0.08546025305986404, + 0.04741435870528221, + 0.27954402565956116, + 0.08832372725009918, + 0.06143859028816223 + ], + "credit_map": [ + 0.019506249576807022, + 0.030759762972593307, + 0.02454080432653427, + 0.03450901433825493, + 0.011210113763809204, + 0.01137036457657814, + 0.017102692276239395, + 0.015039163641631603, + 0.006033594254404306, + 0.004670286551117897, + 0.004158227238804102, + 0.005521133542060852, + 0.003675664309412241, + 0.004555064719170332, + 0.005006582010537386, + 0.004811019171029329, + 0.027492724359035492, + 0.006820394657552242, + 0.020073724910616875, + 0.024121835827827454, + 0.005790960043668747, + 0.02305181510746479, + 0.00895656831562519, + 0.0121097881346941, + 0.008628838695585728, + 0.005782719701528549, + 0.013026345521211624, + 0.007266807369887829, + 0.006666310131549835, + 0.008663373999297619, + 0.00721239997074008, + 0.006248680874705315, + 0.0034258514642715454, + 0.007317210081964731, + 0.004947689361870289, + 0.005332895088940859, + 0.00948786549270153, + 0.003648327197879553, + 0.00507662259042263, + 0.004199531860649586, + 0.010720492340624332, + 0.01237537432461977, + 0.011178547516465187, + 0.015547161921858788, + 0.01177707314491272, + 0.004643021151423454, + 0.011158362030982971, + 0.008060219697654247, + 0.008100944571197033, + 0.006021092180162668, + 0.005432192236185074, + 0.006230065133422613, + 0.006880426779389381, + 0.003436401719227433, + 0.004954131785780191, + 0.006359104998409748, + 0.0861860066652298, + 0.01504364050924778, + 0.0283298771828413, + 0.021661635488271713, + 0.008392652496695518, + 0.08907634019851685, + 0.015072396025061607, + 0.015781471505761147, + 0.010314998216927052, + 0.019979069009423256, + 0.016668740659952164, + 0.008187755011022091, + 0.009090574458241463, + 0.010362111032009125, + 0.007518388330936432, + 0.006202093791216612, + 0.007533501833677292, + 0.010007090866565704, + 0.009283216670155525, + 0.008183751255273819, + 0.005017094314098358, + 0.012388119474053383, + 0.0042805965058505535, + 0.0047452207654714584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.029781080782413483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808043152093887 + ], + "opposite_logprob_tokens": [ + -3.2880804538726807 + ], + "logprob_reward": -0.03808043152093887, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808043152093887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04261241480708122, + 0.01862972415983677, + 0.0580221563577652, + 0.02654344215989113, + 0.22201094031333923, + 0.5402441620826721, + 0.01809852570295334, + 0.035080939531326294, + 0.013677346520125866, + 0.025080319494009018 + ], + "credit_map": [ + 0.004648187197744846, + 0.010314756073057652, + 0.0028720039408653975, + 0.00518522085621953, + 0.004424461163580418, + 0.0032993322238326073, + 0.006561185698956251, + 0.005307272542268038, + 0.0023859154898673296, + 0.0015168852405622602, + 0.0037567708641290665, + 0.0027697726618498564, + 0.001578221214003861, + 0.0015712457243353128, + 0.0030276437755674124, + 0.0020232689566910267, + 0.011113972403109074, + 0.001841462217271328, + 0.003918129485100508, + 0.006937739904969931, + 0.003479958977550268, + 0.017940452322363853, + 0.003314491594210267, + 0.009475955739617348, + 0.0033459493424743414, + 0.002317593665793538, + 0.003183887340128422, + 0.002773925196379423, + 0.004998016636818647, + 0.0037352945655584335, + 0.004076533019542694, + 0.002112243790179491, + 0.010380636900663376, + 0.06092597544193268, + 0.015054906718432903, + 0.025070251896977425, + 0.026180757209658623, + 0.054856982082128525, + 0.01800079643726349, + 0.011540635488927364, + 0.059883590787649155, + 0.08544684946537018, + 0.07833318412303925, + 0.09922520816326141, + 0.056210700422525406, + 0.04037286341190338, + 0.07845724374055862, + 0.04231453686952591, + 0.0014297731686383486, + 0.0031169038265943527, + 0.001700677559711039, + 0.0020676127169281244, + 0.0027311285957694054, + 0.0019699118565768003, + 0.0023341658525168896, + 0.002748352475464344, + 0.007757381536066532, + 0.0021540652960538864, + 0.006274840794503689, + 0.004359770100563765, + 0.002128933323547244, + 0.0075482758693397045, + 0.002856184495612979, + 0.0020014923065900803, + 0.0018446826143190265, + 0.002276951214298606, + 0.001788690104149282, + 0.0013234398793429136, + 0.0016871836269274354, + 0.001200348837301135, + 0.002012367360293865, + 0.001543683698400855, + 0.0020556719973683357, + 0.009009958244860172, + 0.0061561125330626965, + 0.001574631780385971, + 0.0014617613051086664, + 0.0020546552259474993, + 0.0013413899578154087, + 0.0014261401956900954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.03808043152093887 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409122696146369 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.00409122696146369, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409122696146369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07579442113637924, + 0.022963913157582283, + 0.10861844569444656, + 0.058364953845739365, + 0.03151446208357811, + 0.27585452795028687, + 0.16291451454162598, + 0.09770536422729492, + 0.0443333275616169, + 0.12193606793880463 + ], + "credit_map": [ + 0.00826907530426979, + 0.011755920015275478, + 0.008284855633974075, + 0.01415967009961605, + 0.006949378177523613, + 0.006498313043266535, + 0.01095083262771368, + 0.00892638135701418, + 0.004387718625366688, + 0.0027622885536402464, + 0.004785294644534588, + 0.002745034173130989, + 0.0014536267844960093, + 0.0020503257401287556, + 0.0022775756660848856, + 0.0025020490866154432, + 0.023991180583834648, + 0.0036787372082471848, + 0.008188064210116863, + 0.01553617138415575, + 0.005063626449555159, + 0.029612809419631958, + 0.0072333416901528835, + 0.015314514748752117, + 0.008500506170094013, + 0.005431962199509144, + 0.007345298305153847, + 0.01150859147310257, + 0.007033377420157194, + 0.008244098164141178, + 0.005571227520704269, + 0.004729892127215862, + 0.0024630811531096697, + 0.004847620613873005, + 0.002251488622277975, + 0.0033474592491984367, + 0.007174970582127571, + 0.0034029490780085325, + 0.004734047222882509, + 0.003292848588898778, + 0.0350557416677475, + 0.04477463662624359, + 0.03672124817967415, + 0.05018702521920204, + 0.02754899114370346, + 0.01349545270204544, + 0.045545801520347595, + 0.022525604814291, + 0.02297540381550789, + 0.026767579838633537, + 0.01731565035879612, + 0.02183404751121998, + 0.015548618510365486, + 0.020510708913207054, + 0.016228413209319115, + 0.0217340886592865, + 0.010764437727630138, + 0.010444381274282932, + 0.019767699763178825, + 0.01353280059993267, + 0.006823939736932516, + 0.015957705676555634, + 0.012599486857652664, + 0.007814907468855381, + 0.0068524070084095, + 0.006389141082763672, + 0.0038594803772866726, + 0.00451489444822073, + 0.006830138154327869, + 0.004190006759017706, + 0.006158200558274984, + 0.005539058707654476, + 0.008974593132734299, + 0.04380662739276886, + 0.03351128101348877, + 0.007448884658515453, + 0.00854719989001751, + 0.008660877123475075, + 0.005215710029006004, + 0.005770895164459944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.00409122696146369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "full-body shot, shot from a slightly elevated perspective, the camera slowly pans rightward, following the subject in a solemn and steady rhythm, capturing the empress's majestic demeanor. Her hair, crowned by a towering, magnificent bun, is adorned with an intricately carved golden phoenix hairpin, which sways gently with her movements, its tassels cascading down her temples. She wears an intricately crafted crimson silk phoenix robe. Its wide sleeves and long, floor-length skirt are embroidered with countless strands of gold thread, creating a vivid image of a phoenix ready to take flight. Each feather gleams with a metallic sheen. She walks solemnly down a grand and profound palace corridor, flanked by massive vermilion columns adorned with carved and painted beams. The floor is paved with a mirror-like, black marble that reflects a hazy figure. Her steps were steady and deliberate, each step sending her voluminous skirt undulating gracefully across the ground like a crimson wave. The delicate phoenix embroidery on her robe shone brilliantly in the dappled light filtering through a carved window pane on one side, while the gemstones and pearls adorned it with shimmering glints. Motes of dust floating in the air were clearly visible in the slanting light, adding texture to the scene. The entire scene, meticulously rendered with cinematic animation, exuded an exquisite magnificence and solemnity, perfectly showcasing her undeniable imperial demeanor.", + "seed": 6127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_8637_seed_6127_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030637554358690977, + "vlm_logprob": -0.030637554358690977, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.030637554358690977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011842658743262291 + ], + "opposite_logprob_tokens": [ + -4.761842727661133 + ], + "logprob_reward": -0.011842658743262291, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011842658743262291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13836480677127838, + 0.0320189967751503, + 0.03921234980225563, + 0.0889848917722702, + 0.04464953392744064, + 0.06076396256685257, + 0.28184249997138977, + 0.11258640140295029, + 0.11704591661691666, + 0.08453059196472168 + ], + "credit_map": [ + 0.012540287338197231, + 0.015831543132662773, + 0.022511910647153854, + 0.0141600975766778, + 0.019503014162182808, + 0.020263558253645897, + 0.015754397958517075, + 0.017800016328692436, + 0.002664524130523205, + 0.003527903463691473, + 0.003039883216843009, + 0.003732563927769661, + 0.005869622807949781, + 0.004170160740613937, + 0.005521194078028202, + 0.0034931476693600416, + 0.005367952864617109, + 0.006999342702329159, + 0.0033713991288095713, + 0.004657987039536238, + 0.005608434788882732, + 0.0065627931617200375, + 0.003150327131152153, + 0.0034941178746521473, + 0.01058950088918209, + 0.02292400784790516, + 0.0081888223066926, + 0.009545392356812954, + 0.01000706572085619, + 0.00996062345802784, + 0.006103137973695993, + 0.011666346341371536, + 0.005278389900922775, + 0.006903346627950668, + 0.0065404693596065044, + 0.004995855037122965, + 0.006240918301045895, + 0.00478290906175971, + 0.004935407545417547, + 0.0049722399562597275, + 0.006118581630289555, + 0.00890057161450386, + 0.006835203152149916, + 0.005708182230591774, + 0.011876087635755539, + 0.009201312437653542, + 0.004554911516606808, + 0.0075691197998821735, + 0.05891336128115654, + 0.023959971964359283, + 0.03331250324845314, + 0.01690293848514557, + 0.017929287627339363, + 0.07525808364152908, + 0.01460994966328144, + 0.040956418961286545, + 0.00799667276442051, + 0.006459671072661877, + 0.00807228498160839, + 0.0053638191893696785, + 0.006775922141969204, + 0.041039951145648956, + 0.004686338361352682, + 0.03219175711274147, + 0.011300512589514256, + 0.008416376076638699, + 0.00721511198207736, + 0.00602669408544898, + 0.010227578692138195, + 0.0294176135212183, + 0.009363500401377678, + 0.035078536719083786, + 0.011054627597332, + 0.01059420220553875, + 0.013897370547056198, + 0.013083292171359062, + 0.008527456782758236, + 0.00850387942045927, + 0.011453548446297646, + 0.007416230160742998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.011842658743262291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263399798423052 + ], + "opposite_logprob_tokens": [ + -5.255263328552246 + ], + "logprob_reward": -0.005263399798423052, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263399798423052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22597798705101013, + 0.10170531272888184, + 0.03558812290430069, + 0.0934104397892952, + 0.05881208926439285, + 0.03489777445793152, + 0.2543678879737854, + 0.06308560818433762, + 0.06805197894573212, + 0.06410285830497742 + ], + "credit_map": [ + 0.019821615889668465, + 0.03639129549264908, + 0.021209193393588066, + 0.022647453472018242, + 0.026596175506711006, + 0.042697206139564514, + 0.033347368240356445, + 0.023267686367034912, + 0.00878208689391613, + 0.015491034835577011, + 0.01117731910198927, + 0.009831472299993038, + 0.01496928557753563, + 0.011754469946026802, + 0.019235452637076378, + 0.010464192368090153, + 0.004441872239112854, + 0.004855672363191843, + 0.0038020277861505747, + 0.0036635054275393486, + 0.0053363884799182415, + 0.004326462745666504, + 0.004548621363937855, + 0.004613571334630251, + 0.01240250188857317, + 0.02011740580201149, + 0.011188159696757793, + 0.009996820241212845, + 0.011187305673956871, + 0.007881082594394684, + 0.009897066280245781, + 0.010740092024207115, + 0.006369172129780054, + 0.009286337532103062, + 0.007680477108806372, + 0.007412682753056288, + 0.007258068770170212, + 0.0077714864164590836, + 0.006899658590555191, + 0.0061342050321400166, + 0.005358757916837931, + 0.005035179201513529, + 0.006563591305166483, + 0.0031110530253499746, + 0.0037001827731728554, + 0.004669691901654005, + 0.002416999777778983, + 0.004042316693812609, + 0.05445118620991707, + 0.02121131867170334, + 0.026953360065817833, + 0.012162666767835617, + 0.018016792833805084, + 0.07074066996574402, + 0.010806775651872158, + 0.040025122463703156, + 0.00605871481820941, + 0.003191830823197961, + 0.006353994831442833, + 0.0027563832700252533, + 0.003137795953080058, + 0.019908802583813667, + 0.0028102328069508076, + 0.018867848441004753, + 0.008726282976567745, + 0.006513361353427172, + 0.007701524533331394, + 0.006100517697632313, + 0.013557625003159046, + 0.007405210752040148, + 0.012273517437279224, + 0.005773934535682201, + 0.01843787543475628, + 0.010300083085894585, + 0.013314641080796719, + 0.004232229199260473, + 0.003982552792876959, + 0.004278125241398811, + 0.003598109120503068, + 0.0059592449106276035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005263399798423052 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327572926878929 + ], + "opposite_logprob_tokens": [ + -3.77327561378479 + ], + "logprob_reward": -0.02327572926878929, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327572926878929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3492880165576935, + 0.13180553913116455, + 0.09004774689674377, + 0.07047173380851746, + 0.059709638357162476, + 0.06625188142061234, + 0.07503141462802887, + 0.059176553040742874, + 0.03493550047278404, + 0.06328193843364716 + ], + "credit_map": [ + 0.031124886125326157, + 0.02447664737701416, + 0.040720097720623016, + 0.02544189617037773, + 0.06061284616589546, + 0.04595590755343437, + 0.06943412125110626, + 0.051521606743335724, + 0.009982999414205551, + 0.007990581914782524, + 0.013849134556949139, + 0.005775876808911562, + 0.028192222118377686, + 0.02434493601322174, + 0.025552380830049515, + 0.016117407009005547, + 0.012804725207388401, + 0.006691140588372946, + 0.007430297788232565, + 0.0052026924677193165, + 0.018770039081573486, + 0.021440858021378517, + 0.008244689553976059, + 0.009463304653763771, + 0.00821441039443016, + 0.016748258844017982, + 0.006023649591952562, + 0.0066704400815069675, + 0.006223881151527166, + 0.008569332771003246, + 0.008079861290752888, + 0.009941901080310345, + 0.007834131829440594, + 0.0052835335955023766, + 0.01074772235006094, + 0.004595145117491484, + 0.007625369820743799, + 0.01040406059473753, + 0.005142316687852144, + 0.00807736162096262, + 0.009838308207690716, + 0.006804178934544325, + 0.011012888513505459, + 0.0044475579634308815, + 0.008889294229447842, + 0.011879455298185349, + 0.0038122516125440598, + 0.009567947126924992, + 0.010443774983286858, + 0.006185371428728104, + 0.01679503545165062, + 0.005574056878685951, + 0.005003907717764378, + 0.015876377001404762, + 0.004990985617041588, + 0.010161908343434334, + 0.007201565429568291, + 0.00566687248647213, + 0.006175356917083263, + 0.004393633920699358, + 0.0071525415405631065, + 0.01345015224069357, + 0.004253908526152372, + 0.010882521979510784, + 0.004561486188322306, + 0.0036224862560629845, + 0.0047159804962575436, + 0.0036515614483505487, + 0.0051919869147241116, + 0.0057733068242669106, + 0.005122342612594366, + 0.0022963492665439844, + 0.015830181539058685, + 0.008709507994353771, + 0.011314108036458492, + 0.005294116213917732, + 0.005897132214158773, + 0.005620237905532122, + 0.005107559729367495, + 0.005509090609848499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02327572926878929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027074813842773 + ], + "opposite_logprob_tokens": [ + -2.3502707481384277 + ], + "logprob_reward": -0.10027074813842773, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027074813842773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10528463870286942, + 0.10962575674057007, + 0.04839443415403366, + 0.057122327387332916, + 0.02730279602110386, + 0.05422240123152733, + 0.3730258643627167, + 0.11055926978588104, + 0.08167432248592377, + 0.032788150012493134 + ], + "credit_map": [ + 0.00918531697243452, + 0.009503454901278019, + 0.012874973937869072, + 0.00827401876449585, + 0.023314107209444046, + 0.01719452068209648, + 0.014790002256631851, + 0.010148248635232449, + 0.008076470345258713, + 0.011122687719762325, + 0.011633888818323612, + 0.011565517634153366, + 0.018055753782391548, + 0.017114626243710518, + 0.01994035579264164, + 0.012116458266973495, + 0.006212661974132061, + 0.00641420716419816, + 0.004875579383224249, + 0.004727199207991362, + 0.008012384176254272, + 0.006699460558593273, + 0.005841190926730633, + 0.0056117526255548, + 0.011054422706365585, + 0.00915526319295168, + 0.007780850399285555, + 0.004053643438965082, + 0.006758016534149647, + 0.006366136949509382, + 0.004774055909365416, + 0.007179941050708294, + 0.0031375756952911615, + 0.002629523631185293, + 0.005729962140321732, + 0.0020079019013792276, + 0.0028774833772331476, + 0.004507354460656643, + 0.0026894924230873585, + 0.0037235028576105833, + 0.006330615840852261, + 0.007415301166474819, + 0.00655805878341198, + 0.005381063092499971, + 0.008134987205266953, + 0.009260744787752628, + 0.004103860817849636, + 0.007037774659693241, + 0.07655974477529526, + 0.02475891262292862, + 0.055666886270046234, + 0.019133003428578377, + 0.02231754921376705, + 0.09900153428316116, + 0.017258692532777786, + 0.05832957476377487, + 0.014553312212228775, + 0.012519649229943752, + 0.020628882572054863, + 0.010958760045468807, + 0.014179891906678677, + 0.014690516516566277, + 0.009327059611678123, + 0.01370120607316494, + 0.012756875716149807, + 0.005339605733752251, + 0.013796362094581127, + 0.004749062471091747, + 0.008007382974028587, + 0.009759295731782913, + 0.006707005202770233, + 0.020558739081025124, + 0.003949970938265324, + 0.003417510772123933, + 0.004694982431828976, + 0.003937581554055214, + 0.004162879660725594, + 0.0033684824593365192, + 0.006352841854095459, + 0.0029039024375379086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.10027074813842773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535235844552517 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.012535235844552517, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535235844552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10277313739061356, + 0.051403023302555084, + 0.033038828521966934, + 0.04339040070772171, + 0.07759902626276016, + 0.04241194948554039, + 0.2835708260536194, + 0.19822683930397034, + 0.07414804399013519, + 0.09343790262937546 + ], + "credit_map": [ + 0.010107114911079407, + 0.010138013400137424, + 0.012372062541544437, + 0.010953704826533794, + 0.017921082675457, + 0.014347405172884464, + 0.012038814835250378, + 0.014894942753016949, + 0.003656171727925539, + 0.00530613400042057, + 0.00831065233796835, + 0.005955973640084267, + 0.007925303652882576, + 0.006842412985861301, + 0.007253443822264671, + 0.00615293113514781, + 0.00374366226606071, + 0.003399317152798176, + 0.0029220059514045715, + 0.003491332521662116, + 0.004297114908695221, + 0.0038687484338879585, + 0.0066407606936991215, + 0.004675887990742922, + 0.00521848164498806, + 0.006842541974037886, + 0.0065429131500422955, + 0.0031692013144493103, + 0.00581706315279007, + 0.004458065610378981, + 0.0061035435646772385, + 0.00523859029635787, + 0.00746171735227108, + 0.012363473884761333, + 0.012928476557135582, + 0.010334399528801441, + 0.009460761211812496, + 0.007551968563348055, + 0.008113398216664791, + 0.009384830482304096, + 0.006580311339348555, + 0.004250622354447842, + 0.007343553472310305, + 0.004267790820449591, + 0.004817403387278318, + 0.00688989507034421, + 0.002587990602478385, + 0.005674383603036404, + 0.058948032557964325, + 0.02585320733487606, + 0.02845592238008976, + 0.014555669389665127, + 0.020306967198848724, + 0.07934118062257767, + 0.01311132125556469, + 0.04299851506948471, + 0.013459017500281334, + 0.008321496658027172, + 0.008106401190161705, + 0.007175196893513203, + 0.012476449832320213, + 0.07580114901065826, + 0.007763576228171587, + 0.06512358039617538, + 0.007592778652906418, + 0.007708447054028511, + 0.007136371452361345, + 0.006619265768676996, + 0.016877124086022377, + 0.009029243141412735, + 0.01308757159858942, + 0.00609723711386323, + 0.017458222806453705, + 0.011232483200728893, + 0.01477044727653265, + 0.01099332794547081, + 0.010185247287154198, + 0.008252509869635105, + 0.010409143753349781, + 0.010136508382856846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.012535235844552517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "full-body shot, shot from a slightly elevated perspective, the camera slowly pans rightward, following the subject in a solemn and steady rhythm, capturing the empress's majestic demeanor. Her hair, crowned by a towering, magnificent bun, is adorned with an intricately carved golden phoenix hairpin, which sways gently with her movements, its tassels cascading down her temples. She wears an intricately crafted crimson silk phoenix robe. Its wide sleeves and long, floor-length skirt are embroidered with countless strands of gold thread, creating a vivid image of a phoenix ready to take flight. Each feather gleams with a metallic sheen. She walks solemnly down a grand and profound palace corridor, flanked by massive vermilion columns adorned with carved and painted beams. The floor is paved with a mirror-like, black marble that reflects a hazy figure. Her steps were steady and deliberate, each step sending her voluminous skirt undulating gracefully across the ground like a crimson wave. The delicate phoenix embroidery on her robe shone brilliantly in the dappled light filtering through a carved window pane on one side, while the gemstones and pearls adorned it with shimmering glints. Motes of dust floating in the air were clearly visible in the slanting light, adding texture to the scene. The entire scene, meticulously rendered with cinematic animation, exuded an exquisite magnificence and solemnity, perfectly showcasing her undeniable imperial demeanor.", + "seed": 6128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_10/index_8637_seed_6128_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02622615620493889, + "vlm_logprob": -0.02622615620493889, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02622615620493889, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0166330523788929 + ], + "opposite_logprob_tokens": [ + -4.391633033752441 + ], + "logprob_reward": -0.0166330523788929, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0166330523788929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2656964361667633, + 0.09496265649795532, + 0.07544867694377899, + 0.06752055883407593, + 0.11689464747905731, + 0.04860866442322731, + 0.14317357540130615, + 0.0708831176161766, + 0.06432489305734634, + 0.05248676612973213 + ], + "credit_map": [ + 0.04526294022798538, + 0.017559394240379333, + 0.038085319101810455, + 0.020437084138393402, + 0.024332629516720772, + 0.02496512420475483, + 0.03475484997034073, + 0.060299102216959, + 0.009077046997845173, + 0.009972155094146729, + 0.011005655862390995, + 0.010202155448496342, + 0.01072630938142538, + 0.014309859834611416, + 0.010766759514808655, + 0.018902717158198357, + 0.004957075696438551, + 0.006207024212926626, + 0.010527750477194786, + 0.0071668317541480064, + 0.004385329317301512, + 0.012626596726477146, + 0.005847523920238018, + 0.023730548098683357, + 0.005009595770388842, + 0.0061897095292806625, + 0.0081125283613801, + 0.006739677861332893, + 0.012189229018986225, + 0.010309095494449139, + 0.01064460352063179, + 0.008326124399900436, + 0.018168600276112556, + 0.008349727839231491, + 0.0169939287006855, + 0.010297352448105812, + 0.012021300382912159, + 0.018656592816114426, + 0.009658175520598888, + 0.022748978808522224, + 0.0029855782631784678, + 0.003847395069897175, + 0.005649044178426266, + 0.00522981584072113, + 0.005767187103629112, + 0.010368690825998783, + 0.007350703235715628, + 0.007410252466797829, + 0.013753758743405342, + 0.0075362324714660645, + 0.01175065990537405, + 0.008056201972067356, + 0.022866182029247284, + 0.033422596752643585, + 0.02219339646399021, + 0.023594563826918602, + 0.004963018000125885, + 0.011662389151751995, + 0.008199364878237247, + 0.00959801860153675, + 0.012626677751541138, + 0.008918109349906445, + 0.009686709381639957, + 0.005228831432759762, + 0.005887606181204319, + 0.004863245878368616, + 0.005926900543272495, + 0.005429981276392937, + 0.007929714396595955, + 0.016106953844428062, + 0.005161341745406389, + 0.013019155710935593, + 0.008331072516739368, + 0.005355103872716427, + 0.010669081471860409, + 0.009540812112390995, + 0.004702155012637377, + 0.0038402345962822437, + 0.006125263869762421, + 0.003923045005649328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0166330523788929 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087238788604736 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087238788604736, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087238788604736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11512608826160431, + 0.036356233060359955, + 0.052525512874126434, + 0.03338067606091499, + 0.04848547279834747, + 0.05023771524429321, + 0.08203154057264328, + 0.015905257314443588, + 0.08676308393478394, + 0.4791884422302246 + ], + "credit_map": [ + 0.013209815137088299, + 0.010241949930787086, + 0.01392888929694891, + 0.010226032696664333, + 0.012014939449727535, + 0.01010136678814888, + 0.014917481690645218, + 0.0304856114089489, + 0.003260158933699131, + 0.0031257339287549257, + 0.0048142410814762115, + 0.003096404019743204, + 0.00218178890645504, + 0.006335955113172531, + 0.003353857435286045, + 0.010188090614974499, + 0.003624270437285304, + 0.0041743856854736805, + 0.006997006945312023, + 0.003844793187454343, + 0.004237452521920204, + 0.00882533099502325, + 0.0036686554085463285, + 0.01715361885726452, + 0.002039214363321662, + 0.002833580831065774, + 0.003441344015300274, + 0.0029625031165778637, + 0.007024776190519333, + 0.005063630640506744, + 0.005579203832894564, + 0.0044364226050674915, + 0.0030312587041407824, + 0.0024512324016541243, + 0.006440051831305027, + 0.004131675232201815, + 0.00573024433106184, + 0.012496785260736942, + 0.005437958054244518, + 0.008766268379986286, + 0.006399326026439667, + 0.002704075537621975, + 0.004963523708283901, + 0.003447472583502531, + 0.005491076968610287, + 0.013855621218681335, + 0.007084560580551624, + 0.006292056757956743, + 0.0057608336210250854, + 0.0036855496000498533, + 0.011138766072690487, + 0.004779517184942961, + 0.029761821031570435, + 0.014772708527743816, + 0.0073343017138540745, + 0.004798044916242361, + 0.0015153629938140512, + 0.0020989272743463516, + 0.0019950279965996742, + 0.001481629558838904, + 0.0023601921275258064, + 0.0018949037184938788, + 0.0018333164043724537, + 0.0027258973568677902, + 0.01329977996647358, + 0.007434045430272818, + 0.0082070492208004, + 0.004951754584908485, + 0.008884944021701813, + 0.017584292218089104, + 0.007102613337337971, + 0.01929861307144165, + 0.07603040337562561, + 0.045906804502010345, + 0.06446710973978043, + 0.10108429938554764, + 0.036221105605363846, + 0.0502891018986702, + 0.07363081723451614, + 0.03155878931283951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.011087238788604736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977538712322712 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.00977538712322712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977538712322712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17517367005348206, + 0.04907841235399246, + 0.043001823127269745, + 0.04364080727100372, + 0.20395036041736603, + 0.07382021099328995, + 0.059354931116104126, + 0.013892164453864098, + 0.057125531136989594, + 0.2809620201587677 + ], + "credit_map": [ + 0.02203121781349182, + 0.011913206428289413, + 0.023876069113612175, + 0.011782070621848106, + 0.01399217452853918, + 0.023427898064255714, + 0.02446632646024227, + 0.043684713542461395, + 0.004951945971697569, + 0.0028174719773232937, + 0.008679444901645184, + 0.004762708209455013, + 0.003205029061064124, + 0.008481735363602638, + 0.004564255941659212, + 0.01161582674831152, + 0.00365865183994174, + 0.0022023976780474186, + 0.006306098774075508, + 0.0029234245885163546, + 0.002753989538177848, + 0.007061951793730259, + 0.003668396268039942, + 0.014426912181079388, + 0.00336967664770782, + 0.003398746717721224, + 0.004147352185100317, + 0.00254230503924191, + 0.0070066675543785095, + 0.006661964580416679, + 0.00643226457759738, + 0.01008183229714632, + 0.03718547523021698, + 0.010800470598042011, + 0.029243338853120804, + 0.010269121266901493, + 0.019205769523978233, + 0.036421604454517365, + 0.01783107779920101, + 0.042993515729904175, + 0.004548782482743263, + 0.002483710879459977, + 0.0061570690013468266, + 0.005308779887855053, + 0.007712662685662508, + 0.027314629405736923, + 0.004537779372185469, + 0.015756800770759583, + 0.005292349494993687, + 0.002841606270521879, + 0.010330278426408768, + 0.0028257884550839663, + 0.016736676916480064, + 0.009361841715872288, + 0.005275483708828688, + 0.006690912880003452, + 0.001531289773993194, + 0.0013004851061850786, + 0.0021519556175917387, + 0.0013502129586413503, + 0.0018074839608743787, + 0.0015005609020590782, + 0.0019339623395353556, + 0.0023162150755524635, + 0.007330418098717928, + 0.005451064091175795, + 0.008093652315437794, + 0.004433412570506334, + 0.01036401093006134, + 0.005696664564311504, + 0.00543182622641325, + 0.010324484668672085, + 0.0441751703619957, + 0.026890844106674194, + 0.038003623485565186, + 0.05895717069506645, + 0.02096528746187687, + 0.029569482430815697, + 0.04339360073208809, + 0.019006852060556412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.00977538712322712 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08899801969528198 + ], + "opposite_logprob_tokens": [ + -2.4639980792999268 + ], + "logprob_reward": -0.08899801969528198, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08899801969528198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05588744580745697, + 0.1099616140127182, + 0.08646685630083084, + 0.10253595560789108, + 0.062178339809179306, + 0.1321297585964203, + 0.11758454889059067, + 0.01964799501001835, + 0.095801442861557, + 0.21780604124069214 + ], + "credit_map": [ + 0.006531472783535719, + 0.0056938487105071545, + 0.007351182401180267, + 0.006965802516788244, + 0.004157927352935076, + 0.005698758643120527, + 0.007327237166464329, + 0.012161215767264366, + 0.012855679728090763, + 0.005608477629721165, + 0.012833504006266594, + 0.009107439778745174, + 0.007732053752988577, + 0.019887397065758705, + 0.009544064290821552, + 0.03239299729466438, + 0.004641147796064615, + 0.003958749584853649, + 0.005259620025753975, + 0.00552680529654026, + 0.003592178225517273, + 0.016852611675858498, + 0.004913635086268187, + 0.04172210395336151, + 0.006492525804787874, + 0.009819237515330315, + 0.006010825280100107, + 0.0054114218801259995, + 0.016019167378544807, + 0.0220516137778759, + 0.014136029407382011, + 0.022595129907131195, + 0.0035158228129148483, + 0.008389913477003574, + 0.00438524829223752, + 0.007450610864907503, + 0.008108370937407017, + 0.010832777246832848, + 0.008355542086064816, + 0.011140057817101479, + 0.011739375069737434, + 0.005082342308014631, + 0.009194628335535526, + 0.008783114142715931, + 0.013300314545631409, + 0.046083588153123856, + 0.013680793344974518, + 0.024265604093670845, + 0.008830771781504154, + 0.004895215388387442, + 0.016062762588262558, + 0.005714642349630594, + 0.03857848793268204, + 0.01781187206506729, + 0.011873845010995865, + 0.013816950842738152, + 0.0017788887489587069, + 0.0019589208532124758, + 0.002184683457016945, + 0.0016596333589404821, + 0.004135653842240572, + 0.0026822432409971952, + 0.0022302328143268824, + 0.0030177372973412275, + 0.017796196043491364, + 0.010095575824379921, + 0.01006376463919878, + 0.007230178453028202, + 0.010981141589581966, + 0.013191795907914639, + 0.008349557407200336, + 0.018093228340148926, + 0.03361548110842705, + 0.021757489070296288, + 0.02933250367641449, + 0.04765421152114868, + 0.017351701855659485, + 0.02071034163236618, + 0.03286181762814522, + 0.014522505924105644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.08899801969528198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman wearing a crimson silk robe with golden phoenix embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a towering bun adorned with a golden phoenix hairpin that sways with her movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a grand palace corridor with massive vermilion columns and a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly elevated perspective and panning rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman walking with steady, deliberate steps causing her skirt to undulate like a wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637083038687706 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637083038687706, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637083038687706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18456421792507172, + 0.1328456848859787, + 0.07194574177265167, + 0.059111326932907104, + 0.08366486430168152, + 0.14680182933807373, + 0.13072994351387024, + 0.0345582477748394, + 0.032559409737586975, + 0.12321864813566208 + ], + "credit_map": [ + 0.024318022653460503, + 0.021600091829895973, + 0.024069171398878098, + 0.01776101440191269, + 0.021844422444701195, + 0.022696996107697487, + 0.023848509415984154, + 0.02842598408460617, + 0.015410075895488262, + 0.0071192956529557705, + 0.01779959537088871, + 0.010146952234208584, + 0.012789612635970116, + 0.019473761320114136, + 0.011915971525013447, + 0.03819042444229126, + 0.005421960260719061, + 0.005086549092084169, + 0.009665151126682758, + 0.006552882958203554, + 0.005808739457279444, + 0.011440637521445751, + 0.0057660858146846294, + 0.022203737869858742, + 0.005659173242747784, + 0.0037991548888385296, + 0.007280032150447369, + 0.0053975386545062065, + 0.007042493671178818, + 0.01000252366065979, + 0.008380955085158348, + 0.011549452319741249, + 0.004209822975099087, + 0.008801767602562904, + 0.008334826678037643, + 0.007424954324960709, + 0.00865370687097311, + 0.016382774338126183, + 0.007853356190025806, + 0.02200365625321865, + 0.019028346985578537, + 0.0072850435972213745, + 0.01271731685847044, + 0.009637031704187393, + 0.016615640372037888, + 0.04012777656316757, + 0.02384079620242119, + 0.017549877986311913, + 0.009280465543270111, + 0.012344547547399998, + 0.017251405864953995, + 0.014015693217515945, + 0.027838781476020813, + 0.016415167599916458, + 0.023661019280552864, + 0.009922857396304607, + 0.0028473646380007267, + 0.006192575674504042, + 0.00323278340511024, + 0.004421886056661606, + 0.005162185523658991, + 0.004139329772442579, + 0.004694820381700993, + 0.00386730395257473, + 0.005092023406177759, + 0.004590868018567562, + 0.003956491127610207, + 0.0033892178907990456, + 0.004580244887620211, + 0.0038299434818327427, + 0.0031383000314235687, + 0.003982321359217167, + 0.01884809136390686, + 0.013741286471486092, + 0.014889989048242569, + 0.020792996510863304, + 0.011001509614288807, + 0.01697542518377304, + 0.017499499022960663, + 0.00946984626352787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004637083038687706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_tensor([3581, 3581])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_tensor([3581, 3581])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e883afb6aabf0f7981235041e5c54fbcdf4095a7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_tensor([3581, 3581])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful close-up, captured from a slightly upward angle and accompanied by a slight handheld camera shake, locks the viewer's gaze on a hand, its knuckles white from extreme exertion. The back of the hand, stained with dust and minute scratches, bears the marks of intense combat. The hand tightly grips a delicate silver hairpin, its head carved with intricate classical patterns, yet in the dim light, its sharp tip reflects a deadly cold gleam. With a resolute and swift thrust, the wrist thrusts forward, unhesitatingly and decisively piercing the opponent in the foreground. The background is blurred by motion, revealing only the tangled folds of clothing and the rising dust. A narrow beam of cold light slants in from the side and above, precisely outlining the taut muscles of the arm and leaving a brief bright mark on the trajectory of the metal hairpin tip cutting through the air. The entire picture is gloomy in tone, full of critical tension and primitive violent aesthetics, freezing the desperate moment of resistance in this fierce fight.", + "seed": 27734, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_3581_seed_27734_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12314080242067575, + "vlm_logprob": -0.12314080242067575, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.12314080242067575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013598897494375706 + ], + "opposite_logprob_tokens": [ + -4.888598918914795 + ], + "logprob_reward": -0.013598897494375706, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013598897494375706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09714718908071518, + 0.23803694546222687, + 0.062329236418008804, + 0.04797627404332161, + 0.10191567242145538, + 0.06278786808252335, + 0.15081001818180084, + 0.09762934595346451, + 0.09630008041858673, + 0.04506737366318703 + ], + "credit_map": [ + 0.01195480115711689, + 0.008392232470214367, + 0.01269365381449461, + 0.012168891727924347, + 0.009395014494657516, + 0.013086012564599514, + 0.01290311012417078, + 0.016553467139601707, + 0.014780271798372269, + 0.010849020443856716, + 0.02488650381565094, + 0.007805290166288614, + 0.006991489324718714, + 0.07970555871725082, + 0.008891110308468342, + 0.08412770181894302, + 0.006592842750251293, + 0.008071907795965672, + 0.008600059896707535, + 0.006988746579736471, + 0.006752331275492907, + 0.011341376230120659, + 0.007123199291527271, + 0.0068587674759328365, + 0.005599597934633493, + 0.0070766182616353035, + 0.006413533352315426, + 0.004475928843021393, + 0.0031034823041409254, + 0.009527524001896381, + 0.005612288601696491, + 0.006167301908135414, + 0.009833481162786484, + 0.009133759886026382, + 0.015611497685313225, + 0.010745087638497353, + 0.014546005055308342, + 0.01580994576215744, + 0.01614246517419815, + 0.010093433782458305, + 0.005576665513217449, + 0.00688167056068778, + 0.007341406308114529, + 0.007652890868484974, + 0.011736198328435421, + 0.008506395854055882, + 0.009233484975993633, + 0.005859155207872391, + 0.012285897508263588, + 0.023165857419371605, + 0.024777648970484734, + 0.016395175829529762, + 0.024424457922577858, + 0.014821558259427547, + 0.017765512689948082, + 0.01717391237616539, + 0.00960781704634428, + 0.007637816481292248, + 0.006098036654293537, + 0.011545488610863686, + 0.015725722536444664, + 0.014386002905666828, + 0.017582297325134277, + 0.015046169050037861, + 0.021542351692914963, + 0.015876812860369682, + 0.00968076940625906, + 0.006635245867073536, + 0.015086556784808636, + 0.009357641451060772, + 0.010916096158325672, + 0.007204605732113123, + 0.004729040432721376, + 0.005685391370207071, + 0.00562436506152153, + 0.004801184870302677, + 0.007759551517665386, + 0.006664014421403408, + 0.006039976608008146, + 0.0037638500798493624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.013598897494375706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038556396961212 + ], + "opposite_logprob_tokens": [ + -1.9103856086730957 + ], + "logprob_reward": -0.16038556396961212, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038556396961212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2288321852684021, + 0.07147541642189026, + 0.05546002835035324, + 0.06009911373257637, + 0.07107953727245331, + 0.08850936591625214, + 0.09335019439458847, + 0.1860121488571167, + 0.06132509931921959, + 0.08385685831308365 + ], + "credit_map": [ + 0.021179499104619026, + 0.022823473438620567, + 0.026151545345783234, + 0.028772441670298576, + 0.02831977978348732, + 0.025958159938454628, + 0.04377228394150734, + 0.03185503929853439, + 0.00868497509509325, + 0.004418343771249056, + 0.016499178484082222, + 0.007585098035633564, + 0.004538542125374079, + 0.01062588021159172, + 0.006297598592936993, + 0.012825802899897099, + 0.004393405746668577, + 0.00648176996037364, + 0.007915579713881016, + 0.007018418982625008, + 0.005558039527386427, + 0.008225047029554844, + 0.007107781246304512, + 0.008759988471865654, + 0.006783700082451105, + 0.0062310597859323025, + 0.0072283935733139515, + 0.006001523230224848, + 0.004270236007869244, + 0.010254072025418282, + 0.007910244166851044, + 0.011419892311096191, + 0.011998996138572693, + 0.006583517882972956, + 0.014707592315971851, + 0.007013689260929823, + 0.007640051189810038, + 0.006805715151131153, + 0.01069403812289238, + 0.005635946989059448, + 0.011610780842602253, + 0.010810009203851223, + 0.01129756961017847, + 0.011621838435530663, + 0.011042861267924309, + 0.013750414364039898, + 0.00908656045794487, + 0.009289337322115898, + 0.007658547721803188, + 0.01591630093753338, + 0.014226997271180153, + 0.00793408416211605, + 0.02173086442053318, + 0.008368183858692646, + 0.009467264637351036, + 0.008047963492572308, + 0.010023838840425014, + 0.009866074658930302, + 0.008168783038854599, + 0.0200033001601696, + 0.026067165657877922, + 0.02636372670531273, + 0.05461388826370239, + 0.030905388295650482, + 0.008913008496165276, + 0.00617372477427125, + 0.006338790524750948, + 0.007918971590697765, + 0.009391671977937222, + 0.006007300224155188, + 0.0117446668446064, + 0.004836970940232277, + 0.00792623683810234, + 0.00914745219051838, + 0.010553095489740372, + 0.010713886469602585, + 0.013120611198246479, + 0.012330769561231136, + 0.01118119154125452, + 0.00888361781835556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.16038556396961212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486531972885132 + ], + "opposite_logprob_tokens": [ + -1.2236531972885132 + ], + "logprob_reward": -0.3486531972885132, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486531972885132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06129729002714157, + 0.24191229045391083, + 0.08464044332504272, + 0.031123751774430275, + 0.1969122290611267, + 0.10583138465881348, + 0.06729860603809357, + 0.07181147485971451, + 0.1210779920220375, + 0.018094617873430252 + ], + "credit_map": [ + 0.0063210101798176765, + 0.0060292319394648075, + 0.007422260008752346, + 0.008137702010571957, + 0.007923297584056854, + 0.005833695642650127, + 0.010445328429341316, + 0.009184754453599453, + 0.017598651349544525, + 0.009673766791820526, + 0.02835642173886299, + 0.009237800724804401, + 0.006872046273201704, + 0.0720805749297142, + 0.00793097447603941, + 0.0901620164513588, + 0.012369555421173573, + 0.011824243701994419, + 0.01241255085915327, + 0.008298477157950401, + 0.007855274714529514, + 0.012715304270386696, + 0.008019313216209412, + 0.011145716533064842, + 0.003570640692487359, + 0.004118660930544138, + 0.0035926089622080326, + 0.003692778991535306, + 0.003441538428887725, + 0.005299299955368042, + 0.004617963917553425, + 0.002790257567539811, + 0.020301880314946175, + 0.02169862575829029, + 0.03146699443459511, + 0.022968463599681854, + 0.02687486819922924, + 0.016252940520644188, + 0.04151361808180809, + 0.01583481766283512, + 0.009126127697527409, + 0.008772043511271477, + 0.01448752824217081, + 0.010656674392521381, + 0.022065160796046257, + 0.00933996494859457, + 0.01728915609419346, + 0.01409471407532692, + 0.004123591352254152, + 0.007613540161401033, + 0.00667357724159956, + 0.0066892062313854694, + 0.01197704579681158, + 0.008924653753638268, + 0.011284558102488518, + 0.010012432001531124, + 0.007997368462383747, + 0.005844231694936752, + 0.004616738762706518, + 0.008967732079327106, + 0.010217645205557346, + 0.010558940470218658, + 0.013185331597924232, + 0.010423479601740837, + 0.026386622339487076, + 0.012933404184877872, + 0.011628477834165096, + 0.011364314705133438, + 0.018596462905406952, + 0.011153053492307663, + 0.019950343295931816, + 0.009065298363566399, + 0.0023149456828832626, + 0.0020809222478419542, + 0.001799134537577629, + 0.002016466110944748, + 0.0029648588970303535, + 0.002550175180658698, + 0.002078138291835785, + 0.002289974596351385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.3486531972885132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810923546552658 + ], + "opposite_logprob_tokens": [ + -3.288109302520752 + ], + "logprob_reward": -0.03810923546552658, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810923546552658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1456625908613205, + 0.25026920437812805, + 0.0658656656742096, + 0.0237773098051548, + 0.05105166509747505, + 0.06698429584503174, + 0.1296977400779724, + 0.16347724199295044, + 0.056257739663124084, + 0.04695658013224602 + ], + "credit_map": [ + 0.013062961399555206, + 0.01631631888449192, + 0.016621127724647522, + 0.021442020311951637, + 0.020762087777256966, + 0.014721758663654327, + 0.024987345561385155, + 0.01774897240102291, + 0.008831040933728218, + 0.008478304371237755, + 0.038445908576250076, + 0.008336765691637993, + 0.005799116566777229, + 0.0984409749507904, + 0.007697057910263538, + 0.0742400661110878, + 0.00726779131218791, + 0.009058477357029915, + 0.00901851523667574, + 0.008996281772851944, + 0.006512730848044157, + 0.008717228658497334, + 0.006730160675942898, + 0.009564480744302273, + 0.002844839822500944, + 0.004121245350688696, + 0.0025117744226008654, + 0.0028566429391503334, + 0.0025799432769417763, + 0.0036699322517961264, + 0.0031780998688191175, + 0.002014832105487585, + 0.005008275154978037, + 0.005756532307714224, + 0.007065724115818739, + 0.006131806876510382, + 0.007758242078125477, + 0.004837379790842533, + 0.00976464245468378, + 0.004729057662189007, + 0.008307605050504208, + 0.007359716109931469, + 0.008263738825917244, + 0.0075000072829425335, + 0.0124142374843359, + 0.00843726471066475, + 0.007704923860728741, + 0.006996802520006895, + 0.007718219887465239, + 0.01780153438448906, + 0.014474749565124512, + 0.01683836244046688, + 0.016692087054252625, + 0.014456653036177158, + 0.0204909797757864, + 0.02122514508664608, + 0.011628388427197933, + 0.008861824870109558, + 0.00971193891018629, + 0.013602709397673607, + 0.023101914674043655, + 0.0219722967594862, + 0.04304829239845276, + 0.031549882143735886, + 0.013142815791070461, + 0.006768994033336639, + 0.00634785508736968, + 0.004900846164673567, + 0.0064028664492070675, + 0.006982817780226469, + 0.007552077993750572, + 0.004159465432167053, + 0.005018901079893112, + 0.004482005722820759, + 0.004742649849504232, + 0.006468536797910929, + 0.006691181566566229, + 0.007841023616492748, + 0.0063744597136974335, + 0.005337824113667011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.03810923546552658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495711788535118 + ], + "opposite_logprob_tokens": [ + -2.929957151412964 + ], + "logprob_reward": -0.05495711788535118, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495711788535118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07683950662612915, + 0.2501565217971802, + 0.07661645114421844, + 0.03448021784424782, + 0.0488588772714138, + 0.027269070968031883, + 0.05668460950255394, + 0.2445524036884308, + 0.11019343137741089, + 0.07434888929128647 + ], + "credit_map": [ + 0.008643199689686298, + 0.0059820422902703285, + 0.008928309194743633, + 0.007952651008963585, + 0.006758198607712984, + 0.010726193897426128, + 0.012085654772818089, + 0.015763262286782265, + 0.008169470354914665, + 0.010834425687789917, + 0.0409691259264946, + 0.008810524828732014, + 0.0072424327954649925, + 0.09887773543596268, + 0.008165846578776836, + 0.06708695739507675, + 0.007681071758270264, + 0.00877357367426157, + 0.011325730942189693, + 0.009502164088189602, + 0.007445595692843199, + 0.013023377396166325, + 0.007966363802552223, + 0.010898576118052006, + 0.0033685709349811077, + 0.005848952103406191, + 0.004278015345335007, + 0.003952821716666222, + 0.003740994492545724, + 0.005808187648653984, + 0.004521993920207024, + 0.0029606823809444904, + 0.004971441812813282, + 0.005114170257002115, + 0.007568072993308306, + 0.004411633592098951, + 0.00893821008503437, + 0.005558045115321875, + 0.008394469507038593, + 0.0039028397295624018, + 0.0030991113744676113, + 0.0033033937215805054, + 0.0037232604809105396, + 0.003119763918220997, + 0.004626748152077198, + 0.003619191236793995, + 0.0029745586216449738, + 0.0028030446264892817, + 0.006204941775649786, + 0.008842777460813522, + 0.009081678465008736, + 0.005104373209178448, + 0.0092061348259449, + 0.005880290176719427, + 0.0065277013927698135, + 0.005836715921759605, + 0.0167537871748209, + 0.015446295961737633, + 0.013050401583313942, + 0.03002462536096573, + 0.03391103819012642, + 0.03168419003486633, + 0.061323132365942, + 0.04235895723104477, + 0.027256378903985023, + 0.013798488304018974, + 0.015430459752678871, + 0.010342960245907307, + 0.012747755274176598, + 0.009537175297737122, + 0.012276940979063511, + 0.008803275413811207, + 0.007989046163856983, + 0.009317474439740181, + 0.010360497049987316, + 0.009423531591892242, + 0.011782159097492695, + 0.00795531552284956, + 0.009888608008623123, + 0.007632264867424965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.05495711788535118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful close-up, captured from a slightly upward angle and accompanied by a slight handheld camera shake, locks the viewer's gaze on a hand, its knuckles white from extreme exertion. The back of the hand, stained with dust and minute scratches, bears the marks of intense combat. The hand tightly grips a delicate silver hairpin, its head carved with intricate classical patterns, yet in the dim light, its sharp tip reflects a deadly cold gleam. With a resolute and swift thrust, the wrist thrusts forward, unhesitatingly and decisively piercing the opponent in the foreground. The background is blurred by motion, revealing only the tangled folds of clothing and the rising dust. A narrow beam of cold light slants in from the side and above, precisely outlining the taut muscles of the arm and leaving a brief bright mark on the trajectory of the metal hairpin tip cutting through the air. The entire picture is gloomy in tone, full of critical tension and primitive violent aesthetics, freezing the desperate moment of resistance in this fierce fight.", + "seed": 27735, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_3581_seed_27735_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12042981162667274, + "vlm_logprob": -0.12042981162667274, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.12042981162667274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015403473749756813 + ], + "opposite_logprob_tokens": [ + -5.0154032707214355 + ], + "logprob_reward": -0.015403473749756813, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015403473749756813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18538697063922882, + 0.157033309340477, + 0.053653765469789505, + 0.04698038846254349, + 0.07402294129133224, + 0.0511748231947422, + 0.15910011529922485, + 0.06338756531476974, + 0.036036502569913864, + 0.17322365939617157 + ], + "credit_map": [ + 0.04775037616491318, + 0.023357126861810684, + 0.01870892569422722, + 0.01753159984946251, + 0.023268159478902817, + 0.016958359628915787, + 0.014433979988098145, + 0.023378439247608185, + 0.02976350486278534, + 0.013037547469139099, + 0.01686994545161724, + 0.027162032201886177, + 0.011377394199371338, + 0.01382116973400116, + 0.028722146525979042, + 0.01627957448363304, + 0.006308678537607193, + 0.004946539644151926, + 0.005587431136518717, + 0.01026483066380024, + 0.005414782557636499, + 0.005122173577547073, + 0.009035654366016388, + 0.006973674055188894, + 0.004695166368037462, + 0.007088261656463146, + 0.007778192404657602, + 0.0058629149571061134, + 0.00732074398547411, + 0.004874964244663715, + 0.005807928275316954, + 0.003552212379872799, + 0.011874827556312084, + 0.0059410640969872475, + 0.012464124709367752, + 0.0066824075765907764, + 0.004783186595886946, + 0.009791180491447449, + 0.00840099435299635, + 0.014085154049098492, + 0.00649245735257864, + 0.008628858253359795, + 0.008612141944468021, + 0.004968720022588968, + 0.004417892079800367, + 0.007790666073560715, + 0.005625709425657988, + 0.004638378508388996, + 0.009766971692442894, + 0.009573342278599739, + 0.05084402114152908, + 0.04744160175323486, + 0.00976945087313652, + 0.010888837277889252, + 0.012748154811561108, + 0.0080677205696702, + 0.0039927479811012745, + 0.004343167878687382, + 0.010037911124527454, + 0.007847046479582787, + 0.007094105239957571, + 0.0059443628415465355, + 0.008733064867556095, + 0.015395157970488071, + 0.004466371610760689, + 0.004619342740625143, + 0.005775728728622198, + 0.003236493794247508, + 0.004406740423291922, + 0.005391821265220642, + 0.003323934506624937, + 0.004816070199012756, + 0.021049348637461662, + 0.04952392354607582, + 0.05591357499361038, + 0.011711311526596546, + 0.009387160651385784, + 0.010441314429044724, + 0.007909448817372322, + 0.007287572603672743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.015403473749756813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799594610929489 + ], + "opposite_logprob_tokens": [ + -1.8049594163894653 + ], + "logprob_reward": -0.1799594610929489, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799594610929489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10538271814584732, + 0.09390803426504135, + 0.16743259131908417, + 0.18357321619987488, + 0.08563832193613052, + 0.04340372979640961, + 0.16716957092285156, + 0.03535953536629677, + 0.031788893043994904, + 0.08634340763092041 + ], + "credit_map": [ + 0.014234881848096848, + 0.015005898661911488, + 0.009887062013149261, + 0.014256870374083519, + 0.017817707732319832, + 0.012330079451203346, + 0.011469708755612373, + 0.010380514897406101, + 0.013222280889749527, + 0.008835136890411377, + 0.011616374365985394, + 0.01568601094186306, + 0.010184669867157936, + 0.010109862312674522, + 0.013583855703473091, + 0.010669841431081295, + 0.023763269186019897, + 0.013131335377693176, + 0.023538222536444664, + 0.027596840634942055, + 0.015193819999694824, + 0.018220100551843643, + 0.02810525707900524, + 0.01788376271724701, + 0.016086231917142868, + 0.06000123545527458, + 0.0476553812623024, + 0.017724480479955673, + 0.015130273066461086, + 0.00760392053052783, + 0.008812730200588703, + 0.010558977723121643, + 0.010188721120357513, + 0.01153504941612482, + 0.013377373106777668, + 0.009501947090029716, + 0.010404229164123535, + 0.01400825846940279, + 0.00897988211363554, + 0.0076428563334047794, + 0.005716396030038595, + 0.006717659067362547, + 0.005845222622156143, + 0.004881895612925291, + 0.0036512992810457945, + 0.005231189075857401, + 0.006046408321708441, + 0.005313660018146038, + 0.008816413581371307, + 0.008618898689746857, + 0.049609407782554626, + 0.04652063921093941, + 0.011154046282172203, + 0.00991035159677267, + 0.0258786678314209, + 0.006661138031631708, + 0.0029470929875969887, + 0.003996651154011488, + 0.004942626692354679, + 0.005013298708945513, + 0.0038393521681427956, + 0.003574367379769683, + 0.0049539110623300076, + 0.006092236842960119, + 0.0031802079174667597, + 0.0033700314816087484, + 0.004208707716315985, + 0.004306824877858162, + 0.002963382052257657, + 0.0046441941522061825, + 0.004487223923206329, + 0.004628323018550873, + 0.005752171389758587, + 0.014597228728234768, + 0.029625188559293747, + 0.004740577656775713, + 0.004856180399656296, + 0.01604355126619339, + 0.004936104640364647, + 0.005792400799691677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.1799594610929489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134627640247345 + ], + "opposite_logprob_tokens": [ + -1.313462734222412 + ], + "logprob_reward": -0.3134627640247345, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134627640247345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17979705333709717, + 0.10193699598312378, + 0.05040261894464493, + 0.24762436747550964, + 0.19066579639911652, + 0.04583338648080826, + 0.06712893396615982, + 0.04832445830106735, + 0.03142724558711052, + 0.036859095096588135 + ], + "credit_map": [ + 0.04156944155693054, + 0.02372623234987259, + 0.01455594226717949, + 0.022168610244989395, + 0.01971857063472271, + 0.012608119286596775, + 0.01920938305556774, + 0.026240766048431396, + 0.01874944381415844, + 0.008932199329137802, + 0.009607454761862755, + 0.016118360683321953, + 0.0080654826015234, + 0.008869181387126446, + 0.020836733281612396, + 0.01075814850628376, + 0.008144394494593143, + 0.0036677622701972723, + 0.0062950123101472855, + 0.006814755033701658, + 0.007259154226630926, + 0.006008091848343611, + 0.006172759458422661, + 0.006040691863745451, + 0.02443784475326538, + 0.07930558919906616, + 0.06703447550535202, + 0.022691845893859863, + 0.019795799627900124, + 0.010312932543456554, + 0.010031769052147865, + 0.01401414442807436, + 0.02816101722419262, + 0.026717906817793846, + 0.03318345174193382, + 0.02127666212618351, + 0.015517527237534523, + 0.027538711205124855, + 0.02464739792048931, + 0.013623136095702648, + 0.005807638168334961, + 0.004993119742721319, + 0.008033150807023048, + 0.006770634092390537, + 0.004125627689063549, + 0.006616562604904175, + 0.005364235490560532, + 0.004122423008084297, + 0.005218773614615202, + 0.006121794227510691, + 0.010239710099995136, + 0.009839741513133049, + 0.006690065376460552, + 0.007159985136240721, + 0.017857415601611137, + 0.004001452587544918, + 0.0038653863593935966, + 0.00492433924227953, + 0.007076906971633434, + 0.006495020352303982, + 0.004998974502086639, + 0.003818577155470848, + 0.00773107074201107, + 0.00941418670117855, + 0.004771166946738958, + 0.003232774790376425, + 0.0045397160574793816, + 0.004373759031295776, + 0.0028611039742827415, + 0.0035147452726960182, + 0.003967325668781996, + 0.004166657570749521, + 0.0032159092370420694, + 0.007769510615617037, + 0.008659951388835907, + 0.003185379784554243, + 0.0037223491817712784, + 0.0027514230459928513, + 0.0030777461361140013, + 0.004476828500628471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.3134627640247345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329820580780506 + ], + "opposite_logprob_tokens": [ + -3.7732982635498047 + ], + "logprob_reward": -0.02329820580780506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329820580780506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11960376799106598, + 0.25585684180259705, + 0.09232185035943985, + 0.09358853101730347, + 0.11042405664920807, + 0.09374683350324631, + 0.0938684418797493, + 0.03170487657189369, + 0.03240010514855385, + 0.07648476958274841 + ], + "credit_map": [ + 0.016661109402775764, + 0.022757703438401222, + 0.012750478461384773, + 0.01631254516541958, + 0.015006983652710915, + 0.008776683360338211, + 0.015692221000790596, + 0.011646036989986897, + 0.06062537431716919, + 0.018292130902409554, + 0.03537479043006897, + 0.039570651948451996, + 0.016548436135053635, + 0.017962973564863205, + 0.043998073786497116, + 0.02348439022898674, + 0.009590310044586658, + 0.007839512079954147, + 0.009126551449298859, + 0.02058226242661476, + 0.00973673164844513, + 0.00934772752225399, + 0.017937637865543365, + 0.00816111359745264, + 0.009917194955050945, + 0.01954447105526924, + 0.020397400483489037, + 0.01011696271598339, + 0.009160011075437069, + 0.006422782316803932, + 0.009141490794718266, + 0.008888211101293564, + 0.012539014220237732, + 0.013563266023993492, + 0.021374493837356567, + 0.00946163758635521, + 0.012773076072335243, + 0.012726313434541225, + 0.013629487715661526, + 0.014356758445501328, + 0.010467187501490116, + 0.00987640768289566, + 0.016814609989523888, + 0.009431636892259121, + 0.012642595916986465, + 0.01706669107079506, + 0.009774402715265751, + 0.007673288695514202, + 0.006671296898275614, + 0.011706395074725151, + 0.01877456344664097, + 0.017963362857699394, + 0.009706291370093822, + 0.007184996735304594, + 0.017160531133413315, + 0.004700995050370693, + 0.003506039036437869, + 0.0027920727152377367, + 0.0051337978802621365, + 0.0037916870787739754, + 0.0034991507418453693, + 0.0031491874251514673, + 0.005232437513768673, + 0.004600502084940672, + 0.002976929070428014, + 0.0032907805871218443, + 0.005226518958806992, + 0.004125334322452545, + 0.0030405432917177677, + 0.004414157010614872, + 0.004299298394471407, + 0.005026537459343672, + 0.006015942431986332, + 0.008953570388257504, + 0.027161631733179092, + 0.00488984864205122, + 0.004715960938483477, + 0.015524648129940033, + 0.004085908643901348, + 0.005137252155691385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.02329820580780506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002515345811844 + ], + "opposite_logprob_tokens": [ + -2.6950252056121826 + ], + "logprob_reward": -0.07002515345811844, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002515345811844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11394034326076508, + 0.05641547217965126, + 0.09888264536857605, + 0.09072861820459366, + 0.11090065538883209, + 0.05305585637688637, + 0.11616973578929901, + 0.04801371693611145, + 0.040595825761556625, + 0.2712970972061157 + ], + "credit_map": [ + 0.011602142825722694, + 0.019450193271040916, + 0.013380314223468304, + 0.013764514587819576, + 0.01859927363693714, + 0.01229066587984562, + 0.013908102177083492, + 0.010945137590169907, + 0.009910442866384983, + 0.003896816400811076, + 0.006673547904938459, + 0.009666132740676403, + 0.005418968386948109, + 0.005582904908806086, + 0.0100098242983222, + 0.005256832577288151, + 0.009888073429465294, + 0.007977757602930069, + 0.01148246694356203, + 0.01893252320587635, + 0.012463883496820927, + 0.01118243858218193, + 0.01586442068219185, + 0.011091080494225025, + 0.008574506267905235, + 0.02714931219816208, + 0.02313089929521084, + 0.008121738210320473, + 0.008124567568302155, + 0.005362992640584707, + 0.004312463570386171, + 0.005952134728431702, + 0.01556980051100254, + 0.010036936961114407, + 0.02514653280377388, + 0.011202199384570122, + 0.00956655852496624, + 0.016278628259897232, + 0.012658720836043358, + 0.010441270656883717, + 0.005703856702893972, + 0.006143843289464712, + 0.008711135946214199, + 0.0073583354242146015, + 0.005899809766560793, + 0.0083156144246459, + 0.006243167445063591, + 0.004680093377828598, + 0.0071949888952076435, + 0.005680480971932411, + 0.03590600937604904, + 0.033155519515275955, + 0.005202633328735828, + 0.007909308187663555, + 0.01673349365592003, + 0.00438730837777257, + 0.004774624947458506, + 0.005100378766655922, + 0.005571594927459955, + 0.005420512054115534, + 0.006500320043414831, + 0.005012170411646366, + 0.007285848259925842, + 0.008348268456757069, + 0.005717648658901453, + 0.004238248337060213, + 0.006139186676591635, + 0.004201236180961132, + 0.004239780828356743, + 0.004707630258053541, + 0.004257079213857651, + 0.007095014210790396, + 0.029513176530599594, + 0.07188361883163452, + 0.0875207930803299, + 0.019662825390696526, + 0.014206212013959885, + 0.019281014800071716, + 0.013448880054056644, + 0.015780562534928322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.07002515345811844 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful close-up, captured from a slightly upward angle and accompanied by a slight handheld camera shake, locks the viewer's gaze on a hand, its knuckles white from extreme exertion. The back of the hand, stained with dust and minute scratches, bears the marks of intense combat. The hand tightly grips a delicate silver hairpin, its head carved with intricate classical patterns, yet in the dim light, its sharp tip reflects a deadly cold gleam. With a resolute and swift thrust, the wrist thrusts forward, unhesitatingly and decisively piercing the opponent in the foreground. The background is blurred by motion, revealing only the tangled folds of clothing and the rising dust. A narrow beam of cold light slants in from the side and above, precisely outlining the taut muscles of the arm and leaving a brief bright mark on the trajectory of the metal hairpin tip cutting through the air. The entire picture is gloomy in tone, full of critical tension and primitive violent aesthetics, freezing the desperate moment of resistance in this fierce fight.", + "seed": 27736, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_3581_seed_27736_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09111605659127235, + "vlm_logprob": -0.09111605659127235, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09111605659127235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01228833757340908 + ], + "opposite_logprob_tokens": [ + -4.762288570404053 + ], + "logprob_reward": -0.01228833757340908, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01228833757340908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13347461819648743, + 0.12660649418830872, + 0.07200689613819122, + 0.07992617785930634, + 0.20379966497421265, + 0.13371653854846954, + 0.12459678947925568, + 0.05105637013912201, + 0.0391254797577858, + 0.03569095954298973 + ], + "credit_map": [ + 0.010260279290378094, + 0.011194057762622833, + 0.010091082192957401, + 0.017460396513342857, + 0.019646963104605675, + 0.019429856911301613, + 0.025440869852900505, + 0.01995111256837845, + 0.01571548357605934, + 0.013566836714744568, + 0.017160983756184578, + 0.016235997900366783, + 0.015226666815578938, + 0.01845756731927395, + 0.012943929992616177, + 0.01729903370141983, + 0.00815308466553688, + 0.005324753932654858, + 0.009785978123545647, + 0.00733410706743598, + 0.010465198196470737, + 0.012670472264289856, + 0.010825581848621368, + 0.007447723764926195, + 0.007791611365973949, + 0.00828510895371437, + 0.006372619420289993, + 0.00903076771646738, + 0.014272701926529408, + 0.009196219965815544, + 0.010569704696536064, + 0.014407442882657051, + 0.010578063316643238, + 0.010422533378005028, + 0.012078270316123962, + 0.01773139089345932, + 0.007890553213655949, + 0.053420618176460266, + 0.015918269753456116, + 0.07575994729995728, + 0.014162992127239704, + 0.017835302278399467, + 0.01670026034116745, + 0.026291970163583755, + 0.01354672946035862, + 0.010224081575870514, + 0.017974169924855232, + 0.016981014981865883, + 0.010297744534909725, + 0.021361446008086205, + 0.012866410426795483, + 0.017318058758974075, + 0.024552810937166214, + 0.01045201811939478, + 0.014712699688971043, + 0.013035599142313004, + 0.005072195082902908, + 0.0051550050266087055, + 0.009517080150544643, + 0.0044680810533463955, + 0.0037131516728550196, + 0.006323993671685457, + 0.005020066164433956, + 0.01178679522126913, + 0.0038236037362366915, + 0.006054640747606754, + 0.006128233391791582, + 0.004859341774135828, + 0.00406659534201026, + 0.0058948625810444355, + 0.0045031774789094925, + 0.0037950235418975353, + 0.002233407460153103, + 0.005485292058438063, + 0.0050034672021865845, + 0.00444694934412837, + 0.0038915665354579687, + 0.006816149223595858, + 0.004148910753428936, + 0.00366521836258471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01228833757340908 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906454592943192 + ], + "opposite_logprob_tokens": [ + -2.464064598083496 + ], + "logprob_reward": -0.08906454592943192, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906454592943192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07022926956415176, + 0.13587099313735962, + 0.12298769503831863, + 0.09799613058567047, + 0.21348953247070312, + 0.07280346751213074, + 0.034823108464479446, + 0.0863775908946991, + 0.1021825298666954, + 0.06323962658643723 + ], + "credit_map": [ + 0.007428294979035854, + 0.006099460646510124, + 0.0066491528414189816, + 0.008915356360375881, + 0.008408455178141594, + 0.009865463711321354, + 0.011303439736366272, + 0.011559653095901012, + 0.016999470070004463, + 0.01449099276214838, + 0.0141197070479393, + 0.018968790769577026, + 0.01481141708791256, + 0.01529335230588913, + 0.022381918504834175, + 0.0188053660094738, + 0.02998463623225689, + 0.011398246511816978, + 0.013158749788999557, + 0.011909869499504566, + 0.008299056440591812, + 0.013599979691207409, + 0.020069433376193047, + 0.014567737467586994, + 0.014738015830516815, + 0.01646190509200096, + 0.010747787542641163, + 0.01154971681535244, + 0.013662301935255527, + 0.008831722661852837, + 0.012261182069778442, + 0.009743506088852882, + 0.007361900061368942, + 0.0041696154512465, + 0.014071004465222359, + 0.015724094584584236, + 0.007425204385071993, + 0.06047039479017258, + 0.017197052016854286, + 0.0870702713727951, + 0.0056678177788853645, + 0.006285575218498707, + 0.009085023775696754, + 0.010742113925516605, + 0.006206316873431206, + 0.005766470916569233, + 0.01380955521017313, + 0.015240596607327461, + 0.003392950166016817, + 0.005644611082971096, + 0.0035266915801912546, + 0.004853968508541584, + 0.004786396399140358, + 0.003141845343634486, + 0.004660576581954956, + 0.004816070664674044, + 0.006228921469300985, + 0.005867811851203442, + 0.01846414804458618, + 0.005379705689847469, + 0.008499694988131523, + 0.008315061219036579, + 0.007096858695149422, + 0.02652539685368538, + 0.010330344550311565, + 0.014375410042703152, + 0.01505726296454668, + 0.011972584761679173, + 0.009997643530368805, + 0.01580760069191456, + 0.012555592693388462, + 0.012086098082363605, + 0.005641276948153973, + 0.004851056728512049, + 0.016419466584920883, + 0.008624372072517872, + 0.006100225728005171, + 0.007469949312508106, + 0.006061525549739599, + 0.008071753196418285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.08906454592943192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135615587234497 + ], + "opposite_logprob_tokens": [ + -1.4063560962677002 + ], + "logprob_reward": -0.28135615587234497, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135615587234497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21800640225410461, + 0.06519093364477158, + 0.07052508741617203, + 0.06098363175988197, + 0.18403035402297974, + 0.08701639622449875, + 0.09900441765785217, + 0.1522960066795349, + 0.032183513045310974, + 0.03076329454779625 + ], + "credit_map": [ + 0.028204087167978287, + 0.017777228727936745, + 0.033496249467134476, + 0.03793025016784668, + 0.028549380600452423, + 0.02631431445479393, + 0.02225480228662491, + 0.023480085656046867, + 0.007409829180687666, + 0.005934945773333311, + 0.009460752829909325, + 0.00907194148749113, + 0.0070461975410580635, + 0.010550401173532009, + 0.006505629047751427, + 0.009211236611008644, + 0.00850257370620966, + 0.004077491816133261, + 0.00938331987708807, + 0.005694596562534571, + 0.008627718314528465, + 0.010663103312253952, + 0.015690665692090988, + 0.007885617204010487, + 0.005730065982788801, + 0.00603186571970582, + 0.005444776266813278, + 0.008086531423032284, + 0.015197275206446648, + 0.007454372942447662, + 0.0071532027795910835, + 0.005885538645088673, + 0.007351877633482218, + 0.006842923350632191, + 0.012087547220289707, + 0.01563130132853985, + 0.008571667596697807, + 0.04902160167694092, + 0.01613553613424301, + 0.06838787347078323, + 0.007117324974387884, + 0.005770024377852678, + 0.011173667386174202, + 0.01152684073895216, + 0.00753254909068346, + 0.00707746297121048, + 0.01757805235683918, + 0.019240478053689003, + 0.007949404418468475, + 0.014545471407473087, + 0.010661435313522816, + 0.015426439233124256, + 0.01834893971681595, + 0.009054061025381088, + 0.013188757933676243, + 0.009829908609390259, + 0.01344107836484909, + 0.009868263266980648, + 0.03718358650803566, + 0.00825931690633297, + 0.009869826957583427, + 0.013228807598352432, + 0.01293854508548975, + 0.04750657454133034, + 0.0037605587858706713, + 0.004919153172522783, + 0.004365903325378895, + 0.002911025658249855, + 0.003867718391120434, + 0.0038509105797857046, + 0.005253911949694157, + 0.0032543297857046127, + 0.002256748965010047, + 0.003132851095870137, + 0.005744872149080038, + 0.005100161302834749, + 0.0034590051509439945, + 0.004001110326498747, + 0.003607736201956868, + 0.0034608077257871628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.28135615587234497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810355976223946 + ], + "opposite_logprob_tokens": [ + -3.5298104286193848 + ], + "logprob_reward": -0.029810355976223946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810355976223946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19699768722057343, + 0.06108560040593147, + 0.0586739107966423, + 0.026364633813500404, + 0.05596127733588219, + 0.1738038808107376, + 0.15024331212043762, + 0.07299033552408218, + 0.060748547315597534, + 0.14313086867332458 + ], + "credit_map": [ + 0.02086055651307106, + 0.015032676979899406, + 0.026840148493647575, + 0.02872444875538349, + 0.02964347042143345, + 0.024969343096017838, + 0.02837446704506874, + 0.02255259081721306, + 0.010375545360147953, + 0.005782171152532101, + 0.007467711344361305, + 0.00922075379639864, + 0.005335431080311537, + 0.0068863616324961185, + 0.00833092164248228, + 0.0076867020688951015, + 0.006315043196082115, + 0.004603599198162556, + 0.007487123366445303, + 0.005597967188805342, + 0.006370502524077892, + 0.008091287687420845, + 0.011794009245932102, + 0.008414376527071, + 0.0035303833428770304, + 0.002866538939997554, + 0.0027229897677898407, + 0.0037145507521927357, + 0.0032856136094778776, + 0.003014521673321724, + 0.003161284141242504, + 0.004068753216415644, + 0.003901445772498846, + 0.0040710726752877235, + 0.005420821253210306, + 0.0046793208457529545, + 0.002368945861235261, + 0.011635121889412403, + 0.005040230695158243, + 0.01884431578218937, + 0.017978636547923088, + 0.020572196692228317, + 0.022782685235142708, + 0.03407152742147446, + 0.018035318702459335, + 0.01591828465461731, + 0.02102227881550789, + 0.023422932252287865, + 0.014751189388334751, + 0.02703992836177349, + 0.015233714133501053, + 0.024976523593068123, + 0.022235695272684097, + 0.01159563660621643, + 0.019536525011062622, + 0.014874094165861607, + 0.008644764311611652, + 0.009956319816410542, + 0.008849135600030422, + 0.008088274858891964, + 0.01100988406687975, + 0.007732067257165909, + 0.008897362276911736, + 0.009812531992793083, + 0.006557008251547813, + 0.007775579579174519, + 0.009102754294872284, + 0.005802790634334087, + 0.005543852690607309, + 0.0089871846139431, + 0.007703387178480625, + 0.00927598774433136, + 0.010604304261505604, + 0.010472369380295277, + 0.024608293548226357, + 0.014275407418608665, + 0.013270145282149315, + 0.037771351635456085, + 0.014946107752621174, + 0.017182882875204086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.029810355976223946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306088760495186 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306088760495186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306088760495186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10579778999090195, + 0.03266307711601257, + 0.03057919070124626, + 0.033138781785964966, + 0.12726356089115143, + 0.0992351621389389, + 0.04206222668290138, + 0.39333871006965637, + 0.09214424341917038, + 0.043777357786893845 + ], + "credit_map": [ + 0.013715619221329689, + 0.009676171466708183, + 0.011653381399810314, + 0.01423915009945631, + 0.013919271528720856, + 0.012663952074944973, + 0.016926055774092674, + 0.013004171662032604, + 0.0038471168372780085, + 0.0026513554621487856, + 0.0039084297604858875, + 0.0038073568139225245, + 0.0038725275080651045, + 0.0066232262179255486, + 0.0034232912585139275, + 0.004529770463705063, + 0.0020112781785428524, + 0.0016244726721197367, + 0.004049711395055056, + 0.0025895223952829838, + 0.0040568336844444275, + 0.005039939656853676, + 0.0069095357321202755, + 0.00429789163172245, + 0.003056757152080536, + 0.003680306253954768, + 0.0032162938732653856, + 0.003953240346163511, + 0.0036305247340351343, + 0.00396482041105628, + 0.004706254228949547, + 0.006930583156645298, + 0.005206909496337175, + 0.004055390600115061, + 0.010107947513461113, + 0.009317517280578613, + 0.005206012632697821, + 0.03257140889763832, + 0.01138992514461279, + 0.0494084358215332, + 0.011438611894845963, + 0.008967967703938484, + 0.014087283052504063, + 0.016761023551225662, + 0.008656601421535015, + 0.007068284787237644, + 0.016265640035271645, + 0.015989746898412704, + 0.0037017259746789932, + 0.007095102220773697, + 0.0041456399485468864, + 0.005880940239876509, + 0.006257577799260616, + 0.004119083285331726, + 0.005750867072492838, + 0.005111285485327244, + 0.04107518121600151, + 0.017963219434022903, + 0.10467927157878876, + 0.015820618718862534, + 0.026212094351649284, + 0.03799237310886383, + 0.022603431716561317, + 0.12699249386787415, + 0.010609665885567665, + 0.013875922188162804, + 0.013023782521486282, + 0.009544069878757, + 0.010209967382252216, + 0.01013064943253994, + 0.00808144360780716, + 0.016668735072016716, + 0.003581664292141795, + 0.0056358822621405125, + 0.006064582150429487, + 0.005881536286324263, + 0.005060323514044285, + 0.006835275795310736, + 0.006247605197131634, + 0.004470482934266329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.04306088760495186 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful close-up, captured from a slightly upward angle and accompanied by a slight handheld camera shake, locks the viewer's gaze on a hand, its knuckles white from extreme exertion. The back of the hand, stained with dust and minute scratches, bears the marks of intense combat. The hand tightly grips a delicate silver hairpin, its head carved with intricate classical patterns, yet in the dim light, its sharp tip reflects a deadly cold gleam. With a resolute and swift thrust, the wrist thrusts forward, unhesitatingly and decisively piercing the opponent in the foreground. The background is blurred by motion, revealing only the tangled folds of clothing and the rising dust. A narrow beam of cold light slants in from the side and above, precisely outlining the taut muscles of the arm and leaving a brief bright mark on the trajectory of the metal hairpin tip cutting through the air. The entire picture is gloomy in tone, full of critical tension and primitive violent aesthetics, freezing the desperate moment of resistance in this fierce fight.", + "seed": 27737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_11/index_3581_seed_27737_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13151477053761482, + "vlm_logprob": -0.13151477053761482, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.13151477053761482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035791318863630295 + ], + "opposite_logprob_tokens": [ + -3.5357913970947266 + ], + "logprob_reward": -0.035791318863630295, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035791318863630295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31498804688453674, + 0.13799072802066803, + 0.0729840025305748, + 0.05600222200155258, + 0.11110076308250427, + 0.05121889337897301, + 0.10766460746526718, + 0.06309723854064941, + 0.049687959253787994, + 0.0352654792368412 + ], + "credit_map": [ + 0.024482756853103638, + 0.021302662789821625, + 0.03774380311369896, + 0.07234977930784225, + 0.04461166262626648, + 0.03017597645521164, + 0.038397300988435745, + 0.045924123376607895, + 0.010601325891911983, + 0.011070081032812595, + 0.016858654096722603, + 0.0128321573138237, + 0.02248033508658409, + 0.018905295059084892, + 0.019642572849988937, + 0.025600308552384377, + 0.007638862822204828, + 0.006155212875455618, + 0.009781382977962494, + 0.011126172728836536, + 0.009654974564909935, + 0.00744925020262599, + 0.007686182390898466, + 0.013491969555616379, + 0.003719004336744547, + 0.004495914094150066, + 0.006570284720510244, + 0.006830788217484951, + 0.009279755875468254, + 0.010633427649736404, + 0.007029639091342688, + 0.0074434117414057255, + 0.011357852257788181, + 0.015980824828147888, + 0.011646708473563194, + 0.016158750280737877, + 0.01953244023025036, + 0.01544604729861021, + 0.01122460886836052, + 0.00975353829562664, + 0.0035375880543142557, + 0.006925010122358799, + 0.005283601116389036, + 0.00671831239014864, + 0.005505003500729799, + 0.005903321318328381, + 0.00816070381551981, + 0.009185356087982655, + 0.010678423568606377, + 0.01727166399359703, + 0.012203115969896317, + 0.011198424734175205, + 0.01999915950000286, + 0.012251408770680428, + 0.009653134271502495, + 0.014409278519451618, + 0.006770337000489235, + 0.005601321347057819, + 0.008455452509224415, + 0.005345387849956751, + 0.00564339617267251, + 0.01573837175965309, + 0.007503594271838665, + 0.008039383217692375, + 0.004267143085598946, + 0.00451308861374855, + 0.009754438884556293, + 0.007412563543766737, + 0.004871353507041931, + 0.0070508141070604324, + 0.005177846644073725, + 0.006640717387199402, + 0.0037707402370870113, + 0.0033451439812779427, + 0.004912503529340029, + 0.005065804813057184, + 0.003601155010983348, + 0.006736022885888815, + 0.004071434028446674, + 0.0037626775447279215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.035791318863630295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604192852973938 + ], + "opposite_logprob_tokens": [ + -1.910419225692749 + ], + "logprob_reward": -0.1604192852973938, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604192852973938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17837749421596527, + 0.18952739238739014, + 0.12637707591056824, + 0.04764784872531891, + 0.05593768134713173, + 0.14964698255062103, + 0.050528425723314285, + 0.06410115957260132, + 0.07298627495765686, + 0.06486962735652924 + ], + "credit_map": [ + 0.01713779754936695, + 0.01434127613902092, + 0.01518095750361681, + 0.03912995383143425, + 0.0261827539652586, + 0.018859295174479485, + 0.02013876847922802, + 0.027406703680753708, + 0.019121984019875526, + 0.015306580811738968, + 0.02952711284160614, + 0.023933589458465576, + 0.030479367822408676, + 0.021607670933008194, + 0.023115258663892746, + 0.026435822248458862, + 0.012203761376440525, + 0.01040143333375454, + 0.017703354358673096, + 0.023455999791622162, + 0.015215485356748104, + 0.012803292833268642, + 0.011176006868481636, + 0.023417744785547256, + 0.0030016994569450617, + 0.0036812429316341877, + 0.004915882833302021, + 0.00496688112616539, + 0.008615563623607159, + 0.007617101073265076, + 0.0067251455038785934, + 0.0081243347376585, + 0.005473506636917591, + 0.008233152329921722, + 0.005528667941689491, + 0.008050338365137577, + 0.010172535665333271, + 0.006166073959320784, + 0.006827781442552805, + 0.005485625006258488, + 0.014276212081313133, + 0.018116140738129616, + 0.018843626603484154, + 0.019742565229535103, + 0.017950167879462242, + 0.018616531044244766, + 0.019482791423797607, + 0.022618958726525307, + 0.005372592713683844, + 0.007730428129434586, + 0.00805627554655075, + 0.005632489454001188, + 0.007158391177654266, + 0.005281449761241674, + 0.005431779660284519, + 0.005865019280463457, + 0.007713199593126774, + 0.007233164738863707, + 0.008904430083930492, + 0.005414078012108803, + 0.006288100965321064, + 0.01546488981693983, + 0.008024930953979492, + 0.005058365408331156, + 0.00637427531182766, + 0.005539810750633478, + 0.008039379492402077, + 0.008408109657466412, + 0.006219425704330206, + 0.010108509100973606, + 0.010800301097333431, + 0.01749645732343197, + 0.005223607178777456, + 0.0062235998921096325, + 0.006062126252800226, + 0.006485539488494396, + 0.007794814649969339, + 0.020640702918171883, + 0.0050801620818674564, + 0.007359078153967857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.1604192852973938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865784645080566 + ], + "opposite_logprob_tokens": [ + -1.2236578464508057 + ], + "logprob_reward": -0.34865784645080566, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865784645080566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07491052150726318, + 0.07379631698131561, + 0.06781679391860962, + 0.09678011387586594, + 0.04711905121803284, + 0.09755110740661621, + 0.2604944109916687, + 0.10928480327129364, + 0.09979892522096634, + 0.07244788110256195 + ], + "credit_map": [ + 0.007647030055522919, + 0.006221150979399681, + 0.009536132216453552, + 0.012461061589419842, + 0.010727672837674618, + 0.00811986718326807, + 0.009207148104906082, + 0.01099045854061842, + 0.007999079301953316, + 0.005827984306961298, + 0.013422524556517601, + 0.010832211002707481, + 0.010854782536625862, + 0.006992402020841837, + 0.008758028037846088, + 0.00910931546241045, + 0.008117864839732647, + 0.005437091458588839, + 0.009313035756349564, + 0.010096787475049496, + 0.010247831232845783, + 0.006981220096349716, + 0.0066697439178824425, + 0.010953220538794994, + 0.007091343868523836, + 0.006698214448988438, + 0.010030428878962994, + 0.015040427446365356, + 0.013443962670862675, + 0.018079392611980438, + 0.012510612607002258, + 0.013885743916034698, + 0.0044221095740795135, + 0.006298679858446121, + 0.0049527352675795555, + 0.006166316103190184, + 0.008146243169903755, + 0.006092076655477285, + 0.006325339432805777, + 0.004715555347502232, + 0.0073197586461901665, + 0.010962280444800854, + 0.01046106405556202, + 0.016720052808523178, + 0.012228766456246376, + 0.011628498323261738, + 0.013918434269726276, + 0.01431225799024105, + 0.020920535549521446, + 0.02393551915884018, + 0.02296813577413559, + 0.030172917991876602, + 0.038556065410375595, + 0.046908944845199585, + 0.036989275366067886, + 0.04004303365945816, + 0.009019153192639351, + 0.008281229063868523, + 0.014327404089272022, + 0.010720656253397465, + 0.010167457163333893, + 0.02635197900235653, + 0.017672395333647728, + 0.012744530104100704, + 0.007467388175427914, + 0.009482420049607754, + 0.014231160283088684, + 0.01612098515033722, + 0.014137540943920612, + 0.013000539503991604, + 0.013373808935284615, + 0.01198509056121111, + 0.008579698391258717, + 0.006980718579143286, + 0.00787790771573782, + 0.008909598924219608, + 0.00739074032753706, + 0.017131825909018517, + 0.008898689411580563, + 0.006678710225969553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.34865784645080566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371838480234146 + ], + "opposite_logprob_tokens": [ + -3.4087183475494385 + ], + "logprob_reward": -0.03371838480234146, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371838480234146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15003745257854462, + 0.053871724754571915, + 0.09722154587507248, + 0.062183842062950134, + 0.1482304036617279, + 0.14021816849708557, + 0.05922641605138779, + 0.06771311908960342, + 0.1459115743637085, + 0.07538578659296036 + ], + "credit_map": [ + 0.018368897959589958, + 0.012834223918616772, + 0.016253408044576645, + 0.01993161253631115, + 0.02309572696685791, + 0.020726602524518967, + 0.018288390710949898, + 0.02053859271109104, + 0.007214164361357689, + 0.005164323840290308, + 0.008189949207007885, + 0.008302949368953705, + 0.006093158852308989, + 0.004834168124943972, + 0.006334586068987846, + 0.0077384281903505325, + 0.013485556468367577, + 0.008945653215050697, + 0.013794739730656147, + 0.01457936316728592, + 0.01148609071969986, + 0.009853770025074482, + 0.009088283404707909, + 0.015988098457455635, + 0.004714570473879576, + 0.005418790969997644, + 0.009852916933596134, + 0.009183697402477264, + 0.009658231399953365, + 0.00822613574564457, + 0.007366155739873648, + 0.007763337809592485, + 0.01759854145348072, + 0.0228189118206501, + 0.018705209717154503, + 0.016309460625052452, + 0.024943044409155846, + 0.019119394943118095, + 0.017247145995497704, + 0.011488695628941059, + 0.011336485855281353, + 0.01645665243268013, + 0.014153189025819302, + 0.019127005711197853, + 0.020246105268597603, + 0.021643323823809624, + 0.0190289169549942, + 0.018226487562060356, + 0.0045696101151406765, + 0.007566523738205433, + 0.006399468518793583, + 0.006233838386833668, + 0.008441964164376259, + 0.007877778261899948, + 0.00904457550495863, + 0.009092660620808601, + 0.007945495657622814, + 0.005941867828369141, + 0.008414068259298801, + 0.00663801608607173, + 0.00970059260725975, + 0.008925152942538261, + 0.010474760085344315, + 0.00967316422611475, + 0.016874754801392555, + 0.013769522309303284, + 0.014237089082598686, + 0.01551230251789093, + 0.014620413072407246, + 0.01434246264398098, + 0.02048691362142563, + 0.03606811910867691, + 0.006703394930809736, + 0.012875651940703392, + 0.010005730204284191, + 0.00844787061214447, + 0.008996219374239445, + 0.00922971311956644, + 0.008483218029141426, + 0.010643990710377693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03371838480234146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a hand gripping a silver hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown with white knuckles, dust stains, and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a swift forward thrust motion piercing an opponent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera angle slightly upward with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a narrow beam of cold light slanting from the side and above visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789870172739029 + ], + "opposite_logprob_tokens": [ + -2.5789871215820312 + ], + "logprob_reward": -0.0789870172739029, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789870172739029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04595406353473663, + 0.13995596766471863, + 0.047221336513757706, + 0.2803277373313904, + 0.09075307101011276, + 0.07499279081821442, + 0.07696320116519928, + 0.06582155078649521, + 0.11523083597421646, + 0.06277942657470703 + ], + "credit_map": [ + 0.0061116572469472885, + 0.005006454885005951, + 0.005163262598216534, + 0.007538340520113707, + 0.0061469231732189655, + 0.0039717513136565685, + 0.0060565657913684845, + 0.005959107540547848, + 0.016700848937034607, + 0.01407370064407587, + 0.02332838997244835, + 0.02108355052769184, + 0.015472844243049622, + 0.015226303599774837, + 0.018801013007760048, + 0.015269311144948006, + 0.003910092171281576, + 0.0031769711058586836, + 0.0061515700072050095, + 0.007865583524107933, + 0.0073408992029726505, + 0.0052884481847286224, + 0.004369067493826151, + 0.009118706919252872, + 0.015753429383039474, + 0.01857858896255493, + 0.028305090963840485, + 0.029483836144208908, + 0.04666474461555481, + 0.05842284485697746, + 0.03597576171159744, + 0.04714342579245567, + 0.009632835164666176, + 0.014325913973152637, + 0.011029992252588272, + 0.011522429995238781, + 0.016127392649650574, + 0.012276659719645977, + 0.00906168483197689, + 0.0067761666141450405, + 0.0062481435015797615, + 0.007670658640563488, + 0.0076360926032066345, + 0.00914595928043127, + 0.006678923033177853, + 0.00936728622764349, + 0.016142813488841057, + 0.012102916836738586, + 0.00648087402805686, + 0.0076825302094221115, + 0.007755008526146412, + 0.00856035202741623, + 0.01710342802107334, + 0.010656604543328285, + 0.008170695044100285, + 0.010553705506026745, + 0.005147109739482403, + 0.004784712567925453, + 0.007549057248979807, + 0.008002705872058868, + 0.006487254053354263, + 0.01362761203199625, + 0.012573232874274254, + 0.007649870589375496, + 0.00794709101319313, + 0.008957713842391968, + 0.013282805681228638, + 0.015663066878914833, + 0.016270650550723076, + 0.020576391369104385, + 0.016119081526994705, + 0.016414033249020576, + 0.005674173589795828, + 0.006329991854727268, + 0.00547278905287385, + 0.005705909803509712, + 0.007900303229689598, + 0.017751144245266914, + 0.0067279706709086895, + 0.007217143662273884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.0789870172739029 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_tensor([5871, 5871])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_tensor([5871, 5871])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..483f07f0437e4e37e042b672fd3a86b765ecbc04 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_tensor([5871, 5871])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle, long shot, shot upward from near sea level, slowly and steadily moving forward, captures a scene of apocalyptic grandeur. Beneath a sky tinged with leaden gray by the storm, a massive wave, like a moving mountain, roars toward the horizon. Its deep, dark green waters churn with pale, violent foam, carrying with them a primal force of destruction. Directly before it, a massive, weathered black reef, its surface covered in dark moss and jagged cracks, stands defiantly like an ancient watchtower amidst the surging sea. As the wave crashes against the reef with thunderous force, a deafening roar echoes, like the smiting of an invisible axe, stirring up a cloud of mist that soars into the sky. A dazzling crack instantly penetrated from the top of the reef to the underwater base. Then, under the suffocating pressure, the entire reef was split in half. The two huge fractured rock bodies collapsed to the sides and fell into the subsequent surging violent white waves and were instantly swallowed up. The whole picture was full of primitive, violent and awe-inspiring natural destructive power under the cold and depressing light, creating a tense and epic atmosphere.", + "seed": 5581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_5871_seed_5581_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10805359240621329, + "vlm_logprob": -0.10805359240621329, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.10805359240621329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07622038573026657 + ], + "opposite_logprob_tokens": [ + -3.2012202739715576 + ], + "logprob_reward": -0.07622038573026657, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07622038573026657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1545960158109665, + 0.07829120010137558, + 0.10078530758619308, + 0.07655385136604309, + 0.043299365788698196, + 0.04894298315048218, + 0.28784918785095215, + 0.03810212016105652, + 0.08597379177808762, + 0.08560624718666077 + ], + "credit_map": [ + 0.012329427525401115, + 0.014048143289983273, + 0.02224876545369625, + 0.018851419910788536, + 0.01183206494897604, + 0.015996074303984642, + 0.02436654269695282, + 0.03492356464266777, + 0.0058170463889837265, + 0.012073386460542679, + 0.00611940398812294, + 0.0067319911904633045, + 0.006715232506394386, + 0.009742938913404942, + 0.015511520206928253, + 0.015579672530293465, + 0.012530345469713211, + 0.01838923804461956, + 0.010822528973221779, + 0.009658846072852612, + 0.014218789525330067, + 0.010191897861659527, + 0.014797892421483994, + 0.010175767354667187, + 0.008323963731527328, + 0.008041719906032085, + 0.012474470771849155, + 0.015495053492486477, + 0.005992064252495766, + 0.0070715234614908695, + 0.010966657660901546, + 0.00818838831037283, + 0.004023298621177673, + 0.005559651646763086, + 0.004679391160607338, + 0.007734846323728561, + 0.0034110962878912687, + 0.0042261965572834015, + 0.008383985608816147, + 0.005280894692987204, + 0.005213945172727108, + 0.005955624394118786, + 0.003933559637516737, + 0.012016171589493752, + 0.003633626503869891, + 0.0063841454684734344, + 0.00790554191917181, + 0.0039003663696348667, + 0.008367566391825676, + 0.014683396555483341, + 0.007164004258811474, + 0.11774420738220215, + 0.00870671309530735, + 0.00838593952357769, + 0.10996472090482712, + 0.01283261738717556, + 0.00391001533716917, + 0.004432786721736193, + 0.005271703004837036, + 0.004093754570931196, + 0.00396971357986331, + 0.004694963339716196, + 0.005556640215218067, + 0.006172540597617626, + 0.00626965519040823, + 0.0034243958070874214, + 0.03131481632590294, + 0.0052086059004068375, + 0.002360242186114192, + 0.004318397492170334, + 0.004320270847529173, + 0.028757404536008835, + 0.00842979084700346, + 0.009688083082437515, + 0.015152549371123314, + 0.010333347134292126, + 0.00821300782263279, + 0.009662341326475143, + 0.013055608607828617, + 0.01107151247560978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.07622038573026657 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865909367799759 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865909367799759, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865909367799759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038373928517103195, + 0.014927041716873646, + 0.059029411524534225, + 0.03164668008685112, + 0.023156486451625824, + 0.031370777636766434, + 0.7093122005462646, + 0.04326586797833443, + 0.028123410418629646, + 0.020794203504920006 + ], + "credit_map": [ + 0.0038652527146041393, + 0.005894442554563284, + 0.004569214768707752, + 0.004162418190389872, + 0.005313584581017494, + 0.004804490599781275, + 0.004886896349489689, + 0.004877625033259392, + 0.0017326794331893325, + 0.0019141408847644925, + 0.0022210581228137016, + 0.0016373132821172476, + 0.001572784734889865, + 0.0009620327618904412, + 0.00266663427464664, + 0.0022203978151082993, + 0.0073866453021764755, + 0.012384746223688126, + 0.006352872122079134, + 0.006041327957063913, + 0.005766923073679209, + 0.00673022773116827, + 0.008432576432824135, + 0.0059340898878872395, + 0.00360593618825078, + 0.002466723555698991, + 0.0033139220904558897, + 0.008342238143086433, + 0.0018411827040836215, + 0.0025649110320955515, + 0.0063788448460400105, + 0.0031329216435551643, + 0.002297624945640564, + 0.002390126697719097, + 0.0029588753823190928, + 0.0047697508707642555, + 0.0022035129368305206, + 0.0017867239657789469, + 0.0032079287339001894, + 0.003541940590366721, + 0.0038016634061932564, + 0.0023257203865796328, + 0.0032679270952939987, + 0.004694854374974966, + 0.0019453326240181923, + 0.0038675551768392324, + 0.00501667195931077, + 0.00645105168223381, + 0.02126455307006836, + 0.04163448140025139, + 0.016417285427451134, + 0.2876160442829132, + 0.01712949387729168, + 0.022084826603531837, + 0.27182021737098694, + 0.03134527429938316, + 0.006960988510400057, + 0.005622317083179951, + 0.006393566261976957, + 0.004597565624862909, + 0.002885352587327361, + 0.0037690699100494385, + 0.0065138679929077625, + 0.0065231360495090485, + 0.0035650015342980623, + 0.0017984752776101232, + 0.0032981312833726406, + 0.0016082939691841602, + 0.001881130039691925, + 0.0028184980619698763, + 0.004319152794778347, + 0.008834723383188248, + 0.001912908279336989, + 0.0024542759638279676, + 0.003207779722288251, + 0.003470719326287508, + 0.0020825043320655823, + 0.00196393602527678, + 0.0027436567470431328, + 0.002958421129733324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.04865909367799759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870023488998413 + ], + "opposite_logprob_tokens": [ + -1.1370023488998413 + ], + "logprob_reward": -0.3870023488998413, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870023488998413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07024125754833221, + 0.017522839829325676, + 0.12010887265205383, + 0.048594579100608826, + 0.025406688451766968, + 0.07590046525001526, + 0.08628310263156891, + 0.030377520248293877, + 0.13609643280506134, + 0.38946816325187683 + ], + "credit_map": [ + 0.004769178573042154, + 0.006415056996047497, + 0.00980253517627716, + 0.008291615173220634, + 0.004794028587639332, + 0.00742953410372138, + 0.01393588725477457, + 0.014803427271544933, + 0.00180150312371552, + 0.002030171686783433, + 0.002458269940689206, + 0.0019769826903939247, + 0.0018505753250792623, + 0.0018031460931524634, + 0.0027118567377328873, + 0.002890335861593485, + 0.013441194780170918, + 0.009363187476992607, + 0.021655196323990822, + 0.01887086033821106, + 0.01157852541655302, + 0.011336552910506725, + 0.018838291987776756, + 0.015025079250335693, + 0.005323518067598343, + 0.004613864701241255, + 0.0041253347881138325, + 0.015030843205749989, + 0.003265007631853223, + 0.0026610693894326687, + 0.010222917422652245, + 0.0033520294819027185, + 0.0020820465870201588, + 0.0031538361217826605, + 0.003007507184520364, + 0.004318146035075188, + 0.0023475345224142075, + 0.0025276606902480125, + 0.00436052493751049, + 0.0036094356328248978, + 0.008983361534774303, + 0.007426764350384474, + 0.007682508789002895, + 0.01783239282667637, + 0.005658467300236225, + 0.006694392766803503, + 0.014314194209873676, + 0.007308396510779858, + 0.004337241407483816, + 0.00708432774990797, + 0.004299709107726812, + 0.025919822975993156, + 0.006797344423830509, + 0.004452557768672705, + 0.028904501348733902, + 0.004487613216042519, + 0.003512295661494136, + 0.0027851660270243883, + 0.002180721377953887, + 0.0028390646912157536, + 0.0019926284439861774, + 0.0024484493769705296, + 0.00781311746686697, + 0.006806079763919115, + 0.010700155049562454, + 0.006218921858817339, + 0.03517254814505577, + 0.00666075199842453, + 0.0050249407067894936, + 0.007928775623440742, + 0.013863063417375088, + 0.05052729696035385, + 0.036683112382888794, + 0.04337092489004135, + 0.0896516740322113, + 0.06899236887693405, + 0.04648604243993759, + 0.03298768028616905, + 0.04201788827776909, + 0.02927851676940918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.3870023488998413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08391392230987549, + 0.042325057089328766, + 0.10899151861667633, + 0.03468480706214905, + 0.08621644973754883, + 0.08516892790794373, + 0.17071524262428284, + 0.019149942323565483, + 0.06856009364128113, + 0.30027392506599426 + ], + "credit_map": [ + 0.007377924397587776, + 0.008044376038014889, + 0.012687350623309612, + 0.010678963735699654, + 0.012484906241297722, + 0.008759083226323128, + 0.012184684164822102, + 0.011696641333401203, + 0.004014299251139164, + 0.005046981852501631, + 0.005890841595828533, + 0.00497167743742466, + 0.004208615515381098, + 0.0050764842890203, + 0.006470984313637018, + 0.006645177025347948, + 0.013554482720792294, + 0.02012753114104271, + 0.016262173652648926, + 0.013272961601614952, + 0.013171621598303318, + 0.0067818667739629745, + 0.015286975540220737, + 0.01053391769528389, + 0.0035115645732730627, + 0.0036879258695989847, + 0.002956226235255599, + 0.008734043687582016, + 0.0024498968850821257, + 0.002615265315398574, + 0.007877998985350132, + 0.002851887373253703, + 0.008052974939346313, + 0.009300713427364826, + 0.008252608589828014, + 0.014514691196382046, + 0.009571878239512444, + 0.008447734639048576, + 0.015133661217987537, + 0.01294220145791769, + 0.0087735028937459, + 0.007458162494003773, + 0.0072570014744997025, + 0.017662901431322098, + 0.005489993374794722, + 0.00891873612999916, + 0.017651882022619247, + 0.011956758797168732, + 0.00886770710349083, + 0.007917331531643867, + 0.010417768731713295, + 0.04350002110004425, + 0.014572560787200928, + 0.013637710362672806, + 0.06111795827746391, + 0.010684218257665634, + 0.0026043886318802834, + 0.0022688182070851326, + 0.0029846145771443844, + 0.0029877990018576384, + 0.001638192916288972, + 0.001551143592223525, + 0.002174082910642028, + 0.0029409024864435196, + 0.006102777551859617, + 0.0056108999997377396, + 0.008542606607079506, + 0.0031713100615888834, + 0.0028740749694406986, + 0.004787420853972435, + 0.01592724584043026, + 0.02154376171529293, + 0.025723718106746674, + 0.04044805094599724, + 0.060146499425172806, + 0.06033145636320114, + 0.03512856364250183, + 0.022589679807424545, + 0.0366940014064312, + 0.019212007522583008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.014193184673786163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06139818951487541, + 0.012629348784685135, + 0.03348775580525398, + 0.02919091284275055, + 0.061624281108379364, + 0.046729929745197296, + 0.08842834830284119, + 0.027892962098121643, + 0.2970558702945709, + 0.3415623903274536 + ], + "credit_map": [ + 0.006141808349639177, + 0.006810306571424007, + 0.006950599141418934, + 0.007351788226515055, + 0.007844158448278904, + 0.009099479764699936, + 0.009564366191625595, + 0.00763568002730608, + 0.0009504900663159788, + 0.0012821040581911802, + 0.0013999835355207324, + 0.0010121907107532024, + 0.0013789933873340487, + 0.0013469085097312927, + 0.0024002506397664547, + 0.002858428517356515, + 0.004211160354316235, + 0.005205492489039898, + 0.004294184036552906, + 0.004082097206264734, + 0.0035789660178124905, + 0.004004387650638819, + 0.0047876168973743916, + 0.0033238520845770836, + 0.0031035703141242266, + 0.0036059212870895863, + 0.0026164723094552755, + 0.00683573679998517, + 0.0023998492397367954, + 0.0027564328629523516, + 0.005171249154955149, + 0.002701680175960064, + 0.005713911727070808, + 0.007636202499270439, + 0.006807295605540276, + 0.011419721879065037, + 0.0048757800832390785, + 0.006181470584124327, + 0.01139112189412117, + 0.007598779629915953, + 0.006455466616898775, + 0.004844704642891884, + 0.003377101384103298, + 0.010544396005570889, + 0.0035179397091269493, + 0.0050829569809138775, + 0.00852594431489706, + 0.00438141543418169, + 0.00413642730563879, + 0.0050343843176960945, + 0.004734918940812349, + 0.032685309648513794, + 0.0049606021493673325, + 0.0035068721044808626, + 0.0290396548807621, + 0.004330182448029518, + 0.003996516112238169, + 0.003729318268597126, + 0.003322065342217684, + 0.0026023576501756907, + 0.001923923147842288, + 0.0033314689062535763, + 0.005041070282459259, + 0.003946241922676563, + 0.020833922550082207, + 0.012591180391609669, + 0.08139218389987946, + 0.014206357300281525, + 0.009224031120538712, + 0.015430675819516182, + 0.0338602252304554, + 0.10951729118824005, + 0.029622923582792282, + 0.04488326236605644, + 0.06895427405834198, + 0.06940504908561707, + 0.03925243765115738, + 0.02509109303355217, + 0.04238894209265709, + 0.021964404731988907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.014192949049174786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot, shot upward from near sea level, slowly and steadily moving forward, captures a scene of apocalyptic grandeur. Beneath a sky tinged with leaden gray by the storm, a massive wave, like a moving mountain, roars toward the horizon. Its deep, dark green waters churn with pale, violent foam, carrying with them a primal force of destruction. Directly before it, a massive, weathered black reef, its surface covered in dark moss and jagged cracks, stands defiantly like an ancient watchtower amidst the surging sea. As the wave crashes against the reef with thunderous force, a deafening roar echoes, like the smiting of an invisible axe, stirring up a cloud of mist that soars into the sky. A dazzling crack instantly penetrated from the top of the reef to the underwater base. Then, under the suffocating pressure, the entire reef was split in half. The two huge fractured rock bodies collapsed to the sides and fell into the subsequent surging violent white waves and were instantly swallowed up. The whole picture was full of primitive, violent and awe-inspiring natural destructive power under the cold and depressing light, creating a tense and epic atmosphere.", + "seed": 5582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_5871_seed_5582_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18046427788212896, + "vlm_logprob": -0.18046427788212896, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.18046427788212896, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11169866472482681 + ], + "opposite_logprob_tokens": [ + -2.611698627471924 + ], + "logprob_reward": -0.11169866472482681, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.11169866472482681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12075003236532211, + 0.025695793330669403, + 0.056364525109529495, + 0.39976027607917786, + 0.08705301582813263, + 0.07901967316865921, + 0.16452869772911072, + 0.01832173578441143, + 0.02978484332561493, + 0.0187214445322752 + ], + "credit_map": [ + 0.02337951771914959, + 0.020403916016221046, + 0.011196172796189785, + 0.01056091021746397, + 0.015425092540681362, + 0.013633822090923786, + 0.019585300236940384, + 0.006565294228494167, + 0.004129249602556229, + 0.004525207448750734, + 0.002703659934923053, + 0.002206307603046298, + 0.003408415475860238, + 0.002564331516623497, + 0.00391200790181756, + 0.0022466147784143686, + 0.010389773175120354, + 0.006281447131186724, + 0.00597233884036541, + 0.004485570825636387, + 0.00579372001811862, + 0.008567197248339653, + 0.010217351838946342, + 0.0046571302227675915, + 0.04075777903199196, + 0.07246909290552139, + 0.0379926972091198, + 0.07605918496847153, + 0.08993896096944809, + 0.03052157163619995, + 0.027418384328484535, + 0.02460259571671486, + 0.010514256544411182, + 0.009318260475993156, + 0.010897664353251457, + 0.01941036991775036, + 0.007125843316316605, + 0.0069503760896623135, + 0.01107113528996706, + 0.01176510937511921, + 0.009811429306864738, + 0.008657733909785748, + 0.011196047067642212, + 0.009504583664238453, + 0.006876822095364332, + 0.009805857203900814, + 0.010164795443415642, + 0.013002409599721432, + 0.020015936344861984, + 0.022116027772426605, + 0.01682911440730095, + 0.02803049050271511, + 0.00966649129986763, + 0.010898396372795105, + 0.024000713601708412, + 0.032971519976854324, + 0.002345068147405982, + 0.0021456300746649504, + 0.003939073067158461, + 0.002078742254525423, + 0.0018853407818824053, + 0.0024797534570097923, + 0.001654505031183362, + 0.001793622155673802, + 0.004777344409376383, + 0.0027816048823297024, + 0.009013969451189041, + 0.0028213858604431152, + 0.0015135055873543024, + 0.0021014546509832144, + 0.001963081769645214, + 0.004812495782971382, + 0.002715730108320713, + 0.0020212542731314898, + 0.0021787125151604414, + 0.003269452601671219, + 0.0015814568614587188, + 0.0024474042002111673, + 0.0012302302056923509, + 0.003277204930782318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.11169866472482681 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001636922359467 + ], + "opposite_logprob_tokens": [ + -2.695016384124756 + ], + "logprob_reward": -0.07001636922359467, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001636922359467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2102084904909134, + 0.06338179111480713, + 0.10109929740428925, + 0.2359621673822403, + 0.06730964034795761, + 0.04272739961743355, + 0.08219439536333084, + 0.08036290109157562, + 0.03661183640360832, + 0.08014199882745743 + ], + "credit_map": [ + 0.029835470020771027, + 0.03680669516324997, + 0.023863397538661957, + 0.025012820959091187, + 0.022732079029083252, + 0.03527050092816353, + 0.024467086419463158, + 0.01222045999020338, + 0.010639027692377567, + 0.006882172077894211, + 0.005288237240165472, + 0.005103702191263437, + 0.013112038373947144, + 0.01018174085766077, + 0.00710478937253356, + 0.005070091225206852, + 0.016663096845149994, + 0.012628759257495403, + 0.01276627741754055, + 0.010495856404304504, + 0.011143025942146778, + 0.013311944901943207, + 0.012079283595085144, + 0.012011061422526836, + 0.019293859601020813, + 0.03100627288222313, + 0.01691988669335842, + 0.038396645337343216, + 0.056356821209192276, + 0.033395521342754364, + 0.01306152157485485, + 0.02753165178000927, + 0.007293623872101307, + 0.007313229143619537, + 0.009077753871679306, + 0.014483140781521797, + 0.005458301864564419, + 0.006046437192708254, + 0.011362435296177864, + 0.006274725776165724, + 0.005790987052023411, + 0.006784610915929079, + 0.006386453285813332, + 0.004400014877319336, + 0.0038358327001333237, + 0.005234065931290388, + 0.00384233589284122, + 0.006453103385865688, + 0.013510272838175297, + 0.017736971378326416, + 0.008369026705622673, + 0.009303282015025616, + 0.007466054055839777, + 0.007333121728152037, + 0.008603179827332497, + 0.009872490540146828, + 0.011424719356000423, + 0.01171350572258234, + 0.016082335263490677, + 0.00765429949387908, + 0.009436612948775291, + 0.007081763353198767, + 0.008465412072837353, + 0.008504250086843967, + 0.005403910297900438, + 0.005421169567853212, + 0.010421882383525372, + 0.002924964064732194, + 0.0023019665386527777, + 0.0023009495344012976, + 0.002423696918413043, + 0.0054132998920977116, + 0.010369042865931988, + 0.008749782107770443, + 0.01633770577609539, + 0.010770785622298717, + 0.0059267207980155945, + 0.01340331882238388, + 0.004643172491341829, + 0.009941481053829193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.07001636922359467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6932950019836426 + ], + "opposite_logprob_tokens": [ + -0.6932950019836426 + ], + "logprob_reward": -0.6932950019836426, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6932950019836426, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1557733714580536, + 0.07168057560920715, + 0.07637889683246613, + 0.09942254424095154, + 0.13987113535404205, + 0.13198018074035645, + 0.25134485960006714, + 0.035055678337812424, + 0.015953250229358673, + 0.022539513185620308 + ], + "credit_map": [ + 0.022865789011120796, + 0.02247116155922413, + 0.018918367102742195, + 0.016617827117443085, + 0.022122574970126152, + 0.01983932964503765, + 0.01927342638373375, + 0.013664892874658108, + 0.009487207978963852, + 0.010495724156498909, + 0.006433370057493448, + 0.007842830382287502, + 0.010040615685284138, + 0.009918322786688805, + 0.008233112283051014, + 0.009229391813278198, + 0.008651820942759514, + 0.010782426223158836, + 0.010533210821449757, + 0.011068028397858143, + 0.009845806285738945, + 0.011969907209277153, + 0.005837743636220694, + 0.007689950056374073, + 0.011157702654600143, + 0.02123476006090641, + 0.011109692975878716, + 0.010759064927697182, + 0.017865166068077087, + 0.014739828184247017, + 0.006181690841913223, + 0.006374645046889782, + 0.016137752681970596, + 0.014697521924972534, + 0.01949254609644413, + 0.022220486775040627, + 0.0167539045214653, + 0.019985904917120934, + 0.014242371544241905, + 0.016340646892786026, + 0.01654243655502796, + 0.016883600503206253, + 0.016569586470723152, + 0.01273423247039318, + 0.009349190630018711, + 0.013299120590090752, + 0.01836269348859787, + 0.028239328414201736, + 0.02875235117971897, + 0.024051224812865257, + 0.026267336681485176, + 0.048679765313863754, + 0.012838467955589294, + 0.01646457612514496, + 0.038031045347452164, + 0.05626009404659271, + 0.004675671458244324, + 0.005851518362760544, + 0.007157151587307453, + 0.0030206176452338696, + 0.004765176214277744, + 0.0028297423850744963, + 0.0033577794674783945, + 0.003398023545742035, + 0.0024064755998551846, + 0.0025598760694265366, + 0.004226971417665482, + 0.001661007758229971, + 0.0011848791036754847, + 0.001041628303937614, + 0.0010891450801864266, + 0.0017832680605351925, + 0.0029893568716943264, + 0.003413903061300516, + 0.0022544099483639, + 0.003819490782916546, + 0.0015206739772111177, + 0.0028664846904575825, + 0.0012357617961242795, + 0.0044394307769834995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.6932950019836426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734768394380808 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734768394380808, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734768394380808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19177640974521637, + 0.04123421385884285, + 0.13228997588157654, + 0.10593605786561966, + 0.0398046150803566, + 0.05864885076880455, + 0.29729604721069336, + 0.0362478606402874, + 0.031506747007369995, + 0.06525921821594238 + ], + "credit_map": [ + 0.02304116077721119, + 0.025684460997581482, + 0.02155313640832901, + 0.017460964620113373, + 0.03311758488416672, + 0.03869272395968437, + 0.01860233210027218, + 0.01362405065447092, + 0.004142086952924728, + 0.004917438607662916, + 0.004531302023679018, + 0.00428243400529027, + 0.006570016499608755, + 0.008113267831504345, + 0.005174045916646719, + 0.003503624349832535, + 0.022664010524749756, + 0.014279778115451336, + 0.014223936013877392, + 0.012073224410414696, + 0.016179893165826797, + 0.01806054450571537, + 0.020483093336224556, + 0.014325491152703762, + 0.009780559688806534, + 0.019850414246320724, + 0.011813094839453697, + 0.015752319246530533, + 0.029030833393335342, + 0.00979139655828476, + 0.005000574979931116, + 0.004916872829198837, + 0.005106725264340639, + 0.005538164172321558, + 0.004315261263400316, + 0.00723115261644125, + 0.0038558747619390488, + 0.003814644180238247, + 0.0059973131865262985, + 0.003945484291762114, + 0.0076865507289767265, + 0.007074957247823477, + 0.005965224467217922, + 0.006619525142014027, + 0.004078566562384367, + 0.005910681560635567, + 0.007529882714152336, + 0.013783467002213001, + 0.033494383096694946, + 0.031299080699682236, + 0.030368946492671967, + 0.05390491709113121, + 0.017692990601062775, + 0.023115461692214012, + 0.04319457709789276, + 0.0642256885766983, + 0.005727163515985012, + 0.004994995892047882, + 0.008497179485857487, + 0.0029059716034680605, + 0.0034029546659439802, + 0.0035713911056518555, + 0.0031952254939824343, + 0.0039529805071651936, + 0.004445534199476242, + 0.0032569472678005695, + 0.009784402325749397, + 0.004061034880578518, + 0.0020526971202343702, + 0.00208193133585155, + 0.0022301445715129375, + 0.0035940571688115597, + 0.008588626980781555, + 0.005888782441616058, + 0.013890095986425877, + 0.00859986711293459, + 0.004384180530905724, + 0.012665399350225925, + 0.0038154267240315676, + 0.007426838390529156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.006734768394380808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576585084199905 + ], + "opposite_logprob_tokens": [ + -3.8955764770507812 + ], + "logprob_reward": -0.020576585084199905, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576585084199905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16778534650802612, + 0.03616457059979439, + 0.06961262971162796, + 0.04703430458903313, + 0.026451904326677322, + 0.03558877855539322, + 0.35655486583709717, + 0.05625177547335625, + 0.09274643659591675, + 0.11180935055017471 + ], + "credit_map": [ + 0.02064313180744648, + 0.02099393494427204, + 0.02114938572049141, + 0.015597337856888771, + 0.028616903349757195, + 0.03250898793339729, + 0.01580716110765934, + 0.012468480505049229, + 0.004499680828303099, + 0.004715043120086193, + 0.004142837133258581, + 0.005034140311181545, + 0.005064935423433781, + 0.0032196994870901108, + 0.004955822601914406, + 0.004532408434897661, + 0.011098727583885193, + 0.007130175828933716, + 0.007510577794164419, + 0.006974825635552406, + 0.007706087548285723, + 0.008032294921576977, + 0.012734315358102322, + 0.008425620384514332, + 0.005971304606646299, + 0.0059548248536884785, + 0.006876795552670956, + 0.009187653660774231, + 0.005743153393268585, + 0.0037886928766965866, + 0.00468064658343792, + 0.004831230267882347, + 0.004049848299473524, + 0.00407036580145359, + 0.0026966675650328398, + 0.002421024953946471, + 0.004399134311825037, + 0.004684458952397108, + 0.001991216093301773, + 0.0021391885820776224, + 0.0057822661474347115, + 0.0051006353460252285, + 0.004034963436424732, + 0.0036476184614002705, + 0.0036177728325128555, + 0.005571956746280193, + 0.004271808546036482, + 0.00356175540946424, + 0.03945363312959671, + 0.03053983673453331, + 0.037636805325746536, + 0.07050036638975143, + 0.019627980887889862, + 0.024544065818190575, + 0.05351588875055313, + 0.08073628693819046, + 0.008423155173659325, + 0.007678459398448467, + 0.010718111880123615, + 0.004424775019288063, + 0.007595982868224382, + 0.005905698984861374, + 0.004807824268937111, + 0.006697765551507473, + 0.013844623230397701, + 0.008397436700761318, + 0.03056701086461544, + 0.010368895716965199, + 0.0057433294132351875, + 0.007991430349647999, + 0.005876713432371616, + 0.009956996887922287, + 0.014564407989382744, + 0.010172863490879536, + 0.022637274116277695, + 0.015023413114249706, + 0.00811050459742546, + 0.021133681759238243, + 0.007572095841169357, + 0.012595105916261673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.020576585084199905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot, shot upward from near sea level, slowly and steadily moving forward, captures a scene of apocalyptic grandeur. Beneath a sky tinged with leaden gray by the storm, a massive wave, like a moving mountain, roars toward the horizon. Its deep, dark green waters churn with pale, violent foam, carrying with them a primal force of destruction. Directly before it, a massive, weathered black reef, its surface covered in dark moss and jagged cracks, stands defiantly like an ancient watchtower amidst the surging sea. As the wave crashes against the reef with thunderous force, a deafening roar echoes, like the smiting of an invisible axe, stirring up a cloud of mist that soars into the sky. A dazzling crack instantly penetrated from the top of the reef to the underwater base. Then, under the suffocating pressure, the entire reef was split in half. The two huge fractured rock bodies collapsed to the sides and fell into the subsequent surging violent white waves and were instantly swallowed up. The whole picture was full of primitive, violent and awe-inspiring natural destructive power under the cold and depressing light, creating a tense and epic atmosphere.", + "seed": 5583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_5871_seed_5583_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1471615206450224, + "vlm_logprob": -0.1471615206450224, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1471615206450224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043834079056978226 + ], + "opposite_logprob_tokens": [ + -3.9188339710235596 + ], + "logprob_reward": -0.043834079056978226, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.043834079056978226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46637043356895447, + 0.053750187158584595, + 0.1247565969824791, + 0.05753115937113762, + 0.024778861552476883, + 0.08905475586652756, + 0.06098536401987076, + 0.04253353551030159, + 0.04246089980006218, + 0.03777821362018585 + ], + "credit_map": [ + 0.08440613001585007, + 0.035375263541936874, + 0.019772563129663467, + 0.028022730723023415, + 0.04255510866641998, + 0.17187485098838806, + 0.03455318883061409, + 0.049810584634542465, + 0.007611083332449198, + 0.013919242657721043, + 0.004564163740724325, + 0.004863408859819174, + 0.0094396211206913, + 0.004057634621858597, + 0.0051452480256557465, + 0.004149783402681351, + 0.015724482014775276, + 0.032090380787849426, + 0.011470557190477848, + 0.009904617443680763, + 0.017517685890197754, + 0.02190379984676838, + 0.008538325317203999, + 0.007606754545122385, + 0.00466930354014039, + 0.00618281913921237, + 0.008355668745934963, + 0.007949442602694035, + 0.00700594624504447, + 0.009498999454081059, + 0.00557024497538805, + 0.008298736996948719, + 0.002432000357657671, + 0.0041006808169186115, + 0.0023893569596111774, + 0.0031344944145530462, + 0.002851952565833926, + 0.003751537064090371, + 0.002657186007127166, + 0.0034616526681929827, + 0.015981154516339302, + 0.006106248591095209, + 0.006131981965154409, + 0.007740566972643137, + 0.00855760183185339, + 0.005667609628289938, + 0.009492132812738419, + 0.029377466067671776, + 0.004967878106981516, + 0.004546424373984337, + 0.011712846346199512, + 0.010006421245634556, + 0.00982220284640789, + 0.005820383317768574, + 0.0072606937028467655, + 0.00684851361438632, + 0.0029860185459256172, + 0.00396736478433013, + 0.0028906078077852726, + 0.006503828801214695, + 0.005768448580056429, + 0.0029683769680559635, + 0.013849781826138496, + 0.0035991084296256304, + 0.003987773787230253, + 0.005363998934626579, + 0.0038646520115435123, + 0.00942233856767416, + 0.0038441873621195555, + 0.005337263457477093, + 0.006962706334888935, + 0.0036779805086553097, + 0.004039844498038292, + 0.004334353841841221, + 0.0028892136178910732, + 0.004001460503786802, + 0.008592634461820126, + 0.004812739789485931, + 0.005663061514496803, + 0.003444904228672385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.043834079056978226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001970708370209 + ], + "opposite_logprob_tokens": [ + -2.6950197219848633 + ], + "logprob_reward": -0.07001970708370209, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001970708370209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08601351082324982, + 0.19674897193908691, + 0.03759610652923584, + 0.11677040904760361, + 0.03829948231577873, + 0.0666598528623581, + 0.1369534730911255, + 0.23254816234111786, + 0.03511620685458183, + 0.05329379066824913 + ], + "credit_map": [ + 0.008120850659906864, + 0.014409949071705341, + 0.008400119841098785, + 0.010360661894083023, + 0.012821199372410774, + 0.010386642999947071, + 0.01238331664353609, + 0.00913078524172306, + 0.023467956110835075, + 0.04115191102027893, + 0.02309011109173298, + 0.028503254055976868, + 0.014912072569131851, + 0.019280491396784782, + 0.027329184114933014, + 0.01901400461792946, + 0.0035823017824441195, + 0.005835371091961861, + 0.003377446671947837, + 0.004251058213412762, + 0.007275710813701153, + 0.005089095793664455, + 0.004595216363668442, + 0.0035899074282497168, + 0.01312610600143671, + 0.015711139887571335, + 0.017558857798576355, + 0.017162667587399483, + 0.014423299580812454, + 0.017465176060795784, + 0.010809637606143951, + 0.010513528250157833, + 0.003628565464168787, + 0.00456195930019021, + 0.004573337268084288, + 0.004362398758530617, + 0.005607909522950649, + 0.00694443890824914, + 0.0035010548308491707, + 0.005119816865772009, + 0.006571535486727953, + 0.007930019870400429, + 0.006698031909763813, + 0.007543303072452545, + 0.007421767804771662, + 0.007718909997493029, + 0.008903284557163715, + 0.013873005285859108, + 0.011799784377217293, + 0.010385059751570225, + 0.031557269394397736, + 0.027906740084290504, + 0.012760342098772526, + 0.008764662779867649, + 0.0167697761207819, + 0.01700984127819538, + 0.016702590510249138, + 0.020363077521324158, + 0.01905101351439953, + 0.038159169256687164, + 0.026874972507357597, + 0.01726364530622959, + 0.07516182959079742, + 0.01897186040878296, + 0.003717907005921006, + 0.004082384053617716, + 0.006292853970080614, + 0.0049703894183039665, + 0.003991414792835712, + 0.005446079187095165, + 0.0044399406760931015, + 0.0021752414759248495, + 0.006584786344319582, + 0.008641660213470459, + 0.0042505995370447636, + 0.005725174676626921, + 0.009759681299328804, + 0.00710629764944315, + 0.007139861583709717, + 0.004085731692612171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.07001970708370209 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760900378227234 + ], + "opposite_logprob_tokens": [ + -0.8260900378227234 + ], + "logprob_reward": -0.5760900378227234, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760900378227234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3432424068450928, + 0.15386322140693665, + 0.07132948189973831, + 0.03485315665602684, + 0.043308064341545105, + 0.06565751135349274, + 0.038880590349435806, + 0.04287319257855415, + 0.11428139358758926, + 0.09171094000339508 + ], + "credit_map": [ + 0.0665077343583107, + 0.021389061585068703, + 0.011208342388272285, + 0.01779230311512947, + 0.03237483277916908, + 0.1340380609035492, + 0.020938362926244736, + 0.038993723690509796, + 0.01682443544268608, + 0.029274268075823784, + 0.018381476402282715, + 0.020293213427066803, + 0.018733981996774673, + 0.0176352858543396, + 0.017296859994530678, + 0.015423704870045185, + 0.007954313419759274, + 0.013277449645102024, + 0.008638964965939522, + 0.00890805758535862, + 0.007459498010575771, + 0.009851375594735146, + 0.007799936458468437, + 0.007439888082444668, + 0.003443765453994274, + 0.005248465575277805, + 0.004266403615474701, + 0.0040943557396531105, + 0.005280754994601011, + 0.006975576281547546, + 0.0025609531439840794, + 0.0029828830156475306, + 0.004277706146240234, + 0.005262009333819151, + 0.005320876836776733, + 0.005021289922297001, + 0.00588814914226532, + 0.005770777817815542, + 0.005297331605106592, + 0.0064699240028858185, + 0.00539639126509428, + 0.0055055380798876286, + 0.010679936036467552, + 0.008700840175151825, + 0.008475814945995808, + 0.005198149476200342, + 0.011134634725749493, + 0.010566212236881256, + 0.0053168125450611115, + 0.0031490647234022617, + 0.0044978610239923, + 0.005866799037903547, + 0.005092855077236891, + 0.004134294111281633, + 0.004757709335535765, + 0.006065198220312595, + 0.003295023925602436, + 0.005738153588026762, + 0.007097029127180576, + 0.007888641208410263, + 0.0048608663491904736, + 0.004490585066378117, + 0.006068830844014883, + 0.0034340645652264357, + 0.016267554834485054, + 0.010135507211089134, + 0.011214982718229294, + 0.015231365337967873, + 0.011289872229099274, + 0.01594608649611473, + 0.021105743944644928, + 0.013090292923152447, + 0.007874357514083385, + 0.014391379430890083, + 0.0071081556379795074, + 0.012540371157228947, + 0.015942487865686417, + 0.01294700801372528, + 0.012265627272427082, + 0.00864156149327755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.5760900378227234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066960990428925 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066960990428925, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066960990428925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22956375777721405, + 0.05890369787812233, + 0.059618618339300156, + 0.046938978135585785, + 0.03348942846059799, + 0.13625238835811615, + 0.044639989733695984, + 0.050508592277765274, + 0.12755794823169708, + 0.21252664923667908 + ], + "credit_map": [ + 0.04109305143356323, + 0.026947475969791412, + 0.00957662146538496, + 0.011256652884185314, + 0.02141302265226841, + 0.08241544663906097, + 0.012840661220252514, + 0.024020813405513763, + 0.005479459185153246, + 0.00684611639007926, + 0.008108521811664104, + 0.008009741082787514, + 0.010208400897681713, + 0.006147234234958887, + 0.007009733933955431, + 0.007094488479197025, + 0.006662468891590834, + 0.012218184769153595, + 0.006212212610989809, + 0.005991847719997168, + 0.008902590721845627, + 0.008834554813802242, + 0.004736852832138538, + 0.00605990132316947, + 0.0047567845322191715, + 0.006682397332042456, + 0.006887114606797695, + 0.005262112710624933, + 0.007112798746675253, + 0.008097226731479168, + 0.0043882462196052074, + 0.003752294695004821, + 0.0029203977901488543, + 0.003198800375685096, + 0.003559964708983898, + 0.00516836903989315, + 0.005832049995660782, + 0.0047815777361392975, + 0.004741380922496319, + 0.003286887425929308, + 0.025974290445446968, + 0.009076208807528019, + 0.009876532480120659, + 0.01081685908138752, + 0.012852958403527737, + 0.007305984850972891, + 0.015261270105838776, + 0.0450882688164711, + 0.004173556808382273, + 0.0036109306383877993, + 0.008700074627995491, + 0.009747326374053955, + 0.004347472917288542, + 0.004237532615661621, + 0.005594316404312849, + 0.004228778649121523, + 0.003579662414267659, + 0.0057164086028933525, + 0.005611743777990341, + 0.013603052124381065, + 0.004249764606356621, + 0.004268939606845379, + 0.009541565552353859, + 0.003937450237572193, + 0.017009690403938293, + 0.010940006002783775, + 0.017135772854089737, + 0.012936972081661224, + 0.012592095881700516, + 0.018069349229335785, + 0.0255840215831995, + 0.013290022499859333, + 0.017587514594197273, + 0.0291488915681839, + 0.015251773409545422, + 0.027122531086206436, + 0.04212404042482376, + 0.028905482962727547, + 0.030056657269597054, + 0.02232975699007511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.016066960990428925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796818271279335 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796818271279335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796818271279335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24752368032932281, + 0.07829694449901581, + 0.04210001230239868, + 0.07209253311157227, + 0.04354892671108246, + 0.06123191863298416, + 0.13255427777767181, + 0.12035597860813141, + 0.0766109898686409, + 0.12568476796150208 + ], + "credit_map": [ + 0.04494619369506836, + 0.023134665563702583, + 0.009431051090359688, + 0.013956083916127682, + 0.029700731858611107, + 0.08412734419107437, + 0.015329993329942226, + 0.026897607371211052, + 0.009930535219609737, + 0.018154196441173553, + 0.008373136632144451, + 0.009559840895235538, + 0.008409918285906315, + 0.0075204926542937756, + 0.009708738885819912, + 0.006640090141445398, + 0.004239640664309263, + 0.006625120062381029, + 0.005091194994747639, + 0.0048081884160637856, + 0.007751129101961851, + 0.005049846135079861, + 0.005025031510740519, + 0.0035098588559776545, + 0.005819974932819605, + 0.011725189164280891, + 0.009842897765338421, + 0.007973823696374893, + 0.012832744978368282, + 0.013730841688811779, + 0.004608717281371355, + 0.005558347329497337, + 0.0052248649299144745, + 0.004377008881419897, + 0.00596423028036952, + 0.0069381846114993095, + 0.005767372902482748, + 0.006556701380759478, + 0.004413096234202385, + 0.004307470750063658, + 0.005852540489286184, + 0.005102785769850016, + 0.010554291307926178, + 0.010883801616728306, + 0.006449188571423292, + 0.005015015136450529, + 0.009295032359659672, + 0.008079265244305134, + 0.010028617456555367, + 0.013933639973402023, + 0.025887222960591316, + 0.017439696937799454, + 0.025424543768167496, + 0.01644139364361763, + 0.012544645927846432, + 0.010854520834982395, + 0.0056747253984212875, + 0.007177081890404224, + 0.012926981784403324, + 0.03411239758133888, + 0.010689844377338886, + 0.005732867866754532, + 0.03839593753218651, + 0.0056461491622030735, + 0.0050324006006121635, + 0.01073069404810667, + 0.005830481182783842, + 0.010069002397358418, + 0.012952952645719051, + 0.012203401885926723, + 0.014241988770663738, + 0.005550063215196133, + 0.012531114742159843, + 0.01395043171942234, + 0.012367750518023968, + 0.015294899232685566, + 0.020909607410430908, + 0.016429010778665543, + 0.020113173872232437, + 0.01408877782523632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.029796818271279335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot, shot upward from near sea level, slowly and steadily moving forward, captures a scene of apocalyptic grandeur. Beneath a sky tinged with leaden gray by the storm, a massive wave, like a moving mountain, roars toward the horizon. Its deep, dark green waters churn with pale, violent foam, carrying with them a primal force of destruction. Directly before it, a massive, weathered black reef, its surface covered in dark moss and jagged cracks, stands defiantly like an ancient watchtower amidst the surging sea. As the wave crashes against the reef with thunderous force, a deafening roar echoes, like the smiting of an invisible axe, stirring up a cloud of mist that soars into the sky. A dazzling crack instantly penetrated from the top of the reef to the underwater base. Then, under the suffocating pressure, the entire reef was split in half. The two huge fractured rock bodies collapsed to the sides and fell into the subsequent surging violent white waves and were instantly swallowed up. The whole picture was full of primitive, violent and awe-inspiring natural destructive power under the cold and depressing light, creating a tense and epic atmosphere.", + "seed": 5584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_12/index_5871_seed_5584_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.18046431923285128, + "vlm_logprob": -0.18046431923285128, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.18046431923285128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17905570566654205 + ], + "opposite_logprob_tokens": [ + -2.179055690765381 + ], + "logprob_reward": -0.17905570566654205, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.17905570566654205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14513365924358368, + 0.05578808858990669, + 0.0548994280397892, + 0.26715749502182007, + 0.07301868498325348, + 0.039800167083740234, + 0.06635689735412598, + 0.04695970565080643, + 0.03383093327283859, + 0.21705493330955505 + ], + "credit_map": [ + 0.018654173240065575, + 0.010857656598091125, + 0.02913261577486992, + 0.017682766541838646, + 0.02889152057468891, + 0.02498067356646061, + 0.005808087531477213, + 0.009126165881752968, + 0.00864187628030777, + 0.0029367918614298105, + 0.006164004560559988, + 0.0034353735391050577, + 0.009865266270935535, + 0.017206471413373947, + 0.003441963577643037, + 0.004096341319382191, + 0.00750597333535552, + 0.0065676080994307995, + 0.007524183019995689, + 0.005287031177431345, + 0.011162693612277508, + 0.0077110580168664455, + 0.00401640310883522, + 0.00512447627261281, + 0.021277792751789093, + 0.055758342146873474, + 0.013962010852992535, + 0.015339785255491734, + 0.04174935072660446, + 0.10243561118841171, + 0.007162618916481733, + 0.009471976198256016, + 0.013242114335298538, + 0.012926091440021992, + 0.009697635658085346, + 0.00886831060051918, + 0.009324601851403713, + 0.007499406114220619, + 0.005797828081995249, + 0.0056626941077411175, + 0.00451856991276145, + 0.004984981846064329, + 0.0033353371545672417, + 0.004329670686274767, + 0.008421450853347778, + 0.004917519632726908, + 0.005928517784923315, + 0.003364117117598653, + 0.008079273626208305, + 0.007084018085151911, + 0.011595510877668858, + 0.00750007014721632, + 0.009905765764415264, + 0.010713236406445503, + 0.005688382312655449, + 0.005790636409074068, + 0.005809071008116007, + 0.0077253845520317554, + 0.006202464923262596, + 0.00421485910192132, + 0.006579242181032896, + 0.006626490503549576, + 0.0038964534178376198, + 0.005905739031732082, + 0.005191047675907612, + 0.006367402151226997, + 0.0035268384963274, + 0.003632814157754183, + 0.0052535682916641235, + 0.003418332664296031, + 0.003353250678628683, + 0.003087678225710988, + 0.08069198578596115, + 0.017982346937060356, + 0.060100678354501724, + 0.01202255580574274, + 0.0103722820058465, + 0.014653310179710388, + 0.009590723551809788, + 0.011641056276857853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.17905570566654205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705278396606445 + ], + "opposite_logprob_tokens": [ + -2.1270527839660645 + ], + "logprob_reward": -0.12705278396606445, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705278396606445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24044570326805115, + 0.16705596446990967, + 0.03413785621523857, + 0.07113808393478394, + 0.046271953731775284, + 0.07050731778144836, + 0.08836866170167923, + 0.02978186309337616, + 0.05440458282828331, + 0.19788800179958344 + ], + "credit_map": [ + 0.020667491480708122, + 0.010739308781921864, + 0.05786745622754097, + 0.014947640709578991, + 0.02484714612364769, + 0.09126630425453186, + 0.007156743668019772, + 0.012953629717230797, + 0.013606726191937923, + 0.007026628591120243, + 0.04158482328057289, + 0.006478725466877222, + 0.017422666773200035, + 0.07118800282478333, + 0.004403820261359215, + 0.0053445808589458466, + 0.004102686885744333, + 0.003661887487396598, + 0.006460476666688919, + 0.0042152563109993935, + 0.00594565412029624, + 0.004431459121406078, + 0.0030669188126921654, + 0.002253517974168062, + 0.010669969022274017, + 0.007279824931174517, + 0.004201272968202829, + 0.0032730207312852144, + 0.011483271606266499, + 0.028827963396906853, + 0.0021474785171449184, + 0.003255286952480674, + 0.0064613064751029015, + 0.003961889073252678, + 0.004134180024266243, + 0.0036837889347225428, + 0.007689351215958595, + 0.014284773729741573, + 0.0029555687215179205, + 0.003101098584011197, + 0.015025187283754349, + 0.007160072680562735, + 0.00580755528062582, + 0.019951026886701584, + 0.0059352414682507515, + 0.00563489506021142, + 0.004928068723529577, + 0.006065275054425001, + 0.017216388136148453, + 0.010685893706977367, + 0.024937348440289497, + 0.00873860064893961, + 0.009073389694094658, + 0.0072232503443956375, + 0.006248078774660826, + 0.004245715215802193, + 0.00371939386241138, + 0.004938642960041761, + 0.004179475363343954, + 0.003400353714823723, + 0.0035859462805092335, + 0.0030032149516046047, + 0.0038889963179826736, + 0.003065840806812048, + 0.009486416354775429, + 0.008446779102087021, + 0.006600039545446634, + 0.006695145275443792, + 0.0079494034871459, + 0.006023538298904896, + 0.005280226469039917, + 0.0039230394177138805, + 0.0794593021273613, + 0.015046225860714912, + 0.06039239466190338, + 0.012581011280417442, + 0.007106625474989414, + 0.0068814377300441265, + 0.007330039981752634, + 0.009090983308851719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.12705278396606445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760691165924072 + ], + "opposite_logprob_tokens": [ + -0.8260691165924072 + ], + "logprob_reward": -0.5760691165924072, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760691165924072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1360016167163849, + 0.2341807782649994, + 0.08380363136529922, + 0.05859437212347984, + 0.05850466713309288, + 0.08878089487552643, + 0.11346085369586945, + 0.05435538664460182, + 0.07693051546812057, + 0.0953872874379158 + ], + "credit_map": [ + 0.020410530269145966, + 0.01083274744451046, + 0.018527723848819733, + 0.01585230976343155, + 0.027386225759983063, + 0.025067070499062538, + 0.00808649230748415, + 0.009838514029979706, + 0.023849481716752052, + 0.010992036201059818, + 0.03326032683253288, + 0.014602204784750938, + 0.04160359874367714, + 0.07534828037023544, + 0.018431082367897034, + 0.016093764454126358, + 0.012826251797378063, + 0.010390184819698334, + 0.01432732306420803, + 0.008817563764750957, + 0.013608784414827824, + 0.012539227493107319, + 0.006111528724431992, + 0.005182771477848291, + 0.007440322544425726, + 0.0065767779015004635, + 0.006527692079544067, + 0.0055968440137803555, + 0.008455093950033188, + 0.0104867834597826, + 0.00681742699816823, + 0.006693430244922638, + 0.0070177000015974045, + 0.004929057788103819, + 0.0042144679464399815, + 0.00488271564245224, + 0.010080606676638126, + 0.017969217151403427, + 0.00392743619158864, + 0.005483465269207954, + 0.016917839646339417, + 0.010026909410953522, + 0.0071244207210838795, + 0.021482743322849274, + 0.010704047977924347, + 0.008195401169359684, + 0.008092595264315605, + 0.006236941087990999, + 0.013824548572301865, + 0.014411277137696743, + 0.02282373607158661, + 0.012131167575716972, + 0.015275852754712105, + 0.016247007995843887, + 0.009031071327626705, + 0.009716189466416836, + 0.006916658487170935, + 0.007190137170255184, + 0.009633473120629787, + 0.00646402919664979, + 0.006379662547260523, + 0.007941815070807934, + 0.003942960407584906, + 0.005886646918952465, + 0.014182768762111664, + 0.012883496470749378, + 0.009106886573135853, + 0.005746013019233942, + 0.012870966456830502, + 0.009066336788237095, + 0.0075617688708007336, + 0.00551227480173111, + 0.02841009944677353, + 0.012042169459164143, + 0.022405266761779785, + 0.0074654314666986465, + 0.006425978150218725, + 0.008322661742568016, + 0.005044370889663696, + 0.005271313712000847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.5760691165924072 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947985220700502 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947985220700502, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947985220700502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054809361696243286, + 0.29679200053215027, + 0.03433277830481529, + 0.11672484129667282, + 0.07715815305709839, + 0.05065670609474182, + 0.04502972587943077, + 0.04931054636836052, + 0.01881815306842327, + 0.2563677430152893 + ], + "credit_map": [ + 0.00733064953237772, + 0.004025705624371767, + 0.008551322855055332, + 0.004200113471597433, + 0.011546709574759007, + 0.011883854866027832, + 0.0028658309020102024, + 0.0044051785953342915, + 0.023437658324837685, + 0.01307724416255951, + 0.07699248194694519, + 0.01038481667637825, + 0.0316445529460907, + 0.12784025073051453, + 0.006039073690772057, + 0.007375931832939386, + 0.005575419869273901, + 0.002655483316630125, + 0.004922020714730024, + 0.003180553438141942, + 0.007268024608492851, + 0.003992707934230566, + 0.003852538298815489, + 0.002886031987145543, + 0.017266280949115753, + 0.012820602394640446, + 0.005644416902214289, + 0.0044787866063416, + 0.018712323158979416, + 0.04754234105348587, + 0.004571225494146347, + 0.005688863340765238, + 0.006690513342618942, + 0.01707855425775051, + 0.00450923340395093, + 0.007014332804828882, + 0.011724257841706276, + 0.022149795666337013, + 0.00420342804864049, + 0.003788038855418563, + 0.009584407322108746, + 0.004320112522691488, + 0.0037588884588330984, + 0.012521971017122269, + 0.00682942196726799, + 0.004669145215302706, + 0.004847514443099499, + 0.004125240258872509, + 0.00447881082072854, + 0.005565103143453598, + 0.008694548159837723, + 0.005484114866703749, + 0.006485907826572657, + 0.00685968529433012, + 0.003829998429864645, + 0.003631556173786521, + 0.007175292819738388, + 0.008772199973464012, + 0.00625158566981554, + 0.005808105692267418, + 0.005095120053738356, + 0.006617482751607895, + 0.005061279982328415, + 0.004529479891061783, + 0.003288174979388714, + 0.002426953287795186, + 0.0019874991849064827, + 0.0019172847969457507, + 0.0027431624475866556, + 0.0026814432349056005, + 0.002082128543406725, + 0.001691507059149444, + 0.10513266921043396, + 0.020015565678477287, + 0.07900405675172806, + 0.013907395303249359, + 0.009044675156474113, + 0.00919214729219675, + 0.008642441593110561, + 0.011428789235651493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.005947985220700502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a wide-angle, long shot perspective looking upward from near sea level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wave with deep dark green water and pale violent foam visible moving toward the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive weathered black reef covered in dark moss and jagged cracks visible before the wave crashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reef split in half and collapse into the surging white waves as the wave crashes against it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196004718542099 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196004718542099, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196004718542099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046348292380571365, + 0.06361277401447296, + 0.01871630921959877, + 0.023653673008084297, + 0.05139830708503723, + 0.02294244058430195, + 0.061491191387176514, + 0.04247574508190155, + 0.03240671753883362, + 0.6369544863700867 + ], + "credit_map": [ + 0.0077396356500685215, + 0.002449134597554803, + 0.009214820340275764, + 0.003823731327429414, + 0.009033636189997196, + 0.008786041289567947, + 0.002107765758410096, + 0.003193530486896634, + 0.004893400706350803, + 0.002973658498376608, + 0.017414284870028496, + 0.0025835956912487745, + 0.005969885736703873, + 0.026066049933433533, + 0.0018055974505841732, + 0.0019063004292547703, + 0.0029291315004229546, + 0.0016710793133825064, + 0.0035025249235332012, + 0.0018435248639434576, + 0.0039200508035719395, + 0.0022464129142463207, + 0.0012947715586051345, + 0.001308813109062612, + 0.005336623173207045, + 0.0027324031107127666, + 0.0018484813626855612, + 0.0017059938982129097, + 0.0028367703780531883, + 0.004900285042822361, + 0.002180247101932764, + 0.002112870803102851, + 0.008725076913833618, + 0.013720978051424026, + 0.0026366785168647766, + 0.005024726036936045, + 0.006535700522363186, + 0.007814597338438034, + 0.00340042682364583, + 0.0035401193890720606, + 0.0029698105063289404, + 0.002588690957054496, + 0.0035465797409415245, + 0.002794125583022833, + 0.0037236541975289583, + 0.0030505582690238953, + 0.0024529332295060158, + 0.0018160889158025384, + 0.009189297445118427, + 0.007411231752485037, + 0.01603786274790764, + 0.006656426936388016, + 0.00797655712813139, + 0.006875547114759684, + 0.0041136411018669605, + 0.003230627626180649, + 0.00579239334911108, + 0.007878217846155167, + 0.006057282444089651, + 0.005953833926469088, + 0.004618371371179819, + 0.0052602095529437065, + 0.0034929022658616304, + 0.0034225378185510635, + 0.004557848442345858, + 0.004185666795819998, + 0.002810810925439, + 0.004978349432349205, + 0.004309120588004589, + 0.00501323351636529, + 0.0038480423390865326, + 0.0027036459650844336, + 0.25789743661880493, + 0.049080006778240204, + 0.19243782758712769, + 0.03506956249475479, + 0.02391154132783413, + 0.0264322217553854, + 0.02312195487320423, + 0.029003966599702835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.014196004718542099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_tensor([527, 527])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_tensor([527, 527])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2f4d25a40b51681a4ca24f0007ec69308d6bb5fa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_tensor([527, 527])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "grainy black-and-white historical film, shot from a fixed high angle, draws the viewer's gaze into the heart of a Victorian-style laboratory cluttered with bizarre devices and scattered blueprints. At the bottom of the frame, an inventor, sporting a thick, carefully trimmed mustache and wearing a dark vest over a white turtleneck shirt, leans slightly forward, his brow furrowed, his gaze sharply focused on the workbench before him. In the center of the workbench sits a mysterious, smooth sphere of unidentified material, emitting a soft, eerie glow, the sole source of light in the dimly lit scene. The inventor, gripping a small brass hammer in his right hand, raises it steadily and deliberately, and with pinpoint precision, gently and rhythmically strikes the sphere's outer shell. With the final blow, a crack appears in the sphere's surface, and countless bright, irregular arcs of electricity burst forth, scattering wildly in all directions like uncontrolled white lightning. Under the dazzling electric light, the inventor's face and the corners of the laboratory are cast in violently shaking, deep shadows. The whole picture is full of strong visual contrast and a sense of tension and awe at the moment of breakthrough discovery.", + "seed": 21887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_527_seed_21887_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04473463359754533, + "vlm_logprob": -0.04473463359754533, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.04473463359754533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17569199204444885 + ], + "opposite_logprob_tokens": [ + -1.9256919622421265 + ], + "logprob_reward": -0.17569199204444885, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17569199204444885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07289288938045502, + 0.019390221685171127, + 0.016183551400899887, + 0.014083421789109707, + 0.03934628516435623, + 0.27099332213401794, + 0.016445420682430267, + 0.1391894668340683, + 0.013880448415875435, + 0.3975948989391327 + ], + "credit_map": [ + 0.013223635964095592, + 0.0090028066188097, + 0.007780378218740225, + 0.010693132877349854, + 0.009534950368106365, + 0.006998386699706316, + 0.006635687779635191, + 0.00902391318231821, + 0.003744479501619935, + 0.002559986198320985, + 0.002313518663868308, + 0.0021781050600111485, + 0.002649951260536909, + 0.0018917181296274066, + 0.0016207036096602678, + 0.002431758213788271, + 0.0030338438227772713, + 0.0018704957328736782, + 0.0024433538783341646, + 0.0015602806815877557, + 0.002052314579486847, + 0.0019321258878335357, + 0.0016233816277235746, + 0.00166775519028306, + 0.002664007944986224, + 0.0015381680568680167, + 0.0024962665047496557, + 0.0013043731451034546, + 0.001180735183879733, + 0.0016443305648863316, + 0.001151051837950945, + 0.0021044898312538862, + 0.006717092357575893, + 0.003306509694084525, + 0.007193697150796652, + 0.004356768447905779, + 0.002784136915579438, + 0.004121331963688135, + 0.0036910935305058956, + 0.007175657898187637, + 0.0265022162348032, + 0.009253515861928463, + 0.10242194682359695, + 0.011852510273456573, + 0.007726818788796663, + 0.017616258934140205, + 0.008552641607820988, + 0.08706743270158768, + 0.0016799132572486997, + 0.0006999462493695319, + 0.005743477959185839, + 0.0008208350045606494, + 0.0005313538713380694, + 0.0013043013168498874, + 0.0006129808025434613, + 0.005052612163126469, + 0.014035037718713284, + 0.04961478337645531, + 0.0052597555331885815, + 0.047721270471811295, + 0.005441930610686541, + 0.0069643184542655945, + 0.005720050539821386, + 0.004432315472513437, + 0.0013665349688380957, + 0.00419112853705883, + 0.0010972546879202127, + 0.003488814225420356, + 0.0008035348146222532, + 0.0014044417766854167, + 0.0007534768083132803, + 0.0007752621313557029, + 0.028412992134690285, + 0.03959712013602257, + 0.020743930712342262, + 0.01067847479134798, + 0.022761911153793335, + 0.16130658984184265, + 0.008594179525971413, + 0.10549972951412201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.17569199204444885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381845086812973 + ], + "opposite_logprob_tokens": [ + -3.288184404373169 + ], + "logprob_reward": -0.0381845086812973, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381845086812973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09445908665657043, + 0.05242509767413139, + 0.03714221715927124, + 0.010698540136218071, + 0.029247183352708817, + 0.4173862934112549, + 0.2528969347476959, + 0.043635230511426926, + 0.019025426357984543, + 0.043083976954221725 + ], + "credit_map": [ + 0.02249060571193695, + 0.010747111402451992, + 0.018127676099538803, + 0.006669172551482916, + 0.0064873844385147095, + 0.009581541642546654, + 0.007263028994202614, + 0.013092560693621635, + 0.009944223798811436, + 0.008197909221053123, + 0.006990186404436827, + 0.005032343324273825, + 0.0059682345017790794, + 0.006709914188832045, + 0.004361215978860855, + 0.0052210683934390545, + 0.008060435764491558, + 0.006093058735132217, + 0.006080621387809515, + 0.003602399956434965, + 0.0035045805852860212, + 0.0035246852785348892, + 0.0024661824572831392, + 0.003810252994298935, + 0.001998904161155224, + 0.0014728278620168567, + 0.0018503091996535659, + 0.001019310555420816, + 0.0013084396487101912, + 0.0010207724990323186, + 0.000836040882859379, + 0.0011919341050088406, + 0.005876485724002123, + 0.0025767649058252573, + 0.006160294637084007, + 0.001935049775056541, + 0.0022136569023132324, + 0.003241360653191805, + 0.0016996451886370778, + 0.005543924868106842, + 0.0386616550385952, + 0.022551896050572395, + 0.14909116923809052, + 0.018821753561496735, + 0.016294660046696663, + 0.02959749847650528, + 0.015185367316007614, + 0.12718230485916138, + 0.025294192135334015, + 0.01023291889578104, + 0.09526102989912033, + 0.010410265997052193, + 0.006912583019584417, + 0.019694380462169647, + 0.00725147407501936, + 0.07784008234739304, + 0.005154927261173725, + 0.008776052854955196, + 0.0032905968837440014, + 0.010006537660956383, + 0.0042319451458752155, + 0.005138876382261515, + 0.003270220011472702, + 0.0037660731468349695, + 0.0024397526867687702, + 0.003707655007019639, + 0.0016492630820721388, + 0.002779935719445348, + 0.002190025756135583, + 0.0021278630010783672, + 0.0018391412450000644, + 0.0022917883470654488, + 0.005703647155314684, + 0.006679876241832972, + 0.003585681552067399, + 0.007309728302061558, + 0.004124603699892759, + 0.008460445329546928, + 0.0036190268583595753, + 0.0036009696777909994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0381845086812973 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013441346818581223 + ], + "opposite_logprob_tokens": [ + -6.626344203948975 + ], + "logprob_reward": -0.0013441346818581223, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013441346818581223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025915171951055527, + 0.006192195229232311, + 0.008178988471627235, + 0.0054300641641020775, + 0.010197230614721775, + 0.02247750572860241, + 0.08141214400529861, + 0.012993859127163887, + 0.021336199715733528, + 0.8058666586875916 + ], + "credit_map": [ + 0.003934804815798998, + 0.003926025703549385, + 0.0027014354709535837, + 0.002518430585041642, + 0.005098690278828144, + 0.0023317027371376753, + 0.0027052536606788635, + 0.0026988298632204533, + 0.0009219431667588651, + 0.0006540059112012386, + 0.0007646472076885402, + 0.0006373196956701577, + 0.0010681793792173266, + 0.0005970128695480525, + 0.0007679261616431177, + 0.0007811607792973518, + 0.0011195468250662088, + 0.0015428104670718312, + 0.0009206162649206817, + 0.0008888865122571588, + 0.0012115684803575277, + 0.0008497897651977837, + 0.0007275815005414188, + 0.0009181878995150328, + 0.0011800993233919144, + 0.00043821585131809115, + 0.0010657069506123662, + 0.0004948355490341783, + 0.0006479172152467072, + 0.0005807715933769941, + 0.00039260799530893564, + 0.0006299100932665169, + 0.0014398019993677735, + 0.0008238103473559022, + 0.002490209648385644, + 0.0009601991041563451, + 0.0006642152438871562, + 0.0007576098432764411, + 0.0005337137845344841, + 0.0025276709347963333, + 0.0021271707955747843, + 0.0007411324186250567, + 0.008166812360286713, + 0.0011256561847403646, + 0.001062738592736423, + 0.001468055066652596, + 0.0007545437547378242, + 0.007031396962702274, + 0.008084667846560478, + 0.0033231396228075027, + 0.030623769387602806, + 0.003273394424468279, + 0.0023287730291485786, + 0.006343909073621035, + 0.002315377350896597, + 0.02511911652982235, + 0.0012872815132141113, + 0.003624311648309231, + 0.0009078684379346669, + 0.002806165488436818, + 0.0011234852718189359, + 0.0010872717248275876, + 0.0012811248889192939, + 0.0008763496880419552, + 0.0019338078564032912, + 0.00679718516767025, + 0.001337356516160071, + 0.005738760344684124, + 0.001193651813082397, + 0.0015948161017149687, + 0.0012703761458396912, + 0.0014702463522553444, + 0.05719539895653725, + 0.07626059651374817, + 0.03946535289287567, + 0.026421038433909416, + 0.04080626368522644, + 0.315560907125473, + 0.017461959272623062, + 0.23269517719745636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0013441346818581223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017162755131721497, + 0.006975612137466669, + 0.004455106798559427, + 0.010478459298610687, + 0.009090401232242584, + 0.08855882287025452, + 0.048766136169433594, + 0.0961398035287857, + 0.02422773651778698, + 0.6941452026367188 + ], + "credit_map": [ + 0.0021818506065756083, + 0.0020479727536439896, + 0.0019448488019406796, + 0.002777478890493512, + 0.002214010339230299, + 0.002263519447296858, + 0.0018793849740177393, + 0.0018536885036155581, + 0.0011771224671974778, + 0.0006853375234641135, + 0.0009337239898741245, + 0.0009067672654055059, + 0.0008478314848616719, + 0.0008816519984975457, + 0.0007592720212414861, + 0.0007839046302251518, + 0.0007972536841407418, + 0.000577692233491689, + 0.0006836084648966789, + 0.00040689329034648836, + 0.00044701166916638613, + 0.0006504188640974462, + 0.0004270027857273817, + 0.0004652251664083451, + 0.002435973146930337, + 0.0009837460238486528, + 0.0023544260766357183, + 0.0006722576217725873, + 0.0006995631847530603, + 0.0012744846753776073, + 0.0006753627676516771, + 0.0013826454523950815, + 0.001028515282087028, + 0.0004925963585264981, + 0.002639099955558777, + 0.0007195684593170881, + 0.00046249807928688824, + 0.0007579231751151383, + 0.000514256302267313, + 0.002475942485034466, + 0.008855725638568401, + 0.0033988661598414183, + 0.03247613459825516, + 0.004038054030388594, + 0.0029531759209930897, + 0.005967909004539251, + 0.0030761309899389744, + 0.02779282256960869, + 0.004912340082228184, + 0.0021072393283247948, + 0.018211357295513153, + 0.0019796432461589575, + 0.0013996859779581428, + 0.0037507901433855295, + 0.001491314498707652, + 0.014913761988282204, + 0.009642195887863636, + 0.03394443541765213, + 0.0038118455559015274, + 0.03310729190707207, + 0.0037682398688048124, + 0.004609590396285057, + 0.004208622965961695, + 0.00304757640697062, + 0.0021860681008547544, + 0.007184344809502363, + 0.0020450239535421133, + 0.006011905614286661, + 0.0013578144134953618, + 0.002043311484158039, + 0.0014056595973670483, + 0.0019936079625040293, + 0.04978164657950401, + 0.06324443221092224, + 0.03231973573565483, + 0.04180125892162323, + 0.03140374645590782, + 0.25164586305618286, + 0.02109992876648903, + 0.20284853875637054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0052584195509552956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194113029167056 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194113029167056, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194113029167056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.008215736597776413, + 0.003683127462863922, + 0.0023259716108441353, + 0.008841830305755138, + 0.004956090357154608, + 0.03284918889403343, + 0.01959485188126564, + 0.09788332134485245, + 0.0072595407254993916, + 0.8143903613090515 + ], + "credit_map": [ + 0.001443691668100655, + 0.0011673462577164173, + 0.00090934889158234, + 0.0009164407383650541, + 0.0010177385993301868, + 0.0008992725634016097, + 0.0008042042027227581, + 0.0010576938511803746, + 0.0006435344112105668, + 0.0003985840594395995, + 0.0005765943205915391, + 0.0003368775360286236, + 0.00038946623681113124, + 0.0003885547921527177, + 0.00037191162118688226, + 0.0005776042817160487, + 0.00044356888975016773, + 0.0002573934616521001, + 0.00036764368996955454, + 0.00026740162866190076, + 0.0002666597138158977, + 0.0002129366621375084, + 0.00017874615150503814, + 0.00033162132604047656, + 0.0021173893474042416, + 0.0008711647824384272, + 0.002006693510338664, + 0.0005854651099070907, + 0.0005120810819789767, + 0.0009286099812015891, + 0.0005108017357997596, + 0.001309625105932355, + 0.0006600585184060037, + 0.00044215642265044153, + 0.0010950975120067596, + 0.0005115176900289953, + 0.00033730315044522285, + 0.000529244658537209, + 0.0003117088635917753, + 0.0010690032504498959, + 0.003177625359967351, + 0.001032607746310532, + 0.0123756667599082, + 0.0014604564057663083, + 0.0009441659203730524, + 0.0021625652443617582, + 0.0010441119084134698, + 0.010651991702616215, + 0.001985106850042939, + 0.0009930786909535527, + 0.007279576733708382, + 0.0008083631983026862, + 0.0006226229015737772, + 0.0014875010820105672, + 0.0006057763239368796, + 0.0058128247037529945, + 0.009844547137618065, + 0.0348319411277771, + 0.0037176120094954967, + 0.03360409289598465, + 0.0037877527065575123, + 0.0049582053907215595, + 0.004012435674667358, + 0.003126736730337143, + 0.0007210559560917318, + 0.0021315074991434813, + 0.00053272710647434, + 0.0017726493533700705, + 0.00041950540617108345, + 0.0004880925116594881, + 0.0005618400755338371, + 0.0006321626715362072, + 0.058498792350292206, + 0.0780683308839798, + 0.04039442539215088, + 0.026493940502405167, + 0.04191150888800621, + 0.3206692636013031, + 0.017745045945048332, + 0.23060902953147888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.003194113029167056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grainy black-and-white historical film, shot from a fixed high angle, draws the viewer's gaze into the heart of a Victorian-style laboratory cluttered with bizarre devices and scattered blueprints. At the bottom of the frame, an inventor, sporting a thick, carefully trimmed mustache and wearing a dark vest over a white turtleneck shirt, leans slightly forward, his brow furrowed, his gaze sharply focused on the workbench before him. In the center of the workbench sits a mysterious, smooth sphere of unidentified material, emitting a soft, eerie glow, the sole source of light in the dimly lit scene. The inventor, gripping a small brass hammer in his right hand, raises it steadily and deliberately, and with pinpoint precision, gently and rhythmically strikes the sphere's outer shell. With the final blow, a crack appears in the sphere's surface, and countless bright, irregular arcs of electricity burst forth, scattering wildly in all directions like uncontrolled white lightning. Under the dazzling electric light, the inventor's face and the corners of the laboratory are cast in violently shaking, deep shadows. The whole picture is full of strong visual contrast and a sense of tension and awe at the moment of breakthrough discovery.", + "seed": 21888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_527_seed_21888_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1840966931078583, + "vlm_logprob": -0.1840966931078583, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.1840966931078583, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8347668051719666 + ], + "opposite_logprob_tokens": [ + -0.5847668051719666 + ], + "logprob_reward": -0.8347668051719666, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8347668051719666, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24404793977737427, + 0.04299578070640564, + 0.03775031492114067, + 0.04096953943371773, + 0.01662583462893963, + 0.09427221864461899, + 0.03941967710852623, + 0.10203896462917328, + 0.31307944655418396, + 0.06880035996437073 + ], + "credit_map": [ + 0.0406833179295063, + 0.029127581045031548, + 0.03774035722017288, + 0.027351265773177147, + 0.026335272938013077, + 0.026296887546777725, + 0.033818963915109634, + 0.02269427664577961, + 0.006567321252077818, + 0.006473541725426912, + 0.0062891338020563126, + 0.003122105496004224, + 0.003910932689905167, + 0.007053793873637915, + 0.003718813881278038, + 0.005860139150172472, + 0.005853874608874321, + 0.0054208217188715935, + 0.005528559908270836, + 0.004256170708686113, + 0.005051614250987768, + 0.004059205297380686, + 0.004225669428706169, + 0.0033543994650244713, + 0.005960081238299608, + 0.005178258288651705, + 0.0037657341454178095, + 0.005502710584551096, + 0.005628417711704969, + 0.0026751419063657522, + 0.007431916426867247, + 0.00482728099450469, + 0.0025090589188039303, + 0.002199822571128607, + 0.002528358716517687, + 0.0022577375639230013, + 0.0019042182248085737, + 0.0018169397953897715, + 0.0016148920403793454, + 0.0017948062159121037, + 0.013195402920246124, + 0.015413755550980568, + 0.011743002571165562, + 0.007199411280453205, + 0.010568968020379543, + 0.008779340423643589, + 0.011827254667878151, + 0.015545080415904522, + 0.0042117987759411335, + 0.006230344995856285, + 0.0044593820348382, + 0.0047217244282364845, + 0.004335736855864525, + 0.003735315753147006, + 0.00683285528793931, + 0.00489251920953393, + 0.017624355852603912, + 0.015316343866288662, + 0.019806023687124252, + 0.01096605509519577, + 0.012208272702991962, + 0.006659097503870726, + 0.012095457874238491, + 0.0073633575811982155, + 0.043399762362241745, + 0.03799876198172569, + 0.02181628718972206, + 0.06215576454997063, + 0.0381818525493145, + 0.019492944702506065, + 0.06750256568193436, + 0.022531498223543167, + 0.009082301519811153, + 0.009944421239197254, + 0.014562821947038174, + 0.006824976298958063, + 0.008980176411569118, + 0.005794258322566748, + 0.007410860620439053, + 0.006200539879500866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.8347668051719666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906202971935272 + ], + "opposite_logprob_tokens": [ + -2.579061985015869 + ], + "logprob_reward": -0.07906202971935272, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906202971935272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26059776544570923, + 0.12095142900943756, + 0.12051523476839066, + 0.04501006379723549, + 0.03624897822737694, + 0.07137364894151688, + 0.0754692479968071, + 0.05642075464129448, + 0.12008985131978989, + 0.09332306683063507 + ], + "credit_map": [ + 0.0393151119351387, + 0.019880525767803192, + 0.037394315004348755, + 0.026738710701465607, + 0.04034138470888138, + 0.03479226306080818, + 0.028421243652701378, + 0.033714212477207184, + 0.021321941167116165, + 0.019069300964474678, + 0.015006114728748798, + 0.011028238572180271, + 0.01126759871840477, + 0.014367892406880856, + 0.014567268081009388, + 0.014323071576654911, + 0.017298728227615356, + 0.010353430174291134, + 0.010630887001752853, + 0.022577036172151566, + 0.016665754839777946, + 0.014304522424936295, + 0.01612764783203602, + 0.012557225301861763, + 0.006281608249992132, + 0.005425267852842808, + 0.005037552677094936, + 0.002961347345262766, + 0.0071923453360795975, + 0.006468731444329023, + 0.005497447680681944, + 0.006145767867565155, + 0.005275313276797533, + 0.0038983921986073256, + 0.005940332543104887, + 0.004006792325526476, + 0.005564445164054632, + 0.003377493703737855, + 0.004893542733043432, + 0.0032926658168435097, + 0.009835940785706043, + 0.012792030349373817, + 0.01026330329477787, + 0.010255624540150166, + 0.005182541906833649, + 0.0046827057376503944, + 0.008730219677090645, + 0.009631283581256866, + 0.006369184702634811, + 0.00743019999936223, + 0.007351263891905546, + 0.010145746171474457, + 0.014562824741005898, + 0.007860349491238594, + 0.01442262064665556, + 0.007327063474804163, + 0.01009279116988182, + 0.008385540917515755, + 0.01074640266597271, + 0.005439846310764551, + 0.00756391417235136, + 0.0029558942187577486, + 0.007564479950815439, + 0.003671883372589946, + 0.016666242852807045, + 0.017190497368574142, + 0.009641719982028008, + 0.021630704402923584, + 0.013241867534816265, + 0.008533922024071217, + 0.022209476679563522, + 0.01097541768103838, + 0.01501128077507019, + 0.012232874520123005, + 0.00921622198075056, + 0.006682140287011862, + 0.012987373396754265, + 0.006171518005430698, + 0.018613023683428764, + 0.012408626265823841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07906202971935272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006422125734388828 + ], + "opposite_logprob_tokens": [ + -7.3756422996521 + ], + "logprob_reward": -0.0006422125734388828, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006422125734388828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24554702639579773, + 0.06734108179807663, + 0.06758110970258713, + 0.023895621299743652, + 0.02205793559551239, + 0.07501766085624695, + 0.06174420937895775, + 0.12288948893547058, + 0.24006250500679016, + 0.07386333495378494 + ], + "credit_map": [ + 0.03511164337396622, + 0.025250978767871857, + 0.032424237579107285, + 0.024462727829813957, + 0.02829243801534176, + 0.03304818645119667, + 0.03662533313035965, + 0.03033146634697914, + 0.009308119304478168, + 0.007168998476117849, + 0.007168194744735956, + 0.004521653521806002, + 0.006031087599694729, + 0.013051705434918404, + 0.0061259218491613865, + 0.013965402729809284, + 0.009335936047136784, + 0.0064872330985963345, + 0.007804847322404385, + 0.012731769122183323, + 0.00833749957382679, + 0.008210737258195877, + 0.007048396859318018, + 0.007624691352248192, + 0.003507126122713089, + 0.0035154155921190977, + 0.0034518844913691282, + 0.002338366350159049, + 0.004508551210165024, + 0.0020675314590334892, + 0.0022931615822017193, + 0.002213586587458849, + 0.003729338524863124, + 0.0026783093344420195, + 0.002857784740626812, + 0.0023470474407076836, + 0.002838065382093191, + 0.0017338513862341642, + 0.003724850481376052, + 0.0021486887708306313, + 0.009473277255892754, + 0.007899814285337925, + 0.010386692360043526, + 0.007534639444202185, + 0.009796638041734695, + 0.006869625300168991, + 0.010579328052699566, + 0.012477646581828594, + 0.005716041196137667, + 0.007834933698177338, + 0.007072580046951771, + 0.008061732165515423, + 0.009472861886024475, + 0.007163586560636759, + 0.008456721901893616, + 0.007965750060975552, + 0.022154223173856735, + 0.01756506972014904, + 0.022095050662755966, + 0.014196526259183884, + 0.014092082157731056, + 0.006719858385622501, + 0.01787533238530159, + 0.008191344328224659, + 0.033537860959768295, + 0.029788268730044365, + 0.016984406858682632, + 0.04671965166926384, + 0.029362419620156288, + 0.015165627934038639, + 0.05059994384646416, + 0.017904315143823624, + 0.010606197640299797, + 0.007565126288682222, + 0.016681959852576256, + 0.00892795342952013, + 0.009063778445124626, + 0.0065968758426606655, + 0.008016240783035755, + 0.0064052073284983635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0006422125734388828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193875541910529 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193875541910529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193875541910529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21608836948871613, + 0.04419982060790062, + 0.055931780487298965, + 0.04960545524954796, + 0.05102401599287987, + 0.13558752834796906, + 0.15289488434791565, + 0.09532742202281952, + 0.06839826703071594, + 0.13094250857830048 + ], + "credit_map": [ + 0.032816678285598755, + 0.022881079465150833, + 0.03216400742530823, + 0.018753642216324806, + 0.02045699954032898, + 0.031005356460809708, + 0.03061591275036335, + 0.027394678443670273, + 0.00654570059850812, + 0.006640246603637934, + 0.00843100156635046, + 0.003209260292351246, + 0.005100927781313658, + 0.00449451245367527, + 0.005237378645688295, + 0.004540788475424051, + 0.007869649678468704, + 0.0063435970805585384, + 0.005494748707860708, + 0.009878511540591717, + 0.009258931502699852, + 0.005205533467233181, + 0.006404463667422533, + 0.005476338788866997, + 0.008349992334842682, + 0.004980660043656826, + 0.006531968712806702, + 0.006275950465351343, + 0.007819507271051407, + 0.004905167501419783, + 0.006092173513025045, + 0.00465003028512001, + 0.006892080418765545, + 0.005662800744175911, + 0.0070377252995967865, + 0.005925019737333059, + 0.007618576753884554, + 0.006032397970557213, + 0.004223315976560116, + 0.0076320962980389595, + 0.014953372068703175, + 0.014567190781235695, + 0.016192808747291565, + 0.019810017198324203, + 0.027209490537643433, + 0.009295910596847534, + 0.020591912791132927, + 0.012966826558113098, + 0.01181683037430048, + 0.010466407984495163, + 0.014554696157574654, + 0.020693425089120865, + 0.020805295556783676, + 0.015456223860383034, + 0.03348715603351593, + 0.025614842772483826, + 0.016852376982569695, + 0.012427601963281631, + 0.016750691458582878, + 0.011200718581676483, + 0.012697783298790455, + 0.005683691240847111, + 0.013855247758328915, + 0.005859307013452053, + 0.010241596959531307, + 0.012584725394845009, + 0.006618979852646589, + 0.009174378588795662, + 0.009500650689005852, + 0.004251323640346527, + 0.00987106841057539, + 0.006155532319098711, + 0.019763007760047913, + 0.025421155616641045, + 0.023378394544124603, + 0.015495343133807182, + 0.015836216509342194, + 0.008544540032744408, + 0.011479288339614868, + 0.01102454960346222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.003193875541910529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818542532622814 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818542532622814, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818542532622814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.132341206073761, + 0.042666956782341, + 0.032286737114191055, + 0.08723250031471252, + 0.02543213963508606, + 0.06076687574386597, + 0.08661337941884995, + 0.10905523598194122, + 0.20801854133605957, + 0.21558640897274017 + ], + "credit_map": [ + 0.018709134310483932, + 0.017133161425590515, + 0.020672429352998734, + 0.016261691227555275, + 0.015381867997348309, + 0.014455866068601608, + 0.016015619039535522, + 0.013711443170905113, + 0.006826628930866718, + 0.005163757596164942, + 0.006458403542637825, + 0.002975242445245385, + 0.003668076591566205, + 0.005775001365691423, + 0.00557166850194335, + 0.006228178273886442, + 0.004810521379113197, + 0.005258309654891491, + 0.003985462710261345, + 0.004201181698590517, + 0.00495357858017087, + 0.002408374333754182, + 0.003403324633836746, + 0.003265982260927558, + 0.012320306152105331, + 0.01050302293151617, + 0.007121351547539234, + 0.012237885035574436, + 0.012264024466276169, + 0.006727761123329401, + 0.017586391419172287, + 0.008471748791635036, + 0.0038111130706965923, + 0.0038291735108941793, + 0.004408135544508696, + 0.00280963908880949, + 0.0027865793090313673, + 0.0026516346260905266, + 0.0026800662744790316, + 0.0024557984434068203, + 0.009615272283554077, + 0.00922023318707943, + 0.009610356763005257, + 0.0047267223708331585, + 0.005977411288768053, + 0.005256073083728552, + 0.007080272305756807, + 0.009280537255108356, + 0.007239834871143103, + 0.007592585403472185, + 0.009508932940661907, + 0.009134938009083271, + 0.013148171827197075, + 0.012072775512933731, + 0.016600238159298897, + 0.011315903626382351, + 0.01953263208270073, + 0.015117302536964417, + 0.020785219967365265, + 0.012694309465587139, + 0.013271227478981018, + 0.005940604023635387, + 0.014873482286930084, + 0.006840453017503023, + 0.02905237302184105, + 0.02150564268231392, + 0.014123405329883099, + 0.045957036316394806, + 0.02437959425151348, + 0.012923644855618477, + 0.04503556340932846, + 0.015041284263134003, + 0.032568059861660004, + 0.04491164907813072, + 0.037351176142692566, + 0.023768987506628036, + 0.027164913713932037, + 0.01350258570164442, + 0.019159719347953796, + 0.017159316688776016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.002818542532622814 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grainy black-and-white historical film, shot from a fixed high angle, draws the viewer's gaze into the heart of a Victorian-style laboratory cluttered with bizarre devices and scattered blueprints. At the bottom of the frame, an inventor, sporting a thick, carefully trimmed mustache and wearing a dark vest over a white turtleneck shirt, leans slightly forward, his brow furrowed, his gaze sharply focused on the workbench before him. In the center of the workbench sits a mysterious, smooth sphere of unidentified material, emitting a soft, eerie glow, the sole source of light in the dimly lit scene. The inventor, gripping a small brass hammer in his right hand, raises it steadily and deliberately, and with pinpoint precision, gently and rhythmically strikes the sphere's outer shell. With the final blow, a crack appears in the sphere's surface, and countless bright, irregular arcs of electricity burst forth, scattering wildly in all directions like uncontrolled white lightning. Under the dazzling electric light, the inventor's face and the corners of the laboratory are cast in violently shaking, deep shadows. The whole picture is full of strong visual contrast and a sense of tension and awe at the moment of breakthrough discovery.", + "seed": 21889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_527_seed_21889_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.29111901309806854, + "vlm_logprob": -0.29111901309806854, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.29111901309806854, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4133155345916748 + ], + "opposite_logprob_tokens": [ + -0.2883155345916748 + ], + "logprob_reward": -1.4133155345916748, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4133155345916748, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1933249831199646, + 0.07345852255821228, + 0.10385945439338684, + 0.2998596131801605, + 0.09879017621278763, + 0.03662547096610069, + 0.01618827134370804, + 0.02112751454114914, + 0.07399719953536987, + 0.08276879042387009 + ], + "credit_map": [ + 0.023949507623910904, + 0.018890995532274246, + 0.05210629105567932, + 0.020786836743354797, + 0.01795160211622715, + 0.02975016087293625, + 0.012968658469617367, + 0.016920899972319603, + 0.007022207602858543, + 0.00929528009146452, + 0.011234907433390617, + 0.00782819278538227, + 0.007074372842907906, + 0.017879674211144447, + 0.007343609351664782, + 0.005780274514108896, + 0.034871432930231094, + 0.008082693442702293, + 0.03534821420907974, + 0.005175980273634195, + 0.0061877816915512085, + 0.006859328597784042, + 0.002235767897218466, + 0.005098252557218075, + 0.0912850946187973, + 0.01810634508728981, + 0.12089511007070541, + 0.013302153907716274, + 0.017381824553012848, + 0.016604222357273102, + 0.007692213635891676, + 0.014592628926038742, + 0.015749679878354073, + 0.01986060105264187, + 0.017657604068517685, + 0.011083555407822132, + 0.0112579595297575, + 0.007901549339294434, + 0.00881656352430582, + 0.0064626531675457954, + 0.006024331320077181, + 0.003004157915711403, + 0.007966904900968075, + 0.0025921843480318785, + 0.003931458108127117, + 0.0062552085146307945, + 0.002668870147317648, + 0.004182353615760803, + 0.0016916814493015409, + 0.0024372257757931948, + 0.0025327049661427736, + 0.002178264083340764, + 0.0017171092331409454, + 0.0021605012007057667, + 0.0014105008449405432, + 0.0020602832082659006, + 0.0026451756712049246, + 0.0027881639543920755, + 0.003708786563947797, + 0.002284958492964506, + 0.0027274927124381065, + 0.0024993454571813345, + 0.00207265536300838, + 0.0024009353946894407, + 0.00906484853476286, + 0.01242463756352663, + 0.010815367102622986, + 0.0063390289433300495, + 0.007603819016367197, + 0.014881395734846592, + 0.006261792965233326, + 0.006606303155422211, + 0.013269992545247078, + 0.010011281818151474, + 0.009320775046944618, + 0.009535372257232666, + 0.010385884903371334, + 0.012414294295012951, + 0.0070772189646959305, + 0.010753972455859184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -1.4133155345916748 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026396818459033966 + ], + "opposite_logprob_tokens": [ + -3.6513967514038086 + ], + "logprob_reward": -0.026396818459033966, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026396818459033966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24360328912734985, + 0.07116641849279404, + 0.1485733836889267, + 0.16075530648231506, + 0.10915327817201614, + 0.06459914892911911, + 0.02682691253721714, + 0.03802695497870445, + 0.044727303087711334, + 0.09256793558597565 + ], + "credit_map": [ + 0.031043274328112602, + 0.026907220482826233, + 0.05982755124568939, + 0.027236150577664375, + 0.022869674488902092, + 0.038123857229948044, + 0.019183143973350525, + 0.018412431702017784, + 0.008907788433134556, + 0.0055544087663292885, + 0.017319930717349052, + 0.005255430471152067, + 0.005075746215879917, + 0.017504045739769936, + 0.004264311399310827, + 0.00728475721552968, + 0.05360064655542374, + 0.00904774572700262, + 0.05029221996665001, + 0.005149629898369312, + 0.0077291675843298435, + 0.013580757193267345, + 0.00262274406850338, + 0.006550489459186792, + 0.035284724086523056, + 0.011382226832211018, + 0.06763780862092972, + 0.007553693372756243, + 0.00984963309019804, + 0.012597735039889812, + 0.005125622730702162, + 0.011323862709105015, + 0.02268831804394722, + 0.018455103039741516, + 0.024460313841700554, + 0.009367634542286396, + 0.010867741890251637, + 0.010003363713622093, + 0.006532036699354649, + 0.0067787691950798035, + 0.011860189028084278, + 0.004990892950445414, + 0.020933400839567184, + 0.0038824845105409622, + 0.005969695746898651, + 0.00886372197419405, + 0.0030240940395742655, + 0.005074675194919109, + 0.0025467569939792156, + 0.0037403765600174665, + 0.004102046135812998, + 0.0039373114705085754, + 0.002986462786793709, + 0.004476532340049744, + 0.0019156364724040031, + 0.0031217921059578657, + 0.005676953587681055, + 0.0036724964156746864, + 0.007353456690907478, + 0.004824487958103418, + 0.004455420654267073, + 0.004083921667188406, + 0.003582117147743702, + 0.004378107842057943, + 0.0051484438590705395, + 0.0057931155897676945, + 0.005771301686763763, + 0.0039368062280118465, + 0.004771095234900713, + 0.010987663641571999, + 0.003907584119588137, + 0.004411294590681791, + 0.014690686017274857, + 0.014083782210946083, + 0.009954976849257946, + 0.00906915869563818, + 0.00960841029882431, + 0.013095139525830746, + 0.008677585981786251, + 0.013388208113610744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.026396818459033966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015174552099779248 + ], + "opposite_logprob_tokens": [ + -6.501517295837402 + ], + "logprob_reward": -0.0015174552099779248, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015174552099779248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16565212607383728, + 0.0664118081331253, + 0.15331457555294037, + 0.3148845434188843, + 0.08405870944261551, + 0.03267376124858856, + 0.05488315224647522, + 0.02252948470413685, + 0.04928579553961754, + 0.05630601570010185 + ], + "credit_map": [ + 0.019671378657221794, + 0.017565730959177017, + 0.04779985547065735, + 0.018302246928215027, + 0.016294343397021294, + 0.022995680570602417, + 0.011045614257454872, + 0.01197730377316475, + 0.0082632414996624, + 0.00871238298714161, + 0.013049698434770107, + 0.004937922116369009, + 0.006437776144593954, + 0.013832093216478825, + 0.005459005478769541, + 0.00571969710290432, + 0.044275883585214615, + 0.008484432473778725, + 0.0617297887802124, + 0.007430128287523985, + 0.008762156590819359, + 0.012404225766658783, + 0.0032463257666677237, + 0.006981641985476017, + 0.0946098044514656, + 0.018832556903362274, + 0.12884396314620972, + 0.014903188683092594, + 0.018487652763724327, + 0.01703360117971897, + 0.006341288331896067, + 0.01583251729607582, + 0.010101955384016037, + 0.017745239660143852, + 0.011338239535689354, + 0.011170797049999237, + 0.009823180735111237, + 0.008172506466507912, + 0.009152778424322605, + 0.006554023362696171, + 0.0038206393364816904, + 0.005130625329911709, + 0.004882462788373232, + 0.00345978164114058, + 0.0028447830118238926, + 0.005514502990990877, + 0.0029032365418970585, + 0.004117733798921108, + 0.005216680932790041, + 0.011624501086771488, + 0.011847808957099915, + 0.008904492482542992, + 0.004719676915556192, + 0.004068782087415457, + 0.0036357762292027473, + 0.0048654405400156975, + 0.003272887086495757, + 0.003708702279254794, + 0.004448477178812027, + 0.002324009547010064, + 0.002683881903067231, + 0.0024600711185485125, + 0.0015201553469523787, + 0.002111302688717842, + 0.007582309190183878, + 0.008582212030887604, + 0.007892516441643238, + 0.0041839927434921265, + 0.004785968456417322, + 0.008096610195934772, + 0.0039658574387431145, + 0.004196335095912218, + 0.007316958624869585, + 0.00670146569609642, + 0.006929483264684677, + 0.007713478524237871, + 0.005397876258939505, + 0.008249694481492043, + 0.007430574391037226, + 0.006566490046679974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0015174552099779248 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628476247191429 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628476247191429, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628476247191429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13107934594154358, + 0.03415212780237198, + 0.09076711535453796, + 0.4803164005279541, + 0.05095229297876358, + 0.024033786728978157, + 0.02816639468073845, + 0.03920496255159378, + 0.07480905205011368, + 0.04651859402656555 + ], + "credit_map": [ + 0.018435906618833542, + 0.015447975136339664, + 0.03796393796801567, + 0.012271115556359291, + 0.012515262700617313, + 0.0173930823802948, + 0.007521857973188162, + 0.00953019317239523, + 0.0036584248300641775, + 0.002943942090496421, + 0.0070194462314248085, + 0.002438680734485388, + 0.0034304752480238676, + 0.008505435660481453, + 0.002953286049887538, + 0.00320243276655674, + 0.03280532732605934, + 0.005551339127123356, + 0.0285839531570673, + 0.003933167550712824, + 0.004618055187165737, + 0.009261242114007473, + 0.001795629388652742, + 0.004218385089188814, + 0.14360302686691284, + 0.028701409697532654, + 0.2019997537136078, + 0.020204946398735046, + 0.02745046652853489, + 0.02589837647974491, + 0.008518991991877556, + 0.023939339444041252, + 0.008121971040964127, + 0.00506594218313694, + 0.011846275068819523, + 0.005475809797644615, + 0.005896854214370251, + 0.005869632121175528, + 0.004556347616016865, + 0.004119453486055136, + 0.00403724517673254, + 0.0036179993767291307, + 0.003953312058001757, + 0.002507190452888608, + 0.0020206645131111145, + 0.003285796847194433, + 0.0021050076466053724, + 0.002506568096578121, + 0.0026427072007209063, + 0.004745902959257364, + 0.004346563946455717, + 0.004894633311778307, + 0.002828615251928568, + 0.0032716665882617235, + 0.001877199043519795, + 0.0035591023042798042, + 0.004851614125072956, + 0.006682677660137415, + 0.006362875457853079, + 0.0053632534109056, + 0.004307331517338753, + 0.004766331519931555, + 0.003394732717424631, + 0.0034761412534862757, + 0.007791133597493172, + 0.012909263372421265, + 0.007415333297103643, + 0.009155129082500935, + 0.009305337443947792, + 0.011272085830569267, + 0.009076344780623913, + 0.00788441114127636, + 0.0074690645560622215, + 0.004183450248092413, + 0.0048039741814136505, + 0.0044031888246536255, + 0.007783167529851198, + 0.010171840898692608, + 0.0039414758794009686, + 0.0037624244578182697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.007628476247191429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736780982464552 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736780982464552, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736780982464552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21115483343601227, + 0.06565430015325546, + 0.1241389587521553, + 0.10100103169679642, + 0.07284273207187653, + 0.03339344635605812, + 0.05620477721095085, + 0.050114955753088, + 0.12167312949895859, + 0.16382178664207458 + ], + "credit_map": [ + 0.026163065806031227, + 0.021159540861845016, + 0.06598034501075745, + 0.020095139741897583, + 0.022109588608145714, + 0.029240449890494347, + 0.012154152616858482, + 0.014252565801143646, + 0.007990418933331966, + 0.005871977657079697, + 0.014863519929349422, + 0.005594967398792505, + 0.005957275629043579, + 0.014358245767652988, + 0.005974265281111002, + 0.005043630953878164, + 0.04329562187194824, + 0.010541579686105251, + 0.03440135717391968, + 0.005596135277301073, + 0.006339747458696365, + 0.01587066240608692, + 0.0029633694794028997, + 0.005130495876073837, + 0.029880406334996223, + 0.009227247908711433, + 0.03159105032682419, + 0.006452819798141718, + 0.0066238874569535255, + 0.007547155488282442, + 0.0043436684645712376, + 0.00533481128513813, + 0.011650561355054379, + 0.01113758236169815, + 0.014773203991353512, + 0.007305883336812258, + 0.006438771262764931, + 0.007247349247336388, + 0.007713053375482559, + 0.006576338782906532, + 0.0032544159330427647, + 0.003927908837795258, + 0.005452231038361788, + 0.003309560241177678, + 0.0037349220365285873, + 0.005968016106635332, + 0.003335844026878476, + 0.004410549998283386, + 0.006055252160876989, + 0.009247017093002796, + 0.009152377024292946, + 0.00919157825410366, + 0.006035540718585253, + 0.00604010745882988, + 0.004032692406326532, + 0.00645021116361022, + 0.0069199213758111, + 0.005307086743414402, + 0.015148020349442959, + 0.005828390829265118, + 0.00494236359372735, + 0.004905931185930967, + 0.003505461849272251, + 0.003557781223207712, + 0.012384939007461071, + 0.021849827840924263, + 0.012106371112167835, + 0.011420803144574165, + 0.016119664534926414, + 0.025942299515008926, + 0.011811396107077599, + 0.01003783755004406, + 0.029732001945376396, + 0.022041989490389824, + 0.01571054942905903, + 0.01660093106329441, + 0.02345927245914936, + 0.02877262607216835, + 0.015004961751401424, + 0.012499461881816387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.006736780982464552 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grainy black-and-white historical film, shot from a fixed high angle, draws the viewer's gaze into the heart of a Victorian-style laboratory cluttered with bizarre devices and scattered blueprints. At the bottom of the frame, an inventor, sporting a thick, carefully trimmed mustache and wearing a dark vest over a white turtleneck shirt, leans slightly forward, his brow furrowed, his gaze sharply focused on the workbench before him. In the center of the workbench sits a mysterious, smooth sphere of unidentified material, emitting a soft, eerie glow, the sole source of light in the dimly lit scene. The inventor, gripping a small brass hammer in his right hand, raises it steadily and deliberately, and with pinpoint precision, gently and rhythmically strikes the sphere's outer shell. With the final blow, a crack appears in the sphere's surface, and countless bright, irregular arcs of electricity burst forth, scattering wildly in all directions like uncontrolled white lightning. Under the dazzling electric light, the inventor's face and the corners of the laboratory are cast in violently shaking, deep shadows. The whole picture is full of strong visual contrast and a sense of tension and awe at the moment of breakthrough discovery.", + "seed": 21890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_13/index_527_seed_21890_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027447611186653377, + "vlm_logprob": -0.027447611186653377, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.027447611186653377, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09288406372070312 + ], + "opposite_logprob_tokens": [ + -2.592884063720703 + ], + "logprob_reward": -0.09288406372070312, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09288406372070312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15393880009651184, + 0.18213079869747162, + 0.12874281406402588, + 0.05156750977039337, + 0.3033572733402252, + 0.032742250710725784, + 0.026724226772785187, + 0.05629913881421089, + 0.04629192128777504, + 0.01820533163845539 + ], + "credit_map": [ + 0.012786048464477062, + 0.01380952913314104, + 0.024726076051592827, + 0.016839556396007538, + 0.03584986925125122, + 0.018308913335204124, + 0.015608283691108227, + 0.016010509803891182, + 0.010002563707530499, + 0.0050959475338459015, + 0.06454519182443619, + 0.005726272240281105, + 0.07395992428064346, + 0.009468047879636288, + 0.007500919513404369, + 0.005831927992403507, + 0.008242283016443253, + 0.006899791304022074, + 0.03162402659654617, + 0.006963091436773539, + 0.03895604610443115, + 0.01201971061527729, + 0.016688887029886246, + 0.007348979823291302, + 0.004496304783970118, + 0.003444864647462964, + 0.00700750108808279, + 0.006011324934661388, + 0.009320920333266258, + 0.006198707036674023, + 0.009790299460291862, + 0.005297582596540451, + 0.024637145921587944, + 0.030211592093110085, + 0.03812424838542938, + 0.03098111040890217, + 0.039132338017225266, + 0.05619119852781296, + 0.052028678357601166, + 0.03205093741416931, + 0.00244816392660141, + 0.0027386671863496304, + 0.0036115897819399834, + 0.00313760619610548, + 0.007818461395800114, + 0.004536312073469162, + 0.005218571051955223, + 0.0032328797969967127, + 0.0037616232875734568, + 0.00436654407531023, + 0.0027873762883245945, + 0.0019094444578513503, + 0.0034209294244647026, + 0.002038999693468213, + 0.004515743814408779, + 0.003923562355339527, + 0.005313885398209095, + 0.0017700075404718518, + 0.01801469177007675, + 0.004651715978980064, + 0.003313923254609108, + 0.00483081117272377, + 0.003119942732155323, + 0.015284156426787376, + 0.006819558329880238, + 0.0045235673896968365, + 0.008355177007615566, + 0.0068547362461686134, + 0.0049825566820800304, + 0.005493699572980404, + 0.005230438429862261, + 0.004032182972878218, + 0.00223147077485919, + 0.0016557818744331598, + 0.002280952176079154, + 0.001755503355525434, + 0.0033537496346980333, + 0.0023200020659714937, + 0.0026820350904017687, + 0.0019258357351645827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.09288406372070312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029879199340939522 + ], + "opposite_logprob_tokens": [ + -3.529879093170166 + ], + "logprob_reward": -0.029879199340939522, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029879199340939522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11258823424577713, + 0.438136488199234, + 0.07072409987449646, + 0.18367958068847656, + 0.057862963527441025, + 0.02637694776058197, + 0.01852230355143547, + 0.04165828973054886, + 0.024123763665556908, + 0.026327334344387054 + ], + "credit_map": [ + 0.009535983204841614, + 0.009260199032723904, + 0.017772221937775612, + 0.010827968828380108, + 0.02510770782828331, + 0.013600600883364677, + 0.01221889816224575, + 0.014264645054936409, + 0.02403450198471546, + 0.012294287793338299, + 0.15734541416168213, + 0.013645727187395096, + 0.18219418823719025, + 0.018591580912470818, + 0.017049266025424004, + 0.012981506995856762, + 0.005391229875385761, + 0.0036506955511868, + 0.017156746238470078, + 0.004137811250984669, + 0.021695811301469803, + 0.008054227568209171, + 0.005927090998739004, + 0.004710488952696323, + 0.01518784649670124, + 0.00966345053166151, + 0.032540462911129, + 0.01448787935078144, + 0.0555601492524147, + 0.019753316417336464, + 0.02187819592654705, + 0.01460826862603426, + 0.005344183649867773, + 0.005272693000733852, + 0.007500057108700275, + 0.00632166163995862, + 0.008165392093360424, + 0.010427232831716537, + 0.009004118852317333, + 0.005827618762850761, + 0.0015659421915188432, + 0.0013047176180407405, + 0.0031303600408136845, + 0.0015247728442773223, + 0.009528884664177895, + 0.0020205096807330847, + 0.005838841665536165, + 0.0014629184734076262, + 0.002184574492275715, + 0.0025519272312521935, + 0.001758966245688498, + 0.0017292029224336147, + 0.0022487491369247437, + 0.0015136987203732133, + 0.003966429270803928, + 0.0025687541346997023, + 0.004170396365225315, + 0.0013730167411267757, + 0.012254034169018269, + 0.0033827535808086395, + 0.0027640964835882187, + 0.0031187215354293585, + 0.0023841573856770992, + 0.012211112305521965, + 0.005513973068445921, + 0.0022060880437493324, + 0.0046680644154548645, + 0.0019704473670572042, + 0.0035264124162495136, + 0.002667699009180069, + 0.0018642314244061708, + 0.001706846640445292, + 0.002314042067155242, + 0.003431893652305007, + 0.002973098773509264, + 0.002693592570722103, + 0.005083673167973757, + 0.003746882313862443, + 0.003650803118944168, + 0.0024333472829312086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.029879199340939522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022044898942112923 + ], + "opposite_logprob_tokens": [ + -6.127204418182373 + ], + "logprob_reward": -0.0022044898942112923, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022044898942112923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.162689670920372, + 0.44416195154190063, + 0.03998580574989319, + 0.040559884160757065, + 0.031113993376493454, + 0.008923863060772419, + 0.056642454117536545, + 0.17854726314544678, + 0.018406502902507782, + 0.01896858960390091 + ], + "credit_map": [ + 0.012027488090097904, + 0.015587754547595978, + 0.025431256741285324, + 0.020267987623810768, + 0.038720108568668365, + 0.017590923234820366, + 0.02013651467859745, + 0.012927637435495853, + 0.024849815294146538, + 0.01188476849347353, + 0.16029572486877441, + 0.01354942936450243, + 0.1852087527513504, + 0.019043946638703346, + 0.01621180586516857, + 0.013117731548845768, + 0.00346036022529006, + 0.0020746286027133465, + 0.009510200470685959, + 0.0018739994848147035, + 0.010467471554875374, + 0.005677009467035532, + 0.0035404011141508818, + 0.0033817351795732975, + 0.0030491906218230724, + 0.0019190985476598144, + 0.007509224582463503, + 0.0034174819011241198, + 0.012788218446075916, + 0.003746827831491828, + 0.0054259453900158405, + 0.0027039002161473036, + 0.002513288054615259, + 0.0031697694212198257, + 0.0038153112400323153, + 0.003029940417036414, + 0.0056397924199700356, + 0.005458148196339607, + 0.004300334490835667, + 0.0031874082051217556, + 0.0007088537095114589, + 0.0009840538259595633, + 0.0011596218682825565, + 0.0008627359638921916, + 0.001074650208465755, + 0.001133979531005025, + 0.0019837275613099337, + 0.0010162403341382742, + 0.004650087561458349, + 0.0036941240541636944, + 0.005907751154154539, + 0.006061794236302376, + 0.008014420978724957, + 0.004087362438440323, + 0.00866713933646679, + 0.015559771098196507, + 0.018088553100824356, + 0.005939406808465719, + 0.061288021504879, + 0.012148371897637844, + 0.009997976943850517, + 0.017847543582320213, + 0.008557047694921494, + 0.04468034207820892, + 0.0018451636424288154, + 0.0022596719209104776, + 0.0024455233942717314, + 0.0031445438507944345, + 0.0032058937940746546, + 0.0019986466504633427, + 0.0017698060255497694, + 0.0017372550209984183, + 0.0018090693047270179, + 0.0015172604471445084, + 0.001686640433035791, + 0.0019023576751351357, + 0.004132579080760479, + 0.0023137314710766077, + 0.00374218518845737, + 0.0018647658871486783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0022044898942112923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640998784452677 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640998784452677, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640998784452677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07849446684122086, + 0.17890435457229614, + 0.01270246971398592, + 0.022243188694119453, + 0.043908070772886276, + 0.043209418654441833, + 0.07038448750972748, + 0.4604947566986084, + 0.06571316719055176, + 0.023945625871419907 + ], + "credit_map": [ + 0.006446131505072117, + 0.005894143600016832, + 0.014688603579998016, + 0.007979038171470165, + 0.016005773097276688, + 0.010457438416779041, + 0.007917310111224651, + 0.009106028825044632, + 0.00999236386269331, + 0.004752254579216242, + 0.06391695886850357, + 0.006052370648831129, + 0.07362274825572968, + 0.008435511030256748, + 0.006611853837966919, + 0.0055202944204211235, + 0.0013852168340235949, + 0.0011131702922284603, + 0.0020249218214303255, + 0.0018785519059747458, + 0.0024172840639948845, + 0.0009724964620545506, + 0.0016687397146597505, + 0.0012420887360349298, + 0.0022052270360291004, + 0.00164332939311862, + 0.0037736522499471903, + 0.0016903533833101392, + 0.005810769274830818, + 0.0026597457472234964, + 0.0026582281570881605, + 0.001801881822757423, + 0.004525939468294382, + 0.004542013164609671, + 0.005437761545181274, + 0.0044031147845089436, + 0.005472310818731785, + 0.008122723549604416, + 0.006893206853419542, + 0.004511001519858837, + 0.002543236128985882, + 0.003149096155539155, + 0.004886424168944359, + 0.0036070114001631737, + 0.012400764971971512, + 0.004465207457542419, + 0.00932235736399889, + 0.0028353186789900064, + 0.005045835394412279, + 0.0029589752666652203, + 0.00717870844528079, + 0.005679645575582981, + 0.008429622277617455, + 0.005350751336663961, + 0.01676814816892147, + 0.01897280290722847, + 0.05658244341611862, + 0.012120801024138927, + 0.15578846633434296, + 0.0322837270796299, + 0.02831595577299595, + 0.029058219864964485, + 0.021180037409067154, + 0.12516508996486664, + 0.007013370748609304, + 0.00689851026982069, + 0.008685193955898285, + 0.007978873327374458, + 0.012331732548773289, + 0.006585498806089163, + 0.009084353223443031, + 0.007135635241866112, + 0.0024795362260192633, + 0.001702035777270794, + 0.0020303118508309126, + 0.002340121427550912, + 0.004984353203326464, + 0.0028938932809978724, + 0.0048468331806361675, + 0.002668541856110096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.004640998784452677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a grainy black-and-white historical film style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a fixed high angle looking down into a Victorian-style laboratory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an inventor with a thick mustache, wearing a dark vest over a white turtleneck, visible leaning over a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the inventor raise a brass hammer in his right hand and strike a mysterious glowing sphere on the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sphere crack upon impact, causing bright electric arcs to burst forth and cast violently shaking deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629304192960262 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629304192960262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629304192960262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09028196334838867, + 0.22592516243457794, + 0.07761524617671967, + 0.07902348041534424, + 0.07034611701965332, + 0.04704621061682701, + 0.037351470440626144, + 0.1769246906042099, + 0.1523519605398178, + 0.04313366860151291 + ], + "credit_map": [ + 0.007390415761619806, + 0.009023858234286308, + 0.012683981098234653, + 0.009747229516506195, + 0.018367568030953407, + 0.013637470081448555, + 0.01169927790760994, + 0.007732170633971691, + 0.01257790345698595, + 0.006101098842918873, + 0.08081834763288498, + 0.007869867607951164, + 0.09317700564861298, + 0.0102579016238451, + 0.008131861686706543, + 0.006991175934672356, + 0.006396914832293987, + 0.004584672395139933, + 0.01637885719537735, + 0.0033310470171272755, + 0.019731542095541954, + 0.01376077625900507, + 0.006831998936831951, + 0.006599446292966604, + 0.007226403336971998, + 0.0038926212582737207, + 0.015192476101219654, + 0.005562674254179001, + 0.025906430557370186, + 0.007137084845453501, + 0.008874809369444847, + 0.005230981390923262, + 0.006825958378612995, + 0.007413512095808983, + 0.008439024910330772, + 0.00617032777518034, + 0.008792536333203316, + 0.013742237351834774, + 0.01105621736496687, + 0.007906308397650719, + 0.004005115944892168, + 0.003417734056711197, + 0.0075937421061098576, + 0.0033749714493751526, + 0.013770215213298798, + 0.0038948783185333014, + 0.007593657821416855, + 0.0033958994317799807, + 0.003963713999837637, + 0.0032000355422496796, + 0.004497681278735399, + 0.005051899701356888, + 0.004533812869340181, + 0.002820831723511219, + 0.008227367885410786, + 0.005056127905845642, + 0.016039131209254265, + 0.006930691655725241, + 0.057924628257751465, + 0.01016248855739832, + 0.009149710647761822, + 0.03351112827658653, + 0.007427209056913853, + 0.035779714584350586, + 0.0182772446423769, + 0.01527527254074812, + 0.02105267345905304, + 0.018401827663183212, + 0.029768580570816994, + 0.01551297027617693, + 0.018569596111774445, + 0.015493815764784813, + 0.0037494574207812548, + 0.0033191684633493423, + 0.003507869550958276, + 0.004724723286926746, + 0.006481769494712353, + 0.005147684831172228, + 0.009928802959620953, + 0.006274195853620768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.007629304192960262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_tensor([2918, 2918])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_tensor([2918, 2918])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59160c11fb52e91a9257220ecc9019310f627442 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_tensor([2918, 2918])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady close-up shot from a slightly overhead angle focuses squarely on a plump lemon, resting against a pure black, non-reflective background. The lemon's vibrant yellow skin is dotted with fine pits, giving it a natural, waxy sheen. A small, dry, green stem still remains at one end. Illuminated by the bright, even, cold studio light, every detail of the lemon is clearly visible. Suddenly, an invisible chill sweeps through, initiating the freezing process. Tiny, needle-like ice crystals form at a visible rate from the base of the lemon and spread upward. These crystalline frosts quickly connect and form, crawling across the entire rind like living vines, completely covering the original vibrant yellow beneath a thick, shimmering layer of white frost. The entire process is captured with unwavering clarity, visually depicting this remarkable natural physical transformation with a calm and objective approach to popular science, creating a visual experience rich in both focused and scientific beauty.", + "seed": 671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_2918_seed_671_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05848467033356428, + "vlm_logprob": -0.05848467033356428, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.05848467033356428, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009461325593292713 + ], + "opposite_logprob_tokens": [ + -7.134461402893066 + ], + "logprob_reward": -0.009461325593292713, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.009461325593292713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13562197983264923, + 0.05885763466358185, + 0.031090058386325836, + 0.06904320418834686, + 0.054553527384996414, + 0.05685049295425415, + 0.17535275220870972, + 0.20000502467155457, + 0.14525753259658813, + 0.07336770743131638 + ], + "credit_map": [ + 0.02957776188850403, + 0.010153988376259804, + 0.02167799137532711, + 0.012474133633077145, + 0.009852028451859951, + 0.01166556030511856, + 0.020658789202570915, + 0.019561735913157463, + 0.010614600963890553, + 0.007180685177445412, + 0.007674123160541058, + 0.006181192584335804, + 0.006042197346687317, + 0.005976413376629353, + 0.010812581516802311, + 0.004375847522169352, + 0.008588211610913277, + 0.002992817899212241, + 0.004542730748653412, + 0.0019334707176312804, + 0.0018468095222488046, + 0.003692994359880686, + 0.0028737965039908886, + 0.004619232378900051, + 0.01026589423418045, + 0.0050758663564920425, + 0.010432645678520203, + 0.006588701158761978, + 0.006485427264124155, + 0.006780967116355896, + 0.017032615840435028, + 0.0063810949213802814, + 0.012401818297803402, + 0.00375940278172493, + 0.011332222260534763, + 0.004731993656605482, + 0.0031171070877462626, + 0.0059609562158584595, + 0.008198803290724754, + 0.005051231011748314, + 0.013710546307265759, + 0.0035138768143951893, + 0.010619868524372578, + 0.005895616952329874, + 0.003470528172329068, + 0.003521020757034421, + 0.009114575572311878, + 0.007004466839134693, + 0.059691764414310455, + 0.007493351586163044, + 0.030152736231684685, + 0.009611217305064201, + 0.006094255484640598, + 0.010916867293417454, + 0.013455718755722046, + 0.037936873733997345, + 0.0336240716278553, + 0.017032738775014877, + 0.03543670475482941, + 0.02670161984860897, + 0.021700266748666763, + 0.01899724453687668, + 0.024843409657478333, + 0.02166898548603058, + 0.02163877710700035, + 0.013815938495099545, + 0.026692550629377365, + 0.01578613743185997, + 0.028012504801154137, + 0.009995019063353539, + 0.016467643901705742, + 0.012848994694650173, + 0.009931249544024467, + 0.008205858059227467, + 0.006735749077051878, + 0.010099001228809357, + 0.010313261300325394, + 0.005829255096614361, + 0.014962338842451572, + 0.00729100126773119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.009461325593292713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561243176460266 + ], + "opposite_logprob_tokens": [ + -1.6006124019622803 + ], + "logprob_reward": -0.22561243176460266, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561243176460266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29379382729530334, + 0.184670552611351, + 0.10294926166534424, + 0.040875278413295746, + 0.05992806702852249, + 0.06111278757452965, + 0.03665296360850334, + 0.03395675867795944, + 0.032419074326753616, + 0.1536414474248886 + ], + "credit_map": [ + 0.08329043537378311, + 0.019154949113726616, + 0.06227929890155792, + 0.024412894621491432, + 0.017002226784825325, + 0.019768448546528816, + 0.0445091612637043, + 0.023376399651169777, + 0.05261006951332092, + 0.013611231930553913, + 0.04123783856630325, + 0.015226646326482296, + 0.010435246862471104, + 0.01663265749812126, + 0.017726043239235878, + 0.017190810292959213, + 0.027439748868346214, + 0.0078114354982972145, + 0.02393590472638607, + 0.008746525272727013, + 0.006144718267023563, + 0.010192782618105412, + 0.00797305628657341, + 0.010705079883337021, + 0.011982418596744537, + 0.004236061591655016, + 0.008479778654873371, + 0.0033827300649136305, + 0.0025192033499479294, + 0.0031234093476086855, + 0.0029346065130084753, + 0.004217072855681181, + 0.019559932872653008, + 0.0033515258692204952, + 0.016217034310102463, + 0.00388968363404274, + 0.002877789316698909, + 0.00420575262978673, + 0.005016400944441557, + 0.004809947684407234, + 0.01683744788169861, + 0.0036046821624040604, + 0.014393068850040436, + 0.005463702138513327, + 0.0037504020147025585, + 0.004391383845359087, + 0.007917315699160099, + 0.004754785448312759, + 0.007778631057590246, + 0.0018859104020521045, + 0.006282269023358822, + 0.003725632792338729, + 0.002062789397314191, + 0.0030068792402744293, + 0.0054828994907438755, + 0.006427953019738197, + 0.005980742629617453, + 0.0032208850607275963, + 0.004631223157048225, + 0.004128727596253157, + 0.004456756170839071, + 0.003387357108294964, + 0.0044145286083221436, + 0.003736536018550396, + 0.0046798125840723515, + 0.0029614907689392567, + 0.0040553174912929535, + 0.003583580255508423, + 0.004814819898456335, + 0.0034393458627164364, + 0.0039324089884757996, + 0.004952300805598497, + 0.03350304439663887, + 0.013472229242324829, + 0.01641669124364853, + 0.017061199992895126, + 0.019807731732726097, + 0.014304762706160545, + 0.023846246302127838, + 0.015229552984237671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.22561243176460266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560899369418621 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012560899369418621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560899369418621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10675667226314545, + 0.08116485923528671, + 0.10998977720737457, + 0.03622013330459595, + 0.03808509185910225, + 0.03236551955342293, + 0.06496712565422058, + 0.10570672899484634, + 0.3532009720802307, + 0.07154306024312973 + ], + "credit_map": [ + 0.020834987983107567, + 0.01249706745147705, + 0.017815839499235153, + 0.009517218917608261, + 0.015321980230510235, + 0.009981985203921795, + 0.012666392140090466, + 0.008121208287775517, + 0.012829978950321674, + 0.005550284869968891, + 0.01329970546066761, + 0.007117439527064562, + 0.01146483514457941, + 0.005791566334664822, + 0.016326667740941048, + 0.008784382604062557, + 0.013099675998091698, + 0.010565049014985561, + 0.017258575186133385, + 0.011136943474411964, + 0.011254793033003807, + 0.007902087643742561, + 0.030350305140018463, + 0.008422361686825752, + 0.004778780974447727, + 0.0034066587686538696, + 0.0055317627266049385, + 0.004454676527529955, + 0.0034025053028017282, + 0.003045854391530156, + 0.008176249451935291, + 0.00342364888638258, + 0.006827483884990215, + 0.0035786728840321302, + 0.006331716198474169, + 0.0033603340853005648, + 0.0029543363489210606, + 0.0037481761537492275, + 0.007610124535858631, + 0.0036742473021149635, + 0.007444240618497133, + 0.0027235012967139482, + 0.005119741428643465, + 0.002769574522972107, + 0.003034501802176237, + 0.0036357862409204245, + 0.004755751695483923, + 0.0028824242763221264, + 0.014046778902411461, + 0.0038344396743923426, + 0.010960545390844345, + 0.007246420253068209, + 0.0036665964871644974, + 0.003867538645863533, + 0.010223289020359516, + 0.011121517047286034, + 0.01669124886393547, + 0.008636785671114922, + 0.01608222909271717, + 0.012950741685926914, + 0.012475959025323391, + 0.009324980899691582, + 0.014548237435519695, + 0.014996560290455818, + 0.06487209349870682, + 0.040083836764097214, + 0.0182331595569849, + 0.02250773459672928, + 0.13039329648017883, + 0.025736220180988312, + 0.036508407443761826, + 0.014866258017718792, + 0.009962080977857113, + 0.006620231084525585, + 0.00590388011187315, + 0.007999640889465809, + 0.011474767699837685, + 0.007277147378772497, + 0.01153956726193428, + 0.01076575182378292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.012560899369418621 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017177602276206017 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017177602276206017, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017177602276206017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09410200268030167, + 0.03797292336821556, + 0.03222774341702461, + 0.02953910268843174, + 0.01541212946176529, + 0.01396461483091116, + 0.0677429735660553, + 0.13049651682376862, + 0.4401358962059021, + 0.1384061574935913 + ], + "credit_map": [ + 0.025037281215190887, + 0.010452461428940296, + 0.020144082605838776, + 0.007477848324924707, + 0.0060538602992892265, + 0.008319517597556114, + 0.00934730377048254, + 0.007269640453159809, + 0.010987285524606705, + 0.0040822415612638, + 0.00879795290529728, + 0.0029574937652796507, + 0.002465530065819621, + 0.003013501176610589, + 0.0026957702357321978, + 0.0029731409158557653, + 0.006617103237658739, + 0.0038272340316325426, + 0.005572854541242123, + 0.0026038079522550106, + 0.002212420105934143, + 0.003677243134006858, + 0.005160689353942871, + 0.0025563864037394524, + 0.004707535728812218, + 0.001978125423192978, + 0.005460695829242468, + 0.0026382652577012777, + 0.0025944227818399668, + 0.0018728886498138309, + 0.006846741307526827, + 0.003440425032749772, + 0.0036613824777305126, + 0.0009557505836710334, + 0.003168505849316716, + 0.001642871997319162, + 0.0009330532629974186, + 0.0019933723378926516, + 0.001518366625532508, + 0.001538824406452477, + 0.0031645367853343487, + 0.0014349387492984533, + 0.002376281889155507, + 0.0011461707763373852, + 0.001495744101703167, + 0.001343173673376441, + 0.0016243678983300924, + 0.001379398861899972, + 0.010417815297842026, + 0.004371676128357649, + 0.012439216487109661, + 0.007760221138596535, + 0.0045647104270756245, + 0.005908623803406954, + 0.009881456382572651, + 0.01239924505352974, + 0.022633157670497894, + 0.01042884774506092, + 0.03416731208562851, + 0.014097635634243488, + 0.009628106839954853, + 0.010321145877242088, + 0.011403972283005714, + 0.0178163293749094, + 0.08154940605163574, + 0.04987117648124695, + 0.025705594569444656, + 0.027123205363750458, + 0.15750275552272797, + 0.03099452331662178, + 0.047127217054367065, + 0.020261984318494797, + 0.015161016024649143, + 0.014740103855729103, + 0.010730648413300514, + 0.02341366931796074, + 0.019668273627758026, + 0.013225870206952095, + 0.03218791261315346, + 0.00927865318953991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0017177602276206017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307093471288681 + ], + "opposite_logprob_tokens": [ + -3.1680710315704346 + ], + "logprob_reward": -0.04307093471288681, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307093471288681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2335306853055954, + 0.057699937373399734, + 0.07027318328619003, + 0.058031871914863586, + 0.08256856352090836, + 0.06205693632364273, + 0.1223001703619957, + 0.11728261411190033, + 0.11177030205726624, + 0.08448575437068939 + ], + "credit_map": [ + 0.039455439895391464, + 0.016238680109381676, + 0.03537334129214287, + 0.021624062210321426, + 0.018222207203507423, + 0.019647791981697083, + 0.0541711151599884, + 0.028798040002584457, + 0.016571538522839546, + 0.004166652914136648, + 0.011574682779610157, + 0.0042662532068789005, + 0.0033300314098596573, + 0.004767712205648422, + 0.007661106530576944, + 0.0053619565442204475, + 0.012314624153077602, + 0.005812180228531361, + 0.01220693439245224, + 0.006443025078624487, + 0.005980319809168577, + 0.0059965322725474834, + 0.01584390364587307, + 0.005675671622157097, + 0.012531829997897148, + 0.0054227691143751144, + 0.010190404020249844, + 0.005427323281764984, + 0.003511612070724368, + 0.008080022409558296, + 0.006412706337869167, + 0.006455198395997286, + 0.023249881342053413, + 0.003951700404286385, + 0.016758274286985397, + 0.007926810532808304, + 0.004507828038185835, + 0.005567003972828388, + 0.01131027564406395, + 0.009296789765357971, + 0.012750068679451942, + 0.004496004898101091, + 0.014308393001556396, + 0.005571457091718912, + 0.00397626543417573, + 0.005452616605907679, + 0.009813912212848663, + 0.005688214674592018, + 0.04054876044392586, + 0.00560458330437541, + 0.023724710568785667, + 0.006364054046571255, + 0.005184524226933718, + 0.0068260193802416325, + 0.01273458730429411, + 0.021312933415174484, + 0.018339047208428383, + 0.012945315800607204, + 0.014255470596253872, + 0.015600770711898804, + 0.021203652024269104, + 0.007917961105704308, + 0.014765945263206959, + 0.012254451401531696, + 0.018331578001379967, + 0.011228174902498722, + 0.0063347467221319675, + 0.008708212524652481, + 0.03536015376448631, + 0.00973519403487444, + 0.011027329601347446, + 0.01104491576552391, + 0.010389944538474083, + 0.01025688648223877, + 0.0068367235362529755, + 0.011471286416053772, + 0.012443901039659977, + 0.006911690812557936, + 0.019303545355796814, + 0.006871778052300215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.04307093471288681 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady close-up shot from a slightly overhead angle focuses squarely on a plump lemon, resting against a pure black, non-reflective background. The lemon's vibrant yellow skin is dotted with fine pits, giving it a natural, waxy sheen. A small, dry, green stem still remains at one end. Illuminated by the bright, even, cold studio light, every detail of the lemon is clearly visible. Suddenly, an invisible chill sweeps through, initiating the freezing process. Tiny, needle-like ice crystals form at a visible rate from the base of the lemon and spread upward. These crystalline frosts quickly connect and form, crawling across the entire rind like living vines, completely covering the original vibrant yellow beneath a thick, shimmering layer of white frost. The entire process is captured with unwavering clarity, visually depicting this remarkable natural physical transformation with a calm and objective approach to popular science, creating a visual experience rich in both focused and scientific beauty.", + "seed": 672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_2918_seed_672_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02240682060364634, + "vlm_logprob": -0.02240682060364634, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.02240682060364634, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008091172203421593 + ], + "opposite_logprob_tokens": [ + -6.633090972900391 + ], + "logprob_reward": -0.008091172203421593, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.008091172203421593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18843835592269897, + 0.07024415582418442, + 0.04460832104086876, + 0.052637435495853424, + 0.11619804799556732, + 0.2326488196849823, + 0.1151772141456604, + 0.06088467687368393, + 0.05486380308866501, + 0.06429912894964218 + ], + "credit_map": [ + 0.014812377281486988, + 0.03390795364975929, + 0.013410203158855438, + 0.05486702173948288, + 0.01698497124016285, + 0.01526391226798296, + 0.020833708345890045, + 0.018358223140239716, + 0.016057629138231277, + 0.008861131966114044, + 0.005010155029594898, + 0.0076505327597260475, + 0.00593428872525692, + 0.008873013779520988, + 0.006409262772649527, + 0.011448143050074577, + 0.005372693296521902, + 0.00826499704271555, + 0.003388013457879424, + 0.006690063048154116, + 0.007764917798340321, + 0.004457914270460606, + 0.005578207317739725, + 0.0030915196985006332, + 0.010870311409235, + 0.003876877250149846, + 0.005701374728232622, + 0.004998802673071623, + 0.004824677482247353, + 0.0061444491147994995, + 0.006794929504394531, + 0.009426020085811615, + 0.022083641961216927, + 0.00566665967926383, + 0.008597067557275295, + 0.012000294402241707, + 0.008251487277448177, + 0.03222176432609558, + 0.009513399563729763, + 0.017863741144537926, + 0.04800849035382271, + 0.015950769186019897, + 0.04582465440034866, + 0.02899039536714554, + 0.012950742617249489, + 0.028210612013936043, + 0.023154284805059433, + 0.02955888956785202, + 0.018932027742266655, + 0.009969225153326988, + 0.026523470878601074, + 0.020013000816106796, + 0.009362530894577503, + 0.00801332388073206, + 0.015165625140070915, + 0.007198017090559006, + 0.008509458042681217, + 0.007260264363139868, + 0.010520475916564465, + 0.006358136422932148, + 0.005674279760569334, + 0.006479890551418066, + 0.007641990203410387, + 0.008440186269581318, + 0.006890134885907173, + 0.005251740105450153, + 0.0070963576436042786, + 0.005566207692027092, + 0.006113939918577671, + 0.004917909391224384, + 0.009516457095742226, + 0.009511065669357777, + 0.005711320787668228, + 0.007266673259437084, + 0.005477539729326963, + 0.0061394209042191505, + 0.008034909144043922, + 0.006694722454994917, + 0.010050691664218903, + 0.01492385659366846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.008091172203421593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048696450889110565 + ], + "opposite_logprob_tokens": [ + -3.048696517944336 + ], + "logprob_reward": -0.048696450889110565, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048696450889110565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0916980430483818, + 0.3105977773666382, + 0.07382532954216003, + 0.0712716281414032, + 0.07062429934740067, + 0.06484275311231613, + 0.03776505962014198, + 0.06193700060248375, + 0.12516100704669952, + 0.09227706491947174 + ], + "credit_map": [ + 0.0048630633391439915, + 0.016392135992646217, + 0.007679284084588289, + 0.026674257591366768, + 0.007499529980123043, + 0.007764801848679781, + 0.010136600583791733, + 0.01068837009370327, + 0.057947348803281784, + 0.03371639922261238, + 0.024740900844335556, + 0.04183463752269745, + 0.03599102050065994, + 0.03181029111146927, + 0.03401714190840721, + 0.05054006725549698, + 0.008608308620750904, + 0.01517312042415142, + 0.005348742939531803, + 0.009448492899537086, + 0.014139905571937561, + 0.004730250220745802, + 0.009057414717972279, + 0.007319094147533178, + 0.00945287849754095, + 0.007813206873834133, + 0.008703489787876606, + 0.007357781752943993, + 0.010145059786736965, + 0.010780008509755135, + 0.00933768879622221, + 0.007681521587073803, + 0.012403074651956558, + 0.0043693119660019875, + 0.007624372374266386, + 0.010481567122042179, + 0.00760122574865818, + 0.013358362019062042, + 0.009459859691560268, + 0.00532653508707881, + 0.009110910817980766, + 0.007251049857586622, + 0.007683966308832169, + 0.012647885829210281, + 0.006727361585944891, + 0.007268358021974564, + 0.008368614129722118, + 0.005784607492387295, + 0.005536938551813364, + 0.003561887424439192, + 0.006385765504091978, + 0.00649236747995019, + 0.0038569504395127296, + 0.0031093063298612833, + 0.004992715083062649, + 0.0038291297387331724, + 0.012729550711810589, + 0.007323438301682472, + 0.013381543569266796, + 0.0066894949413836, + 0.005789593327790499, + 0.004400159697979689, + 0.006505596451461315, + 0.0051176236011087894, + 0.019335981458425522, + 0.010603830218315125, + 0.015173584222793579, + 0.015381868928670883, + 0.021319374442100525, + 0.010036291554570198, + 0.01884518191218376, + 0.014464901760220528, + 0.00862107053399086, + 0.014001132920384407, + 0.007631084881722927, + 0.011900718323886395, + 0.010079558938741684, + 0.008621130138635635, + 0.013562532141804695, + 0.017859848216176033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.048696450889110565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646694287657738 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646694287657738, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646694287657738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14940334856510162, + 0.08148659020662308, + 0.04128548502922058, + 0.11540064960718155, + 0.21692423522472382, + 0.049537044018507004, + 0.14564773440361023, + 0.11365144699811935, + 0.045492663979530334, + 0.04117078334093094 + ], + "credit_map": [ + 0.011502468027174473, + 0.028384605422616005, + 0.014915681444108486, + 0.021199971437454224, + 0.01596819795668125, + 0.017769942060112953, + 0.022903455421328545, + 0.016759036108851433, + 0.007132717873901129, + 0.010937456041574478, + 0.01132298819720745, + 0.008578086271882057, + 0.012232881970703602, + 0.010305575095117092, + 0.010864829644560814, + 0.010112058371305466, + 0.0030596626456826925, + 0.007354486268013716, + 0.0037135391030460596, + 0.005704647861421108, + 0.008039109408855438, + 0.004242601338773966, + 0.005540905054658651, + 0.003630533814430237, + 0.01116718165576458, + 0.013325167819857597, + 0.012009991332888603, + 0.010272327810525894, + 0.025874989107251167, + 0.01651405543088913, + 0.015272821299731731, + 0.010964120738208294, + 0.04326207563281059, + 0.012906420975923538, + 0.02654709480702877, + 0.02345566637814045, + 0.016139840707182884, + 0.05315990373492241, + 0.01934828609228134, + 0.02210495062172413, + 0.0077454508282244205, + 0.004307234194129705, + 0.004675344098359346, + 0.008522783406078815, + 0.004453137516975403, + 0.006563990376889706, + 0.00807114876806736, + 0.005197955295443535, + 0.01917557790875435, + 0.010896308347582817, + 0.0199339147657156, + 0.034674037247896194, + 0.016600817441940308, + 0.010551775805652142, + 0.01919703185558319, + 0.01461826078593731, + 0.02738434076309204, + 0.011624165810644627, + 0.026819758117198944, + 0.010957383550703526, + 0.008158644661307335, + 0.009435116313397884, + 0.010671064257621765, + 0.00860096886754036, + 0.004942385014146566, + 0.004482120741158724, + 0.005983099807053804, + 0.006298400927335024, + 0.005212575662881136, + 0.0048598237335681915, + 0.007540172431617975, + 0.006174085661768913, + 0.0034210204612463713, + 0.00354596134275198, + 0.002981623401865363, + 0.005839622113853693, + 0.004476027563214302, + 0.004896197002381086, + 0.0056740352883934975, + 0.010336295701563358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.004646694287657738 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00194193993229419 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.00194193993229419, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00194193993229419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06528809666633606, + 0.09491259604692459, + 0.04508979246020317, + 0.06732838600873947, + 0.06452356278896332, + 0.1623481810092926, + 0.13601240515708923, + 0.20292896032333374, + 0.11005385220050812, + 0.0515141487121582 + ], + "credit_map": [ + 0.006766388192772865, + 0.009153731167316437, + 0.006298969499766827, + 0.009890335611999035, + 0.0075384038500487804, + 0.008625449612736702, + 0.009137662127614021, + 0.00787715706974268, + 0.013357899151742458, + 0.011435802094638348, + 0.009120667353272438, + 0.011103127151727676, + 0.012864544056355953, + 0.012445634230971336, + 0.013608451932668686, + 0.010976467281579971, + 0.004056195262819529, + 0.006409475114196539, + 0.003518357640132308, + 0.007391964551061392, + 0.01022100355476141, + 0.003917268477380276, + 0.005384375806897879, + 0.004191155545413494, + 0.00721586961299181, + 0.008829567581415176, + 0.0075838095508515835, + 0.006884091068059206, + 0.012856975197792053, + 0.007794572971761227, + 0.009194428101181984, + 0.006969069596379995, + 0.010768220759928226, + 0.0044892700389027596, + 0.00912520382553339, + 0.008738727308809757, + 0.004559131804853678, + 0.011320948600769043, + 0.0068495385348796844, + 0.008672519586980343, + 0.02715255878865719, + 0.013209839351475239, + 0.03622923791408539, + 0.025856642052531242, + 0.012996694073081017, + 0.014895942993462086, + 0.020544687286019325, + 0.011462589725852013, + 0.020736172795295715, + 0.011625833809375763, + 0.027862491086125374, + 0.020654108375310898, + 0.013295701704919338, + 0.01103239692747593, + 0.014232506975531578, + 0.016573190689086914, + 0.04054136946797371, + 0.019558273255825043, + 0.035443905740976334, + 0.0218089297413826, + 0.019645188003778458, + 0.01932957023382187, + 0.02557668648660183, + 0.021025050431489944, + 0.011876074597239494, + 0.009259120561182499, + 0.010163577273488045, + 0.01269523799419403, + 0.015703042969107628, + 0.009860397316515446, + 0.021523434668779373, + 0.0189729705452919, + 0.00447556609287858, + 0.004763305187225342, + 0.0037522988859564066, + 0.006833284627646208, + 0.006159315351396799, + 0.005091164726763964, + 0.006593388970941305, + 0.013845824636518955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.00194193993229419 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048657845705747604 + ], + "opposite_logprob_tokens": [ + -3.0486578941345215 + ], + "logprob_reward": -0.048657845705747604, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048657845705747604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0768919438123703, + 0.06277827173471451, + 0.28908291459083557, + 0.19247184693813324, + 0.06407945603132248, + 0.06381862610578537, + 0.0449845977127552, + 0.04872407764196396, + 0.07755123823881149, + 0.07961709797382355 + ], + "credit_map": [ + 0.007062313612550497, + 0.011638833209872246, + 0.006396813318133354, + 0.022383900359272957, + 0.005764196626842022, + 0.006143377628177404, + 0.007309650536626577, + 0.01019285898655653, + 0.008027787320315838, + 0.011272303760051727, + 0.0067550684325397015, + 0.008025613613426685, + 0.010061223059892654, + 0.0059830667451024055, + 0.006262809969484806, + 0.00639039883390069, + 0.028100797906517982, + 0.0466863214969635, + 0.020157206803560257, + 0.0522664412856102, + 0.059662457555532455, + 0.022659767419099808, + 0.03825793415307999, + 0.02129199728369713, + 0.05150061845779419, + 0.009768806397914886, + 0.011857960373163223, + 0.011531876400113106, + 0.015566807240247726, + 0.01962815597653389, + 0.013839172199368477, + 0.05877844616770744, + 0.007553476840257645, + 0.006822568830102682, + 0.004275130573660135, + 0.015600571408867836, + 0.004657652694731951, + 0.008826100267469883, + 0.008884072303771973, + 0.007459887769073248, + 0.009595208801329136, + 0.0034512411803007126, + 0.007690600119531155, + 0.008158368058502674, + 0.00588233582675457, + 0.005563455168157816, + 0.008198464289307594, + 0.01527895126491785, + 0.006923429202288389, + 0.005707388278096914, + 0.008152921684086323, + 0.007347781211137772, + 0.005143326707184315, + 0.0033216082956641912, + 0.005335111171007156, + 0.0030530302319675684, + 0.009288677014410496, + 0.005581263918429613, + 0.009595971554517746, + 0.006311019416898489, + 0.00520735327154398, + 0.003572887508198619, + 0.0060795946046710014, + 0.0030873077921569347, + 0.007757595740258694, + 0.006671193987131119, + 0.008884181268513203, + 0.00917265098541975, + 0.01004322525113821, + 0.0069963023997843266, + 0.015148106031119823, + 0.012877983041107655, + 0.007005543448030949, + 0.009762972593307495, + 0.007233426906168461, + 0.00904969871044159, + 0.008727410808205605, + 0.008183469995856285, + 0.011324445717036724, + 0.018330130726099014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.048657845705747604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady close-up shot from a slightly overhead angle focuses squarely on a plump lemon, resting against a pure black, non-reflective background. The lemon's vibrant yellow skin is dotted with fine pits, giving it a natural, waxy sheen. A small, dry, green stem still remains at one end. Illuminated by the bright, even, cold studio light, every detail of the lemon is clearly visible. Suddenly, an invisible chill sweeps through, initiating the freezing process. Tiny, needle-like ice crystals form at a visible rate from the base of the lemon and spread upward. These crystalline frosts quickly connect and form, crawling across the entire rind like living vines, completely covering the original vibrant yellow beneath a thick, shimmering layer of white frost. The entire process is captured with unwavering clarity, visually depicting this remarkable natural physical transformation with a calm and objective approach to popular science, creating a visual experience rich in both focused and scientific beauty.", + "seed": 673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_2918_seed_673_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027328991144895554, + "vlm_logprob": -0.027328991144895554, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.027328991144895554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007791721727699041 + ], + "opposite_logprob_tokens": [ + -6.882791519165039 + ], + "logprob_reward": -0.007791721727699041, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.007791721727699041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10650471597909927, + 0.14673402905464172, + 0.10245517641305923, + 0.05129702016711235, + 0.09082071483135223, + 0.050594646483659744, + 0.13140058517456055, + 0.15017083287239075, + 0.10073579102754593, + 0.06928651034832001 + ], + "credit_map": [ + 0.01657583750784397, + 0.0077942414209246635, + 0.019242379814386368, + 0.009127969853579998, + 0.019458334892988205, + 0.011135846376419067, + 0.01160179078578949, + 0.011568308807909489, + 0.01830635406076908, + 0.011111418716609478, + 0.026110714301466942, + 0.010523875243961811, + 0.023752836510539055, + 0.018589820712804794, + 0.01896498166024685, + 0.019374027848243713, + 0.013192698359489441, + 0.00757972477003932, + 0.018472589552402496, + 0.006569241173565388, + 0.014558564871549606, + 0.013177068904042244, + 0.013707255944609642, + 0.015198031440377235, + 0.0073247929103672504, + 0.004214177373796701, + 0.007856568321585655, + 0.006227714475244284, + 0.005713647697120905, + 0.006323734764009714, + 0.005534816533327103, + 0.008101568557322025, + 0.013472840189933777, + 0.010367922484874725, + 0.015571190975606441, + 0.014818545430898666, + 0.00982306431978941, + 0.0105365552008152, + 0.007936730980873108, + 0.00829386431723833, + 0.008770609274506569, + 0.007329915184527636, + 0.007643172517418861, + 0.005553633440285921, + 0.005171558354049921, + 0.005167817231267691, + 0.005576174706220627, + 0.005381766706705093, + 0.0164199061691761, + 0.012183336541056633, + 0.018817218020558357, + 0.01983155868947506, + 0.015834873542189598, + 0.015089664608240128, + 0.016652531921863556, + 0.01657150126993656, + 0.019962536171078682, + 0.015088304877281189, + 0.02669351175427437, + 0.02271873876452446, + 0.015021554194390774, + 0.014208587817847729, + 0.02428748644888401, + 0.012190092355012894, + 0.014582105912268162, + 0.012714860029518604, + 0.01611780747771263, + 0.018915100023150444, + 0.007741441484540701, + 0.00801690760999918, + 0.011081100441515446, + 0.011566465720534325, + 0.005781168583780527, + 0.007562172133475542, + 0.007774587254971266, + 0.006584418471902609, + 0.008517520502209663, + 0.011955968104302883, + 0.011788387782871723, + 0.00932228472083807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.007791721727699041 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006505131721497 + ], + "opposite_logprob_tokens": [ + -2.6950650215148926 + ], + "logprob_reward": -0.07006505131721497, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006505131721497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17790170013904572, + 0.11155498027801514, + 0.04454758018255234, + 0.04239920526742935, + 0.14229990541934967, + 0.11651730537414551, + 0.12171409279108047, + 0.13913436233997345, + 0.031159667298197746, + 0.07277122139930725 + ], + "credit_map": [ + 0.04801115393638611, + 0.014379410073161125, + 0.0387682281434536, + 0.016990747302770615, + 0.014329013414680958, + 0.014067589305341244, + 0.011162030510604382, + 0.020193524658679962, + 0.02529434859752655, + 0.009754023514688015, + 0.02183966524899006, + 0.014941945672035217, + 0.007526490371674299, + 0.011861938051879406, + 0.007894576527178288, + 0.012441991828382015, + 0.008211106061935425, + 0.0035239653661847115, + 0.009026844054460526, + 0.004707221873104572, + 0.005357629153877497, + 0.0047821104526519775, + 0.0035601051058620214, + 0.005378596484661102, + 0.007593672722578049, + 0.003535892814397812, + 0.00807228870689869, + 0.0035094586201012135, + 0.005249107256531715, + 0.0053619565442204475, + 0.0038307199720293283, + 0.005246104672551155, + 0.02268286421895027, + 0.011633472517132759, + 0.03034292161464691, + 0.03421734273433685, + 0.011523285880684853, + 0.011353449895977974, + 0.010583640076220036, + 0.00996293593198061, + 0.025364495813846588, + 0.0184068251401186, + 0.021348582580685616, + 0.012289505451917648, + 0.010839954949915409, + 0.008803527802228928, + 0.00933063868433237, + 0.010133776813745499, + 0.015081025660037994, + 0.010432947427034378, + 0.018382282927632332, + 0.015727488324046135, + 0.0180022232234478, + 0.013740960508584976, + 0.015163588337600231, + 0.015183581039309502, + 0.015415032394230366, + 0.015413704328238964, + 0.01947338506579399, + 0.019430721178650856, + 0.01310659758746624, + 0.02219257690012455, + 0.017552318051457405, + 0.01655002124607563, + 0.003997846972197294, + 0.0031900345347821712, + 0.005470736417919397, + 0.004419451113790274, + 0.0024912201333791018, + 0.002673343289643526, + 0.003741546766832471, + 0.005175490397959948, + 0.008942541666328907, + 0.011319536715745926, + 0.00824460107833147, + 0.00751916877925396, + 0.008423631079494953, + 0.012103238143026829, + 0.009191140532493591, + 0.007027363404631615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.07006505131721497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641016039997339 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007641016039997339, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641016039997339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14668215811252594, + 0.26365962624549866, + 0.0747087374329567, + 0.030695199966430664, + 0.10408434271812439, + 0.09630025178194046, + 0.10910311341285706, + 0.08356712013483047, + 0.04433608427643776, + 0.0468633770942688 + ], + "credit_map": [ + 0.019456367939710617, + 0.011236775666475296, + 0.021188456565141678, + 0.01577134057879448, + 0.017705222591757774, + 0.01851300708949566, + 0.02176406979560852, + 0.021046919748187065, + 0.031166652217507362, + 0.018480760976672173, + 0.04492096230387688, + 0.01736900582909584, + 0.04494190588593483, + 0.03252431005239487, + 0.034260984510183334, + 0.03999504819512367, + 0.011989675462245941, + 0.005227704532444477, + 0.013591242954134941, + 0.007390795275568962, + 0.010845487006008625, + 0.00926075130701065, + 0.008044752292335033, + 0.00835832767188549, + 0.0060343979857862, + 0.002526308177039027, + 0.00506557198241353, + 0.003522082231938839, + 0.0029457206837832928, + 0.003988892305642366, + 0.002529750345274806, + 0.004082477651536465, + 0.012989767827093601, + 0.010515390895307064, + 0.017272094264626503, + 0.019352274015545845, + 0.01237772312015295, + 0.011353515088558197, + 0.00989479385316372, + 0.010328785516321659, + 0.02087760902941227, + 0.009912933222949505, + 0.017498552799224854, + 0.01146281510591507, + 0.01093747466802597, + 0.009581412188708782, + 0.007481240667402744, + 0.008548213168978691, + 0.01177884265780449, + 0.00984592828899622, + 0.014853890053927898, + 0.01587369292974472, + 0.016325611621141434, + 0.01126424502581358, + 0.014896879903972149, + 0.014264029450714588, + 0.009358378127217293, + 0.009634610265493393, + 0.010935547761619091, + 0.013469691388309002, + 0.008416321128606796, + 0.011163990013301373, + 0.011405734345316887, + 0.009182844310998917, + 0.007535088341683149, + 0.004871939308941364, + 0.0066901142708957195, + 0.006454240530729294, + 0.0035537753719836473, + 0.0048575312830507755, + 0.005108054261654615, + 0.005265339743345976, + 0.006963290739804506, + 0.005278550088405609, + 0.00620378740131855, + 0.0044563584960997105, + 0.004794078879058361, + 0.0053904103115201, + 0.007242136634886265, + 0.006534762214869261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.007641016039997339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024917051196098328 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024917051196098328, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024917051196098328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07435077428817749, + 0.0641152486205101, + 0.0211793202906847, + 0.02030067704617977, + 0.03603630140423775, + 0.10601059347391129, + 0.2691447138786316, + 0.2776368260383606, + 0.07377403974533081, + 0.057451508939266205 + ], + "credit_map": [ + 0.013906043954193592, + 0.00499409344047308, + 0.013364393264055252, + 0.006168034393340349, + 0.008260655216872692, + 0.008534212596714497, + 0.008339052088558674, + 0.010784288868308067, + 0.008710620924830437, + 0.005054502747952938, + 0.011183549650013447, + 0.004192923195660114, + 0.009745701216161251, + 0.008527899160981178, + 0.007726841606199741, + 0.00897321105003357, + 0.0039032595232129097, + 0.001511530950665474, + 0.004317867569625378, + 0.001817192416638136, + 0.00243549095466733, + 0.002809203462675214, + 0.0017244494520127773, + 0.002660326426848769, + 0.004007273353636265, + 0.0013511816505342722, + 0.003681536065414548, + 0.0019098360789939761, + 0.0028176980558782816, + 0.0020592391956597567, + 0.002317572245374322, + 0.0021563423797488213, + 0.005430852994322777, + 0.002911898074671626, + 0.00704191904515028, + 0.009402063675224781, + 0.0027765294071286917, + 0.0024175578728318214, + 0.002924507251009345, + 0.0031309728510677814, + 0.020727086812257767, + 0.012643164955079556, + 0.01720600575208664, + 0.01687166839838028, + 0.009521923027932644, + 0.008739695884287357, + 0.01080016978085041, + 0.009500882588326931, + 0.032236482948064804, + 0.023777974769473076, + 0.037068672478199005, + 0.040683843195438385, + 0.035695698112249374, + 0.02935933694243431, + 0.036173876374959946, + 0.03414883092045784, + 0.03290640562772751, + 0.03344980627298355, + 0.04099339619278908, + 0.04726720601320267, + 0.026835566386580467, + 0.031551044434309006, + 0.03802262991666794, + 0.02661077305674553, + 0.012421802617609501, + 0.009958754293620586, + 0.01088045071810484, + 0.011588629335165024, + 0.005705792922526598, + 0.006759782321751118, + 0.007357173599302769, + 0.009101655334234238, + 0.0066079916432499886, + 0.005957288667559624, + 0.006974102929234505, + 0.007040339056402445, + 0.006438163109123707, + 0.007405873853713274, + 0.008358610793948174, + 0.008669138886034489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.0024917051196098328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865546151995659 + ], + "opposite_logprob_tokens": [ + -3.0486555099487305 + ], + "logprob_reward": -0.04865546151995659, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865546151995659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19260485470294952, + 0.08945447206497192, + 0.07462860643863678, + 0.05619046464562416, + 0.05942762270569801, + 0.08766023814678192, + 0.18941916525363922, + 0.13240081071853638, + 0.056688908487558365, + 0.06152491644024849 + ], + "credit_map": [ + 0.05528537556529045, + 0.013706113211810589, + 0.04563223570585251, + 0.018391713500022888, + 0.013132077641785145, + 0.014315925538539886, + 0.012659937143325806, + 0.019481442868709564, + 0.01966891810297966, + 0.004616159945726395, + 0.01764335110783577, + 0.006786854937672615, + 0.010371184907853603, + 0.009366743266582489, + 0.008994643576443195, + 0.012006601318717003, + 0.01242255698889494, + 0.006258785724639893, + 0.013997598551213741, + 0.005673415958881378, + 0.010607639327645302, + 0.007914639078080654, + 0.00811070203781128, + 0.009643256664276123, + 0.00850992277264595, + 0.005871681962162256, + 0.008702854625880718, + 0.006055207457393408, + 0.005735631100833416, + 0.008350399322807789, + 0.0048686605878174305, + 0.008096098899841309, + 0.006800814997404814, + 0.0054186489433050156, + 0.007381219882518053, + 0.0067514595575630665, + 0.008380568586289883, + 0.007018689531832933, + 0.008547198958694935, + 0.009129016660153866, + 0.01672990433871746, + 0.014201202429831028, + 0.01422322727739811, + 0.011745256371796131, + 0.007565527688711882, + 0.0066457693465054035, + 0.008425378240644932, + 0.008123967796564102, + 0.02421075478196144, + 0.019222985953092575, + 0.02897283062338829, + 0.02811177261173725, + 0.022874994203448296, + 0.019278528168797493, + 0.02326032891869545, + 0.023486942052841187, + 0.019059903919696808, + 0.013797003775835037, + 0.022981641814112663, + 0.024019604548811913, + 0.012141254730522633, + 0.011968574486672878, + 0.018958915024995804, + 0.009473911486566067, + 0.01086733303964138, + 0.012398315593600273, + 0.008326364681124687, + 0.006196958478540182, + 0.004763325210660696, + 0.004293461795896292, + 0.00478913402184844, + 0.00505400775000453, + 0.007119219750165939, + 0.01085501816123724, + 0.005469982046633959, + 0.005571257323026657, + 0.006497795693576336, + 0.00942693930119276, + 0.007942271418869495, + 0.008642425760626793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.04865546151995659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady close-up shot from a slightly overhead angle focuses squarely on a plump lemon, resting against a pure black, non-reflective background. The lemon's vibrant yellow skin is dotted with fine pits, giving it a natural, waxy sheen. A small, dry, green stem still remains at one end. Illuminated by the bright, even, cold studio light, every detail of the lemon is clearly visible. Suddenly, an invisible chill sweeps through, initiating the freezing process. Tiny, needle-like ice crystals form at a visible rate from the base of the lemon and spread upward. These crystalline frosts quickly connect and form, crawling across the entire rind like living vines, completely covering the original vibrant yellow beneath a thick, shimmering layer of white frost. The entire process is captured with unwavering clarity, visually depicting this remarkable natural physical transformation with a calm and objective approach to popular science, creating a visual experience rich in both focused and scientific beauty.", + "seed": 674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_14/index_2918_seed_674_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022704481007531287, + "vlm_logprob": -0.022704481007531287, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.022704481007531287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007302023936063051 + ], + "opposite_logprob_tokens": [ + -6.6323018074035645 + ], + "logprob_reward": -0.007302023936063051, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.007302023936063051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08901838213205338, + 0.060023896396160126, + 0.05966833978891373, + 0.07404603064060211, + 0.31731128692626953, + 0.08146370947360992, + 0.14092779159545898, + 0.09006956964731216, + 0.0275897067040205, + 0.059881240129470825 + ], + "credit_map": [ + 0.010401028208434582, + 0.008128081448376179, + 0.007590660825371742, + 0.012723919935524464, + 0.008333412930369377, + 0.017763013020157814, + 0.0091343242675066, + 0.014943943358957767, + 0.006780165713280439, + 0.0068188318982720375, + 0.003788535250350833, + 0.012900027446448803, + 0.0042421468533575535, + 0.009038534946739674, + 0.009059653617441654, + 0.007396007888019085, + 0.0045930249616503716, + 0.010401672683656216, + 0.003076010849326849, + 0.015829265117645264, + 0.004421378020197153, + 0.004173922818154097, + 0.010471872054040432, + 0.006701203528791666, + 0.00876386184245348, + 0.004916603676974773, + 0.005758016370236874, + 0.0102189090102911, + 0.006316985934972763, + 0.018266700208187103, + 0.010452506132423878, + 0.009352453984320164, + 0.019492803141474724, + 0.0224470142275095, + 0.00763706536963582, + 0.06271785497665405, + 0.018660100176930428, + 0.028217289596796036, + 0.1235073134303093, + 0.034631870687007904, + 0.0049433717504143715, + 0.008554489351809025, + 0.005831313785165548, + 0.01639251969754696, + 0.006992023903876543, + 0.007487139198929071, + 0.02413884364068508, + 0.00712401932105422, + 0.00735644344240427, + 0.009975257329642773, + 0.008024501614272594, + 0.019589999690651894, + 0.01249577384442091, + 0.034557659178972244, + 0.035699281841516495, + 0.013228880241513252, + 0.006540738511830568, + 0.006406331434845924, + 0.006124296225607395, + 0.008137340657413006, + 0.006714290007948875, + 0.024844884872436523, + 0.01777970790863037, + 0.013521980494260788, + 0.0030742911621928215, + 0.004394231829792261, + 0.0027783613186329603, + 0.002989219967275858, + 0.003505924018099904, + 0.003081155475229025, + 0.004885957110673189, + 0.002880569314584136, + 0.005595371592789888, + 0.009362376295030117, + 0.006579138338565826, + 0.008063724264502525, + 0.010334465652704239, + 0.005651571787893772, + 0.008103827014565468, + 0.006190769840031862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.007302023936063051 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983858808875084 + ], + "opposite_logprob_tokens": [ + -3.5298385620117188 + ], + "logprob_reward": -0.02983858808875084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983858808875084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023952685296535492, + 0.015110954642295837, + 0.02094387821853161, + 0.09671945869922638, + 0.6915162801742554, + 0.0718855932354927, + 0.027818573638796806, + 0.015783241018652916, + 0.00774031737819314, + 0.02852897346019745 + ], + "credit_map": [ + 0.0031505741644650698, + 0.0020286235958337784, + 0.00218210113234818, + 0.0031850235536694527, + 0.001967495772987604, + 0.00378032005392015, + 0.0026338386815041304, + 0.005024708807468414, + 0.0012349220924079418, + 0.0018817922100424767, + 0.0010295602260157466, + 0.003866196144372225, + 0.0012121351901441813, + 0.0017717110458761454, + 0.0024170244578272104, + 0.0016976135084405541, + 0.0023754320573061705, + 0.0024343577679246664, + 0.0013014012947678566, + 0.0043623787350952625, + 0.0020459985826164484, + 0.0030333080794662237, + 0.0037111137062311172, + 0.0016798877622932196, + 0.005373306106775999, + 0.010703859850764275, + 0.005016733892261982, + 0.02969067543745041, + 0.0067996671423316, + 0.008882339112460613, + 0.022821152582764626, + 0.007431723643094301, + 0.04394259303808212, + 0.05017378181219101, + 0.015430698171257973, + 0.13832999765872955, + 0.042329248040914536, + 0.05334635078907013, + 0.2695406377315521, + 0.07842294871807098, + 0.00543456245213747, + 0.00507111381739378, + 0.0021574741695076227, + 0.012555309571325779, + 0.004787044133991003, + 0.006952874828130007, + 0.02437075786292553, + 0.010556451976299286, + 0.002361315069720149, + 0.0031740120612084866, + 0.003971274942159653, + 0.004114499781280756, + 0.003920523449778557, + 0.004354314412921667, + 0.003334927139803767, + 0.0025877063162624836, + 0.0013442683266475797, + 0.0011404611868783832, + 0.0019656436052173376, + 0.0014050665777176619, + 0.0015225305687636137, + 0.004097816534340382, + 0.0017704739002510905, + 0.0025369806680828333, + 0.0007037871400825679, + 0.0012412925716489553, + 0.0007646008161827922, + 0.0007993219769559801, + 0.0010418618330731988, + 0.0013747303746640682, + 0.0007411384140141308, + 0.0010735848918557167, + 0.0030406364239752293, + 0.004312069155275822, + 0.003833930939435959, + 0.0033430142793804407, + 0.004067705944180489, + 0.004427750129252672, + 0.002540683140978217, + 0.0029631832148879766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.02983858808875084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499378964304924 + ], + "opposite_logprob_tokens": [ + -2.9299938678741455 + ], + "logprob_reward": -0.05499378964304924, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499378964304924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16688059270381927, + 0.15467292070388794, + 0.09686513245105743, + 0.12081038951873779, + 0.148445725440979, + 0.09923127293586731, + 0.1045856773853302, + 0.041659582406282425, + 0.027474507689476013, + 0.03937411680817604 + ], + "credit_map": [ + 0.023894187062978745, + 0.016781887039542198, + 0.011561102233827114, + 0.027563946321606636, + 0.012725203298032284, + 0.03815372660756111, + 0.013669552281498909, + 0.022531015798449516, + 0.008444732055068016, + 0.023260675370693207, + 0.008685851469635963, + 0.04166986793279648, + 0.010370503179728985, + 0.020885342732071877, + 0.026922818273305893, + 0.014433152042329311, + 0.009626089595258236, + 0.008805100806057453, + 0.0049620093777775764, + 0.015263759531080723, + 0.006455897353589535, + 0.014674575999379158, + 0.02342105470597744, + 0.013656657189130783, + 0.0063339713960886, + 0.01240005437284708, + 0.007974634878337383, + 0.031070688739418983, + 0.008448046632111073, + 0.012356007471680641, + 0.031806185841560364, + 0.010420825332403183, + 0.010777579620480537, + 0.011306404136121273, + 0.0049127331003546715, + 0.02725343033671379, + 0.008578908629715443, + 0.014346997253596783, + 0.05104557052254677, + 0.02022412419319153, + 0.0068841963075101376, + 0.01504990179091692, + 0.009756140410900116, + 0.017042838037014008, + 0.010104266926646233, + 0.008532905019819736, + 0.02081315591931343, + 0.01104788389056921, + 0.008967384696006775, + 0.009318627417087555, + 0.01061371061950922, + 0.012374787591397762, + 0.01318263541907072, + 0.020315714180469513, + 0.01034543476998806, + 0.019467394798994064, + 0.002993204165250063, + 0.0036694363225251436, + 0.004837057087570429, + 0.006189452949911356, + 0.005720983259379864, + 0.0051346090622246265, + 0.006450170651078224, + 0.00666467472910881, + 0.0027035598177462816, + 0.004052241798490286, + 0.001985230715945363, + 0.002458174480125308, + 0.0036490564234554768, + 0.004446127917617559, + 0.004462145734578371, + 0.0037179759237915277, + 0.00424452219158411, + 0.005722739268094301, + 0.004816104657948017, + 0.0043817562982439995, + 0.006509073078632355, + 0.005126165226101875, + 0.004923578351736069, + 0.003650183789432049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05499378964304924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196133067831397 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196133067831397, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196133067831397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05394702032208443, + 0.03956340625882149, + 0.01689263805747032, + 0.021934200078248978, + 0.30777275562286377, + 0.059956345707178116, + 0.3161943852901459, + 0.05684203281998634, + 0.030993595719337463, + 0.09590362012386322 + ], + "credit_map": [ + 0.00648445775732398, + 0.0038897902704775333, + 0.004041621461510658, + 0.00675906240940094, + 0.004997442476451397, + 0.014581630937755108, + 0.004672749433666468, + 0.008520270697772503, + 0.00375765492208302, + 0.0050001684576272964, + 0.0022860614117234945, + 0.01151346042752266, + 0.0032399301417171955, + 0.003209250746294856, + 0.006778389681130648, + 0.003778489539399743, + 0.0017158534610643983, + 0.0019367526983842254, + 0.0009653473971411586, + 0.0034317325334995985, + 0.001979196909815073, + 0.001536558847874403, + 0.0031445082277059555, + 0.0021826880984008312, + 0.0032449099235236645, + 0.00247773714363575, + 0.0012644213857129216, + 0.004532246384769678, + 0.0015770737081766129, + 0.003690381534397602, + 0.002506602555513382, + 0.0026408270932734013, + 0.01939723640680313, + 0.02170264907181263, + 0.007064356934279203, + 0.061089709401130676, + 0.017928844317793846, + 0.021998142823576927, + 0.12154816091060638, + 0.03704364225268364, + 0.0037927927915006876, + 0.0042235367000103, + 0.0015821268316358328, + 0.013016683980822563, + 0.0038521280512213707, + 0.004593027289956808, + 0.021848106756806374, + 0.007047944236546755, + 0.015518797561526299, + 0.021159907802939415, + 0.014540856704115868, + 0.04779312387108803, + 0.030175069347023964, + 0.07009529322385788, + 0.0881195142865181, + 0.02879182994365692, + 0.004925512243062258, + 0.005476284772157669, + 0.004456133581697941, + 0.006859895307570696, + 0.006181051954627037, + 0.010566444136202335, + 0.007308709900826216, + 0.011067998595535755, + 0.0033701774664223194, + 0.003976052161306143, + 0.0036788785364478827, + 0.00282464106567204, + 0.004481626208871603, + 0.00479610450565815, + 0.003427563700824976, + 0.004438553936779499, + 0.009049824438989162, + 0.01651073805987835, + 0.008719328790903091, + 0.011294065974652767, + 0.0161221195012331, + 0.011230087839066982, + 0.014066441915929317, + 0.008911016397178173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.003196133067831397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady close-up of a plump lemon from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure black and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lemon have vibrant yellow skin with fine pits and a small dry green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do needle-like ice crystals form at the base of the lemon and spread upward to cover the rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even, cold studio light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053813423961400986, + 0.03275105357170105, + 0.035459745675325394, + 0.08963637799024582, + 0.6397838592529297, + 0.012418307363986969, + 0.07106705754995346, + 0.027841269969940186, + 0.010761003009974957, + 0.02646787278354168 + ], + "credit_map": [ + 0.006864900700747967, + 0.0037116874009370804, + 0.004308044444769621, + 0.005056052468717098, + 0.005163206253200769, + 0.01577448472380638, + 0.003992713987827301, + 0.008942343294620514, + 0.0028824019245803356, + 0.0037703693378716707, + 0.001807848340831697, + 0.010660055093467236, + 0.0024913381785154343, + 0.0025791062507778406, + 0.005422140471637249, + 0.003137800144031644, + 0.0028328171465545893, + 0.0037503179628401995, + 0.0019456155132502317, + 0.007942772470414639, + 0.003557579591870308, + 0.004492218140512705, + 0.008014064282178879, + 0.002924362663179636, + 0.005904090125113726, + 0.006015212740749121, + 0.003154391422867775, + 0.018483972176909447, + 0.006183416582643986, + 0.009403178468346596, + 0.031586453318595886, + 0.008905679918825626, + 0.04032096266746521, + 0.04677669331431389, + 0.016551554203033447, + 0.12397230416536331, + 0.038892924785614014, + 0.04735527187585831, + 0.2440672367811203, + 0.08184698224067688, + 0.0010212004417553544, + 0.0022523512598127127, + 0.0007117727072909474, + 0.0018319478258490562, + 0.0011982141295447946, + 0.0016818171134218574, + 0.001999054802581668, + 0.001721950713545084, + 0.004017752595245838, + 0.008157703094184399, + 0.004612456075847149, + 0.011067602783441544, + 0.006873698905110359, + 0.011818152852356434, + 0.020032662898302078, + 0.004487028811126947, + 0.0018079194705933332, + 0.002345685614272952, + 0.002080606995150447, + 0.003405900439247489, + 0.0040944418869912624, + 0.006208228413015604, + 0.0054473476484417915, + 0.0024511420633643866, + 0.0008889604359865189, + 0.0013595864875242114, + 0.0009907857747748494, + 0.0009004397434182465, + 0.0016850773245096207, + 0.0019161507952958345, + 0.001328463084064424, + 0.0016915409360080957, + 0.0027395254001021385, + 0.00340127176605165, + 0.003815151983872056, + 0.0028051589615643024, + 0.004113261122256517, + 0.004101107828319073, + 0.002699238946661353, + 0.002793159568682313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.0181918703019619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_tensor([2372, 2372])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_tensor([2372, 2372])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb6376b0cbad93b595eaff896bd64dad20381f62 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_tensor([2372, 2372])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, the camera panning left in a slow, graceful arc, captures the depths of a quiet, ancient courtyard. The floor is paved with mottled bluestone, covered in a thick, iridescent blanket of fallen leaves, primarily golden ginkgo and crimson maple. A woman dressed in flowing Hanfu (Chinese traditional clothing) takes center stage. She wears an elegant jade-green double-breasted ruqun (skirt), with wide silk sleeves delicately embroidered with flowing cloud motifs in silver thread, gleaming softly in the gentle afternoon sun. Her long, jet-black hair is tied into a classically refined bun, secured with a simple jade hairpin. Initially facing away from the camera, she then, with remarkable slowness, begins to turn. Her wide skirt and sleeves trace graceful arcs through the air, causing the fallen leaves at her feet to swirl gently. Her movements are smooth and uninterrupted. Finally, when halfway around, she tilts her head slightly, casting a long, affectionate glance back. A hint of melancholy graces her delicate features, and her eyes, as if telling a thousand-year-old story, evoke a sense of tranquility, nostalgia, and poetry. The entire scene is bathed in warm golden light, and the air is filled with a tranquil, nostalgic, and poetic atmosphere.", + "seed": 22698, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_2372_seed_22698_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03570554181933403, + "vlm_logprob": -0.03570554181933403, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03570554181933403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014260408468544483 + ], + "opposite_logprob_tokens": [ + -4.389260292053223 + ], + "logprob_reward": -0.014260408468544483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014260408468544483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25584810972213745, + 0.05043408274650574, + 0.08632103353738785, + 0.04099242389202118, + 0.09908394515514374, + 0.05256952717900276, + 0.08642474561929703, + 0.12136860936880112, + 0.16704177856445312, + 0.03991571068763733 + ], + "credit_map": [ + 0.025038547813892365, + 0.024615487083792686, + 0.01762091927230358, + 0.027290409430861473, + 0.060564812272787094, + 0.030011022463440895, + 0.046582769602537155, + 0.02412411756813526, + 0.005507089197635651, + 0.006283306051045656, + 0.004640177823603153, + 0.006180098280310631, + 0.006087954621762037, + 0.00727442279458046, + 0.010143022052943707, + 0.004318013321608305, + 0.0113453334197402, + 0.007394281681627035, + 0.01020277664065361, + 0.007503374479711056, + 0.008058563806116581, + 0.017198845744132996, + 0.01179138757288456, + 0.012826467864215374, + 0.006909291259944439, + 0.006176031660288572, + 0.0038158949464559555, + 0.0035629530902951956, + 0.004281671717762947, + 0.005888585466891527, + 0.005842672195285559, + 0.004515326116234064, + 0.006725465413182974, + 0.006624990608543158, + 0.0077599044889211655, + 0.006999879609793425, + 0.0073328772559762, + 0.029043979942798615, + 0.026844734326004982, + 0.007752119563519955, + 0.0035324611235409975, + 0.0029465288389474154, + 0.004409314598888159, + 0.0030796106439083815, + 0.003990132827311754, + 0.018443934619426727, + 0.003195755183696747, + 0.012971785850822926, + 0.009257212281227112, + 0.008853340521454811, + 0.010014229454100132, + 0.012213902547955513, + 0.007794426754117012, + 0.016476046293973923, + 0.010255873203277588, + 0.011559722013771534, + 0.01509274635463953, + 0.02297351323068142, + 0.011004497297108173, + 0.014081701636314392, + 0.021626755595207214, + 0.009502414613962173, + 0.008910085074603558, + 0.01817690022289753, + 0.013568834401667118, + 0.01603761501610279, + 0.01921098679304123, + 0.011114523746073246, + 0.010144500061869621, + 0.03726004436612129, + 0.012851973995566368, + 0.046853307634592056, + 0.0041177175007760525, + 0.00637849373742938, + 0.004791846964508295, + 0.007290975656360388, + 0.004534793086349964, + 0.004892969503998756, + 0.0034423444885760546, + 0.004466566257178783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014260408468544483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279580116271973 + ], + "opposite_logprob_tokens": [ + -2.0177958011627197 + ], + "logprob_reward": -0.14279580116271973, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279580116271973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10681366920471191, + 0.04848027601838112, + 0.0921306163072586, + 0.02281644195318222, + 0.1285381317138672, + 0.13129407167434692, + 0.049164868891239166, + 0.06452315300703049, + 0.07086478918790817, + 0.2853739261627197 + ], + "credit_map": [ + 0.012339540757238865, + 0.012687583453953266, + 0.009934928268194199, + 0.012184184975922108, + 0.013134869746863842, + 0.012490535154938698, + 0.019524946808815002, + 0.014517086558043957, + 0.004256685264408588, + 0.005192748736590147, + 0.0034773279912769794, + 0.004519389010965824, + 0.010967351496219635, + 0.007044999394565821, + 0.008454825729131699, + 0.004566953517496586, + 0.011037123389542103, + 0.009656660258769989, + 0.01085706427693367, + 0.007820925675332546, + 0.009983756579458714, + 0.018660295754671097, + 0.011527376249432564, + 0.012587424367666245, + 0.002614251570776105, + 0.003468351671472192, + 0.0022275715600699186, + 0.0028736258391290903, + 0.002479164395481348, + 0.0030872358474880457, + 0.0033408452291041613, + 0.0027253972366452217, + 0.008884943090379238, + 0.006474072579294443, + 0.008758234791457653, + 0.007779260631650686, + 0.008332707919180393, + 0.03946781903505325, + 0.03869355842471123, + 0.010147564113140106, + 0.009318748489022255, + 0.00881236232817173, + 0.010427056811749935, + 0.00792815163731575, + 0.011068073101341724, + 0.044409528374671936, + 0.009405482560396194, + 0.029924679547548294, + 0.00535865593701601, + 0.0033862616401165724, + 0.008415614254772663, + 0.004328758455812931, + 0.0025707040913403034, + 0.013769304379820824, + 0.0029732342809438705, + 0.008362342603504658, + 0.010927928611636162, + 0.007079781033098698, + 0.009261074475944042, + 0.006830706261098385, + 0.005885602440685034, + 0.007044265978038311, + 0.005174472462385893, + 0.012319328263401985, + 0.013422607444226742, + 0.011633007787168026, + 0.006884671747684479, + 0.006429300177842379, + 0.007573653012514114, + 0.007207922637462616, + 0.007107352372258902, + 0.010606281459331512, + 0.02966536208987236, + 0.028103100135922432, + 0.04470174014568329, + 0.02943306230008602, + 0.024818597361445427, + 0.028615659102797508, + 0.020145060494542122, + 0.07989136874675751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.14279580116271973 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739859934896231 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006739859934896231, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739859934896231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.126710906624794, + 0.059215743094682693, + 0.06180514767765999, + 0.014096914790570736, + 0.5165948867797852, + 0.0507323294878006, + 0.019073791801929474, + 0.05390692129731178, + 0.03042813390493393, + 0.06743526458740234 + ], + "credit_map": [ + 0.01970371976494789, + 0.016400981694459915, + 0.01276822853833437, + 0.012062455527484417, + 0.016704875975847244, + 0.014726578257977962, + 0.016630833968520164, + 0.017713233828544617, + 0.008389797061681747, + 0.0070680794306099415, + 0.008423373103141785, + 0.006432309281080961, + 0.008734184317290783, + 0.005847711581736803, + 0.006884329952299595, + 0.007435958832502365, + 0.008670933544635773, + 0.00580106396228075, + 0.008205152116715908, + 0.006902162916958332, + 0.006402625702321529, + 0.009790328331291676, + 0.007942348718643188, + 0.008090534247457981, + 0.002008153358474374, + 0.0015345163410529494, + 0.001643526367843151, + 0.0015214537270367146, + 0.0017706257058307528, + 0.0018239226192235947, + 0.0021518331486731768, + 0.001642883289605379, + 0.03362821042537689, + 0.021496979519724846, + 0.03425110876560211, + 0.02180502377450466, + 0.030107397586107254, + 0.17353400588035583, + 0.1644807606935501, + 0.03729139640927315, + 0.003990368917584419, + 0.004743317607790232, + 0.004254598636180162, + 0.004898719023913145, + 0.003103561233729124, + 0.011551326140761375, + 0.005275677889585495, + 0.012914757244288921, + 0.0022168736904859543, + 0.0022198809310793877, + 0.0025406889617443085, + 0.002443697303533554, + 0.0016731804935261607, + 0.003073917469009757, + 0.0019065901869907975, + 0.002998961601406336, + 0.0061848838813602924, + 0.006729032378643751, + 0.007891657762229443, + 0.004847546573728323, + 0.0060927774757146835, + 0.007881767116487026, + 0.004020536784082651, + 0.010258720256388187, + 0.004615269601345062, + 0.0055582765489816666, + 0.0032365897204726934, + 0.0026992373168468475, + 0.00356091046705842, + 0.004308982752263546, + 0.0035009190905839205, + 0.0029479486402124166, + 0.0048910886980593204, + 0.004117265343666077, + 0.007268082816153765, + 0.00532153807580471, + 0.007680033799260855, + 0.018779508769512177, + 0.005528551992028952, + 0.013849197886884212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.006739859934896231 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253500021994114 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253500021994114, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253500021994114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15334632992744446, + 0.0398225374519825, + 0.049104440957307816, + 0.04172026365995407, + 0.08324746787548065, + 0.08197586983442307, + 0.1494179219007492, + 0.2250290811061859, + 0.07844892889261246, + 0.09788714349269867 + ], + "credit_map": [ + 0.01115351915359497, + 0.0123972836881876, + 0.011481787078082561, + 0.016016898676753044, + 0.033997390419244766, + 0.01767670549452305, + 0.02838730625808239, + 0.022235441952943802, + 0.004127834923565388, + 0.0038479336071759462, + 0.005129947327077389, + 0.003534061834216118, + 0.0064266338013112545, + 0.005596648436039686, + 0.005404787603765726, + 0.005754689686000347, + 0.004935757257044315, + 0.005214889533817768, + 0.005818701349198818, + 0.00624571880325675, + 0.00616066437214613, + 0.006217241752892733, + 0.007890450768172741, + 0.0066210185177624226, + 0.0036525309551507235, + 0.006612817756831646, + 0.0031625262927263975, + 0.0048120105639100075, + 0.006122587714344263, + 0.006140148267149925, + 0.00840715877711773, + 0.002810484729707241, + 0.008014403283596039, + 0.012052543461322784, + 0.007422333583235741, + 0.01288628950715065, + 0.010000805370509624, + 0.011964894831180573, + 0.014526871964335442, + 0.006379322148859501, + 0.005024518817663193, + 0.00865188892930746, + 0.0061575425788760185, + 0.007797181140631437, + 0.006152010522782803, + 0.02164195477962494, + 0.00881685595959425, + 0.017733916640281677, + 0.014355607330799103, + 0.017480477690696716, + 0.020848050713539124, + 0.024249320849776268, + 0.014548935927450657, + 0.023696068674325943, + 0.018867578357458115, + 0.015371887944638729, + 0.02385094203054905, + 0.05224446579813957, + 0.021452223882079124, + 0.019780736416578293, + 0.039534974843263626, + 0.018060075119137764, + 0.013647000305354595, + 0.036458659917116165, + 0.008556757122278214, + 0.016270890831947327, + 0.0071916659362614155, + 0.010199886746704578, + 0.010168636217713356, + 0.0067439330741763115, + 0.009364448487758636, + 0.009952709078788757, + 0.008534091524779797, + 0.009928041137754917, + 0.010536536574363708, + 0.010511152446269989, + 0.0110222939401865, + 0.02310669608414173, + 0.007186473347246647, + 0.017061861231923103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01253500021994114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196639310568571 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.002196639310568571, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196639310568571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1572456657886505, + 0.06260475516319275, + 0.10415486991405487, + 0.18481068313121796, + 0.06513644754886627, + 0.08760412782430649, + 0.06160963326692581, + 0.13000817596912384, + 0.059930384159088135, + 0.08689525723457336 + ], + "credit_map": [ + 0.015211866237223148, + 0.015750547870993614, + 0.01717744581401348, + 0.016412438824772835, + 0.02703131176531315, + 0.02000921219587326, + 0.027496106922626495, + 0.018156733363866806, + 0.005982506554573774, + 0.006470950786024332, + 0.009402519091963768, + 0.006486318074166775, + 0.008203762583434582, + 0.008052241057157516, + 0.010366644710302353, + 0.007639813236892223, + 0.015246731229126453, + 0.009570956230163574, + 0.014181322418153286, + 0.00986568070948124, + 0.012264067307114601, + 0.016390031203627586, + 0.015219752676784992, + 0.011416330933570862, + 0.032293956726789474, + 0.02293049916625023, + 0.020174352452158928, + 0.016865618526935577, + 0.018288154155015945, + 0.030932003632187843, + 0.0271159578114748, + 0.016210149973630905, + 0.006049038842320442, + 0.005931415129452944, + 0.00967676006257534, + 0.008216578513383865, + 0.0073047722689807415, + 0.010483605787158012, + 0.007281804922968149, + 0.010192472487688065, + 0.004500043112784624, + 0.004138329531997442, + 0.00565520441159606, + 0.003846625564619899, + 0.0031678921077400446, + 0.030988318845629692, + 0.0058682700619101524, + 0.029439453035593033, + 0.0061079030856490135, + 0.00603020004928112, + 0.007483689580112696, + 0.007315369322896004, + 0.005283059552311897, + 0.01329928170889616, + 0.006437827832996845, + 0.009652301669120789, + 0.014051134698092937, + 0.021415987983345985, + 0.014209765009582043, + 0.012727582827210426, + 0.023502586409449577, + 0.009722301736474037, + 0.008109189569950104, + 0.02626962773501873, + 0.006922389846295118, + 0.01338626816868782, + 0.0056917318142950535, + 0.0071870614774525166, + 0.006769412197172642, + 0.006595110986381769, + 0.0071921502240002155, + 0.006186256650835276, + 0.007634440436959267, + 0.008504591882228851, + 0.010957646183669567, + 0.006886260583996773, + 0.010854946449398994, + 0.021666210144758224, + 0.005996385123580694, + 0.014394779689610004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.002196639310568571 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, the camera panning left in a slow, graceful arc, captures the depths of a quiet, ancient courtyard. The floor is paved with mottled bluestone, covered in a thick, iridescent blanket of fallen leaves, primarily golden ginkgo and crimson maple. A woman dressed in flowing Hanfu (Chinese traditional clothing) takes center stage. She wears an elegant jade-green double-breasted ruqun (skirt), with wide silk sleeves delicately embroidered with flowing cloud motifs in silver thread, gleaming softly in the gentle afternoon sun. Her long, jet-black hair is tied into a classically refined bun, secured with a simple jade hairpin. Initially facing away from the camera, she then, with remarkable slowness, begins to turn. Her wide skirt and sleeves trace graceful arcs through the air, causing the fallen leaves at her feet to swirl gently. Her movements are smooth and uninterrupted. Finally, when halfway around, she tilts her head slightly, casting a long, affectionate glance back. A hint of melancholy graces her delicate features, and her eyes, as if telling a thousand-year-old story, evoke a sense of tranquility, nostalgia, and poetry. The entire scene is bathed in warm golden light, and the air is filled with a tranquil, nostalgic, and poetic atmosphere.", + "seed": 22699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_2372_seed_22699_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023110732110217212, + "vlm_logprob": -0.023110732110217212, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.023110732110217212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005809209309518337 + ], + "opposite_logprob_tokens": [ + -5.505809307098389 + ], + "logprob_reward": -0.005809209309518337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005809209309518337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4174044132232666, + 0.05261053517460823, + 0.04704011604189873, + 0.017117317765951157, + 0.029793579131364822, + 0.03423916921019554, + 0.10708524286746979, + 0.031605158001184464, + 0.19581320881843567, + 0.06729131191968918 + ], + "credit_map": [ + 0.03915299102663994, + 0.07309555262327194, + 0.059236492961645126, + 0.08094268292188644, + 0.030052250251173973, + 0.05597645789384842, + 0.050687696784734726, + 0.028260281309485435, + 0.006204287987202406, + 0.006196674425154924, + 0.012941262684762478, + 0.004304776433855295, + 0.004996247589588165, + 0.006762592121958733, + 0.005873551592230797, + 0.00533114280551672, + 0.005032493732869625, + 0.00466984324157238, + 0.005183064378798008, + 0.006931658368557692, + 0.005894944537431002, + 0.0064175911247730255, + 0.0069460030645132065, + 0.00596451573073864, + 0.002018734347075224, + 0.0018081320449709892, + 0.0016827223589643836, + 0.004151360597461462, + 0.0015455188695341349, + 0.0016580007504671812, + 0.0021406388841569424, + 0.0021122100297361612, + 0.004143824800848961, + 0.002843170426785946, + 0.0050826058723032475, + 0.004079267382621765, + 0.004012845456600189, + 0.002719210460782051, + 0.003673235420137644, + 0.003239418612793088, + 0.005359930917620659, + 0.002721521770581603, + 0.010377174243330956, + 0.007526041939854622, + 0.0017519746907055378, + 0.003147540846839547, + 0.0014245858183130622, + 0.0019303986337035894, + 0.007711536716669798, + 0.004547284450381994, + 0.044395580887794495, + 0.005102802533656359, + 0.005129907745867968, + 0.01451420970261097, + 0.003312864573672414, + 0.02237105183303356, + 0.0030600365716964006, + 0.004442483652383089, + 0.005570000037550926, + 0.0044303955510258675, + 0.003607743652537465, + 0.0034667160362005234, + 0.002256824867799878, + 0.004770958796143532, + 0.013012848794460297, + 0.009569955989718437, + 0.07121233642101288, + 0.011475063860416412, + 0.007062317803502083, + 0.012319826520979404, + 0.00686302874237299, + 0.06429781764745712, + 0.0068487441167235374, + 0.007047548890113831, + 0.007315278053283691, + 0.007346299476921558, + 0.01638616994023323, + 0.006778260227292776, + 0.00735999271273613, + 0.008209018036723137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.005809209309518337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897159457206726 + ], + "opposite_logprob_tokens": [ + -2.5789716243743896 + ], + "logprob_reward": -0.07897159457206726, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897159457206726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4267996549606323, + 0.048219066113233566, + 0.0326547771692276, + 0.04776753485202789, + 0.041045065969228745, + 0.06809931993484497, + 0.10867627710103989, + 0.055858638137578964, + 0.1306942254304886, + 0.04018552228808403 + ], + "credit_map": [ + 0.0401146374642849, + 0.07524768263101578, + 0.06302216649055481, + 0.08192197233438492, + 0.03132586181163788, + 0.058489348739385605, + 0.04809792339801788, + 0.02858000621199608, + 0.006144221872091293, + 0.007295750081539154, + 0.008396552875638008, + 0.005779774859547615, + 0.005991306621581316, + 0.006338489707559347, + 0.00511359516531229, + 0.003159369807690382, + 0.002751074032858014, + 0.00296641793102026, + 0.004261134658008814, + 0.0046308827586472034, + 0.003745556343346834, + 0.005517878569662571, + 0.0035050949081778526, + 0.0052767349407076836, + 0.004472671542316675, + 0.003453699639067054, + 0.005519714206457138, + 0.014806198887526989, + 0.004279621411114931, + 0.0035105140414088964, + 0.007172736339271069, + 0.004552374593913555, + 0.0076734949834644794, + 0.004244470968842506, + 0.0061230529099702835, + 0.004400279372930527, + 0.005186515394598246, + 0.00480447243899107, + 0.004258551634848118, + 0.004354225005954504, + 0.008153784088790417, + 0.004806696902960539, + 0.019527234137058258, + 0.016592541709542274, + 0.004652732517570257, + 0.003948942758142948, + 0.006034122314304113, + 0.004383261315524578, + 0.008270163089036942, + 0.01045510359108448, + 0.026500068604946136, + 0.00989444274455309, + 0.012755648232996464, + 0.016949927434325218, + 0.009252061136066914, + 0.0145988455042243, + 0.008496611379086971, + 0.011516633443534374, + 0.010405748151242733, + 0.005315972492098808, + 0.004892114084213972, + 0.005788322538137436, + 0.0030927842017263174, + 0.006350445095449686, + 0.008444811217486858, + 0.006269461940973997, + 0.04829021915793419, + 0.007420072332024574, + 0.005150770302861929, + 0.00804111547768116, + 0.0046793874353170395, + 0.0423983670771122, + 0.003910484723746777, + 0.0035002408549189568, + 0.003954828251153231, + 0.005778224673122168, + 0.0075646801851689816, + 0.0036973620299249887, + 0.006435705814510584, + 0.00534399226307869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.07897159457206726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635037526488304 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635037526488304, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635037526488304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38805755972862244, + 0.1868927925825119, + 0.050721555948257446, + 0.1591930240392685, + 0.020363183692097664, + 0.018861668184399605, + 0.046892642974853516, + 0.04769224673509598, + 0.06086469814181328, + 0.020460620522499084 + ], + "credit_map": [ + 0.037222739309072495, + 0.07059112936258316, + 0.059109434485435486, + 0.07342422008514404, + 0.027784438803792, + 0.05279373750090599, + 0.04039769247174263, + 0.02673419564962387, + 0.015696993097662926, + 0.027001822367310524, + 0.05105776712298393, + 0.020001858472824097, + 0.018479902297258377, + 0.022177571430802345, + 0.014909719116985798, + 0.017567157745361328, + 0.005642378702759743, + 0.004770951811224222, + 0.007152203470468521, + 0.006186863407492638, + 0.005236049648374319, + 0.005750065203756094, + 0.008036545477807522, + 0.00794649962335825, + 0.014533508569002151, + 0.010770323686301708, + 0.02079523727297783, + 0.056836098432540894, + 0.009166100062429905, + 0.010884765535593033, + 0.022002775222063065, + 0.014204206876456738, + 0.002696962095797062, + 0.002396197058260441, + 0.0034311756025999784, + 0.002626624656841159, + 0.0018443086883053184, + 0.0017690472304821014, + 0.00341656431555748, + 0.002182305557653308, + 0.004389127250760794, + 0.0021973527036607265, + 0.0035355789586901665, + 0.002665163716301322, + 0.0016444657230749726, + 0.0017096331575885415, + 0.0011089597828686237, + 0.0016113885212689638, + 0.0031944834627211094, + 0.003706267336383462, + 0.014411472715437412, + 0.0035659174900501966, + 0.003505734261125326, + 0.011896596290171146, + 0.002265802351757884, + 0.0043463679030537605, + 0.0028647487051784992, + 0.0025943375658243895, + 0.014906073920428753, + 0.003094289219006896, + 0.004348380956798792, + 0.0031227516010403633, + 0.0022215298376977444, + 0.014540135860443115, + 0.0037850707303732634, + 0.003495987970381975, + 0.0222004447132349, + 0.003283214522525668, + 0.00264962506480515, + 0.00397174758836627, + 0.0019432998960837722, + 0.019535312429070473, + 0.002484261756762862, + 0.0026306083891540766, + 0.0019891364499926567, + 0.002301522297784686, + 0.004992776084691286, + 0.0018363569397479296, + 0.001988685689866543, + 0.0022372736129909754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02635037526488304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036117101553827524 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.0036117101553827524, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036117101553827524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.204520121216774, + 0.038247354328632355, + 0.037031389772892, + 0.03712988272309303, + 0.06255301088094711, + 0.03341710567474365, + 0.16181430220603943, + 0.10263752937316895, + 0.18411438167095184, + 0.13853493332862854 + ], + "credit_map": [ + 0.01976887509226799, + 0.029648054391145706, + 0.028649333864450455, + 0.03238741680979729, + 0.017093708738684654, + 0.029287926852703094, + 0.0314849354326725, + 0.0161998700350523, + 0.004528636112809181, + 0.004858312662690878, + 0.006243762094527483, + 0.0046319677494466305, + 0.004271415527909994, + 0.004505917429924011, + 0.005158453714102507, + 0.0040488881058990955, + 0.003814938711002469, + 0.0034760027192533016, + 0.004525091499090195, + 0.00497221015393734, + 0.003275908762589097, + 0.004507487639784813, + 0.0062627349980175495, + 0.006197015289217234, + 0.0027068734634667635, + 0.0030764457769691944, + 0.004504064098000526, + 0.01256195455789566, + 0.0024962283205240965, + 0.002810732927173376, + 0.004053378943353891, + 0.004920205567032099, + 0.012384201399981976, + 0.006596332881599665, + 0.008914603851735592, + 0.0065291766077280045, + 0.007124754134565592, + 0.005906726233661175, + 0.005744311027228832, + 0.009352898225188255, + 0.003090714570134878, + 0.003460102481767535, + 0.005006926134228706, + 0.0061000632122159, + 0.0059806788340210915, + 0.0028427033685147762, + 0.004450347740203142, + 0.002485565608367324, + 0.010186929255723953, + 0.005911626853048801, + 0.06780342757701874, + 0.00862110499292612, + 0.008848580531775951, + 0.03663055598735809, + 0.003848646068945527, + 0.019963430240750313, + 0.006680530495941639, + 0.005401562433689833, + 0.034377116709947586, + 0.005987401120364666, + 0.006994251161813736, + 0.006452028173953295, + 0.00357818603515625, + 0.03316645324230194, + 0.01045157015323639, + 0.008242127485573292, + 0.0654817670583725, + 0.010817148722708225, + 0.011504251509904861, + 0.011729849502444267, + 0.0073224822990596294, + 0.058565184473991394, + 0.013738320209085941, + 0.01690385676920414, + 0.015598803758621216, + 0.01508841197937727, + 0.03128938376903534, + 0.011837664060294628, + 0.018057938665151596, + 0.016020549461245537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0036117101553827524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008107712492346764 + ], + "opposite_logprob_tokens": [ + -7.125810623168945 + ], + "logprob_reward": -0.0008107712492346764, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008107712492346764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3576320707798004, + 0.04848312959074974, + 0.02956729754805565, + 0.02472536452114582, + 0.025617899373173714, + 0.021023882552981377, + 0.15705423057079315, + 0.10488037019968033, + 0.19691690802574158, + 0.03409881889820099 + ], + "credit_map": [ + 0.03472490236163139, + 0.06488324701786041, + 0.05222493037581444, + 0.06847501546144485, + 0.02825000509619713, + 0.04754215106368065, + 0.03671152889728546, + 0.024820294231176376, + 0.00501388031989336, + 0.006703803315758705, + 0.01029182318598032, + 0.004844870883971453, + 0.004694774281233549, + 0.006245041731745005, + 0.004419845994561911, + 0.006269093137234449, + 0.004340780898928642, + 0.0030475284438580275, + 0.00394822983071208, + 0.002967642853036523, + 0.0038061598315835, + 0.002616716781631112, + 0.004260816145688295, + 0.004579424858093262, + 0.003940856084227562, + 0.0018393122591078281, + 0.0026674154214560986, + 0.006135250441730022, + 0.0020091442856937647, + 0.002678412478417158, + 0.0029059688095003366, + 0.0025490059051662683, + 0.005143382586538792, + 0.002548773540183902, + 0.003511653980240226, + 0.0028697748202830553, + 0.0024804205168038607, + 0.0025783334858715534, + 0.0021365443244576454, + 0.0043490175157785416, + 0.005016139708459377, + 0.0030673809815198183, + 0.0031892939005047083, + 0.0022781556472182274, + 0.0020743899513036013, + 0.002362341620028019, + 0.0014578342670574784, + 0.0015783491544425488, + 0.010248957201838493, + 0.0061804503202438354, + 0.0666242465376854, + 0.008527245372533798, + 0.007136822212487459, + 0.040818654000759125, + 0.002751824678853154, + 0.014766041189432144, + 0.006867712829262018, + 0.004240887705236673, + 0.03924424201250076, + 0.005124404560774565, + 0.003804080421105027, + 0.006328857969492674, + 0.002974956063553691, + 0.03629523143172264, + 0.016382113099098206, + 0.014240480028092861, + 0.06533205509185791, + 0.012313964776694775, + 0.008527922444045544, + 0.015630444511771202, + 0.006356978323310614, + 0.05813296511769295, + 0.005487890914082527, + 0.003814115421846509, + 0.004368605092167854, + 0.004421376157552004, + 0.004058588296175003, + 0.003859856864437461, + 0.004267975687980652, + 0.0038204132579267025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0008107712492346764 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, the camera panning left in a slow, graceful arc, captures the depths of a quiet, ancient courtyard. The floor is paved with mottled bluestone, covered in a thick, iridescent blanket of fallen leaves, primarily golden ginkgo and crimson maple. A woman dressed in flowing Hanfu (Chinese traditional clothing) takes center stage. She wears an elegant jade-green double-breasted ruqun (skirt), with wide silk sleeves delicately embroidered with flowing cloud motifs in silver thread, gleaming softly in the gentle afternoon sun. Her long, jet-black hair is tied into a classically refined bun, secured with a simple jade hairpin. Initially facing away from the camera, she then, with remarkable slowness, begins to turn. Her wide skirt and sleeves trace graceful arcs through the air, causing the fallen leaves at her feet to swirl gently. Her movements are smooth and uninterrupted. Finally, when halfway around, she tilts her head slightly, casting a long, affectionate glance back. A hint of melancholy graces her delicate features, and her eyes, as if telling a thousand-year-old story, evoke a sense of tranquility, nostalgia, and poetry. The entire scene is bathed in warm golden light, and the air is filled with a tranquil, nostalgic, and poetic atmosphere.", + "seed": 22700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_2372_seed_22700_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01723517505452037, + "vlm_logprob": -0.01723517505452037, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.01723517505452037, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005431537050753832 + ], + "opposite_logprob_tokens": [ + -5.505431652069092 + ], + "logprob_reward": -0.005431537050753832, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005431537050753832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09632815420627594, + 0.06782243400812149, + 0.05678267404437065, + 0.06719274073839188, + 0.07651383429765701, + 0.0470745675265789, + 0.08418209105730057, + 0.2295738011598587, + 0.22390758991241455, + 0.050622135400772095 + ], + "credit_map": [ + 0.00770857697352767, + 0.008880304172635078, + 0.01377507857978344, + 0.015090188942849636, + 0.017114033922553062, + 0.01078988891094923, + 0.012573354877531528, + 0.010396720841526985, + 0.006818532012403011, + 0.006090604700148106, + 0.00894106738269329, + 0.008216644637286663, + 0.009109500795602798, + 0.007288065738976002, + 0.01227759663015604, + 0.00908042211085558, + 0.005338899325579405, + 0.005951927974820137, + 0.006729677319526672, + 0.007817772217094898, + 0.0069949389435350895, + 0.009721989743411541, + 0.007203507237136364, + 0.007023959420621395, + 0.00772217009216547, + 0.006467015482485294, + 0.006957335863262415, + 0.007262269966304302, + 0.008671198971569538, + 0.010149340145289898, + 0.013013020157814026, + 0.006950397975742817, + 0.006550726946443319, + 0.008668622933328152, + 0.008285506628453732, + 0.009312518872320652, + 0.012790578417479992, + 0.009445139206945896, + 0.013375341892242432, + 0.008085396140813828, + 0.004015788435935974, + 0.0072797262109816074, + 0.006672354880720377, + 0.008086655288934708, + 0.005092293489724398, + 0.005530685186386108, + 0.006313986610621214, + 0.004083074163645506, + 0.005585685838013887, + 0.006882188841700554, + 0.013091577216982841, + 0.02166776917874813, + 0.010195926763117313, + 0.006883870344609022, + 0.01155057456344366, + 0.008324498310685158, + 0.018768133595585823, + 0.022958751767873764, + 0.03088144212961197, + 0.029087048023939133, + 0.036434803158044815, + 0.013988873921334743, + 0.0525335930287838, + 0.02492116205394268, + 0.023819684982299805, + 0.026878109201788902, + 0.02197663113474846, + 0.03540268540382385, + 0.02862771414220333, + 0.02310660481452942, + 0.0418781042098999, + 0.022218063473701477, + 0.006625354755669832, + 0.005600898060947657, + 0.0053881434723734856, + 0.005618717055767775, + 0.007751257624477148, + 0.00597855169326067, + 0.009054630063474178, + 0.004604582209140062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.005431537050753832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000380754470825 + ], + "opposite_logprob_tokens": [ + -2.6950037479400635 + ], + "logprob_reward": -0.07000380754470825, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000380754470825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10972005128860474, + 0.036739084869623184, + 0.04157216474413872, + 0.07902076095342636, + 0.06487563252449036, + 0.030697129666805267, + 0.10930587351322174, + 0.09699486196041107, + 0.35984277725219727, + 0.0712316483259201 + ], + "credit_map": [ + 0.007341546472162008, + 0.0072507900185883045, + 0.014411823824048042, + 0.016081007197499275, + 0.01801127940416336, + 0.009595136158168316, + 0.020221715793013573, + 0.016806766390800476, + 0.004416486248373985, + 0.0036524822935462, + 0.004547858145087957, + 0.004175553098320961, + 0.004604615271091461, + 0.0036767139099538326, + 0.006900522392243147, + 0.004764854907989502, + 0.006416862830519676, + 0.004938345402479172, + 0.005043648183345795, + 0.0067086974158883095, + 0.003303108736872673, + 0.004196220077574253, + 0.006103475112468004, + 0.004861804656684399, + 0.008072183467447758, + 0.010050017386674881, + 0.008501577191054821, + 0.00998089462518692, + 0.009353597648441792, + 0.009469225071370602, + 0.015528286807239056, + 0.00806498620659113, + 0.0059833768755197525, + 0.005805144552141428, + 0.009360667318105698, + 0.00791233591735363, + 0.011200481094419956, + 0.007890407927334309, + 0.010957262478768826, + 0.005765954963862896, + 0.0031274608336389065, + 0.0046511306427419186, + 0.0034298065584152937, + 0.004179858602583408, + 0.004322112072259188, + 0.0034251101315021515, + 0.004787897691130638, + 0.0027737536001950502, + 0.012576647102832794, + 0.014379060827195644, + 0.013503972440958023, + 0.030200256034731865, + 0.006525853648781776, + 0.007469369564205408, + 0.011881869286298752, + 0.012768848799169064, + 0.008141294121742249, + 0.009275011718273163, + 0.011560576036572456, + 0.013799292035400867, + 0.013776605017483234, + 0.006592718884348869, + 0.022818680852651596, + 0.011030685156583786, + 0.0880122110247612, + 0.030368341132998466, + 0.025967584922909737, + 0.025338847190141678, + 0.02886730432510376, + 0.12152399122714996, + 0.02172655425965786, + 0.018037954345345497, + 0.010707424022257328, + 0.005853279028087854, + 0.018744580447673798, + 0.005773760844022036, + 0.005465395282953978, + 0.009443123824894428, + 0.0045895385555922985, + 0.010654548183083534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07000380754470825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019414640264585614 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019414640264585614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019414640264585614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793079823255539, + 0.035013437271118164, + 0.05600738152861595, + 0.09397388994693756, + 0.1635928750038147, + 0.044350799173116684, + 0.07811025530099869, + 0.10542761534452438, + 0.2071896195411682, + 0.037026166915893555 + ], + "credit_map": [ + 0.016369014978408813, + 0.010344226844608784, + 0.024797197431325912, + 0.019592314958572388, + 0.031088534742593765, + 0.017637833952903748, + 0.036793116480112076, + 0.022685734555125237, + 0.004488627426326275, + 0.0032600504346191883, + 0.003829434048384428, + 0.0040860166773200035, + 0.004519312642514706, + 0.004142799414694309, + 0.0067182802595198154, + 0.003968913573771715, + 0.003913360647857189, + 0.005590822547674179, + 0.006021811161190271, + 0.006180388852953911, + 0.007157281041145325, + 0.011700225062668324, + 0.008963428437709808, + 0.006480056792497635, + 0.007274114526808262, + 0.00915834866464138, + 0.010602700524032116, + 0.011696123518049717, + 0.015221163630485535, + 0.009251605719327927, + 0.01940282993018627, + 0.011366997845470905, + 0.0175807885825634, + 0.019012680277228355, + 0.021296489983797073, + 0.021525854244828224, + 0.019564932212233543, + 0.023473788052797318, + 0.022420553490519524, + 0.01871778629720211, + 0.004091281443834305, + 0.006556818261742592, + 0.006350396201014519, + 0.00768329156562686, + 0.003071653889492154, + 0.006793906446546316, + 0.005387117154896259, + 0.004416332580149174, + 0.008737549185752869, + 0.0076837316155433655, + 0.01147220004349947, + 0.018461452797055244, + 0.007899043150246143, + 0.005117130000144243, + 0.012851317413151264, + 0.00588782224804163, + 0.009424526244401932, + 0.010621621273458004, + 0.015224066562950611, + 0.01329723745584488, + 0.016175497323274612, + 0.008697561919689178, + 0.021143482998013496, + 0.010843616910278797, + 0.04818752408027649, + 0.01886218972504139, + 0.013593936339020729, + 0.014610890299081802, + 0.017412032932043076, + 0.07122602313756943, + 0.01108236238360405, + 0.012214651331305504, + 0.003792852396145463, + 0.004376554861664772, + 0.004046299494802952, + 0.004974171053618193, + 0.005269339773803949, + 0.0040173218585550785, + 0.005715769715607166, + 0.004833855666220188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0019414640264585614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762161472812295 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.00762161472812295, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762161472812295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04982994496822357, + 0.027384614571928978, + 0.04261518269777298, + 0.1173897534608841, + 0.07822664082050323, + 0.034711454063653946, + 0.06828776001930237, + 0.28419169783592224, + 0.2586131989955902, + 0.038749780505895615 + ], + "credit_map": [ + 0.003980100620537996, + 0.004044405650347471, + 0.010088914074003696, + 0.007400585804134607, + 0.007982772774994373, + 0.005163745954632759, + 0.006046650465577841, + 0.005122767295688391, + 0.002063515828922391, + 0.0022687306627631187, + 0.0030829980969429016, + 0.004008590709418058, + 0.0035669105127453804, + 0.002549059921875596, + 0.006122988648712635, + 0.0037218204233795404, + 0.003729752032086253, + 0.003271484049037099, + 0.00610874779522419, + 0.006749557331204414, + 0.00453604944050312, + 0.005190342664718628, + 0.008174031041562557, + 0.00485521974042058, + 0.010717252269387245, + 0.014354255981743336, + 0.010629541240632534, + 0.01265034545212984, + 0.01648557558655739, + 0.010782971046864986, + 0.028232675045728683, + 0.013537134975194931, + 0.007445075549185276, + 0.007882311940193176, + 0.009602396748960018, + 0.009068809449672699, + 0.012441865168511868, + 0.008702198043465614, + 0.012935115024447441, + 0.010148866102099419, + 0.001994396559894085, + 0.00481319148093462, + 0.004172590561211109, + 0.006259199697524309, + 0.004523071460425854, + 0.0030829969327896833, + 0.006303439848124981, + 0.003562565892934799, + 0.008354758843779564, + 0.006139070726931095, + 0.00612178398296237, + 0.011155168525874615, + 0.01090526394546032, + 0.006158883683383465, + 0.012024547904729843, + 0.00742828194051981, + 0.022793129086494446, + 0.02780691720545292, + 0.03706739842891693, + 0.03694771230220795, + 0.04444466531276703, + 0.0176630150526762, + 0.06584590673446655, + 0.03162294998764992, + 0.023796623572707176, + 0.02820584364235401, + 0.022810064256191254, + 0.037401486188173294, + 0.03553187847137451, + 0.03247396647930145, + 0.049550969153642654, + 0.028842365369200706, + 0.0032055540941655636, + 0.005766899324953556, + 0.0034049353562295437, + 0.005670133512467146, + 0.005575152114033699, + 0.00453780684620142, + 0.007213730830699205, + 0.0033755695912986994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.00762161472812295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177451922558248 + ], + "opposite_logprob_tokens": [ + -6.7511773109436035 + ], + "logprob_reward": -0.001177451922558248, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177451922558248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09204518795013428, + 0.030889643356204033, + 0.042129989713430405, + 0.07468805462121964, + 0.10596320033073425, + 0.04650535434484482, + 0.14086206257343292, + 0.16875343024730682, + 0.23758113384246826, + 0.060581985861063004 + ], + "credit_map": [ + 0.01058957353234291, + 0.007060326635837555, + 0.019990133121609688, + 0.009136258624494076, + 0.011683246120810509, + 0.01337064802646637, + 0.009824682958424091, + 0.01039031520485878, + 0.005173635669052601, + 0.0027370192110538483, + 0.004539148882031441, + 0.003316901158541441, + 0.003690767101943493, + 0.0026098883245140314, + 0.005213206168264151, + 0.00360907637514174, + 0.005371293984353542, + 0.005488555412739515, + 0.005454464349895716, + 0.00711941160261631, + 0.004038751125335693, + 0.004067517817020416, + 0.005137530155479908, + 0.005452457815408707, + 0.008476799353957176, + 0.0063475994393229485, + 0.006908601615577936, + 0.00804965477436781, + 0.01227992307394743, + 0.008245514705777168, + 0.015617218799889088, + 0.008762737736105919, + 0.011819246225059032, + 0.011174214072525501, + 0.013441760092973709, + 0.012373792938888073, + 0.014448439702391624, + 0.013753724284470081, + 0.016213135793805122, + 0.012738875113427639, + 0.003471566829830408, + 0.007516790181398392, + 0.005907625891268253, + 0.009139890782535076, + 0.00419084494933486, + 0.0047263652086257935, + 0.006872082129120827, + 0.004680183716118336, + 0.020140767097473145, + 0.023617813363671303, + 0.0172190610319376, + 0.017285142093896866, + 0.01602023094892502, + 0.010871612466871738, + 0.017377691343426704, + 0.01832972653210163, + 0.019428778439760208, + 0.02402684837579727, + 0.022121258080005646, + 0.022210728377103806, + 0.022204222157597542, + 0.016371432691812515, + 0.02934117242693901, + 0.013048984110355377, + 0.02417781390249729, + 0.028801236301660538, + 0.023969406262040138, + 0.03761107847094536, + 0.031549278646707535, + 0.023073354735970497, + 0.043781109154224396, + 0.024617841467261314, + 0.01129450835287571, + 0.006428816355764866, + 0.0060849799774587154, + 0.006684694439172745, + 0.007357277907431126, + 0.007142691873013973, + 0.010223610326647758, + 0.005365407094359398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.001177451922558248 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, the camera panning left in a slow, graceful arc, captures the depths of a quiet, ancient courtyard. The floor is paved with mottled bluestone, covered in a thick, iridescent blanket of fallen leaves, primarily golden ginkgo and crimson maple. A woman dressed in flowing Hanfu (Chinese traditional clothing) takes center stage. She wears an elegant jade-green double-breasted ruqun (skirt), with wide silk sleeves delicately embroidered with flowing cloud motifs in silver thread, gleaming softly in the gentle afternoon sun. Her long, jet-black hair is tied into a classically refined bun, secured with a simple jade hairpin. Initially facing away from the camera, she then, with remarkable slowness, begins to turn. Her wide skirt and sleeves trace graceful arcs through the air, causing the fallen leaves at her feet to swirl gently. Her movements are smooth and uninterrupted. Finally, when halfway around, she tilts her head slightly, casting a long, affectionate glance back. A hint of melancholy graces her delicate features, and her eyes, as if telling a thousand-year-old story, evoke a sense of tranquility, nostalgia, and poetry. The entire scene is bathed in warm golden light, and the air is filled with a tranquil, nostalgic, and poetic atmosphere.", + "seed": 22701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_15/index_2372_seed_22701_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019905389368068425, + "vlm_logprob": -0.019905389368068425, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.019905389368068425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006173706613481045 + ], + "opposite_logprob_tokens": [ + -5.256173610687256 + ], + "logprob_reward": -0.006173706613481045, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006173706613481045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2819729447364807, + 0.05584803223609924, + 0.05514386668801308, + 0.06052098423242569, + 0.07713974267244339, + 0.03939451277256012, + 0.03481387346982956, + 0.18377572298049927, + 0.09993939101696014, + 0.11145099252462387 + ], + "credit_map": [ + 0.031507253646850586, + 0.044114887714385986, + 0.04371782764792442, + 0.06627130508422852, + 0.01928994432091713, + 0.02280563674867153, + 0.031237266957759857, + 0.023028802126646042, + 0.0064094471745193005, + 0.008105437271296978, + 0.006693487521260977, + 0.010157999582588673, + 0.00689638452604413, + 0.0050381021574139595, + 0.008148721419274807, + 0.004398446995764971, + 0.007440927904099226, + 0.010154283605515957, + 0.0066508278250694275, + 0.006271855439990759, + 0.0069052125327289104, + 0.006205815356224775, + 0.006091322749853134, + 0.005423618480563164, + 0.006690893322229385, + 0.007656374014914036, + 0.009191527031362057, + 0.006986030377447605, + 0.006653375457972288, + 0.00726049579679966, + 0.006599195767194033, + 0.009483087807893753, + 0.010477377101778984, + 0.014135481789708138, + 0.009638680145144463, + 0.0079582454636693, + 0.007689378224313259, + 0.008195992559194565, + 0.009400776587426662, + 0.009643802419304848, + 0.005682023707777262, + 0.005305737257003784, + 0.007043627090752125, + 0.0030884447041898966, + 0.005248010158538818, + 0.004131588619202375, + 0.005525125190615654, + 0.0033699534833431244, + 0.007400230970233679, + 0.0055139739997684956, + 0.008334251120686531, + 0.003942388575524092, + 0.0022021334152668715, + 0.0016588745638728142, + 0.0022022873163223267, + 0.0035597304813563824, + 0.02358606457710266, + 0.021549906581640244, + 0.06453869491815567, + 0.040561970323324203, + 0.008759340271353722, + 0.006283259019255638, + 0.008925871923565865, + 0.009570591151714325, + 0.01767296902835369, + 0.013330526649951935, + 0.018376445397734642, + 0.020105961710214615, + 0.0076569789089262486, + 0.008358586579561234, + 0.005968739278614521, + 0.008469171822071075, + 0.012888082303106785, + 0.006498444359749556, + 0.007899852469563484, + 0.026109954342246056, + 0.006752582732588053, + 0.007324921432882547, + 0.008988538756966591, + 0.03498860448598862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.006173706613481045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896597683429718 + ], + "opposite_logprob_tokens": [ + -2.578965902328491 + ], + "logprob_reward": -0.07896597683429718, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896597683429718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3416680693626404, + 0.09533580392599106, + 0.15321989357471466, + 0.1036171242594719, + 0.030860930681228638, + 0.02641317993402481, + 0.01370882336050272, + 0.03326576575636864, + 0.07015112787485123, + 0.13175922632217407 + ], + "credit_map": [ + 0.03835815191268921, + 0.0582151859998703, + 0.07031562179327011, + 0.0644151046872139, + 0.023205071687698364, + 0.024757282808423042, + 0.032108332961797714, + 0.030293330550193787, + 0.00944528542459011, + 0.014419840648770332, + 0.012858130037784576, + 0.018214045092463493, + 0.011691455729305744, + 0.007161040790379047, + 0.011900662444531918, + 0.009645359590649605, + 0.020117728039622307, + 0.021152786910533905, + 0.018986739218235016, + 0.015273100696504116, + 0.01731540448963642, + 0.024668116122484207, + 0.017008649185299873, + 0.018697388470172882, + 0.013868898153305054, + 0.0109562324360013, + 0.013096075505018234, + 0.009489452466368675, + 0.009758340194821358, + 0.016820063814520836, + 0.009119379334151745, + 0.020508693531155586, + 0.003113748272880912, + 0.002749726874753833, + 0.006666752975434065, + 0.003731333650648594, + 0.002740840893238783, + 0.0029473723843693733, + 0.0027711435686796904, + 0.006140018347650766, + 0.003487840062007308, + 0.004506708588451147, + 0.004302928224205971, + 0.002240516943857074, + 0.0023906573187559843, + 0.0031429489608854055, + 0.0024527842178940773, + 0.0038887965492904186, + 0.002546315547078848, + 0.002323566935956478, + 0.0015507404459640384, + 0.0014798205811530352, + 0.0017429470317438245, + 0.0011094725923612714, + 0.0013760554138571024, + 0.0015799063257873058, + 0.003691211575642228, + 0.0037679639644920826, + 0.007040478754788637, + 0.007749209180474281, + 0.003670299658551812, + 0.0023663302417844534, + 0.002884838031604886, + 0.002095436677336693, + 0.010388880036771297, + 0.009575135074555874, + 0.012865962460637093, + 0.013665415346622467, + 0.007408598903566599, + 0.006062369793653488, + 0.0045001450926065445, + 0.005684634670615196, + 0.014740892685949802, + 0.009273366071283817, + 0.012936905026435852, + 0.029962584376335144, + 0.005339296068996191, + 0.010066679678857327, + 0.009216477163136005, + 0.04022303968667984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.07896597683429718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067384387366473675 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.0067384387366473675, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067384387366473675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10272382199764252, + 0.04943636804819107, + 0.04819253087043762, + 0.08214463293552399, + 0.02712199278175831, + 0.04413270950317383, + 0.10010992735624313, + 0.1308647096157074, + 0.2948438823223114, + 0.12042948603630066 + ], + "credit_map": [ + 0.014051622711122036, + 0.015400069765746593, + 0.016590040177106857, + 0.01788683608174324, + 0.0075450860895216465, + 0.01211460493505001, + 0.011274364776909351, + 0.007861189544200897, + 0.0061432672664523125, + 0.006908654700964689, + 0.0052176774479448795, + 0.00913641881197691, + 0.004709166474640369, + 0.004589851945638657, + 0.008720459416508675, + 0.0040108696557581425, + 0.005167306400835514, + 0.008442847058176994, + 0.006876752246171236, + 0.006725696846842766, + 0.004202211741358042, + 0.005313858389854431, + 0.005534225609153509, + 0.005929631181061268, + 0.010820186696946621, + 0.007614207454025745, + 0.014631181955337524, + 0.007200739812105894, + 0.007307529449462891, + 0.013073056004941463, + 0.006761233322322369, + 0.014736494049429893, + 0.003204334992915392, + 0.0035671258810907602, + 0.004518134519457817, + 0.003909487277269363, + 0.00357986893504858, + 0.0023616694379597902, + 0.0032331382390111685, + 0.002748229308053851, + 0.006519521586596966, + 0.004263778682798147, + 0.008051782846450806, + 0.0042546596378088, + 0.004998826887458563, + 0.004757291171699762, + 0.005221672356128693, + 0.006065174005925655, + 0.030631113797426224, + 0.017537575215101242, + 0.02214476279914379, + 0.010924150235950947, + 0.0044774520210921764, + 0.004030295647680759, + 0.004114956595003605, + 0.006249612662941217, + 0.018457895144820213, + 0.018029751256108284, + 0.039643775671720505, + 0.024755753576755524, + 0.007499785628169775, + 0.007895897142589092, + 0.006346745416522026, + 0.008235096000134945, + 0.05292823538184166, + 0.03454636409878731, + 0.06344003975391388, + 0.05304308235645294, + 0.026210054755210876, + 0.017616454511880875, + 0.021271152421832085, + 0.025788480415940285, + 0.030458776280283928, + 0.014192271046340466, + 0.01709032990038395, + 0.02680147998034954, + 0.005657368339598179, + 0.011054329574108124, + 0.006422741338610649, + 0.008752180263400078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0067384387366473675 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731334142386913 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731334142386913, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731334142386913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441546380519867, + 0.07208463549613953, + 0.10800030827522278, + 0.041950199753046036, + 0.11702331900596619, + 0.09232695400714874, + 0.02397819049656391, + 0.1072605550289154, + 0.07549931108951569, + 0.21772189438343048 + ], + "credit_map": [ + 0.01701737754046917, + 0.023215418681502342, + 0.02491278201341629, + 0.03416798636317253, + 0.011515078134834766, + 0.010610614903271198, + 0.014572098851203918, + 0.008143282495439053, + 0.005376708228141069, + 0.007928591221570969, + 0.010361026041209698, + 0.022825704887509346, + 0.00865628756582737, + 0.00435546925291419, + 0.008609558455646038, + 0.003971289377659559, + 0.009045745246112347, + 0.012003735639154911, + 0.016019243746995926, + 0.014859471470117569, + 0.022240348160266876, + 0.01096775196492672, + 0.013026325963437557, + 0.00983767956495285, + 0.004424791317433119, + 0.003957829438149929, + 0.005822043400257826, + 0.004114992916584015, + 0.004579438827931881, + 0.005989255383610725, + 0.006863401737064123, + 0.006198444869369268, + 0.015192619524896145, + 0.014753560535609722, + 0.019318651407957077, + 0.01131780631840229, + 0.01254936121404171, + 0.009337501600384712, + 0.018120916560292244, + 0.016432909294962883, + 0.011861938051879406, + 0.014878200367093086, + 0.017800260335206985, + 0.006922757253050804, + 0.008468667045235634, + 0.009736007079482079, + 0.012728272005915642, + 0.009930853731930256, + 0.0033685213420540094, + 0.0028675785288214684, + 0.003137075575068593, + 0.0027058206032961607, + 0.0032987960148602724, + 0.002279892098158598, + 0.0029975087381899357, + 0.003322998760268092, + 0.016214042901992798, + 0.012534016743302345, + 0.02429802343249321, + 0.0156569667160511, + 0.009127607569098473, + 0.010857830755412579, + 0.010609813034534454, + 0.007962249219417572, + 0.00918320007622242, + 0.01187523826956749, + 0.011588147841393948, + 0.014394775032997131, + 0.006755608133971691, + 0.010418243706226349, + 0.004872622434049845, + 0.0064114760607481, + 0.016232816502451897, + 0.01007138006389141, + 0.016462309285998344, + 0.06647898256778717, + 0.01153278723359108, + 0.010254199616611004, + 0.01354318019002676, + 0.07314624637365341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.006731334142386913 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot featuring a woman in Hanfu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning left in a slow, graceful arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor paved with mottled bluestone covered in golden ginkgo and crimson maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman wear a jade-green double-breasted ruqun with silver cloud motifs and turn slowly from facing away to facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light within an ancient courtyard setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009174905135296285 + ], + "opposite_logprob_tokens": [ + -7.000917434692383 + ], + "logprob_reward": -0.0009174905135296285, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009174905135296285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23174090683460236, + 0.11407781392335892, + 0.0332944430410862, + 0.05200276896357536, + 0.1263553649187088, + 0.05568401888012886, + 0.10612019896507263, + 0.0518006831407547, + 0.07124002277851105, + 0.15768373012542725 + ], + "credit_map": [ + 0.02763991244137287, + 0.038004811853170395, + 0.031322382390499115, + 0.046614840626716614, + 0.017393676564097404, + 0.022577617317438126, + 0.028002040460705757, + 0.02018563449382782, + 0.011309024877846241, + 0.015160653740167618, + 0.01503098662942648, + 0.03196045756340027, + 0.010831679217517376, + 0.007866453379392624, + 0.01244672853499651, + 0.009471822530031204, + 0.0025742819998413324, + 0.003251135116443038, + 0.0049340687692165375, + 0.0036159923765808344, + 0.004870923236012459, + 0.00507225701585412, + 0.005570882465690374, + 0.0034049050882458687, + 0.008737228810787201, + 0.007971215061843395, + 0.0082168597728014, + 0.004990097600966692, + 0.005895139649510384, + 0.00473658787086606, + 0.004308832343667746, + 0.007146806921809912, + 0.02013804018497467, + 0.022922972217202187, + 0.018851155415177345, + 0.010562577284872532, + 0.007729827892035246, + 0.016878603026270866, + 0.007643300108611584, + 0.021628892049193382, + 0.008828038349747658, + 0.011704028584063053, + 0.008152943104505539, + 0.003585674101486802, + 0.004705589730292559, + 0.008677566424012184, + 0.003046513069421053, + 0.0069836643524467945, + 0.027586424723267555, + 0.02036345563828945, + 0.02488187700510025, + 0.012265187688171864, + 0.005173513200134039, + 0.005059977062046528, + 0.004417605232447386, + 0.006372158881276846, + 0.006339319050312042, + 0.006508416961878538, + 0.016496922820806503, + 0.011933262459933758, + 0.002211512764915824, + 0.0028101515490561724, + 0.002475285902619362, + 0.0030258127953857183, + 0.01195506937801838, + 0.009637250564992428, + 0.017080796882510185, + 0.011958695016801357, + 0.005054392386227846, + 0.007878214120864868, + 0.0026569778565317392, + 0.005018629599362612, + 0.01361492183059454, + 0.012278982438147068, + 0.01623503491282463, + 0.03937793895602226, + 0.008431346155703068, + 0.010728118009865284, + 0.0097865154966712, + 0.04723086580634117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0009174905135296285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_tensor([5998, 5998])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_tensor([5998, 5998])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1a3777c7526690304a4726863da23a2d726c7566 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_tensor([5998, 5998])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot captures a teenage boy bicycling frantically along a narrow, rain-slicked sidewalk in a sprawling metropolis. The scene is rendered in a classic 1980s anime style, characterized by bold outlines and a stark, hand-drawn aesthetic. The teenager's dark hair is matted to his forehead by rain, his face a mask of determination, his teeth gnashed. He wears a bright yellow nylon windbreaker that gleams in the city lights, paired with dark, soaked jeans. Hunched low over the handlebars of his simple road bike, his legs a blur as he pedals relentlessly, kicking up spectacular splashes and capturing a riot of color. The alley is flanked by towering, dark buildings, their surfaces slick with rain. Vibrant neon signs in vibrant pink, cyan, and Japanese script cast long, distorted reflections across the massive puddle on the asphalt. As the camera keeps pace, it captures the dramatic anamorphic lens flare cast horizontally by the overhead streetlights, while a fine, grainy film texture overlays the image, completing the urgent and visually saturated retro atmosphere.", + "seed": 32826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_5998_seed_32826_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054449799051508305, + "vlm_logprob": -0.054449799051508305, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.054449799051508305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25504186749458313 + ], + "opposite_logprob_tokens": [ + -1.5050418376922607 + ], + "logprob_reward": -0.25504186749458313, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25504186749458313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21556000411510468, + 0.059724144637584686, + 0.08448794484138489, + 0.06991875171661377, + 0.0404350571334362, + 0.17419122159481049, + 0.13034199178218842, + 0.07814180105924606, + 0.10132995992898941, + 0.0458691269159317 + ], + "credit_map": [ + 0.017968766391277313, + 0.01421296689659357, + 0.041309528052806854, + 0.018475254997611046, + 0.030066069215536118, + 0.04227067157626152, + 0.019205262884497643, + 0.032051488757133484, + 0.009084447287023067, + 0.009305683895945549, + 0.008458021096885204, + 0.0060839527286589146, + 0.009031700901687145, + 0.006240952759981155, + 0.0055266255512833595, + 0.005992757622152567, + 0.012771223671734333, + 0.008818680420517921, + 0.01840769313275814, + 0.010141086764633656, + 0.007278751581907272, + 0.0061006383039057255, + 0.007669029291719198, + 0.013300838880240917, + 0.012594074942171574, + 0.007028299383819103, + 0.007250574883073568, + 0.00883579533547163, + 0.009407789446413517, + 0.00767090218141675, + 0.009061155840754509, + 0.008070159703493118, + 0.0036640488542616367, + 0.004960896447300911, + 0.004459073301404715, + 0.0050685107707977295, + 0.007081704679876566, + 0.003952853847295046, + 0.00673122052103281, + 0.004516750108450651, + 0.029686393216252327, + 0.007622775621712208, + 0.03499206155538559, + 0.015076894313097, + 0.014563746750354767, + 0.025907572358846664, + 0.021710826084017754, + 0.024630945175886154, + 0.013437921181321144, + 0.010781859047710896, + 0.015409226529300213, + 0.030995197594165802, + 0.012422103434801102, + 0.012586196884512901, + 0.018528183922171593, + 0.016181303188204765, + 0.012129251845180988, + 0.006227518431842327, + 0.018035678192973137, + 0.005132265388965607, + 0.010628608986735344, + 0.010332301259040833, + 0.006020852364599705, + 0.009635328315198421, + 0.014786699786782265, + 0.008836413733661175, + 0.011650286614894867, + 0.009680432267487049, + 0.009382428601384163, + 0.017864318564534187, + 0.01259208656847477, + 0.016537288203835487, + 0.005160615313798189, + 0.005377049092203379, + 0.005969640798866749, + 0.005176831036806107, + 0.004685614723712206, + 0.006739716045558453, + 0.004575758706778288, + 0.008183897472918034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.25504186749458313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006757265888154507 + ], + "opposite_logprob_tokens": [ + -5.0067572593688965 + ], + "logprob_reward": -0.006757265888154507, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006757265888154507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2663617730140686, + 0.05614778771996498, + 0.09406303614377975, + 0.059839457273483276, + 0.047186195850372314, + 0.13235291838645935, + 0.16178086400032043, + 0.06824008375406265, + 0.04969213902950287, + 0.06433571875095367 + ], + "credit_map": [ + 0.01813756488263607, + 0.015538697130978107, + 0.04714229702949524, + 0.02510710246860981, + 0.03787679597735405, + 0.060320910066366196, + 0.02245219051837921, + 0.03978623449802399, + 0.0078019569627940655, + 0.007593754678964615, + 0.007846237160265446, + 0.005917267873883247, + 0.008596150204539299, + 0.005034940782934427, + 0.006333706434816122, + 0.0070237815380096436, + 0.01330078300088644, + 0.00866609811782837, + 0.018865427002310753, + 0.011513754725456238, + 0.008825377561151981, + 0.008344435133039951, + 0.012391028925776482, + 0.012156140990555286, + 0.012143956497311592, + 0.004537841770797968, + 0.006666924804449081, + 0.006647421512752771, + 0.006409314461052418, + 0.0067919110879302025, + 0.008232026360929012, + 0.008410067297518253, + 0.004764074459671974, + 0.005954081658273935, + 0.006059957668185234, + 0.006739629898220301, + 0.0062462580390274525, + 0.005224095191806555, + 0.00505576329305768, + 0.007142337504774332, + 0.02192435972392559, + 0.005551803857088089, + 0.02324819192290306, + 0.007377144880592823, + 0.009003756567835808, + 0.03183390572667122, + 0.011205621995031834, + 0.022208137437701225, + 0.023725753650069237, + 0.013643013313412666, + 0.02206927165389061, + 0.030296167358756065, + 0.01548535842448473, + 0.011839678511023521, + 0.018226955085992813, + 0.026494676247239113, + 0.009133237414062023, + 0.005481858737766743, + 0.01345943845808506, + 0.006805597338825464, + 0.009594564326107502, + 0.00772431492805481, + 0.00849419180303812, + 0.00754688261076808, + 0.006608284078538418, + 0.004491386469453573, + 0.007273899856954813, + 0.004593143705278635, + 0.004949823021888733, + 0.007597276009619236, + 0.0058923992328345776, + 0.008285933174192905, + 0.006941904313862324, + 0.008085520938038826, + 0.01178495679050684, + 0.007644939702004194, + 0.006292938720434904, + 0.007404085248708725, + 0.006727396976202726, + 0.00945398397743702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.006757265888154507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21736779808998108, + 0.032602161169052124, + 0.286022424697876, + 0.056630056351423264, + 0.01865026168525219, + 0.1838912069797516, + 0.061676062643527985, + 0.019027477130293846, + 0.020533092319965363, + 0.10359940677881241 + ], + "credit_map": [ + 0.01720607280731201, + 0.01228400319814682, + 0.03794512897729874, + 0.02246958017349243, + 0.02581970952451229, + 0.04214741662144661, + 0.02443751133978367, + 0.03505839407444, + 0.006702783051878214, + 0.004668648820370436, + 0.00534159317612648, + 0.003436575410887599, + 0.003537794342264533, + 0.003476251382380724, + 0.0023217545822262764, + 0.003116762265563011, + 0.04025183245539665, + 0.036719124764204025, + 0.05241160839796066, + 0.030109668150544167, + 0.03506501764059067, + 0.019967732951045036, + 0.03146718814969063, + 0.04003027081489563, + 0.015693560242652893, + 0.003693995764479041, + 0.006993923336267471, + 0.007647211197763681, + 0.004035408142954111, + 0.0053178598172962666, + 0.0059486813843250275, + 0.007299421355128288, + 0.002388493623584509, + 0.0013894162839278579, + 0.003185190260410309, + 0.0024665440432727337, + 0.0016341264126822352, + 0.0021299454383552074, + 0.001299262628890574, + 0.004157283343374729, + 0.033584944903850555, + 0.00988505594432354, + 0.04287686198949814, + 0.011361628770828247, + 0.009229482151567936, + 0.034676168113946915, + 0.012005436234176159, + 0.03027164749801159, + 0.00622321804985404, + 0.004961788654327393, + 0.008580398745834827, + 0.014537958428263664, + 0.005527149420231581, + 0.005905382335186005, + 0.008910011500120163, + 0.007030160166323185, + 0.002656474942341447, + 0.0015387657331302762, + 0.0040197595953941345, + 0.0017703084740787745, + 0.002441475633531809, + 0.0025970647111535072, + 0.0014903004048392177, + 0.002513329265639186, + 0.0030992734245955944, + 0.0019073656294494867, + 0.0027152192778885365, + 0.0018805429572239518, + 0.0020866028498858213, + 0.0031149741262197495, + 0.0020681950263679028, + 0.003660920774564147, + 0.014418518170714378, + 0.01788703352212906, + 0.01821800507605076, + 0.013021369464695454, + 0.00690247118473053, + 0.009981670416891575, + 0.006793526466935873, + 0.016376810148358345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017151420470327139 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.0017151420470327139, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017151420470327139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13999998569488525, + 0.06691057980060577, + 0.06584415584802628, + 0.06328526139259338, + 0.0868832916021347, + 0.22485706210136414, + 0.15447334945201874, + 0.034593019634485245, + 0.04046892747282982, + 0.12268438935279846 + ], + "credit_map": [ + 0.014100526459515095, + 0.007835769094526768, + 0.02184026874601841, + 0.012172745540738106, + 0.022008270025253296, + 0.029608668759465218, + 0.010835214518010616, + 0.021598519757390022, + 0.013081789016723633, + 0.008586126379668713, + 0.011751743964850903, + 0.008445481769740582, + 0.006436942145228386, + 0.006469446234405041, + 0.005081988871097565, + 0.0070570590905845165, + 0.0072924052365124226, + 0.007818826474249363, + 0.008907315321266651, + 0.006283117458224297, + 0.009345636703073978, + 0.005472388584166765, + 0.011308088898658752, + 0.009416371583938599, + 0.010913491249084473, + 0.003489779308438301, + 0.008898524567484856, + 0.006447117310017347, + 0.006159941200166941, + 0.010086972266435623, + 0.006529678590595722, + 0.010759752243757248, + 0.01228388398885727, + 0.007170071359723806, + 0.012776830233633518, + 0.010099854320287704, + 0.007186776492744684, + 0.013080205768346786, + 0.00907857809215784, + 0.015207086689770222, + 0.03570764511823654, + 0.010710085742175579, + 0.054117877036333084, + 0.011573473922908306, + 0.011322975158691406, + 0.04834368824958801, + 0.015206840820610523, + 0.03787445276975632, + 0.01737573742866516, + 0.012683584354817867, + 0.021294917911291122, + 0.03658490628004074, + 0.012230174615979195, + 0.013910573907196522, + 0.022943411022424698, + 0.01745002530515194, + 0.00404707994312048, + 0.003120609326288104, + 0.0068097724579274654, + 0.0034952196292579174, + 0.0052600158378481865, + 0.004329727962613106, + 0.003155214013531804, + 0.004375377669930458, + 0.005542511120438576, + 0.003466993337497115, + 0.007360977120697498, + 0.003196640871465206, + 0.003538125893101096, + 0.005869537591934204, + 0.004466167651116848, + 0.007027972023934126, + 0.019129678606987, + 0.018268628045916557, + 0.020024366676807404, + 0.013946964405477047, + 0.010159717872738838, + 0.01265695970505476, + 0.01007930003106594, + 0.01841876097023487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0017151420470327139 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040955008007586 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.0040955008007586, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040955008007586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1589486449956894, + 0.07853332161903381, + 0.07371565699577332, + 0.10635828226804733, + 0.1165323480963707, + 0.15791329741477966, + 0.09792567044496536, + 0.05456296354532242, + 0.06185651570558548, + 0.09365324676036835 + ], + "credit_map": [ + 0.01978335715830326, + 0.010449964553117752, + 0.033872541040182114, + 0.014425897970795631, + 0.01446540281176567, + 0.023162513971328735, + 0.017265858128666878, + 0.025523114949464798, + 0.006748219020664692, + 0.008053740486502647, + 0.011995567940175533, + 0.010595609433948994, + 0.007641243282705545, + 0.008143753744661808, + 0.007286535110324621, + 0.018068663775920868, + 0.011438034474849701, + 0.010365703143179417, + 0.010805194266140461, + 0.009858681820333004, + 0.007788006216287613, + 0.0055273002944886684, + 0.009662152267992496, + 0.008270587772130966, + 0.023886829614639282, + 0.0067292433232069016, + 0.013923745602369308, + 0.017065763473510742, + 0.008611543104052544, + 0.007821758277714252, + 0.012488847598433495, + 0.01583055593073368, + 0.013587500900030136, + 0.011735612526535988, + 0.01737058535218239, + 0.015152371488511562, + 0.012697095982730389, + 0.01248813234269619, + 0.01661108247935772, + 0.01688997447490692, + 0.022678129374980927, + 0.0073583414778113365, + 0.031607888638973236, + 0.010033047758042812, + 0.01223009917885065, + 0.03102969564497471, + 0.017080454155802727, + 0.025895651429891586, + 0.016572555527091026, + 0.007152135483920574, + 0.014338742941617966, + 0.01988583244383335, + 0.0066167376935482025, + 0.009110631421208382, + 0.012686626054346561, + 0.011562418192625046, + 0.009314518421888351, + 0.003963606897741556, + 0.014463372528553009, + 0.004109119530767202, + 0.004790384788066149, + 0.007056194823235273, + 0.004105442203581333, + 0.006760330870747566, + 0.0112169673666358, + 0.005383679643273354, + 0.007135821972042322, + 0.005144448950886726, + 0.005575672257691622, + 0.01025037094950676, + 0.006819534115493298, + 0.010330025106668472, + 0.012555500492453575, + 0.012147543951869011, + 0.016723880544304848, + 0.009759010747075081, + 0.00805969163775444, + 0.008914503268897533, + 0.011005193926393986, + 0.01448793150484562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0040955008007586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot captures a teenage boy bicycling frantically along a narrow, rain-slicked sidewalk in a sprawling metropolis. The scene is rendered in a classic 1980s anime style, characterized by bold outlines and a stark, hand-drawn aesthetic. The teenager's dark hair is matted to his forehead by rain, his face a mask of determination, his teeth gnashed. He wears a bright yellow nylon windbreaker that gleams in the city lights, paired with dark, soaked jeans. Hunched low over the handlebars of his simple road bike, his legs a blur as he pedals relentlessly, kicking up spectacular splashes and capturing a riot of color. The alley is flanked by towering, dark buildings, their surfaces slick with rain. Vibrant neon signs in vibrant pink, cyan, and Japanese script cast long, distorted reflections across the massive puddle on the asphalt. As the camera keeps pace, it captures the dramatic anamorphic lens flare cast horizontally by the overhead streetlights, while a fine, grainy film texture overlays the image, completing the urgent and visually saturated retro atmosphere.", + "seed": 32827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_5998_seed_32827_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07503972144331783, + "vlm_logprob": -0.07503972144331783, + "vlm_yes_no_margin": 4.475000011920929, + "vlm_token_credit": -0.07503972144331783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35208624601364136 + ], + "opposite_logprob_tokens": [ + -1.2270863056182861 + ], + "logprob_reward": -0.35208624601364136, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.35208624601364136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15037433803081512, + 0.06267526000738144, + 0.11840105056762695, + 0.1934000551700592, + 0.09768769890069962, + 0.05742223933339119, + 0.026547877117991447, + 0.15996916592121124, + 0.05956462025642395, + 0.07395767420530319 + ], + "credit_map": [ + 0.0325143039226532, + 0.019905319437384605, + 0.012155105359852314, + 0.03375367820262909, + 0.010914314538240433, + 0.00883991364389658, + 0.022018801420927048, + 0.010272903367877007, + 0.007056135684251785, + 0.005394516978412867, + 0.008042914792895317, + 0.006302062422037125, + 0.00859296228736639, + 0.010346276685595512, + 0.011846556328237057, + 0.005093839950859547, + 0.012568475678563118, + 0.015624096617102623, + 0.014887139201164246, + 0.01079447939991951, + 0.015901220962405205, + 0.016190573573112488, + 0.020819274708628654, + 0.011615793220698833, + 0.02502983994781971, + 0.020991012454032898, + 0.024928351864218712, + 0.02542024664580822, + 0.02293863333761692, + 0.023977359756827354, + 0.029478225857019424, + 0.020636405795812607, + 0.008819639682769775, + 0.01382688619196415, + 0.006541327107697725, + 0.011143825016915798, + 0.014320988208055496, + 0.014470091089606285, + 0.017714951187372208, + 0.010850002989172935, + 0.009198462590575218, + 0.006728389300405979, + 0.007675092667341232, + 0.006306643132120371, + 0.007683394942432642, + 0.006873027887195349, + 0.00830843299627304, + 0.004648799542337656, + 0.00309366500005126, + 0.001915079657919705, + 0.00309331645257771, + 0.001404820941388607, + 0.003359334310516715, + 0.0049502975307404995, + 0.004881669767200947, + 0.003849694039672613, + 0.025915803387761116, + 0.025107260793447495, + 0.01146653387695551, + 0.014404748566448689, + 0.014246523380279541, + 0.032651640474796295, + 0.02088395692408085, + 0.015292714349925518, + 0.005833358969539404, + 0.019777996465563774, + 0.003520344849675894, + 0.011247576214373112, + 0.004801543429493904, + 0.004900819156318903, + 0.005434032529592514, + 0.004048953298479319, + 0.005078250542283058, + 0.012500163167715073, + 0.005834400188177824, + 0.007402052171528339, + 0.011678947135806084, + 0.008899918757379055, + 0.014001863077282906, + 0.008562089875340462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.35208624601364136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270515102893114 + ], + "opposite_logprob_tokens": [ + -5.255270481109619 + ], + "logprob_reward": -0.005270515102893114, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270515102893114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30127590894699097, + 0.09086275845766068, + 0.038980256766080856, + 0.20803986489772797, + 0.11713539808988571, + 0.10598286241292953, + 0.01645166613161564, + 0.04308552294969559, + 0.030570000410079956, + 0.047615762799978256 + ], + "credit_map": [ + 0.0797886848449707, + 0.03906479850411415, + 0.028297092765569687, + 0.05464326590299606, + 0.015233765356242657, + 0.03921741247177124, + 0.028721198439598083, + 0.01630968600511551, + 0.010379733517765999, + 0.012016893364489079, + 0.007418982684612274, + 0.009897136129438877, + 0.014098326675593853, + 0.01660294272005558, + 0.013385890051722527, + 0.007062868680804968, + 0.00439402274787426, + 0.004487267695367336, + 0.0065713138319551945, + 0.0035812323912978172, + 0.004121960606426001, + 0.004558619111776352, + 0.006434515584260225, + 0.004831328988075256, + 0.024744588881731033, + 0.019749769940972328, + 0.025941351428627968, + 0.023618420585989952, + 0.02742406353354454, + 0.023600922897458076, + 0.04136063531041145, + 0.02160012535750866, + 0.008527718484401703, + 0.010590339079499245, + 0.00784590095281601, + 0.012958220206201077, + 0.017793001607060432, + 0.018515178933739662, + 0.025131629779934883, + 0.01577342301607132, + 0.00999943632632494, + 0.012907006777822971, + 0.00855263415724039, + 0.013371225446462631, + 0.018604891374707222, + 0.016513584181666374, + 0.01729714125394821, + 0.008736949414014816, + 0.00256750313565135, + 0.001653120736591518, + 0.0022085723467171192, + 0.0018553741974756122, + 0.0018848214531317353, + 0.0018295708578079939, + 0.0025855228304862976, + 0.0018671825528144836, + 0.006337563972920179, + 0.007462530396878719, + 0.003404587274417281, + 0.004927070811390877, + 0.005333938170224428, + 0.006463987752795219, + 0.004864356014877558, + 0.004291491117328405, + 0.0028966686222702265, + 0.0055656880140304565, + 0.00404376583173871, + 0.0050008767284452915, + 0.002667108317837119, + 0.0022754399105906487, + 0.004079158417880535, + 0.0040412950329482555, + 0.004352688789367676, + 0.006523634772747755, + 0.005885272286832333, + 0.007369934115558863, + 0.006585736759006977, + 0.0041642808355391026, + 0.006947317160665989, + 0.005786900408565998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005270515102893114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072147637605667 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072147637605667, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072147637605667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26967620849609375, + 0.06089775264263153, + 0.07953859120607376, + 0.1966504156589508, + 0.15345996618270874, + 0.057985980063676834, + 0.0356983058154583, + 0.03915304318070412, + 0.055242400616407394, + 0.05169731751084328 + ], + "credit_map": [ + 0.0316600427031517, + 0.040947094559669495, + 0.02697024494409561, + 0.08563652634620667, + 0.012295562773942947, + 0.02378290891647339, + 0.03279362991452217, + 0.015590229071676731, + 0.00937630608677864, + 0.008925841189920902, + 0.00854480266571045, + 0.004566790070384741, + 0.006623087916523218, + 0.009847110137343407, + 0.0072951167821884155, + 0.005718699656426907, + 0.0076980008743703365, + 0.0137472627684474, + 0.011416623368859291, + 0.006674310192465782, + 0.011839793995022774, + 0.009621099568903446, + 0.011731099337339401, + 0.006810402497649193, + 0.02522711455821991, + 0.019574442878365517, + 0.026583600789308548, + 0.022132590413093567, + 0.02283046953380108, + 0.025577431544661522, + 0.031433187425136566, + 0.023291591554880142, + 0.019650069996714592, + 0.035897888243198395, + 0.011285477317869663, + 0.02055652067065239, + 0.013146239332854748, + 0.02202482335269451, + 0.015674395486712456, + 0.01522454246878624, + 0.00588617566972971, + 0.007563126273453236, + 0.004327470436692238, + 0.006443846505135298, + 0.008791401982307434, + 0.009581861086189747, + 0.009766926057636738, + 0.005625172518193722, + 0.004473627079278231, + 0.0028167616110295057, + 0.00420416472479701, + 0.002766334218904376, + 0.00364862778224051, + 0.005892280954867601, + 0.005936842877417803, + 0.00595966586843133, + 0.0046742158010602, + 0.0051613375544548035, + 0.003166103269904852, + 0.004357384517788887, + 0.005042956676334143, + 0.00745137594640255, + 0.005914225243031979, + 0.0033854444045573473, + 0.004897505044937134, + 0.017437944188714027, + 0.00371262663975358, + 0.013575263321399689, + 0.004182970151305199, + 0.003236955963075161, + 0.005021100398153067, + 0.0031780365388840437, + 0.004896465688943863, + 0.006986507214605808, + 0.0067916568368673325, + 0.0051948437467217445, + 0.006737791933119297, + 0.006155513226985931, + 0.008677473291754723, + 0.0062570637091994286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.011072147637605667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015148365637287498 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015148365637287498, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015148365637287498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17023950815200806, + 0.08808015286922455, + 0.08573923259973526, + 0.33474868535995483, + 0.06334244459867477, + 0.048130232840776443, + 0.02361893840134144, + 0.04211480915546417, + 0.0466899536550045, + 0.09729605168104172 + ], + "credit_map": [ + 0.042841836810112, + 0.01649656891822815, + 0.015541540458798409, + 0.018869031220674515, + 0.011138325557112694, + 0.03576982766389847, + 0.01587684266269207, + 0.013705512508749962, + 0.009319511242210865, + 0.010191160254180431, + 0.011660261079668999, + 0.009659143164753914, + 0.01435824017971754, + 0.01343869511038065, + 0.013986130245029926, + 0.005467006471008062, + 0.012852871790528297, + 0.011153957806527615, + 0.016873423010110855, + 0.006377630401402712, + 0.007474880199879408, + 0.010661590844392776, + 0.011501110158860683, + 0.008843758143484592, + 0.03984866663813591, + 0.03103136457502842, + 0.042597949504852295, + 0.04496100917458534, + 0.03724472597241402, + 0.042137663811445236, + 0.059838585555553436, + 0.037088725715875626, + 0.0081798005849123, + 0.006503966171294451, + 0.004711218643933535, + 0.007052020635455847, + 0.0077460771426558495, + 0.010761733166873455, + 0.010854030959308147, + 0.007533594965934753, + 0.006971955299377441, + 0.00599381048232317, + 0.00497010350227356, + 0.0046678707003593445, + 0.005910573527216911, + 0.007196507416665554, + 0.007154613267630339, + 0.0052647944539785385, + 0.0018577160080894828, + 0.0016281720018014312, + 0.0028839975129812956, + 0.0018271425506100059, + 0.002909442875534296, + 0.00416418444365263, + 0.0043790871277451515, + 0.003969194833189249, + 0.006254263687878847, + 0.007940228097140789, + 0.0031015092972666025, + 0.00489756278693676, + 0.005413247738033533, + 0.005768875125795603, + 0.005222084000706673, + 0.0035170356277376413, + 0.005040592513978481, + 0.010835986584424973, + 0.005449071992188692, + 0.005522770341485739, + 0.005266422405838966, + 0.00434916140511632, + 0.005384163931012154, + 0.004841777961701155, + 0.011160465888679028, + 0.017683055251836777, + 0.018965935334563255, + 0.010253209620714188, + 0.008822205476462841, + 0.008062900975346565, + 0.011289210990071297, + 0.011059064418077469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0015148365637287498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052548618987202644 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.0052548618987202644, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052548618987202644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39607125520706177, + 0.05365176126360893, + 0.04874849319458008, + 0.07781217247247696, + 0.17680031061172485, + 0.04309231415390968, + 0.038219939917325974, + 0.05080873891711235, + 0.07115552574396133, + 0.04363945871591568 + ], + "credit_map": [ + 0.09622661024332047, + 0.06309976428747177, + 0.024555189535021782, + 0.10688919574022293, + 0.02079981565475464, + 0.021665051579475403, + 0.04595658555626869, + 0.016879064962267876, + 0.006828164216130972, + 0.006454895716160536, + 0.0073096598498523235, + 0.005524837411940098, + 0.00601376686245203, + 0.008424988016486168, + 0.007999868132174015, + 0.005095584783703089, + 0.0059699867852032185, + 0.006162961013615131, + 0.008153790608048439, + 0.004175011999905109, + 0.004766132216900587, + 0.007225356064736843, + 0.006614229641854763, + 0.005681028123944998, + 0.00742476899176836, + 0.007628733292222023, + 0.012615055777132511, + 0.008014986291527748, + 0.011190793476998806, + 0.00818562414497137, + 0.015665654093027115, + 0.007086552679538727, + 0.016515269875526428, + 0.0188919510692358, + 0.011506265960633755, + 0.02390478551387787, + 0.02414553239941597, + 0.02598833292722702, + 0.03380785137414932, + 0.02204032614827156, + 0.006737470626831055, + 0.0063835326582193375, + 0.003263631369918585, + 0.0057118674740195274, + 0.004702048841863871, + 0.006512518506497145, + 0.004460921045392752, + 0.00532032223418355, + 0.007824849337339401, + 0.005684358067810535, + 0.004445770289748907, + 0.003244023770093918, + 0.004425677005201578, + 0.003899517236277461, + 0.005134851671755314, + 0.00356089323759079, + 0.005570476874709129, + 0.0051769595593214035, + 0.005627899430692196, + 0.005261568818241358, + 0.005846832413226366, + 0.010267063044011593, + 0.006113843992352486, + 0.006944095250219107, + 0.012572893872857094, + 0.00908138882368803, + 0.007138397078961134, + 0.016887227073311806, + 0.006609030067920685, + 0.006212228909134865, + 0.008043148554861546, + 0.004611209500581026, + 0.004620676394551992, + 0.006159752141684294, + 0.0061132232658565044, + 0.006828266195952892, + 0.005311992485076189, + 0.005230217706412077, + 0.004051473923027515, + 0.005323860328644514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0052548618987202644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot captures a teenage boy bicycling frantically along a narrow, rain-slicked sidewalk in a sprawling metropolis. The scene is rendered in a classic 1980s anime style, characterized by bold outlines and a stark, hand-drawn aesthetic. The teenager's dark hair is matted to his forehead by rain, his face a mask of determination, his teeth gnashed. He wears a bright yellow nylon windbreaker that gleams in the city lights, paired with dark, soaked jeans. Hunched low over the handlebars of his simple road bike, his legs a blur as he pedals relentlessly, kicking up spectacular splashes and capturing a riot of color. The alley is flanked by towering, dark buildings, their surfaces slick with rain. Vibrant neon signs in vibrant pink, cyan, and Japanese script cast long, distorted reflections across the massive puddle on the asphalt. As the camera keeps pace, it captures the dramatic anamorphic lens flare cast horizontally by the overhead streetlights, while a fine, grainy film texture overlays the image, completing the urgent and visually saturated retro atmosphere.", + "seed": 32828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_5998_seed_32828_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04880546978674829, + "vlm_logprob": -0.04880546978674829, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.04880546978674829, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22785654664039612 + ], + "opposite_logprob_tokens": [ + -1.6028565168380737 + ], + "logprob_reward": -0.22785654664039612, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22785654664039612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17514613270759583, + 0.11150514334440231, + 0.13574141263961792, + 0.24223200976848602, + 0.0740787461400032, + 0.04404696822166443, + 0.08753480017185211, + 0.0671238899230957, + 0.03945935517549515, + 0.02313154935836792 + ], + "credit_map": [ + 0.01579427160322666, + 0.02568456530570984, + 0.015098721720278263, + 0.014393772929906845, + 0.02935284748673439, + 0.02995849773287773, + 0.018356245011091232, + 0.02650720812380314, + 0.01073842030018568, + 0.014399152249097824, + 0.010139480233192444, + 0.011706291697919369, + 0.015513667836785316, + 0.015722231939435005, + 0.01583085022866726, + 0.017455050721764565, + 0.01129115466028452, + 0.007676216773688793, + 0.013379799202084541, + 0.013095928356051445, + 0.012070918455719948, + 0.039809875190258026, + 0.015409301966428757, + 0.02300821803510189, + 0.02474752813577652, + 0.02503896690905094, + 0.02649141289293766, + 0.014559406787157059, + 0.030967039987444878, + 0.05877934396266937, + 0.021237200126051903, + 0.04041111096739769, + 0.01157335564494133, + 0.005338509567081928, + 0.011930334381759167, + 0.005854682996869087, + 0.007582360412925482, + 0.00944532360881567, + 0.010449344292283058, + 0.011904831975698471, + 0.005492213182151318, + 0.003067175392061472, + 0.009324873797595501, + 0.004014506936073303, + 0.004778974689543247, + 0.004962867591530085, + 0.005379435606300831, + 0.007026920560747385, + 0.008272686041891575, + 0.0049516428261995316, + 0.008952740579843521, + 0.008747701533138752, + 0.0075052580796182156, + 0.01669810526072979, + 0.006478237919509411, + 0.02592843398451805, + 0.006264663301408291, + 0.005453579593449831, + 0.010033990256488323, + 0.00917587149888277, + 0.009233836084604263, + 0.00657529104501009, + 0.009881103411316872, + 0.010505557991564274, + 0.003660668386146426, + 0.003940141294151545, + 0.005237925332039595, + 0.004664091393351555, + 0.0046353572979569435, + 0.004050165414810181, + 0.007286811713129282, + 0.005984194111078978, + 0.0029466014821082354, + 0.0030332060996443033, + 0.0029604381415992975, + 0.0018065691692754626, + 0.004096084274351597, + 0.0027435612864792347, + 0.0022789412178099155, + 0.0032661480363458395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.22785654664039612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651158772408962 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651158772408962, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651158772408962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19652721285820007, + 0.20199082791805267, + 0.1023053526878357, + 0.12602762877941132, + 0.07178264856338501, + 0.06176205351948738, + 0.05507848784327507, + 0.02420997992157936, + 0.14223617315292358, + 0.018079616129398346 + ], + "credit_map": [ + 0.010429940186440945, + 0.024887749925255775, + 0.023033559322357178, + 0.017964815720915794, + 0.037446267902851105, + 0.028742577880620956, + 0.02603614330291748, + 0.027986155822873116, + 0.015359021723270416, + 0.025867152959108353, + 0.017116820439696312, + 0.022687382996082306, + 0.03087422251701355, + 0.03409583121538162, + 0.02482195757329464, + 0.03116842918097973, + 0.00718469824641943, + 0.006507143843919039, + 0.01211125124245882, + 0.008096356876194477, + 0.009617013856768608, + 0.030345069244503975, + 0.011843734420835972, + 0.016600074246525764, + 0.01195440161973238, + 0.007970757782459259, + 0.012597083114087582, + 0.007062744814902544, + 0.01861448772251606, + 0.03332795575261116, + 0.012505115009844303, + 0.02199508436024189, + 0.011655273847281933, + 0.006104369647800922, + 0.010303140617907047, + 0.005730089731514454, + 0.008723102509975433, + 0.008788025937974453, + 0.009576389566063881, + 0.010902254842221737, + 0.007160428911447525, + 0.003952614031732082, + 0.013430866412818432, + 0.0056876433081924915, + 0.0060688406229019165, + 0.007769730407744646, + 0.007210453040897846, + 0.010481477715075016, + 0.006344262510538101, + 0.006749872583895922, + 0.006209672428667545, + 0.007184858899563551, + 0.004839535802602768, + 0.007704441901296377, + 0.003871070919558406, + 0.012174771167337894, + 0.0020714725833386183, + 0.0029009159188717604, + 0.002966111060231924, + 0.0023337346501648426, + 0.0042464095167815685, + 0.0023834072053432465, + 0.004187094978988171, + 0.0031208328437060118, + 0.02846352569758892, + 0.014065643772482872, + 0.02524319477379322, + 0.01186116598546505, + 0.012237698771059513, + 0.01820327155292034, + 0.011488735675811768, + 0.020672934129834175, + 0.001977811101824045, + 0.0020517627708613873, + 0.003094976767897606, + 0.0019699374679476023, + 0.002105462597683072, + 0.002548407996073365, + 0.0018247542902827263, + 0.0025065033696591854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.008651158772408962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815808402374387 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.002815808402374387, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815808402374387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3315533697605133, + 0.12090152502059937, + 0.06651754677295685, + 0.10032688826322556, + 0.03657820448279381, + 0.09209021925926208, + 0.062471963465213776, + 0.05928032100200653, + 0.09797501564025879, + 0.032304923981428146 + ], + "credit_map": [ + 0.043135933578014374, + 0.029471511021256447, + 0.04888506978750229, + 0.031288836151361465, + 0.03504440560936928, + 0.05687985196709633, + 0.033072445541620255, + 0.053775325417518616, + 0.008822153322398663, + 0.01663883775472641, + 0.013106820173561573, + 0.01189049705862999, + 0.020643973723053932, + 0.01801702193915844, + 0.014196616597473621, + 0.01758560724556446, + 0.005659852642565966, + 0.007300774101167917, + 0.007805690634995699, + 0.006216160487383604, + 0.008688973262906075, + 0.01174150314182043, + 0.008783436380326748, + 0.010321157053112984, + 0.010743594728410244, + 0.008948126807808876, + 0.0122936200350523, + 0.007817362435162067, + 0.014734317548573017, + 0.02098565734922886, + 0.009789482690393925, + 0.015014725737273693, + 0.005156942177563906, + 0.002892846940085292, + 0.006811958272010088, + 0.00293880607932806, + 0.0030405325815081596, + 0.004999636672437191, + 0.004590360447764397, + 0.0061471243388950825, + 0.01279421616345644, + 0.006544781848788261, + 0.0206520427018404, + 0.00873149186372757, + 0.008053019642829895, + 0.011767280288040638, + 0.0076881106942892075, + 0.01585928350687027, + 0.004257592372596264, + 0.0037193605676293373, + 0.005897097755223513, + 0.004869375377893448, + 0.006221294403076172, + 0.01372579950839281, + 0.005193255841732025, + 0.018588190898299217, + 0.004274642560631037, + 0.006411244627088308, + 0.005631172098219395, + 0.0066087213344872, + 0.01117120310664177, + 0.005561556667089462, + 0.011211647652089596, + 0.00841013714671135, + 0.01742727868258953, + 0.009167950600385666, + 0.01653609238564968, + 0.008415250107645988, + 0.00919086392968893, + 0.013974973000586033, + 0.009576027281582355, + 0.013686584308743477, + 0.003986081574112177, + 0.005544205661863089, + 0.003025061683729291, + 0.002683746162801981, + 0.006361912004649639, + 0.0031943838112056255, + 0.0030441710259765387, + 0.0044653634540736675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.002815808402374387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015119798481464386 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015119798481464386, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015119798481464386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18520721793174744, + 0.1266978234052658, + 0.1072402372956276, + 0.20287245512008667, + 0.055027078837156296, + 0.05838410183787346, + 0.041748471558094025, + 0.07038062065839767, + 0.126643568277359, + 0.0257983710616827 + ], + "credit_map": [ + 0.017816921696066856, + 0.02416186034679413, + 0.02367052622139454, + 0.013451305218040943, + 0.03246406838297844, + 0.024625549092888832, + 0.02410382218658924, + 0.02491316758096218, + 0.00999179482460022, + 0.011391004547476768, + 0.018461134284734726, + 0.014977207407355309, + 0.011408036574721336, + 0.015463560819625854, + 0.021667150780558586, + 0.0233379565179348, + 0.00747756427153945, + 0.005579615011811256, + 0.008695515803992748, + 0.00801011547446251, + 0.010072549805045128, + 0.036354538053274155, + 0.011541612446308136, + 0.01950872503221035, + 0.02617921493947506, + 0.022784601897001266, + 0.021949227899312973, + 0.013051619753241539, + 0.027018798515200615, + 0.0455198772251606, + 0.014072722755372524, + 0.03229639306664467, + 0.0066971308551728725, + 0.005320039112120867, + 0.006746404338628054, + 0.005321602802723646, + 0.007620951160788536, + 0.007642240263521671, + 0.007367212325334549, + 0.008311498910188675, + 0.00487808370962739, + 0.004517873749136925, + 0.009919037111103535, + 0.006779032293707132, + 0.006794323679059744, + 0.008556291460990906, + 0.006936539895832539, + 0.01000292506068945, + 0.00472424877807498, + 0.0027964944019913673, + 0.0057137529365718365, + 0.005655298940837383, + 0.0035488258581608534, + 0.004086999222636223, + 0.0055711581371724606, + 0.009651697240769863, + 0.004535861313343048, + 0.007848127745091915, + 0.01032730657607317, + 0.007260322105139494, + 0.01310458779335022, + 0.005348538048565388, + 0.01149151474237442, + 0.010464373975992203, + 0.02408166602253914, + 0.012102949433028698, + 0.021971991285681725, + 0.011190305463969707, + 0.010800565592944622, + 0.01770727150142193, + 0.010876353830099106, + 0.017912482842803, + 0.003563971258699894, + 0.0023845175746828318, + 0.002642204286530614, + 0.002333523938432336, + 0.00255381828173995, + 0.0056323581375181675, + 0.0027944203466176987, + 0.003893559332937002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0015119798481464386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031918552704155445 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031918552704155445, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031918552704155445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21023789048194885, + 0.11221513897180557, + 0.09041490405797958, + 0.07721084356307983, + 0.060727689415216446, + 0.10936073213815689, + 0.07168567925691605, + 0.03755684942007065, + 0.19517655670642853, + 0.035413771867752075 + ], + "credit_map": [ + 0.014476815238595009, + 0.030248034745454788, + 0.024061402305960655, + 0.019703606143593788, + 0.038573600351810455, + 0.024675622582435608, + 0.027812769636511803, + 0.03068603202700615, + 0.009157312102615833, + 0.010702398605644703, + 0.01749536208808422, + 0.009607164189219475, + 0.014256968162953854, + 0.02180315926671028, + 0.010751388035714626, + 0.01844138093292713, + 0.00841556116938591, + 0.005501292180269957, + 0.008119546808302402, + 0.006410968955606222, + 0.009153072722256184, + 0.027937917038798332, + 0.008892239071428776, + 0.015984313562512398, + 0.005536457058042288, + 0.004893584176898003, + 0.00920675229281187, + 0.0070344130508601665, + 0.009554633870720863, + 0.017775680869817734, + 0.010386698879301548, + 0.012822617776691914, + 0.008187856525182724, + 0.005955865606665611, + 0.010469851084053516, + 0.005690041929483414, + 0.007023521699011326, + 0.007958454079926014, + 0.005774195771664381, + 0.009667901322245598, + 0.011986189521849155, + 0.009978768415749073, + 0.015618966892361641, + 0.012496380135416985, + 0.012778296135365963, + 0.018695969134569168, + 0.011606041342020035, + 0.016200121492147446, + 0.003426765324547887, + 0.006134570110589266, + 0.005172774661332369, + 0.005650582257658243, + 0.007287789136171341, + 0.01631651446223259, + 0.006865216419100761, + 0.02083146944642067, + 0.0027869506739079952, + 0.00452909292653203, + 0.003802877850830555, + 0.004471926484256983, + 0.007785588037222624, + 0.003263851860538125, + 0.005273931659758091, + 0.005642629228532314, + 0.03571131452918053, + 0.01852530613541603, + 0.0332663357257843, + 0.016452137380838394, + 0.017556751146912575, + 0.02680571936070919, + 0.019479794427752495, + 0.027379194274544716, + 0.004951978102326393, + 0.003977237269282341, + 0.004530200269073248, + 0.0032962693367153406, + 0.00475679524242878, + 0.005044385325163603, + 0.003741380525752902, + 0.005115526728332043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0031918552704155445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot captures a teenage boy bicycling frantically along a narrow, rain-slicked sidewalk in a sprawling metropolis. The scene is rendered in a classic 1980s anime style, characterized by bold outlines and a stark, hand-drawn aesthetic. The teenager's dark hair is matted to his forehead by rain, his face a mask of determination, his teeth gnashed. He wears a bright yellow nylon windbreaker that gleams in the city lights, paired with dark, soaked jeans. Hunched low over the handlebars of his simple road bike, his legs a blur as he pedals relentlessly, kicking up spectacular splashes and capturing a riot of color. The alley is flanked by towering, dark buildings, their surfaces slick with rain. Vibrant neon signs in vibrant pink, cyan, and Japanese script cast long, distorted reflections across the massive puddle on the asphalt. As the camera keeps pace, it captures the dramatic anamorphic lens flare cast horizontally by the overhead streetlights, while a fine, grainy film texture overlays the image, completing the urgent and visually saturated retro atmosphere.", + "seed": 32829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_2/index_5998_seed_32829_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.38711395720019937, + "vlm_logprob": -0.38711395720019937, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.38711395720019937, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8051542043685913 + ], + "opposite_logprob_tokens": [ + -0.1801542192697525 + ], + "logprob_reward": -1.8051542043685913, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8051542043685913, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.43332090973854065, + 0.06962291896343231, + 0.04423472657799721, + 0.07248704135417938, + 0.04126296937465668, + 0.036087267100811005, + 0.09548605978488922, + 0.04620014876127243, + 0.05849938094615936, + 0.10279855877161026 + ], + "credit_map": [ + 0.0414140522480011, + 0.057505834847688675, + 0.04719031602144241, + 0.04755745828151703, + 0.06710734218358994, + 0.06830181926488876, + 0.05997413769364357, + 0.044270001351833344, + 0.005299037788063288, + 0.006292218808084726, + 0.014619509689509869, + 0.013335042633116245, + 0.006475289352238178, + 0.010435030795633793, + 0.006658957805484533, + 0.006507842801511288, + 0.002957976423203945, + 0.006233674474060535, + 0.00525991665199399, + 0.00373460934497416, + 0.006047664675861597, + 0.008014151826500893, + 0.007469673175364733, + 0.004517062101513147, + 0.005074168089777231, + 0.010433021001517773, + 0.011752903461456299, + 0.008731394074857235, + 0.010873670689761639, + 0.01264710258692503, + 0.007472951430827379, + 0.005501830484718084, + 0.004481397569179535, + 0.007473947945982218, + 0.005882165860384703, + 0.005192444194108248, + 0.0037831284571439028, + 0.0037728752940893173, + 0.006024997215718031, + 0.004652014933526516, + 0.0022510886192321777, + 0.00505455257371068, + 0.0037569187115877867, + 0.0033873070497065783, + 0.006012342870235443, + 0.002728776540607214, + 0.008074813522398472, + 0.004821473266929388, + 0.005875753238797188, + 0.015455481596291065, + 0.02026522345840931, + 0.009775213897228241, + 0.008569163270294666, + 0.015704941004514694, + 0.00964401289820671, + 0.010196279734373093, + 0.0029663247987627983, + 0.00401021633297205, + 0.009710858575999737, + 0.0035846165847033262, + 0.005094659980386496, + 0.008294424042105675, + 0.007231397088617086, + 0.005307655781507492, + 0.002026631496846676, + 0.00861061830073595, + 0.003584063844755292, + 0.004725551698356867, + 0.015337448567152023, + 0.00967101939022541, + 0.009420357644557953, + 0.005123695358633995, + 0.011969690211117268, + 0.015343599021434784, + 0.010284793563187122, + 0.011327644810080528, + 0.014680333435535431, + 0.015306707471609116, + 0.014119073748588562, + 0.009766722097992897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -1.8051542043685913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07014685124158859 + ], + "opposite_logprob_tokens": [ + -2.6951467990875244 + ], + "logprob_reward": -0.07014685124158859, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07014685124158859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44511663913726807, + 0.07136034220457077, + 0.03949183225631714, + 0.06109985336661339, + 0.04441800341010094, + 0.03383416309952736, + 0.1106070950627327, + 0.036514751613140106, + 0.04346387833356857, + 0.11409348249435425 + ], + "credit_map": [ + 0.04529630392789841, + 0.05944860726594925, + 0.048702437430620193, + 0.048479001969099045, + 0.06737039983272552, + 0.07064832746982574, + 0.0595494881272316, + 0.04562205821275711, + 0.006896488834172487, + 0.010547400452196598, + 0.017076056450605392, + 0.007143975235521793, + 0.006401540711522102, + 0.011442173272371292, + 0.006348771508783102, + 0.005503940396010876, + 0.003632616950199008, + 0.005997060798108578, + 0.004897490609437227, + 0.003385283285751939, + 0.004883201327174902, + 0.006616111844778061, + 0.006117792800068855, + 0.003962273709475994, + 0.0041243890300393105, + 0.008916361257433891, + 0.00993636529892683, + 0.006893759127706289, + 0.009093287400901318, + 0.010360266081988811, + 0.006693588569760323, + 0.0050818356685340405, + 0.005051689222455025, + 0.008607935160398483, + 0.006611802149564028, + 0.004991396330296993, + 0.004264679737389088, + 0.004530672915279865, + 0.006204614881426096, + 0.004155209753662348, + 0.0031764383893460035, + 0.0034864365588873625, + 0.006420902907848358, + 0.0030821841210126877, + 0.0038683502934873104, + 0.0037273159250617027, + 0.005645775236189365, + 0.004426759202033281, + 0.007319280412048101, + 0.019441908225417137, + 0.023309871554374695, + 0.011142055504024029, + 0.010937823913991451, + 0.017924126237630844, + 0.010021631605923176, + 0.010510393418371677, + 0.0022896751761436462, + 0.004928102251142263, + 0.004641929175704718, + 0.003009019186720252, + 0.006164849270135164, + 0.005949975922703743, + 0.00620686262845993, + 0.00332433870062232, + 0.001938301371410489, + 0.0072416081093251705, + 0.0032783891074359417, + 0.0035402688663452864, + 0.009802916087210178, + 0.0072596725076437, + 0.006567765958607197, + 0.0038349535316228867, + 0.012918521650135517, + 0.017526382580399513, + 0.010693049058318138, + 0.013763398863375187, + 0.016421575099229813, + 0.01666395179927349, + 0.015484359115362167, + 0.010622242465615273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.07014685124158859 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043075159192085266 + ], + "opposite_logprob_tokens": [ + -3.1680750846862793 + ], + "logprob_reward": -0.043075159192085266, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043075159192085266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42609384655952454, + 0.09009325504302979, + 0.07154545187950134, + 0.06465424597263336, + 0.03850306570529938, + 0.04720574617385864, + 0.07993581891059875, + 0.10814561694860458, + 0.021758241578936577, + 0.05206472426652908 + ], + "credit_map": [ + 0.046056609600782394, + 0.06063909828662872, + 0.05284936726093292, + 0.048419415950775146, + 0.05931396782398224, + 0.06710734218358994, + 0.052064355462789536, + 0.03964367136359215, + 0.00780839566141367, + 0.010319305583834648, + 0.02428346313536167, + 0.01254265196621418, + 0.005124960094690323, + 0.015525954775512218, + 0.007296620402485132, + 0.007191913202404976, + 0.006841697730123997, + 0.011657590977847576, + 0.012363777495920658, + 0.007332370150834322, + 0.005611028987914324, + 0.012624788098037243, + 0.008180410601198673, + 0.006933792028576136, + 0.008027824573218822, + 0.010156860575079918, + 0.011183029040694237, + 0.006349990144371986, + 0.00718627218157053, + 0.010668160393834114, + 0.006400474812835455, + 0.004681641235947609, + 0.004647661466151476, + 0.006509475409984589, + 0.005489933770149946, + 0.0031508863903582096, + 0.0039607807993888855, + 0.0048083048313856125, + 0.0048562586307525635, + 0.00507976533845067, + 0.003970383200794458, + 0.005854321178048849, + 0.008450446650385857, + 0.004393588751554489, + 0.005642974283546209, + 0.005809126887470484, + 0.006790221203118563, + 0.006294685415923595, + 0.005923905875533819, + 0.012417241930961609, + 0.014090280048549175, + 0.009164384566247463, + 0.007225827779620886, + 0.01261853612959385, + 0.008081058040261269, + 0.010414591990411282, + 0.005689500365406275, + 0.01355172973126173, + 0.00972212478518486, + 0.009444558061659336, + 0.01990453526377678, + 0.017138097435235977, + 0.02132296934723854, + 0.0113720977678895, + 0.002349389251321554, + 0.003247127402573824, + 0.00249205413274467, + 0.0026491025928407907, + 0.0035493785981088877, + 0.0031765263993293047, + 0.0019651358015835285, + 0.0023295267019420862, + 0.005626766011118889, + 0.006655882112681866, + 0.006840821355581284, + 0.005051096901297569, + 0.006570145953446627, + 0.008017243817448616, + 0.00789471622556448, + 0.005408052820712328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.043075159192085266 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645033273845911 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004645033273845911, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645033273845911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4468284547328949, + 0.0812024176120758, + 0.05754461884498596, + 0.060429658740758896, + 0.05263728275895119, + 0.04288030415773392, + 0.07787840813398361, + 0.04823824018239975, + 0.04504048451781273, + 0.08732013404369354 + ], + "credit_map": [ + 0.044840723276138306, + 0.0586886890232563, + 0.04973721131682396, + 0.049164220690727234, + 0.06583769619464874, + 0.07021689414978027, + 0.06250908970832825, + 0.04583393409848213, + 0.007497302256524563, + 0.00883096270263195, + 0.015081556513905525, + 0.014911781065165997, + 0.006186153274029493, + 0.012149212881922722, + 0.007404109463095665, + 0.009141342714428902, + 0.003977619111537933, + 0.009911958128213882, + 0.008392809890210629, + 0.007815234363079071, + 0.005435168277472258, + 0.009703736752271652, + 0.0059784273616969585, + 0.006329669617116451, + 0.0053151329047977924, + 0.00856638140976429, + 0.011291188187897205, + 0.006398988887667656, + 0.008072607219219208, + 0.010387900285422802, + 0.0060319481417536736, + 0.004365511238574982, + 0.0050172582268714905, + 0.01238533016294241, + 0.007028417661786079, + 0.006772138644009829, + 0.0035935642663389444, + 0.00467563746497035, + 0.007985588163137436, + 0.005179347470402718, + 0.0037695374339818954, + 0.004809502046555281, + 0.006757744122296572, + 0.00431419862434268, + 0.005115089472383261, + 0.0036955191753804684, + 0.0061245751567184925, + 0.008294142782688141, + 0.005464841146022081, + 0.014295432716608047, + 0.015252149663865566, + 0.007874196395277977, + 0.00710076792165637, + 0.012139840982854366, + 0.00821235403418541, + 0.007538834121078253, + 0.00310440338216722, + 0.00542886508628726, + 0.010042150504887104, + 0.0036864711437374353, + 0.005002608522772789, + 0.008210205473005772, + 0.007585025392472744, + 0.005178509280085564, + 0.001935861655510962, + 0.006973725277930498, + 0.0031610263977199793, + 0.003518229117617011, + 0.011768020689487457, + 0.00685854023322463, + 0.0065504563972353935, + 0.0042746285907924175, + 0.00919304694980383, + 0.012671629898250103, + 0.009284860454499722, + 0.009272350929677486, + 0.012848377227783203, + 0.012770299799740314, + 0.012559622526168823, + 0.00871995184570551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.004645033273845911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage boy bicycling frantically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a bright yellow nylon windbreaker and dark soaked jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow, rain-slicked sidewalk in a sprawling metropolis with towering dark buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vibrant neon signs in pink, cyan, and Japanese script visible casting reflections on a massive puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic 1980s anime style with bold outlines, a grainy film texture, and horizontal anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254853792488575 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.01254853792488575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254853792488575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4412336051464081, + 0.060403287410736084, + 0.046743281185626984, + 0.07087726891040802, + 0.05409028008580208, + 0.03587407246232033, + 0.07573065161705017, + 0.04717688634991646, + 0.05647465959191322, + 0.1113959401845932 + ], + "credit_map": [ + 0.0440189391374588, + 0.059320591390132904, + 0.0501907616853714, + 0.0490875169634819, + 0.06649572402238846, + 0.07033567130565643, + 0.05837942287325859, + 0.043404992669820786, + 0.005620282143354416, + 0.006989530753344297, + 0.015885449945926666, + 0.00582207040861249, + 0.0047700172290205956, + 0.01051320880651474, + 0.006231429986655712, + 0.004571294877678156, + 0.0035821148194372654, + 0.006889331620186567, + 0.007616857532411814, + 0.00450631370767951, + 0.004730553831905127, + 0.008160542696714401, + 0.006225615739822388, + 0.005031949374824762, + 0.005313123110681772, + 0.009553625248372555, + 0.012230725027620792, + 0.009418674744665623, + 0.009720515459775925, + 0.012093866243958473, + 0.006961869075894356, + 0.00558486906811595, + 0.004957933910191059, + 0.013083389960229397, + 0.0068615204654634, + 0.0065239169634878635, + 0.003976453561335802, + 0.004860015120357275, + 0.007999411784112453, + 0.005827638320624828, + 0.0037511729169636965, + 0.0037974275182932615, + 0.005790846422314644, + 0.003973305691033602, + 0.004019973799586296, + 0.0036592702381312847, + 0.006337114609777927, + 0.004544962663203478, + 0.004994107875972986, + 0.01401266548782587, + 0.01603058911859989, + 0.007227113004773855, + 0.007277771830558777, + 0.011782160960137844, + 0.007045719772577286, + 0.007360521238297224, + 0.0020677009597420692, + 0.006928774993866682, + 0.0056274994276463985, + 0.0029162163846194744, + 0.008318409323692322, + 0.00792020931839943, + 0.009196894243359566, + 0.00420118123292923, + 0.0026163870934396982, + 0.008753172121942043, + 0.0036730326246470213, + 0.00439341738820076, + 0.014081627130508423, + 0.009443740360438824, + 0.008459760807454586, + 0.005053522530943155, + 0.012958994135260582, + 0.017729496583342552, + 0.010705668479204178, + 0.013549750670790672, + 0.01588275656104088, + 0.016373245045542717, + 0.014220633544027805, + 0.009975395165383816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.01254853792488575 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_tensor([8720, 8720])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_tensor([8720, 8720])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..677b3703b87ea22bc6cc50f5f7e8cc59002d7678 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_tensor([8720, 8720])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a man with short, slightly tousled brown hair, sitting astride a black, ergonomic office chair on five wheels, as if steering an out-of-control race car. The camera captures his facial expression in close, ecstatic, and playful close-up: his eyes wide with excitement, his eyebrows raised dramatically, his mouth stretched into a wide, uncontrollable grin, his cheek muscles tensed with laughter. He wears a light blue Oxford shirt, his tie slouched to one side, his sleeves casually rolled up to his elbows. He leans forward, his hands gripping the armrests of his chair. He pushes off the ground with alternating force, propelling the chair in a comical sprint across the polished office floor. In the background, neat rows of office cubicles and the glass conference room walls blur in the background's rapid movement. The fluorescent lights on the ceiling cast even and bright light, highlighting every exaggerated muscle movement and the blush on his face caused by excitement. The whole scene was full of unconventional playfulness and pure joy, creating an infectious humorous atmosphere that suddenly broke out in the tense work.", + "seed": 10991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_8720_seed_10991_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05130310649983585, + "vlm_logprob": -0.05130310649983585, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.05130310649983585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01290870551019907 + ], + "opposite_logprob_tokens": [ + -4.887908935546875 + ], + "logprob_reward": -0.01290870551019907, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01290870551019907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08778524398803711, + 0.06873385608196259, + 0.06948105245828629, + 0.09068863838911057, + 0.05199892818927765, + 0.4825016260147095, + 0.045427024364471436, + 0.03210213780403137, + 0.026513168588280678, + 0.044768236577510834 + ], + "credit_map": [ + 0.009757138788700104, + 0.015391610562801361, + 0.011740239337086678, + 0.00446757348254323, + 0.010471721179783344, + 0.02048557810485363, + 0.007753178011626005, + 0.007718219887465239, + 0.012748130597174168, + 0.007535943761467934, + 0.008336950093507767, + 0.005623658653348684, + 0.006619848310947418, + 0.01190897636115551, + 0.007681035902351141, + 0.008279320783913136, + 0.0055394466035068035, + 0.007044112775474787, + 0.004223838914185762, + 0.013002187944948673, + 0.010163925588130951, + 0.007265256717801094, + 0.010902313515543938, + 0.011339979246258736, + 0.008547757752239704, + 0.0056782984174788, + 0.01309051550924778, + 0.009528725408017635, + 0.010165230371057987, + 0.011924748308956623, + 0.01589564047753811, + 0.015857728198170662, + 0.0062583922408521175, + 0.0038848177064210176, + 0.009426436387002468, + 0.005269782617688179, + 0.007136615924537182, + 0.006232581567019224, + 0.006924963556230068, + 0.006865347735583782, + 0.03293873742222786, + 0.181090846657753, + 0.01826949045062065, + 0.04958382248878479, + 0.026430336758494377, + 0.016985861584544182, + 0.02150532230734825, + 0.13569730520248413, + 0.004437901545315981, + 0.00733577273786068, + 0.006160958670079708, + 0.003709269454702735, + 0.002201345982030034, + 0.003792309667915106, + 0.0046840268187224865, + 0.013105446472764015, + 0.006002115551382303, + 0.0042113917879760265, + 0.004309429787099361, + 0.0022427267394959927, + 0.004829565528780222, + 0.004183520097285509, + 0.003628013888373971, + 0.002695380011573434, + 0.0021797369699925184, + 0.0042418185621500015, + 0.0037307750899344683, + 0.002324372762814164, + 0.0030014675576239824, + 0.004343559965491295, + 0.004546781536191702, + 0.002144661033526063, + 0.007101813331246376, + 0.0055978139862418175, + 0.005282436031848192, + 0.004232659935951233, + 0.00547462934628129, + 0.005003357771784067, + 0.004968591500073671, + 0.00710693746805191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.01290870551019907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035905987024307 + ], + "opposite_logprob_tokens": [ + -2.3503589630126953 + ], + "logprob_reward": -0.10035905987024307, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035905987024307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15862376987934113, + 0.10513964295387268, + 0.06518081575632095, + 0.10804944485425949, + 0.06910158693790436, + 0.2617257535457611, + 0.09057779610157013, + 0.05212811380624771, + 0.02794477716088295, + 0.0615282766520977 + ], + "credit_map": [ + 0.021729106083512306, + 0.02164601720869541, + 0.02008305862545967, + 0.010283431969583035, + 0.021093379706144333, + 0.03540967032313347, + 0.014855241402983665, + 0.013523858971893787, + 0.02624492347240448, + 0.013258361257612705, + 0.017109308391809464, + 0.008072247728705406, + 0.009008921682834625, + 0.012229381129145622, + 0.010872391983866692, + 0.008344113826751709, + 0.004911718890070915, + 0.0072153834626078606, + 0.003952170256525278, + 0.010399406775832176, + 0.009881737641990185, + 0.007910988293588161, + 0.009536318480968475, + 0.01137309055775404, + 0.010797310620546341, + 0.006945276632905006, + 0.016411131247878075, + 0.011884406208992004, + 0.010542232543230057, + 0.01679125241935253, + 0.014779925346374512, + 0.019897906109690666, + 0.0137136559933424, + 0.009497452527284622, + 0.011239051818847656, + 0.004056261852383614, + 0.007783364970237017, + 0.009191381745040417, + 0.005986957810819149, + 0.0076334611512720585, + 0.01743137277662754, + 0.09686005860567093, + 0.011272904463112354, + 0.02891828864812851, + 0.014767328277230263, + 0.009807376191020012, + 0.013038803823292255, + 0.06962961703538895, + 0.007923571392893791, + 0.010023456998169422, + 0.01902317814528942, + 0.011780941858887672, + 0.0049236430786550045, + 0.008524163626134396, + 0.009458193555474281, + 0.018920646980404854, + 0.011050752364099026, + 0.007858756929636002, + 0.0056242626160383224, + 0.003944635856896639, + 0.007902037352323532, + 0.0067632789723575115, + 0.005681073293089867, + 0.003303318051621318, + 0.0028346360195428133, + 0.004619235638529062, + 0.0028775574173778296, + 0.0023777694441378117, + 0.003622958669438958, + 0.002807335229590535, + 0.00595359830185771, + 0.002851686906069517, + 0.008415443822741508, + 0.011456004343926907, + 0.006848408840596676, + 0.00695525947958231, + 0.007866994477808475, + 0.007125667296350002, + 0.0048844837583601475, + 0.00797601044178009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.10035905987024307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709856033325195 + ], + "opposite_logprob_tokens": [ + -2.127098560333252 + ], + "logprob_reward": -0.12709856033325195, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709856033325195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11321475356817245, + 0.0662592351436615, + 0.12543541193008423, + 0.15028174221515656, + 0.030732518061995506, + 0.2855255603790283, + 0.05208108201622963, + 0.08024294674396515, + 0.049417056143283844, + 0.04680972173810005 + ], + "credit_map": [ + 0.011057251133024693, + 0.011623264290392399, + 0.016253912821412086, + 0.00701537262648344, + 0.022220658138394356, + 0.023680929094552994, + 0.01159360446035862, + 0.009769764728844166, + 0.01309644989669323, + 0.007866235449910164, + 0.0074764941819012165, + 0.004323306959122419, + 0.00994872860610485, + 0.008299371227622032, + 0.008572794497013092, + 0.006675849203020334, + 0.00772453099489212, + 0.01331328134983778, + 0.007306527346372604, + 0.017575077712535858, + 0.020032215863466263, + 0.01425565592944622, + 0.019916726276278496, + 0.02531140297651291, + 0.012009351514279842, + 0.008922573179006577, + 0.015796935185790062, + 0.01583791896700859, + 0.01902477629482746, + 0.017791811376810074, + 0.03138703852891922, + 0.029511326923966408, + 0.0036267454270273447, + 0.00378330796957016, + 0.0038742253091186285, + 0.0024989796802401543, + 0.004168831277638674, + 0.004107228480279446, + 0.004738060757517815, + 0.003935140557587147, + 0.019178563728928566, + 0.1076526939868927, + 0.01204688847064972, + 0.03166097775101662, + 0.01710316352546215, + 0.009381943382322788, + 0.012782188132405281, + 0.07571914792060852, + 0.005068120080977678, + 0.006139910779893398, + 0.00798292737454176, + 0.0039040104020386934, + 0.002729725791141391, + 0.005585621111094952, + 0.007431371603161097, + 0.013239393942058086, + 0.01199085358530283, + 0.01029733382165432, + 0.006417963653802872, + 0.007653096225112677, + 0.01380185503512621, + 0.011195420287549496, + 0.011570611968636513, + 0.007315817754715681, + 0.004558449145406485, + 0.006322970148175955, + 0.0045652734115719795, + 0.0055630323477089405, + 0.007582451682537794, + 0.0056824879720807076, + 0.008963518775999546, + 0.006178874988108873, + 0.009666845202445984, + 0.007273444440215826, + 0.0039070225320756435, + 0.0032405522651970387, + 0.0046746558509767056, + 0.005131253506988287, + 0.004249660763889551, + 0.008666285313665867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.12709856033325195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614442190155387 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.003614442190155387, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614442190155387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06599318981170654, + 0.033070847392082214, + 0.032199110835790634, + 0.04464564099907875, + 0.06487993896007538, + 0.5221764445304871, + 0.09603211283683777, + 0.09762360900640488, + 0.010240546427667141, + 0.03313855454325676 + ], + "credit_map": [ + 0.012061678804457188, + 0.0070887054316699505, + 0.014404316432774067, + 0.003422264941036701, + 0.007554433308541775, + 0.011248227208852768, + 0.005616615526378155, + 0.004596950486302376, + 0.00788014754652977, + 0.003533514216542244, + 0.0052525936625897884, + 0.002695139730349183, + 0.0027625691145658493, + 0.003273541107773781, + 0.004030817653983831, + 0.0036425236612558365, + 0.0024749727454036474, + 0.0029952472541481256, + 0.0015814376529306173, + 0.004674105439335108, + 0.004177658353000879, + 0.0036451220512390137, + 0.005723655689507723, + 0.006926910486072302, + 0.004789651837199926, + 0.002938398625701666, + 0.005373271647840738, + 0.004105378873646259, + 0.004905798006802797, + 0.008879211731255054, + 0.00557359866797924, + 0.008080334402620792, + 0.016903560608625412, + 0.01383756473660469, + 0.010428096167743206, + 0.004109843168407679, + 0.004953807219862938, + 0.006568103563040495, + 0.004540873691439629, + 0.003538089105859399, + 0.03612980991601944, + 0.19459085166454315, + 0.017959753051400185, + 0.05265427753329277, + 0.028538592159748077, + 0.01800733245909214, + 0.023905588313937187, + 0.15039023756980896, + 0.008698798716068268, + 0.01086043007671833, + 0.016585351899266243, + 0.015623876824975014, + 0.005400874186307192, + 0.007491766475141048, + 0.00958241242915392, + 0.02178860642015934, + 0.018762845546007156, + 0.02527426928281784, + 0.010242486372590065, + 0.007100805640220642, + 0.007946939207613468, + 0.01543086301535368, + 0.005541557911783457, + 0.007323842030018568, + 0.0010160881793126464, + 0.0019136855844408274, + 0.001065755495801568, + 0.001049780286848545, + 0.0014527859166264534, + 0.0012028596829622984, + 0.0013507329858839512, + 0.0011888580629602075, + 0.005090377293527126, + 0.008634927682578564, + 0.0033478247933089733, + 0.0035416774917393923, + 0.004201306030154228, + 0.002826903946697712, + 0.0024202754721045494, + 0.003075260668992996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.003614442190155387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534764595329762 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534764595329762, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534764595329762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08007746934890747, + 0.06443117558956146, + 0.026830291375517845, + 0.046672556549310684, + 0.023479582741856575, + 0.6051976084709167, + 0.05896356701850891, + 0.033129457384347916, + 0.032282814383506775, + 0.028935441747307777 + ], + "credit_map": [ + 0.010769860818982124, + 0.010734348557889462, + 0.00987374410033226, + 0.004838686436414719, + 0.011271861381828785, + 0.019071632996201515, + 0.006814995314925909, + 0.006702349055558443, + 0.015802033245563507, + 0.00787342619150877, + 0.008629848249256611, + 0.004727205727249384, + 0.0054857064969837666, + 0.010377907194197178, + 0.0064310599118471146, + 0.005103993695229292, + 0.00195829919539392, + 0.0031291802879422903, + 0.0015631415881216526, + 0.0029305669013410807, + 0.005537078715860844, + 0.003323412500321865, + 0.004412376321852207, + 0.003976238425821066, + 0.005103864707052708, + 0.003329614410176873, + 0.0062233987264335155, + 0.004388549830764532, + 0.0060133980587124825, + 0.007892186753451824, + 0.006028786301612854, + 0.0076927621848881245, + 0.004170166794210672, + 0.0023719065356999636, + 0.0033451321069151163, + 0.002227179706096649, + 0.0035055382177233696, + 0.0026847952976822853, + 0.0028235730715095997, + 0.002351292409002781, + 0.041201140731573105, + 0.22414617240428925, + 0.02523413673043251, + 0.0666949525475502, + 0.034677110612392426, + 0.02274031564593315, + 0.0291952695697546, + 0.16130855679512024, + 0.004155504517257214, + 0.005111780948936939, + 0.01559442188590765, + 0.005215040873736143, + 0.002214935841038823, + 0.004645221866667271, + 0.005259981844574213, + 0.016766682267189026, + 0.004271896556019783, + 0.0070087187923491, + 0.00333891692571342, + 0.0022389087826013565, + 0.00527968117967248, + 0.004848427139222622, + 0.003534771502017975, + 0.0026081399992108345, + 0.0030636966694146395, + 0.005857956595718861, + 0.0030243771616369486, + 0.003361603245139122, + 0.005401833448559046, + 0.00351674179546535, + 0.005452015437185764, + 0.0026045937556773424, + 0.004701168276369572, + 0.00745558924973011, + 0.0030375064816325903, + 0.0023273457773029804, + 0.003740347223356366, + 0.00287875859066844, + 0.00222528213635087, + 0.0025694468058645725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012534764595329762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a man with short, slightly tousled brown hair, sitting astride a black, ergonomic office chair on five wheels, as if steering an out-of-control race car. The camera captures his facial expression in close, ecstatic, and playful close-up: his eyes wide with excitement, his eyebrows raised dramatically, his mouth stretched into a wide, uncontrollable grin, his cheek muscles tensed with laughter. He wears a light blue Oxford shirt, his tie slouched to one side, his sleeves casually rolled up to his elbows. He leans forward, his hands gripping the armrests of his chair. He pushes off the ground with alternating force, propelling the chair in a comical sprint across the polished office floor. In the background, neat rows of office cubicles and the glass conference room walls blur in the background's rapid movement. The fluorescent lights on the ceiling cast even and bright light, highlighting every exaggerated muscle movement and the blush on his face caused by excitement. The whole scene was full of unconventional playfulness and pure joy, creating an infectious humorous atmosphere that suddenly broke out in the tense work.", + "seed": 10992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_8720_seed_10992_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017873570160008966, + "vlm_logprob": -0.017873570160008966, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.017873570160008966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009936983697116375 + ], + "opposite_logprob_tokens": [ + -5.759936809539795 + ], + "logprob_reward": -0.009936983697116375, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009936983697116375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144977405667305, + 0.12489300221204758, + 0.03874490410089493, + 0.07150329649448395, + 0.07543575763702393, + 0.03472329303622246, + 0.05081222578883171, + 0.23520177602767944, + 0.07322775572538376, + 0.15048053860664368 + ], + "credit_map": [ + 0.023182207718491554, + 0.01912762224674225, + 0.013883158564567566, + 0.01644551381468773, + 0.015004128217697144, + 0.02112787961959839, + 0.013721898198127747, + 0.022484999150037766, + 0.022331178188323975, + 0.029607178643345833, + 0.02496231347322464, + 0.012009207159280777, + 0.011187988333404064, + 0.009838330559432507, + 0.007612958084791899, + 0.007343851029872894, + 0.0051338342018425465, + 0.004613227676600218, + 0.007595882751047611, + 0.005327526014298201, + 0.003357494017109275, + 0.0036488575860857964, + 0.004465381149202585, + 0.0046027000062167645, + 0.011392510496079922, + 0.010850640945136547, + 0.010069120675325394, + 0.01039096713066101, + 0.008134594187140465, + 0.009159508161246777, + 0.005975373554974794, + 0.005530585069209337, + 0.007656434550881386, + 0.0069357845932245255, + 0.010615801438689232, + 0.007877633906900883, + 0.008719260804355145, + 0.012313161045312881, + 0.013223562389612198, + 0.008094116114079952, + 0.003230373840779066, + 0.0039024564903229475, + 0.005953249055892229, + 0.00630953349173069, + 0.0038449601270258427, + 0.00293015711940825, + 0.004575327504426241, + 0.0039772335439920425, + 0.005681391805410385, + 0.00824782159179449, + 0.009466308169066906, + 0.008845893666148186, + 0.00436171842738986, + 0.005129294935613871, + 0.004272486083209515, + 0.0048073092475533485, + 0.02369629219174385, + 0.018669821321964264, + 0.018584273755550385, + 0.019458891823887825, + 0.022158876061439514, + 0.01606408879160881, + 0.06469827890396118, + 0.051871273666620255, + 0.008751176297664642, + 0.008310792036354542, + 0.011150307022035122, + 0.012331563048064709, + 0.006169331260025501, + 0.008195143193006516, + 0.006583888083696365, + 0.011735553853213787, + 0.011707250028848648, + 0.012152559123933315, + 0.051503363996744156, + 0.008949379436671734, + 0.00880392175167799, + 0.010298710316419601, + 0.010522949509322643, + 0.03654240071773529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.009936983697116375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813149780035019 + ], + "opposite_logprob_tokens": [ + -3.2881314754486084 + ], + "logprob_reward": -0.03813149780035019, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813149780035019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1779678612947464, + 0.0834093689918518, + 0.037885624915361404, + 0.23286646604537964, + 0.10604432225227356, + 0.04627576097846031, + 0.03395838662981987, + 0.10730307549238205, + 0.09303813427686691, + 0.08125100284814835 + ], + "credit_map": [ + 0.03355860710144043, + 0.02080599218606949, + 0.029742207378149033, + 0.01823711395263672, + 0.017802368849515915, + 0.01892011985182762, + 0.019160594791173935, + 0.019740847870707512, + 0.013858255930244923, + 0.01571459136903286, + 0.009799305349588394, + 0.007507422938942909, + 0.01086373534053564, + 0.01396486721932888, + 0.005866123363375664, + 0.005835068877786398, + 0.003727406496182084, + 0.00416297884657979, + 0.0069144065491855145, + 0.0047432007268071175, + 0.003994997590780258, + 0.003107507014647126, + 0.005550520494580269, + 0.005684605799615383, + 0.04476598650217056, + 0.02974748983979225, + 0.048338551074266434, + 0.04036479815840721, + 0.02480853535234928, + 0.017384294420480728, + 0.014084913767874241, + 0.013371885754168034, + 0.01392690185457468, + 0.012962467968463898, + 0.0152647215873003, + 0.014603789895772934, + 0.01670140027999878, + 0.010046908631920815, + 0.015361140482127666, + 0.007176988758146763, + 0.0036339156795293093, + 0.004597327671945095, + 0.012758187018334866, + 0.00863412395119667, + 0.005475251004099846, + 0.003061925293877721, + 0.005454733967781067, + 0.0026602912694215775, + 0.0039669121615588665, + 0.00472244992852211, + 0.0053682331927120686, + 0.004317909013479948, + 0.003796306671574712, + 0.004697246011346579, + 0.003676500404253602, + 0.0034128243569284678, + 0.01111050508916378, + 0.009368675760924816, + 0.015667608007788658, + 0.0137870442122221, + 0.009005190804600716, + 0.007483041379600763, + 0.027601664885878563, + 0.013279343023896217, + 0.004599608480930328, + 0.006344444118440151, + 0.027634816244244576, + 0.010974842123687267, + 0.006413040217012167, + 0.006356610916554928, + 0.0063232071697711945, + 0.024391567334532738, + 0.005933527834713459, + 0.005862347315996885, + 0.02431606687605381, + 0.006363534834235907, + 0.004865713883191347, + 0.007007082458585501, + 0.0060160341672599316, + 0.020886698737740517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.03813149780035019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023295527324080467 + ], + "opposite_logprob_tokens": [ + -3.7732956409454346 + ], + "logprob_reward": -0.023295527324080467, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295527324080467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14869043231010437, + 0.05960076302289963, + 0.034170880913734436, + 0.12923826277256012, + 0.07440870255231857, + 0.03412129357457161, + 0.026149630546569824, + 0.15540888905525208, + 0.18557070195674896, + 0.1526404172182083 + ], + "credit_map": [ + 0.02643582969903946, + 0.017208337783813477, + 0.016369342803955078, + 0.014047893695533276, + 0.021665744483470917, + 0.025138335302472115, + 0.013845132663846016, + 0.01397982332855463, + 0.011975042521953583, + 0.014070909470319748, + 0.007712597958743572, + 0.005433165933936834, + 0.005687435623258352, + 0.005641527473926544, + 0.003976819105446339, + 0.00510327098891139, + 0.003990476485341787, + 0.0039007263258099556, + 0.003234714502468705, + 0.004750451073050499, + 0.0049609593115746975, + 0.004966282285749912, + 0.0038955959025770426, + 0.004471678286790848, + 0.03190615400671959, + 0.020425016060471535, + 0.008119476959109306, + 0.013489949516952038, + 0.022578485310077667, + 0.01403195969760418, + 0.010759140364825726, + 0.007928079925477505, + 0.011093656532466412, + 0.006914650555700064, + 0.00938613060861826, + 0.00959349051117897, + 0.011476024053990841, + 0.00987730361521244, + 0.010133597999811172, + 0.005933852866292, + 0.00310837896540761, + 0.0037576875183731318, + 0.004056334495544434, + 0.003933428321033716, + 0.007241502869874239, + 0.0033604453783482313, + 0.004885347094386816, + 0.0037781731225550175, + 0.0027887243777513504, + 0.0046627214178442955, + 0.003424653084948659, + 0.004659587983042002, + 0.002405922394245863, + 0.0034418711438775063, + 0.0027289700228720903, + 0.0020371826831251383, + 0.01654481701552868, + 0.01880069449543953, + 0.018868595361709595, + 0.014370162971317768, + 0.015243303962051868, + 0.013292280957102776, + 0.032743487507104874, + 0.02554556168615818, + 0.01848737709224224, + 0.02656942419707775, + 0.02923627570271492, + 0.023557117208838463, + 0.02455124631524086, + 0.018416408449411392, + 0.01638234592974186, + 0.028370510786771774, + 0.0102564487606287, + 0.008081912063062191, + 0.05090278014540672, + 0.008374227210879326, + 0.01015668734908104, + 0.01184793934226036, + 0.00904830638319254, + 0.04397213086485863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.023295527324080467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940988120622933 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.001940988120622933, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940988120622933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06387249380350113, + 0.07647460699081421, + 0.02793658897280693, + 0.09731430560350418, + 0.16946840286254883, + 0.023242777213454247, + 0.13147281110286713, + 0.3356623351573944, + 0.05619581416249275, + 0.018359903246164322 + ], + "credit_map": [ + 0.013098864816129208, + 0.007411513011902571, + 0.009498865343630314, + 0.006974073126912117, + 0.006325289141386747, + 0.008519662544131279, + 0.005804126150906086, + 0.00624009920284152, + 0.014976057223975658, + 0.015596960671246052, + 0.017917055636644363, + 0.009908433072268963, + 0.005533682648092508, + 0.004107722546905279, + 0.004415468312799931, + 0.004019227344542742, + 0.0031870573293417692, + 0.0030938873533159494, + 0.006918524391949177, + 0.0036150605883449316, + 0.002400363562628627, + 0.002788369543850422, + 0.0030378263909369707, + 0.0028955007437616587, + 0.023614883422851562, + 0.015948830172419548, + 0.012692536227405071, + 0.014717831276357174, + 0.012139580212533474, + 0.007332777604460716, + 0.005472850054502487, + 0.005395012442022562, + 0.02666066773235798, + 0.016748182475566864, + 0.029278593137860298, + 0.021794980391860008, + 0.02598174847662449, + 0.018862832337617874, + 0.01834646612405777, + 0.011794939637184143, + 0.0025291857309639454, + 0.00426019262522459, + 0.0034168334677815437, + 0.003613813314586878, + 0.0032977210357785225, + 0.0022509333211928606, + 0.002421918325126171, + 0.0014521775301545858, + 0.01757550612092018, + 0.01652207039296627, + 0.028209656476974487, + 0.015017323195934296, + 0.01117374561727047, + 0.01310172863304615, + 0.01635454036295414, + 0.013518235646188259, + 0.040343862026929855, + 0.02920287847518921, + 0.03210398927330971, + 0.03401955962181091, + 0.031262222677469254, + 0.02691977471113205, + 0.0737474337220192, + 0.06806258857250214, + 0.0033966340124607086, + 0.0031939633190631866, + 0.01918691024184227, + 0.005682225804775953, + 0.0028081098571419716, + 0.00414943415671587, + 0.00298411981202662, + 0.01479441486299038, + 0.002000779379159212, + 0.0017676533898338675, + 0.0022195789497345686, + 0.0028034516144543886, + 0.0027209350373595953, + 0.0018519354052841663, + 0.003201541258022189, + 0.001794026349671185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.001940988120622933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606285385787487 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.01606285385787487, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606285385787487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1824590414762497, + 0.05514562129974365, + 0.04897462576627731, + 0.1509069800376892, + 0.0940900444984436, + 0.04716482013463974, + 0.0471678301692009, + 0.08872871845960617, + 0.06206340715289116, + 0.22329896688461304 + ], + "credit_map": [ + 0.03348449245095253, + 0.025914520025253296, + 0.021956833079457283, + 0.02469562739133835, + 0.016867443919181824, + 0.02634451352059841, + 0.01510001253336668, + 0.01809559389948845, + 0.00685367826372385, + 0.009005160070955753, + 0.010095516219735146, + 0.006605071946978569, + 0.007332020904868841, + 0.005523872096091509, + 0.004809188190847635, + 0.0049211117438972, + 0.005791032686829567, + 0.005898750852793455, + 0.012179081328213215, + 0.005855378694832325, + 0.004514473490417004, + 0.00471864640712738, + 0.00422319583594799, + 0.005794066935777664, + 0.030501583591103554, + 0.018986595794558525, + 0.028573492541909218, + 0.024719733744859695, + 0.018571432679891586, + 0.010844103991985321, + 0.009678923524916172, + 0.009031112305819988, + 0.013556965626776218, + 0.01484528835862875, + 0.009127776138484478, + 0.01425257883965969, + 0.009523107670247555, + 0.01397272851318121, + 0.011687575839459896, + 0.0071240137331187725, + 0.003759636078029871, + 0.004605186637490988, + 0.008276118896901608, + 0.009529097005724907, + 0.00605526939034462, + 0.004226689226925373, + 0.005594260524958372, + 0.005118558183312416, + 0.006283159833401442, + 0.006251802202314138, + 0.010144556872546673, + 0.005134629551321268, + 0.005675852298736572, + 0.004038789309561253, + 0.004938540048897266, + 0.004700496327131987, + 0.010913826525211334, + 0.007614564150571823, + 0.01829112507402897, + 0.010168993845582008, + 0.007136814761906862, + 0.007783115841448307, + 0.01910501904785633, + 0.007715259678661823, + 0.00468186242505908, + 0.006011929828673601, + 0.013298220001161098, + 0.007638161536306143, + 0.007546633947640657, + 0.005127459764480591, + 0.00519452802836895, + 0.012564617209136486, + 0.016608217731118202, + 0.012842514552175999, + 0.07488882541656494, + 0.01105580199509859, + 0.011796030215919018, + 0.01706152968108654, + 0.013369318097829819, + 0.0656767189502716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.01606285385787487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a man with short, slightly tousled brown hair, sitting astride a black, ergonomic office chair on five wheels, as if steering an out-of-control race car. The camera captures his facial expression in close, ecstatic, and playful close-up: his eyes wide with excitement, his eyebrows raised dramatically, his mouth stretched into a wide, uncontrollable grin, his cheek muscles tensed with laughter. He wears a light blue Oxford shirt, his tie slouched to one side, his sleeves casually rolled up to his elbows. He leans forward, his hands gripping the armrests of his chair. He pushes off the ground with alternating force, propelling the chair in a comical sprint across the polished office floor. In the background, neat rows of office cubicles and the glass conference room walls blur in the background's rapid movement. The fluorescent lights on the ceiling cast even and bright light, highlighting every exaggerated muscle movement and the blush on his face caused by excitement. The whole scene was full of unconventional playfulness and pure joy, creating an infectious humorous atmosphere that suddenly broke out in the tense work.", + "seed": 10993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_8720_seed_10993_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018607903504744172, + "vlm_logprob": -0.018607903504744172, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.018607903504744172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010612716898322105 + ], + "opposite_logprob_tokens": [ + -5.385612487792969 + ], + "logprob_reward": -0.010612716898322105, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010612716898322105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15757887065410614, + 0.12062054872512817, + 0.05995316803455353, + 0.10205584764480591, + 0.04071221500635147, + 0.07206190377473831, + 0.15318743884563446, + 0.12858738005161285, + 0.07519250363111496, + 0.09005014598369598 + ], + "credit_map": [ + 0.04355940595269203, + 0.02437332645058632, + 0.01724935695528984, + 0.016167020425200462, + 0.020698903128504753, + 0.012399435043334961, + 0.013284793123602867, + 0.009846627712249756, + 0.017630400136113167, + 0.021943997591733932, + 0.007664045784622431, + 0.009111519902944565, + 0.009943057782948017, + 0.04136400297284126, + 0.008442017249763012, + 0.004521514754742384, + 0.009576498530805111, + 0.0057041640393435955, + 0.007863231003284454, + 0.009678064845502377, + 0.007098923437297344, + 0.0060433740727603436, + 0.010261540301144123, + 0.0037273727357387543, + 0.007933095097541809, + 0.014412487857043743, + 0.008823541924357414, + 0.00982134509831667, + 0.021865127608180046, + 0.019119204953312874, + 0.011429752223193645, + 0.008651289157569408, + 0.007154037710279226, + 0.005853036884218454, + 0.004053526557981968, + 0.004048585891723633, + 0.007524904329329729, + 0.004566280636936426, + 0.00434294855222106, + 0.003168895374983549, + 0.009874832816421986, + 0.009919331409037113, + 0.009220040403306484, + 0.008094522170722485, + 0.010544047690927982, + 0.008968242444097996, + 0.008063049055635929, + 0.007377838250249624, + 0.010057286359369755, + 0.01598576083779335, + 0.0166703462600708, + 0.010204924270510674, + 0.05572936311364174, + 0.014558259397745132, + 0.020888973027467728, + 0.009092530235648155, + 0.019812749698758125, + 0.023502230644226074, + 0.023762214928865433, + 0.009254831820726395, + 0.009258223697543144, + 0.02570665255188942, + 0.008180350065231323, + 0.009110123850405216, + 0.012521752156317234, + 0.015948578715324402, + 0.008658742532134056, + 0.006360698025673628, + 0.010034133680164814, + 0.005103146191686392, + 0.009045051410794258, + 0.00752039672806859, + 0.015709849074482918, + 0.011050126515328884, + 0.011222658678889275, + 0.004786380100995302, + 0.008387796580791473, + 0.021808525547385216, + 0.007829993031919003, + 0.0092548169195652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.010612716898322105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373117744922638 + ], + "opposite_logprob_tokens": [ + -3.40873122215271 + ], + "logprob_reward": -0.03373117744922638, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373117744922638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09295832365751266, + 0.05249285697937012, + 0.04859049990773201, + 0.04012574255466461, + 0.051554180681705475, + 0.05038178339600563, + 0.057820096611976624, + 0.5610279440879822, + 0.021957479417324066, + 0.023091137409210205 + ], + "credit_map": [ + 0.014382678084075451, + 0.012963689863681793, + 0.010158094577491283, + 0.011880872771143913, + 0.014262635260820389, + 0.00939920637756586, + 0.011870335787534714, + 0.00804081466048956, + 0.010954148136079311, + 0.01250244490802288, + 0.003556656651198864, + 0.005014346446841955, + 0.005640110000967979, + 0.006203849334269762, + 0.005294084548950195, + 0.003327218582853675, + 0.008221054449677467, + 0.005713149439543486, + 0.007428632117807865, + 0.007762295659631491, + 0.0039043880533427, + 0.005166706163436174, + 0.006794764194637537, + 0.0035995077341794968, + 0.0037934198044240475, + 0.007199503947049379, + 0.003953552804887295, + 0.0036251910496503115, + 0.008971825242042542, + 0.004405617248266935, + 0.004793958738446236, + 0.003382673254236579, + 0.00925011932849884, + 0.0071485755033791065, + 0.0060384850949049, + 0.005606047809123993, + 0.006571915000677109, + 0.0062871286645531654, + 0.0063095674850046635, + 0.004342339001595974, + 0.00492874626070261, + 0.011288175359368324, + 0.004970645997673273, + 0.00555346067994833, + 0.007488240022212267, + 0.006002143956720829, + 0.0056810881942510605, + 0.0044692805968225, + 0.003415813436731696, + 0.005090159829705954, + 0.0088660204783082, + 0.004195305984467268, + 0.018833838403224945, + 0.005838330835103989, + 0.005839990917593241, + 0.005740640684962273, + 0.07675448060035706, + 0.09399939328432083, + 0.12608523666858673, + 0.03865586221218109, + 0.040730610489845276, + 0.11160953342914581, + 0.03177889436483383, + 0.041413936764001846, + 0.0035086439456790686, + 0.002295895479619503, + 0.003964169416576624, + 0.002092997543513775, + 0.0033391823526471853, + 0.0022129653953015804, + 0.002353392308577895, + 0.0021902332082390785, + 0.004269617144018412, + 0.002576389815658331, + 0.0030230125412344933, + 0.0023499378003180027, + 0.0024311465676873922, + 0.0032588017638772726, + 0.002343331463634968, + 0.002838901476934552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.03373117744922638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029817529022693634 + ], + "opposite_logprob_tokens": [ + -3.529817581176758 + ], + "logprob_reward": -0.029817529022693634, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029817529022693634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11839582026004791, + 0.1277351826429367, + 0.12594862282276154, + 0.06573159992694855, + 0.19195187091827393, + 0.18055360019207, + 0.07161286473274231, + 0.033971045166254044, + 0.04612228274345398, + 0.03797703608870506 + ], + "credit_map": [ + 0.028117679059505463, + 0.014431686140596867, + 0.012879718095064163, + 0.012957734987139702, + 0.01613483391702175, + 0.011996098794043064, + 0.013004804030060768, + 0.008873279206454754, + 0.018614748492836952, + 0.02534301206469536, + 0.008609462529420853, + 0.01289722602814436, + 0.011285080574452877, + 0.03541485220193863, + 0.010085796006023884, + 0.005485015921294689, + 0.014534501358866692, + 0.011715586297214031, + 0.016861271113157272, + 0.01292209979146719, + 0.028972292318940163, + 0.019837383180856705, + 0.013057040050625801, + 0.00804846454411745, + 0.008075153455138206, + 0.011643294245004654, + 0.005359870381653309, + 0.0038964946288615465, + 0.010105916298925877, + 0.014603860676288605, + 0.006602528505027294, + 0.005444482434540987, + 0.03571246936917305, + 0.024474140256643295, + 0.022927014157176018, + 0.021037720143795013, + 0.026250410825014114, + 0.024073641747236252, + 0.022540973499417305, + 0.0149355074390769, + 0.02091812714934349, + 0.026548253372311592, + 0.026515478268265724, + 0.01994304172694683, + 0.03374238684773445, + 0.01812782883644104, + 0.019759729504585266, + 0.014998773112893105, + 0.008131048642098904, + 0.007670823018997908, + 0.010267077945172787, + 0.0067907157354056835, + 0.010732647031545639, + 0.01126400101929903, + 0.008053109049797058, + 0.008703446947038174, + 0.006354105658829212, + 0.003187676193192601, + 0.008806176483631134, + 0.00275219464674592, + 0.003622674383223057, + 0.004595200996845961, + 0.002481559757143259, + 0.002171461470425129, + 0.006102561950683594, + 0.008903623558580875, + 0.005982580594718456, + 0.004680249374359846, + 0.006784458179026842, + 0.003605830715969205, + 0.005429472774267197, + 0.004633511416614056, + 0.005573404021561146, + 0.005183623172342777, + 0.003703699680045247, + 0.0022656877990812063, + 0.005002395249903202, + 0.009777966886758804, + 0.003319642972201109, + 0.003150618402287364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.029817529022693634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817115979269147 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817115979269147, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817115979269147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448410451412201, + 0.056496512144804, + 0.06956129521131516, + 0.11453165113925934, + 0.03889075294137001, + 0.07208387553691864, + 0.0583336316049099, + 0.24981831014156342, + 0.09378515928983688, + 0.10165783762931824 + ], + "credit_map": [ + 0.02361530438065529, + 0.021260084584355354, + 0.01448144018650055, + 0.02722102589905262, + 0.019030503928661346, + 0.01375215407460928, + 0.016872189939022064, + 0.008608345873653889, + 0.012216776609420776, + 0.009715549647808075, + 0.0039011461194604635, + 0.005504176486283541, + 0.006896789185702801, + 0.00801890715956688, + 0.006816681008785963, + 0.003426486160606146, + 0.011484147049486637, + 0.006793062202632427, + 0.0121367909014225, + 0.006381571292877197, + 0.012076681479811668, + 0.009206865914165974, + 0.0071298410184681416, + 0.004352341406047344, + 0.008531308732926846, + 0.01744196005165577, + 0.007358337286859751, + 0.007895750924944878, + 0.02519839257001877, + 0.026878846809267998, + 0.012562987394630909, + 0.008664069697260857, + 0.007495139725506306, + 0.004509639460593462, + 0.004894023761153221, + 0.003609214210882783, + 0.005940124858170748, + 0.00537889776751399, + 0.003646890167146921, + 0.0034168215934187174, + 0.0056479815393686295, + 0.009200427681207657, + 0.008230126462876797, + 0.010149055160582066, + 0.014116821810603142, + 0.009965411387383938, + 0.009166608564555645, + 0.005607436876744032, + 0.005658039823174477, + 0.007830808870494366, + 0.007562515791505575, + 0.003976942505687475, + 0.017594432458281517, + 0.005361527670174837, + 0.006052143406122923, + 0.00429722061380744, + 0.03615401312708855, + 0.042596571147441864, + 0.053666774183511734, + 0.017157377675175667, + 0.018058322370052338, + 0.050076838582754135, + 0.014286866411566734, + 0.017821531742811203, + 0.015486998483538628, + 0.014067305251955986, + 0.012746419757604599, + 0.005627820733934641, + 0.012148172594606876, + 0.016664400696754456, + 0.007706004660576582, + 0.009338031522929668, + 0.022448532283306122, + 0.01280584279447794, + 0.011564447544515133, + 0.007747931871563196, + 0.013408905826508999, + 0.01657678745687008, + 0.008946630172431469, + 0.008158765733242035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.002817115979269147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060978174209595 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.016060978174209595, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060978174209595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1290847212076187, + 0.16347908973693848, + 0.07210744917392731, + 0.10873283445835114, + 0.09420447051525116, + 0.09372439235448837, + 0.09958717226982117, + 0.10262623429298401, + 0.11257478594779968, + 0.023878909647464752 + ], + "credit_map": [ + 0.034080080687999725, + 0.019206158816814423, + 0.01849084347486496, + 0.014868647791445255, + 0.016309991478919983, + 0.00943880993872881, + 0.01111152395606041, + 0.00557865621522069, + 0.03493444249033928, + 0.02902151457965374, + 0.012337733060121536, + 0.009842036291956902, + 0.014938147738575935, + 0.045605480670928955, + 0.010080467909574509, + 0.006719269324094057, + 0.010179191827774048, + 0.007912796922028065, + 0.008068783208727837, + 0.008628801442682743, + 0.014778601005673409, + 0.010479388758540154, + 0.007091295439749956, + 0.004968596622347832, + 0.013287301175296307, + 0.015038756653666496, + 0.01165226474404335, + 0.009600911289453506, + 0.014379208907485008, + 0.024078505113720894, + 0.01148527767509222, + 0.009210603311657906, + 0.019670920446515083, + 0.012065095826983452, + 0.012901428155601025, + 0.009195134975016117, + 0.011912103742361069, + 0.012195027433335781, + 0.008973401971161366, + 0.007291353307664394, + 0.010861418209969997, + 0.012118188664317131, + 0.015302681364119053, + 0.010236406698822975, + 0.016218749806284904, + 0.010203534737229347, + 0.010922664776444435, + 0.007860739715397358, + 0.01405416801571846, + 0.013260599225759506, + 0.015517711639404297, + 0.008579185232520103, + 0.01995980739593506, + 0.007803434040397406, + 0.010720111429691315, + 0.009692155756056309, + 0.013018053956329823, + 0.015197102911770344, + 0.026427341625094414, + 0.006665569264441729, + 0.008514579385519028, + 0.019067412242293358, + 0.006366014014929533, + 0.007370166014879942, + 0.015084444545209408, + 0.019035454839468002, + 0.009260260500013828, + 0.00981368962675333, + 0.01439682673662901, + 0.017894184216856956, + 0.01412270125001669, + 0.012967219576239586, + 0.003977647516876459, + 0.0028675857465714216, + 0.002262139692902565, + 0.0021566571667790413, + 0.0028683932032436132, + 0.0034964687656611204, + 0.002632289892062545, + 0.0036177278961986303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.016060978174209595 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a man with short, slightly tousled brown hair, sitting astride a black, ergonomic office chair on five wheels, as if steering an out-of-control race car. The camera captures his facial expression in close, ecstatic, and playful close-up: his eyes wide with excitement, his eyebrows raised dramatically, his mouth stretched into a wide, uncontrollable grin, his cheek muscles tensed with laughter. He wears a light blue Oxford shirt, his tie slouched to one side, his sleeves casually rolled up to his elbows. He leans forward, his hands gripping the armrests of his chair. He pushes off the ground with alternating force, propelling the chair in a comical sprint across the polished office floor. In the background, neat rows of office cubicles and the glass conference room walls blur in the background's rapid movement. The fluorescent lights on the ceiling cast even and bright light, highlighting every exaggerated muscle movement and the blush on his face caused by excitement. The whole scene was full of unconventional playfulness and pure joy, creating an infectious humorous atmosphere that suddenly broke out in the tense work.", + "seed": 10994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_3/index_8720_seed_10994_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012513652769848704, + "vlm_logprob": -0.012513652769848704, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.012513652769848704, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00957550946623087 + ], + "opposite_logprob_tokens": [ + -5.634575366973877 + ], + "logprob_reward": -0.00957550946623087, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00957550946623087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15777629613876343, + 0.09936759620904922, + 0.07236618548631668, + 0.07944103330373764, + 0.1642027199268341, + 0.030600756406784058, + 0.03080751933157444, + 0.06977272033691406, + 0.09099259972572327, + 0.20467256009578705 + ], + "credit_map": [ + 0.02751399762928486, + 0.019690820947289467, + 0.017570506781339645, + 0.014899112284183502, + 0.018619194626808167, + 0.02526116743683815, + 0.019036302343010902, + 0.015185198746621609, + 0.009257485158741474, + 0.009005221538245678, + 0.01023656316101551, + 0.01275072805583477, + 0.013926761224865913, + 0.02322263829410076, + 0.008367177098989487, + 0.01260102167725563, + 0.00791673082858324, + 0.008084136061370373, + 0.007698195055127144, + 0.010761029087007046, + 0.01315575372427702, + 0.0069139194674789906, + 0.008254417218267918, + 0.009582006372511387, + 0.01173133123666048, + 0.015642566606402397, + 0.00663938419893384, + 0.005168340168893337, + 0.013457171618938446, + 0.0129435108974576, + 0.005968851502984762, + 0.007889878004789352, + 0.010599449276924133, + 0.051376648247241974, + 0.0425582192838192, + 0.011235862970352173, + 0.013346239924430847, + 0.016220739111304283, + 0.008253010921180248, + 0.010612554848194122, + 0.00399142736569047, + 0.003990519791841507, + 0.002998060081154108, + 0.003249378176406026, + 0.004332283511757851, + 0.004821951035410166, + 0.0036754263564944267, + 0.003541712649166584, + 0.003457798855379224, + 0.005016681272536516, + 0.0034798940178006887, + 0.002545482013374567, + 0.004463035613298416, + 0.0031592596787959337, + 0.004491913132369518, + 0.004193455446511507, + 0.007625773549079895, + 0.006411935668438673, + 0.005901894997805357, + 0.005530870519578457, + 0.010753282345831394, + 0.010165651328861713, + 0.013981454074382782, + 0.009401856921613216, + 0.011398641392588615, + 0.012127501890063286, + 0.010502553544938564, + 0.010806342586874962, + 0.014577805995941162, + 0.009990244172513485, + 0.00888159591704607, + 0.012707909569144249, + 0.05537136271595955, + 0.008149348199367523, + 0.029914600774645805, + 0.020869867876172066, + 0.006101400125771761, + 0.06294631212949753, + 0.010498659685254097, + 0.01082100160419941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.00957550946623087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210716545581818 + ], + "opposite_logprob_tokens": [ + -4.0182108879089355 + ], + "logprob_reward": -0.018210716545581818, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210716545581818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16867567598819733, + 0.12161161750555038, + 0.07597789168357849, + 0.059987831860780716, + 0.09367883950471878, + 0.03974452614784241, + 0.04902682453393936, + 0.044885337352752686, + 0.04399069771170616, + 0.30242079496383667 + ], + "credit_map": [ + 0.024722356349229813, + 0.017361002042889595, + 0.017969755455851555, + 0.016920993104577065, + 0.021359289065003395, + 0.03162175416946411, + 0.022175192832946777, + 0.016545336693525314, + 0.012124728411436081, + 0.011432534083724022, + 0.012415426783263683, + 0.02644738368690014, + 0.014769055880606174, + 0.014749559573829174, + 0.016046715900301933, + 0.013626214116811752, + 0.012028463184833527, + 0.006100732367485762, + 0.005645750090479851, + 0.009663213044404984, + 0.013800108805298805, + 0.008147264830768108, + 0.008404038846492767, + 0.012188312597572803, + 0.01354873925447464, + 0.014543641358613968, + 0.0052626165561378, + 0.003871751483529806, + 0.005778349470347166, + 0.008107239380478859, + 0.004031190648674965, + 0.0048442985862493515, + 0.006171126384288073, + 0.028272604569792747, + 0.0251875389367342, + 0.006017802283167839, + 0.008273721672594547, + 0.008973564021289349, + 0.004114549141377211, + 0.00666793854907155, + 0.0045495592057704926, + 0.003209319431334734, + 0.0029962819535285234, + 0.004019432235509157, + 0.005626051686704159, + 0.008895330131053925, + 0.005441932938992977, + 0.005006617866456509, + 0.008228398859500885, + 0.005969909951090813, + 0.0051637510769069195, + 0.003834004048258066, + 0.00484426086768508, + 0.010819239541888237, + 0.004148031119257212, + 0.006019225809723139, + 0.004645606502890587, + 0.004322174470871687, + 0.005087820813059807, + 0.005176188889890909, + 0.0066154468804597855, + 0.007892393507063389, + 0.006235362496227026, + 0.0049103484489023685, + 0.0037788087502121925, + 0.006583943031728268, + 0.004555779974907637, + 0.005119196139276028, + 0.010044600814580917, + 0.005375256761908531, + 0.0032788950484246016, + 0.005254220683127642, + 0.08172024041414261, + 0.011802843771874905, + 0.10446853935718536, + 0.047831904143095016, + 0.008213087916374207, + 0.027820536866784096, + 0.010310026817023754, + 0.010253602638840675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.018210716545581818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294012993574142 + ], + "opposite_logprob_tokens": [ + -3.773293972015381 + ], + "logprob_reward": -0.023294012993574142, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294012993574142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16460858285427094, + 0.09853197634220123, + 0.13588327169418335, + 0.08192098885774612, + 0.08080368489027023, + 0.11557230353355408, + 0.03732465207576752, + 0.10580199956893921, + 0.02610360085964203, + 0.15344896912574768 + ], + "credit_map": [ + 0.024167822673916817, + 0.01490735448896885, + 0.017828308045864105, + 0.02141731232404709, + 0.018112359568476677, + 0.02825234830379486, + 0.02149762213230133, + 0.01842546835541725, + 0.012505901046097279, + 0.00825093500316143, + 0.011039537377655506, + 0.013156627304852009, + 0.013628600165247917, + 0.01750342734158039, + 0.01074521616101265, + 0.01170173566788435, + 0.01412765309214592, + 0.011713262647390366, + 0.008972580544650555, + 0.010527991689741611, + 0.04300401359796524, + 0.018060093745589256, + 0.01498433668166399, + 0.014493352733552456, + 0.011326699517667294, + 0.01901540718972683, + 0.007761431857943535, + 0.006888772826641798, + 0.013703695498406887, + 0.011277035810053349, + 0.005709744058549404, + 0.00623820535838604, + 0.0050902413204312325, + 0.02102845534682274, + 0.01985102705657482, + 0.004818273708224297, + 0.00949098914861679, + 0.009734929539263248, + 0.004217894747853279, + 0.006571877747774124, + 0.01684902422130108, + 0.010070741176605225, + 0.007670292630791664, + 0.005846560467034578, + 0.015076962299644947, + 0.028107361868023872, + 0.013295091688632965, + 0.018656261265277863, + 0.005495661869645119, + 0.004164418671280146, + 0.004690604284405708, + 0.0038207254838198423, + 0.0033626616932451725, + 0.007073896937072277, + 0.004072780255228281, + 0.004643903113901615, + 0.007362884469330311, + 0.011433123610913754, + 0.008671589195728302, + 0.009058886207640171, + 0.01985052600502968, + 0.018023932352662086, + 0.016960812732577324, + 0.014440249651670456, + 0.002404631581157446, + 0.004344229120761156, + 0.002373363822698593, + 0.0023722609039396048, + 0.005158666521310806, + 0.003590289270505309, + 0.00241515738889575, + 0.00344500201754272, + 0.060029443353414536, + 0.009357644245028496, + 0.01432353351265192, + 0.013165069743990898, + 0.006913476623594761, + 0.03594110533595085, + 0.006227264180779457, + 0.007491437718272209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.023294012993574142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017119289841502905 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017119289841502905, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017119289841502905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17888681590557098, + 0.07233340293169022, + 0.0297950878739357, + 0.01880422793328762, + 0.07315639406442642, + 0.02923757955431938, + 0.023205183446407318, + 0.03234285116195679, + 0.0412876196205616, + 0.5009508728981018 + ], + "credit_map": [ + 0.02795899659395218, + 0.026310069486498833, + 0.02207816392183304, + 0.015929898247122765, + 0.019273757934570312, + 0.023070456460118294, + 0.01838153600692749, + 0.025883933529257774, + 0.011147585697472095, + 0.0057440754026174545, + 0.007672884967178106, + 0.008695613592863083, + 0.010061351582407951, + 0.015106548555195332, + 0.00425996957346797, + 0.009645365178585052, + 0.005685650277882814, + 0.003705320879817009, + 0.0022527207620441914, + 0.0029282337054610252, + 0.006053608376532793, + 0.0035224894527345896, + 0.0027619218453764915, + 0.002885141409933567, + 0.0026117623783648014, + 0.00214423518627882, + 0.0019019035389646888, + 0.0015137582086026669, + 0.003932932857424021, + 0.003322034841403365, + 0.0017292893026024103, + 0.0016483088256791234, + 0.004075984936207533, + 0.02024141699075699, + 0.01772596500813961, + 0.005141251254826784, + 0.006499180104583502, + 0.00868256390094757, + 0.003247735323384404, + 0.007542291656136513, + 0.0038147985469549894, + 0.0033412391785532236, + 0.0020207460038363934, + 0.002033429918810725, + 0.0039822980761528015, + 0.005981883034110069, + 0.004274819511920214, + 0.0037883634213358164, + 0.00316647463478148, + 0.0024087815545499325, + 0.002012077020481229, + 0.002127743326127529, + 0.003127311822026968, + 0.0031333845108747482, + 0.002616103272885084, + 0.004613306373357773, + 0.002850392134860158, + 0.0030215983279049397, + 0.004336763173341751, + 0.0043990155681967735, + 0.0046623460948467255, + 0.005580190569162369, + 0.00406311359256506, + 0.003429427742958069, + 0.0049774982035160065, + 0.0053358059376478195, + 0.004758188501000404, + 0.004097357392311096, + 0.00960580725222826, + 0.0049908701330423355, + 0.003497651545330882, + 0.004024443682283163, + 0.1790994107723236, + 0.02236424945294857, + 0.09935523569583893, + 0.051056619733572006, + 0.015091933310031891, + 0.10209153592586517, + 0.013681970536708832, + 0.018209896981716156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0017119289841502905 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man with short, slightly tousled brown hair wearing a light blue Oxford shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting astride a black ergonomic office chair on five wheels and pushing it across a polished office floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man display an ecstatic facial expression with wide eyes, raised eyebrows, and a wide grin while leaning forward with hands gripping armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of neat rows of office cubicles and glass conference room walls that appear blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot that follows the man's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776095859706402 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776095859706402, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776095859706402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15038473904132843, + 0.09231139719486237, + 0.04496953636407852, + 0.03445512428879738, + 0.07298443466424942, + 0.051458340138196945, + 0.03038475662469864, + 0.039472416043281555, + 0.07773254811763763, + 0.40584680438041687 + ], + "credit_map": [ + 0.02160640060901642, + 0.017573032528162003, + 0.017513219267129898, + 0.0141148054972291, + 0.022362040355801582, + 0.025697210803627968, + 0.015039564110338688, + 0.016478439792990685, + 0.009486783295869827, + 0.0076290313154459, + 0.010454380884766579, + 0.019352920353412628, + 0.012400994077324867, + 0.012560611590743065, + 0.009336376562714577, + 0.011090290732681751, + 0.00497286394238472, + 0.00563134765252471, + 0.001935626845806837, + 0.005863659083843231, + 0.010781087912619114, + 0.005516344681382179, + 0.0051119364798069, + 0.005156663712114096, + 0.005660157185047865, + 0.008267841301858425, + 0.003086912678554654, + 0.002289425814524293, + 0.0044634342193603516, + 0.005733251105993986, + 0.0025229903403669596, + 0.0024311121087521315, + 0.005978049244731665, + 0.015356816351413727, + 0.01203610748052597, + 0.003816218115389347, + 0.012861610390245914, + 0.011659919284284115, + 0.003413378493860364, + 0.007862328551709652, + 0.006248300429433584, + 0.00535836024209857, + 0.0034830267541110516, + 0.004476405680179596, + 0.005566715728491545, + 0.011059830896556377, + 0.00888345018029213, + 0.006382239982485771, + 0.002926232758909464, + 0.004056296311318874, + 0.004253804218024015, + 0.0033477875404059887, + 0.004217670299112797, + 0.0021962481550872326, + 0.004695952404290438, + 0.004690759815275669, + 0.0032746591605246067, + 0.004969494417309761, + 0.004242070019245148, + 0.004975331947207451, + 0.006972895003855228, + 0.005635756533592939, + 0.0046719396486878395, + 0.004730266984552145, + 0.009634834714233875, + 0.006653404328972101, + 0.009590977802872658, + 0.005060816183686256, + 0.014083546586334705, + 0.014227673411369324, + 0.008311842568218708, + 0.010169445537030697, + 0.11871166527271271, + 0.01213696226477623, + 0.12824279069900513, + 0.07190301269292831, + 0.006042166613042355, + 0.05346156656742096, + 0.007955890148878098, + 0.00739271892234683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.009776095859706402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_tensor([6984, 6984])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_tensor([6984, 6984])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aee478e5fd42fb29fca267a092f51e34a14b9d0b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_tensor([6984, 6984])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up meticulously depicts an imperfectly shaped cube of ice, resting directly on the polished, reflective surface of a dark granite countertop. The cube is translucent, its crystalline structure riddled with a delicate network of internal cracks and tiny, trapped bubbles. Its initially sharp edges gleam under a single, warm, focused spotlight emanating from an unseen source. At first, the cube begins to melt imperceptibly, then gradually; beads of condensation form and merge into small streams that flow down the cube's sides, forming a clear, shallow puddle that slowly expands from the bottom. As the action progresses, the cube's sharp geometry inexorably softens, its corners rounded, and its mass noticeably diminishing as it appears to float in its own meltwater, which spreads across the dark, smooth surface in ever-expanding, irregular shapes. The focused light creates brilliant, mirror-like highlights on the wet ice surface and the expanding puddle, while casting soft, distorted shadows that shift as the cube's shape shifts, creating a quiet, contemplative atmosphere of complete absorption in the mesmerizing and inevitable process of transformation.", + "seed": 30067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_6984_seed_30067_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.26116978470236063, + "vlm_logprob": -0.26116978470236063, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.26116978470236063, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11808763444423676 + ], + "opposite_logprob_tokens": [ + -2.2430875301361084 + ], + "logprob_reward": -0.11808763444423676, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11808763444423676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09852034598588943, + 0.040289394557476044, + 0.031271301209926605, + 0.16636797785758972, + 0.35229578614234924, + 0.05442263185977936, + 0.03613883629441261, + 0.014934144914150238, + 0.18245309591293335, + 0.02330649457871914 + ], + "credit_map": [ + 0.010455857962369919, + 0.00966795813292265, + 0.010337263345718384, + 0.009728962555527687, + 0.01613696664571762, + 0.013703595846891403, + 0.012629888020455837, + 0.01585986278951168, + 0.005494535434991121, + 0.003205859800800681, + 0.008910653181374073, + 0.0033515021204948425, + 0.0038514744956046343, + 0.006285709794610739, + 0.004086898174136877, + 0.005102761089801788, + 0.004717766772955656, + 0.002527112141251564, + 0.0059381211176514626, + 0.002820180496200919, + 0.0034862447064369917, + 0.00415254570543766, + 0.0036583407782018185, + 0.003970987629145384, + 0.028160974383354187, + 0.015537356026470661, + 0.023938041180372238, + 0.017721064388751984, + 0.022379880771040916, + 0.020152507349848747, + 0.020296799018979073, + 0.018181363120675087, + 0.057478517293930054, + 0.029661476612091064, + 0.058758556842803955, + 0.02564595267176628, + 0.0458504781126976, + 0.05610949918627739, + 0.04688064008951187, + 0.031910672783851624, + 0.010627341456711292, + 0.005170504096895456, + 0.007700772490352392, + 0.0031643726397305727, + 0.00728616351261735, + 0.01043692883104086, + 0.00520246010273695, + 0.004834088496863842, + 0.003271036082878709, + 0.003284572623670101, + 0.003237759228795767, + 0.0027885213494300842, + 0.006509459111839533, + 0.0051370179280638695, + 0.00543905608355999, + 0.006471414119005203, + 0.0026783549692481756, + 0.0014777859905734658, + 0.0017626070184633136, + 0.0012213080190122128, + 0.002184500452131033, + 0.0019416346913203597, + 0.002027251059189439, + 0.001640702597796917, + 0.04284528270363808, + 0.015322514809668064, + 0.01581406407058239, + 0.008133810013532639, + 0.02633066475391388, + 0.019198698922991753, + 0.011177541688084602, + 0.04363052546977997, + 0.0028084004297852516, + 0.0027977938298135996, + 0.0025207907892763615, + 0.0022060431074351072, + 0.0036484699230641127, + 0.0029238902498036623, + 0.002662709215655923, + 0.0037383995950222015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.11808763444423676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905167877674103 + ], + "opposite_logprob_tokens": [ + -2.4640517234802246 + ], + "logprob_reward": -0.08905167877674103, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905167877674103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12802699208259583, + 0.06415697932243347, + 0.03513858839869499, + 0.045097772032022476, + 0.1958814561367035, + 0.1848170906305313, + 0.03639018163084984, + 0.024744698777794838, + 0.19783751666545868, + 0.08790873736143112 + ], + "credit_map": [ + 0.01842646487057209, + 0.008509335108101368, + 0.023102430626749992, + 0.009765351191163063, + 0.013270684517920017, + 0.023693213239312172, + 0.01226230151951313, + 0.018997205421328545, + 0.00817769393324852, + 0.004353250842541456, + 0.012768421322107315, + 0.005275614093989134, + 0.006707827094942331, + 0.011743861250579357, + 0.006543676368892193, + 0.008586636744439602, + 0.006839600391685963, + 0.002675166353583336, + 0.006023947149515152, + 0.002390642650425434, + 0.003775030141696334, + 0.004617190919816494, + 0.004022309090942144, + 0.00479469820857048, + 0.00598488375544548, + 0.003906779456883669, + 0.008974846452474594, + 0.004869766999036074, + 0.004521768074482679, + 0.006427004933357239, + 0.0037227473221719265, + 0.006689978297799826, + 0.03384332358837128, + 0.017861323431134224, + 0.0375494547188282, + 0.01313153374940157, + 0.02311660163104534, + 0.03054455853998661, + 0.019386693835258484, + 0.020447954535484314, + 0.03916144371032715, + 0.02171282283961773, + 0.027152711525559425, + 0.013217703439295292, + 0.019482681527733803, + 0.030145397409796715, + 0.014922757633030415, + 0.019021566957235336, + 0.006394264288246632, + 0.004426812287420034, + 0.005084210075438023, + 0.003465658985078335, + 0.00428402703255415, + 0.005194799043238163, + 0.0038809827528893948, + 0.003659429494291544, + 0.004005197901278734, + 0.002522001275792718, + 0.0034806584008038044, + 0.0022616260685026646, + 0.003328873310238123, + 0.0034830558579415083, + 0.0028709189500659704, + 0.0027923681773245335, + 0.04055699333548546, + 0.021009016782045364, + 0.03195207938551903, + 0.016961341723799706, + 0.017369752749800682, + 0.01996678300201893, + 0.015061888843774796, + 0.03495965152978897, + 0.02580435946583748, + 0.01264488697052002, + 0.006422766949981451, + 0.005111798644065857, + 0.0072293514385819435, + 0.008380684070289135, + 0.007016254588961601, + 0.015298630110919476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.08905167877674103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539205074310303 + ], + "opposite_logprob_tokens": [ + -0.42892053723335266 + ], + "logprob_reward": -1.0539205074310303, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539205074310303, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15272672474384308, + 0.1078067421913147, + 0.08291162550449371, + 0.1429152488708496, + 0.074226975440979, + 0.13691173493862152, + 0.050388701260089874, + 0.03137170523405075, + 0.09726814180612564, + 0.1234724149107933 + ], + "credit_map": [ + 0.0237432774156332, + 0.012213070876896381, + 0.022811340168118477, + 0.015351281501352787, + 0.014727890491485596, + 0.026033764705061913, + 0.013800431974232197, + 0.024045675992965698, + 0.01734532229602337, + 0.009130792692303658, + 0.016843877732753754, + 0.008932339958846569, + 0.014752516523003578, + 0.015032345429062843, + 0.01011031772941351, + 0.015659229829907417, + 0.014529290609061718, + 0.007762623950839043, + 0.013043048791587353, + 0.007192668039351702, + 0.008910126984119415, + 0.010896295309066772, + 0.009784982539713383, + 0.01079259067773819, + 0.021330390125513077, + 0.01420169323682785, + 0.023257708176970482, + 0.012079047039151192, + 0.0181080661714077, + 0.019353196024894714, + 0.017221948131918907, + 0.017363203689455986, + 0.011442537419497967, + 0.009357945062220097, + 0.009320481680333614, + 0.0067538972944021225, + 0.00847356952726841, + 0.011684631928801537, + 0.008799242787063122, + 0.008394668810069561, + 0.02617531828582287, + 0.01340487040579319, + 0.01860189624130726, + 0.009431757032871246, + 0.017812995240092278, + 0.027844134718179703, + 0.010393484495580196, + 0.013247265480458736, + 0.006856222171336412, + 0.004928095731884241, + 0.006590905599296093, + 0.005789557937532663, + 0.007547326851636171, + 0.006823659408837557, + 0.004595452919602394, + 0.00725747924298048, + 0.004291519057005644, + 0.002985247876495123, + 0.0036623484920710325, + 0.002436813898384571, + 0.006603301502764225, + 0.003972490783780813, + 0.004291103221476078, + 0.0031288794707506895, + 0.013162603601813316, + 0.01168365590274334, + 0.01307766418904066, + 0.009061696007847786, + 0.01669926568865776, + 0.011828895658254623, + 0.012405848130583763, + 0.009348515421152115, + 0.018316209316253662, + 0.021584872156381607, + 0.018915444612503052, + 0.01617865078151226, + 0.011827748268842697, + 0.012277820147573948, + 0.011281638406217098, + 0.013090025633573532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -1.0539205074310303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085470207035542 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085470207035542, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085470207035542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12930141389369965, + 0.16862206161022186, + 0.03548327833414078, + 0.04205257445573807, + 0.17248989641666412, + 0.0996878445148468, + 0.08041684329509735, + 0.051554612815380096, + 0.07897721976041794, + 0.14141424000263214 + ], + "credit_map": [ + 0.02679522894322872, + 0.010041404515504837, + 0.018273741006851196, + 0.00925225205719471, + 0.012163624167442322, + 0.02017388679087162, + 0.014331577345728874, + 0.018269697204232216, + 0.028718039393424988, + 0.012189800851047039, + 0.037474483251571655, + 0.011410215869545937, + 0.01204331498593092, + 0.03075544349849224, + 0.01183613482862711, + 0.0241946242749691, + 0.006970004644244909, + 0.0032024276442825794, + 0.006987936794757843, + 0.0023728806991130114, + 0.0032321151811629534, + 0.00493133720010519, + 0.0037868639919906855, + 0.003999712411314249, + 0.005488662514835596, + 0.004070190247148275, + 0.006088410969823599, + 0.0040105790831148624, + 0.006742819678038359, + 0.005122194066643715, + 0.005068717524409294, + 0.0054610008373856544, + 0.0227860976010561, + 0.018628599122166634, + 0.022507181391119957, + 0.015907520428299904, + 0.027527717873454094, + 0.021398277953267097, + 0.02729102410376072, + 0.016443483531475067, + 0.017870476469397545, + 0.010410681366920471, + 0.017603226006031036, + 0.007677490822970867, + 0.007708019111305475, + 0.01611410453915596, + 0.009298797696828842, + 0.01300505455583334, + 0.012368704192340374, + 0.007684615440666676, + 0.015937719494104385, + 0.0087388651445508, + 0.008498042821884155, + 0.010888888500630856, + 0.007290699053555727, + 0.009009311906993389, + 0.008997094817459583, + 0.005076560191810131, + 0.005127579439431429, + 0.0026246327906847, + 0.008780151605606079, + 0.009989546611905098, + 0.006090853828936815, + 0.004868196323513985, + 0.011096992529928684, + 0.009922715835273266, + 0.012047895230352879, + 0.00678503792732954, + 0.011417848989367485, + 0.008141729980707169, + 0.011900353245437145, + 0.007664646953344345, + 0.03500418737530708, + 0.0166135523468256, + 0.015365995466709137, + 0.011940078809857368, + 0.0105506656691432, + 0.013818093575537205, + 0.009324803948402405, + 0.028796866536140442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.011085470207035542 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370363265275955 + ], + "opposite_logprob_tokens": [ + -3.408703565597534 + ], + "logprob_reward": -0.03370363265275955, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370363265275955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21428129076957703, + 0.05426253378391266, + 0.0754377543926239, + 0.08152119815349579, + 0.044107478111982346, + 0.10168551653623581, + 0.06257500499486923, + 0.08051788806915283, + 0.15589188039302826, + 0.12971951067447662 + ], + "credit_map": [ + 0.030554641038179398, + 0.015087664127349854, + 0.036412183195352554, + 0.013546032831072807, + 0.021665459498763084, + 0.04219067841768265, + 0.021875115111470222, + 0.032949477434158325, + 0.00951986014842987, + 0.0040119183249771595, + 0.00849191751331091, + 0.00397007679566741, + 0.005270044319331646, + 0.009995084255933762, + 0.005054599605500698, + 0.007949029095470905, + 0.009725376032292843, + 0.006026421207934618, + 0.011320379562675953, + 0.00826640147715807, + 0.010681519284844398, + 0.009931300766766071, + 0.01043061725795269, + 0.009055731818079948, + 0.008934937417507172, + 0.00729932077229023, + 0.009884321130812168, + 0.008103176951408386, + 0.011306194588541985, + 0.013236693106591702, + 0.01039888896048069, + 0.012357661500573158, + 0.007036176044493914, + 0.003400271525606513, + 0.009471814148128033, + 0.003152638440951705, + 0.004784525837749243, + 0.0068697454407811165, + 0.004878763575106859, + 0.004513540305197239, + 0.018440429121255875, + 0.011706708930432796, + 0.016419515013694763, + 0.007680004928261042, + 0.011216367594897747, + 0.014554465189576149, + 0.011743304319679737, + 0.00992470420897007, + 0.012291966006159782, + 0.007370541803538799, + 0.010619839653372765, + 0.009129561483860016, + 0.005063246935606003, + 0.008322188630700111, + 0.0047279829159379005, + 0.005049674306064844, + 0.008905017748475075, + 0.008836011402308941, + 0.0077659450471401215, + 0.0050843809731304646, + 0.01808100752532482, + 0.010441004298627377, + 0.012687460519373417, + 0.008717049844563007, + 0.0217962134629488, + 0.016386600211262703, + 0.017835555598139763, + 0.011850678361952305, + 0.03263174369931221, + 0.02333737351000309, + 0.013856876641511917, + 0.018196819350123405, + 0.03829675912857056, + 0.021754832938313484, + 0.007649590726941824, + 0.006358741316944361, + 0.010955394245684147, + 0.010378857143223286, + 0.010688425973057747, + 0.02363690361380577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03370363265275955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up meticulously depicts an imperfectly shaped cube of ice, resting directly on the polished, reflective surface of a dark granite countertop. The cube is translucent, its crystalline structure riddled with a delicate network of internal cracks and tiny, trapped bubbles. Its initially sharp edges gleam under a single, warm, focused spotlight emanating from an unseen source. At first, the cube begins to melt imperceptibly, then gradually; beads of condensation form and merge into small streams that flow down the cube's sides, forming a clear, shallow puddle that slowly expands from the bottom. As the action progresses, the cube's sharp geometry inexorably softens, its corners rounded, and its mass noticeably diminishing as it appears to float in its own meltwater, which spreads across the dark, smooth surface in ever-expanding, irregular shapes. The focused light creates brilliant, mirror-like highlights on the wet ice surface and the expanding puddle, while casting soft, distorted shadows that shift as the cube's shape shifts, creating a quiet, contemplative atmosphere of complete absorption in the mesmerizing and inevitable process of transformation.", + "seed": 30068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_6984_seed_30068_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.27507966086268426, + "vlm_logprob": -0.27507966086268426, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.27507966086268426, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08487015962600708 + ], + "opposite_logprob_tokens": [ + -2.5848701000213623 + ], + "logprob_reward": -0.08487015962600708, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08487015962600708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062081415206193924, + 0.029314333572983742, + 0.034221719950437546, + 0.021250171586871147, + 0.5412390232086182, + 0.0967075303196907, + 0.06913971900939941, + 0.10054877400398254, + 0.026344161480665207, + 0.019153108820319176 + ], + "credit_map": [ + 0.011070309206843376, + 0.004714554641395807, + 0.011080021969974041, + 0.00435312744230032, + 0.006281532347202301, + 0.011673861183226109, + 0.004273869562894106, + 0.00863413605839014, + 0.0031963346991688013, + 0.002928894478827715, + 0.005931375548243523, + 0.0028610783629119396, + 0.0025725916493684053, + 0.0038566405419260263, + 0.0031838680151849985, + 0.00478355074301362, + 0.0056688603945076466, + 0.0032846988178789616, + 0.007388826925307512, + 0.002287857001647353, + 0.0024063815362751484, + 0.005608069710433483, + 0.002111877780407667, + 0.005465145688503981, + 0.002156822709366679, + 0.0017701559700071812, + 0.004973852541297674, + 0.001509744324721396, + 0.0016575503395870328, + 0.0048657734878361225, + 0.00193918461445719, + 0.0023770881816744804, + 0.04470793530344963, + 0.021825645118951797, + 0.04008466750383377, + 0.028207628056406975, + 0.031172137707471848, + 0.18227386474609375, + 0.027213547378778458, + 0.16575361788272858, + 0.02078194171190262, + 0.009476345963776112, + 0.028129857033491135, + 0.00595275079831481, + 0.007821658626198769, + 0.008753223344683647, + 0.007398012559860945, + 0.008393744938075542, + 0.009037845768034458, + 0.0071830060333013535, + 0.012357904575765133, + 0.005511707626283169, + 0.014308104291558266, + 0.005430625285953283, + 0.0073459879495203495, + 0.00796453282237053, + 0.02816358394920826, + 0.007338699419051409, + 0.032202254980802536, + 0.006861008703708649, + 0.007453331258147955, + 0.005569145083427429, + 0.006261472124606371, + 0.006699277088046074, + 0.004321104846894741, + 0.0021788410376757383, + 0.0075777266174554825, + 0.0020558496471494436, + 0.002265595830976963, + 0.0033027525059878826, + 0.0016713192453607917, + 0.002970971167087555, + 0.003702090587466955, + 0.0011501617264002562, + 0.006606960669159889, + 0.0011878462973982096, + 0.00156550039537251, + 0.0024986055213958025, + 0.0013479795306921005, + 0.001093963859602809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.08487015962600708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031679272651672 + ], + "opposite_logprob_tokens": [ + -2.3503167629241943 + ], + "logprob_reward": -0.10031679272651672, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031679272651672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03156673535704613, + 0.009877117350697517, + 0.013571498915553093, + 0.014194024726748466, + 0.6200046539306641, + 0.05508634075522423, + 0.0199438389390707, + 0.09018544107675552, + 0.04363277181982994, + 0.10193760693073273 + ], + "credit_map": [ + 0.005407604854553938, + 0.002241872949525714, + 0.006916092708706856, + 0.0019356644479557872, + 0.0018566467333585024, + 0.004561976529657841, + 0.0019500328926369548, + 0.006696845404803753, + 0.0012239237548783422, + 0.0006503405747935176, + 0.0024774083867669106, + 0.0007663042633794248, + 0.0006048965733498335, + 0.001370572834275663, + 0.0007342912722378969, + 0.002049380913376808, + 0.0015827976167201996, + 0.0008350797579623759, + 0.0037183589302003384, + 0.0007356966962106526, + 0.000834792444948107, + 0.0043169124983251095, + 0.0005095411324873567, + 0.0010383217595517635, + 0.0016333047533407807, + 0.0008489241590723395, + 0.003400651505216956, + 0.0009751301258802414, + 0.0008257189183495939, + 0.002865948947146535, + 0.0012362366542220116, + 0.002408111933618784, + 0.050645627081394196, + 0.023873230442404747, + 0.04552935063838959, + 0.032161250710487366, + 0.03336206078529358, + 0.21028387546539307, + 0.03163276985287666, + 0.1925164759159088, + 0.019502397626638412, + 0.0023105614818632603, + 0.005216412246227264, + 0.0015084006590768695, + 0.0013498449698090553, + 0.007569516077637672, + 0.0018006094032898545, + 0.015828602015972137, + 0.0037237280048429966, + 0.0024936331901699305, + 0.002971273846924305, + 0.002061083447188139, + 0.0034568477421998978, + 0.0014153484953567386, + 0.0021054898388683796, + 0.0017164374003186822, + 0.033560946583747864, + 0.00414496473968029, + 0.033273473381996155, + 0.0043204957619309425, + 0.003318934701383114, + 0.005290865432471037, + 0.0030381185933947563, + 0.0032376525923609734, + 0.005962801165878773, + 0.002948555862531066, + 0.015824859961867332, + 0.002974448259919882, + 0.003566071391105652, + 0.004756434820592403, + 0.002589645329862833, + 0.00500995758920908, + 0.03673427551984787, + 0.003474448574706912, + 0.03523275628685951, + 0.0036992577370256186, + 0.003110733116045594, + 0.010061630979180336, + 0.0032811614219099283, + 0.006343354936689138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.10031679272651672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370904445648193 + ], + "opposite_logprob_tokens": [ + -0.38709041476249695 + ], + "logprob_reward": -1.1370904445648193, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370904445648193, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06121435388922691, + 0.05034622177481651, + 0.015713663771748543, + 0.039038579910993576, + 0.11912129819393158, + 0.06254283338785172, + 0.08141444623470306, + 0.5285470485687256, + 0.01734498329460621, + 0.02471655234694481 + ], + "credit_map": [ + 0.010551744140684605, + 0.0036020094994455576, + 0.017610901966691017, + 0.00394587405025959, + 0.004665016196668148, + 0.00861078966408968, + 0.003364522010087967, + 0.008863498456776142, + 0.007167842239141464, + 0.004222598858177662, + 0.010567519813776016, + 0.003968095872551203, + 0.0038486607372760773, + 0.008126137778162956, + 0.003600283293053508, + 0.008845084346830845, + 0.0020125077571719885, + 0.0015631080605089664, + 0.0025332916993647814, + 0.0023114560171961784, + 0.0011780301574617624, + 0.0024402164854109287, + 0.0011700213653966784, + 0.002505031181499362, + 0.004624747671186924, + 0.002248225500807166, + 0.01179245114326477, + 0.0027667658869177103, + 0.0016517577460035682, + 0.009821716696023941, + 0.0018413111101835966, + 0.004291602410376072, + 0.010494938120245934, + 0.0049507212825119495, + 0.013169851154088974, + 0.006054888479411602, + 0.0083377156406641, + 0.03521353378891945, + 0.006213831715285778, + 0.03468581661581993, + 0.023257466033101082, + 0.0038273234385997057, + 0.00534220552071929, + 0.0021987860091030598, + 0.0022367562633007765, + 0.00861781369894743, + 0.002270948840305209, + 0.014791537076234818, + 0.011657760478556156, + 0.011337287724018097, + 0.014141361229121685, + 0.007711461745202541, + 0.011108490638434887, + 0.00762855913490057, + 0.008824927732348442, + 0.009004591964185238, + 0.1978970468044281, + 0.01956840045750141, + 0.20585663616657257, + 0.023576699197292328, + 0.01796373911201954, + 0.028786908835172653, + 0.016224153339862823, + 0.018673492595553398, + 0.003229838330298662, + 0.0014005855191498995, + 0.0038596822414547205, + 0.0015793615020811558, + 0.0019615842029452324, + 0.002311589429154992, + 0.0014672501711174846, + 0.0015350915491580963, + 0.0024728192947804928, + 0.0013524213572964072, + 0.00805723387748003, + 0.00164229073561728, + 0.0017058264929801226, + 0.006395010743290186, + 0.001444170717149973, + 0.0016467786626890302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -1.1370904445648193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023310784250497818 + ], + "opposite_logprob_tokens": [ + -3.773310899734497 + ], + "logprob_reward": -0.023310784250497818, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023310784250497818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04221826419234276, + 0.0394008569419384, + 0.0818125531077385, + 0.06515774875879288, + 0.26186031103134155, + 0.04337596520781517, + 0.050322797149419785, + 0.3799961507320404, + 0.018074113875627518, + 0.017781198024749756 + ], + "credit_map": [ + 0.006700924132019281, + 0.0029095166828483343, + 0.007954790256917477, + 0.0030436848755925894, + 0.003039297414943576, + 0.006447122897952795, + 0.002782694762572646, + 0.009340236894786358, + 0.006098917219787836, + 0.002083365572616458, + 0.008902156725525856, + 0.002737050876021385, + 0.0017138333059847355, + 0.006499460898339748, + 0.0022058782633394003, + 0.009160199202597141, + 0.013017972931265831, + 0.005229955539107323, + 0.013086375780403614, + 0.005812923423945904, + 0.00508180120959878, + 0.016578495502471924, + 0.004448045510798693, + 0.0185569915920496, + 0.006888740696012974, + 0.004276381805539131, + 0.020733755081892014, + 0.004106664564460516, + 0.003807331435382366, + 0.01197764277458191, + 0.005398259963840246, + 0.007968978956341743, + 0.02083941549062729, + 0.009997167624533176, + 0.019082650542259216, + 0.013296769000589848, + 0.014148695394396782, + 0.08874005824327469, + 0.013531738892197609, + 0.08222384005784988, + 0.014998032711446285, + 0.0020741287153214216, + 0.004137885756790638, + 0.0012445165775716305, + 0.0012798430398106575, + 0.005555356852710247, + 0.0016052990686148405, + 0.01248090248554945, + 0.011920093558728695, + 0.004434168804436922, + 0.01211561169475317, + 0.0035518757067620754, + 0.007368892431259155, + 0.0030196080915629864, + 0.004346428904682398, + 0.003566121216863394, + 0.14516301453113556, + 0.01395349483937025, + 0.14809586107730865, + 0.016536492854356766, + 0.011849110014736652, + 0.021323466673493385, + 0.011145592667162418, + 0.011929132975637913, + 0.002573599573224783, + 0.0011972415959462523, + 0.005997723899781704, + 0.0013334420509636402, + 0.0013497764011844993, + 0.0022141002118587494, + 0.0014259322779253125, + 0.001982299843803048, + 0.005755995400249958, + 0.0007939243223518133, + 0.006381577346473932, + 0.0006724923732690513, + 0.0008788654231466353, + 0.0011839335784316063, + 0.0008183704339899123, + 0.0012960416497662663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.023310784250497818 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810123145580292 + ], + "opposite_logprob_tokens": [ + -3.5298101902008057 + ], + "logprob_reward": -0.029810123145580292, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810123145580292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05568169429898262, + 0.01310969889163971, + 0.010932689532637596, + 0.020087536424398422, + 0.4370400905609131, + 0.07959292083978653, + 0.015687398612499237, + 0.32101428508758545, + 0.016566574573516846, + 0.030287139117717743 + ], + "credit_map": [ + 0.011952796950936317, + 0.004411817528307438, + 0.007223257329314947, + 0.0034668806474655867, + 0.0037999211344867945, + 0.016707107424736023, + 0.003821716411039233, + 0.0042981961742043495, + 0.001971533987671137, + 0.0012206851970404387, + 0.00326728168874979, + 0.0008392948657274246, + 0.0008614733815193176, + 0.0018167966045439243, + 0.0007942162337712944, + 0.0023384173400700092, + 0.0015784638235345483, + 0.0013469138648360968, + 0.0016830278327688575, + 0.001200466533191502, + 0.0010267860488966107, + 0.0021071238443255424, + 0.0008381633087992668, + 0.0011517447419464588, + 0.001955080544576049, + 0.0017682441975921392, + 0.0033875922672450542, + 0.003151262179017067, + 0.00133324577473104, + 0.0032989848405122757, + 0.0029652006924152374, + 0.0022279273252934217, + 0.03627535700798035, + 0.017487071454524994, + 0.03161926195025444, + 0.022992804646492004, + 0.024345654994249344, + 0.14859424531459808, + 0.022128546610474586, + 0.13359712064266205, + 0.030457453802227974, + 0.004058138467371464, + 0.0064882924780249596, + 0.0023376017343252897, + 0.0022719711996614933, + 0.010447665117681026, + 0.002557565923780203, + 0.020974228158593178, + 0.0016712152864784002, + 0.0020066017750650644, + 0.002473521512001753, + 0.0014075998915359378, + 0.002311691641807556, + 0.0015278265345841646, + 0.002144539961591363, + 0.0021444011945277452, + 0.12048015743494034, + 0.012302766554057598, + 0.12379420548677444, + 0.014816347509622574, + 0.010955688543617725, + 0.018062936142086983, + 0.010045411996543407, + 0.010556777939200401, + 0.002690175548195839, + 0.0016813665861263871, + 0.004159106407314539, + 0.0016185806598514318, + 0.002038322389125824, + 0.0017102357232943177, + 0.0010913392761722207, + 0.0015774468192830682, + 0.00887716468423605, + 0.0010464017977938056, + 0.012097042985260487, + 0.0013764358591288328, + 0.0012570578837767243, + 0.0024740383960306644, + 0.0013368127401918173, + 0.0018221838399767876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.029810123145580292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up meticulously depicts an imperfectly shaped cube of ice, resting directly on the polished, reflective surface of a dark granite countertop. The cube is translucent, its crystalline structure riddled with a delicate network of internal cracks and tiny, trapped bubbles. Its initially sharp edges gleam under a single, warm, focused spotlight emanating from an unseen source. At first, the cube begins to melt imperceptibly, then gradually; beads of condensation form and merge into small streams that flow down the cube's sides, forming a clear, shallow puddle that slowly expands from the bottom. As the action progresses, the cube's sharp geometry inexorably softens, its corners rounded, and its mass noticeably diminishing as it appears to float in its own meltwater, which spreads across the dark, smooth surface in ever-expanding, irregular shapes. The focused light creates brilliant, mirror-like highlights on the wet ice surface and the expanding puddle, while casting soft, distorted shadows that shift as the cube's shape shifts, creating a quiet, contemplative atmosphere of complete absorption in the mesmerizing and inevitable process of transformation.", + "seed": 30069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_6984_seed_30069_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.094362036511302, + "vlm_logprob": -0.094362036511302, + "vlm_yes_no_margin": 3.200000011920929, + "vlm_token_credit": -0.094362036511302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011408921331167221 + ], + "opposite_logprob_tokens": [ + -5.011408805847168 + ], + "logprob_reward": -0.011408921331167221, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011408921331167221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13632020354270935, + 0.07932915538549423, + 0.022454939782619476, + 0.032892920076847076, + 0.08497761934995651, + 0.2424839287996292, + 0.1581529974937439, + 0.16302278637886047, + 0.02225102297961712, + 0.0581144243478775 + ], + "credit_map": [ + 0.01667879708111286, + 0.010983351618051529, + 0.027854029089212418, + 0.009529958479106426, + 0.015346276573836803, + 0.021287478506565094, + 0.015451867133378983, + 0.019188450649380684, + 0.010202313773334026, + 0.009202838875353336, + 0.01104552298784256, + 0.00971957203000784, + 0.009318789467215538, + 0.01168647687882185, + 0.007165107410401106, + 0.010988534428179264, + 0.003777159843593836, + 0.002174608875066042, + 0.00363503722473979, + 0.0019143985118716955, + 0.001988764852285385, + 0.002920500934123993, + 0.002011986216530204, + 0.004032484255731106, + 0.004439945332705975, + 0.0028679317329078913, + 0.005987243726849556, + 0.002375858137384057, + 0.0024611835833638906, + 0.005060961935669184, + 0.0027612692210823298, + 0.006938527803868055, + 0.012688243761658669, + 0.008569666184484959, + 0.012655424885451794, + 0.009698109701275826, + 0.010661506094038486, + 0.009447257034480572, + 0.012466798536479473, + 0.00879061222076416, + 0.030463648959994316, + 0.008951052092015743, + 0.06855911761522293, + 0.01251899916678667, + 0.011732636950910091, + 0.07633693516254425, + 0.016127945855259895, + 0.017793606966733932, + 0.029369691386818886, + 0.013250480405986309, + 0.03232746943831444, + 0.010916930623352528, + 0.015279103070497513, + 0.025067944079637527, + 0.0120509322732687, + 0.019890442490577698, + 0.033018093556165695, + 0.012628189288079739, + 0.029797784984111786, + 0.011053632013499737, + 0.01282760314643383, + 0.024692602455615997, + 0.012994110584259033, + 0.026010766625404358, + 0.0029629890341311693, + 0.0020556072704494, + 0.0038607625756412745, + 0.0016985713737085462, + 0.003738989122211933, + 0.0030846144072711468, + 0.002406378509476781, + 0.0024431105703115463, + 0.011669579893350601, + 0.0032740659080445766, + 0.011368487030267715, + 0.003372503211721778, + 0.006041047163307667, + 0.00962116103619337, + 0.004899539519101381, + 0.007868041284382343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.011408921331167221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636326290667057 + ], + "opposite_logprob_tokens": [ + -3.6513633728027344 + ], + "logprob_reward": -0.02636326290667057, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636326290667057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.063062384724617, + 0.03722193092107773, + 0.012656115926802158, + 0.026881536468863487, + 0.04741387069225311, + 0.5620200634002686, + 0.08322595804929733, + 0.07324224710464478, + 0.05169036239385605, + 0.0425855778157711 + ], + "credit_map": [ + 0.009099073708057404, + 0.006639366969466209, + 0.010747436434030533, + 0.006832873914390802, + 0.005725148133933544, + 0.006971483118832111, + 0.008458280004560947, + 0.008588718250393867, + 0.00670833932235837, + 0.0033211286645382643, + 0.005349744111299515, + 0.003783800406381488, + 0.0037470622919499874, + 0.006227721925824881, + 0.0028923763893544674, + 0.005191755015403032, + 0.00198119948618114, + 0.0010793573455885053, + 0.0024334590416401625, + 0.0010166491847485304, + 0.0012227896368131042, + 0.0018044405151158571, + 0.0010521291987970471, + 0.0020660902373492718, + 0.0045045241713523865, + 0.003003257792443037, + 0.004415086470544338, + 0.002891280222684145, + 0.002900136634707451, + 0.0025561563670635223, + 0.0038888331037014723, + 0.0027222582139074802, + 0.006688007619231939, + 0.0064238859340548515, + 0.006393577437847853, + 0.005176957231014967, + 0.006719676312059164, + 0.004922829568386078, + 0.00583803653717041, + 0.005250899121165276, + 0.08707085996866226, + 0.03825071081519127, + 0.11229322105646133, + 0.04105783998966217, + 0.03709934279322624, + 0.1483706831932068, + 0.05502839758992195, + 0.04284891486167908, + 0.01654144562780857, + 0.0064740413799881935, + 0.017624324187636375, + 0.008041848428547382, + 0.005193824414163828, + 0.01229113806039095, + 0.005831519607454538, + 0.011227808892726898, + 0.013395394198596478, + 0.005781153216958046, + 0.01615322008728981, + 0.0045347027480602264, + 0.005750914569944143, + 0.010056324303150177, + 0.00511465547606349, + 0.012455875985324383, + 0.008821588940918446, + 0.00334919523447752, + 0.011532250791788101, + 0.0033434710931032896, + 0.003951265476644039, + 0.007918023504316807, + 0.004362457897514105, + 0.008412105031311512, + 0.009321070276200771, + 0.002708018757402897, + 0.007436753716319799, + 0.002579853404313326, + 0.004914587363600731, + 0.006806795950978994, + 0.003881384152919054, + 0.004937110468745232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.02636326290667057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34868282079696655 + ], + "opposite_logprob_tokens": [ + -1.2236828804016113 + ], + "logprob_reward": -0.34868282079696655, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34868282079696655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07159139215946198, + 0.03414557874202728, + 0.014519001357257366, + 0.018368056043982506, + 0.058037906885147095, + 0.46035444736480713, + 0.1727565973997116, + 0.10352925211191177, + 0.0220233965665102, + 0.044674430042505264 + ], + "credit_map": [ + 0.010534243658185005, + 0.004859759006649256, + 0.011949775740504265, + 0.005361033137887716, + 0.007625597529113293, + 0.01287195086479187, + 0.005148602183908224, + 0.013240429572761059, + 0.004564025904983282, + 0.002351202070713043, + 0.006157031282782555, + 0.002514281775802374, + 0.0031112905126065016, + 0.005315212532877922, + 0.00212362059392035, + 0.008008912205696106, + 0.0023791969288140535, + 0.0011018069926649332, + 0.00232270872220397, + 0.0012843880103901029, + 0.0012332549085840583, + 0.001881863921880722, + 0.0013138360809534788, + 0.0030019443947821856, + 0.0028151338919997215, + 0.0018884435994550586, + 0.002484550466760993, + 0.0017446859274059534, + 0.0015589476097375154, + 0.002357992809265852, + 0.0014566556783393025, + 0.004061645362526178, + 0.009016718715429306, + 0.00725860521197319, + 0.009030419401824474, + 0.00588183430954814, + 0.00700297299772501, + 0.00653522415086627, + 0.00848064012825489, + 0.004831487778574228, + 0.06645224243402481, + 0.02466973289847374, + 0.10835464298725128, + 0.028197260573506355, + 0.02558176964521408, + 0.1363813430070877, + 0.03728744015097618, + 0.033429987728595734, + 0.03256509080529213, + 0.012428400106728077, + 0.041386693716049194, + 0.012047423049807549, + 0.010919620282948017, + 0.028397148475050926, + 0.010973229072988033, + 0.024038970470428467, + 0.019097359851002693, + 0.0077785830944776535, + 0.019347194582223892, + 0.0076421708799898624, + 0.008225655183196068, + 0.015883035957813263, + 0.00821679923683405, + 0.017338450998067856, + 0.003749921452254057, + 0.002134584588930011, + 0.0038825843948870897, + 0.0024796698708087206, + 0.00202481378801167, + 0.002711724489927292, + 0.002314627869054675, + 0.0027254694141447544, + 0.009260960854589939, + 0.0029724070336669683, + 0.00816067773848772, + 0.002975544659420848, + 0.0046825832687318325, + 0.007129407953470945, + 0.004073174670338631, + 0.005419666413217783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.34868282079696655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302748799324036 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302748799324036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302748799324036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13864286243915558, + 0.02037939988076687, + 0.011730888858437538, + 0.03542673960328102, + 0.035524189472198486, + 0.5307125449180603, + 0.04927351325750351, + 0.012766562402248383, + 0.05034099891781807, + 0.1152023896574974 + ], + "credit_map": [ + 0.018361996859312057, + 0.009047817438840866, + 0.022918617352843285, + 0.009684642776846886, + 0.012557702139019966, + 0.028437415137887, + 0.010110371746122837, + 0.02752428501844406, + 0.0026650400832295418, + 0.0022635976783931255, + 0.003163117216899991, + 0.0012312965700402856, + 0.002123055048286915, + 0.003717245301231742, + 0.0013581993989646435, + 0.0038578466046601534, + 0.0019485880620777607, + 0.0007701761787757277, + 0.001993676647543907, + 0.0005686759250238538, + 0.0009699559886939824, + 0.0015735970810055733, + 0.0006831894279457629, + 0.003223028965294361, + 0.006094059906899929, + 0.0028643168043345213, + 0.006892327684909105, + 0.0024553609546273947, + 0.0023711416870355606, + 0.004453692585229874, + 0.0027576154097914696, + 0.00753822410479188, + 0.006271486636251211, + 0.002974137431010604, + 0.00536966510117054, + 0.003079631831496954, + 0.004355159588158131, + 0.006283309310674667, + 0.0035786035005003214, + 0.0036121918819844723, + 0.07890896499156952, + 0.03143387287855148, + 0.1169828325510025, + 0.03539028391242027, + 0.030866961926221848, + 0.15131708979606628, + 0.046724241226911545, + 0.03908826410770416, + 0.010816480964422226, + 0.0030673271976411343, + 0.010990158654749393, + 0.003753339173272252, + 0.0037075059954077005, + 0.007577507756650448, + 0.0036797139327973127, + 0.005681472364813089, + 0.002130257897078991, + 0.0009094014531001449, + 0.002715341979637742, + 0.0009196464670822024, + 0.0012892518425360322, + 0.0018023911397904158, + 0.0011926222359761596, + 0.0018076479900628328, + 0.0115186283364892, + 0.003350758459419012, + 0.008330284617841244, + 0.0034434504341334105, + 0.0047337510623037815, + 0.008156000636518002, + 0.005150160286575556, + 0.005657954141497612, + 0.024673203006386757, + 0.006202978547662497, + 0.02349439077079296, + 0.005770504008978605, + 0.011305302381515503, + 0.018350008875131607, + 0.0100935660302639, + 0.015312433242797852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.023302748799324036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205242872238159 + ], + "opposite_logprob_tokens": [ + -2.8120524883270264 + ], + "logprob_reward": -0.06205242872238159, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205242872238159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12832693755626678, + 0.07805047184228897, + 0.031131234019994736, + 0.06228907033801079, + 0.11759975552558899, + 0.3405662178993225, + 0.1505890041589737, + 0.054006874561309814, + 0.0228611808270216, + 0.014579183422029018 + ], + "credit_map": [ + 0.01774454489350319, + 0.009757230058312416, + 0.020047783851623535, + 0.009239861741662025, + 0.012746443971991539, + 0.021561067551374435, + 0.01150344219058752, + 0.025726573541760445, + 0.010188421234488487, + 0.004457395523786545, + 0.01468134205788374, + 0.003926490433514118, + 0.005015084519982338, + 0.015820994973182678, + 0.003800129983574152, + 0.02016061171889305, + 0.005706015974283218, + 0.0025762435980141163, + 0.006382565945386887, + 0.002036800142377615, + 0.003399089677259326, + 0.003775825025513768, + 0.0021259684581309557, + 0.005128726363182068, + 0.009901290759444237, + 0.004874917678534985, + 0.011245997622609138, + 0.004519200883805752, + 0.004500743933022022, + 0.0082554267719388, + 0.005229609087109566, + 0.013761881738901138, + 0.021318772807717323, + 0.011543438769876957, + 0.02075190283358097, + 0.01000775583088398, + 0.014405442401766777, + 0.014555992558598518, + 0.01488810870796442, + 0.010128348134458065, + 0.046604495495557785, + 0.01827714592218399, + 0.08533830940723419, + 0.021887561306357384, + 0.02077559381723404, + 0.09723131358623505, + 0.025086507201194763, + 0.025365279987454414, + 0.024438703432679176, + 0.010252424515783787, + 0.028475649654865265, + 0.016350576654076576, + 0.015106912702322006, + 0.026080871000885963, + 0.015325588174164295, + 0.014558281749486923, + 0.011721749790012836, + 0.003961244132369757, + 0.008251016028225422, + 0.00435117119923234, + 0.003936389461159706, + 0.00854603573679924, + 0.005108345299959183, + 0.008130927570164204, + 0.003785749664530158, + 0.001740973093546927, + 0.004880794323980808, + 0.0015401217387989163, + 0.001972829457372427, + 0.003433100413531065, + 0.002467576414346695, + 0.0030400347895920277, + 0.0020116539672017097, + 0.0015520789893344045, + 0.0020618976559489965, + 0.001497339690104127, + 0.002326370682567358, + 0.0019101996440440416, + 0.0015750539023429155, + 0.0016445887740701437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.06205242872238159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up meticulously depicts an imperfectly shaped cube of ice, resting directly on the polished, reflective surface of a dark granite countertop. The cube is translucent, its crystalline structure riddled with a delicate network of internal cracks and tiny, trapped bubbles. Its initially sharp edges gleam under a single, warm, focused spotlight emanating from an unseen source. At first, the cube begins to melt imperceptibly, then gradually; beads of condensation form and merge into small streams that flow down the cube's sides, forming a clear, shallow puddle that slowly expands from the bottom. As the action progresses, the cube's sharp geometry inexorably softens, its corners rounded, and its mass noticeably diminishing as it appears to float in its own meltwater, which spreads across the dark, smooth surface in ever-expanding, irregular shapes. The focused light creates brilliant, mirror-like highlights on the wet ice surface and the expanding puddle, while casting soft, distorted shadows that shift as the cube's shape shifts, creating a quiet, contemplative atmosphere of complete absorption in the mesmerizing and inevitable process of transformation.", + "seed": 30070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_4/index_6984_seed_30070_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09475998394191265, + "vlm_logprob": -0.09475998394191265, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.09475998394191265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05858302116394043 + ], + "opposite_logprob_tokens": [ + -2.9335830211639404 + ], + "logprob_reward": -0.05858302116394043, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05858302116394043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23576687276363373, + 0.3080700635910034, + 0.06079208105802536, + 0.04556869715452194, + 0.06760125607252121, + 0.035317856818437576, + 0.05840630456805229, + 0.12048343569040298, + 0.03996248543262482, + 0.028031010180711746 + ], + "credit_map": [ + 0.034005362540483475, + 0.019132884219288826, + 0.038456935435533524, + 0.02257559634745121, + 0.017416514456272125, + 0.028082897886633873, + 0.02217387594282627, + 0.05392277613282204, + 0.0635753944516182, + 0.026625726372003555, + 0.05210575461387634, + 0.024034803733229637, + 0.04026486724615097, + 0.022712374106049538, + 0.030797135084867477, + 0.04795395955443382, + 0.008291816338896751, + 0.009947563521564007, + 0.0074238223023712635, + 0.008000058121979237, + 0.009033444337546825, + 0.005389140918850899, + 0.006092195864766836, + 0.00661403639242053, + 0.006825446151196957, + 0.005281779449433088, + 0.0061798919923603535, + 0.004603604786098003, + 0.007732302416115999, + 0.004008125513792038, + 0.006236952729523182, + 0.004700587131083012, + 0.017665894702076912, + 0.006712312810122967, + 0.015489686280488968, + 0.005916627123951912, + 0.004280498716980219, + 0.004857337102293968, + 0.007306172978132963, + 0.0053727212361991405, + 0.006801206152886152, + 0.002761663869023323, + 0.0045476690866053104, + 0.003417240222916007, + 0.0030392762273550034, + 0.006622802931815386, + 0.003660357790067792, + 0.004467634484171867, + 0.00851610116660595, + 0.002721351571381092, + 0.005406080279499292, + 0.0028499357867985964, + 0.0027653477154672146, + 0.019100366160273552, + 0.0026588502805680037, + 0.014388266019523144, + 0.009298551827669144, + 0.006571229547262192, + 0.04488908499479294, + 0.00828761886805296, + 0.004694169852882624, + 0.025043267756700516, + 0.005256976466625929, + 0.01644253171980381, + 0.0025272781495004892, + 0.00406954251229763, + 0.007332551293075085, + 0.0043056984432041645, + 0.004126242361962795, + 0.007115012034773827, + 0.0038197487592697144, + 0.006666409317404032, + 0.004469277337193489, + 0.003383112605661154, + 0.003787721274420619, + 0.004010196775197983, + 0.0029101052787154913, + 0.0030758073553442955, + 0.0033641078043729067, + 0.0030306782573461533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.05858302116394043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054961517453193665 + ], + "opposite_logprob_tokens": [ + -2.9299614429473877 + ], + "logprob_reward": -0.054961517453193665, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054961517453193665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2854059338569641, + 0.2545433044433594, + 0.0874524712562561, + 0.06520535051822662, + 0.0758344754576683, + 0.0424983985722065, + 0.048442304134368896, + 0.06847487390041351, + 0.05390625819563866, + 0.018236562609672546 + ], + "credit_map": [ + 0.06335717439651489, + 0.01894908957183361, + 0.06923335045576096, + 0.0190895963460207, + 0.0158680509775877, + 0.03131324425339699, + 0.019324587658047676, + 0.04827084764838219, + 0.05485467612743378, + 0.019139884039759636, + 0.05673370510339737, + 0.016706842929124832, + 0.026696324348449707, + 0.020833821967244148, + 0.020103493705391884, + 0.039474599063396454, + 0.013685126788914204, + 0.010572414845228195, + 0.011637641116976738, + 0.011072150431573391, + 0.00896592065691948, + 0.01253129169344902, + 0.00737738236784935, + 0.011610559187829494, + 0.01092442125082016, + 0.008834578096866608, + 0.010669529438018799, + 0.008536126464605331, + 0.007412545382976532, + 0.005916096735745668, + 0.005564211402088404, + 0.007347852922976017, + 0.019752494990825653, + 0.0113176628947258, + 0.01556198950856924, + 0.006091499701142311, + 0.004556422121822834, + 0.007146462332457304, + 0.006295391824096441, + 0.005112556740641594, + 0.008490013889968395, + 0.003455282188951969, + 0.005626488011330366, + 0.0035065629053860903, + 0.0032228315249085426, + 0.007932609878480434, + 0.004257651977241039, + 0.006006961688399315, + 0.005947197321802378, + 0.003119315253570676, + 0.0039625926874578, + 0.00454488443210721, + 0.0025310604833066463, + 0.013185733929276466, + 0.003094396786764264, + 0.012057128362357616, + 0.004685870837420225, + 0.003185443114489317, + 0.02521219290792942, + 0.004505970515310764, + 0.002677586628124118, + 0.017056826502084732, + 0.0029834010638296604, + 0.008167590945959091, + 0.004267823416739702, + 0.007031031418591738, + 0.006085666362196207, + 0.0074114263989031315, + 0.006836299318820238, + 0.008831212297081947, + 0.005901412572711706, + 0.007541394792497158, + 0.002595234662294388, + 0.0017383095109835267, + 0.003459794679656625, + 0.0025029715616256, + 0.0019401380559429526, + 0.002040486317127943, + 0.0016803811304271221, + 0.002279249019920826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.054961517453193665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255980670452118 + ], + "opposite_logprob_tokens": [ + -1.6005980968475342 + ], + "logprob_reward": -0.2255980670452118, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255980670452118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3352883458137512, + 0.14090292155742645, + 0.0679517611861229, + 0.07823557406663895, + 0.08389067649841309, + 0.054030053317546844, + 0.09211261570453644, + 0.04486145079135895, + 0.04695691168308258, + 0.05576961860060692 + ], + "credit_map": [ + 0.052518803626298904, + 0.027923060581088066, + 0.05938149243593216, + 0.025534022599458694, + 0.024948688223958015, + 0.03754431754350662, + 0.031924180686473846, + 0.0755138173699379, + 0.028518633916974068, + 0.011270599439740181, + 0.030230557546019554, + 0.012740934267640114, + 0.015016560442745686, + 0.011871489696204662, + 0.01274166814982891, + 0.018512489274144173, + 0.01025056280195713, + 0.006734288763254881, + 0.010270931757986546, + 0.006291443947702646, + 0.013479073531925678, + 0.00803582463413477, + 0.006434164009988308, + 0.006455480586737394, + 0.013087655417621136, + 0.008123107254505157, + 0.01655210368335247, + 0.008721230551600456, + 0.009555540047585964, + 0.0075054955668747425, + 0.005881011951714754, + 0.008809435181319714, + 0.019413499161601067, + 0.009776701219379902, + 0.019780263304710388, + 0.007401836570352316, + 0.007414608262479305, + 0.0054559167474508286, + 0.007979284040629864, + 0.006668574642390013, + 0.008945642970502377, + 0.00527930585667491, + 0.007215582765638828, + 0.0054060365073382854, + 0.004987580236047506, + 0.010114084929227829, + 0.004795913118869066, + 0.007285913452506065, + 0.00993364304304123, + 0.004087263718247414, + 0.005887965206056833, + 0.0045684524811804295, + 0.00467927660793066, + 0.03494967147707939, + 0.004736115224659443, + 0.02327023819088936, + 0.0032533539924770594, + 0.0032910373993217945, + 0.0113169364631176, + 0.0050926776602864265, + 0.0027954448014497757, + 0.007824563421308994, + 0.00444247480481863, + 0.006844968535006046, + 0.004912266042083502, + 0.005426272749900818, + 0.009956303983926773, + 0.005748880561441183, + 0.00417497381567955, + 0.006124279927462339, + 0.004009169526398182, + 0.006604769267141819, + 0.006385127082467079, + 0.005684222094714642, + 0.011118462309241295, + 0.00733397388830781, + 0.005841297097504139, + 0.007086422760039568, + 0.005454556550830603, + 0.006865563802421093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.2255980670452118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628476247191429 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628476247191429, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628476247191429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24989533424377441, + 0.09380334615707397, + 0.1753964126110077, + 0.048339273780584335, + 0.09725508093833923, + 0.06466035544872284, + 0.0614757314324379, + 0.1266312450170517, + 0.04281362146139145, + 0.03972959890961647 + ], + "credit_map": [ + 0.03957055136561394, + 0.019132418558001518, + 0.041450195014476776, + 0.026947248727083206, + 0.024030260741710663, + 0.03357724845409393, + 0.02058349922299385, + 0.04460391774773598, + 0.017499279230833054, + 0.009028829634189606, + 0.02032683789730072, + 0.00844064075499773, + 0.005082759540528059, + 0.012367608025670052, + 0.006573590449988842, + 0.0144838010892272, + 0.02472888119518757, + 0.025833897292613983, + 0.028485100716352463, + 0.018362563103437424, + 0.024150852113962173, + 0.01845385693013668, + 0.016005542129278183, + 0.01937573030591011, + 0.00725523428991437, + 0.004562539514154196, + 0.008962887339293957, + 0.0064805918373167515, + 0.004141794051975012, + 0.005148594733327627, + 0.004488700069487095, + 0.007298931013792753, + 0.023216871544718742, + 0.011779793538153172, + 0.016691269353032112, + 0.008666183799505234, + 0.007313546258956194, + 0.007785899098962545, + 0.010293066501617432, + 0.011508448980748653, + 0.014188033528625965, + 0.005250521469861269, + 0.00867857038974762, + 0.005184120498597622, + 0.005074838642030954, + 0.013061410747468472, + 0.004404008854180574, + 0.00881885550916195, + 0.005105818621814251, + 0.002689097775146365, + 0.0093400739133358, + 0.0032975540962070227, + 0.0029091499745845795, + 0.023840967565774918, + 0.002084852661937475, + 0.012208218686282635, + 0.006717852782458067, + 0.004618937149643898, + 0.0350748710334301, + 0.005772126838564873, + 0.005118068773299456, + 0.049001093953847885, + 0.004300825297832489, + 0.016027478501200676, + 0.0029034544713795185, + 0.0032952744513750076, + 0.012780035845935345, + 0.004504051990807056, + 0.0026812718715518713, + 0.007119128480553627, + 0.0030919432174414396, + 0.006438463926315308, + 0.004746014252305031, + 0.005518623162060976, + 0.006505225319415331, + 0.003937581554055214, + 0.00578693114221096, + 0.004619757644832134, + 0.004030924290418625, + 0.00458454480394721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.007628476247191429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an imperfectly shaped, translucent cube of ice with internal cracks and bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, reflective dark granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the ice cube melting, with condensation forming streams and a puddle expanding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single warm focused spotlight creating mirror-like highlights and distorted shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static high-angle close-up throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702883780002594 + ], + "opposite_logprob_tokens": [ + -2.1270289421081543 + ], + "logprob_reward": -0.12702883780002594, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702883780002594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18053443729877472, + 0.03953075036406517, + 0.1482672244310379, + 0.07748109102249146, + 0.15747730433940887, + 0.18740755319595337, + 0.08057835698127747, + 0.0844963788986206, + 0.02288644015789032, + 0.021340422332286835 + ], + "credit_map": [ + 0.0318894125521183, + 0.020074263215065002, + 0.028883807361125946, + 0.016075126826763153, + 0.01544968318194151, + 0.01785067282617092, + 0.01939217746257782, + 0.030919304117560387, + 0.007425195537507534, + 0.004147508647292852, + 0.006149388384073973, + 0.0034035753924399614, + 0.004232261795550585, + 0.004593022167682648, + 0.0042039453983306885, + 0.005375859327614307, + 0.02134433388710022, + 0.022394640371203423, + 0.02220778912305832, + 0.015645461156964302, + 0.021521873772144318, + 0.015743527561426163, + 0.014024339616298676, + 0.015385271981358528, + 0.010639206506311893, + 0.007428095210343599, + 0.014435230754315853, + 0.008839495480060577, + 0.014043795876204967, + 0.006384890526533127, + 0.007953490130603313, + 0.007756893988698721, + 0.04098590090870857, + 0.02207438461482525, + 0.03611615300178528, + 0.01402868703007698, + 0.00965991709381342, + 0.012074804864823818, + 0.012788834981620312, + 0.009748631156980991, + 0.04073676839470863, + 0.014042917639017105, + 0.023643875494599342, + 0.01661897636950016, + 0.013952824287116528, + 0.038543764501810074, + 0.013165957294404507, + 0.02670248970389366, + 0.009706476703286171, + 0.005610897671431303, + 0.006655848119407892, + 0.0039809830486774445, + 0.004316514357924461, + 0.029561882838606834, + 0.0038521122187376022, + 0.016893655061721802, + 0.004715528804808855, + 0.0030614985153079033, + 0.02489343471825123, + 0.003545921528711915, + 0.003414105623960495, + 0.03259244188666344, + 0.002436983399093151, + 0.009836474433541298, + 0.0017633677925914526, + 0.003006383078172803, + 0.0035478146746754646, + 0.003969644196331501, + 0.0029136564116925, + 0.0031352234072983265, + 0.002336090663447976, + 0.002214262494817376, + 0.002373721683397889, + 0.0018179414328187704, + 0.003196415491402149, + 0.0024037822149693966, + 0.0034074594732373953, + 0.0024989829398691654, + 0.0021995508577674627, + 0.003442571498453617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.12702883780002594 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_tensor([8638, 8638])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_tensor([8638, 8638])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..78535a749f8f7a6b52e7014bbabde597d78612eb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_tensor([8638, 8638])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot begins from behind the subject, then seamlessly transitions into an extremely slow, sustained zoom out. At dusk, a lone, unidentified figure, slender and unidentified, dressed in a faded, dark gray hoodie with the hood pulled back from his head and dark trousers, runs vigorously along the center of two infinitely long, slightly rusted railroad tracks. With a determined, rhythmic gait, he swings his arms vigorously, leans forward, and places each step precisely on the cold railroad ties. The setting sun paints the sky a magnificent panorama of warm orange, soft pink, and deepening indigo. The low-angle light illuminates the two rails into long, shimmering golden lines, casting a long, solitary shadow behind him. As the camera steadily pulls back and up, his figure rapidly shrinks in size, from a running individual to a tiny black dot against the vast wilderness. Surrounded by a desolate wilderness of withered yellow weeds, the distant horizon almost merges with the vanishing point of the railroad tracks. Ultimately, this tiny figure is completely swallowed up by the vast sky and endless earth, disappearing completely from sight. Only the empty tracks and the gradually cooling twilight remain, creating a strong sense of fate, an endless journey, and a profound atmosphere of loneliness.", + "seed": 31652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_8638_seed_31652_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02099960520863533, + "vlm_logprob": -0.02099960520863533, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.02099960520863533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02069849707186222 + ], + "opposite_logprob_tokens": [ + -5.395698547363281 + ], + "logprob_reward": -0.02069849707186222, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.02069849707186222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06839720159769058, + 0.02870793826878071, + 0.023661138489842415, + 0.21791504323482513, + 0.0915580689907074, + 0.17775961756706238, + 0.15330207347869873, + 0.09334920346736908, + 0.072553351521492, + 0.07279638200998306 + ], + "credit_map": [ + 0.008562039583921432, + 0.00814273115247488, + 0.00677573774009943, + 0.007477724924683571, + 0.00873438362032175, + 0.013134929351508617, + 0.008399670012295246, + 0.007169988937675953, + 0.0035007521510124207, + 0.0033492993097752333, + 0.0031722141429781914, + 0.0030372673645615578, + 0.003882930614054203, + 0.0059177386574447155, + 0.003694084007292986, + 0.0021536522544920444, + 0.00296196760609746, + 0.0032193369697779417, + 0.0024548927322030067, + 0.0016886013327166438, + 0.002522589871659875, + 0.00509372865781188, + 0.0020140523556619883, + 0.003705967217683792, + 0.021851180121302605, + 0.062223806977272034, + 0.011804248206317425, + 0.017154840752482414, + 0.013560343533754349, + 0.06447916477918625, + 0.01662074215710163, + 0.010220691561698914, + 0.006747544277459383, + 0.026777876541018486, + 0.003971241880208254, + 0.004837408661842346, + 0.0034776886459439993, + 0.03356311097741127, + 0.0075494288466870785, + 0.004633770789951086, + 0.013737305998802185, + 0.05720217153429985, + 0.008047470822930336, + 0.008614958263933659, + 0.010005065239965916, + 0.059055086225271225, + 0.013633082620799541, + 0.007464469876140356, + 0.021690022200345993, + 0.04286162555217743, + 0.014795119874179363, + 0.010590997524559498, + 0.015737567096948624, + 0.028850391507148743, + 0.010322813875973225, + 0.00845352839678526, + 0.013093740679323673, + 0.023918163031339645, + 0.006837206892669201, + 0.007136064115911722, + 0.014372330158948898, + 0.012878132052719593, + 0.009712050668895245, + 0.005401511210948229, + 0.01048576645553112, + 0.021594630554318428, + 0.008406279608607292, + 0.0063160015270113945, + 0.007906089536845684, + 0.00674629770219326, + 0.005110231228172779, + 0.0059880539774894714, + 0.00925682857632637, + 0.02029048651456833, + 0.005787179339677095, + 0.0038757130969315767, + 0.007962721399962902, + 0.016329539939761162, + 0.006023976486176252, + 0.0032699331641197205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.02069849707186222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108088076114655 + ], + "opposite_logprob_tokens": [ + -3.2881081104278564 + ], + "logprob_reward": -0.038108088076114655, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108088076114655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0641600489616394, + 0.045235391706228256, + 0.030990876257419586, + 0.11453662812709808, + 0.01436106488108635, + 0.47481316328048706, + 0.049554064869880676, + 0.04255043715238571, + 0.09427551925182343, + 0.06952282041311264 + ], + "credit_map": [ + 0.005266705993562937, + 0.007162759080529213, + 0.008368391543626785, + 0.008970759809017181, + 0.007892887108027935, + 0.008359403349459171, + 0.010047403164207935, + 0.008091744035482407, + 0.005245199427008629, + 0.005272616166621447, + 0.0043134684674441814, + 0.003981748130172491, + 0.008751547895371914, + 0.00839249137789011, + 0.0055275470949709415, + 0.0037507733795791864, + 0.0025628635194152594, + 0.0029909585136920214, + 0.003038574941456318, + 0.0024279779754579067, + 0.008387505076825619, + 0.004563942085951567, + 0.0030038265977054834, + 0.0040152291767299175, + 0.01265737321227789, + 0.02720990777015686, + 0.008351718075573444, + 0.009484791196882725, + 0.010194573551416397, + 0.029384709894657135, + 0.009748377837240696, + 0.007505177520215511, + 0.0016171714523807168, + 0.002313146833330393, + 0.001388730015605688, + 0.001393118523992598, + 0.001518351025879383, + 0.0031735352240502834, + 0.0013946936232969165, + 0.0015623184153810143, + 0.05252420902252197, + 0.17327503859996796, + 0.012973280623555183, + 0.019460244104266167, + 0.021351326256990433, + 0.15332472324371338, + 0.03139134496450424, + 0.010513007640838623, + 0.008679777383804321, + 0.013742182403802872, + 0.0038205874152481556, + 0.003267487045377493, + 0.005205191671848297, + 0.007132330909371376, + 0.003629421116784215, + 0.004077089484781027, + 0.005592824891209602, + 0.011621755547821522, + 0.0029113348573446274, + 0.003301379969343543, + 0.00484250346198678, + 0.006657179445028305, + 0.004222757648676634, + 0.00340070528909564, + 0.010436688549816608, + 0.019960680976510048, + 0.01219795923680067, + 0.010230373591184616, + 0.01000277791172266, + 0.010121859610080719, + 0.008467373438179493, + 0.012857804074883461, + 0.008153428323566914, + 0.01790165901184082, + 0.008137984201312065, + 0.0069482759572565556, + 0.010654772631824017, + 0.0085735023021698, + 0.003935764543712139, + 0.005217432510107756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.038108088076114655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979855239391327 + ], + "opposite_logprob_tokens": [ + -3.5297985076904297 + ], + "logprob_reward": -0.02979855239391327, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979855239391327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037442196160554886, + 0.15456920862197876, + 0.014977449551224709, + 0.09014701843261719, + 0.021003136411309242, + 0.5324227809906006, + 0.019094642251729965, + 0.026637542992830276, + 0.04926380515098572, + 0.05444221571087837 + ], + "credit_map": [ + 0.004085139371454716, + 0.004528234247118235, + 0.005635206587612629, + 0.005056742578744888, + 0.005430434364825487, + 0.0050737615674734116, + 0.004379038233309984, + 0.003253635950386524, + 0.014326217584311962, + 0.01954270526766777, + 0.01524865161627531, + 0.012004618532955647, + 0.03166532516479492, + 0.029235823079943657, + 0.01618104986846447, + 0.016364797949790955, + 0.0018236769828945398, + 0.0018983401823788881, + 0.0012139210011810064, + 0.0009782613487914205, + 0.0037306661251932383, + 0.0029232443775981665, + 0.0013774947728961706, + 0.0010318445274606347, + 0.016646454110741615, + 0.025037894025444984, + 0.0075252787210047245, + 0.009097200818359852, + 0.007263946812599897, + 0.012475036084651947, + 0.006209457293152809, + 0.005891748704016209, + 0.0020515897776931524, + 0.002952607348561287, + 0.0018757105572149158, + 0.0021164934150874615, + 0.003554187249392271, + 0.0036481935530900955, + 0.0028602450620383024, + 0.0019441078184172511, + 0.05493655428290367, + 0.19276666641235352, + 0.016203783452510834, + 0.020957252010703087, + 0.023051977157592773, + 0.1757584512233734, + 0.03547082096338272, + 0.01327725313603878, + 0.0023544365540146828, + 0.002344781532883644, + 0.002164804143831134, + 0.0024674367159605026, + 0.0022728347685188055, + 0.0028832568787038326, + 0.001607897225767374, + 0.002999191638082266, + 0.002405901439487934, + 0.004518667235970497, + 0.0028406870551407337, + 0.0024258040357381105, + 0.003418001811951399, + 0.006975037977099419, + 0.002614394761621952, + 0.0014390451833605766, + 0.004340915009379387, + 0.009615757502615452, + 0.01038439106196165, + 0.0050302231684327126, + 0.004485799465328455, + 0.006696552038192749, + 0.003984908107668161, + 0.004725256469100714, + 0.004846665542572737, + 0.012208209373056889, + 0.005754453130066395, + 0.006103329826146364, + 0.008674749173223972, + 0.007474707439541817, + 0.004441355355083942, + 0.0049387430772185326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.02979855239391327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195535331964493 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195535331964493, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195535331964493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03170193359255791, + 0.014200735837221146, + 0.014188271015882492, + 0.2054971158504486, + 0.026156838983297348, + 0.44117650389671326, + 0.09443099796772003, + 0.052676063030958176, + 0.07369082421064377, + 0.04628075659275055 + ], + "credit_map": [ + 0.005648345220834017, + 0.0033969825599342585, + 0.0031398958526551723, + 0.0031736737582832575, + 0.003851846558973193, + 0.0054485551081597805, + 0.003713893238455057, + 0.0033287424594163895, + 0.0018181208288297057, + 0.0018283883109688759, + 0.0020151289645582438, + 0.0014583944575861096, + 0.001712640281766653, + 0.0023113854695111513, + 0.0015787662705406547, + 0.001477911020629108, + 0.002512221457436681, + 0.0015070628141984344, + 0.001665256917476654, + 0.0010687486501410604, + 0.002185204066336155, + 0.002628548536449671, + 0.0015053278766572475, + 0.0011159005807712674, + 0.021509915590286255, + 0.06895357370376587, + 0.00929923914372921, + 0.01075531356036663, + 0.008468604646623135, + 0.06480950862169266, + 0.014443849213421345, + 0.007257118355482817, + 0.0026678848080337048, + 0.0049271234311163425, + 0.002502604154869914, + 0.003433989593759179, + 0.003707226598635316, + 0.003651715349406004, + 0.0027122662868350744, + 0.0025540278293192387, + 0.04499000683426857, + 0.16269268095493317, + 0.01146840862929821, + 0.016726359724998474, + 0.01980249211192131, + 0.14750687777996063, + 0.028315957635641098, + 0.009673701599240303, + 0.017425602301955223, + 0.014460993930697441, + 0.00954587385058403, + 0.0070440988056361675, + 0.009812130592763424, + 0.022147679701447487, + 0.007267392706125975, + 0.006727226078510284, + 0.0053481594659388065, + 0.004573488142341375, + 0.0065043335780501366, + 0.0067918249405920506, + 0.008363324217498302, + 0.007692495360970497, + 0.008752715773880482, + 0.004649720620363951, + 0.006223307456821203, + 0.007815971039235592, + 0.008170952089130878, + 0.00680781714618206, + 0.008619106374680996, + 0.014888583682477474, + 0.009209314361214638, + 0.01195577159523964, + 0.0043476903811097145, + 0.011943317018449306, + 0.0039808922447264194, + 0.0034838588908314705, + 0.0062428186647593975, + 0.007727047894150019, + 0.0039047065656632185, + 0.004650423768907785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.014195535331964493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021010806784033775, + 0.016935696825385094, + 0.005244223866611719, + 0.10170494019985199, + 0.03635299205780029, + 0.5795429348945618, + 0.05952393636107445, + 0.02031789906322956, + 0.12511107325553894, + 0.03425547108054161 + ], + "credit_map": [ + 0.0031309265177696943, + 0.0022461144253611565, + 0.002449928317219019, + 0.002546857111155987, + 0.00211119232699275, + 0.0031760463025420904, + 0.003061052644625306, + 0.0022886875085532665, + 0.002002142136916518, + 0.0017641234444454312, + 0.001689219381660223, + 0.001422966830432415, + 0.0025134205352514982, + 0.0035703761968761683, + 0.001909928279928863, + 0.002063518390059471, + 0.00040399751742370427, + 0.0005677695735357702, + 0.0008372760494239628, + 0.0004233081708662212, + 0.0005952738574706018, + 0.001364571275189519, + 0.0005143712041899562, + 0.0005376560729928315, + 0.01766783744096756, + 0.026231953874230385, + 0.009554069489240646, + 0.010408214293420315, + 0.00926496647298336, + 0.014506060630083084, + 0.007804490625858307, + 0.006267343647778034, + 0.003838506294414401, + 0.011112954467535019, + 0.0017734899884089828, + 0.0022331788204610348, + 0.002458845963701606, + 0.010591386817395687, + 0.0023431244771927595, + 0.0020014995243400335, + 0.06158128380775452, + 0.2117498815059662, + 0.01668311096727848, + 0.021973200142383575, + 0.024602342396974564, + 0.19169792532920837, + 0.03799432888627052, + 0.013260849751532078, + 0.011995591223239899, + 0.01804877631366253, + 0.003325756872072816, + 0.003784158965572715, + 0.004344099201261997, + 0.01039609219878912, + 0.004304241389036179, + 0.0033252136781811714, + 0.00184816459659487, + 0.003891432425007224, + 0.0025001470930874348, + 0.0018670657882466912, + 0.0024325395934283733, + 0.0034325774759054184, + 0.0026219438295811415, + 0.0017240274464711547, + 0.01659514755010605, + 0.03595341369509697, + 0.01529958937317133, + 0.011041119694709778, + 0.011900932528078556, + 0.013137444853782654, + 0.010261332616209984, + 0.010922090150415897, + 0.0035120747052133083, + 0.00742931617423892, + 0.0034124141093343496, + 0.003521178150549531, + 0.0038284645415842533, + 0.006021616514772177, + 0.0033521929290145636, + 0.0031782116275280714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.002197353169322014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot begins from behind the subject, then seamlessly transitions into an extremely slow, sustained zoom out. At dusk, a lone, unidentified figure, slender and unidentified, dressed in a faded, dark gray hoodie with the hood pulled back from his head and dark trousers, runs vigorously along the center of two infinitely long, slightly rusted railroad tracks. With a determined, rhythmic gait, he swings his arms vigorously, leans forward, and places each step precisely on the cold railroad ties. The setting sun paints the sky a magnificent panorama of warm orange, soft pink, and deepening indigo. The low-angle light illuminates the two rails into long, shimmering golden lines, casting a long, solitary shadow behind him. As the camera steadily pulls back and up, his figure rapidly shrinks in size, from a running individual to a tiny black dot against the vast wilderness. Surrounded by a desolate wilderness of withered yellow weeds, the distant horizon almost merges with the vanishing point of the railroad tracks. Ultimately, this tiny figure is completely swallowed up by the vast sky and endless earth, disappearing completely from sight. Only the empty tracks and the gradually cooling twilight remain, creating a strong sense of fate, an endless journey, and a profound atmosphere of loneliness.", + "seed": 31653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_8638_seed_31653_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03372150254435837, + "vlm_logprob": -0.03372150254435837, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03372150254435837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023583153262734413 + ], + "opposite_logprob_tokens": [ + -4.398582935333252 + ], + "logprob_reward": -0.023583153262734413, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.023583153262734413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11648698896169662, + 0.2242727130651474, + 0.08867525309324265, + 0.05886290967464447, + 0.09339385479688644, + 0.1372491419315338, + 0.11048821359872818, + 0.10708789527416229, + 0.038400258868932724, + 0.025082753971219063 + ], + "credit_map": [ + 0.017022671177983284, + 0.010759321972727776, + 0.016126004979014397, + 0.011525409296154976, + 0.011075541377067566, + 0.013852187432348728, + 0.0239099208265543, + 0.01221594400703907, + 0.022647876292467117, + 0.0146824661642313, + 0.014555856585502625, + 0.04031999036669731, + 0.01960628107190132, + 0.019220322370529175, + 0.08173467218875885, + 0.01150527037680149, + 0.009213699959218502, + 0.010446439497172832, + 0.013461614958941936, + 0.00889409426599741, + 0.01639077626168728, + 0.010082590393722057, + 0.010606988333165646, + 0.009579059667885303, + 0.007888324558734894, + 0.006138679105788469, + 0.007560765836387873, + 0.006022448651492596, + 0.010941087268292904, + 0.010468008928000927, + 0.006062639411538839, + 0.003780955681577325, + 0.018225235864520073, + 0.015533347614109516, + 0.008548692800104618, + 0.005189650226384401, + 0.02015041932463646, + 0.01107175461947918, + 0.008829459547996521, + 0.005845308303833008, + 0.016025450080633163, + 0.03534868359565735, + 0.015045234002172947, + 0.008373203687369823, + 0.02072509564459324, + 0.018911970779299736, + 0.012234159745275974, + 0.010585361160337925, + 0.013828936964273453, + 0.03069355897605419, + 0.011828066781163216, + 0.008124012500047684, + 0.01270205620676279, + 0.016636665910482407, + 0.008922969922423363, + 0.007751955185085535, + 0.014810909517109394, + 0.023409705609083176, + 0.009053410962224007, + 0.008505938574671745, + 0.014718892052769661, + 0.013067969121038914, + 0.014097968116402626, + 0.009423107840120792, + 0.004812359344214201, + 0.005275098141282797, + 0.004503915086388588, + 0.00292428326793015, + 0.005912730935961008, + 0.00447192182764411, + 0.00603841058909893, + 0.0044615441001951694, + 0.0020540521945804358, + 0.0031611251179128885, + 0.003919729497283697, + 0.002549807308241725, + 0.003896898590028286, + 0.0026614784728735685, + 0.0034366557374596596, + 0.003403007984161377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.023583153262734413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898305356502533 + ], + "opposite_logprob_tokens": [ + -2.5789830684661865 + ], + "logprob_reward": -0.07898305356502533, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898305356502533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440456062555313, + 0.1917162537574768, + 0.2293788194656372, + 0.13518019020557404, + 0.13343319296836853, + 0.08242334425449371, + 0.029702749103307724, + 0.029686037451028824, + 0.013449493795633316, + 0.010984249413013458 + ], + "credit_map": [ + 0.017405256628990173, + 0.014412977732717991, + 0.016649050638079643, + 0.01990244910120964, + 0.009570358321070671, + 0.019084051251411438, + 0.031198520213365555, + 0.015822947025299072, + 0.02103954739868641, + 0.01183019857853651, + 0.012722104787826538, + 0.03399365022778511, + 0.015842560678720474, + 0.017515478655695915, + 0.06810048967599869, + 0.010672229342162609, + 0.018598556518554688, + 0.0150842210277915, + 0.034000691026449203, + 0.022203216329216957, + 0.043347444385290146, + 0.025007324293255806, + 0.02968878485262394, + 0.04144859313964844, + 0.020699698477983475, + 0.01644705981016159, + 0.008957904763519764, + 0.006776763126254082, + 0.017075717449188232, + 0.047974880784749985, + 0.010223680175840855, + 0.007024483289569616, + 0.014638460241258144, + 0.051603127270936966, + 0.00400602025911212, + 0.0029535521753132343, + 0.009106510318815708, + 0.04268503561615944, + 0.004732570145279169, + 0.0037079141475260258, + 0.010330845601856709, + 0.022488372400403023, + 0.006999178323894739, + 0.0037968922406435013, + 0.014984611421823502, + 0.012182041071355343, + 0.007526057306677103, + 0.004115346819162369, + 0.0029447264969348907, + 0.004753801506012678, + 0.003837442258372903, + 0.0026441060472279787, + 0.0056229084730148315, + 0.002971269888803363, + 0.0037017101421952248, + 0.0032267849892377853, + 0.0050109620206058025, + 0.006070578470826149, + 0.0031184314284473658, + 0.0016401233151555061, + 0.004972006194293499, + 0.003679115790873766, + 0.003217402845621109, + 0.0019774180836975574, + 0.001687395852059126, + 0.0015283646062016487, + 0.0020640073344111443, + 0.0013739651767536998, + 0.0014386889524757862, + 0.0017241068417206407, + 0.0019850796088576317, + 0.0016478844918310642, + 0.0013104546815156937, + 0.0016362256137654185, + 0.0012751271715387702, + 0.0011783158406615257, + 0.001606245874427259, + 0.0011610020883381367, + 0.001479042344726622, + 0.0013378357980400324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.07898305356502533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865546151995659 + ], + "opposite_logprob_tokens": [ + -3.0486555099487305 + ], + "logprob_reward": -0.04865546151995659, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865546151995659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14565275609493256, + 0.07276450842618942, + 0.2504025995731354, + 0.10426884144544601, + 0.11442279815673828, + 0.07420902699232101, + 0.09167400002479553, + 0.09070389717817307, + 0.04556708782911301, + 0.010334460064768791 + ], + "credit_map": [ + 0.0156658124178648, + 0.018196143209934235, + 0.015318591147661209, + 0.01813122071325779, + 0.02257579192519188, + 0.015408441424369812, + 0.029085857793688774, + 0.01127089187502861, + 0.007303548511117697, + 0.008118193596601486, + 0.006656008772552013, + 0.010903137736022472, + 0.011209411546587944, + 0.009632192552089691, + 0.01319731492549181, + 0.005744702648371458, + 0.030458709225058556, + 0.022336559370160103, + 0.03109794855117798, + 0.02516886219382286, + 0.05758896842598915, + 0.024615485221147537, + 0.02739105001091957, + 0.031745027750730515, + 0.018244702368974686, + 0.010260582901537418, + 0.007145628798753023, + 0.007363195065408945, + 0.011465055868029594, + 0.03526740148663521, + 0.008375297300517559, + 0.006146972067654133, + 0.015439982526004314, + 0.018537847325205803, + 0.013092956505715847, + 0.006958554033190012, + 0.024831311777234077, + 0.010995340533554554, + 0.013703477568924427, + 0.010863333009183407, + 0.01014138013124466, + 0.024588534608483315, + 0.0069503472186625, + 0.0035129194147884846, + 0.007010055240243673, + 0.013008889742195606, + 0.005504528526216745, + 0.0034923667553812265, + 0.008417594246566296, + 0.018394220620393753, + 0.0108225978910923, + 0.005992192775011063, + 0.01887763850390911, + 0.011023525148630142, + 0.011508714407682419, + 0.006637511774897575, + 0.009785394184291363, + 0.015332667157053947, + 0.007942885160446167, + 0.009173683822154999, + 0.017834236845374107, + 0.009921249933540821, + 0.010878029279410839, + 0.009835753589868546, + 0.00411264318972826, + 0.007131378166377544, + 0.004065599292516708, + 0.0030584472697228193, + 0.01054369006305933, + 0.004492876119911671, + 0.008198690600693226, + 0.003963765222579241, + 0.0009956699796020985, + 0.0021062325686216354, + 0.0009767879964783788, + 0.0009353237110190094, + 0.00199448736384511, + 0.0012533540138974786, + 0.0011225342750549316, + 0.0009500703890807927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.04865546151995659 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195652678608894 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195652678608894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18559154868125916, + 0.06667112559080124, + 0.06605593115091324, + 0.0999647006392479, + 0.21655313670635223, + 0.25680163502693176, + 0.048227693885564804, + 0.030680803582072258, + 0.02238789200782776, + 0.007065556477755308 + ], + "credit_map": [ + 0.018523184582591057, + 0.017668088898062706, + 0.02554379589855671, + 0.015117941424250603, + 0.030614353716373444, + 0.03557372838258743, + 0.028894538059830666, + 0.013655914925038815, + 0.0108161810785532, + 0.00455944100394845, + 0.0056145815178751945, + 0.00889470987021923, + 0.005689682904630899, + 0.009887858293950558, + 0.016753360629081726, + 0.004455311689525843, + 0.007920078001916409, + 0.0062864962965250015, + 0.008753656409680843, + 0.006233814172446728, + 0.013318744488060474, + 0.006521586794406176, + 0.010540713556110859, + 0.006480840500444174, + 0.010391922667622566, + 0.010400190949440002, + 0.00828542746603489, + 0.009005913510918617, + 0.026475928723812103, + 0.017694806680083275, + 0.010738419368863106, + 0.006972093600779772, + 0.014696372672915459, + 0.0789773091673851, + 0.007014357950538397, + 0.006539162714034319, + 0.011415884830057621, + 0.08295693248510361, + 0.007278416771441698, + 0.00767470384016633, + 0.03188750892877579, + 0.07018063217401505, + 0.02772841416299343, + 0.014456482604146004, + 0.04050499200820923, + 0.0360565260052681, + 0.021709833294153214, + 0.014277252368628979, + 0.005279630422592163, + 0.011561602354049683, + 0.004928201902657747, + 0.0039192670956254005, + 0.006954431999474764, + 0.006721061654388905, + 0.004763668403029442, + 0.004099832847714424, + 0.004729326814413071, + 0.005110172089189291, + 0.002869145479053259, + 0.003320830874145031, + 0.0039899940602481365, + 0.0032669594511389732, + 0.0038316305726766586, + 0.0035627437755465508, + 0.0028153869789093733, + 0.005147948861122131, + 0.0028212948236614466, + 0.0016462428029626608, + 0.002671749098226428, + 0.0031359668355435133, + 0.002028364222496748, + 0.002120936755090952, + 0.0007581082754768431, + 0.001736968057230115, + 0.0006313954945653677, + 0.000515829015057534, + 0.001074025989510119, + 0.0011949019972234964, + 0.0006878730491735041, + 0.0004664551524911076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.014195652678608894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031901916954666376 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.0031901916954666376, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031901916954666376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21530430018901825, + 0.0721418634057045, + 0.14979447424411774, + 0.17326685786247253, + 0.13484814763069153, + 0.13541573286056519, + 0.03680911660194397, + 0.024960774928331375, + 0.033239126205444336, + 0.024219635874032974 + ], + "credit_map": [ + 0.029459411278367043, + 0.02819538675248623, + 0.02655564248561859, + 0.030017875134944916, + 0.018841248005628586, + 0.02765948325395584, + 0.03507692366838455, + 0.019498324021697044, + 0.00612292904406786, + 0.006924754474312067, + 0.004760355688631535, + 0.010906349867582321, + 0.007388342637568712, + 0.009758503176271915, + 0.02150357887148857, + 0.004777042660862207, + 0.019123729318380356, + 0.013245604000985622, + 0.023094195872545242, + 0.019990917295217514, + 0.022219238802790642, + 0.014402219094336033, + 0.0177453625947237, + 0.019973214715719223, + 0.027069680392742157, + 0.0181744284927845, + 0.01811855472624302, + 0.013285125605762005, + 0.016606096178293228, + 0.05474374443292618, + 0.014325115829706192, + 0.010944114066660404, + 0.011135700158774853, + 0.046222880482673645, + 0.005605404265224934, + 0.003238365985453129, + 0.016974732279777527, + 0.03964564576745033, + 0.007446435280144215, + 0.0045789871364831924, + 0.016247570514678955, + 0.029806677252054214, + 0.013569515198469162, + 0.010038859210908413, + 0.02686968818306923, + 0.01638100855052471, + 0.013533850200474262, + 0.008968575857579708, + 0.004324076231569052, + 0.005806980654597282, + 0.005357616581022739, + 0.0035090718884021044, + 0.004912492353469133, + 0.006151693873107433, + 0.0028864936903119087, + 0.0038606938906013966, + 0.0035985996946692467, + 0.003201842773705721, + 0.0029824681114405394, + 0.0016160305822268128, + 0.004871620796620846, + 0.003441426670178771, + 0.003281863173469901, + 0.001966922078281641, + 0.00402037613093853, + 0.0060492041520774364, + 0.006509986240416765, + 0.002336799632757902, + 0.0024410246405750513, + 0.004644661210477352, + 0.0033857780508697033, + 0.0038512954488396645, + 0.0027039763517677784, + 0.003303087316453457, + 0.0029064342379570007, + 0.0028418286237865686, + 0.003168985014781356, + 0.0033206474035978317, + 0.0031652036122977734, + 0.0028094726148992777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0031901916954666376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot begins from behind the subject, then seamlessly transitions into an extremely slow, sustained zoom out. At dusk, a lone, unidentified figure, slender and unidentified, dressed in a faded, dark gray hoodie with the hood pulled back from his head and dark trousers, runs vigorously along the center of two infinitely long, slightly rusted railroad tracks. With a determined, rhythmic gait, he swings his arms vigorously, leans forward, and places each step precisely on the cold railroad ties. The setting sun paints the sky a magnificent panorama of warm orange, soft pink, and deepening indigo. The low-angle light illuminates the two rails into long, shimmering golden lines, casting a long, solitary shadow behind him. As the camera steadily pulls back and up, his figure rapidly shrinks in size, from a running individual to a tiny black dot against the vast wilderness. Surrounded by a desolate wilderness of withered yellow weeds, the distant horizon almost merges with the vanishing point of the railroad tracks. Ultimately, this tiny figure is completely swallowed up by the vast sky and endless earth, disappearing completely from sight. Only the empty tracks and the gradually cooling twilight remain, creating a strong sense of fate, an endless journey, and a profound atmosphere of loneliness.", + "seed": 31654, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_8638_seed_31654_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.032768393168225884, + "vlm_logprob": -0.032768393168225884, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.032768393168225884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018490446731448174 + ], + "opposite_logprob_tokens": [ + -5.143490314483643 + ], + "logprob_reward": -0.018490446731448174, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.018490446731448174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06202584132552147, + 0.022359080612659454, + 0.031336862593889236, + 0.07663898169994354, + 0.033975131809711456, + 0.1907719224691391, + 0.29143303632736206, + 0.1162705346941948, + 0.13198864459991455, + 0.043199993669986725 + ], + "credit_map": [ + 0.011976745910942554, + 0.008825480006635189, + 0.005199729464948177, + 0.006716637406498194, + 0.004086112603545189, + 0.004728465806692839, + 0.010445987805724144, + 0.010046684183180332, + 0.003456967184320092, + 0.0027599669992923737, + 0.002480939729139209, + 0.002278057159855962, + 0.0032097790390253067, + 0.0033377371728420258, + 0.0028858641162514687, + 0.0019497694447636604, + 0.003975827246904373, + 0.0034840726293623447, + 0.002380052115768194, + 0.0027773540932685137, + 0.00700040627270937, + 0.006425965111702681, + 0.0035333815030753613, + 0.0017598080448806286, + 0.018949301913380623, + 0.028133220970630646, + 0.0039327493868768215, + 0.0042115855030715466, + 0.004575354978442192, + 0.009531965479254723, + 0.004274588078260422, + 0.003030221676453948, + 0.004763671662658453, + 0.0054024588316679, + 0.003896518377587199, + 0.004152348730713129, + 0.003917355090379715, + 0.0060846139676868916, + 0.003307263134047389, + 0.002450902247801423, + 0.00765962665900588, + 0.07303944230079651, + 0.008342272602021694, + 0.005558283533900976, + 0.011327392421662807, + 0.0727418065071106, + 0.007989280857145786, + 0.004113811068236828, + 0.010756460949778557, + 0.055063582956790924, + 0.023379206657409668, + 0.01144498586654663, + 0.10913196951150894, + 0.05636097118258476, + 0.013838684186339378, + 0.011457185260951519, + 0.010401890613138676, + 0.018988093361258507, + 0.01342057902365923, + 0.01078048162162304, + 0.015988202765583992, + 0.021267546340823174, + 0.017316626384854317, + 0.008107110857963562, + 0.006069609429687262, + 0.04981314018368721, + 0.005619332194328308, + 0.004792739637196064, + 0.0471133254468441, + 0.008988750167191029, + 0.006062762811779976, + 0.0035289947409182787, + 0.00386147596873343, + 0.006142245139926672, + 0.004874828737229109, + 0.004746692720800638, + 0.007240359205752611, + 0.0067265271209180355, + 0.006359083112329245, + 0.0032487825956195593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.018490446731448174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003159821033478 + ], + "opposite_logprob_tokens": [ + -2.6950316429138184 + ], + "logprob_reward": -0.07003159821033478, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003159821033478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18992790579795837, + 0.046533260494470596, + 0.17228014767169952, + 0.043569598346948624, + 0.050476040691137314, + 0.17652827501296997, + 0.06458985805511475, + 0.1472432017326355, + 0.0888373851776123, + 0.020014381036162376 + ], + "credit_map": [ + 0.04234548285603523, + 0.021066082641482353, + 0.016120560467243195, + 0.018994521349668503, + 0.01589077152311802, + 0.01692351885139942, + 0.027499550953507423, + 0.03108738735318184, + 0.005771095864474773, + 0.006118451245129108, + 0.0049052247777581215, + 0.004210448823869228, + 0.0066886781714856625, + 0.007391079794615507, + 0.007557795383036137, + 0.0038904843386262655, + 0.036168064922094345, + 0.027214648202061653, + 0.012230695225298405, + 0.008846268057823181, + 0.01185871846973896, + 0.06178571656346321, + 0.0074279471300542355, + 0.006748082581907511, + 0.0058342935517430305, + 0.004818672779947519, + 0.004366499371826649, + 0.0037287285085767508, + 0.010555985383689404, + 0.006345789413899183, + 0.004549525678157806, + 0.003370097605511546, + 0.007253690622746944, + 0.007528768386691809, + 0.004252749960869551, + 0.006459804717451334, + 0.007736857049167156, + 0.008440262638032436, + 0.005120383575558662, + 0.0036835179198533297, + 0.006719481665641069, + 0.0706239715218544, + 0.003615021239966154, + 0.00422405730932951, + 0.010307295247912407, + 0.07130485028028488, + 0.006289237644523382, + 0.0034443330951035023, + 0.003912817221134901, + 0.010855861008167267, + 0.009059076197445393, + 0.004948425572365522, + 0.015017558820545673, + 0.007498750928789377, + 0.008380786515772343, + 0.004916575737297535, + 0.008463017642498016, + 0.04425307363271713, + 0.010760382749140263, + 0.008828939869999886, + 0.025912677869200706, + 0.016189318150281906, + 0.0235891193151474, + 0.009246645495295525, + 0.004543560091406107, + 0.025225205346941948, + 0.009070566855370998, + 0.006840913090854883, + 0.02290850505232811, + 0.008039352484047413, + 0.007099911104887724, + 0.00510936276987195, + 0.0024138870649039745, + 0.0020963484421372414, + 0.0025552972219884396, + 0.0023895357735455036, + 0.00255151791498065, + 0.002391289221122861, + 0.003197130747139454, + 0.002419373020529747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.07003159821033478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204582005739212 + ], + "opposite_logprob_tokens": [ + -2.8120458126068115 + ], + "logprob_reward": -0.06204582005739212, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204582005739212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13430091738700867, + 0.054129406809806824, + 0.07069471478462219, + 0.1121697798371315, + 0.024149324744939804, + 0.3068780303001404, + 0.028046797960996628, + 0.05178067833185196, + 0.1670132279396057, + 0.05083708092570305 + ], + "credit_map": [ + 0.02773071825504303, + 0.016539499163627625, + 0.011722414754331112, + 0.012870806269347668, + 0.012669120915234089, + 0.012554999440908432, + 0.019992612302303314, + 0.02022075280547142, + 0.006527534686028957, + 0.005645720288157463, + 0.006142038386315107, + 0.005379915703088045, + 0.007308146916329861, + 0.01102734636515379, + 0.007202768698334694, + 0.004895937163382769, + 0.009089755825698376, + 0.008645536378026009, + 0.004860675428062677, + 0.004794567823410034, + 0.01488781999796629, + 0.017380930483341217, + 0.006658874452114105, + 0.004376561380922794, + 0.026278194040060043, + 0.037356119602918625, + 0.007296576630324125, + 0.005203745793551207, + 0.009299924597144127, + 0.014085058122873306, + 0.008201300166547298, + 0.004448863677680492, + 0.002305838977918029, + 0.003535226220265031, + 0.0029370433185249567, + 0.003509428584948182, + 0.0032134803477674723, + 0.003299308242276311, + 0.003173787845298648, + 0.0021752119064331055, + 0.011450571939349174, + 0.12865810096263885, + 0.0034366711042821407, + 0.0049148621037602425, + 0.016065645962953568, + 0.12998078763484955, + 0.007663108874112368, + 0.004708313383162022, + 0.0022444212809205055, + 0.004246048629283905, + 0.005526203196495771, + 0.0020727659575641155, + 0.0049912696704268456, + 0.003872968489304185, + 0.0032605736050754786, + 0.0018325503915548325, + 0.0038962499238550663, + 0.0071664610877633095, + 0.004504525102674961, + 0.004956681281328201, + 0.008996274322271347, + 0.009274428710341454, + 0.008929375559091568, + 0.0040566879324615, + 0.006826713215559721, + 0.0572044774889946, + 0.01169644109904766, + 0.007804878056049347, + 0.056365568190813065, + 0.01371384784579277, + 0.006675055716186762, + 0.006726251915097237, + 0.005962932948023081, + 0.0073954821564257145, + 0.004387822467833757, + 0.00492162024602294, + 0.008748604916036129, + 0.009221717715263367, + 0.006948044989258051, + 0.0032508571166545153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.06204582005739212 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1024758368730545, + 0.023547165095806122, + 0.054277367889881134, + 0.09186824411153793, + 0.019421877339482307, + 0.21339206397533417, + 0.19060617685317993, + 0.07646247744560242, + 0.1379021257162094, + 0.09004667401313782 + ], + "credit_map": [ + 0.018677840009331703, + 0.012586052529513836, + 0.009564172476530075, + 0.010877796448767185, + 0.012296363711357117, + 0.00906619243323803, + 0.015276565216481686, + 0.014130854047834873, + 0.003669683588668704, + 0.002472266089171171, + 0.002888128627091646, + 0.0023518921807408333, + 0.0028789874631911516, + 0.002457511378452182, + 0.004032900556921959, + 0.0027957954443991184, + 0.009249993599951267, + 0.006255872081965208, + 0.005620692856609821, + 0.004363862797617912, + 0.007941706106066704, + 0.013135408982634544, + 0.004615187644958496, + 0.003094644285738468, + 0.018070444464683533, + 0.023092908784747124, + 0.009375328198075294, + 0.007392923813313246, + 0.009748758748173714, + 0.011894340626895428, + 0.0071558039635419846, + 0.005137730855494738, + 0.002083452418446541, + 0.0031256815418601036, + 0.0027204297948628664, + 0.0025398253928869963, + 0.0027506626211106777, + 0.0027399719692766666, + 0.0020178891718387604, + 0.0014439626829698682, + 0.008351928554475307, + 0.09191004931926727, + 0.0030758895445615053, + 0.002676123520359397, + 0.010800482705235481, + 0.09109651297330856, + 0.00238739768974483, + 0.0030936829280108213, + 0.0050545986741781235, + 0.037246305495500565, + 0.010257546789944172, + 0.006511902902275324, + 0.07713764160871506, + 0.04070518910884857, + 0.00752040883526206, + 0.00617258483543992, + 0.006552898325026035, + 0.014768019318580627, + 0.008016049861907959, + 0.006214533466845751, + 0.01316289696842432, + 0.010755419731140137, + 0.011074233800172806, + 0.005918428301811218, + 0.007810214068740606, + 0.032394956797361374, + 0.01437586359679699, + 0.011137314140796661, + 0.032532744109630585, + 0.013984211720526218, + 0.01521346066147089, + 0.010453351773321629, + 0.009994478896260262, + 0.01496502012014389, + 0.008554522879421711, + 0.007325645070523024, + 0.017622074112296104, + 0.014202314428985119, + 0.008306172676384449, + 0.00907644722610712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021975908894091845 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021975908894091845, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021975908894091845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03818404674530029, + 0.02113119326531887, + 0.023969285190105438, + 0.03602474927902222, + 0.03968299552798271, + 0.17474770545959473, + 0.07493389397859573, + 0.24070659279823303, + 0.33457598090171814, + 0.016043594107031822 + ], + "credit_map": [ + 0.0073917219415307045, + 0.004219874739646912, + 0.004688405431807041, + 0.003914792090654373, + 0.0023426478728652, + 0.0034810190554708242, + 0.005290613975375891, + 0.006854969542473555, + 0.003447531955316663, + 0.0021852629724889994, + 0.0028541036881506443, + 0.0023893516045063734, + 0.0024830636102706194, + 0.0033686060924082994, + 0.0024800824467092752, + 0.0019231920596212149, + 0.0034766446333378553, + 0.003135428996756673, + 0.0020658797584474087, + 0.0020322641357779503, + 0.00370308686979115, + 0.005978164728730917, + 0.002067662077024579, + 0.001510154572315514, + 0.005844591185450554, + 0.005923734046518803, + 0.005635838955640793, + 0.003350350074470043, + 0.004232622683048248, + 0.004533965140581131, + 0.003382427152246237, + 0.003121222136542201, + 0.004748031962662935, + 0.004259634297341108, + 0.005526632070541382, + 0.006752307526767254, + 0.005409664940088987, + 0.0051346952095627785, + 0.004530014004558325, + 0.0033220176119357347, + 0.006636185105890036, + 0.07260666787624359, + 0.0028928653337061405, + 0.0037351250648498535, + 0.009363912045955658, + 0.07397802174091339, + 0.002639512065798044, + 0.0028954080771654844, + 0.004756610374897718, + 0.013745543546974659, + 0.00775274308398366, + 0.005389702506363392, + 0.019609199836850166, + 0.010275062173604965, + 0.00815732590854168, + 0.005247712600976229, + 0.017580104991793633, + 0.06578933447599411, + 0.02448648028075695, + 0.015994159504771233, + 0.050833169370889664, + 0.025050317868590355, + 0.026765385642647743, + 0.014207644388079643, + 0.013565181754529476, + 0.12123975902795792, + 0.018427323549985886, + 0.013118051923811436, + 0.12338699400424957, + 0.025157663971185684, + 0.010172137059271336, + 0.009508861228823662, + 0.0018864109879359603, + 0.0019032404525205493, + 0.0016970253782346845, + 0.0015223531518131495, + 0.0026026854757219553, + 0.003094536019489169, + 0.0018958001164719462, + 0.0014415429905056953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.0021975908894091845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot begins from behind the subject, then seamlessly transitions into an extremely slow, sustained zoom out. At dusk, a lone, unidentified figure, slender and unidentified, dressed in a faded, dark gray hoodie with the hood pulled back from his head and dark trousers, runs vigorously along the center of two infinitely long, slightly rusted railroad tracks. With a determined, rhythmic gait, he swings his arms vigorously, leans forward, and places each step precisely on the cold railroad ties. The setting sun paints the sky a magnificent panorama of warm orange, soft pink, and deepening indigo. The low-angle light illuminates the two rails into long, shimmering golden lines, casting a long, solitary shadow behind him. As the camera steadily pulls back and up, his figure rapidly shrinks in size, from a running individual to a tiny black dot against the vast wilderness. Surrounded by a desolate wilderness of withered yellow weeds, the distant horizon almost merges with the vanishing point of the railroad tracks. Ultimately, this tiny figure is completely swallowed up by the vast sky and endless earth, disappearing completely from sight. Only the empty tracks and the gradually cooling twilight remain, creating a strong sense of fate, an endless journey, and a profound atmosphere of loneliness.", + "seed": 31655, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_5/index_8638_seed_31655_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024189228005707265, + "vlm_logprob": -0.024189228005707265, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.024189228005707265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027902014553546906 + ], + "opposite_logprob_tokens": [ + -5.402902126312256 + ], + "logprob_reward": -0.027902014553546906, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.027902014553546906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13539282977581024, + 0.09849930554628372, + 0.04521799832582474, + 0.06652023643255234, + 0.08067895472049713, + 0.047569070011377335, + 0.06460703164339066, + 0.0964578315615654, + 0.2523864209651947, + 0.11267031729221344 + ], + "credit_map": [ + 0.019227642565965652, + 0.01957191899418831, + 0.008505769073963165, + 0.012214886955916882, + 0.019970864057540894, + 0.020096655935049057, + 0.01833251677453518, + 0.017472578212618828, + 0.013480047695338726, + 0.014332525432109833, + 0.006636554375290871, + 0.00706463074311614, + 0.014440386556088924, + 0.018897224217653275, + 0.011747083626687527, + 0.011900849640369415, + 0.007705360651016235, + 0.005748205352574587, + 0.0038238526321947575, + 0.0036683869548141956, + 0.006952386349439621, + 0.005754283629357815, + 0.007192939054220915, + 0.004372580908238888, + 0.011663516983389854, + 0.008872436359524727, + 0.0033580290619283915, + 0.003979680128395557, + 0.008792010135948658, + 0.013202781789004803, + 0.010660546831786633, + 0.005991233978420496, + 0.012747210450470448, + 0.009216147474944592, + 0.0038584319408982992, + 0.006375249475240707, + 0.007084574084728956, + 0.022895121946930885, + 0.011967510916292667, + 0.006534704007208347, + 0.005901140160858631, + 0.005649518221616745, + 0.002802761271595955, + 0.003581370459869504, + 0.011005549691617489, + 0.006395266391336918, + 0.007451517973095179, + 0.004781944211572409, + 0.007540440186858177, + 0.009629614651203156, + 0.0034673649352043867, + 0.005920093972235918, + 0.007118802517652512, + 0.012872043997049332, + 0.010294439271092415, + 0.007764231413602829, + 0.012681900523602962, + 0.010287805460393429, + 0.004106316715478897, + 0.004757601767778397, + 0.014769041910767555, + 0.028939049690961838, + 0.0105876624584198, + 0.010328450240194798, + 0.023796159774065018, + 0.03113817423582077, + 0.011577898636460304, + 0.016139697283506393, + 0.07718189060688019, + 0.03609213978052139, + 0.039113253355026245, + 0.017347203567624092, + 0.016919173300266266, + 0.01159551553428173, + 0.006789105944335461, + 0.006055345293134451, + 0.020741667598485947, + 0.02167433127760887, + 0.01652132347226143, + 0.012373854406177998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.027902014553546906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033711470663547516 + ], + "opposite_logprob_tokens": [ + -3.4087114334106445 + ], + "logprob_reward": -0.033711470663547516, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033711470663547516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270807147026062, + 0.10000570863485336, + 0.0577654130756855, + 0.08904972672462463, + 0.11451990902423859, + 0.041004735976457596, + 0.06072121113538742, + 0.07192854583263397, + 0.06779735535383224, + 0.12640026211738586 + ], + "credit_map": [ + 0.036298852413892746, + 0.03171021118760109, + 0.013454054482281208, + 0.021654849871993065, + 0.044994283467531204, + 0.04213045537471771, + 0.04038018360733986, + 0.04018425941467285, + 0.011743159033358097, + 0.012807724066078663, + 0.004606197588145733, + 0.007860730402171612, + 0.020085442811250687, + 0.02273624576628208, + 0.012066112831234932, + 0.00810009054839611, + 0.007479310035705566, + 0.006012577563524246, + 0.005196096375584602, + 0.005062832031399012, + 0.011849482543766499, + 0.007971870712935925, + 0.009073777124285698, + 0.005119464825838804, + 0.010446401312947273, + 0.011575362645089626, + 0.0067574260756373405, + 0.006905633024871349, + 0.01928200013935566, + 0.01352672278881073, + 0.013852271251380444, + 0.006703914143145084, + 0.013983492739498615, + 0.015168138779699802, + 0.007208756636828184, + 0.00901490356773138, + 0.015276879072189331, + 0.027712807059288025, + 0.017947908490896225, + 0.008207013830542564, + 0.006518401205539703, + 0.005210007540881634, + 0.00213568564504385, + 0.0038206407334655523, + 0.006121173035353422, + 0.006706261076033115, + 0.007188333664089441, + 0.0033042319118976593, + 0.007270700763911009, + 0.007905196398496628, + 0.004017176106572151, + 0.004552823957055807, + 0.007254370022565126, + 0.017821289598941803, + 0.006970398593693972, + 0.004929251503199339, + 0.012816693633794785, + 0.010620600543916225, + 0.003864925354719162, + 0.003924299962818623, + 0.01034168154001236, + 0.016030197963118553, + 0.007841275073587894, + 0.006488869898021221, + 0.007443729322403669, + 0.008504264056682587, + 0.002352326875552535, + 0.0035344366915524006, + 0.01474529318511486, + 0.02005608379840851, + 0.007839532569050789, + 0.0033216841984540224, + 0.013952876441180706, + 0.011608703061938286, + 0.010024027898907661, + 0.0072828056290745735, + 0.02300998568534851, + 0.020167667418718338, + 0.029006799682974815, + 0.011347392573952675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.033711470663547516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979438751935959 + ], + "opposite_logprob_tokens": [ + -3.529794454574585 + ], + "logprob_reward": -0.02979438751935959, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979438751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42565613985061646, + 0.05280807241797447, + 0.06057434901595116, + 0.06305143237113953, + 0.06253625452518463, + 0.02956860326230526, + 0.04913629591464996, + 0.06809050589799881, + 0.04580879956483841, + 0.14276942610740662 + ], + "credit_map": [ + 0.038296882063150406, + 0.08715671300888062, + 0.02479238249361515, + 0.056572552770376205, + 0.07605704665184021, + 0.052791256457567215, + 0.05631694570183754, + 0.033672381192445755, + 0.007963540032505989, + 0.006407012231647968, + 0.003182116663083434, + 0.004107024986296892, + 0.008701435290277004, + 0.009307454340159893, + 0.00584960775449872, + 0.0072898888029158115, + 0.007583316415548325, + 0.008475060574710369, + 0.0032472293823957443, + 0.004445839673280716, + 0.014727137051522732, + 0.009740471839904785, + 0.008284621872007847, + 0.004070676397532225, + 0.005481140688061714, + 0.008139274083077908, + 0.003841803641989827, + 0.004776692949235439, + 0.016375306993722916, + 0.011469128541648388, + 0.008523778058588505, + 0.004444307181984186, + 0.006353936158120632, + 0.009223924949765205, + 0.004147016443312168, + 0.00564880296587944, + 0.010036378167569637, + 0.009851292707026005, + 0.011034537106752396, + 0.006240376271307468, + 0.0034538162872195244, + 0.00412476621568203, + 0.0014951471239328384, + 0.0022419572342187166, + 0.005317491479218006, + 0.006185329053550959, + 0.004363708663731813, + 0.002386390930041671, + 0.005861080717295408, + 0.005276472773402929, + 0.003434674581512809, + 0.005431183148175478, + 0.007588938809931278, + 0.006811445578932762, + 0.00942869670689106, + 0.005303810816258192, + 0.0072614653035998344, + 0.009308652952313423, + 0.00470319390296936, + 0.005618144758045673, + 0.013909651897847652, + 0.00857037864625454, + 0.011037501506507397, + 0.007681525312364101, + 0.006544992793351412, + 0.005238011479377747, + 0.0024985899217426777, + 0.00375408330000937, + 0.009754011407494545, + 0.005493043921887875, + 0.006358630955219269, + 0.006167438346892595, + 0.014688539318740368, + 0.01473388820886612, + 0.00890417117625475, + 0.012593013234436512, + 0.019532984122633934, + 0.02644515410065651, + 0.025986533612012863, + 0.01988515816628933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02979438751935959 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347124949097633 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347124949097633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347124949097633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07025246322154999, + 0.03415064141154289, + 0.06012840196490288, + 0.07382714003324509, + 0.030326038599014282, + 0.05036776140332222, + 0.06838817894458771, + 0.09298621118068695, + 0.11850849539041519, + 0.4010646343231201 + ], + "credit_map": [ + 0.008422289043664932, + 0.006803830619901419, + 0.0068419864401221275, + 0.006470849737524986, + 0.007866132073104382, + 0.009945034049451351, + 0.011657251976430416, + 0.012245092540979385, + 0.003673600498586893, + 0.004308999516069889, + 0.0028370190411806107, + 0.003828185610473156, + 0.005486097186803818, + 0.004022802226245403, + 0.004855250474065542, + 0.005138687323778868, + 0.007342323660850525, + 0.00924130342900753, + 0.004006727132946253, + 0.005247729364782572, + 0.007905120961368084, + 0.009218322113156319, + 0.010696354322135448, + 0.00647052563726902, + 0.010546712204813957, + 0.009699225425720215, + 0.0041669816710054874, + 0.004346157424151897, + 0.011010278016328812, + 0.014636224135756493, + 0.010169744491577148, + 0.009251815266907215, + 0.0045991684310138226, + 0.005337482318282127, + 0.0020240959711372852, + 0.002355639822781086, + 0.004058123100548983, + 0.0047592646442353725, + 0.004245292395353317, + 0.0029469712171703577, + 0.00890344101935625, + 0.008818531408905983, + 0.0027079428546130657, + 0.002971133915707469, + 0.009141725488007069, + 0.007404303643852472, + 0.00574516924098134, + 0.00467551639303565, + 0.007642186712473631, + 0.009523853659629822, + 0.00432167062535882, + 0.004866416100412607, + 0.012618159875273705, + 0.013523678295314312, + 0.009092516265809536, + 0.006799694150686264, + 0.010389383882284164, + 0.011569569818675518, + 0.00468492042273283, + 0.005421510897576809, + 0.01800919882953167, + 0.0193014033138752, + 0.015310216695070267, + 0.008300011977553368, + 0.01378642488270998, + 0.01304470282047987, + 0.004320414271205664, + 0.006480128038674593, + 0.023459261283278465, + 0.032084282487630844, + 0.01761419139802456, + 0.007719096727669239, + 0.05130500718951225, + 0.03785298764705658, + 0.0277569442987442, + 0.020819630473852158, + 0.06999292969703674, + 0.08448772877454758, + 0.07369287312030792, + 0.03515654057264328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.026347124949097633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot starting from a low angle behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a lone figure wearing a faded dark gray hoodie with the hood pulled back and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure running vigorously along the center of two slightly rusted railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera transition into an extremely slow, sustained zoom out while pulling back and up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate wilderness with withered yellow weeds and a sky painted in warm orange, soft pink, and deepening indigo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031911423429846764 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031911423429846764, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031911423429846764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18640202283859253, + 0.060342103242874146, + 0.04483235999941826, + 0.07330308109521866, + 0.14066772162914276, + 0.039713673293590546, + 0.11384523659944534, + 0.11322643607854843, + 0.07675442844629288, + 0.15091289579868317 + ], + "credit_map": [ + 0.0238743145018816, + 0.029361827298998833, + 0.010205226019024849, + 0.019907068461179733, + 0.02943991869688034, + 0.027232185006141663, + 0.02452581562101841, + 0.0218556746840477, + 0.00881149061024189, + 0.007199011743068695, + 0.0036697944160550833, + 0.005336611066013575, + 0.007802108768373728, + 0.012002094648778439, + 0.00643108831718564, + 0.009089915081858635, + 0.006732947193086147, + 0.006368372123688459, + 0.0027718665078282356, + 0.0039029582403600216, + 0.007391817402094603, + 0.007730334997177124, + 0.005327570717781782, + 0.004606497008353472, + 0.009411866776645184, + 0.010350562632083893, + 0.0046889628283679485, + 0.0070639485493302345, + 0.008991660550236702, + 0.013647886924445629, + 0.009899776428937912, + 0.009248415939509869, + 0.019487889483571053, + 0.023564236238598824, + 0.00901048630475998, + 0.01360057108104229, + 0.01410002913326025, + 0.025001341477036476, + 0.02423941157758236, + 0.01166378054767847, + 0.004965686239302158, + 0.0067738438956439495, + 0.0021491139195859432, + 0.0026019432116299868, + 0.007767189759761095, + 0.005533373448997736, + 0.005461723543703556, + 0.004460800904780626, + 0.01516707893460989, + 0.01918035000562668, + 0.01096617802977562, + 0.011524428613483906, + 0.012671336531639099, + 0.015371298417448997, + 0.016159553080797195, + 0.012805026024580002, + 0.015679923817515373, + 0.020604198798537254, + 0.00671403156593442, + 0.007077147252857685, + 0.016387049108743668, + 0.01911751925945282, + 0.016938401386141777, + 0.010708182118833065, + 0.011946350336074829, + 0.009867683984339237, + 0.002811447950080037, + 0.004726482089608908, + 0.011116390116512775, + 0.01994144357740879, + 0.008358984254300594, + 0.007985648699104786, + 0.01757829636335373, + 0.01637067086994648, + 0.009217771701514721, + 0.011492705903947353, + 0.020236307755112648, + 0.035255152732133865, + 0.027196230366826057, + 0.013565788976848125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0031911423429846764 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_tensor([9009, 9009])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_tensor([9009, 9009])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d38c864d381e52fd17add2b42702872b24fa5ee6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_tensor([9009, 9009])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot immerses the viewer in a mysterious nocturnal world, focusing with piercing clarity on a vibrant praying mantis occupying the foreground. The creature's body is a stunning canvas of color: its primary exoskeleton is a vibrant emerald green with a subtle waxy sheen, while its upper thorax and wing casings are adorned with intricate petal-like patterns of pure white and delicate pink. It perches on the edge of a large, black leaf, shimmering with tiny dewdrops. The mantis stalks its unseen prey with intense focus; its triangular head, dominated by large, multi-faceted eyes, rotates with meticulous slowness to track something off-screen. The entire scene is bathed in cool, low-key lighting, creating a color scheme of deep blue and velvety purple, casting the out-of-focus background of lush foliage into shadow and mystery. As the mantis prepares to move, it stealthily raises one of its serrated front legs, a movement so slow and deliberate as to be almost imperceptible, heightening the palpable tension and predatory focus of its nocturnal hunt.", + "seed": 15530, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_9009_seed_15530_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.053150149434804915, + "vlm_logprob": -0.053150149434804915, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.053150149434804915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005751015152782202 + ], + "opposite_logprob_tokens": [ + -5.255751132965088 + ], + "logprob_reward": -0.005751015152782202, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005751015152782202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16951884329319, + 0.021264271810650826, + 0.05626207962632179, + 0.03435959294438362, + 0.20108506083488464, + 0.09780220687389374, + 0.17780283093452454, + 0.02323668636381626, + 0.1806323230266571, + 0.038036126643419266 + ], + "credit_map": [ + 0.028981054201722145, + 0.019419901072978973, + 0.020092276856303215, + 0.017937252297997475, + 0.014337781816720963, + 0.03393777459859848, + 0.01660696230828762, + 0.018205838277935982, + 0.003350375220179558, + 0.0027265914250165224, + 0.0033271070569753647, + 0.0023883210960775614, + 0.0018053395906463265, + 0.0025276776868849993, + 0.001722846645861864, + 0.0034160141367465258, + 0.009641370736062527, + 0.008089207112789154, + 0.006711417343467474, + 0.0068659852258861065, + 0.008108249865472317, + 0.005285564810037613, + 0.006350244395434856, + 0.005210039671510458, + 0.007592510432004929, + 0.0027542009484022856, + 0.002789293648675084, + 0.003727699164301157, + 0.0026321171317249537, + 0.009835826233029366, + 0.0017926337895914912, + 0.00323531124740839, + 0.04538901522755623, + 0.02954949252307415, + 0.019374581053853035, + 0.024389658123254776, + 0.01540360040962696, + 0.020444152876734734, + 0.026633763685822487, + 0.019900798797607422, + 0.014904160052537918, + 0.026351308450102806, + 0.007249678485095501, + 0.008900437504053116, + 0.010007582604885101, + 0.009494693018496037, + 0.013186028227210045, + 0.007708316668868065, + 0.06036563217639923, + 0.011905818246304989, + 0.006994936615228653, + 0.006895578000694513, + 0.0068409498780965805, + 0.06973043829202652, + 0.007829315960407257, + 0.0072401766665279865, + 0.0044009387493133545, + 0.0027631307020783424, + 0.005314158275723457, + 0.0023219056893140078, + 0.0016592452302575111, + 0.0022895827423781157, + 0.0021799157839268446, + 0.002307810354977846, + 0.06280482560396194, + 0.013050098903477192, + 0.013379447162151337, + 0.0077501460909843445, + 0.007801140192896128, + 0.011261502280831337, + 0.010192937217652798, + 0.05439222976565361, + 0.004637307953089476, + 0.005044137593358755, + 0.004716348368674517, + 0.00564234983175993, + 0.004266390576958656, + 0.00520394928753376, + 0.004587017931044102, + 0.003938623704016209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.005751015152782202 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546183541417122 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012546183541417122, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546183541417122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24873076379299164, + 0.02987702563405037, + 0.059243083000183105, + 0.019904373213648796, + 0.028774444013834, + 0.04936474561691284, + 0.18450818955898285, + 0.02493875101208687, + 0.31540021300315857, + 0.03925842046737671 + ], + "credit_map": [ + 0.07152104377746582, + 0.019770938903093338, + 0.01559369545429945, + 0.01519680954515934, + 0.01407037302851677, + 0.08650467544794083, + 0.014717780984938145, + 0.011355435475707054, + 0.004638765472918749, + 0.004013220779597759, + 0.004004030022770166, + 0.00355204613879323, + 0.003127720672637224, + 0.003112945007160306, + 0.0023816393222659826, + 0.0050466591492295265, + 0.012298577465116978, + 0.007786821108311415, + 0.007260183338075876, + 0.009525323286652565, + 0.005183481611311436, + 0.005369212478399277, + 0.005435336381196976, + 0.006384146399796009, + 0.0036656090524047613, + 0.0019184881821274757, + 0.0019474462606012821, + 0.0034372189547866583, + 0.0017083940329030156, + 0.004585042595863342, + 0.001422240980900824, + 0.0012199326884001493, + 0.005498419050127268, + 0.0039672343991696835, + 0.0038046801928430796, + 0.003736047074198723, + 0.0031404602341353893, + 0.0028146596159785986, + 0.003401185153052211, + 0.0024117599241435528, + 0.0082825543358922, + 0.012166695669293404, + 0.0036505190655589104, + 0.004376769531518221, + 0.0046035246923565865, + 0.005496324971318245, + 0.0063380287028849125, + 0.00445032911375165, + 0.06417106837034225, + 0.010351810604333878, + 0.007450364995747805, + 0.005651028826832771, + 0.005406243726611137, + 0.07768897712230682, + 0.006913542281836271, + 0.006875146646052599, + 0.004400516394525766, + 0.003276312258094549, + 0.004967018961906433, + 0.002680242760106921, + 0.001617837231606245, + 0.003076039021834731, + 0.0023790232371538877, + 0.002541762078180909, + 0.11096843332052231, + 0.022154057398438454, + 0.022998034954071045, + 0.013391278684139252, + 0.013532271608710289, + 0.019565127789974213, + 0.017781708389520645, + 0.09500934928655624, + 0.0045165070332586765, + 0.005214344244450331, + 0.004782659932971001, + 0.004629815928637981, + 0.0058586131781339645, + 0.0060842642560601234, + 0.004478756804019213, + 0.003693459089845419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.012546183541417122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305894672870636 + ], + "opposite_logprob_tokens": [ + -3.1680588722229004 + ], + "logprob_reward": -0.04305894672870636, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305894672870636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2732014060020447, + 0.022938555106520653, + 0.04143071919679642, + 0.07790021598339081, + 0.10038349777460098, + 0.12650233507156372, + 0.1468803584575653, + 0.07264934480190277, + 0.10958575457334518, + 0.02852782793343067 + ], + "credit_map": [ + 0.08645717054605484, + 0.017497839406132698, + 0.016012096777558327, + 0.012972884811460972, + 0.012993546202778816, + 0.10198801010847092, + 0.013353330083191395, + 0.011926534585654736, + 0.0035075200721621513, + 0.0035244456958025694, + 0.002966325730085373, + 0.002569482196122408, + 0.0024998909793794155, + 0.002429528860375285, + 0.002000680658966303, + 0.003440679283812642, + 0.006851480342447758, + 0.007358288858085871, + 0.005525740794837475, + 0.004364830441772938, + 0.005302665755152702, + 0.004205354489386082, + 0.004222718067467213, + 0.0035996397491544485, + 0.019104717299342155, + 0.005796249955892563, + 0.006575106177479029, + 0.006999323144555092, + 0.005815494805574417, + 0.026730569079518318, + 0.002860485343262553, + 0.004018272273242474, + 0.020364485681056976, + 0.014838592149317265, + 0.012698166072368622, + 0.012791644781827927, + 0.008543060161173344, + 0.009348616003990173, + 0.01211391668766737, + 0.009685016237199306, + 0.018904322758316994, + 0.03444661200046539, + 0.009972386993467808, + 0.011882573366165161, + 0.011892040260136127, + 0.013454957865178585, + 0.016013873741030693, + 0.009935564361512661, + 0.04920453578233719, + 0.0077562169171869755, + 0.007481698412448168, + 0.004788696765899658, + 0.004512757062911987, + 0.06151316687464714, + 0.0055760894902050495, + 0.006047199014574289, + 0.01388976164162159, + 0.006135314237326384, + 0.022502198815345764, + 0.007099604234099388, + 0.0036750740837305784, + 0.008487490005791187, + 0.005782251711934805, + 0.005077647045254707, + 0.03777896612882614, + 0.007910148240625858, + 0.0080634243786335, + 0.004907661583274603, + 0.004962322302162647, + 0.0067382813431322575, + 0.006316797342151403, + 0.03290815278887749, + 0.003480536863207817, + 0.00338594033382833, + 0.0031651859171688557, + 0.004183500539511442, + 0.0035172952339053154, + 0.003305182559415698, + 0.0040010735392570496, + 0.003489113412797451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.04305894672870636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157784223556519 + ], + "opposite_logprob_tokens": [ + -1.70157790184021 + ], + "logprob_reward": -0.20157784223556519, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157784223556519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11703764647245407, + 0.025179842486977577, + 0.045911625027656555, + 0.015100772492587566, + 0.021237773820757866, + 0.028776489198207855, + 0.23980942368507385, + 0.01849953643977642, + 0.3853274881839752, + 0.10311947762966156 + ], + "credit_map": [ + 0.03471747040748596, + 0.008608627133071423, + 0.007067773025482893, + 0.006513148546218872, + 0.008372669108211994, + 0.04030941426753998, + 0.0065700397826731205, + 0.004878494888544083, + 0.004872176330536604, + 0.0029159290716052055, + 0.002665546489879489, + 0.0019851711113005877, + 0.002439772477373481, + 0.0039822375401854515, + 0.002005519112572074, + 0.004313488956540823, + 0.0061063822358846664, + 0.009485429152846336, + 0.004920443054288626, + 0.005999157205224037, + 0.006114579271525145, + 0.004729731474071741, + 0.004832759033888578, + 0.003723140573129058, + 0.002757769776508212, + 0.0016627887962386012, + 0.0019661581609398127, + 0.002581593580543995, + 0.0012942001922056079, + 0.0020292799454182386, + 0.001323803560808301, + 0.0014851774321869016, + 0.003978123422712088, + 0.003092875936999917, + 0.002122059464454651, + 0.0030973521061241627, + 0.0016151232412084937, + 0.0018297286005690694, + 0.0032707813661545515, + 0.002231728285551071, + 0.00411092396825552, + 0.0073594143614172935, + 0.0016619995003566146, + 0.003172958269715309, + 0.0031794689130038023, + 0.0030454425141215324, + 0.003966637887060642, + 0.0022796415723860264, + 0.08363771438598633, + 0.013547304086387157, + 0.009406210854649544, + 0.007682535331696272, + 0.006916779093444347, + 0.10055510699748993, + 0.00888812355697155, + 0.00917564332485199, + 0.003421876346692443, + 0.0018473940435796976, + 0.0053821736946702, + 0.001971765421330929, + 0.0010361061431467533, + 0.0015462302835658193, + 0.001993888057768345, + 0.0013001005863770843, + 0.13567723333835602, + 0.02719295583665371, + 0.027716802433133125, + 0.01640608161687851, + 0.01646043360233307, + 0.02382168360054493, + 0.021591592580080032, + 0.11646067351102829, + 0.01234507281333208, + 0.013707148842513561, + 0.011121049523353577, + 0.011690972372889519, + 0.016383521258831024, + 0.015539185144007206, + 0.012302757240831852, + 0.010029759258031845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.20157784223556519 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816759515553713 + ], + "opposite_logprob_tokens": [ + -5.877816677093506 + ], + "logprob_reward": -0.002816759515553713, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816759515553713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49892932176589966, + 0.04452255368232727, + 0.04018757492303848, + 0.057600975036621094, + 0.08741936087608337, + 0.0837910994887352, + 0.03521336615085602, + 0.09066321700811386, + 0.023848749697208405, + 0.03782385587692261 + ], + "credit_map": [ + 0.1270035356283188, + 0.04121549800038338, + 0.04125440865755081, + 0.037659961730241776, + 0.035972725600004196, + 0.15511244535446167, + 0.032489970326423645, + 0.02822074107825756, + 0.008277983404695988, + 0.00535466056317091, + 0.005336981266736984, + 0.004149383865296841, + 0.005349235143512487, + 0.005677972454577684, + 0.003687357297167182, + 0.006688977126032114, + 0.00541440024971962, + 0.004490294959396124, + 0.00575324147939682, + 0.005523166619241238, + 0.004587809555232525, + 0.004166947212070227, + 0.005647317972034216, + 0.004604397341609001, + 0.014410099945962429, + 0.004173130728304386, + 0.004400956444442272, + 0.005793932359665632, + 0.004364064894616604, + 0.018647221848368645, + 0.002585167996585369, + 0.0032263940665870905, + 0.01549325231462717, + 0.013320772908627987, + 0.014487550593912601, + 0.01166245061904192, + 0.0066229114308953285, + 0.008068246766924858, + 0.008480728603899479, + 0.009283437393605709, + 0.01313393097370863, + 0.02062872424721718, + 0.007125552278012037, + 0.00786877702921629, + 0.00771172484382987, + 0.00877121090888977, + 0.011179275810718536, + 0.007371899206191301, + 0.007360019721090794, + 0.0037821021396666765, + 0.003147368784993887, + 0.0020500801037997007, + 0.0025150266010314226, + 0.010166974738240242, + 0.003611366730183363, + 0.002580423140898347, + 0.017499273642897606, + 0.007101382594555616, + 0.02274210937321186, + 0.009469437412917614, + 0.004948388319462538, + 0.01365814357995987, + 0.007167880889028311, + 0.008076597936451435, + 0.003284454345703125, + 0.0048985653556883335, + 0.0025629911106079817, + 0.0022908400278538465, + 0.002657633274793625, + 0.002635359298437834, + 0.0020401696674525738, + 0.0034787352196872234, + 0.005536430981010199, + 0.004162659868597984, + 0.005824585445225239, + 0.005308032035827637, + 0.002941236598417163, + 0.004684271290898323, + 0.004982893355190754, + 0.004383740946650505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.002816759515553713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot immerses the viewer in a mysterious nocturnal world, focusing with piercing clarity on a vibrant praying mantis occupying the foreground. The creature's body is a stunning canvas of color: its primary exoskeleton is a vibrant emerald green with a subtle waxy sheen, while its upper thorax and wing casings are adorned with intricate petal-like patterns of pure white and delicate pink. It perches on the edge of a large, black leaf, shimmering with tiny dewdrops. The mantis stalks its unseen prey with intense focus; its triangular head, dominated by large, multi-faceted eyes, rotates with meticulous slowness to track something off-screen. The entire scene is bathed in cool, low-key lighting, creating a color scheme of deep blue and velvety purple, casting the out-of-focus background of lush foliage into shadow and mystery. As the mantis prepares to move, it stealthily raises one of its serrated front legs, a movement so slow and deliberate as to be almost imperceptible, heightening the palpable tension and predatory focus of its nocturnal hunt.", + "seed": 15531, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_9009_seed_15531_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07027485789731144, + "vlm_logprob": -0.07027485789731144, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.07027485789731144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02211574651300907 + ], + "opposite_logprob_tokens": [ + -3.897115707397461 + ], + "logprob_reward": -0.02211574651300907, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02211574651300907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12686419486999512, + 0.08165187388658524, + 0.02514786273241043, + 0.061351191252470016, + 0.2750145494937897, + 0.03338926285505295, + 0.09405526518821716, + 0.08538050204515457, + 0.07417799532413483, + 0.14296728372573853 + ], + "credit_map": [ + 0.015868550166487694, + 0.012652735225856304, + 0.012800006195902824, + 0.0182186346501112, + 0.014050069265067577, + 0.020652571693062782, + 0.01586703024804592, + 0.01675461232662201, + 0.012812292203307152, + 0.012877789326012135, + 0.00825153011828661, + 0.009180490858852863, + 0.008747964166104794, + 0.00825069285929203, + 0.0075156125240027905, + 0.014015506021678448, + 0.0031489089597016573, + 0.0024462519213557243, + 0.0024725075345486403, + 0.003532072762027383, + 0.0025657478254288435, + 0.004355757497251034, + 0.0028506165836006403, + 0.003776001278311014, + 0.004877799656242132, + 0.007009967230260372, + 0.004500007722526789, + 0.010117379017174244, + 0.008302331902086735, + 0.007247077766805887, + 0.010477309115231037, + 0.008819323033094406, + 0.020835569128394127, + 0.020207645371556282, + 0.021561864763498306, + 0.04254525899887085, + 0.05874871835112572, + 0.0309296865016222, + 0.038415420800447464, + 0.04177039861679077, + 0.0036663447972387075, + 0.0029159989207983017, + 0.003209020709618926, + 0.005482418462634087, + 0.002712791319936514, + 0.0034060808829963207, + 0.005688127595931292, + 0.006308484822511673, + 0.009543280117213726, + 0.008621479384601116, + 0.01390085183084011, + 0.008085522800683975, + 0.014897755347192287, + 0.012677934020757675, + 0.014899492263793945, + 0.011428951285779476, + 0.006865161471068859, + 0.007731013000011444, + 0.007016303483396769, + 0.010187157429754734, + 0.009317860007286072, + 0.012478352524340153, + 0.013958859257400036, + 0.017825806513428688, + 0.004779520910233259, + 0.0062974607571959496, + 0.007608611136674881, + 0.007767665199935436, + 0.008130237460136414, + 0.01251280214637518, + 0.014603981748223305, + 0.012477720156311989, + 0.011299003846943378, + 0.012246646918356419, + 0.018653543666005135, + 0.011208186857402325, + 0.011883712373673916, + 0.032608721405267715, + 0.02151741087436676, + 0.02355005592107773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.02211574651300907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060367725789547 + ], + "opposite_logprob_tokens": [ + -3.895603656768799 + ], + "logprob_reward": -0.02060367725789547, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060367725789547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0884043276309967, + 0.06330274045467377, + 0.055590081959962845, + 0.058267150074243546, + 0.41737017035484314, + 0.03116084262728691, + 0.08255818486213684, + 0.06208622828125954, + 0.0466679185628891, + 0.09459231793880463 + ], + "credit_map": [ + 0.011758573353290558, + 0.011248435825109482, + 0.010222791694104671, + 0.01262497529387474, + 0.009239674545824528, + 0.01376595813781023, + 0.00785565935075283, + 0.011688251048326492, + 0.01230466365814209, + 0.007811828516423702, + 0.0071207499131560326, + 0.005683045834302902, + 0.005951893050223589, + 0.006713546812534332, + 0.005947765428572893, + 0.011769238859415054, + 0.007880100049078465, + 0.004864176269620657, + 0.003897619666531682, + 0.006346824113279581, + 0.005346136633306742, + 0.009288119152188301, + 0.005863362457603216, + 0.012103741057217121, + 0.0054413750767707825, + 0.005093410145491362, + 0.005507597699761391, + 0.0053498875349760056, + 0.008344976231455803, + 0.010727410204708576, + 0.008114353753626347, + 0.009688131511211395, + 0.03178952634334564, + 0.029744798317551613, + 0.03307923302054405, + 0.060102496296167374, + 0.0922236517071724, + 0.047738127410411835, + 0.060324784368276596, + 0.06236753612756729, + 0.0032808417454361916, + 0.002599857747554779, + 0.0019613606855273247, + 0.006030535791069269, + 0.0020899521186947823, + 0.0036327987909317017, + 0.006082913838326931, + 0.005482577718794346, + 0.006330227013677359, + 0.008867830038070679, + 0.009655768983066082, + 0.01041601225733757, + 0.010235770605504513, + 0.013295472599565983, + 0.013156543485820293, + 0.01060054823756218, + 0.007655505556613207, + 0.007423977367579937, + 0.006741492077708244, + 0.005136304069310427, + 0.006627687253057957, + 0.009303072467446327, + 0.009118936024606228, + 0.010079249739646912, + 0.003924018703401089, + 0.0047066169790923595, + 0.005734897218644619, + 0.004505606833845377, + 0.006055004894733429, + 0.00798073597252369, + 0.00761076295748353, + 0.006150273140519857, + 0.009017092175781727, + 0.008395770564675331, + 0.011583179235458374, + 0.008041853085160255, + 0.009072315879166126, + 0.020667674019932747, + 0.013652993366122246, + 0.01416142750531435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.02060367725789547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031517595052719 + ], + "opposite_logprob_tokens": [ + -2.3503150939941406 + ], + "logprob_reward": -0.10031517595052719, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031517595052719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09154878556728363, + 0.05077109485864639, + 0.08279721438884735, + 0.045775189995765686, + 0.4193013906478882, + 0.043199311941862106, + 0.06844799220561981, + 0.08129318058490753, + 0.046215541660785675, + 0.0706503614783287 + ], + "credit_map": [ + 0.01052228081971407, + 0.009867659769952297, + 0.009295362047851086, + 0.011250089854001999, + 0.009301752783358097, + 0.01569286920130253, + 0.011382085271179676, + 0.014236685819923878, + 0.006540924310684204, + 0.007202718872576952, + 0.005285666324198246, + 0.006994984112679958, + 0.004416176583617926, + 0.007423229515552521, + 0.005017891526222229, + 0.00788949616253376, + 0.010231207124888897, + 0.0072321281768381596, + 0.007025651168078184, + 0.013695493340492249, + 0.009221083484590054, + 0.011833042837679386, + 0.011307216249406338, + 0.012251388281583786, + 0.003966065123677254, + 0.0030700990464538336, + 0.0042601111344993114, + 0.005669579841196537, + 0.0056703356094658375, + 0.01066821999847889, + 0.005507790949195623, + 0.006962982472032309, + 0.03142400085926056, + 0.029471708461642265, + 0.033393919467926025, + 0.06111614778637886, + 0.09046309441328049, + 0.05055389180779457, + 0.059310413897037506, + 0.06356816738843918, + 0.004136907868087292, + 0.002983519108965993, + 0.0031652781181037426, + 0.00750712538138032, + 0.0032999394461512566, + 0.004882997367531061, + 0.009453441947698593, + 0.007770093623548746, + 0.009244894608855247, + 0.005264557898044586, + 0.0064549753442406654, + 0.011757303029298782, + 0.005070915911346674, + 0.009535453282296658, + 0.008726559579372406, + 0.012393329292535782, + 0.01203618198633194, + 0.009807048365473747, + 0.007615780457854271, + 0.007393003907054663, + 0.009247411042451859, + 0.011561401188373566, + 0.009982757270336151, + 0.01364959217607975, + 0.004201618488878012, + 0.004023355431854725, + 0.005699168890714645, + 0.00472302595153451, + 0.006401462014764547, + 0.007316296920180321, + 0.007591969799250364, + 0.006258643697947264, + 0.006182552315294743, + 0.005889487452805042, + 0.008440492674708366, + 0.006418671924620867, + 0.007171037141233683, + 0.014191452413797379, + 0.011016729287803173, + 0.011339938268065453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.10031517595052719 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016032636165619 + ], + "opposite_logprob_tokens": [ + -1.7016032934188843 + ], + "logprob_reward": -0.2016032636165619, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016032636165619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1468961536884308, + 0.07064846903085709, + 0.09038512408733368, + 0.06480613350868225, + 0.15552005171775818, + 0.10316309332847595, + 0.14333559572696686, + 0.05352821573615074, + 0.04259872063994408, + 0.12911845743656158 + ], + "credit_map": [ + 0.015017200261354446, + 0.012832420878112316, + 0.01654643751680851, + 0.01654856652021408, + 0.015113623812794685, + 0.024053286761045456, + 0.01828647032380104, + 0.02849813736975193, + 0.008399791084229946, + 0.011596993543207645, + 0.007632260676473379, + 0.008138864301145077, + 0.007236028090119362, + 0.00921841710805893, + 0.008268251083791256, + 0.010157860815525055, + 0.011295674368739128, + 0.009337362833321095, + 0.008604355156421661, + 0.0136449309065938, + 0.011606397107243538, + 0.008975829929113388, + 0.015212679281830788, + 0.011707888916134834, + 0.005165079142898321, + 0.006232673767954111, + 0.005549597553908825, + 0.01203872263431549, + 0.008819110691547394, + 0.008970766328275204, + 0.009672456420958042, + 0.008357725106179714, + 0.012394925579428673, + 0.011686275713145733, + 0.009078392758965492, + 0.023268219083547592, + 0.03498025983572006, + 0.02007872611284256, + 0.02025994099676609, + 0.023773305118083954, + 0.01024120394140482, + 0.009447342716157436, + 0.008477930910885334, + 0.017290277406573296, + 0.011104910634458065, + 0.011293036863207817, + 0.01936657913029194, + 0.015941806137561798, + 0.017103202641010284, + 0.010863980278372765, + 0.019528234377503395, + 0.02082081511616707, + 0.014512651599943638, + 0.01983300969004631, + 0.015806663781404495, + 0.024867035448551178, + 0.006380751729011536, + 0.005238774232566357, + 0.005096452310681343, + 0.004995123017579317, + 0.005605935584753752, + 0.008422017097473145, + 0.008981888182461262, + 0.008807268925011158, + 0.0035808205138891935, + 0.0037059530150145292, + 0.0056512742303311825, + 0.00443930784240365, + 0.005812841467559338, + 0.006716342642903328, + 0.007404209580272436, + 0.005287969950586557, + 0.011771658435463905, + 0.011088493280112743, + 0.01671772450208664, + 0.008385462686419487, + 0.010850263759493828, + 0.03374816104769707, + 0.020536530762910843, + 0.01602015644311905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.2016032636165619 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736426148563623 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736426148563623, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736426148563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07522941380739212, + 0.07982935756444931, + 0.19738315045833588, + 0.07559464126825333, + 0.36080995202064514, + 0.0654233992099762, + 0.04784667491912842, + 0.031810883432626724, + 0.02498013526201248, + 0.04109232872724533 + ], + "credit_map": [ + 0.008741823956370354, + 0.010340929962694645, + 0.006824934855103493, + 0.01274047140032053, + 0.007697299122810364, + 0.010168679058551788, + 0.007754785940051079, + 0.010960492305457592, + 0.01739867962896824, + 0.007343706209212542, + 0.0060007465071976185, + 0.007999313063919544, + 0.00545675540342927, + 0.008450187742710114, + 0.008703211322426796, + 0.018476758152246475, + 0.02950550988316536, + 0.012502304278314114, + 0.01413036696612835, + 0.019920749589800835, + 0.018838901072740555, + 0.03392929211258888, + 0.02278137393295765, + 0.0457746721804142, + 0.008033317513763905, + 0.007294279523193836, + 0.006026824004948139, + 0.007240853272378445, + 0.010569198988378048, + 0.011580746620893478, + 0.011175845749676228, + 0.01367358211427927, + 0.026584791019558907, + 0.025671379640698433, + 0.027899978682398796, + 0.051428817212581635, + 0.07831770181655884, + 0.04266710579395294, + 0.053585559129714966, + 0.054654624313116074, + 0.005780578125268221, + 0.005613240413367748, + 0.006113702896982431, + 0.007637562695890665, + 0.006310831289738417, + 0.00784672237932682, + 0.014832394197583199, + 0.011288367211818695, + 0.007900984026491642, + 0.004556042607873678, + 0.006377779878675938, + 0.006039842497557402, + 0.004415876232087612, + 0.005422552116215229, + 0.00496268505230546, + 0.008170912973582745, + 0.004035510588437319, + 0.003741186112165451, + 0.002960837446153164, + 0.0031440239399671555, + 0.0033748960122466087, + 0.004501648712903261, + 0.004656397271901369, + 0.005396386608481407, + 0.0025321475695818663, + 0.0023430860601365566, + 0.0031080974731594324, + 0.0031069384422153234, + 0.0032311263494193554, + 0.0038935095071792603, + 0.003896866925060749, + 0.00286836177110672, + 0.006469889543950558, + 0.004657289013266563, + 0.006609858945012093, + 0.0033603133633732796, + 0.004419291857630014, + 0.007802099920809269, + 0.004208866506814957, + 0.0035647174809128046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.006736426148563623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot immerses the viewer in a mysterious nocturnal world, focusing with piercing clarity on a vibrant praying mantis occupying the foreground. The creature's body is a stunning canvas of color: its primary exoskeleton is a vibrant emerald green with a subtle waxy sheen, while its upper thorax and wing casings are adorned with intricate petal-like patterns of pure white and delicate pink. It perches on the edge of a large, black leaf, shimmering with tiny dewdrops. The mantis stalks its unseen prey with intense focus; its triangular head, dominated by large, multi-faceted eyes, rotates with meticulous slowness to track something off-screen. The entire scene is bathed in cool, low-key lighting, creating a color scheme of deep blue and velvety purple, casting the out-of-focus background of lush foliage into shadow and mystery. As the mantis prepares to move, it stealthily raises one of its serrated front legs, a movement so slow and deliberate as to be almost imperceptible, heightening the palpable tension and predatory focus of its nocturnal hunt.", + "seed": 15532, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_9009_seed_15532_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03882931265980005, + "vlm_logprob": -0.03882931265980005, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03882931265980005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006982804741710424 + ], + "opposite_logprob_tokens": [ + -5.256982803344727 + ], + "logprob_reward": -0.006982804741710424, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006982804741710424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4744555354118347, + 0.11418453603982925, + 0.05380561947822571, + 0.055801305919885635, + 0.061455681920051575, + 0.02550964429974556, + 0.05269945040345192, + 0.03673156723380089, + 0.022902153432369232, + 0.10245449095964432 + ], + "credit_map": [ + 0.029183341190218925, + 0.020510949194431305, + 0.041568323969841, + 0.054514359682798386, + 0.027213923633098602, + 0.11755027621984482, + 0.03160072863101959, + 0.15231367945671082, + 0.012358062900602818, + 0.01256991270929575, + 0.01598837412893772, + 0.008361943997442722, + 0.015890764072537422, + 0.016692698001861572, + 0.01207859069108963, + 0.020244203507900238, + 0.0070978873409330845, + 0.005105314310640097, + 0.008591963909566402, + 0.005744304973632097, + 0.005882235709577799, + 0.006423855200409889, + 0.00638381764292717, + 0.008576244115829468, + 0.010181312449276447, + 0.004270398989319801, + 0.006176257971674204, + 0.006064366549253464, + 0.0052045839838683605, + 0.008078004233539104, + 0.007494950667023659, + 0.008331437595188618, + 0.0074002705514431, + 0.006127683911472559, + 0.008360154926776886, + 0.006163674406707287, + 0.007715959567576647, + 0.0076265232637524605, + 0.008244139142334461, + 0.009817278943955898, + 0.003860211232677102, + 0.0024776507634669542, + 0.003103001741692424, + 0.0029578376561403275, + 0.0030263883527368307, + 0.0041855680756270885, + 0.002294981386512518, + 0.003604008350521326, + 0.003737587947398424, + 0.002651093527674675, + 0.004981157835572958, + 0.0035237695556133986, + 0.0032316413708031178, + 0.014586333185434341, + 0.0024335940834134817, + 0.017554279416799545, + 0.005624100565910339, + 0.0037562204524874687, + 0.004742857534438372, + 0.005202848929911852, + 0.004271604586392641, + 0.004192175343632698, + 0.003955997992306948, + 0.004985766485333443, + 0.002457243390381336, + 0.0017951942281797528, + 0.003108978271484375, + 0.0017223575850948691, + 0.0031316704116761684, + 0.004414813593029976, + 0.0029677017591893673, + 0.003304198384284973, + 0.010873557068407536, + 0.007011315785348415, + 0.01918460987508297, + 0.007160233799368143, + 0.015322774648666382, + 0.01697201281785965, + 0.013362416997551918, + 0.012567582540214062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.006982804741710424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087474413216114 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087474413216114, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087474413216114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44894394278526306, + 0.05788862705230713, + 0.07716760784387589, + 0.10997151583433151, + 0.027175799012184143, + 0.04434344917535782, + 0.05460071936249733, + 0.06771637499332428, + 0.030549099668860435, + 0.08164297789335251 + ], + "credit_map": [ + 0.02269897609949112, + 0.013434755615890026, + 0.03918500989675522, + 0.03535042703151703, + 0.021542200818657875, + 0.12436550855636597, + 0.017050106078386307, + 0.1753169149160385, + 0.008763219229876995, + 0.005504432134330273, + 0.007302131503820419, + 0.004929773975163698, + 0.005507775582373142, + 0.009642343036830425, + 0.007293256931006908, + 0.008945685811340809, + 0.010676699690520763, + 0.008414057083427906, + 0.010956100188195705, + 0.00833671074360609, + 0.010757100768387318, + 0.010266141034662724, + 0.0073476447723805904, + 0.01041314098984003, + 0.026271197944879532, + 0.008043292909860611, + 0.011133170686662197, + 0.009935983456671238, + 0.0086640240624547, + 0.013709700666368008, + 0.01638960652053356, + 0.015824535861611366, + 0.003401776310056448, + 0.0019605013076215982, + 0.003435528138652444, + 0.002373890718445182, + 0.0022103721275925636, + 0.0043455311097204685, + 0.003107942407950759, + 0.006340254098176956, + 0.007013161666691303, + 0.00391959585249424, + 0.008302121423184872, + 0.0044622235000133514, + 0.004669702146202326, + 0.005977187305688858, + 0.004383498337119818, + 0.005615951959043741, + 0.006894473917782307, + 0.0028569709975272417, + 0.005025289952754974, + 0.002944167936220765, + 0.0035266669001430273, + 0.015038566663861275, + 0.003120241453871131, + 0.015194334089756012, + 0.01227740291506052, + 0.00670950161293149, + 0.010118523612618446, + 0.008290845900774002, + 0.00738716684281826, + 0.008410662412643433, + 0.005851611960679293, + 0.00867065042257309, + 0.004493445158004761, + 0.002529046032577753, + 0.003616226837038994, + 0.0029748070519417524, + 0.0038892130833119154, + 0.003976903390139341, + 0.004413110204041004, + 0.004656344652175903, + 0.010396190918982029, + 0.00695544108748436, + 0.019424619153141975, + 0.007024754304438829, + 0.006770349107682705, + 0.008659426122903824, + 0.00875026173889637, + 0.013661928474903107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.011087474413216114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306123033165932 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306123033165932, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306123033165932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5265308618545532, + 0.08122396469116211, + 0.11871648579835892, + 0.04125654324889183, + 0.038766343146562576, + 0.02151251956820488, + 0.06380261480808258, + 0.04625552520155907, + 0.01393789704889059, + 0.04799722135066986 + ], + "credit_map": [ + 0.03986234962940216, + 0.037367139011621475, + 0.07441016286611557, + 0.09063071757555008, + 0.04605747014284134, + 0.06900251656770706, + 0.06483743339776993, + 0.10436306148767471, + 0.007021870464086533, + 0.009761563502252102, + 0.011818734928965569, + 0.008001895621418953, + 0.011384681798517704, + 0.010233941487967968, + 0.008063754066824913, + 0.014937524683773518, + 0.01734132133424282, + 0.010252919048070908, + 0.018604110926389694, + 0.011181800626218319, + 0.013677512295544147, + 0.016947975382208824, + 0.012267310172319412, + 0.018443532288074493, + 0.004017820581793785, + 0.003534128190949559, + 0.004621738102287054, + 0.007405327633023262, + 0.003253744915127754, + 0.004790063016116619, + 0.007872741669416428, + 0.005760981235653162, + 0.003161755157634616, + 0.0038892498705536127, + 0.004850026220083237, + 0.005332959815859795, + 0.0047376216389238834, + 0.00438632583245635, + 0.006266249809414148, + 0.006142157595604658, + 0.004019464831799269, + 0.001743791624903679, + 0.0035020196810364723, + 0.0019788851495832205, + 0.0025671725161373615, + 0.002258577151224017, + 0.0021782254334539175, + 0.0032643848098814487, + 0.007892372086644173, + 0.002362116239964962, + 0.005258691031485796, + 0.0028615377377718687, + 0.003911597654223442, + 0.01720433123409748, + 0.0030403146520256996, + 0.021271653473377228, + 0.007228997070342302, + 0.003702190238982439, + 0.007276500575244427, + 0.005205363035202026, + 0.0063244071789085865, + 0.006909594405442476, + 0.003972748294472694, + 0.005635726731270552, + 0.002139107324182987, + 0.0010582600953057408, + 0.0015624837251380086, + 0.0012990748509764671, + 0.0018499421421438456, + 0.0018724314868450165, + 0.001997661776840687, + 0.002158936345949769, + 0.00588853657245636, + 0.003271157620474696, + 0.008303201757371426, + 0.003498980076983571, + 0.006885684095323086, + 0.01005703117698431, + 0.005161033011972904, + 0.004931593779474497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04306123033165932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706612050533295 + ], + "opposite_logprob_tokens": [ + -2.127066135406494 + ], + "logprob_reward": -0.12706612050533295, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706612050533295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49155640602111816, + 0.0890093594789505, + 0.04666650667786598, + 0.12478580325841904, + 0.058294251561164856, + 0.031104356050491333, + 0.029688004404306412, + 0.035729169845581055, + 0.03285928815603256, + 0.06030680984258652 + ], + "credit_map": [ + 0.026845920830965042, + 0.03713126480579376, + 0.053802214562892914, + 0.056910883635282516, + 0.046596307307481766, + 0.08709125965833664, + 0.04530101642012596, + 0.13787758350372314, + 0.008267498575150967, + 0.008222625590860844, + 0.010607042349874973, + 0.010803872719407082, + 0.007851216942071915, + 0.012382411397993565, + 0.011943875811994076, + 0.018930822610855103, + 0.006200466305017471, + 0.004191532731056213, + 0.005157691892236471, + 0.006726918276399374, + 0.005653165280818939, + 0.004764655604958534, + 0.007793229538947344, + 0.006178850308060646, + 0.016730552539229393, + 0.010111329145729542, + 0.016704605892300606, + 0.0191599503159523, + 0.013081304728984833, + 0.010346202179789543, + 0.021056750789284706, + 0.017595116049051285, + 0.006873299367725849, + 0.004654251970350742, + 0.006674584932625294, + 0.011059711687266827, + 0.0060188849456608295, + 0.005864531267434359, + 0.008922739885747433, + 0.008226255886256695, + 0.0056907255202531815, + 0.002590022748336196, + 0.0036139951553195715, + 0.003646829631179571, + 0.00381977716460824, + 0.0030758753418922424, + 0.004273673985153437, + 0.004393457435071468, + 0.003997400868684053, + 0.0027507664635777473, + 0.004412439651787281, + 0.003638983704149723, + 0.002989604603499174, + 0.004248054698109627, + 0.0024499488063156605, + 0.005200808402150869, + 0.007176278159022331, + 0.0033314030151814222, + 0.005350781138986349, + 0.0038327714428305626, + 0.003639110829681158, + 0.004453989211469889, + 0.0030732073355466127, + 0.004871634300798178, + 0.005203021224588156, + 0.002421916462481022, + 0.004454691428691149, + 0.0027106665074825287, + 0.004260360728949308, + 0.004279870539903641, + 0.005380980670452118, + 0.004147781059145927, + 0.005443877074867487, + 0.004418646916747093, + 0.008896359242498875, + 0.005609199404716492, + 0.006769454572349787, + 0.007833247072994709, + 0.011166723445057869, + 0.010169303975999355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.12706612050533295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948933307081461 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948933307081461, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948933307081461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45195186138153076, + 0.18195737898349762, + 0.07997258007526398, + 0.031723566353321075, + 0.03116365149617195, + 0.027424940839409828, + 0.11361244320869446, + 0.020544935017824173, + 0.030065413564443588, + 0.03158322349190712 + ], + "credit_map": [ + 0.034756701439619064, + 0.029014118015766144, + 0.038761623203754425, + 0.03130035102367401, + 0.03829953074455261, + 0.10656284540891647, + 0.02453896962106228, + 0.1487177163362503, + 0.01653217524290085, + 0.02018877863883972, + 0.02354229800403118, + 0.020388903096318245, + 0.02771523967385292, + 0.020891277119517326, + 0.026259442791342735, + 0.026439275592565536, + 0.008993910625576973, + 0.006359731312841177, + 0.008780314587056637, + 0.012384355068206787, + 0.009470789693295956, + 0.013270747847855091, + 0.009889466688036919, + 0.010823266580700874, + 0.003542032092809677, + 0.002558192005380988, + 0.003353312611579895, + 0.004670679569244385, + 0.0032329431269317865, + 0.0029677762649953365, + 0.0062758526764810085, + 0.005122777074575424, + 0.004436444956809282, + 0.0029195768292993307, + 0.0037395881954580545, + 0.0042896089144051075, + 0.0027775170747190714, + 0.0034869376104325056, + 0.004599059000611305, + 0.004914918448776007, + 0.007257408928126097, + 0.002189639490097761, + 0.003446312854066491, + 0.0019106046529486775, + 0.003324441146105528, + 0.0024654900189489126, + 0.003488154150545597, + 0.0033428878523409367, + 0.00621547969058156, + 0.0025893966667354107, + 0.007880193181335926, + 0.0030793307814747095, + 0.00486761936917901, + 0.03936517611145973, + 0.0038961018435657024, + 0.04571914300322533, + 0.002577517880126834, + 0.001566457562148571, + 0.002988928696140647, + 0.00297079817391932, + 0.0035251632798463106, + 0.00209797453135252, + 0.0021678260527551174, + 0.00265026674605906, + 0.003717836458235979, + 0.00273225549608469, + 0.003559633856639266, + 0.0021688658744096756, + 0.004103115294128656, + 0.004084296058863401, + 0.005743654910475016, + 0.0039557539857923985, + 0.0037956528831273317, + 0.0024881893768906593, + 0.0062805647030472755, + 0.0031399691943079233, + 0.004747749771922827, + 0.003640388138592243, + 0.004089674912393093, + 0.0034010345116257668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.005948933307081461 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot immerses the viewer in a mysterious nocturnal world, focusing with piercing clarity on a vibrant praying mantis occupying the foreground. The creature's body is a stunning canvas of color: its primary exoskeleton is a vibrant emerald green with a subtle waxy sheen, while its upper thorax and wing casings are adorned with intricate petal-like patterns of pure white and delicate pink. It perches on the edge of a large, black leaf, shimmering with tiny dewdrops. The mantis stalks its unseen prey with intense focus; its triangular head, dominated by large, multi-faceted eyes, rotates with meticulous slowness to track something off-screen. The entire scene is bathed in cool, low-key lighting, creating a color scheme of deep blue and velvety purple, casting the out-of-focus background of lush foliage into shadow and mystery. As the mantis prepares to move, it stealthily raises one of its serrated front legs, a movement so slow and deliberate as to be almost imperceptible, heightening the palpable tension and predatory focus of its nocturnal hunt.", + "seed": 15533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_6/index_9009_seed_15533_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027168582566082476, + "vlm_logprob": -0.027168582566082476, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.027168582566082476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001995002618059516 + ], + "opposite_logprob_tokens": [ + -6.751995086669922 + ], + "logprob_reward": -0.001995002618059516, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001995002618059516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19911429286003113, + 0.08906461298465729, + 0.28731364011764526, + 0.028332706540822983, + 0.0663621798157692, + 0.20560184121131897, + 0.034337855875492096, + 0.028380136936903, + 0.035741645842790604, + 0.02575107477605343 + ], + "credit_map": [ + 0.029264600947499275, + 0.01702784188091755, + 0.019588204100728035, + 0.02011890895664692, + 0.019571522250771523, + 0.043954819440841675, + 0.019651448354125023, + 0.02993697114288807, + 0.012500731274485588, + 0.012164531275629997, + 0.013224366120994091, + 0.011435924097895622, + 0.006311842706054449, + 0.010236721485853195, + 0.007976580411195755, + 0.01521393097937107, + 0.10523739457130432, + 0.011197212152183056, + 0.010311087593436241, + 0.011527794413268566, + 0.013763324357569218, + 0.11208155006170273, + 0.01104851346462965, + 0.012146785855293274, + 0.004761360120028257, + 0.0023987817112356424, + 0.003726171562448144, + 0.003113160375505686, + 0.0028250899631530046, + 0.0033906197641044855, + 0.004036076366901398, + 0.004081450402736664, + 0.0092517901211977, + 0.006817555986344814, + 0.007861698046326637, + 0.010518291033804417, + 0.009826580993831158, + 0.00842756312340498, + 0.006374991498887539, + 0.007283716928213835, + 0.022974660620093346, + 0.028556130826473236, + 0.03166291490197182, + 0.023163923993706703, + 0.0331987701356411, + 0.02491079643368721, + 0.016652105376124382, + 0.024482544511556625, + 0.007288190070539713, + 0.0029685304034501314, + 0.002752459142357111, + 0.004033960402011871, + 0.0025422496255487204, + 0.0043833330273628235, + 0.0035426211543381214, + 0.006826515309512615, + 0.0035575784277170897, + 0.0031910736579447985, + 0.003204470034688711, + 0.0030086496844887733, + 0.003667819779366255, + 0.00391582865267992, + 0.002938586985692382, + 0.004896130878478289, + 0.0036204331554472446, + 0.0040257591754198074, + 0.004747232887893915, + 0.006058491766452789, + 0.0031554161105304956, + 0.005296305287629366, + 0.003403920214623213, + 0.005434089805930853, + 0.0026532639749348164, + 0.0028069987893104553, + 0.003893566317856312, + 0.0035380113404244184, + 0.0027383214328438044, + 0.0037306430749595165, + 0.002760746283456683, + 0.0036295263562351465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.001995002618059516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16663233935832977, + 0.07202009111642838, + 0.24973809719085693, + 0.04363666847348213, + 0.13430152833461761, + 0.06432628631591797, + 0.07013486325740814, + 0.05806534364819527, + 0.05101259425282478, + 0.09013219177722931 + ], + "credit_map": [ + 0.024807311594486237, + 0.01096804067492485, + 0.023792024701833725, + 0.01870192028582096, + 0.01527918316423893, + 0.025984007865190506, + 0.02201451174914837, + 0.025085344910621643, + 0.010858427733182907, + 0.005574748385697603, + 0.014174574054777622, + 0.00892875250428915, + 0.00482076033949852, + 0.008858774788677692, + 0.006941882893443108, + 0.011862172745168209, + 0.08272676914930344, + 0.009587408974766731, + 0.015617795288562775, + 0.011865861713886261, + 0.01610453426837921, + 0.0892207995057106, + 0.009988061152398586, + 0.014626877382397652, + 0.005878852214664221, + 0.003379064379259944, + 0.005201431456953287, + 0.006883755326271057, + 0.0056595574133098125, + 0.005643766839057207, + 0.006043927278369665, + 0.004946314729750156, + 0.01975042186677456, + 0.015814945101737976, + 0.01440558023750782, + 0.017641542479395866, + 0.024376284331083298, + 0.01775745488703251, + 0.012531831860542297, + 0.012023475021123886, + 0.0062798368744552135, + 0.007301153149455786, + 0.007784656248986721, + 0.008424606174230576, + 0.009652483277022839, + 0.007983592338860035, + 0.007797709200531244, + 0.009102249518036842, + 0.018274640664458275, + 0.008468043059110641, + 0.006672905758023262, + 0.005049366969615221, + 0.005793217569589615, + 0.008658399805426598, + 0.0056538451462984085, + 0.011564443819224834, + 0.007317794486880302, + 0.005996109452098608, + 0.005631475243717432, + 0.0063180821016430855, + 0.007935266010463238, + 0.006314417347311974, + 0.010173057205975056, + 0.00837914552539587, + 0.007064510136842728, + 0.007326559163630009, + 0.007102858740836382, + 0.007260638289153576, + 0.006426097825169563, + 0.004947929177433252, + 0.003915166947990656, + 0.006968833040446043, + 0.011122137308120728, + 0.009567530825734138, + 0.017371803522109985, + 0.010987927205860615, + 0.009569717571139336, + 0.010713533498346806, + 0.008551344275474548, + 0.012248203158378601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979126386344433 + ], + "opposite_logprob_tokens": [ + -3.5297913551330566 + ], + "logprob_reward": -0.02979126386344433, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979126386344433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09857010841369629, + 0.06656017899513245, + 0.3410000801086426, + 0.1486157327890396, + 0.15196169912815094, + 0.09191306680440903, + 0.022695213556289673, + 0.01646488346159458, + 0.029484787955880165, + 0.03273429721593857 + ], + "credit_map": [ + 0.017182091251015663, + 0.007562186568975449, + 0.0139535591006279, + 0.016538726165890694, + 0.008226907812058926, + 0.010390093550086021, + 0.015377373434603214, + 0.009339169599115849, + 0.009751876816153526, + 0.008840983733534813, + 0.00806222204118967, + 0.011363120749592781, + 0.003922832664102316, + 0.010318326763808727, + 0.006014201324433088, + 0.008286609314382076, + 0.08778416365385056, + 0.026973603293299675, + 0.03787567839026451, + 0.03541777282953262, + 0.017192328348755836, + 0.08122314512729645, + 0.020076731219887733, + 0.03445664793252945, + 0.01868271641433239, + 0.010588092729449272, + 0.020394498482346535, + 0.02742369845509529, + 0.01958375610411167, + 0.018750997260212898, + 0.017978595569729805, + 0.015213379636406898, + 0.022907542064785957, + 0.01890738680958748, + 0.014728240668773651, + 0.017031937837600708, + 0.03014366701245308, + 0.02127435803413391, + 0.01311906985938549, + 0.013849501498043537, + 0.010561364702880383, + 0.01161333080381155, + 0.011683150194585323, + 0.010508506558835506, + 0.015816660597920418, + 0.010734600946307182, + 0.008515199646353722, + 0.012480255216360092, + 0.004757819231599569, + 0.002354751341044903, + 0.00252689141780138, + 0.0022267538588494062, + 0.0024912876542657614, + 0.0028821383602917194, + 0.002142405603080988, + 0.00331316702067852, + 0.0026373977307230234, + 0.0015081822639331222, + 0.0017973246285691857, + 0.002125644823536277, + 0.0016630206955596805, + 0.002322894986718893, + 0.0016180488746613264, + 0.002792369108647108, + 0.0027442786376923323, + 0.003228053217753768, + 0.0038594864308834076, + 0.004550713114440441, + 0.0021522126626223326, + 0.005536834709346294, + 0.0027927709743380547, + 0.004620438441634178, + 0.0027305216062813997, + 0.0016920669004321098, + 0.005603722762316465, + 0.0028942814096808434, + 0.0024137396831065416, + 0.008047516457736492, + 0.004004953894764185, + 0.005347494501620531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.02979126386344433 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902681618928909 + ], + "opposite_logprob_tokens": [ + -2.464026927947998 + ], + "logprob_reward": -0.08902681618928909, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902681618928909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08191625028848648, + 0.05966769903898239, + 0.4707084596157074, + 0.09172126650810242, + 0.04906592518091202, + 0.05388002470135689, + 0.06184828281402588, + 0.03912441432476044, + 0.05471699312329292, + 0.03735065832734108 + ], + "credit_map": [ + 0.013269620016217232, + 0.004305916838347912, + 0.01385208684951067, + 0.008531446568667889, + 0.006773201748728752, + 0.01360382605344057, + 0.007892357185482979, + 0.013687804341316223, + 0.008917725645005703, + 0.00608867546543479, + 0.013408907689154148, + 0.006692846305668354, + 0.005396573804318905, + 0.009788992814719677, + 0.004663430154323578, + 0.0047105527482926846, + 0.16040056943893433, + 0.01765565574169159, + 0.020455049350857735, + 0.019987618550658226, + 0.03579346835613251, + 0.173471599817276, + 0.020011091604828835, + 0.022933457046747208, + 0.011427950114011765, + 0.0077811903320252895, + 0.011705169454216957, + 0.012198220938444138, + 0.015104436315596104, + 0.009796704165637493, + 0.015880484133958817, + 0.00782711897045374, + 0.007284124847501516, + 0.006320300977677107, + 0.005416924599558115, + 0.006437918171286583, + 0.008155254647135735, + 0.006269811186939478, + 0.004249547608196735, + 0.004932046867907047, + 0.004841206129640341, + 0.005714009050279856, + 0.00819831620901823, + 0.006031875964254141, + 0.007988028228282928, + 0.006602257490158081, + 0.005623172968626022, + 0.0088811619207263, + 0.011975188739597797, + 0.006531903985887766, + 0.006337882950901985, + 0.005292344372719526, + 0.005776313599199057, + 0.0074844202026724815, + 0.007160670589655638, + 0.011289558373391628, + 0.0051888092420995235, + 0.0038841988425701857, + 0.004677930381149054, + 0.004672774579375982, + 0.005513366777449846, + 0.004465462174266577, + 0.004516580607742071, + 0.00620529567822814, + 0.005517452955245972, + 0.005534759256988764, + 0.004300808999687433, + 0.0056828283704817295, + 0.00702173775061965, + 0.011600430123507977, + 0.004888712894171476, + 0.010170268826186657, + 0.0027604415081441402, + 0.002040698193013668, + 0.003922474570572376, + 0.003713494399562478, + 0.0032737681176513433, + 0.011282955296337605, + 0.004500619135797024, + 0.0058562131598591805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.08902681618928909 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a vibrant praying mantis in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis's exoskeleton primarily emerald green with a waxy sheen and white/pink petal-like patterns on the thorax and wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis perched on the edge of a large black leaf covered in tiny dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis slowly rotate its triangular head with large multi-faceted eyes to track something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool low-key lighting with a deep blue and velvety purple color scheme?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486353972926736 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.002486353972926736, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486353972926736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11503195762634277, + 0.10137531906366348, + 0.11534272879362106, + 0.05955880135297775, + 0.0531146302819252, + 0.21142823994159698, + 0.036051515489816666, + 0.09693486988544464, + 0.09890758246183395, + 0.11225442588329315 + ], + "credit_map": [ + 0.01894572749733925, + 0.011912934482097626, + 0.01894645392894745, + 0.010405370965600014, + 0.008575412444770336, + 0.014177903532981873, + 0.015132567845284939, + 0.0169355608522892, + 0.01561849843710661, + 0.010697265155613422, + 0.014639283530414104, + 0.013554149307310581, + 0.0065171136520802975, + 0.019322577863931656, + 0.008335068821907043, + 0.01269135344773531, + 0.03322085365653038, + 0.004809855949133635, + 0.009601633995771408, + 0.008463983424007893, + 0.009516316466033459, + 0.034823838621377945, + 0.006297955755144358, + 0.008608277887105942, + 0.00857815146446228, + 0.0053919958882033825, + 0.006812389008700848, + 0.009248990565538406, + 0.005492332857102156, + 0.007033915258944035, + 0.010110597126185894, + 0.006890423130244017, + 0.007069517392665148, + 0.004402382764965296, + 0.006918902508914471, + 0.006827234756201506, + 0.007348767481744289, + 0.007248884066939354, + 0.0052100662142038345, + 0.008088871836662292, + 0.026836790144443512, + 0.026919370517134666, + 0.038704998791217804, + 0.020754374563694, + 0.03364821895956993, + 0.02537156082689762, + 0.01808803528547287, + 0.021104862913489342, + 0.005966108292341232, + 0.0033897904213517904, + 0.004382620565593243, + 0.004140854813158512, + 0.0038964704144746065, + 0.004327135626226664, + 0.00509917177259922, + 0.004849357530474663, + 0.011033930815756321, + 0.0114264702424407, + 0.011449635028839111, + 0.01055174507200718, + 0.015505758114159107, + 0.009890281595289707, + 0.010684844106435776, + 0.016392193734645844, + 0.009487655945122242, + 0.010844131000339985, + 0.01149296760559082, + 0.012158327735960484, + 0.010626381263136864, + 0.017905572429299355, + 0.010357815772294998, + 0.016034724190831184, + 0.0073962463065981865, + 0.005720360204577446, + 0.01510175596922636, + 0.00912155769765377, + 0.008346345275640488, + 0.03541751950979233, + 0.01162195485085249, + 0.019528677687048912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.002486353972926736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_tensor([7523, 7523])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_tensor([7523, 7523])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..31babe312b41ac8bf23496bf1f44fe430a0b8895 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_tensor([7523, 7523])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic medium shot, shot in the style of sports photography, slowly arcs around a figure seated on the massive, dark wooden judge's bench within the moot courtroom. Clad in a heavy, matte-fabric black judge's robe, the figure maintains a posture of intense concentration, their face partially obscured by dramatic, high-contrast lighting. A harsh spotlight, shining from above and sideways, carves them out of the surrounding darkness, highlighting the tension in their powerful shoulders and forearms. A sheet of white paper rests on the bench's polished surface. The figure's hand, gripping a traditional bamboo-handled calligraphy brush, moves with the explosive grace of a high-performance athlete; they execute powerful, deliberate strokes, the brush tip gliding across the paper with controlled force. The camera captures every microscopic detail with sharp, dramatic focus: the slight tension of the knuckles, the placement of the glistening black ink, and the final, quick flick of the wrist that completes the act with championship-winning precision, creating an atmosphere of intense concentration and physical prowess.", + "seed": 26468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26468_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.052646773681044576, + "vlm_logprob": -0.052646773681044576, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.052646773681044576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10763495415449142 + ], + "opposite_logprob_tokens": [ + -2.357635021209717 + ], + "logprob_reward": -0.10763495415449142, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10763495415449142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15326137840747833, + 0.08093681186437607, + 0.08799510449171066, + 0.04588933289051056, + 0.04724537953734398, + 0.09088078141212463, + 0.028135672211647034, + 0.14886334538459778, + 0.08154213428497314, + 0.2352500706911087 + ], + "credit_map": [ + 0.019034139811992645, + 0.013776086270809174, + 0.019415173679590225, + 0.016939813271164894, + 0.017593592405319214, + 0.02046748250722885, + 0.018021689727902412, + 0.028013404458761215, + 0.011309337802231312, + 0.012890215031802654, + 0.0103702861815691, + 0.007359582465142012, + 0.00938801746815443, + 0.010375547222793102, + 0.008590687066316605, + 0.010653135366737843, + 0.00925360806286335, + 0.015086779370903969, + 0.010612402111291885, + 0.013403032906353474, + 0.00980938971042633, + 0.010354186408221722, + 0.009493638761341572, + 0.009982064366340637, + 0.005478832870721817, + 0.007884366437792778, + 0.006327491719275713, + 0.004754003137350082, + 0.009789561852812767, + 0.004933895543217659, + 0.0031539280898869038, + 0.003567254403606057, + 0.005599710624665022, + 0.005654729902744293, + 0.005396174732595682, + 0.007849372923374176, + 0.005943320691585541, + 0.006830977741628885, + 0.0047198268584907055, + 0.005251267459243536, + 0.010196409188210964, + 0.012108472175896168, + 0.01052551157772541, + 0.008495531976222992, + 0.01694893091917038, + 0.011066057719290257, + 0.009011904709041119, + 0.012527958489954472, + 0.0044142440892755985, + 0.005892795044928789, + 0.0046730670146644115, + 0.002597573446109891, + 0.002321823500096798, + 0.0033775062765926123, + 0.0022324740421026945, + 0.0026261901948601007, + 0.020695241168141365, + 0.022141067311167717, + 0.0204570684581995, + 0.02026345022022724, + 0.025267627090215683, + 0.014769149012863636, + 0.014106471091508865, + 0.01116326916962862, + 0.011440031230449677, + 0.009157082997262478, + 0.010156442411243916, + 0.008162160404026508, + 0.009545786306262016, + 0.010562377981841564, + 0.012280785478651524, + 0.010237465612590313, + 0.033676937222480774, + 0.043857600539922714, + 0.030626393854618073, + 0.03464236482977867, + 0.025969741865992546, + 0.024864597246050835, + 0.021247340366244316, + 0.020365096628665924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.10763495415449142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03375411406159401 + ], + "opposite_logprob_tokens": [ + -3.4087541103363037 + ], + "logprob_reward": -0.03375411406159401, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03375411406159401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21717621386051178, + 0.1230214312672615, + 0.059022918343544006, + 0.01930171065032482, + 0.04110891371965408, + 0.11506982147693634, + 0.05702952295541763, + 0.18238894641399384, + 0.08335742354393005, + 0.10252311080694199 + ], + "credit_map": [ + 0.02612212672829628, + 0.04493236169219017, + 0.023156950250267982, + 0.02103579044342041, + 0.03047090955078602, + 0.02838689647614956, + 0.01704307645559311, + 0.026028096675872803, + 0.017570823431015015, + 0.03153116628527641, + 0.012044738046824932, + 0.010714716278016567, + 0.01685771532356739, + 0.013247779570519924, + 0.00902493204921484, + 0.012029563076794147, + 0.006882774643599987, + 0.01167835108935833, + 0.0065926359966397285, + 0.009395982138812542, + 0.007414219435304403, + 0.005575378425419331, + 0.005774720571935177, + 0.0057088579051196575, + 0.002314465120434761, + 0.0030758939683437347, + 0.0024693638551980257, + 0.0022779337596148252, + 0.0025727739557623863, + 0.002401374513283372, + 0.0016285240417346358, + 0.002561381086707115, + 0.00619925744831562, + 0.006899064872413874, + 0.0048768953420221806, + 0.0051742070354521275, + 0.007099418435245752, + 0.004087773617357016, + 0.003356912173330784, + 0.003415387589484453, + 0.01577635295689106, + 0.029136115685105324, + 0.012222790159285069, + 0.010286953300237656, + 0.017748933285474777, + 0.011913664638996124, + 0.008438348770141602, + 0.009546658024191856, + 0.009496025741100311, + 0.01435129065066576, + 0.009922469966113567, + 0.003945452626794577, + 0.005443574860692024, + 0.006040563341230154, + 0.00349979754537344, + 0.004330351948738098, + 0.02809029631316662, + 0.03272218257188797, + 0.0233549065887928, + 0.02171080745756626, + 0.031837496906518936, + 0.019785188138484955, + 0.01398991234600544, + 0.010898169130086899, + 0.01120700128376484, + 0.012184246443212032, + 0.012008171528577805, + 0.010085845366120338, + 0.010490026324987411, + 0.01029409933835268, + 0.008347754366695881, + 0.00874028354883194, + 0.011526287533342838, + 0.01181443314999342, + 0.011122176423668861, + 0.019052080810070038, + 0.011473658494651318, + 0.012922156602144241, + 0.01377103105187416, + 0.010841288603842258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03375411406159401 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498656630516052 + ], + "opposite_logprob_tokens": [ + -2.9299864768981934 + ], + "logprob_reward": -0.05498656630516052, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498656630516052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3198903203010559, + 0.08268054574728012, + 0.06973611563444138, + 0.08157399296760559, + 0.051794469356536865, + 0.15030845999717712, + 0.029851576313376427, + 0.08619966357946396, + 0.0710868388414383, + 0.05687794089317322 + ], + "credit_map": [ + 0.034787971526384354, + 0.04188956320285797, + 0.042886022478342056, + 0.034390486776828766, + 0.03125728294253349, + 0.05915315821766853, + 0.03712001442909241, + 0.03840585798025131, + 0.010252520442008972, + 0.012407910078763962, + 0.00994657538831234, + 0.0066183144226670265, + 0.015387095510959625, + 0.012993849813938141, + 0.00755641283467412, + 0.007517884019762278, + 0.009416806511580944, + 0.007302809506654739, + 0.009950816631317139, + 0.008951012045145035, + 0.007351192645728588, + 0.010610303841531277, + 0.006340527441352606, + 0.009812652133405209, + 0.010501040145754814, + 0.011197344399988651, + 0.011323426850140095, + 0.014658802188932896, + 0.006769356783479452, + 0.010346267372369766, + 0.008593980222940445, + 0.008183789439499378, + 0.008028115145862103, + 0.006596121937036514, + 0.008016671054065228, + 0.0064371079206466675, + 0.007108727004379034, + 0.006071865558624268, + 0.004236576613038778, + 0.005299288779497147, + 0.020556850358843803, + 0.034487295895814896, + 0.015469073317945004, + 0.013312322087585926, + 0.02268167771399021, + 0.017804905772209167, + 0.012216699309647083, + 0.013779652304947376, + 0.004599489737302065, + 0.003610314102843404, + 0.004759881645441055, + 0.0023035851772874594, + 0.0033543650060892105, + 0.004442809149622917, + 0.0032456449698656797, + 0.0035354893188923597, + 0.01294910442084074, + 0.014449630863964558, + 0.010486889630556107, + 0.010383679531514645, + 0.016099920496344566, + 0.008926383219659328, + 0.007341325748711824, + 0.005562742706388235, + 0.008839622139930725, + 0.008797862567007542, + 0.008954471908509731, + 0.007008697371929884, + 0.010733569040894508, + 0.008784039877355099, + 0.008708485402166843, + 0.009260090067982674, + 0.007470388431102037, + 0.00824876967817545, + 0.007236021105200052, + 0.005312463268637657, + 0.009673134423792362, + 0.0065473648719489574, + 0.006401464808732271, + 0.0059883370995521545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.05498656630516052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203342333436012 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.018203342333436012, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018203342333436012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11298958957195282, + 0.0781199187040329, + 0.04784927889704704, + 0.05951203778386116, + 0.1616729348897934, + 0.07289665937423706, + 0.028548529371619225, + 0.23843920230865479, + 0.09015076607465744, + 0.1098211258649826 + ], + "credit_map": [ + 0.014709166251122952, + 0.018443290144205093, + 0.014826883561909199, + 0.01562410593032837, + 0.014001204632222652, + 0.01649879477918148, + 0.008582391776144505, + 0.010303739458322525, + 0.011951852589845657, + 0.012315000407397747, + 0.009227166883647442, + 0.006677054800093174, + 0.007333723362535238, + 0.009801866486668587, + 0.008957909420132637, + 0.011855334043502808, + 0.006954637821763754, + 0.00805282685905695, + 0.006753204856067896, + 0.005792115814983845, + 0.004814322106540203, + 0.005485900677740574, + 0.004572959616780281, + 0.005423306487500668, + 0.007754450663924217, + 0.008536601439118385, + 0.0082870963960886, + 0.005987127777189016, + 0.013187285512685776, + 0.006205886136740446, + 0.005285424645990133, + 0.004268157761543989, + 0.024078959599137306, + 0.022231604903936386, + 0.025630444288253784, + 0.019396668300032616, + 0.02276681736111641, + 0.019366327673196793, + 0.014637326821684837, + 0.013564764522016048, + 0.010288568213582039, + 0.018007023259997368, + 0.007958346977829933, + 0.007035592570900917, + 0.01216975785791874, + 0.00828058272600174, + 0.004939838778227568, + 0.004216940142214298, + 0.004560536239296198, + 0.005400386638939381, + 0.004695713985711336, + 0.0021991783287376165, + 0.0030524141620844603, + 0.003731180215254426, + 0.0023967279121279716, + 0.0025123883970081806, + 0.03667141869664192, + 0.04060707613825798, + 0.029615415260195732, + 0.028993314132094383, + 0.028002409264445305, + 0.024619033560156822, + 0.025042971596121788, + 0.024887530133128166, + 0.015104448422789574, + 0.011721326969563961, + 0.009701982140541077, + 0.009041175246238708, + 0.011258551850914955, + 0.011660008691251278, + 0.010663188062608242, + 0.011000066995620728, + 0.013905703090131283, + 0.014933649450540543, + 0.011843160726130009, + 0.018017293885350227, + 0.01565135270357132, + 0.01438247598707676, + 0.013806049712002277, + 0.007281426340341568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.018203342333436012 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654891550540924 + ], + "opposite_logprob_tokens": [ + -3.048654794692993 + ], + "logprob_reward": -0.048654891550540924, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654891550540924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13344532251358032, + 0.08407263457775116, + 0.06246859207749367, + 0.04721248522400856, + 0.046099063009023666, + 0.07791201025247574, + 0.03715537488460541, + 0.05735676363110542, + 0.12132473289966583, + 0.3329530358314514 + ], + "credit_map": [ + 0.016490519046783447, + 0.023607147857546806, + 0.019020894542336464, + 0.012859538197517395, + 0.020362185314297676, + 0.017891287803649902, + 0.009263991378247738, + 0.013949763961136341, + 0.011616803705692291, + 0.008746281266212463, + 0.010203060694038868, + 0.0060033174231648445, + 0.016267355531454086, + 0.012724959291517735, + 0.007221372332423925, + 0.011289473623037338, + 0.007955851964652538, + 0.008953837677836418, + 0.007631633896380663, + 0.0066362242214381695, + 0.008627167902886868, + 0.006265451665967703, + 0.006598401349037886, + 0.009800024330615997, + 0.007189437281340361, + 0.005584998521953821, + 0.00725769205018878, + 0.004111667163670063, + 0.00465519679710269, + 0.005778114311397076, + 0.0058897389099001884, + 0.006745639722794294, + 0.0069341291673481464, + 0.005300577264279127, + 0.006075432524085045, + 0.005777217913419008, + 0.005891906097531319, + 0.005430085118860006, + 0.006341137457638979, + 0.004348579328507185, + 0.009152078069746494, + 0.012887260876595974, + 0.008977469056844711, + 0.0059030367992818356, + 0.011169332079589367, + 0.012714516371488571, + 0.007625761907547712, + 0.009482556954026222, + 0.008299537934362888, + 0.003284843871369958, + 0.00915342103689909, + 0.0029562488198280334, + 0.0026570504996925592, + 0.004657200537621975, + 0.0025543137453496456, + 0.0035927630960941315, + 0.00849138479679823, + 0.008821435272693634, + 0.006033937446773052, + 0.00506991520524025, + 0.00965319573879242, + 0.005756756290793419, + 0.0061330185271799564, + 0.007397117093205452, + 0.01591048575937748, + 0.019301367923617363, + 0.012522699311375618, + 0.013548675924539566, + 0.02456553839147091, + 0.015405913814902306, + 0.012714128009974957, + 0.007355918642133474, + 0.050888266414403915, + 0.0645780861377716, + 0.027790099382400513, + 0.03836967051029205, + 0.04765210300683975, + 0.04369376227259636, + 0.03133619204163551, + 0.02864483930170536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.048654891550540924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium shot, shot in the style of sports photography, slowly arcs around a figure seated on the massive, dark wooden judge's bench within the moot courtroom. Clad in a heavy, matte-fabric black judge's robe, the figure maintains a posture of intense concentration, their face partially obscured by dramatic, high-contrast lighting. A harsh spotlight, shining from above and sideways, carves them out of the surrounding darkness, highlighting the tension in their powerful shoulders and forearms. A sheet of white paper rests on the bench's polished surface. The figure's hand, gripping a traditional bamboo-handled calligraphy brush, moves with the explosive grace of a high-performance athlete; they execute powerful, deliberate strokes, the brush tip gliding across the paper with controlled force. The camera captures every microscopic detail with sharp, dramatic focus: the slight tension of the knuckles, the placement of the glistening black ink, and the final, quick flick of the wrist that completes the act with championship-winning precision, creating an atmosphere of intense concentration and physical prowess.", + "seed": 26469, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26469_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14488984197378157, + "vlm_logprob": -0.14488984197378157, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.14488984197378157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43670573830604553 + ], + "opposite_logprob_tokens": [ + -1.0617057085037231 + ], + "logprob_reward": -0.43670573830604553, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43670573830604553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05278193950653076, + 0.10284951329231262, + 0.09683401882648468, + 0.042265329509973526, + 0.06490093469619751, + 0.028501471504569054, + 0.01860714517533779, + 0.4157051742076874, + 0.12167207896709442, + 0.05588240921497345 + ], + "credit_map": [ + 0.009854724630713463, + 0.003415188752114773, + 0.009505153633654118, + 0.005404842086136341, + 0.008141735568642616, + 0.006108500994741917, + 0.005854566115885973, + 0.004497228190302849, + 0.01657521165907383, + 0.009701880626380444, + 0.020311297848820686, + 0.008836198598146439, + 0.012148210778832436, + 0.012367596849799156, + 0.009463954716920853, + 0.01344516035169363, + 0.01718309335410595, + 0.013995105400681496, + 0.018740851432085037, + 0.010274081490933895, + 0.00606201495975256, + 0.008437152951955795, + 0.011887171305716038, + 0.010254550725221634, + 0.008143510669469833, + 0.004159797448664904, + 0.0035860235802829266, + 0.00441390136256814, + 0.006091498304158449, + 0.005560116842389107, + 0.005380980204790831, + 0.004929501563310623, + 0.011859819293022156, + 0.005297995172441006, + 0.016811516135931015, + 0.006849342025816441, + 0.0047035240568220615, + 0.007954117842018604, + 0.00604011956602335, + 0.005384499207139015, + 0.004757200833410025, + 0.002284199930727482, + 0.005408887285739183, + 0.0024087727069854736, + 0.002495063468813896, + 0.007147442549467087, + 0.001582103199325502, + 0.002417799085378647, + 0.003158824285492301, + 0.0031192742753773928, + 0.003198083722963929, + 0.0018352307379245758, + 0.0017822436057031155, + 0.003007655730471015, + 0.0014683614717796445, + 0.001037471811287105, + 0.07018215954303741, + 0.05695067718625069, + 0.08146476745605469, + 0.04418683052062988, + 0.03709428384900093, + 0.050039105117321014, + 0.031087912619113922, + 0.044699449092149734, + 0.02468663454055786, + 0.00940710213035345, + 0.03188302740454674, + 0.01617247797548771, + 0.006808804348111153, + 0.01006573997437954, + 0.012653373181819916, + 0.009994917549192905, + 0.01073188241571188, + 0.006554102059453726, + 0.013858459889888763, + 0.0066150096245110035, + 0.004225708078593016, + 0.006150151137262583, + 0.0037354195956140757, + 0.004011672455817461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.43670573830604553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034731030464172 + ], + "opposite_logprob_tokens": [ + -2.3503472805023193 + ], + "logprob_reward": -0.10034731030464172, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034731030464172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14892850816249847, + 0.06607528775930405, + 0.1219899132847786, + 0.0669560506939888, + 0.0722108706831932, + 0.0322657972574234, + 0.07202788442373276, + 0.13830538094043732, + 0.21362382173538208, + 0.06761645525693893 + ], + "credit_map": [ + 0.024795830249786377, + 0.0158386267721653, + 0.01740209572017193, + 0.025176063179969788, + 0.02285926789045334, + 0.013447081670165062, + 0.015457672998309135, + 0.013951878994703293, + 0.012127017602324486, + 0.004971824120730162, + 0.007480875123292208, + 0.007075271103531122, + 0.004520425572991371, + 0.007844476960599422, + 0.0047129616141319275, + 0.01734243519604206, + 0.01822231151163578, + 0.01445033773779869, + 0.019895367324352264, + 0.013997890055179596, + 0.009726431220769882, + 0.013722100295126438, + 0.021238457411527634, + 0.010737014934420586, + 0.015870727598667145, + 0.0049691228196024895, + 0.007565449923276901, + 0.00482773594558239, + 0.006160512566566467, + 0.006907710339874029, + 0.008827348239719868, + 0.011827446520328522, + 0.011098943650722504, + 0.00949513353407383, + 0.014346295967698097, + 0.007874015718698502, + 0.008653134107589722, + 0.008935749530792236, + 0.00587258068844676, + 0.005935018416494131, + 0.005605025216937065, + 0.0023230428341776133, + 0.0056749871000647545, + 0.003616475500166416, + 0.0024278946220874786, + 0.00814303569495678, + 0.0018754297634586692, + 0.0025999073404818773, + 0.00969379860907793, + 0.012321744114160538, + 0.010861280374228954, + 0.006491218693554401, + 0.007053843233734369, + 0.007445266470313072, + 0.010960186831653118, + 0.007200548890978098, + 0.024775156751275063, + 0.016731785610318184, + 0.03020014986395836, + 0.015829550102353096, + 0.013539418578147888, + 0.01520475372672081, + 0.00915557611733675, + 0.012868989259004593, + 0.03164808452129364, + 0.036441005766391754, + 0.02061939798295498, + 0.03172072395682335, + 0.026182105764746666, + 0.022460997104644775, + 0.028976745903491974, + 0.015574750490486622, + 0.011463671922683716, + 0.010252539068460464, + 0.008524596691131592, + 0.010240532457828522, + 0.0066298386082053185, + 0.00856104027479887, + 0.005820749793201685, + 0.006123484577983618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.10034731030464172 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428384780883789 + ], + "opposite_logprob_tokens": [ + -2.017838478088379 + ], + "logprob_reward": -0.1428384780883789, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428384780883789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09232860803604126, + 0.18985609710216522, + 0.05390836298465729, + 0.1633569747209549, + 0.08581603318452835, + 0.04239872843027115, + 0.05571913346648216, + 0.09147608280181885, + 0.13331730663776398, + 0.09182260185480118 + ], + "credit_map": [ + 0.02194577269256115, + 0.007391604129225016, + 0.014641126617789268, + 0.007122321054339409, + 0.010714896023273468, + 0.01035299152135849, + 0.009100651368498802, + 0.011059259064495564, + 0.030229652300477028, + 0.01976270042359829, + 0.03695892542600632, + 0.024550212547183037, + 0.021605126559734344, + 0.024164803326129913, + 0.014025650918483734, + 0.01855904422700405, + 0.0085750138387084, + 0.005504349246621132, + 0.011974147520959377, + 0.006981255952268839, + 0.004128011874854565, + 0.005409990902990103, + 0.006615922786295414, + 0.004719679243862629, + 0.03238685801625252, + 0.014860467053949833, + 0.016596641391515732, + 0.019136112183332443, + 0.01704445295035839, + 0.018516624346375465, + 0.025116734206676483, + 0.01969909854233265, + 0.014346186071634293, + 0.0101256612688303, + 0.011972182430326939, + 0.011916044168174267, + 0.008543712086975574, + 0.010819167830049992, + 0.010631872341036797, + 0.007461222819983959, + 0.00816928781569004, + 0.003219051519408822, + 0.0048864553682506084, + 0.0037061823531985283, + 0.003619480412453413, + 0.012722187675535679, + 0.0028696563094854355, + 0.003206432331353426, + 0.008508756756782532, + 0.005494678858667612, + 0.010626645758748055, + 0.007313136477023363, + 0.005930545274168253, + 0.007037059403955936, + 0.006595316808670759, + 0.004213002510368824, + 0.013909335248172283, + 0.00861285999417305, + 0.015842067077755928, + 0.011106180027127266, + 0.009256419725716114, + 0.013400819152593613, + 0.00649907486513257, + 0.012849335558712482, + 0.023455822840332985, + 0.01255478709936142, + 0.0289524607360363, + 0.01905742473900318, + 0.01011136919260025, + 0.012524754740297794, + 0.016706138849258423, + 0.009954553097486496, + 0.01810462772846222, + 0.008502761833369732, + 0.027444925159215927, + 0.009417557157576084, + 0.007280110381543636, + 0.01007117610424757, + 0.004829935729503632, + 0.006171520333737135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.1428384780883789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366515085101128 + ], + "opposite_logprob_tokens": [ + -3.6513664722442627 + ], + "logprob_reward": -0.026366515085101128, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366515085101128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11116144806146622, + 0.07452967762947083, + 0.05168897286057472, + 0.07439077645540237, + 0.21269206702709198, + 0.08373808860778809, + 0.07868430763483047, + 0.10971833765506744, + 0.06572884321212769, + 0.13766741752624512 + ], + "credit_map": [ + 0.02585916966199875, + 0.008587608113884926, + 0.015315718948841095, + 0.015601739287376404, + 0.010496347211301327, + 0.012300691567361355, + 0.009317041374742985, + 0.013683135621249676, + 0.010379270650446415, + 0.008039368316531181, + 0.007780973333865404, + 0.014295781031250954, + 0.0071767899207770824, + 0.010464812628924847, + 0.006139254197478294, + 0.010253428481519222, + 0.00836100336164236, + 0.006437043659389019, + 0.008030727505683899, + 0.00542002497240901, + 0.004355550743639469, + 0.00636838935315609, + 0.004928586073219776, + 0.007787647191435099, + 0.012775440700352192, + 0.006634618621319532, + 0.0063654836267232895, + 0.006457023322582245, + 0.012975405901670456, + 0.011668258346617222, + 0.00902267824858427, + 0.008491862565279007, + 0.03755304962396622, + 0.014967598021030426, + 0.045678332448005676, + 0.026430610567331314, + 0.014608091674745083, + 0.03251226246356964, + 0.021092602983117104, + 0.01984952576458454, + 0.011938519775867462, + 0.0074667190201580524, + 0.013022465631365776, + 0.01008567400276661, + 0.008845416828989983, + 0.02042514458298683, + 0.00485090259462595, + 0.007103244308382273, + 0.013197550550103188, + 0.009473246522247791, + 0.014372475445270538, + 0.0057260929606854916, + 0.007037405855953693, + 0.012655586004257202, + 0.008882371708750725, + 0.0073395781219005585, + 0.015896379947662354, + 0.017107317224144936, + 0.0176301971077919, + 0.01202380657196045, + 0.00956527516245842, + 0.01539179403334856, + 0.008344735018908978, + 0.013758829794824123, + 0.012650436721742153, + 0.005756277125328779, + 0.01717268116772175, + 0.007046074606478214, + 0.005644341465085745, + 0.005351744592189789, + 0.007143441122025251, + 0.004963840823620558, + 0.024368243291974068, + 0.01946159638464451, + 0.027911147102713585, + 0.017906885594129562, + 0.01477882545441389, + 0.014393758028745651, + 0.009179990738630295, + 0.009666967205703259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.026366515085101128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819116808474064 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819116808474064, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819116808474064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24189528822898865, + 0.0978582575917244, + 0.06624650955200195, + 0.07948678731918335, + 0.1319696456193924, + 0.06420744210481644, + 0.044270772486925125, + 0.10134091228246689, + 0.051936425268650055, + 0.12078795582056046 + ], + "credit_map": [ + 0.051706235855817795, + 0.012131460942327976, + 0.024529483169317245, + 0.07554437220096588, + 0.028404243290424347, + 0.020383894443511963, + 0.01782357506453991, + 0.01137202512472868, + 0.016051385551691055, + 0.012685182504355907, + 0.009304448030889034, + 0.015744835138320923, + 0.008579854853451252, + 0.01175185851752758, + 0.0065384008921682835, + 0.017202289775013924, + 0.009827664121985435, + 0.007072308100759983, + 0.008932311087846756, + 0.00820369552820921, + 0.00821615383028984, + 0.008498131297528744, + 0.008117693476378918, + 0.007378550246357918, + 0.01128715742379427, + 0.010284407064318657, + 0.00953617598861456, + 0.01179183553904295, + 0.008640961721539497, + 0.008317905478179455, + 0.01252500619739294, + 0.00710334163159132, + 0.021613379940390587, + 0.014376482926309109, + 0.02567509189248085, + 0.01793355867266655, + 0.010033921338617802, + 0.014957262203097343, + 0.01664084568619728, + 0.010739102959632874, + 0.009575012139976025, + 0.005877038463950157, + 0.008771331049501896, + 0.008667157031595707, + 0.00875120423734188, + 0.010975873097777367, + 0.005605871323496103, + 0.005983952432870865, + 0.00807682704180479, + 0.0062002576887607574, + 0.004674952942878008, + 0.004083422012627125, + 0.004415576346218586, + 0.009805536828935146, + 0.004128491505980492, + 0.0028857067227363586, + 0.015963247045874596, + 0.008345029316842556, + 0.016658281907439232, + 0.012585208751261234, + 0.012579523026943207, + 0.014646015129983425, + 0.007685855031013489, + 0.0128777539357543, + 0.008221505209803581, + 0.007297234144061804, + 0.006648905575275421, + 0.007247550878673792, + 0.005052512977272272, + 0.005399385467171669, + 0.006964141968637705, + 0.005105191841721535, + 0.020760629326105118, + 0.018651815131306648, + 0.023778842762112617, + 0.017202865332365036, + 0.01225326582789421, + 0.011353245005011559, + 0.00892604235559702, + 0.007861255668103695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.01819116808474064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium shot, shot in the style of sports photography, slowly arcs around a figure seated on the massive, dark wooden judge's bench within the moot courtroom. Clad in a heavy, matte-fabric black judge's robe, the figure maintains a posture of intense concentration, their face partially obscured by dramatic, high-contrast lighting. A harsh spotlight, shining from above and sideways, carves them out of the surrounding darkness, highlighting the tension in their powerful shoulders and forearms. A sheet of white paper rests on the bench's polished surface. The figure's hand, gripping a traditional bamboo-handled calligraphy brush, moves with the explosive grace of a high-performance athlete; they execute powerful, deliberate strokes, the brush tip gliding across the paper with controlled force. The camera captures every microscopic detail with sharp, dramatic focus: the slight tension of the knuckles, the placement of the glistening black ink, and the final, quick flick of the wrist that completes the act with championship-winning precision, creating an atmosphere of intense concentration and physical prowess.", + "seed": 26470, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26470_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07925971560180187, + "vlm_logprob": -0.07925971560180187, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.07925971560180187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2251890003681183 + ], + "opposite_logprob_tokens": [ + -1.725188970565796 + ], + "logprob_reward": -0.2251890003681183, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2251890003681183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15723583102226257, + 0.07059713453054428, + 0.06269461661577225, + 0.01798129640519619, + 0.14756517112255096, + 0.052511315792798996, + 0.06128283962607384, + 0.043531641364097595, + 0.19809532165527344, + 0.1885048747062683 + ], + "credit_map": [ + 0.04528720676898956, + 0.014052373357117176, + 0.01075481902807951, + 0.011579380370676517, + 0.009883476421236992, + 0.0417676642537117, + 0.011863332241773605, + 0.012047572061419487, + 0.012081952765583992, + 0.012629015371203423, + 0.0069425636902451515, + 0.008080408908426762, + 0.005850267130881548, + 0.008828196674585342, + 0.007120269816368818, + 0.009064468555152416, + 0.015759848058223724, + 0.009448613040149212, + 0.011669526807963848, + 0.00890171155333519, + 0.0045072054490447044, + 0.004369931761175394, + 0.003936049062758684, + 0.004101734142750502, + 0.0018434192752465606, + 0.00240513333119452, + 0.0019659006502479315, + 0.002368417801335454, + 0.002355943899601698, + 0.0022223512642085552, + 0.002419955562800169, + 0.002400174969807267, + 0.00595130817964673, + 0.00973069854080677, + 0.006984503008425236, + 0.007064264267683029, + 0.05443133786320686, + 0.03472601994872093, + 0.012649782933294773, + 0.01602725312113762, + 0.006180077791213989, + 0.006420268211513758, + 0.006927583832293749, + 0.004121956415474415, + 0.0062074377201497555, + 0.009153694845736027, + 0.008024709299206734, + 0.005475586745887995, + 0.00476127490401268, + 0.006596841849386692, + 0.008867273107171059, + 0.00805233046412468, + 0.010256431996822357, + 0.012558802030980587, + 0.0044718156568706036, + 0.005718069616705179, + 0.004253792576491833, + 0.003550406778231263, + 0.0030360061209648848, + 0.005669247359037399, + 0.008269603364169598, + 0.00795770063996315, + 0.0036798773799091578, + 0.007115007843822241, + 0.018002038821578026, + 0.017749523743987083, + 0.016543397679924965, + 0.011394746601581573, + 0.031690508127212524, + 0.036485131829977036, + 0.01779131032526493, + 0.0484386570751667, + 0.017467450350522995, + 0.03241121023893356, + 0.009474353864789009, + 0.04912180453538895, + 0.014449555426836014, + 0.029825540259480476, + 0.015856977552175522, + 0.019897980615496635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.2251890003681183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07010828703641891 + ], + "opposite_logprob_tokens": [ + -2.69510817527771 + ], + "logprob_reward": -0.07010828703641891, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07010828703641891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11877591162919998, + 0.2108619660139084, + 0.03680671751499176, + 0.10880051553249359, + 0.08657824248075485, + 0.04258866235613823, + 0.08573451638221741, + 0.05021068826317787, + 0.20134593546390533, + 0.05829690024256706 + ], + "credit_map": [ + 0.019827771931886673, + 0.01568220928311348, + 0.017425769940018654, + 0.014779550954699516, + 0.013242078013718128, + 0.017301829531788826, + 0.012578418478369713, + 0.00793828722089529, + 0.034925393760204315, + 0.04852520674467087, + 0.030591320246458054, + 0.04454466700553894, + 0.019418111070990562, + 0.012526247650384903, + 0.010479927062988281, + 0.009851088747382164, + 0.006597992032766342, + 0.005048956722021103, + 0.007549905683845282, + 0.004945404827594757, + 0.0033455844968557358, + 0.003586942795664072, + 0.0036057494580745697, + 0.0021261796355247498, + 0.015366582199931145, + 0.019117610529065132, + 0.011156064458191395, + 0.009269819594919682, + 0.013121260330080986, + 0.020462719723582268, + 0.010176439769566059, + 0.010130020789802074, + 0.0027010985650122166, + 0.005756639409810305, + 0.00410899892449379, + 0.0029463586397469044, + 0.03297053277492523, + 0.021597087383270264, + 0.0066774627193808556, + 0.009820065461099148, + 0.004969008732587099, + 0.004035204648971558, + 0.004072312265634537, + 0.0026151123456656933, + 0.004819789901375771, + 0.007780739106237888, + 0.007510418072342873, + 0.006786078214645386, + 0.00428913114592433, + 0.009843706153333187, + 0.005162287503480911, + 0.008965735323727131, + 0.019762178882956505, + 0.025735093280673027, + 0.0069545903243124485, + 0.005021801218390465, + 0.004664708394557238, + 0.004679568577557802, + 0.003927657846361399, + 0.005914333276450634, + 0.008923360146582127, + 0.008787369355559349, + 0.0037770159542560577, + 0.009536677040159702, + 0.019824540242552757, + 0.017997268587350845, + 0.017038092017173767, + 0.011875098571181297, + 0.032209575176239014, + 0.036946851760149, + 0.0175632257014513, + 0.0478912852704525, + 0.00620473176240921, + 0.005856939125806093, + 0.00278919143602252, + 0.013386199250817299, + 0.008839552290737629, + 0.009367899969220161, + 0.005746331997215748, + 0.006106051150709391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.07010828703641891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208626553416252 + ], + "opposite_logprob_tokens": [ + -2.812086343765259 + ], + "logprob_reward": -0.06208626553416252, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208626553416252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12489014863967896, + 0.07687132805585861, + 0.0619964562356472, + 0.02765568532049656, + 0.07824390381574631, + 0.17144203186035156, + 0.0563189759850502, + 0.06433124840259552, + 0.25576263666152954, + 0.08248762786388397 + ], + "credit_map": [ + 0.02819213643670082, + 0.01206512376666069, + 0.022029530256986618, + 0.011256147176027298, + 0.00891862716525793, + 0.026538370177149773, + 0.009312283247709274, + 0.006577932741492987, + 0.008688909001648426, + 0.01268057432025671, + 0.008781981654465199, + 0.009817364625632763, + 0.007337781134992838, + 0.01259366050362587, + 0.00744866207242012, + 0.009522395208477974, + 0.010472810827195644, + 0.009866246953606606, + 0.009505472145974636, + 0.0075929355807602406, + 0.008469409309327602, + 0.006062676664441824, + 0.006338606588542461, + 0.0036882967688143253, + 0.0030321013182401657, + 0.0033628239762037992, + 0.003087002085521817, + 0.004402196500450373, + 0.0035595567896962166, + 0.004174272995442152, + 0.0023148655891418457, + 0.0037228669971227646, + 0.0040434580296278, + 0.006165571045130491, + 0.005196231417357922, + 0.005094114225357771, + 0.02636723965406418, + 0.01882954314351082, + 0.0072525362484157085, + 0.005295207723975182, + 0.018478138372302055, + 0.02441731095314026, + 0.018598832190036774, + 0.01567971520125866, + 0.021859614178538322, + 0.02460354007780552, + 0.02900351770222187, + 0.018801357597112656, + 0.005373177584260702, + 0.007146920543164015, + 0.007802012376487255, + 0.006447274703532457, + 0.009123292751610279, + 0.008051346987485886, + 0.005903021898120642, + 0.006471931468695402, + 0.006290065124630928, + 0.006961232051253319, + 0.0051868692971765995, + 0.007639773655682802, + 0.012510955333709717, + 0.011233910918235779, + 0.0054651470854878426, + 0.009043295867741108, + 0.025661151856184006, + 0.022388052195310593, + 0.021425852552056313, + 0.01447159331291914, + 0.03897280991077423, + 0.04801937937736511, + 0.022715093567967415, + 0.06210869923233986, + 0.011715099215507507, + 0.01435103826224804, + 0.004433506168425083, + 0.00940434355288744, + 0.01566300541162491, + 0.011623503640294075, + 0.007329660467803478, + 0.007967475801706314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.06208626553416252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563489377498627 + ], + "opposite_logprob_tokens": [ + -4.387563705444336 + ], + "logprob_reward": -0.012563489377498627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563489377498627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11500979214906693, + 0.047594547271728516, + 0.06998179852962494, + 0.03893246129155159, + 0.24602672457695007, + 0.0924631878733635, + 0.03667375445365906, + 0.04880344495177269, + 0.23587976396083832, + 0.06863456219434738 + ], + "credit_map": [ + 0.020008377730846405, + 0.009885619394481182, + 0.01774396188557148, + 0.008840742520987988, + 0.018425079062581062, + 0.020353982225060463, + 0.010057070292532444, + 0.009694949723780155, + 0.006920219864696264, + 0.008700649254024029, + 0.005677576642483473, + 0.006316406186670065, + 0.005398524925112724, + 0.005232488736510277, + 0.004037714097648859, + 0.005310962907969952, + 0.00930831115692854, + 0.013561656698584557, + 0.008885322138667107, + 0.01204767357558012, + 0.009130548685789108, + 0.006355199031531811, + 0.006494775880128145, + 0.004198301583528519, + 0.004287631716579199, + 0.003774831537157297, + 0.004555956460535526, + 0.004050555173307657, + 0.006356508005410433, + 0.00684399576857686, + 0.0037677143700420856, + 0.00529526500031352, + 0.008468222804367542, + 0.013977828435599804, + 0.008751455694437027, + 0.007954714819788933, + 0.09957357496023178, + 0.06752276420593262, + 0.018281707540154457, + 0.021496456116437912, + 0.008655250072479248, + 0.01126206386834383, + 0.006931204814463854, + 0.010083537548780441, + 0.010944169014692307, + 0.016631944105029106, + 0.015524530783295631, + 0.012430476024746895, + 0.003830358386039734, + 0.0039251395501196384, + 0.006079561542719603, + 0.005082970019429922, + 0.00457776291295886, + 0.00537464814260602, + 0.003852518741041422, + 0.003950791899114847, + 0.004570562392473221, + 0.005751009099185467, + 0.004280791152268648, + 0.005844261031597853, + 0.010000013746321201, + 0.007815140299499035, + 0.004376950673758984, + 0.006164709106087685, + 0.025357874110341072, + 0.028798777610063553, + 0.020289655774831772, + 0.0183254461735487, + 0.0467727892100811, + 0.034830961376428604, + 0.021923616528511047, + 0.03958062455058098, + 0.009827018715441227, + 0.009287812747061253, + 0.0037631916347891092, + 0.010083100758492947, + 0.010598049499094486, + 0.01076669991016388, + 0.006368103437125683, + 0.00794057734310627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.012563489377498627 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351535692811012 + ], + "opposite_logprob_tokens": [ + -3.6513514518737793 + ], + "logprob_reward": -0.026351535692811012, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351535692811012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10790059715509415, + 0.042325545102357864, + 0.04501057788729668, + 0.02404232695698738, + 0.2791467308998108, + 0.058370087295770645, + 0.06596536189317703, + 0.06695867329835892, + 0.21491257846355438, + 0.09536749869585037 + ], + "credit_map": [ + 0.021917397156357765, + 0.009918642230331898, + 0.015317265875637531, + 0.00909215398132801, + 0.01119461003690958, + 0.02151571214199066, + 0.009537795558571815, + 0.009407013654708862, + 0.0071852002292871475, + 0.006646363064646721, + 0.006863338872790337, + 0.007054812740534544, + 0.00353577290661633, + 0.004705773200839758, + 0.0035013998858630657, + 0.0028328849002718925, + 0.008068421855568886, + 0.007653420325368643, + 0.005812902469187975, + 0.007040038704872131, + 0.00541316531598568, + 0.0037565515376627445, + 0.004207757767289877, + 0.0030583194456994534, + 0.0028668909799307585, + 0.0028980837669223547, + 0.002842199755832553, + 0.0025931226555258036, + 0.003099556313827634, + 0.0033993299584835768, + 0.0028139944188296795, + 0.0035291477106511593, + 0.010611304081976414, + 0.02002842165529728, + 0.012345564551651478, + 0.009144717827439308, + 0.10689098387956619, + 0.06590854376554489, + 0.019232431426644325, + 0.0349847674369812, + 0.008575043641030788, + 0.006908138748258352, + 0.0050635128282010555, + 0.00441158190369606, + 0.007157228421419859, + 0.008894383907318115, + 0.011258581653237343, + 0.006101613864302635, + 0.005312598776072264, + 0.0075301737524569035, + 0.00964322965592146, + 0.007352101616561413, + 0.01195810828357935, + 0.013524391688406467, + 0.004806418437510729, + 0.005838337354362011, + 0.004719472490251064, + 0.006716713774949312, + 0.004913972225040197, + 0.008941580541431904, + 0.01416509598493576, + 0.010994135402143002, + 0.005107633303850889, + 0.011400071904063225, + 0.02124839648604393, + 0.01718074455857277, + 0.02060089446604252, + 0.014156284742057323, + 0.03156591206789017, + 0.03867214918136597, + 0.020021334290504456, + 0.05146687105298042, + 0.011790461838245392, + 0.01313320454210043, + 0.006023861467838287, + 0.01662031002342701, + 0.017442788928747177, + 0.014416145160794258, + 0.007296712137758732, + 0.008644010871648788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.026351535692811012 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic medium shot, shot in the style of sports photography, slowly arcs around a figure seated on the massive, dark wooden judge's bench within the moot courtroom. Clad in a heavy, matte-fabric black judge's robe, the figure maintains a posture of intense concentration, their face partially obscured by dramatic, high-contrast lighting. A harsh spotlight, shining from above and sideways, carves them out of the surrounding darkness, highlighting the tension in their powerful shoulders and forearms. A sheet of white paper rests on the bench's polished surface. The figure's hand, gripping a traditional bamboo-handled calligraphy brush, moves with the explosive grace of a high-performance athlete; they execute powerful, deliberate strokes, the brush tip gliding across the paper with controlled force. The camera captures every microscopic detail with sharp, dramatic focus: the slight tension of the knuckles, the placement of the glistening black ink, and the final, quick flick of the wrist that completes the act with championship-winning precision, creating an atmosphere of intense concentration and physical prowess.", + "seed": 26471, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_7/index_7523_seed_26471_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03824483621865511, + "vlm_logprob": -0.03824483621865511, + "vlm_yes_no_margin": 3.549999952316284, + "vlm_token_credit": -0.03824483621865511, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08201292902231216 + ], + "opposite_logprob_tokens": [ + -2.832012891769409 + ], + "logprob_reward": -0.08201292902231216, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08201292902231216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10653867572546005, + 0.11158376932144165, + 0.12505406141281128, + 0.025760576128959656, + 0.2588344216346741, + 0.031103407964110374, + 0.05696175619959831, + 0.02089160494506359, + 0.03346628323197365, + 0.22980542480945587 + ], + "credit_map": [ + 0.01729327253997326, + 0.014551523141562939, + 0.009671801701188087, + 0.010122111067175865, + 0.014213158749043941, + 0.017180021852254868, + 0.010875011794269085, + 0.01263177115470171, + 0.008375654928386211, + 0.021804263815283775, + 0.006613614037632942, + 0.0074392324313521385, + 0.015697121620178223, + 0.020657997578382492, + 0.015698503702878952, + 0.01529738586395979, + 0.008782860822975636, + 0.01600188948214054, + 0.01642351597547531, + 0.02489025704562664, + 0.017031481489539146, + 0.01940067484974861, + 0.012721147388219833, + 0.009802229702472687, + 0.0029151130001991987, + 0.0032320984173566103, + 0.003399098291993141, + 0.0032022001687437296, + 0.0044927531853318214, + 0.0038541124667972326, + 0.002417956944555044, + 0.0022472431883215904, + 0.031012270599603653, + 0.0400010347366333, + 0.021447908133268356, + 0.02424156665802002, + 0.04947516694664955, + 0.04826138913631439, + 0.01955508626997471, + 0.02484002523124218, + 0.005810389760881662, + 0.003801519749686122, + 0.0018393220379948616, + 0.002351027447730303, + 0.004792128689587116, + 0.0046025230549275875, + 0.0034346990287303925, + 0.004471798427402973, + 0.005326007958501577, + 0.007448342628777027, + 0.0025268320459872484, + 0.004125380888581276, + 0.014843111857771873, + 0.011292093433439732, + 0.004579123109579086, + 0.00682086544111371, + 0.002506174612790346, + 0.0023939672391861677, + 0.0012411788338795304, + 0.0013285083696246147, + 0.005813802592456341, + 0.0026536516379565, + 0.0018340138485655189, + 0.003120308741927147, + 0.004322766326367855, + 0.0032555609941482544, + 0.0022809214424341917, + 0.0032221719156950712, + 0.005074638407677412, + 0.003941184375435114, + 0.006868462078273296, + 0.004500579088926315, + 0.008017257787287235, + 0.01035461388528347, + 0.016045788303017616, + 0.07368192076683044, + 0.02283201739192009, + 0.0176845695823431, + 0.025743141770362854, + 0.055446114391088486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.08201292902231216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018237756565213203 + ], + "opposite_logprob_tokens": [ + -4.018237590789795 + ], + "logprob_reward": -0.018237756565213203, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018237756565213203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06562793999910355, + 0.12631115317344666, + 0.20057575404644012, + 0.07491055876016617, + 0.16171564161777496, + 0.027137577533721924, + 0.044943396002054214, + 0.02389855869114399, + 0.026586130261421204, + 0.24829325079917908 + ], + "credit_map": [ + 0.008907395415008068, + 0.007350360509008169, + 0.0052726902067661285, + 0.005750798154622316, + 0.009148523211479187, + 0.013148597441613674, + 0.006761564873158932, + 0.009288007393479347, + 0.008179725147783756, + 0.02282540686428547, + 0.007186947390437126, + 0.010286916047334671, + 0.019506651908159256, + 0.018118008971214294, + 0.021880576387047768, + 0.01832692325115204, + 0.013274155557155609, + 0.024376610293984413, + 0.021983018144965172, + 0.039772626012563705, + 0.03315098211169243, + 0.03188461437821388, + 0.019790813326835632, + 0.016342930495738983, + 0.0063519966788589954, + 0.010615960694849491, + 0.006380376871675253, + 0.011017851531505585, + 0.014260669238865376, + 0.011797209270298481, + 0.008076030761003494, + 0.006410463247448206, + 0.012062217108905315, + 0.025060715153813362, + 0.00946750771254301, + 0.013757636770606041, + 0.04135121777653694, + 0.029187755659222603, + 0.013393212109804153, + 0.01743539236485958, + 0.0061567057855427265, + 0.002375552663579583, + 0.0012637368636205792, + 0.0020524419378489256, + 0.005077403038740158, + 0.004382622428238392, + 0.0025350372307002544, + 0.003294077469035983, + 0.004205480683594942, + 0.006164089310914278, + 0.0035913288593292236, + 0.003261259291321039, + 0.010876627638936043, + 0.007306579500436783, + 0.0034480937756597996, + 0.006089936941862106, + 0.0036366500426083803, + 0.004388553090393543, + 0.0012652111472561955, + 0.0017753159627318382, + 0.005721427500247955, + 0.0027859113179147243, + 0.0021328881848603487, + 0.0021926022600382566, + 0.002948894863948226, + 0.0033177926670759916, + 0.0018416487146168947, + 0.003308448940515518, + 0.0038782411720603704, + 0.0032657997217029333, + 0.004870429169386625, + 0.0031548752449452877, + 0.008450382389128208, + 0.010921943001449108, + 0.01578027568757534, + 0.07499749958515167, + 0.01906072162091732, + 0.02256549336016178, + 0.024746043607592583, + 0.07177089154720306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.018237756565213203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083611875772476 + ], + "opposite_logprob_tokens": [ + -3.168083667755127 + ], + "logprob_reward": -0.043083611875772476, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083611875772476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16456452012062073, + 0.09967568516731262, + 0.18187566101551056, + 0.14030958712100983, + 0.08157886564731598, + 0.029265640303492546, + 0.07209806144237518, + 0.03529737889766693, + 0.03400550037622452, + 0.16132909059524536 + ], + "credit_map": [ + 0.011820938438177109, + 0.01598457247018814, + 0.021400244906544685, + 0.02293795347213745, + 0.033156607300043106, + 0.02938036061823368, + 0.015266014263033867, + 0.014617819339036942, + 0.011212940327823162, + 0.021017974242568016, + 0.009327463805675507, + 0.012569214217364788, + 0.013857348822057247, + 0.013917251490056515, + 0.0071678608655929565, + 0.010605634190142155, + 0.012810098938643932, + 0.020517272874712944, + 0.0216520968824625, + 0.036878205835819244, + 0.025748714804649353, + 0.027611564844846725, + 0.019621390849351883, + 0.01703631319105625, + 0.01727810502052307, + 0.01293489895761013, + 0.0237048901617527, + 0.016398588195443153, + 0.022349905222654343, + 0.020958181470632553, + 0.016136441379785538, + 0.010548576712608337, + 0.00732446601614356, + 0.010633664205670357, + 0.005500799976289272, + 0.008082369342446327, + 0.017973631620407104, + 0.015124124474823475, + 0.0077035510912537575, + 0.009236257523298264, + 0.005831636488437653, + 0.002596966689452529, + 0.0021286143455654383, + 0.003007146529853344, + 0.005662740208208561, + 0.004688064567744732, + 0.002446949714794755, + 0.0029035210609436035, + 0.00732521852478385, + 0.01322179939597845, + 0.004465486854314804, + 0.00541646359488368, + 0.015160594135522842, + 0.011127272620797157, + 0.0053155249916017056, + 0.010065698064863682, + 0.005565524101257324, + 0.007331477012485266, + 0.0019856763537973166, + 0.0035228997003287077, + 0.006851461715996265, + 0.00365524273365736, + 0.0034057958982884884, + 0.002979300683364272, + 0.004194588400423527, + 0.0041087111458182335, + 0.00225111679174006, + 0.003299871226772666, + 0.004808561876416206, + 0.00429570022970438, + 0.00635952316224575, + 0.004687428940087557, + 0.00746107567101717, + 0.011182642541825771, + 0.009851482696831226, + 0.04406649246811867, + 0.01268930546939373, + 0.01684032566845417, + 0.015913361683487892, + 0.043324392288923264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.043083611875772476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795336984097958 + ], + "opposite_logprob_tokens": [ + -4.634795188903809 + ], + "logprob_reward": -0.009795336984097958, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795336984097958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09197740256786346, + 0.05688204616308212, + 0.03330637142062187, + 0.19036492705345154, + 0.22660072147846222, + 0.12474987655878067, + 0.029093975201249123, + 0.031247040256857872, + 0.07582100480794907, + 0.1399565488100052 + ], + "credit_map": [ + 0.007309687323868275, + 0.011897464282810688, + 0.012008097022771835, + 0.010051630437374115, + 0.01435455959290266, + 0.013048301450908184, + 0.0118415467441082, + 0.01146612223237753, + 0.005345925688743591, + 0.011238758452236652, + 0.003678393317386508, + 0.0038334380369633436, + 0.007592867594212294, + 0.009149317629635334, + 0.00574320275336504, + 0.010300152003765106, + 0.0038632277864962816, + 0.004830034449696541, + 0.0030642710626125336, + 0.002745596459135413, + 0.006635844707489014, + 0.004806557670235634, + 0.0031619705259799957, + 0.004198872484266758, + 0.02616288512945175, + 0.021766379475593567, + 0.021707884967327118, + 0.02174626663327217, + 0.03222452476620674, + 0.030395735055208206, + 0.016034187749028206, + 0.020327087491750717, + 0.027250289916992188, + 0.03885192796587944, + 0.01769646257162094, + 0.022771205753087997, + 0.03898422420024872, + 0.04097948223352432, + 0.018462911248207092, + 0.021604245528578758, + 0.022881491109728813, + 0.011564606800675392, + 0.006178451236337423, + 0.008865824900567532, + 0.0332133024930954, + 0.021318349987268448, + 0.00970807857811451, + 0.011019792407751083, + 0.001987097319215536, + 0.003146744566038251, + 0.0018753142794594169, + 0.002832618309184909, + 0.0042138490825891495, + 0.0047281947918236256, + 0.002986041596159339, + 0.0073241181671619415, + 0.00412428192794323, + 0.005799568258225918, + 0.0019201774848625064, + 0.003402834292501211, + 0.006450718734413385, + 0.002945853164419532, + 0.003207023488357663, + 0.003396584652364254, + 0.007523122243583202, + 0.008219867013394833, + 0.006274367216974497, + 0.008157123811542988, + 0.009821081534028053, + 0.008663508109748363, + 0.01566649042069912, + 0.011495454236865044, + 0.007490243297070265, + 0.008978191763162613, + 0.010217010974884033, + 0.04296733811497688, + 0.01532665453851223, + 0.012582189403474331, + 0.018708515912294388, + 0.023686423897743225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.009795336984097958 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a moot courtroom featuring a massive, dark wooden judge's bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a heavy, matte-fabric black judge's robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a traditional bamboo-handled calligraphy brush and making powerful strokes on a sheet of white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by a harsh spotlight from above and sideways creating high contrast and obscuring the figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly arc around the seated figure in a dynamic medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094546645879745 + ], + "opposite_logprob_tokens": [ + -3.2880945205688477 + ], + "logprob_reward": -0.038094546645879745, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094546645879745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10324722528457642, + 0.034397661685943604, + 0.17392463982105255, + 0.058925289660692215, + 0.11727472394704819, + 0.06448034197092056, + 0.03489304706454277, + 0.05987679213285446, + 0.09950210899114609, + 0.2534782290458679 + ], + "credit_map": [ + 0.010716981254518032, + 0.012923477217555046, + 0.015120182186365128, + 0.012238605879247189, + 0.015061993151903152, + 0.014847557991743088, + 0.01033913902938366, + 0.011999291367828846, + 0.0026232439558953047, + 0.007525152992457151, + 0.0028181837406009436, + 0.003632535459473729, + 0.004227014258503914, + 0.004797077737748623, + 0.004529428202658892, + 0.004245022777467966, + 0.012982913292944431, + 0.020289026200771332, + 0.016881810501217842, + 0.031711645424366, + 0.032866936177015305, + 0.029408952221274376, + 0.017277885228395462, + 0.01250546332448721, + 0.006676414981484413, + 0.005811039824038744, + 0.007956826128065586, + 0.0067484015598893166, + 0.009915489703416824, + 0.008611260913312435, + 0.005595620255917311, + 0.00761023024097085, + 0.013045544736087322, + 0.019549619406461716, + 0.009578922763466835, + 0.014332534745335579, + 0.019314700737595558, + 0.020596271380782127, + 0.009953321889042854, + 0.010903800837695599, + 0.007087388541549444, + 0.00545352790504694, + 0.004650224931538105, + 0.00544308265671134, + 0.012122043408453465, + 0.010686728172004223, + 0.00765342079102993, + 0.011383912526071072, + 0.0028603726532310247, + 0.004228208679705858, + 0.002221128437668085, + 0.0029595589730888605, + 0.00808071531355381, + 0.005998584907501936, + 0.003410112811252475, + 0.005134362727403641, + 0.006650603376328945, + 0.008244171738624573, + 0.0024810070171952248, + 0.00785729568451643, + 0.013793354853987694, + 0.007065553218126297, + 0.004895966034382582, + 0.008888832293450832, + 0.00868261605501175, + 0.007677880581468344, + 0.010034772567451, + 0.008599509485065937, + 0.01155814714729786, + 0.016003241762518883, + 0.01817290112376213, + 0.018773037940263748, + 0.011384270153939724, + 0.01778840459883213, + 0.0177568718791008, + 0.07528857886791229, + 0.03156473860144615, + 0.03044181689620018, + 0.02801493927836418, + 0.04123857989907265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.038094546645879745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_tensor([2982, 2982])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_tensor([2982, 2982])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..efb11ace46d5667db8e1fb25325e54487c4ec071 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_tensor([2982, 2982])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle close-up slowly zooms in on a tiny, articulated figure meticulously crafted from polished brass, standing atop the vast, reflective surface of a dark-varnished mahogany desk. A single, warm light source, emanating from the off-camera right, casts long shadows and creates brilliant golden shimmers along the figure's bent back and limbs, highlighting the smooth metallic texture against the dark brown of the wood. With mechanically graceful and reverent slowness, the figure begins to prostrate; its torso articulated forward from the waist, its articulated arms steadily extended, its small, palm-down hands resting flat on the polished surface. The movement culminates as its simple, spherical head bends downward, gently resting on the desk. Its entire deliberate movement is perfectly mirrored in the dark, polished wood below, creating a stunning, distorted reflection that doubles the silent, solemn ritual and imbues the scene with a profound atmosphere of quiet contemplation.", + "seed": 18772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_2982_seed_18772_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01519549764925614, + "vlm_logprob": -0.01519549764925614, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.01519549764925614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024165021255612373 + ], + "opposite_logprob_tokens": [ + -4.024165153503418 + ], + "logprob_reward": -0.024165021255612373, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024165021255612373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14158985018730164, + 0.16412168741226196, + 0.08408625423908234, + 0.05801156535744667, + 0.01995180919766426, + 0.10893377661705017, + 0.08283472806215286, + 0.15716685354709625, + 0.0440121553838253, + 0.13929136097431183 + ], + "credit_map": [ + 0.01823478937149048, + 0.016361309215426445, + 0.017741581425070763, + 0.011658814735710621, + 0.012936320155858994, + 0.022614292800426483, + 0.01881875842809677, + 0.023223984986543655, + 0.019857939332723618, + 0.022038334980607033, + 0.01808912307024002, + 0.021080737933516502, + 0.014311039820313454, + 0.019264627248048782, + 0.020002637058496475, + 0.029477238655090332, + 0.0062346081249415874, + 0.011195829138159752, + 0.009716746397316456, + 0.007095036096870899, + 0.01421015989035368, + 0.012806618586182594, + 0.008493507280945778, + 0.014333744533360004, + 0.004428009036928415, + 0.006222289055585861, + 0.008587748743593693, + 0.005947544239461422, + 0.008481678552925587, + 0.008022109977900982, + 0.007540930528193712, + 0.008781259879469872, + 0.002170614432543516, + 0.0018262012163177133, + 0.0031152316369116306, + 0.0023795627057552338, + 0.0022229840978980064, + 0.002481639850884676, + 0.002781619317829609, + 0.002973957220092416, + 0.00821902696043253, + 0.011272829957306385, + 0.017663398757576942, + 0.011790110729634762, + 0.01772184669971466, + 0.014229316264390945, + 0.013657850213348866, + 0.014379399828612804, + 0.005921820178627968, + 0.008426617830991745, + 0.01499059610068798, + 0.0074083940126001835, + 0.008005749434232712, + 0.01311855111271143, + 0.01088140532374382, + 0.014081590808928013, + 0.01298773568123579, + 0.019378045573830605, + 0.01681448519229889, + 0.017912253737449646, + 0.03458477556705475, + 0.018765069544315338, + 0.017970552667975426, + 0.018753938376903534, + 0.004912498407065868, + 0.006444364320486784, + 0.006384904030710459, + 0.005717087537050247, + 0.005633439403027296, + 0.004779943730682135, + 0.005837867967784405, + 0.004302048124372959, + 0.011473924852907658, + 0.01602405123412609, + 0.021769670769572258, + 0.011560100130736828, + 0.02363429218530655, + 0.020911816507577896, + 0.015245180577039719, + 0.018672334030270576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.024165021255612373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549597769975662 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549597769975662, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549597769975662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08174759894609451, + 0.08808755874633789, + 0.11779666692018509, + 0.03308599814772606, + 0.039248231798410416, + 0.08416412025690079, + 0.0879271999001503, + 0.17097114026546478, + 0.04885013401508331, + 0.24812130630016327 + ], + "credit_map": [ + 0.01102926954627037, + 0.012174120172858238, + 0.010581249371170998, + 0.008059575222432613, + 0.006295871455222368, + 0.010713274590671062, + 0.009634946472942829, + 0.01325930468738079, + 0.009823447093367577, + 0.01004953496158123, + 0.011147642508149147, + 0.013147692196071148, + 0.006728848442435265, + 0.015376104041934013, + 0.006312808487564325, + 0.015501493588089943, + 0.008489005267620087, + 0.01682354137301445, + 0.012585446238517761, + 0.011479430831968784, + 0.017506958916783333, + 0.01721024699509144, + 0.013601453974843025, + 0.02010059729218483, + 0.001953310798853636, + 0.0030860970728099346, + 0.004384891130030155, + 0.0030110932420939207, + 0.004714881535619497, + 0.00715363584458828, + 0.0034395172260701656, + 0.005342575255781412, + 0.004074159078299999, + 0.0031377808190882206, + 0.005360663868486881, + 0.0038574556820094585, + 0.005036246497184038, + 0.004527601413428783, + 0.007246820721775293, + 0.006007509306073189, + 0.007412400096654892, + 0.00836154818534851, + 0.014323145151138306, + 0.010279672220349312, + 0.012319211848080158, + 0.009300746954977512, + 0.011381053365767002, + 0.010786347091197968, + 0.008950671181082726, + 0.010372977703809738, + 0.012303548865020275, + 0.008593502454459667, + 0.01068031508475542, + 0.012222004123032093, + 0.011322399601340294, + 0.013481788337230682, + 0.02345370501279831, + 0.015410109423100948, + 0.023516034707427025, + 0.03321133553981781, + 0.0146184042096138, + 0.015809819102287292, + 0.025071123614907265, + 0.019880615174770355, + 0.004620784427970648, + 0.006480210926383734, + 0.007950739935040474, + 0.007576274685561657, + 0.006234589032828808, + 0.005251394119113684, + 0.005049786064773798, + 0.005686361342668533, + 0.03009219467639923, + 0.022662946954369545, + 0.040463559329509735, + 0.03609410300850868, + 0.018124189227819443, + 0.02048899605870247, + 0.03308111056685448, + 0.04711423069238663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.012549597769975662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369913622736931 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.03369913622736931, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369913622736931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0845068022608757, + 0.0690796971321106, + 0.08014701306819916, + 0.07190937548875809, + 0.06565240770578384, + 0.2900557219982147, + 0.16286695003509521, + 0.09120482951402664, + 0.04683035612106323, + 0.037746794521808624 + ], + "credit_map": [ + 0.006945382338017225, + 0.00994910392910242, + 0.007345042657107115, + 0.01129663921892643, + 0.01082120556384325, + 0.00880458578467369, + 0.015157912857830524, + 0.014186940155923367, + 0.005659159738570452, + 0.005164372269064188, + 0.006990951020270586, + 0.009564831852912903, + 0.0069329384714365005, + 0.01240348257124424, + 0.008569842204451561, + 0.013794125989079475, + 0.005925279576331377, + 0.007941764779388905, + 0.009188971482217312, + 0.011176057159900665, + 0.009519057348370552, + 0.00943957269191742, + 0.01384150143712759, + 0.01311482209712267, + 0.0051133399829268456, + 0.006273046135902405, + 0.006854402832686901, + 0.004458938725292683, + 0.01996731385588646, + 0.01308162696659565, + 0.006956629920750856, + 0.009204086847603321, + 0.0066529070027172565, + 0.006982648279517889, + 0.011171722784638405, + 0.008637918159365654, + 0.004554197192192078, + 0.007380875758826733, + 0.00843874178826809, + 0.01183339674025774, + 0.024647461250424385, + 0.031022988259792328, + 0.053431276232004166, + 0.028633877635002136, + 0.03792781010270119, + 0.04256405308842659, + 0.030658727511763573, + 0.04116956889629364, + 0.015731552615761757, + 0.017409952357411385, + 0.023466261103749275, + 0.015242834575474262, + 0.020366238430142403, + 0.018140116706490517, + 0.0188894160091877, + 0.033620599657297134, + 0.00932393129914999, + 0.008185351267457008, + 0.0156268123537302, + 0.009040682576596737, + 0.012454933486878872, + 0.013334211893379688, + 0.01172749511897564, + 0.01151142455637455, + 0.0038828346878290176, + 0.004437194671481848, + 0.007085271179676056, + 0.007659852039068937, + 0.005815559066832066, + 0.0057627069763839245, + 0.006242868956178427, + 0.005944074131548405, + 0.0037170478608459234, + 0.004799181595444679, + 0.00528414873406291, + 0.004847806878387928, + 0.005179575178772211, + 0.004113653674721718, + 0.0042044539004564285, + 0.005600932519882917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.03369913622736931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24225689470767975, + 0.0703236386179924, + 0.13451869785785675, + 0.1973119080066681, + 0.03408019244670868, + 0.09956391155719757, + 0.07863996922969818, + 0.04574248567223549, + 0.07240758091211319, + 0.025154724717140198 + ], + "credit_map": [ + 0.023808462545275688, + 0.020418018102645874, + 0.029723988845944405, + 0.028324607759714127, + 0.029537491500377655, + 0.025426559150218964, + 0.046198777854442596, + 0.03881899639964104, + 0.0055580479092895985, + 0.006553370039910078, + 0.007240549195557833, + 0.011989007703959942, + 0.007177966646850109, + 0.008352077566087246, + 0.00980081595480442, + 0.013651804998517036, + 0.010697457008063793, + 0.014281372539699078, + 0.019813239574432373, + 0.017982468008995056, + 0.02020578272640705, + 0.019168401136994362, + 0.013252915814518929, + 0.019117066636681557, + 0.01649346947669983, + 0.01837252825498581, + 0.01800171099603176, + 0.0148294223472476, + 0.04655207321047783, + 0.033749230206012726, + 0.02461719512939453, + 0.024696288630366325, + 0.00314838788472116, + 0.002210342325270176, + 0.005843249149620533, + 0.003046248108148575, + 0.003616215195506811, + 0.004111927468329668, + 0.005828951019793749, + 0.006274868734180927, + 0.006529198493808508, + 0.009050657972693443, + 0.01172303780913353, + 0.01319375354796648, + 0.011718391440808773, + 0.017445743083953857, + 0.01623961515724659, + 0.013663511723279953, + 0.009319692850112915, + 0.008022675290703773, + 0.012085039168596268, + 0.0076398118399083614, + 0.010656493715941906, + 0.009345179423689842, + 0.009705085307359695, + 0.011865993030369282, + 0.00395759753882885, + 0.0044091870076954365, + 0.007869157008826733, + 0.006212348584085703, + 0.005680253263562918, + 0.0058848559856414795, + 0.004896839149296284, + 0.00683224480599165, + 0.01046143751591444, + 0.006747884210199118, + 0.011356502771377563, + 0.008709284476935863, + 0.007291500456631184, + 0.010918526910245419, + 0.00697523495182395, + 0.009947205893695354, + 0.0021121161989867687, + 0.003453596029430628, + 0.0037293361965566874, + 0.0025016965810209513, + 0.004470647778362036, + 0.003574166912585497, + 0.0025454622227698565, + 0.00276770256459713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.004640049301087856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009236836922354996 + ], + "opposite_logprob_tokens": [ + -7.0009236335754395 + ], + "logprob_reward": -0.0009236836922354996, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009236836922354996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15666714310646057, + 0.09299936890602112, + 0.09174052625894547, + 0.11679208278656006, + 0.18209686875343323, + 0.06372592598199844, + 0.08306051045656204, + 0.08829624205827713, + 0.057657741010189056, + 0.06696358323097229 + ], + "credit_map": [ + 0.014763586223125458, + 0.01421796903014183, + 0.016635047271847725, + 0.020817704498767853, + 0.01837611198425293, + 0.017148450016975403, + 0.02857745811343193, + 0.026130815967917442, + 0.008702949620783329, + 0.009601226076483727, + 0.009999117814004421, + 0.012001236900687218, + 0.010334459133446217, + 0.014096121303737164, + 0.012572783045470715, + 0.015691466629505157, + 0.008039298467338085, + 0.010420942679047585, + 0.011602952145040035, + 0.009430485777556896, + 0.012414418160915375, + 0.013091611675918102, + 0.010283263400197029, + 0.016457557678222656, + 0.007899615913629532, + 0.00998231116682291, + 0.01148622203618288, + 0.017515650019049644, + 0.019945232197642326, + 0.01763051562011242, + 0.016931891441345215, + 0.015400652773678303, + 0.01745668053627014, + 0.009895936585962772, + 0.025816112756729126, + 0.0251658596098423, + 0.01622595638036728, + 0.01852472685277462, + 0.04057140275835991, + 0.0284401997923851, + 0.005057819653302431, + 0.006819157395511866, + 0.007123843766748905, + 0.00857225526124239, + 0.009063362143933773, + 0.008073220029473305, + 0.011073729954659939, + 0.007942537777125835, + 0.00897156074643135, + 0.008524304255843163, + 0.01530040055513382, + 0.007618539966642857, + 0.00792392622679472, + 0.012422865256667137, + 0.008400317281484604, + 0.013898596167564392, + 0.007230828050523996, + 0.00807294249534607, + 0.012663960456848145, + 0.010199030861258507, + 0.011865000240504742, + 0.010683356784284115, + 0.013261147774755955, + 0.014319979585707188, + 0.00692918011918664, + 0.006250737700611353, + 0.00868204329162836, + 0.006174576468765736, + 0.007084429729729891, + 0.009913839399814606, + 0.0048548756167292595, + 0.007768059615045786, + 0.00510796345770359, + 0.0066856020130217075, + 0.01068613026291132, + 0.008713515475392342, + 0.009561766870319843, + 0.007351386826485395, + 0.009191924706101418, + 0.00966529082506895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0009236836922354996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle close-up slowly zooms in on a tiny, articulated figure meticulously crafted from polished brass, standing atop the vast, reflective surface of a dark-varnished mahogany desk. A single, warm light source, emanating from the off-camera right, casts long shadows and creates brilliant golden shimmers along the figure's bent back and limbs, highlighting the smooth metallic texture against the dark brown of the wood. With mechanically graceful and reverent slowness, the figure begins to prostrate; its torso articulated forward from the waist, its articulated arms steadily extended, its small, palm-down hands resting flat on the polished surface. The movement culminates as its simple, spherical head bends downward, gently resting on the desk. Its entire deliberate movement is perfectly mirrored in the dark, polished wood below, creating a stunning, distorted reflection that doubles the silent, solemn ritual and imbues the scene with a profound atmosphere of quiet contemplation.", + "seed": 18773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_2982_seed_18773_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011719550145789981, + "vlm_logprob": -0.011719550145789981, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.011719550145789981, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0184331014752388 + ], + "opposite_logprob_tokens": [ + -4.768433094024658 + ], + "logprob_reward": -0.0184331014752388, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0184331014752388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04695897176861763, + 0.07472295314073563, + 0.05770416930317879, + 0.20579633116722107, + 0.2734107971191406, + 0.05850313976407051, + 0.040103472769260406, + 0.03754754364490509, + 0.14525525271892548, + 0.05999737232923508 + ], + "credit_map": [ + 0.0062419213354587555, + 0.006983471568673849, + 0.009979094378650188, + 0.003270233515650034, + 0.0037799489218741655, + 0.007353122811764479, + 0.003690144745633006, + 0.00566103495657444, + 0.008922098204493523, + 0.012458874844014645, + 0.011951567605137825, + 0.0033676475286483765, + 0.010098562575876713, + 0.008295484818518162, + 0.003931536339223385, + 0.0156971774995327, + 0.005250350572168827, + 0.006349631119519472, + 0.015293802134692669, + 0.003262478159740567, + 0.004779300652444363, + 0.004958712495863438, + 0.0026806178502738476, + 0.015129277482628822, + 0.019803082570433617, + 0.010942453518509865, + 0.07460299134254456, + 0.010539642535150051, + 0.012705392204225063, + 0.01187413651496172, + 0.008156870491802692, + 0.05717175453901291, + 0.04622684046626091, + 0.013846919871866703, + 0.05739691108465195, + 0.014849932864308357, + 0.020347587764263153, + 0.02091493085026741, + 0.0131036676466465, + 0.08672400563955307, + 0.008718576282262802, + 0.006102702580392361, + 0.008885563351213932, + 0.012184888124465942, + 0.006329959724098444, + 0.0059638721868395805, + 0.004672591108828783, + 0.005644986405968666, + 0.006430603563785553, + 0.005541943944990635, + 0.010426654480397701, + 0.0036109634675085545, + 0.0027161615435034037, + 0.0029321920592337847, + 0.0029411709401756525, + 0.0055037811398506165, + 0.006903356406837702, + 0.004672776907682419, + 0.007725072093307972, + 0.0051908232271671295, + 0.0020956804510205984, + 0.002840383443981409, + 0.003336693625897169, + 0.004782760050147772, + 0.020058274269104004, + 0.015724454075098038, + 0.019377127289772034, + 0.021553516387939453, + 0.02020340971648693, + 0.009412993676960468, + 0.02326306886970997, + 0.01566240005195141, + 0.014347899705171585, + 0.006694603245705366, + 0.009462517686188221, + 0.007434165105223656, + 0.007577876560389996, + 0.005170759744942188, + 0.004939809907227755, + 0.004369737580418587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0184331014752388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022129351273179054 + ], + "opposite_logprob_tokens": [ + -6.127213001251221 + ], + "logprob_reward": -0.0022129351273179054, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022129351273179054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15080192685127258, + 0.12196458131074905, + 0.07233794778585434, + 0.07300513237714767, + 0.20746703445911407, + 0.06837640702724457, + 0.07294737547636032, + 0.07885977625846863, + 0.10751967132091522, + 0.04672016203403473 + ], + "credit_map": [ + 0.020029481500387192, + 0.026561103761196136, + 0.020083611831068993, + 0.015578758902847767, + 0.015888607129454613, + 0.015950020402669907, + 0.018199659883975983, + 0.018510695546865463, + 0.01657607965171337, + 0.008050375618040562, + 0.021929051727056503, + 0.006587379612028599, + 0.01277301087975502, + 0.016566220670938492, + 0.0071112774312496185, + 0.03237118199467659, + 0.006836554501205683, + 0.006603254470974207, + 0.022351376712322235, + 0.003607736434787512, + 0.005265045445412397, + 0.004583248868584633, + 0.0026720671448856592, + 0.02041865698993206, + 0.007976816035807133, + 0.004494012799113989, + 0.019352683797478676, + 0.006034187041223049, + 0.005058009177446365, + 0.01088248286396265, + 0.004681084770709276, + 0.014525855891406536, + 0.028116760775446892, + 0.011068413034081459, + 0.048558853566646576, + 0.015294711105525494, + 0.013047140091657639, + 0.019319450482726097, + 0.00835391879081726, + 0.06370777636766434, + 0.012957335449755192, + 0.0056792739778757095, + 0.011292688548564911, + 0.016047311946749687, + 0.0041931807063519955, + 0.006159678101539612, + 0.00462565990164876, + 0.007421274669468403, + 0.015289616771042347, + 0.0058832657523453236, + 0.019334742799401283, + 0.008378105238080025, + 0.0047764559276402, + 0.004197925329208374, + 0.004298035986721516, + 0.010789230465888977, + 0.013278358615934849, + 0.010234775952994823, + 0.015173199586570263, + 0.011110039427876472, + 0.006722056306898594, + 0.007853763177990913, + 0.006653396412730217, + 0.007834184914827347, + 0.023757463321089745, + 0.01599276438355446, + 0.015305154025554657, + 0.019467344507575035, + 0.007717359345406294, + 0.005797313526272774, + 0.008645662106573582, + 0.010836610570549965, + 0.0094650499522686, + 0.004683531820774078, + 0.007505226880311966, + 0.005820396821945906, + 0.006496046204119921, + 0.004116390831768513, + 0.005411603022366762, + 0.003221913706511259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0022129351273179054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033696141093969345 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.033696141093969345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033696141093969345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11650890111923218, + 0.05559314787387848, + 0.08616729080677032, + 0.27361732721328735, + 0.3388368487358093, + 0.01875060237944126, + 0.03443498909473419, + 0.04041039198637009, + 0.021768955513834953, + 0.013911587186157703 + ], + "credit_map": [ + 0.015169601887464523, + 0.023576702922582626, + 0.016090650111436844, + 0.011281986720860004, + 0.01379782147705555, + 0.009734008461236954, + 0.011573073454201221, + 0.015285063534975052, + 0.007557420060038567, + 0.009922645054757595, + 0.010771697387099266, + 0.002552282065153122, + 0.0059136757627129555, + 0.005746586713939905, + 0.003825852880254388, + 0.009302986785769463, + 0.00754153449088335, + 0.008780018426477909, + 0.024143103510141373, + 0.004663317929953337, + 0.007165844086557627, + 0.006775643676519394, + 0.0036204010248184204, + 0.023477423936128616, + 0.04098058119416237, + 0.010962017811834812, + 0.08041542023420334, + 0.01612209714949131, + 0.018261464312672615, + 0.039306242018938065, + 0.00974042434245348, + 0.057829055935144424, + 0.05364350974559784, + 0.019612228497862816, + 0.06973514705896378, + 0.03520895913243294, + 0.028966430574655533, + 0.04957117885351181, + 0.022219736129045486, + 0.059879645705223083, + 0.002401295118033886, + 0.0020017053466290236, + 0.0030242002103477716, + 0.00243099476210773, + 0.00321207451634109, + 0.0019641125109046698, + 0.0017126124585047364, + 0.0020036085043102503, + 0.00628080265596509, + 0.005768606439232826, + 0.008312513120472431, + 0.003643125295639038, + 0.002570411656051874, + 0.0019052040297538042, + 0.0016755464021116495, + 0.004278781358152628, + 0.00879402831196785, + 0.004955703858286142, + 0.007506371010094881, + 0.0048731854185462, + 0.003738729516044259, + 0.004081320017576218, + 0.0034390510991215706, + 0.0030220006592571735, + 0.0036455730441957712, + 0.002897571539506316, + 0.0038188400212675333, + 0.002906959969550371, + 0.002113350434228778, + 0.0013575665652751923, + 0.003133984049782157, + 0.0018951102392747998, + 0.0031178821809589863, + 0.0016487004468217492, + 0.0024784384295344353, + 0.002162329154089093, + 0.0013736555119976401, + 0.0011519138934090734, + 0.0010782722383737564, + 0.0009003955055959523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.033696141093969345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11838079243898392, + 0.1307075172662735, + 0.11129604279994965, + 0.16606470942497253, + 0.12240787595510483, + 0.17794740200042725, + 0.051725588738918304, + 0.0221562460064888, + 0.059472598135471344, + 0.039841312915086746 + ], + "credit_map": [ + 0.017963619902729988, + 0.01282308716326952, + 0.02094164863228798, + 0.009813803248107433, + 0.009910195134580135, + 0.015821082517504692, + 0.01506820134818554, + 0.01603914424777031, + 0.015609879046678543, + 0.03030327335000038, + 0.01934153214097023, + 0.006774775218218565, + 0.017124822363257408, + 0.011986790224909782, + 0.007725271862000227, + 0.02184116095304489, + 0.010384056717157364, + 0.00948398932814598, + 0.03640782833099365, + 0.005221259780228138, + 0.00594277773052454, + 0.007428287994116545, + 0.003919020760804415, + 0.03250880911946297, + 0.015126594342291355, + 0.010546342469751835, + 0.048299722373485565, + 0.013693156652152538, + 0.015743941068649292, + 0.013430532068014145, + 0.01052277535200119, + 0.0387016236782074, + 0.03485405072569847, + 0.007632013875991106, + 0.016642898321151733, + 0.006540198810398579, + 0.010345123708248138, + 0.010775388218462467, + 0.005550100933760405, + 0.030068086460232735, + 0.03218439966440201, + 0.014649251475930214, + 0.030505698174238205, + 0.04018186777830124, + 0.013635854236781597, + 0.016813479363918304, + 0.012628366239368916, + 0.01734844408929348, + 0.008181612007319927, + 0.007147366181015968, + 0.008523222990334034, + 0.006821413990110159, + 0.005201119929552078, + 0.003968445118516684, + 0.0043165916576981544, + 0.007565812673419714, + 0.004132942296564579, + 0.0018358489032834768, + 0.005374820437282324, + 0.0024205425288528204, + 0.0018049994250759482, + 0.0024299947544932365, + 0.0019078750628978014, + 0.0022492199204862118, + 0.008543591015040874, + 0.005924934055656195, + 0.00795653648674488, + 0.009232845157384872, + 0.007640620227903128, + 0.0047263531014323235, + 0.008348191156983376, + 0.007099523209035397, + 0.005918588954955339, + 0.003434795420616865, + 0.006058516912162304, + 0.005162564106285572, + 0.006552672479301691, + 0.003935919608920813, + 0.0050462642684578896, + 0.0037319858092814684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.0036176491994410753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006379238329827785 + ], + "opposite_logprob_tokens": [ + -7.375638008117676 + ], + "logprob_reward": -0.0006379238329827785, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006379238329827785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08395490795373917, + 0.07532021403312683, + 0.1874280869960785, + 0.14161251485347748, + 0.19184798002243042, + 0.03748849779367447, + 0.06263101100921631, + 0.06587610393762589, + 0.08238961547613144, + 0.07145107537508011 + ], + "credit_map": [ + 0.011063093319535255, + 0.011548473499715328, + 0.01325203850865364, + 0.006661220453679562, + 0.0067140343599021435, + 0.012650012038648129, + 0.007932444103062153, + 0.014133589342236519, + 0.011846084147691727, + 0.014473960734903812, + 0.00991904828697443, + 0.00516592338681221, + 0.00801155623048544, + 0.010636323131620884, + 0.004621060565114021, + 0.010646258480846882, + 0.01979833096265793, + 0.011551927775144577, + 0.065882608294487, + 0.00815337710082531, + 0.009259669110178947, + 0.011992791667580605, + 0.006368336267769337, + 0.054421041160821915, + 0.025211147964000702, + 0.005826624110341072, + 0.033349934965372086, + 0.007578175514936447, + 0.008404264226555824, + 0.024883650243282318, + 0.005172717850655317, + 0.031186003237962723, + 0.028940455988049507, + 0.011768316850066185, + 0.03526176139712334, + 0.020958958193659782, + 0.0149643924087286, + 0.027644596993923187, + 0.013752609491348267, + 0.038556892424821854, + 0.007213703356683254, + 0.0032599642872810364, + 0.006634716875851154, + 0.008431660942733288, + 0.0023729519452899694, + 0.003356337547302246, + 0.0024464621674269438, + 0.003772701835259795, + 0.011075162328779697, + 0.007848279550671577, + 0.01501734834164381, + 0.006076504476368427, + 0.0050094579346477985, + 0.004929365124553442, + 0.003511522663757205, + 0.009163366630673409, + 0.014196528121829033, + 0.008125616237521172, + 0.011606650426983833, + 0.008226488716900349, + 0.005333234090358019, + 0.006107761058956385, + 0.006836297456175089, + 0.0054435222409665585, + 0.016295041888952255, + 0.012235235422849655, + 0.011504396796226501, + 0.01216503232717514, + 0.007718689274042845, + 0.005340734962373972, + 0.009978968650102615, + 0.007151514291763306, + 0.014534653164446354, + 0.008768596686422825, + 0.01093275286257267, + 0.009112541563808918, + 0.007908100262284279, + 0.006080003455281258, + 0.007108396850526333, + 0.007006026804447174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.0006379238329827785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle close-up slowly zooms in on a tiny, articulated figure meticulously crafted from polished brass, standing atop the vast, reflective surface of a dark-varnished mahogany desk. A single, warm light source, emanating from the off-camera right, casts long shadows and creates brilliant golden shimmers along the figure's bent back and limbs, highlighting the smooth metallic texture against the dark brown of the wood. With mechanically graceful and reverent slowness, the figure begins to prostrate; its torso articulated forward from the waist, its articulated arms steadily extended, its small, palm-down hands resting flat on the polished surface. The movement culminates as its simple, spherical head bends downward, gently resting on the desk. Its entire deliberate movement is perfectly mirrored in the dark, polished wood below, creating a stunning, distorted reflection that doubles the silent, solemn ritual and imbues the scene with a profound atmosphere of quiet contemplation.", + "seed": 18774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_2982_seed_18774_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01490254258387722, + "vlm_logprob": -0.01490254258387722, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.01490254258387722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02017197199165821 + ], + "opposite_logprob_tokens": [ + -4.395172119140625 + ], + "logprob_reward": -0.02017197199165821, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02017197199165821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15272393822669983, + 0.03327195718884468, + 0.07631082087755203, + 0.16363216936588287, + 0.15171372890472412, + 0.06457558274269104, + 0.11543650925159454, + 0.07017756253480911, + 0.10736364126205444, + 0.06479407101869583 + ], + "credit_map": [ + 0.01813262514770031, + 0.008255495689809322, + 0.048565834760665894, + 0.010846092365682125, + 0.009251849725842476, + 0.011925366707146168, + 0.007569588255137205, + 0.03817707672715187, + 0.005738572217524052, + 0.004762963857501745, + 0.00523673091083765, + 0.00295936013571918, + 0.003557639429345727, + 0.003210431430488825, + 0.0028393417596817017, + 0.0049669151194393635, + 0.01120775006711483, + 0.00852971337735653, + 0.011364996433258057, + 0.004269529599696398, + 0.005406384356319904, + 0.013689257204532623, + 0.004721555393189192, + 0.017121639102697372, + 0.01294661033898592, + 0.02756686322391033, + 0.019694745540618896, + 0.009023280814290047, + 0.021448347717523575, + 0.03582369536161423, + 0.011654666624963284, + 0.02547396719455719, + 0.013487362302839756, + 0.03451201692223549, + 0.020879261195659637, + 0.018179934471845627, + 0.01577666960656643, + 0.01157229021191597, + 0.015448741614818573, + 0.021857455372810364, + 0.007538588251918554, + 0.00824253261089325, + 0.010797114111483097, + 0.006127538625150919, + 0.007006728556007147, + 0.006730599328875542, + 0.006687656044960022, + 0.011444821022450924, + 0.015408081002533436, + 0.015582246705889702, + 0.01916429027915001, + 0.011109747923910618, + 0.009923933073878288, + 0.009371999651193619, + 0.011671838350594044, + 0.023204373195767403, + 0.012021913193166256, + 0.006362122017890215, + 0.01243249699473381, + 0.009965188801288605, + 0.00455375574529171, + 0.006595536135137081, + 0.006721431855112314, + 0.01152511965483427, + 0.019504200667142868, + 0.012020397931337357, + 0.014069327153265476, + 0.012460694648325443, + 0.012475810013711452, + 0.011130262166261673, + 0.010039632208645344, + 0.015663323923945427, + 0.009481695480644703, + 0.004721927456557751, + 0.01072667632251978, + 0.010614106431603432, + 0.007182999514043331, + 0.0071310545317828655, + 0.00832371972501278, + 0.0066118910908699036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.02017197199165821 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002207106677815318 + ], + "opposite_logprob_tokens": [ + -6.127207279205322 + ], + "logprob_reward": -0.002207106677815318, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002207106677815318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22241836786270142, + 0.1103738397359848, + 0.10252781212329865, + 0.1631307750940323, + 0.09397532790899277, + 0.10614190995693207, + 0.08921186625957489, + 0.03563656285405159, + 0.04448596015572548, + 0.032097652554512024 + ], + "credit_map": [ + 0.03460764139890671, + 0.012885396368801594, + 0.07141715288162231, + 0.018469421193003654, + 0.014025195501744747, + 0.021452289074659348, + 0.012269573286175728, + 0.037291694432497025, + 0.011451718397438526, + 0.01258822437375784, + 0.026978565379977226, + 0.009681884199380875, + 0.010321356356143951, + 0.008050457574427128, + 0.009254330769181252, + 0.022047284990549088, + 0.00787044595927, + 0.0063575441017746925, + 0.024504229426383972, + 0.004199041519314051, + 0.005468824412673712, + 0.012498334050178528, + 0.004588395357131958, + 0.037040986120700836, + 0.013314335606992245, + 0.025622785091400146, + 0.016391200944781303, + 0.007890427485108376, + 0.016659848392009735, + 0.04011363163590431, + 0.009165024384856224, + 0.033973511308431625, + 0.010947608388960361, + 0.017112359404563904, + 0.013325482606887817, + 0.00973113439977169, + 0.009028354659676552, + 0.0068212952464818954, + 0.009772847406566143, + 0.01723623275756836, + 0.011635667644441128, + 0.01027773879468441, + 0.018313409760594368, + 0.010600277222692966, + 0.012585523538291454, + 0.012183036655187607, + 0.011523852124810219, + 0.019022392109036446, + 0.01089723501354456, + 0.012278092093765736, + 0.015304879285395145, + 0.009296063333749771, + 0.007542078848928213, + 0.007237270474433899, + 0.008178788237273693, + 0.01847744733095169, + 0.0052727446891367435, + 0.005047147627919912, + 0.005790469702333212, + 0.00500058988109231, + 0.003036064561456442, + 0.0031508735846728086, + 0.0032606611493974924, + 0.005078009329736233, + 0.009942954406142235, + 0.005533046554774046, + 0.006615892518311739, + 0.0051216366700828075, + 0.004540550522506237, + 0.003898088587448001, + 0.003512472612783313, + 0.005321311764419079, + 0.0038131382316350937, + 0.003011730732396245, + 0.005152455996721983, + 0.005644295830279589, + 0.0033780145458877087, + 0.0030869527254253626, + 0.004403108265250921, + 0.0036079534329473972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002207106677815318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043054379522800446 + ], + "opposite_logprob_tokens": [ + -3.1680543422698975 + ], + "logprob_reward": -0.043054379522800446, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043054379522800446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12046843767166138, + 0.049202222377061844, + 0.12242481112480164, + 0.18937791883945465, + 0.049218375235795975, + 0.07958950102329254, + 0.1756371408700943, + 0.07169688493013382, + 0.03985881805419922, + 0.10252593457698822 + ], + "credit_map": [ + 0.017179392278194427, + 0.008872796781361103, + 0.030443847179412842, + 0.010265432298183441, + 0.011438754387199879, + 0.014039912261068821, + 0.009557466022670269, + 0.018670838326215744, + 0.00920116063207388, + 0.00544495927169919, + 0.008414419367909431, + 0.005708584561944008, + 0.005150730721652508, + 0.004777407739311457, + 0.004359253216534853, + 0.006145704071968794, + 0.018615249544382095, + 0.019592009484767914, + 0.019343357533216476, + 0.008389400318264961, + 0.012101167812943459, + 0.016900066286325455, + 0.010416614823043346, + 0.017066949978470802, + 0.017080260440707207, + 0.03275012969970703, + 0.022195346653461456, + 0.009971171617507935, + 0.020565632730722427, + 0.04137779399752617, + 0.013536377809941769, + 0.031901199370622635, + 0.006541660986840725, + 0.006720354314893484, + 0.007526863366365433, + 0.0071914177387952805, + 0.0040565370582044125, + 0.005121895112097263, + 0.005261124577373266, + 0.006798522546887398, + 0.01149373222142458, + 0.007307006977498531, + 0.01736840419471264, + 0.00800317246466875, + 0.005285897292196751, + 0.008444557897746563, + 0.005776531994342804, + 0.015910200774669647, + 0.027976522222161293, + 0.023307375609874725, + 0.02937418408691883, + 0.01786341890692711, + 0.018050506711006165, + 0.01400347426533699, + 0.017324455082416534, + 0.027737192809581757, + 0.015189036726951599, + 0.007404708303511143, + 0.012775756418704987, + 0.00788355153053999, + 0.004378635436296463, + 0.0075485482811927795, + 0.005633493419736624, + 0.010883151553571224, + 0.011704104952514172, + 0.002715727314352989, + 0.005694893188774586, + 0.003795953467488289, + 0.0022856304422020912, + 0.0035442113876342773, + 0.0037619094364345074, + 0.006356386002153158, + 0.012635722756385803, + 0.006517759524285793, + 0.017351854592561722, + 0.019006313756108284, + 0.010812195017933846, + 0.011896537616848946, + 0.010452846996486187, + 0.013852696865797043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.043054379522800446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638158440589905 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638158440589905, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638158440589905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2334006130695343, + 0.06867445260286331, + 0.1832878738641739, + 0.10856562107801437, + 0.0885833129286766, + 0.04746626690030098, + 0.07056977599859238, + 0.05942278727889061, + 0.04651300981640816, + 0.09351633489131927 + ], + "credit_map": [ + 0.03033430688083172, + 0.011361857876181602, + 0.06873144209384918, + 0.019864274188876152, + 0.01095718052238226, + 0.023058483377099037, + 0.011185801587998867, + 0.057907260954380035, + 0.011172247119247913, + 0.006083582062274218, + 0.01383083313703537, + 0.006583749782294035, + 0.0046801273711025715, + 0.006629436742514372, + 0.004415011033415794, + 0.015279455110430717, + 0.022650793194770813, + 0.020682720467448235, + 0.03700292855501175, + 0.009321620687842369, + 0.013081726618111134, + 0.020751945674419403, + 0.011824502609670162, + 0.04797162488102913, + 0.013219100423157215, + 0.01748683676123619, + 0.014655997976660728, + 0.010868488810956478, + 0.01224955078214407, + 0.014766139909625053, + 0.01300481241196394, + 0.012314680963754654, + 0.010905817151069641, + 0.013524217531085014, + 0.011299039237201214, + 0.012119108811020851, + 0.011449997313320637, + 0.007426193915307522, + 0.009832028299570084, + 0.012026896700263023, + 0.004537302069365978, + 0.0043042623437941074, + 0.008147465996444225, + 0.005468870513141155, + 0.007290765643119812, + 0.004885236732661724, + 0.005121916998177767, + 0.007710443809628487, + 0.017441196367144585, + 0.00450573954731226, + 0.012827536091208458, + 0.005388111341744661, + 0.007660422474145889, + 0.006108672358095646, + 0.007743484806269407, + 0.008894608356058598, + 0.0063768657855689526, + 0.005957257002592087, + 0.010104479268193245, + 0.00832733977586031, + 0.005848416592925787, + 0.008471596986055374, + 0.005602427292615175, + 0.008734401315450668, + 0.009129611775279045, + 0.005135263316333294, + 0.006159601733088493, + 0.006824323441833258, + 0.003793039359152317, + 0.004073516000062227, + 0.005559694953262806, + 0.0058379569090902805, + 0.011082599870860577, + 0.008815379813313484, + 0.015341361984610558, + 0.017652133479714394, + 0.008964486420154572, + 0.007816154509782791, + 0.013059069402515888, + 0.010785134509205818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008638158440589905 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00044109628652222455 + ], + "opposite_logprob_tokens": [ + -7.750441074371338 + ], + "logprob_reward": -0.00044109628652222455, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00044109628652222455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15091827511787415, + 0.1541425734758377, + 0.2505340874195099, + 0.05764824151992798, + 0.07792122662067413, + 0.060577500611543655, + 0.043815698474645615, + 0.0813385471701622, + 0.05460396036505699, + 0.06849981844425201 + ], + "credit_map": [ + 0.01948021911084652, + 0.010118653066456318, + 0.048489976674318314, + 0.009847362525761127, + 0.007831267081201077, + 0.01248602382838726, + 0.0068041239865124226, + 0.03586067631840706, + 0.01357563678175211, + 0.024253621697425842, + 0.02962479367852211, + 0.014242425560951233, + 0.018595745787024498, + 0.013966133818030357, + 0.01484215259552002, + 0.025042081251740456, + 0.015129015780985355, + 0.020524799823760986, + 0.05297143757343292, + 0.019300656393170357, + 0.011026058346033096, + 0.03346343711018562, + 0.016115915030241013, + 0.08200281113386154, + 0.006807715632021427, + 0.009475260972976685, + 0.008855116553604603, + 0.004983246326446533, + 0.008560199290513992, + 0.006935946177691221, + 0.0050883302465081215, + 0.00694243423640728, + 0.0070340330712497234, + 0.017945749685168266, + 0.009868109598755836, + 0.006508242804557085, + 0.012381107546389103, + 0.006612478289753199, + 0.00951990857720375, + 0.008051605895161629, + 0.007323583122342825, + 0.011590336449444294, + 0.008015603758394718, + 0.007576941978186369, + 0.005868217907845974, + 0.006583413574844599, + 0.006069765891879797, + 0.007549643982201815, + 0.0073448969051241875, + 0.002445822348818183, + 0.00820652861148119, + 0.003489079186692834, + 0.002455435460433364, + 0.004194506444036961, + 0.004244979936629534, + 0.011434457264840603, + 0.009330171160399914, + 0.009448935277760029, + 0.013524889014661312, + 0.013314476236701012, + 0.009825651533901691, + 0.008067773655056953, + 0.008047418668866158, + 0.009779246523976326, + 0.00924079678952694, + 0.004312107805162668, + 0.007477786857634783, + 0.009417716413736343, + 0.004197664558887482, + 0.005223041400313377, + 0.005690925754606724, + 0.009043929167091846, + 0.008906187489628792, + 0.004609567578881979, + 0.011751679703593254, + 0.013845033943653107, + 0.006556542590260506, + 0.00684888893738389, + 0.007727585267275572, + 0.008254346437752247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00044109628652222455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle close-up slowly zooms in on a tiny, articulated figure meticulously crafted from polished brass, standing atop the vast, reflective surface of a dark-varnished mahogany desk. A single, warm light source, emanating from the off-camera right, casts long shadows and creates brilliant golden shimmers along the figure's bent back and limbs, highlighting the smooth metallic texture against the dark brown of the wood. With mechanically graceful and reverent slowness, the figure begins to prostrate; its torso articulated forward from the waist, its articulated arms steadily extended, its small, palm-down hands resting flat on the polished surface. The movement culminates as its simple, spherical head bends downward, gently resting on the desk. Its entire deliberate movement is perfectly mirrored in the dark, polished wood below, creating a stunning, distorted reflection that doubles the silent, solemn ritual and imbues the scene with a profound atmosphere of quiet contemplation.", + "seed": 18775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_8/index_2982_seed_18775_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027403177600353955, + "vlm_logprob": -0.027403177600353955, + "vlm_yes_no_margin": 4.624999952316284, + "vlm_token_credit": -0.027403177600353955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03238120675086975 + ], + "opposite_logprob_tokens": [ + -4.032381057739258 + ], + "logprob_reward": -0.03238120675086975, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03238120675086975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1182132139801979, + 0.05721988156437874, + 0.06775712966918945, + 0.047792673110961914, + 0.06278443336486816, + 0.06987787783145905, + 0.3559863567352295, + 0.06555373966693878, + 0.06250970810651779, + 0.09230493009090424 + ], + "credit_map": [ + 0.020320193842053413, + 0.01002964936196804, + 0.01369719859212637, + 0.010869637131690979, + 0.014724362641572952, + 0.014232355169951916, + 0.0153009332716465, + 0.01903887651860714, + 0.008057920262217522, + 0.005037994589656591, + 0.007203062996268272, + 0.005734513513743877, + 0.008039536885917187, + 0.006831781938672066, + 0.007172968704253435, + 0.009142100811004639, + 0.01138643454760313, + 0.005231275223195553, + 0.007740377448499203, + 0.005427803378552198, + 0.004988126922398806, + 0.010231893509626389, + 0.010175389237701893, + 0.012575826607644558, + 0.009458932094275951, + 0.004712898284196854, + 0.00896974466741085, + 0.0038989053573459387, + 0.00411607138812542, + 0.005169795360416174, + 0.004868289455771446, + 0.0065980348736047745, + 0.009793882258236408, + 0.00644785538315773, + 0.010116053745150566, + 0.007008713204413652, + 0.007581172976642847, + 0.005755912978202105, + 0.006079438608139753, + 0.010001408867537975, + 0.010345619171857834, + 0.006423593033105135, + 0.009468493051826954, + 0.0096355015411973, + 0.005087064579129219, + 0.0063646347261965275, + 0.010198172181844711, + 0.012354797683656216, + 0.04483741149306297, + 0.019015435129404068, + 0.059557367116212845, + 0.01885213889181614, + 0.02204531989991665, + 0.024723399430513382, + 0.045995473861694336, + 0.1209598109126091, + 0.0069746351800858974, + 0.004473768640309572, + 0.01169816218316555, + 0.005454239435493946, + 0.007911340333521366, + 0.005382316652685404, + 0.012394740246236324, + 0.011264542117714882, + 0.006502562668174505, + 0.003619441529735923, + 0.010582230053842068, + 0.005907214712351561, + 0.005719780921936035, + 0.004671226721256971, + 0.011853637173771858, + 0.013653616420924664, + 0.006250323262065649, + 0.006759257055819035, + 0.011501085013151169, + 0.009998073801398277, + 0.006912937853485346, + 0.009634953923523426, + 0.021008523181080818, + 0.020239779725670815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.03238120675086975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046508475206792355 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.0046508475206792355, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046508475206792355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19103281199932098, + 0.10667584836483002, + 0.0810265764594078, + 0.03538862615823746, + 0.026851799339056015, + 0.08227373659610748, + 0.2760578393936157, + 0.04554106667637825, + 0.0394621342420578, + 0.11568955332040787 + ], + "credit_map": [ + 0.024830326437950134, + 0.01368209533393383, + 0.022801928222179413, + 0.01682390458881855, + 0.032832078635692596, + 0.020877836272120476, + 0.03209135681390762, + 0.02709328755736351, + 0.020110495388507843, + 0.00848570466041565, + 0.01641038805246353, + 0.009814063087105751, + 0.011464863084256649, + 0.010775657370686531, + 0.014067823067307472, + 0.01554686389863491, + 0.016504589468240738, + 0.006885690148919821, + 0.010993584990501404, + 0.007978195324540138, + 0.008344277739524841, + 0.007887161336839199, + 0.009752810932695866, + 0.012680265121161938, + 0.005906563252210617, + 0.00228508235886693, + 0.0047769444063305855, + 0.0029788969550281763, + 0.004540529102087021, + 0.0038999447133392096, + 0.004176074638962746, + 0.006824590731412172, + 0.00550623144954443, + 0.001910894294269383, + 0.005060662515461445, + 0.0025708507746458054, + 0.002391734393313527, + 0.002752948785200715, + 0.002723704557865858, + 0.003934774547815323, + 0.012995559722185135, + 0.00833880715072155, + 0.011944759637117386, + 0.010137662291526794, + 0.008996570482850075, + 0.006945514120161533, + 0.010340264067053795, + 0.01257460005581379, + 0.03679131343960762, + 0.015398186631500721, + 0.04625667259097099, + 0.014055358245968819, + 0.016779959201812744, + 0.019009102135896683, + 0.03524570167064667, + 0.0925215631723404, + 0.005503302440047264, + 0.003575065406039357, + 0.0075563304126262665, + 0.003935220185667276, + 0.006854469422250986, + 0.0034792430233210325, + 0.007383862044662237, + 0.007253573276102543, + 0.004272453486919403, + 0.002734291600063443, + 0.006337110418826342, + 0.0035347994416952133, + 0.003962046932429075, + 0.003419621381908655, + 0.008299683220684528, + 0.006902128923684359, + 0.010558515787124634, + 0.005506632849574089, + 0.018881414085626602, + 0.010521846823394299, + 0.00952096562832594, + 0.007441410794854164, + 0.02881922386586666, + 0.024439549073576927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0046508475206792355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901950716972351 + ], + "opposite_logprob_tokens": [ + -2.464019536972046 + ], + "logprob_reward": -0.08901950716972351, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901950716972351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09264302253723145, + 0.08462300151586533, + 0.05306493863463402, + 0.09331532567739487, + 0.037574417889118195, + 0.06933797895908356, + 0.2234342396259308, + 0.19031530618667603, + 0.12005458772182465, + 0.03563716635107994 + ], + "credit_map": [ + 0.013104412704706192, + 0.00708454754203558, + 0.014087901450693607, + 0.014003781601786613, + 0.007634327281266451, + 0.010877321474254131, + 0.012987912632524967, + 0.012862821109592915, + 0.014648284763097763, + 0.006613848265260458, + 0.010277917608618736, + 0.009572307579219341, + 0.010829324834048748, + 0.010420379228889942, + 0.009847644716501236, + 0.012413307093083858, + 0.010200652293860912, + 0.003944169729948044, + 0.006576151587069035, + 0.0069845253601670265, + 0.004327402915805578, + 0.00557054253295064, + 0.006634692661464214, + 0.008826804347336292, + 0.017295587807893753, + 0.006787035148590803, + 0.012706134468317032, + 0.009232327342033386, + 0.009141892194747925, + 0.010719901882112026, + 0.009331075474619865, + 0.018101388588547707, + 0.006085187196731567, + 0.0033099621068686247, + 0.005290737841278315, + 0.00455115083605051, + 0.00433620810508728, + 0.003434203565120697, + 0.00503780459985137, + 0.005529169458895922, + 0.01046352181583643, + 0.007044604513794184, + 0.009221785701811314, + 0.008280404843389988, + 0.007019221782684326, + 0.007386087439954281, + 0.008713453076779842, + 0.011208905838429928, + 0.027528954669833183, + 0.02104751020669937, + 0.03677577152848244, + 0.018964292481541634, + 0.023199040442705154, + 0.01814749650657177, + 0.024193257093429565, + 0.05357793718576431, + 0.019642705097794533, + 0.014657355844974518, + 0.03335343301296234, + 0.018277129158377647, + 0.019112158566713333, + 0.013518823310732841, + 0.03848057612776756, + 0.03327313810586929, + 0.011490990407764912, + 0.007021548692137003, + 0.02059515379369259, + 0.010665706358850002, + 0.008961614221334457, + 0.00945967435836792, + 0.030377322807908058, + 0.02148258313536644, + 0.003850271226838231, + 0.0024616457521915436, + 0.005346871446818113, + 0.0034267886076122522, + 0.0029302777256816626, + 0.00347382971085608, + 0.006600674707442522, + 0.007546810433268547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.08901950716972351 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782588109374046 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782588109374046, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782588109374046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1159665584564209, + 0.1214025542140007, + 0.023676393553614616, + 0.02060222439467907, + 0.029818089678883553, + 0.04385789856314659, + 0.3275512754917145, + 0.11906057596206665, + 0.1206371933221817, + 0.0774272084236145 + ], + "credit_map": [ + 0.01529712975025177, + 0.009302855469286442, + 0.014873813837766647, + 0.016690803691744804, + 0.015760889276862144, + 0.011175473220646381, + 0.017259817570447922, + 0.015605783089995384, + 0.016344377771019936, + 0.008750612847507, + 0.014047622680664062, + 0.011998143047094345, + 0.025128209963440895, + 0.012333786115050316, + 0.01782185770571232, + 0.0149779561907053, + 0.003509160131216049, + 0.0020594201050698757, + 0.00264924974180758, + 0.0022288025356829166, + 0.0024965223856270313, + 0.0042129517532885075, + 0.0023277776781469584, + 0.0041925134137272835, + 0.004614114761352539, + 0.0014800771605223417, + 0.0028751734644174576, + 0.0024596666917204857, + 0.0019353559473529458, + 0.001639311434701085, + 0.0024602217599749565, + 0.0031383067835122347, + 0.005163918249309063, + 0.003052466781809926, + 0.004857304506003857, + 0.003140775253996253, + 0.003138614585623145, + 0.002671098103746772, + 0.003120936220511794, + 0.004672977142035961, + 0.0060517857782542706, + 0.004703131970018148, + 0.005613513756543398, + 0.005164068192243576, + 0.004416046664118767, + 0.005483599845319986, + 0.006113841198384762, + 0.00631191348657012, + 0.04404650256037712, + 0.01804903894662857, + 0.054510947316884995, + 0.01671876758337021, + 0.021950606256723404, + 0.023639213293790817, + 0.041872456669807434, + 0.10676376521587372, + 0.012310801073908806, + 0.007811606861650944, + 0.020523028448224068, + 0.009304055944085121, + 0.018846699967980385, + 0.009272808209061623, + 0.021233338862657547, + 0.01975824683904648, + 0.01197259221225977, + 0.006125140469521284, + 0.019978318363428116, + 0.010536161251366138, + 0.01034380029886961, + 0.009809853509068489, + 0.028672685846686363, + 0.023198647424578667, + 0.005585326813161373, + 0.004122873768210411, + 0.010852023027837276, + 0.007573670707643032, + 0.006218378897756338, + 0.007085143588483334, + 0.018739990890026093, + 0.017249807715415955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.009782588109374046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny, articulated figure crafted from polished brass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure standing on a dark-varnished mahogany desk with a reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform a prostration motion where the torso, arms, and head move sequentially to rest on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible reflection of the figure's movement in the polished desk surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011817384511232376 + ], + "opposite_logprob_tokens": [ + -6.751181602478027 + ], + "logprob_reward": -0.0011817384511232376, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011817384511232376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2391694188117981, + 0.16648732125759125, + 0.02407802827656269, + 0.03327068313956261, + 0.030421938747167587, + 0.06460576504468918, + 0.19054444134235382, + 0.05718952789902687, + 0.08356093615293503, + 0.11067187041044235 + ], + "credit_map": [ + 0.03556245192885399, + 0.019688783213496208, + 0.033873431384563446, + 0.029342124238610268, + 0.029560791328549385, + 0.025036878883838654, + 0.037084393203258514, + 0.029020581394433975, + 0.030713733285665512, + 0.013002271763980389, + 0.0235435850918293, + 0.01587853766977787, + 0.016616951674222946, + 0.022430971264839172, + 0.019505377858877182, + 0.024795901030302048, + 0.004670513793826103, + 0.002503383671864867, + 0.0026985155418515205, + 0.0022713730577379465, + 0.0018767737783491611, + 0.0023916992358863354, + 0.0033912721555680037, + 0.0042744968086481094, + 0.0060950350016355515, + 0.002774448599666357, + 0.005562177859246731, + 0.0027220584452152252, + 0.0025538315530866385, + 0.003335273824632168, + 0.004047134891152382, + 0.0061807213351130486, + 0.005140607710927725, + 0.002636752324178815, + 0.0052383714355528355, + 0.002977177733555436, + 0.003308413317427039, + 0.0027583129703998566, + 0.0035296031273901463, + 0.004832699429243803, + 0.01150483451783657, + 0.00652717798948288, + 0.008001349866390228, + 0.008660619147121906, + 0.006514687556773424, + 0.005939876195043325, + 0.007854575291275978, + 0.009602641686797142, + 0.0237606018781662, + 0.021299319341778755, + 0.030798031017184258, + 0.017514731734991074, + 0.020865635946393013, + 0.01889139972627163, + 0.0219408068805933, + 0.03547392785549164, + 0.0059378501027822495, + 0.004856270272284746, + 0.008920819498598576, + 0.0051816413179039955, + 0.009061944670975208, + 0.003745561232790351, + 0.009436365216970444, + 0.010049078613519669, + 0.006983059924095869, + 0.004060869570821524, + 0.012117363512516022, + 0.00895498413592577, + 0.007780586369335651, + 0.008301378227770329, + 0.01976906880736351, + 0.015593630261719227, + 0.00852670893073082, + 0.007823651656508446, + 0.014991858042776585, + 0.01185284461826086, + 0.009236910380423069, + 0.008010190911591053, + 0.02694113552570343, + 0.023288574069738388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0011817384511232376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_tensor([6247, 6247])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_tensor([6247, 6247])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..22ca829f92c21c9d0cb67b2a608531e9d84c4b41 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_tensor([6247, 6247])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-altitude drone shot slowly descends and advances, passing through layers of acid-soaked metal overpasses and a maze of pipes, finally focusing on the floor of an urban canyon. There, a scholar moves silently, his face thin and grave, his signature long, graying beard swaying slightly in the humid air. His silhouette bears a striking resemblance to the late Qing Dynasty thinker Kang Youwei. He wears a thick, dark gray traditional cotton and linen robe, its hems darkened by dirt from the floor, a stark contrast to the cold, shimmering metal surrounding him. Giant holographic billboards flicker and morph silently on the skyscraper facades, casting dynamic splatters of magenta and cyan light that reflect in his wary eyes. He hunched over the shadows of the wall, stealing forward with almost silent, cautious steps, each steady and deliberate. He paused periodically, glancing up at the searchlights of patrolling drones overhead before decisively merging into the deeper darkness. The entire scene was shrouded in a sense of high-tech oppression and the tension of solitary risk. Amidst the cold tones of the canvas, the vibrant neon lights and the quaint silhouettes of the figures combined to create a cyberpunk canvas brimming with dramatic tension.", + "seed": 23113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_6247_seed_23113_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15847868341952562, + "vlm_logprob": -0.15847868341952562, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.15847868341952562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43848466873168945 + ], + "opposite_logprob_tokens": [ + -1.0634846687316895 + ], + "logprob_reward": -0.43848466873168945, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43848466873168945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16316594183444977, + 0.1482195109128952, + 0.07748693972826004, + 0.13604284822940826, + 0.039511967450380325, + 0.046125225722789764, + 0.04493646323680878, + 0.11120785027742386, + 0.20393997430801392, + 0.029363231733441353 + ], + "credit_map": [ + 0.018688805401325226, + 0.01918511465191841, + 0.013018805533647537, + 0.027386069297790527, + 0.020982930436730385, + 0.017241548746824265, + 0.03187216445803642, + 0.01479050051420927, + 0.016039352864027023, + 0.030671898275613785, + 0.018447430804371834, + 0.01357227936387062, + 0.025672519579529762, + 0.01103205792605877, + 0.019973214715719223, + 0.012810766696929932, + 0.008177671581506729, + 0.01433116476982832, + 0.012889605946838856, + 0.010607968084514141, + 0.006340137217193842, + 0.009820832870900631, + 0.00591639569029212, + 0.009403160773217678, + 0.011317034251987934, + 0.012460581958293915, + 0.01304591540247202, + 0.013420503586530685, + 0.014574194326996803, + 0.030203184112906456, + 0.019710492342710495, + 0.021310938522219658, + 0.00329353054985404, + 0.0031870491802692413, + 0.0056709591299295425, + 0.005713431164622307, + 0.004629178438335657, + 0.0047126528806984425, + 0.006477212067693472, + 0.00582795450463891, + 0.0036747774574905634, + 0.00265013100579381, + 0.00642809271812439, + 0.004491409752517939, + 0.003247286658734083, + 0.005716479383409023, + 0.006948606576770544, + 0.01296844333410263, + 0.0028638476505875587, + 0.002938498742878437, + 0.006906249560415745, + 0.00704852445051074, + 0.00559380603954196, + 0.006747361738234758, + 0.006208732258528471, + 0.006629444193094969, + 0.009539788588881493, + 0.009485693648457527, + 0.023464519530534744, + 0.01637105643749237, + 0.018931500613689423, + 0.010846258141100407, + 0.012982279062271118, + 0.009586761705577374, + 0.011718221008777618, + 0.010577535256743431, + 0.022362351417541504, + 0.020851654931902885, + 0.032286595553159714, + 0.043526146560907364, + 0.021296318620443344, + 0.04132116213440895, + 0.0021817334927618504, + 0.001976537285372615, + 0.002583353081718087, + 0.0029337711166590452, + 0.006003593094646931, + 0.005216242279857397, + 0.0039337691850960255, + 0.004534231033176184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.43848466873168945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902336865663528 + ], + "opposite_logprob_tokens": [ + -2.5790233612060547 + ], + "logprob_reward": -0.07902336865663528, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902336865663528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15554766356945038, + 0.06137705221772194, + 0.1055123433470726, + 0.1796979308128357, + 0.11702316999435425, + 0.06024555489420891, + 0.07195784151554108, + 0.12710192799568176, + 0.0924098864197731, + 0.029126634821295738 + ], + "credit_map": [ + 0.02080194279551506, + 0.020370598882436752, + 0.011316079646348953, + 0.02451823465526104, + 0.02024187333881855, + 0.014674208126962185, + 0.028012214228510857, + 0.015612506307661533, + 0.005621334072202444, + 0.01342273224145174, + 0.008340934291481972, + 0.009763599373400211, + 0.006095804739743471, + 0.004274135455489159, + 0.00946466438472271, + 0.004393848590552807, + 0.010735523886978626, + 0.011018822900950909, + 0.026280976831912994, + 0.007022842299193144, + 0.006686363369226456, + 0.011210777796804905, + 0.005855341907590628, + 0.02670169435441494, + 0.016450004652142525, + 0.015673253685235977, + 0.015024702064692974, + 0.018159044906497, + 0.01830463670194149, + 0.04109188914299011, + 0.026191914454102516, + 0.028802499175071716, + 0.011201905086636543, + 0.011183704249560833, + 0.010281792841851711, + 0.007701259106397629, + 0.010152283124625683, + 0.02381245419383049, + 0.01247973833233118, + 0.03021002560853958, + 0.005384117364883423, + 0.006027601193636656, + 0.01161025371402502, + 0.005222564097493887, + 0.006033278536051512, + 0.007849544286727905, + 0.006407115142792463, + 0.01171108242124319, + 0.0052528684027493, + 0.007267529144883156, + 0.00839037075638771, + 0.014709662646055222, + 0.010319292545318604, + 0.007752439472824335, + 0.01062819268554449, + 0.007637489587068558, + 0.010290347971022129, + 0.008155027404427528, + 0.013869540765881538, + 0.037528276443481445, + 0.008899505250155926, + 0.010737520642578602, + 0.018292395398020744, + 0.019329311326146126, + 0.0047699748538434505, + 0.005939089693129063, + 0.01078805886209011, + 0.011007732711732388, + 0.02065117098391056, + 0.014703861437737942, + 0.011313896626234055, + 0.01323609333485365, + 0.003973789047449827, + 0.0018415628001093864, + 0.004484294913709164, + 0.0023006412666291, + 0.003681329544633627, + 0.0038615441881120205, + 0.0037379246205091476, + 0.00524554867297411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.07902336865663528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035257279872894 + ], + "opposite_logprob_tokens": [ + -1.9103525876998901 + ], + "logprob_reward": -0.16035257279872894, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035257279872894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06906286627054214, + 0.06589294224977493, + 0.05617998540401459, + 0.09865705668926239, + 0.07214590907096863, + 0.11285370588302612, + 0.18478019535541534, + 0.09170672297477722, + 0.18159162998199463, + 0.0671289712190628 + ], + "credit_map": [ + 0.005034279078245163, + 0.010995312593877316, + 0.006777717266231775, + 0.009860222227871418, + 0.008718778379261494, + 0.008755064569413662, + 0.011944648809731007, + 0.006976844742894173, + 0.0037200963124632835, + 0.018883053213357925, + 0.008478756062686443, + 0.005293380934745073, + 0.005740465130656958, + 0.013818208128213882, + 0.0051473150961101055, + 0.0048116641119122505, + 0.006481219083070755, + 0.005693501327186823, + 0.013230584561824799, + 0.004352889489382505, + 0.0036765718832612038, + 0.005754935555160046, + 0.003579868469387293, + 0.013410408981144428, + 0.0063168322667479515, + 0.011193333193659782, + 0.008735241368412971, + 0.007891606539487839, + 0.01222313940525055, + 0.02310909703373909, + 0.013190163299441338, + 0.015997642651200294, + 0.004292500205338001, + 0.0045274789445102215, + 0.00909409485757351, + 0.009532270953059196, + 0.009021361358463764, + 0.010606314986944199, + 0.011382585391402245, + 0.013689294457435608, + 0.010241976007819176, + 0.006984263192862272, + 0.013110796920955181, + 0.01389062125235796, + 0.015442132018506527, + 0.013381186872720718, + 0.015402304008603096, + 0.024400420486927032, + 0.011849910020828247, + 0.010491868481040001, + 0.01838003844022751, + 0.03083759918808937, + 0.022873269394040108, + 0.027347011491656303, + 0.023402651771903038, + 0.03959782421588898, + 0.007234715390950441, + 0.007629389874637127, + 0.014613804407417774, + 0.014282663352787495, + 0.013739220798015594, + 0.011424675583839417, + 0.013230111449956894, + 0.009552137926220894, + 0.006990734022110701, + 0.007200014777481556, + 0.01660998724400997, + 0.016311004757881165, + 0.0163098257035017, + 0.05327126383781433, + 0.015471501275897026, + 0.049427296966314316, + 0.005737551022320986, + 0.0053420290350914, + 0.006536452099680901, + 0.007369018625468016, + 0.00740802101790905, + 0.011873522773385048, + 0.009485720656812191, + 0.013376658782362938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.16035257279872894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207170344889164 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207170344889164, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207170344889164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12681271135807037, + 0.08545936644077301, + 0.10502419620752335, + 0.22168947756290436, + 0.19234289228916168, + 0.04282226786017418, + 0.05482424050569534, + 0.06255371868610382, + 0.0499686524271965, + 0.05850246548652649 + ], + "credit_map": [ + 0.00963150430470705, + 0.0144853126257658, + 0.00982637144625187, + 0.019654108211398125, + 0.014827561564743519, + 0.013992620632052422, + 0.029884202405810356, + 0.014511045999825, + 0.0070594982244074345, + 0.013221192173659801, + 0.009544609114527702, + 0.008396157994866371, + 0.01108709629625082, + 0.020450815558433533, + 0.007901842705905437, + 0.007798154838383198, + 0.010722965002059937, + 0.011566532775759697, + 0.018269607797265053, + 0.014523901045322418, + 0.008080988191068172, + 0.010870225727558136, + 0.01164812408387661, + 0.019341858103871346, + 0.022046251222491264, + 0.02159200794994831, + 0.024578813463449478, + 0.019954683259129524, + 0.03026421368122101, + 0.04194123297929764, + 0.030941158533096313, + 0.03037111647427082, + 0.017937960103154182, + 0.017461605370044708, + 0.020937321707606316, + 0.020209219306707382, + 0.017844200134277344, + 0.035282883793115616, + 0.022023718804121017, + 0.040645964443683624, + 0.0031878878362476826, + 0.004979017190635204, + 0.009265213273465633, + 0.006903015077114105, + 0.0033222134225070477, + 0.004486910067498684, + 0.004839492496103048, + 0.005838519893586636, + 0.004148552659898996, + 0.0060836998745799065, + 0.006402894854545593, + 0.009937035851180553, + 0.005261735990643501, + 0.00877609197050333, + 0.0073944032192230225, + 0.0068198260851204395, + 0.008685196749866009, + 0.0051827747374773026, + 0.008142157457768917, + 0.01042196899652481, + 0.007791677489876747, + 0.006651353091001511, + 0.008431833237409592, + 0.007246760185807943, + 0.003617749083787203, + 0.003531175898388028, + 0.007547702174633741, + 0.009235245175659657, + 0.004713638219982386, + 0.007490957621484995, + 0.005535679869353771, + 0.008296503685414791, + 0.006374594289809465, + 0.004073768854141235, + 0.007700241170823574, + 0.006618433631956577, + 0.00737714534625411, + 0.009690137580037117, + 0.006834155879914761, + 0.009833992458879948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.014207170344889164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032563656568527 + ], + "opposite_logprob_tokens": [ + -2.350325584411621 + ], + "logprob_reward": -0.10032563656568527, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032563656568527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1785108894109726, + 0.06742656975984573, + 0.04517674073576927, + 0.10401773452758789, + 0.10527060180902481, + 0.07212913781404495, + 0.0962982252240181, + 0.1374513953924179, + 0.1122584268450737, + 0.08146031200885773 + ], + "credit_map": [ + 0.01805354654788971, + 0.018424198031425476, + 0.014534113928675652, + 0.026496069505810738, + 0.021132852882146835, + 0.01815480925142765, + 0.04116305708885193, + 0.02055223472416401, + 0.007221661042422056, + 0.015017767436802387, + 0.010697273537516594, + 0.006863330025225878, + 0.008541780523955822, + 0.007651563733816147, + 0.006966739427298307, + 0.004466460086405277, + 0.007083606440573931, + 0.007527165114879608, + 0.004284258466213942, + 0.005207134410738945, + 0.004491977393627167, + 0.005688855890184641, + 0.005309505853801966, + 0.005584235768765211, + 0.008226806297898293, + 0.009480884298682213, + 0.011591028422117233, + 0.010017159394919872, + 0.014113967306911945, + 0.01876799389719963, + 0.013924104161560535, + 0.017895789816975594, + 0.007171778008341789, + 0.006122849881649017, + 0.008383585140109062, + 0.011029794812202454, + 0.009521503932774067, + 0.02296518161892891, + 0.011793502606451511, + 0.02828240767121315, + 0.00565585820004344, + 0.004779417999088764, + 0.007528690155595541, + 0.004408060573041439, + 0.004653507377952337, + 0.009982007555663586, + 0.01133416872471571, + 0.023787427693605423, + 0.007866411469876766, + 0.009856540709733963, + 0.00963746476918459, + 0.0195479616522789, + 0.00992128811776638, + 0.011104797944426537, + 0.014627473428845406, + 0.013736287131905556, + 0.009968721307814121, + 0.008607909083366394, + 0.013808446004986763, + 0.0321904793381691, + 0.012914099730551243, + 0.01883893460035324, + 0.017802316695451736, + 0.023320483043789864, + 0.004723475780338049, + 0.004131542518734932, + 0.010718682780861855, + 0.012617048807442188, + 0.010883684270083904, + 0.03015323542058468, + 0.010216054506599903, + 0.02881469577550888, + 0.008246508426964283, + 0.005492744036018848, + 0.011444308795034885, + 0.00891159474849701, + 0.008172939531505108, + 0.015692658722400665, + 0.008949263952672482, + 0.014550291933119297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.10032563656568527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude drone shot slowly descends and advances, passing through layers of acid-soaked metal overpasses and a maze of pipes, finally focusing on the floor of an urban canyon. There, a scholar moves silently, his face thin and grave, his signature long, graying beard swaying slightly in the humid air. His silhouette bears a striking resemblance to the late Qing Dynasty thinker Kang Youwei. He wears a thick, dark gray traditional cotton and linen robe, its hems darkened by dirt from the floor, a stark contrast to the cold, shimmering metal surrounding him. Giant holographic billboards flicker and morph silently on the skyscraper facades, casting dynamic splatters of magenta and cyan light that reflect in his wary eyes. He hunched over the shadows of the wall, stealing forward with almost silent, cautious steps, each steady and deliberate. He paused periodically, glancing up at the searchlights of patrolling drones overhead before decisively merging into the deeper darkness. The entire scene was shrouded in a sense of high-tech oppression and the tension of solitary risk. Amidst the cold tones of the canvas, the vibrant neon lights and the quaint silhouettes of the figures combined to create a cyberpunk canvas brimming with dramatic tension.", + "seed": 23114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_6247_seed_23114_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12699960954487324, + "vlm_logprob": -0.12699960954487324, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.12699960954487324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2600020170211792 + ], + "opposite_logprob_tokens": [ + -1.5100020170211792 + ], + "logprob_reward": -0.2600020170211792, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2600020170211792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19026461243629456, + 0.20337802171707153, + 0.12395012378692627, + 0.09676701575517654, + 0.06509950011968613, + 0.05054010450839996, + 0.03678318113088608, + 0.11187681555747986, + 0.06334606558084488, + 0.0579945407807827 + ], + "credit_map": [ + 0.01576174609363079, + 0.02966981567442417, + 0.023732515051960945, + 0.02352728135883808, + 0.032525185495615005, + 0.014690595678985119, + 0.03722161054611206, + 0.01313585415482521, + 0.027167964726686478, + 0.030538201332092285, + 0.021491801366209984, + 0.02745375782251358, + 0.023377643898129463, + 0.02865750528872013, + 0.023612217977643013, + 0.021078936755657196, + 0.021100345999002457, + 0.01409888919442892, + 0.017210762947797775, + 0.01267068088054657, + 0.014650089666247368, + 0.012397686019539833, + 0.013858543708920479, + 0.017963120713829994, + 0.013297961093485355, + 0.004446862731128931, + 0.011452988721430302, + 0.01788683980703354, + 0.00582842854782939, + 0.004613450262695551, + 0.02806730754673481, + 0.011173175647854805, + 0.005822845734655857, + 0.004349652677774429, + 0.00882621668279171, + 0.01288760919123888, + 0.007183765526860952, + 0.007263191509991884, + 0.011868554167449474, + 0.006897662300616503, + 0.003443717258051038, + 0.0048410966992378235, + 0.007992230355739594, + 0.007842573337256908, + 0.010440880432724953, + 0.00419329758733511, + 0.008570089936256409, + 0.003216217504814267, + 0.006303658243268728, + 0.004301410634070635, + 0.003630537772551179, + 0.005482939537614584, + 0.005650524515658617, + 0.0029087653383612633, + 0.005345095414668322, + 0.003160251537337899, + 0.010256334207952023, + 0.010480175726115704, + 0.01020922139286995, + 0.013261520303785801, + 0.01780429482460022, + 0.014088730327785015, + 0.010551449842751026, + 0.025225084275007248, + 0.004409754648804665, + 0.012045356445014477, + 0.007144669070839882, + 0.00898683164268732, + 0.011046746745705605, + 0.00556759350001812, + 0.008011220954358578, + 0.006133893504738808, + 0.0027801499236375093, + 0.005512085277587175, + 0.0076944343745708466, + 0.010200178250670433, + 0.010614296421408653, + 0.007337022572755814, + 0.008519690483808517, + 0.005336683243513107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.2600020170211792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907108753919601 + ], + "opposite_logprob_tokens": [ + -2.464071035385132 + ], + "logprob_reward": -0.08907108753919601, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907108753919601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1698741912841797, + 0.11757437139749527, + 0.12525680661201477, + 0.12317177653312683, + 0.11106009036302567, + 0.04543308541178703, + 0.07118995487689972, + 0.049349695444107056, + 0.14835980534553528, + 0.03873017057776451 + ], + "credit_map": [ + 0.013220001012086868, + 0.024269018322229385, + 0.01758052594959736, + 0.02581842802464962, + 0.026165319606661797, + 0.01089826226234436, + 0.04096910357475281, + 0.010953537188470364, + 0.01381621789187193, + 0.010007207281887531, + 0.010636378079652786, + 0.014748324640095234, + 0.01298245694488287, + 0.016134873032569885, + 0.0178952869027853, + 0.02135363221168518, + 0.015863733366131783, + 0.016754750162363052, + 0.019549978896975517, + 0.01289731077849865, + 0.019471604377031326, + 0.012206452898681164, + 0.013925958424806595, + 0.014587010256946087, + 0.010409533977508545, + 0.008797633461654186, + 0.015590477734804153, + 0.020033255219459534, + 0.01230127178132534, + 0.005424549803137779, + 0.03806094080209732, + 0.012554117478430271, + 0.008068571798503399, + 0.008241254836320877, + 0.017766643315553665, + 0.0209429282695055, + 0.008883701637387276, + 0.014868629164993763, + 0.01929241232573986, + 0.012995950877666473, + 0.004395363852381706, + 0.0043085734359920025, + 0.006712376140058041, + 0.0065292962826788425, + 0.008876250125467777, + 0.0040647611021995544, + 0.00649650115519762, + 0.00404996145516634, + 0.009471205063164234, + 0.0054901703260838985, + 0.013392594642937183, + 0.00906286295503378, + 0.00802372582256794, + 0.006697287317365408, + 0.01109791174530983, + 0.007954202592372894, + 0.006725997664034367, + 0.006428466644138098, + 0.006785281002521515, + 0.00732785277068615, + 0.005185642279684544, + 0.004096299409866333, + 0.008195734582841396, + 0.004604418762028217, + 0.0070780678652226925, + 0.032715920358896255, + 0.015304519794881344, + 0.01537331473082304, + 0.01630445197224617, + 0.013718019239604473, + 0.03052528016269207, + 0.01734022982418537, + 0.0021845963783562183, + 0.00498233875259757, + 0.00559989782050252, + 0.005847157444804907, + 0.005038249772042036, + 0.003942633513361216, + 0.006162179168313742, + 0.004973115399479866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.08907108753919601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278598129749298 + ], + "opposite_logprob_tokens": [ + -2.0177860260009766 + ], + "logprob_reward": -0.14278598129749298, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278598129749298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08656246215105057, + 0.0331331230700016, + 0.09034202992916107, + 0.17961537837982178, + 0.17226585745811462, + 0.02146776393055916, + 0.0693259984254837, + 0.11648886650800705, + 0.15493878722190857, + 0.07585974782705307 + ], + "credit_map": [ + 0.006544414442032576, + 0.011397899128496647, + 0.009328472428023815, + 0.012395945377647877, + 0.013969243504106998, + 0.005882692523300648, + 0.019388819113373756, + 0.007654975634068251, + 0.004226282238960266, + 0.005143995862454176, + 0.004797359462827444, + 0.003977666143327951, + 0.0038242535665631294, + 0.003934978507459164, + 0.003664962947368622, + 0.0035636257380247116, + 0.015705158933997154, + 0.009728985838592052, + 0.014372585341334343, + 0.011428313329815865, + 0.012108503840863705, + 0.008115922100841999, + 0.008867799304425716, + 0.010014762170612812, + 0.021523363888263702, + 0.00824282318353653, + 0.023339835926890373, + 0.032516371458768845, + 0.011384682729840279, + 0.00743907643482089, + 0.061792220920324326, + 0.013377005234360695, + 0.011992988176643848, + 0.013292166404426098, + 0.027998432517051697, + 0.0407298319041729, + 0.010175837203860283, + 0.014053610153496265, + 0.03685971349477768, + 0.017163271084427834, + 0.003781764768064022, + 0.0022504220250993967, + 0.0038318976294249296, + 0.0031175692565739155, + 0.0016872192500159144, + 0.0020120488479733467, + 0.0027553688269108534, + 0.0020314729772508144, + 0.0072884694673120975, + 0.007504846900701523, + 0.013234943151473999, + 0.009965571574866772, + 0.008009649813175201, + 0.005480417516082525, + 0.010567545890808105, + 0.007274552248418331, + 0.007989801466464996, + 0.013506735675036907, + 0.012148489244282246, + 0.012294110842049122, + 0.02160797454416752, + 0.013139201328158379, + 0.013212871737778187, + 0.022589676082134247, + 0.008180360309779644, + 0.03268541023135185, + 0.02331359311938286, + 0.01616431027650833, + 0.021047601476311684, + 0.012905252166092396, + 0.017828624695539474, + 0.022813649848103523, + 0.003006023121997714, + 0.009016894735395908, + 0.011199120432138443, + 0.013986397534608841, + 0.008712522685527802, + 0.0071377987042069435, + 0.017260447144508362, + 0.0055405450984835625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.14278598129749298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608009822666645 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.01608009822666645, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608009822666645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22492997348308563, + 0.11380813270807266, + 0.04579777643084526, + 0.1012195274233818, + 0.08867476880550385, + 0.017164643853902817, + 0.11354131996631622, + 0.06628945469856262, + 0.0929923728108406, + 0.1355820596218109 + ], + "credit_map": [ + 0.014763738960027695, + 0.02263202890753746, + 0.02691493183374405, + 0.03442889079451561, + 0.03429504111409187, + 0.018169276416301727, + 0.04068849980831146, + 0.03303758427500725, + 0.018923450261354446, + 0.017423279583454132, + 0.013870933093130589, + 0.015527608804404736, + 0.01501141395419836, + 0.013579687103629112, + 0.011979096569120884, + 0.007492665201425552, + 0.007310612592846155, + 0.006070231087505817, + 0.0066293072886765, + 0.004107265267521143, + 0.004398515913635492, + 0.003489177441224456, + 0.005484502296894789, + 0.008308161981403828, + 0.015446376986801624, + 0.007567836903035641, + 0.012215551920235157, + 0.013835074380040169, + 0.009055964648723602, + 0.0075439405627548695, + 0.023021169006824493, + 0.012533613480627537, + 0.006516594439744949, + 0.005560379009693861, + 0.013265549205243587, + 0.017627369612455368, + 0.005989463999867439, + 0.008344346657395363, + 0.01680736057460308, + 0.014563708566129208, + 0.002073717536404729, + 0.0018469237256795168, + 0.0027819282840937376, + 0.00257171131670475, + 0.0022408058866858482, + 0.0018288271967321634, + 0.002261678921058774, + 0.0015590504044666886, + 0.021234454587101936, + 0.013911119662225246, + 0.014696769416332245, + 0.01522589661180973, + 0.011172765865921974, + 0.013180558569729328, + 0.014365151524543762, + 0.009754601866006851, + 0.004556423053145409, + 0.008581297472119331, + 0.007023116108030081, + 0.007758531253784895, + 0.008263354189693928, + 0.007097169291228056, + 0.008115366101264954, + 0.014894194900989532, + 0.004296360537409782, + 0.011280067265033722, + 0.008394879288971424, + 0.012256293557584286, + 0.01241947989910841, + 0.006395126227289438, + 0.02083328180015087, + 0.017116885632276535, + 0.008735215291380882, + 0.012111817486584187, + 0.013808224350214005, + 0.01957610249519348, + 0.025848792865872383, + 0.018850404769182205, + 0.02247699908912182, + 0.014174506068229675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01608009822666645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705886363983154 + ], + "opposite_logprob_tokens": [ + -2.127058982849121 + ], + "logprob_reward": -0.12705886363983154, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705886363983154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14160366356372833, + 0.07031526416540146, + 0.08289144188165665, + 0.11225384473800659, + 0.08828262239694595, + 0.021155819296836853, + 0.09235668182373047, + 0.09432809799909592, + 0.16488538682460785, + 0.1319272369146347 + ], + "credit_map": [ + 0.01059014443308115, + 0.014528704807162285, + 0.01692105084657669, + 0.01902906969189644, + 0.022384872660040855, + 0.010268841870129108, + 0.03144463896751404, + 0.016436344012618065, + 0.008460449054837227, + 0.009324315004050732, + 0.008760812692344189, + 0.009082498028874397, + 0.007831121794879436, + 0.009501445107161999, + 0.011006275191903114, + 0.006348337512463331, + 0.014849575236439705, + 0.011796717531979084, + 0.01130700670182705, + 0.009634837508201599, + 0.00758853554725647, + 0.006238970439881086, + 0.010360057465732098, + 0.01111573539674282, + 0.014702319167554379, + 0.0067031970247626305, + 0.013812940567731857, + 0.01963796466588974, + 0.010014105588197708, + 0.006786799058318138, + 0.03278956562280655, + 0.0078069535084068775, + 0.005426990799605846, + 0.007591425906866789, + 0.01255330815911293, + 0.01912195049226284, + 0.008519758470356464, + 0.009768044576048851, + 0.016149140894412994, + 0.009152003563940525, + 0.0033698093611747026, + 0.0022522963117808104, + 0.0023810553830116987, + 0.0032272059470415115, + 0.003178956685587764, + 0.0018720177467912436, + 0.002369970316067338, + 0.00250450661405921, + 0.012213379144668579, + 0.012266013771295547, + 0.018883883953094482, + 0.013303564861416817, + 0.008324404247105122, + 0.007669398095458746, + 0.012249111197888851, + 0.007446924690157175, + 0.01022124569863081, + 0.010234972462058067, + 0.009416964836418629, + 0.012197836302220821, + 0.019568469375371933, + 0.008125261403620243, + 0.013922970741987228, + 0.010640377178788185, + 0.008504034020006657, + 0.03893395885825157, + 0.02028016187250614, + 0.01793690398335457, + 0.024687347933650017, + 0.014019649475812912, + 0.02310306392610073, + 0.017420269548892975, + 0.005794005002826452, + 0.01684586890041828, + 0.02079438418149948, + 0.024918867275118828, + 0.019586078822612762, + 0.013209962286055088, + 0.021658750250935555, + 0.009119316004216671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.12705886363983154 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude drone shot slowly descends and advances, passing through layers of acid-soaked metal overpasses and a maze of pipes, finally focusing on the floor of an urban canyon. There, a scholar moves silently, his face thin and grave, his signature long, graying beard swaying slightly in the humid air. His silhouette bears a striking resemblance to the late Qing Dynasty thinker Kang Youwei. He wears a thick, dark gray traditional cotton and linen robe, its hems darkened by dirt from the floor, a stark contrast to the cold, shimmering metal surrounding him. Giant holographic billboards flicker and morph silently on the skyscraper facades, casting dynamic splatters of magenta and cyan light that reflect in his wary eyes. He hunched over the shadows of the wall, stealing forward with almost silent, cautious steps, each steady and deliberate. He paused periodically, glancing up at the searchlights of patrolling drones overhead before decisively merging into the deeper darkness. The entire scene was shrouded in a sense of high-tech oppression and the tension of solitary risk. Amidst the cold tones of the canvas, the vibrant neon lights and the quaint silhouettes of the figures combined to create a cyberpunk canvas brimming with dramatic tension.", + "seed": 23115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_6247_seed_23115_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3026447555050254, + "vlm_logprob": -0.3026447555050254, + "vlm_yes_no_margin": 1.825, + "vlm_token_credit": -0.3026447555050254, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9086757898330688 + ], + "opposite_logprob_tokens": [ + -0.5336757898330688 + ], + "logprob_reward": -0.9086757898330688, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9086757898330688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22811105847358704, + 0.08621474355459213, + 0.08488772809505463, + 0.1803950071334839, + 0.06937242299318314, + 0.0872865840792656, + 0.04909354820847511, + 0.1143074557185173, + 0.06072654575109482, + 0.03960492089390755 + ], + "credit_map": [ + 0.02409937046468258, + 0.03287387266755104, + 0.03243861347436905, + 0.026529191061854362, + 0.02896290458738804, + 0.02849419414997101, + 0.01898263394832611, + 0.03573029115796089, + 0.007447720039635897, + 0.012918789871037006, + 0.013417028822004795, + 0.01867080107331276, + 0.009355434216558933, + 0.008183200843632221, + 0.008041881956160069, + 0.008179886266589165, + 0.006627158261835575, + 0.015787122771143913, + 0.006646411027759314, + 0.007305664476007223, + 0.010936077684164047, + 0.011685175821185112, + 0.015219290740787983, + 0.010680828243494034, + 0.00823448970913887, + 0.009539611637592316, + 0.009460371918976307, + 0.011553925462067127, + 0.00977391004562378, + 0.05673430860042572, + 0.01322764903306961, + 0.061870746314525604, + 0.0064017027616500854, + 0.009776980616152287, + 0.008545435965061188, + 0.007935929112136364, + 0.008908499963581562, + 0.008013628423213959, + 0.00787424948066473, + 0.01191599853336811, + 0.01297750510275364, + 0.007853923365473747, + 0.009039792232215405, + 0.01179646048694849, + 0.010153577663004398, + 0.010547785088419914, + 0.01290248055011034, + 0.01201505959033966, + 0.006290463265031576, + 0.005418816581368446, + 0.00673889834433794, + 0.006139887031167746, + 0.004417405929416418, + 0.00477774441242218, + 0.007131977938115597, + 0.008178356103599072, + 0.012976428493857384, + 0.013711929321289062, + 0.014399741776287556, + 0.010345022194087505, + 0.019398069009184837, + 0.017435934394598007, + 0.013227710500359535, + 0.012812625616788864, + 0.0059646484442055225, + 0.006271016784012318, + 0.010650222189724445, + 0.008046737872064114, + 0.005586021114140749, + 0.008007947355508804, + 0.00799936056137085, + 0.008200585842132568, + 0.0034254849888384342, + 0.005655897315591574, + 0.00688896095380187, + 0.004100124817341566, + 0.00452084606513381, + 0.0050202528946101665, + 0.004618804901838303, + 0.005374548025429249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.9086757898330688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163455605506897 + ], + "opposite_logprob_tokens": [ + -1.7016345262527466 + ], + "logprob_reward": -0.20163455605506897, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163455605506897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14979498088359833, + 0.05737011134624481, + 0.055998723953962326, + 0.08525049686431885, + 0.1391662359237671, + 0.06308525800704956, + 0.07078428566455841, + 0.08114106953144073, + 0.2809318006038666, + 0.016477074474096298 + ], + "credit_map": [ + 0.019048934802412987, + 0.019333763048052788, + 0.028333917260169983, + 0.02163834497332573, + 0.01767786778509617, + 0.013382205739617348, + 0.013258527033030987, + 0.01712142862379551, + 0.003300206968560815, + 0.006563328206539154, + 0.008341609500348568, + 0.010518889874219894, + 0.008675409480929375, + 0.007831832394003868, + 0.00686911353841424, + 0.005269728135317564, + 0.006674834527075291, + 0.01016327366232872, + 0.004365850239992142, + 0.004980068188160658, + 0.006422536913305521, + 0.006759422365576029, + 0.012200990691781044, + 0.004431745503097773, + 0.006617074832320213, + 0.005728263407945633, + 0.006107851397246122, + 0.01199024822562933, + 0.005042130593210459, + 0.02275904081761837, + 0.00856980960816145, + 0.01843607798218727, + 0.01621149480342865, + 0.02015380747616291, + 0.02610914781689644, + 0.019597619771957397, + 0.012205216102302074, + 0.014738977886736393, + 0.01487013790756464, + 0.015279832296073437, + 0.008710996247828007, + 0.009325259365141392, + 0.00789395160973072, + 0.008489934727549553, + 0.007683634292334318, + 0.007322000339627266, + 0.006390347145497799, + 0.007269137538969517, + 0.007415443193167448, + 0.009147637523710728, + 0.00991516187787056, + 0.0069469003938138485, + 0.00522783724591136, + 0.00791302602738142, + 0.00857573188841343, + 0.015642544254660606, + 0.008355339989066124, + 0.013790024444460869, + 0.011613015085458755, + 0.007629387080669403, + 0.009522838518023491, + 0.01172049529850483, + 0.010375000536441803, + 0.00813497044146061, + 0.027877725660800934, + 0.03553479164838791, + 0.05537703260779381, + 0.03728225827217102, + 0.02600950002670288, + 0.0302959643304348, + 0.02924085594713688, + 0.03931368887424469, + 0.0035526519641280174, + 0.002025183057412505, + 0.002431998262181878, + 0.0015040179714560509, + 0.001402615220285952, + 0.0016562287928536534, + 0.0018722880631685257, + 0.002032090211287141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.20163455605506897 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813147008419037 + ], + "opposite_logprob_tokens": [ + -1.406314730644226 + ], + "logprob_reward": -0.2813147008419037, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813147008419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10173976421356201, + 0.11842229217290878, + 0.0960998386144638, + 0.057496994733810425, + 0.08555775880813599, + 0.31097936630249023, + 0.0584445521235466, + 0.0676218718290329, + 0.0541364811360836, + 0.04950110986828804 + ], + "credit_map": [ + 0.008459054864943027, + 0.015215682797133923, + 0.012377424165606499, + 0.016776349395513535, + 0.01437569409608841, + 0.014845866709947586, + 0.010638981126248837, + 0.009050705470144749, + 0.009664434008300304, + 0.01958800107240677, + 0.019823288545012474, + 0.024971894919872284, + 0.015121053904294968, + 0.011247922666370869, + 0.00989651307463646, + 0.008109177462756634, + 0.005233545787632465, + 0.014272579923272133, + 0.004970467183738947, + 0.006012024357914925, + 0.008102848194539547, + 0.021033748984336853, + 0.011247864924371243, + 0.02522675320506096, + 0.004947559908032417, + 0.0032851025462150574, + 0.005264210514724255, + 0.005793385673314333, + 0.0037420992739498615, + 0.015130703337490559, + 0.004943205043673515, + 0.014390721917152405, + 0.009041910991072655, + 0.005997004918754101, + 0.008811783976852894, + 0.019358208402991295, + 0.013354630209505558, + 0.004758193623274565, + 0.015014834702014923, + 0.009221184067428112, + 0.021719638258218765, + 0.01746501587331295, + 0.016908565536141396, + 0.027052802965044975, + 0.06139831990003586, + 0.04927390068769455, + 0.07322561740875244, + 0.04393548518419266, + 0.007556075695902109, + 0.005289261229336262, + 0.00964359287172556, + 0.005495868157595396, + 0.005898807663470507, + 0.006798308342695236, + 0.00759500078856945, + 0.010167633183300495, + 0.005916409660130739, + 0.009513852186501026, + 0.008589770644903183, + 0.007696869317442179, + 0.007401648908853531, + 0.010581186041235924, + 0.010152933187782764, + 0.007769191637635231, + 0.006700016092509031, + 0.0067069134674966335, + 0.009608940221369267, + 0.006394512485712767, + 0.00520307756960392, + 0.005909012630581856, + 0.005951508414000273, + 0.007662491407245398, + 0.004866043105721474, + 0.00577917555347085, + 0.00536895077675581, + 0.004210421349853277, + 0.008061511442065239, + 0.006325226742774248, + 0.007606530096381903, + 0.00728324381634593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.2813147008419037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864903163164854 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.00864903163164854, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864903163164854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40719762444496155, + 0.03697461634874344, + 0.14647641777992249, + 0.13080242276191711, + 0.040107134729623795, + 0.04604879021644592, + 0.06212088093161583, + 0.04533923417329788, + 0.039318349212408066, + 0.045614514499902725 + ], + "credit_map": [ + 0.04732389748096466, + 0.054134294390678406, + 0.06346435844898224, + 0.06959107518196106, + 0.05084514617919922, + 0.03509313613176346, + 0.03672691807150841, + 0.05001877248287201, + 0.0035914129111915827, + 0.007532685063779354, + 0.004230121616274118, + 0.0045661781914532185, + 0.005308662075549364, + 0.004663051106035709, + 0.004584678914397955, + 0.002497827634215355, + 0.014493708498775959, + 0.04197176918387413, + 0.009833390824496746, + 0.012779749929904938, + 0.02060246467590332, + 0.016136256977915764, + 0.018371036276221275, + 0.012288033962249756, + 0.006353460717946291, + 0.006427181418985128, + 0.004314768128097057, + 0.006232231855392456, + 0.008213884197175503, + 0.04559924826025963, + 0.006017274688929319, + 0.04764437675476074, + 0.004663227591663599, + 0.003782696323469281, + 0.003789826761931181, + 0.007438782136887312, + 0.005059618968516588, + 0.0035219446290284395, + 0.00625962857156992, + 0.005591412540525198, + 0.0051672388799488544, + 0.004210616461932659, + 0.003424677299335599, + 0.004850644152611494, + 0.009483068250119686, + 0.006389516871422529, + 0.006964729633182287, + 0.005558301694691181, + 0.008060231804847717, + 0.010365410707890987, + 0.006483713164925575, + 0.006460823584347963, + 0.007677717600017786, + 0.0092872753739357, + 0.006943800952285528, + 0.006841908674687147, + 0.00522518390789628, + 0.0064645446836948395, + 0.006883750669658184, + 0.004418903961777687, + 0.0060857986100018024, + 0.004817043896764517, + 0.006073789671063423, + 0.005370218772441149, + 0.006033001001924276, + 0.006914213765412569, + 0.0050755091942846775, + 0.004438317380845547, + 0.0026361700147390366, + 0.005053323227912188, + 0.004144343547523022, + 0.005023470148444176, + 0.007473088335245848, + 0.004912231117486954, + 0.007649398874491453, + 0.004440302960574627, + 0.004184667021036148, + 0.006899519823491573, + 0.004416301380842924, + 0.0056390054523944855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.00864903163164854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294969916343689 + ], + "opposite_logprob_tokens": [ + -2.2379496097564697 + ], + "logprob_reward": -0.11294969916343689, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294969916343689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1941228210926056, + 0.0712992250919342, + 0.0716945081949234, + 0.12167846411466599, + 0.0566125214099884, + 0.0765056237578392, + 0.10666926950216293, + 0.15001706779003143, + 0.0639743059873581, + 0.08742616325616837 + ], + "credit_map": [ + 0.022639906033873558, + 0.033385273069143295, + 0.030976427718997, + 0.03192880004644394, + 0.023012926802039146, + 0.016698362305760384, + 0.015920808538794518, + 0.019560320302844048, + 0.006214227993041277, + 0.012105853296816349, + 0.006690371781587601, + 0.007998570799827576, + 0.009178819134831429, + 0.010124430991709232, + 0.00867237988859415, + 0.010314570739865303, + 0.008081507869064808, + 0.011454463005065918, + 0.005637256428599358, + 0.007788481656461954, + 0.008053090423345566, + 0.008907369337975979, + 0.014187960885465145, + 0.007584377191960812, + 0.006366214249283075, + 0.005777248181402683, + 0.004891824908554554, + 0.0075218710117042065, + 0.008684009313583374, + 0.042069729417562485, + 0.0062893894501030445, + 0.040078189224004745, + 0.005564415827393532, + 0.004464106634259224, + 0.007747573312371969, + 0.005966040305793285, + 0.009999681264162064, + 0.004758288152515888, + 0.009274061769247055, + 0.008838360197842121, + 0.008643020875751972, + 0.00511748855933547, + 0.0047541274689137936, + 0.008007029071450233, + 0.012518275529146194, + 0.009014872834086418, + 0.018112340942025185, + 0.010338467545807362, + 0.00877244770526886, + 0.012257955968379974, + 0.014422689564526081, + 0.00999824982136488, + 0.010904372669756413, + 0.012562522664666176, + 0.014184746891260147, + 0.023566290736198425, + 0.01878870092332363, + 0.03348884731531143, + 0.01914192922413349, + 0.01206604577600956, + 0.02253229357302189, + 0.014568706974387169, + 0.015415547415614128, + 0.014014994725584984, + 0.008617345243692398, + 0.006368826609104872, + 0.011419277638196945, + 0.009251237846910954, + 0.007424269337207079, + 0.005582891870290041, + 0.0064479270949959755, + 0.008862525224685669, + 0.010865418240427971, + 0.012025164440274239, + 0.010013516992330551, + 0.008592834696173668, + 0.008032631129026413, + 0.013173024170100689, + 0.013320546597242355, + 0.011403028853237629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.11294969916343689 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude drone shot slowly descends and advances, passing through layers of acid-soaked metal overpasses and a maze of pipes, finally focusing on the floor of an urban canyon. There, a scholar moves silently, his face thin and grave, his signature long, graying beard swaying slightly in the humid air. His silhouette bears a striking resemblance to the late Qing Dynasty thinker Kang Youwei. He wears a thick, dark gray traditional cotton and linen robe, its hems darkened by dirt from the floor, a stark contrast to the cold, shimmering metal surrounding him. Giant holographic billboards flicker and morph silently on the skyscraper facades, casting dynamic splatters of magenta and cyan light that reflect in his wary eyes. He hunched over the shadows of the wall, stealing forward with almost silent, cautious steps, each steady and deliberate. He paused periodically, glancing up at the searchlights of patrolling drones overhead before decisively merging into the deeper darkness. The entire scene was shrouded in a sense of high-tech oppression and the tension of solitary risk. Amidst the cold tones of the canvas, the vibrant neon lights and the quaint silhouettes of the figures combined to create a cyberpunk canvas brimming with dramatic tension.", + "seed": 23116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000165/group_9/index_6247_seed_23116_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.29122621901333334, + "vlm_logprob": -0.29122621901333334, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.29122621901333334, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9827442765235901 + ], + "opposite_logprob_tokens": [ + -0.4827442765235901 + ], + "logprob_reward": -0.9827442765235901, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9827442765235901, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09923633188009262, + 0.07732819020748138, + 0.09540429711341858, + 0.09085441380739212, + 0.13380476832389832, + 0.10188545286655426, + 0.11665983498096466, + 0.08888936787843704, + 0.06493952125310898, + 0.13099776208400726 + ], + "credit_map": [ + 0.012342221103608608, + 0.016384726390242577, + 0.012888760305941105, + 0.01374487392604351, + 0.010106522589921951, + 0.013643795624375343, + 0.011242474429309368, + 0.008882969617843628, + 0.012404765002429485, + 0.013568466529250145, + 0.007528898771852255, + 0.010980919003486633, + 0.007515561301261187, + 0.007561650592833757, + 0.008912025019526482, + 0.008855914697051048, + 0.00917066354304552, + 0.02000420168042183, + 0.005587216466665268, + 0.013794212602078915, + 0.009065323509275913, + 0.010549289174377918, + 0.011779236607253551, + 0.015454158186912537, + 0.009383944794535637, + 0.018726440146565437, + 0.01081977691501379, + 0.01181949395686388, + 0.01246610376983881, + 0.010334036312997341, + 0.007447725627571344, + 0.00985688902437687, + 0.014278248883783817, + 0.0165025033056736, + 0.017669109627604485, + 0.014863100834190845, + 0.017991947010159492, + 0.018662894144654274, + 0.01675332896411419, + 0.017083637416362762, + 0.012303600087761879, + 0.011456524953246117, + 0.009935279376804829, + 0.009832805953919888, + 0.019808171316981316, + 0.01718440279364586, + 0.009698087349534035, + 0.011666588485240936, + 0.01936762034893036, + 0.011196902021765709, + 0.021152744069695473, + 0.01509562972933054, + 0.021495947614312172, + 0.010817121714353561, + 0.009209669195115566, + 0.008324208669364452, + 0.009673681110143661, + 0.011687814258038998, + 0.010153462179005146, + 0.014954876154661179, + 0.006146748550236225, + 0.008539034053683281, + 0.01084896456450224, + 0.016884788870811462, + 0.008629737421870232, + 0.0071595702320337296, + 0.007692547980695963, + 0.009696423076093197, + 0.00661015510559082, + 0.008811136707663536, + 0.007613915950059891, + 0.00872603990137577, + 0.017039591446518898, + 0.010291372425854206, + 0.013386007398366928, + 0.012192122638225555, + 0.016727779060602188, + 0.025391004979610443, + 0.015263107605278492, + 0.020706793293356895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.9827442765235901 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042283177375793 + ], + "opposite_logprob_tokens": [ + -1.9104228019714355 + ], + "logprob_reward": -0.16042283177375793, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042283177375793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1860215812921524, + 0.06699102371931076, + 0.10058653354644775, + 0.06673703342676163, + 0.11974599212408066, + 0.06960242241621017, + 0.06066342443227768, + 0.07620955258607864, + 0.10048111528158188, + 0.15296132862567902 + ], + "credit_map": [ + 0.03162718191742897, + 0.03324994072318077, + 0.02703149802982807, + 0.02289203554391861, + 0.016680140048265457, + 0.019012384116649628, + 0.01931590586900711, + 0.016212476417422295, + 0.0129529545083642, + 0.009560363367199898, + 0.008375351317226887, + 0.009848983958363533, + 0.005194399505853653, + 0.005775945261120796, + 0.007082287687808275, + 0.008200736716389656, + 0.013278028927743435, + 0.014784252271056175, + 0.008189486339688301, + 0.008942718617618084, + 0.012828035280108452, + 0.009789405390620232, + 0.011961119249463081, + 0.020813487470149994, + 0.007227750960737467, + 0.008477944880723953, + 0.012730865739285946, + 0.008093664422631264, + 0.006478297524154186, + 0.007934524677693844, + 0.006890113931149244, + 0.008903877809643745, + 0.011485467664897442, + 0.01488455943763256, + 0.011649818159639835, + 0.011489972472190857, + 0.013241949491202831, + 0.02726701647043228, + 0.015097536146640778, + 0.014629671350121498, + 0.006658065598458052, + 0.008192437700927258, + 0.0058424267917871475, + 0.006975483149290085, + 0.012477340176701546, + 0.013689691200852394, + 0.0064883725717663765, + 0.00927861500531435, + 0.005998923443257809, + 0.008654030971229076, + 0.009948755614459515, + 0.009315344505012035, + 0.008729182183742523, + 0.004942683037370443, + 0.005542959552258253, + 0.007531545590609312, + 0.007328568026423454, + 0.006967020686715841, + 0.008615761995315552, + 0.010077604092657566, + 0.007591822184622288, + 0.011754876933991909, + 0.008247315883636475, + 0.015626581385731697, + 0.006810418795794249, + 0.006653506774455309, + 0.007368133869022131, + 0.013131987303495407, + 0.005315941292792559, + 0.02699599601328373, + 0.010919155552983284, + 0.02328597754240036, + 0.019244009628891945, + 0.009056465700268745, + 0.009411186911165714, + 0.008176916278898716, + 0.0087586035951972, + 0.04902644082903862, + 0.008779055438935757, + 0.040508657693862915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.16042283177375793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155532658100128 + ], + "opposite_logprob_tokens": [ + -1.7015553712844849 + ], + "logprob_reward": -0.20155532658100128, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155532658100128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07071531563997269, + 0.05098074674606323, + 0.06516985595226288, + 0.22518417239189148, + 0.15648019313812256, + 0.08897463232278824, + 0.11199194937944412, + 0.05221875384449959, + 0.13946533203125, + 0.03881910443305969 + ], + "credit_map": [ + 0.008468138985335827, + 0.018651815131306648, + 0.005590990651398897, + 0.009150712750852108, + 0.007278627250343561, + 0.00862615555524826, + 0.006977159529924393, + 0.005971712991595268, + 0.010729633271694183, + 0.008512286469340324, + 0.007259435020387173, + 0.005903610028326511, + 0.004041189327836037, + 0.005122085101902485, + 0.0034047679509967566, + 0.0060077328234910965, + 0.004693792667239904, + 0.010182887315750122, + 0.0040252478793263435, + 0.007880675606429577, + 0.00948687270283699, + 0.006773198489099741, + 0.008284490555524826, + 0.013842683285474777, + 0.022978520020842552, + 0.03545267879962921, + 0.027218662202358246, + 0.02234998159110546, + 0.04177055507898331, + 0.03540178760886192, + 0.017500530928373337, + 0.0225114393979311, + 0.01688195765018463, + 0.0218038447201252, + 0.023132912814617157, + 0.01986987702548504, + 0.017786171287298203, + 0.02353847026824951, + 0.014971002005040646, + 0.018495948985219002, + 0.007315861526876688, + 0.009039475582540035, + 0.009249345399439335, + 0.011620329692959785, + 0.01416324544698, + 0.014041513204574585, + 0.011958590708673, + 0.011586266569793224, + 0.014963875524699688, + 0.011346080340445042, + 0.020208630710840225, + 0.016339940950274467, + 0.01568691059947014, + 0.010764733888208866, + 0.00905951764434576, + 0.013622254133224487, + 0.0061654881574213505, + 0.004541042726486921, + 0.005824064835906029, + 0.009450213983654976, + 0.003587497631087899, + 0.006102194078266621, + 0.006514080334454775, + 0.01003416907042265, + 0.016848061233758926, + 0.012241853401064873, + 0.015630057081580162, + 0.016524510458111763, + 0.010158108547329903, + 0.029499376192688942, + 0.010715199634432793, + 0.02784816361963749, + 0.007726145442575216, + 0.004719509743154049, + 0.0028531518764793873, + 0.004767023958265781, + 0.0027525760233402252, + 0.006327169016003609, + 0.0041719782166182995, + 0.005501546431332827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.20155532658100128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089714244008064 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089714244008064, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089714244008064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10083140432834625, + 0.10621374845504761, + 0.06027320772409439, + 0.10441847145557404, + 0.13619396090507507, + 0.04416308179497719, + 0.07029446214437485, + 0.09295421093702316, + 0.17491427063941956, + 0.10974320024251938 + ], + "credit_map": [ + 0.015130124054849148, + 0.01344718411564827, + 0.019636305049061775, + 0.014979409985244274, + 0.00757830124348402, + 0.012445156462490559, + 0.00879537034779787, + 0.008819546550512314, + 0.013321522623300552, + 0.014766587875783443, + 0.013997089117765427, + 0.01589955948293209, + 0.007663985714316368, + 0.013017266988754272, + 0.013104826211929321, + 0.014442907646298409, + 0.007318912073969841, + 0.014637977816164494, + 0.0038819971960037947, + 0.007926455698907375, + 0.005599096417427063, + 0.006727828644216061, + 0.006287192460149527, + 0.007893736474215984, + 0.006383073981851339, + 0.008327147923409939, + 0.009628430008888245, + 0.009497249498963356, + 0.018899988383054733, + 0.028966208919882774, + 0.008669506758451462, + 0.014046850614249706, + 0.012158854864537716, + 0.015132518485188484, + 0.015250074677169323, + 0.01295069232583046, + 0.02226627804338932, + 0.025901662185788155, + 0.017087653279304504, + 0.015446207486093044, + 0.004372346680611372, + 0.004825251642614603, + 0.0038453456945717335, + 0.004105257801711559, + 0.009041594341397285, + 0.008841359987854958, + 0.003295893082395196, + 0.005836030002683401, + 0.008709650486707687, + 0.010516502894461155, + 0.01106752548366785, + 0.012000377289950848, + 0.01001550629734993, + 0.0051818047650158405, + 0.006547518074512482, + 0.006255568470805883, + 0.014352362602949142, + 0.006691924296319485, + 0.013710587285459042, + 0.012600383721292019, + 0.006705297622829676, + 0.011689843609929085, + 0.008910694159567356, + 0.0182931087911129, + 0.013827924616634846, + 0.008036716841161251, + 0.01113826408982277, + 0.011365550570189953, + 0.009117355570197105, + 0.05868807062506676, + 0.009183415211737156, + 0.053556960076093674, + 0.008995220996439457, + 0.006574080791324377, + 0.006767545826733112, + 0.00846914667636156, + 0.005342353601008654, + 0.03768862411379814, + 0.005675199907273054, + 0.030231023207306862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.011089714244008064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a high-altitude drone shot that slowly descends and advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed of acid-soaked metal overpasses, a maze of pipes, and an urban canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scholar visible wearing a thick, dark gray traditional cotton and linen robe with dirt-stained hems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant holographic billboards flickering on skyscraper facades, casting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scholar move with cautious, deliberate steps while glancing up at searchlights from patrolling drones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031894594430923 + ], + "opposite_logprob_tokens": [ + -2.3503189086914062 + ], + "logprob_reward": -0.10031894594430923, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031894594430923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07271893322467804, + 0.06070387363433838, + 0.07409019768238068, + 0.15371890366077423, + 0.10839539021253586, + 0.13816028833389282, + 0.0670962780714035, + 0.11657106876373291, + 0.12247353047132492, + 0.08607162535190582 + ], + "credit_map": [ + 0.010123559273779392, + 0.01263738889247179, + 0.010029659606516361, + 0.010632836259901524, + 0.005942999385297298, + 0.009426986798644066, + 0.007258312311023474, + 0.006667180452495813, + 0.007397040259093046, + 0.009042346850037575, + 0.005280954297631979, + 0.00922213215380907, + 0.005374135449528694, + 0.007984072901308537, + 0.005655196495354176, + 0.01074798870831728, + 0.00848365481942892, + 0.013328162021934986, + 0.005463856738060713, + 0.010763490572571754, + 0.009891366586089134, + 0.006491243373602629, + 0.007535998709499836, + 0.012132414616644382, + 0.009185625240206718, + 0.017064671963453293, + 0.014066698029637337, + 0.011929717846214771, + 0.02416165918111801, + 0.03608202934265137, + 0.00894614402204752, + 0.03228233382105827, + 0.006533550564199686, + 0.011466423980891705, + 0.009391793981194496, + 0.010578877292573452, + 0.025304406881332397, + 0.022734874859452248, + 0.009116950444877148, + 0.013268506154417992, + 0.011593668721616268, + 0.018595099449157715, + 0.013434034772217274, + 0.01692797616124153, + 0.019628506153821945, + 0.0209656972438097, + 0.018795276060700417, + 0.01822001487016678, + 0.008088751696050167, + 0.009170289151370525, + 0.011170981451869011, + 0.010310385376214981, + 0.01027525682002306, + 0.005396849010139704, + 0.006364126689732075, + 0.006319630425423384, + 0.0106956260278821, + 0.015298563987016678, + 0.011980877257883549, + 0.022781793028116226, + 0.00906192697584629, + 0.009149530902504921, + 0.01721736043691635, + 0.020385373383760452, + 0.01675358973443508, + 0.009492320008575916, + 0.014093473553657532, + 0.010025442577898502, + 0.008085559122264385, + 0.026176661252975464, + 0.012819532305002213, + 0.025026943534612656, + 0.013288739137351513, + 0.019019387662410736, + 0.0049699959345161915, + 0.012559929862618446, + 0.0045782774686813354, + 0.013630864210426807, + 0.005648751277476549, + 0.0123756667599082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.10031894594430923 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_tensor([6348, 6348])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_tensor([6348, 6348])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d9f998cc5a3cd066c9e7967c27e80ece1ad3c1e8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_tensor([6348, 6348])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking low-angle close-up captures a massive white tiger at the very center of the frame, its massive bulk practically filling the entire frame. The beast's thick, snow-white fur is marked with distinct inky black stripes, and its icy blue eyes gleam with a fierce glint, fixed intently on the camera. Its maw, gaping, reveals its dagger-sharp canines, its barbed pink tongue curling with intensity as it unleashes a heartfelt, ear-piercing roar. Its bristly whiskers tremble with fury as it roars, its powerful sonic waves seemingly shaking the very air. The hot breath from its throat condenses into a white mist in the cool air. The background is the gloomy jungle. A few faint rays of light penetrate the dense canopy, sculpting the tiger's muscular shoulders and reflecting off its wet teeth. The entire picture is full of primitive power. The low-angle upward perspective greatly enhances its size and majesty, creating a suffocating, wild and threatening sense of oppression.", + "seed": 27062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_6348_seed_27062_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02201551999896765, + "vlm_logprob": -0.02201551999896765, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02201551999896765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006845711264759302 + ], + "opposite_logprob_tokens": [ + -5.381845474243164 + ], + "logprob_reward": -0.006845711264759302, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006845711264759302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1304815113544464, + 0.37102067470550537, + 0.034502431750297546, + 0.21272790431976318, + 0.07243940979242325, + 0.015997283160686493, + 0.05022981017827988, + 0.03658084571361542, + 0.027307990938425064, + 0.048712100833654404 + ], + "credit_map": [ + 0.016206201165914536, + 0.01010376401245594, + 0.03359007462859154, + 0.016381001099944115, + 0.01817837543785572, + 0.011187566444277763, + 0.015836456790566444, + 0.008998075500130653, + 0.08653298765420914, + 0.01958802528679371, + 0.13774998486042023, + 0.024444803595542908, + 0.03086448274552822, + 0.043087977916002274, + 0.016459960490465164, + 0.01229248009622097, + 0.0043630716390907764, + 0.0035237683914601803, + 0.006756247952580452, + 0.004786787088960409, + 0.00512994360178709, + 0.0037665378767997026, + 0.0030767994467169046, + 0.003099277149885893, + 0.025674184784293175, + 0.023460745811462402, + 0.029214458540081978, + 0.03241371363401413, + 0.017565209418535233, + 0.021594438701868057, + 0.02476208098232746, + 0.03804308548569679, + 0.004614746663719416, + 0.003865676000714302, + 0.0244402214884758, + 0.00453567411750555, + 0.004661313723772764, + 0.004244458395987749, + 0.004409812856465578, + 0.021667513996362686, + 0.0023139554541558027, + 0.002410829532891512, + 0.0017772301798686385, + 0.0018710297299548984, + 0.001735529163852334, + 0.002295198617503047, + 0.001564473146572709, + 0.0020290380343794823, + 0.007313982583582401, + 0.004745439160615206, + 0.005209451541304588, + 0.007618552073836327, + 0.00689130462706089, + 0.00805286318063736, + 0.0057797543704509735, + 0.004618466831743717, + 0.005304635502398014, + 0.0032716344576328993, + 0.0036932467482984066, + 0.00453962292522192, + 0.005640231538563967, + 0.004062716383486986, + 0.005600251257419586, + 0.004468510393053293, + 0.009138199500739574, + 0.0023660066071897745, + 0.0045700580812990665, + 0.002402752870693803, + 0.0019213201012462378, + 0.0019467435777187347, + 0.0023249706719070673, + 0.002637941623106599, + 0.0047872429713606834, + 0.004770450759679079, + 0.00631722854450345, + 0.006897277198731899, + 0.003637690795585513, + 0.004544671159237623, + 0.011283167637884617, + 0.006474373396486044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.006845711264759302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004659153055399656 + ], + "opposite_logprob_tokens": [ + -5.379659175872803 + ], + "logprob_reward": -0.004659153055399656, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004659153055399656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1961570382118225, + 0.32043230533599854, + 0.021672062575817108, + 0.03306013345718384, + 0.1111844927072525, + 0.11805179715156555, + 0.038870420306921005, + 0.03294038400053978, + 0.05834514647722244, + 0.0692863017320633 + ], + "credit_map": [ + 0.03399207815527916, + 0.01495625264942646, + 0.024308662861585617, + 0.018152570351958275, + 0.03575131297111511, + 0.028541097417473793, + 0.022373611107468605, + 0.018081437796354294, + 0.0819489061832428, + 0.016110002994537354, + 0.1263243407011032, + 0.01884276606142521, + 0.02102598361670971, + 0.03253438323736191, + 0.01521016750484705, + 0.008435732685029507, + 0.0033052079379558563, + 0.001320435432717204, + 0.0023695267736911774, + 0.0018425193848088384, + 0.0022390822414308786, + 0.002119730692356825, + 0.006527131889015436, + 0.0019484275253489614, + 0.005505343899130821, + 0.0032176291570067406, + 0.002035904908552766, + 0.0025093944277614355, + 0.005391222424805164, + 0.0031389680225402117, + 0.004600116983056068, + 0.006661551538854837, + 0.00701581547036767, + 0.005681127775460482, + 0.03866977617144585, + 0.005892560817301273, + 0.006821895483881235, + 0.00732673704624176, + 0.006743974518030882, + 0.03303259611129761, + 0.008335426449775696, + 0.02083524502813816, + 0.011853127740323544, + 0.01674884557723999, + 0.01637556217610836, + 0.018878890201449394, + 0.012111557647585869, + 0.012913135811686516, + 0.005154088605195284, + 0.003851566230878234, + 0.0034316040109843016, + 0.005087409168481827, + 0.006160789169371128, + 0.005554391536861658, + 0.0061158230528235435, + 0.003514743410050869, + 0.004875365179032087, + 0.0028385533951222897, + 0.0041346768848598, + 0.003869150299578905, + 0.005216090939939022, + 0.0032232683151960373, + 0.005532620940357447, + 0.003250653389841318, + 0.00869678519666195, + 0.005797011312097311, + 0.007172148209065199, + 0.010850333608686924, + 0.005404829513281584, + 0.006766654551029205, + 0.006171524059027433, + 0.007485856767743826, + 0.007186178117990494, + 0.006795060820877552, + 0.008992016315460205, + 0.010372794233262539, + 0.005546682048588991, + 0.005989336408674717, + 0.014817505143582821, + 0.00958672259002924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.004659153055399656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620482861995697 + ], + "opposite_logprob_tokens": [ + -2.8120481967926025 + ], + "logprob_reward": -0.0620482861995697, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620482861995697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2270127683877945, + 0.09511131793260574, + 0.09864376485347748, + 0.036354660987854004, + 0.11900578439235687, + 0.08734879642724991, + 0.13492800295352936, + 0.050524868071079254, + 0.05314180999994278, + 0.0979282408952713 + ], + "credit_map": [ + 0.03718751668930054, + 0.016631843522191048, + 0.030155625194311142, + 0.02141791209578514, + 0.04003789275884628, + 0.03224380314350128, + 0.02579480968415737, + 0.023543357849121094, + 0.015735303983092308, + 0.005184291396290064, + 0.02901667356491089, + 0.008057010360062122, + 0.007975568063557148, + 0.011484557762742043, + 0.0114892004057765, + 0.006168711930513382, + 0.015062129124999046, + 0.011699039489030838, + 0.0182011891156435, + 0.010867899283766747, + 0.01031443290412426, + 0.007593412883579731, + 0.016656238585710526, + 0.008249416947364807, + 0.003552438458427787, + 0.0031244345009326935, + 0.004681607708334923, + 0.0038791666738688946, + 0.006644830107688904, + 0.003581336699426174, + 0.006320411339402199, + 0.00457043619826436, + 0.006957305129617453, + 0.005881716031581163, + 0.03906258940696716, + 0.0063328733667731285, + 0.00860334187746048, + 0.006603396963328123, + 0.008290726691484451, + 0.037273842841386795, + 0.007364083081483841, + 0.011344736441969872, + 0.013415282592177391, + 0.01141273695975542, + 0.009279120713472366, + 0.009921778924763203, + 0.011394788511097431, + 0.013216269202530384, + 0.01943923346698284, + 0.011112992651760578, + 0.012859259732067585, + 0.017921816557645798, + 0.022223778069019318, + 0.020186306908726692, + 0.01959526352584362, + 0.011589355766773224, + 0.009580378420650959, + 0.0038470979779958725, + 0.005764690227806568, + 0.004764250945299864, + 0.006737655494362116, + 0.004294591024518013, + 0.0061243511736392975, + 0.009411850944161415, + 0.01059897430241108, + 0.005154103972017765, + 0.006896392907947302, + 0.008887269534170628, + 0.005250591319054365, + 0.005533560179173946, + 0.00460425578057766, + 0.0062166606076061726, + 0.011789757758378983, + 0.011211372911930084, + 0.013698477298021317, + 0.014540762640535831, + 0.00747483316808939, + 0.006885406095534563, + 0.023120593279600143, + 0.009207039140164852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0620482861995697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793115332722664 + ], + "opposite_logprob_tokens": [ + -3.5297930240631104 + ], + "logprob_reward": -0.029793115332722664, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793115332722664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10538823157548904, + 0.1791384071111679, + 0.11147564649581909, + 0.05987031012773514, + 0.18390542268753052, + 0.06699289381504059, + 0.04629479721188545, + 0.1772015243768692, + 0.04965852573513985, + 0.020074255764484406 + ], + "credit_map": [ + 0.01452519092708826, + 0.007654577028006315, + 0.02066526561975479, + 0.009197408333420753, + 0.016134632751345634, + 0.012960375286638737, + 0.011439588852226734, + 0.012811187654733658, + 0.04505326971411705, + 0.00883077085018158, + 0.06614912301301956, + 0.009839643724262714, + 0.013689011335372925, + 0.01744878850877285, + 0.010138080455362797, + 0.007989735342562199, + 0.020391933619976044, + 0.013896836899220943, + 0.02402195893228054, + 0.011304493993520737, + 0.01208167988806963, + 0.008900327607989311, + 0.012006878852844238, + 0.008871539495885372, + 0.011926316656172276, + 0.005173567216843367, + 0.00437754625454545, + 0.00550273759290576, + 0.005665611941367388, + 0.0053609441965818405, + 0.006508623715490103, + 0.015354962088167667, + 0.010756609961390495, + 0.008189959451556206, + 0.06243209168314934, + 0.011478406377136707, + 0.012315821833908558, + 0.012652156874537468, + 0.011060459539294243, + 0.055019915103912354, + 0.010012365877628326, + 0.00929952785372734, + 0.006370968651026487, + 0.007969368249177933, + 0.01120896264910698, + 0.006583665497601032, + 0.008358326740562916, + 0.007189704570919275, + 0.006567924749106169, + 0.003065308555960655, + 0.004933317191898823, + 0.003790727350860834, + 0.009819748811423779, + 0.006233572028577328, + 0.006804481614381075, + 0.00507971690967679, + 0.05394811928272247, + 0.015287978574633598, + 0.032564274966716766, + 0.01526164822280407, + 0.018334418535232544, + 0.010029107332229614, + 0.01815996877849102, + 0.013616006821393967, + 0.019162863492965698, + 0.003326279344037175, + 0.009652391076087952, + 0.003689610632136464, + 0.0029627876356244087, + 0.00381223252043128, + 0.003734593279659748, + 0.0033177663572132587, + 0.0029213293455541134, + 0.0024047570768743753, + 0.002377820899710059, + 0.002577222650870681, + 0.0021446894388645887, + 0.0024501350708305836, + 0.0024301474913954735, + 0.0027681533247232437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.029793115332722664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731334142386913 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731334142386913, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731334142386913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09881209582090378, + 0.052302755415439606, + 0.05599880963563919, + 0.055996123701334, + 0.30193623900413513, + 0.04976916313171387, + 0.02450818382203579, + 0.03331474959850311, + 0.11612246930599213, + 0.21123948693275452 + ], + "credit_map": [ + 0.023401491343975067, + 0.007259158883243799, + 0.012176547199487686, + 0.008732562884688377, + 0.01725785806775093, + 0.012501494958996773, + 0.010549947619438171, + 0.00693301809951663, + 0.010926368646323681, + 0.002243828261271119, + 0.019102396443486214, + 0.002937223296612501, + 0.0036096242256462574, + 0.00781315378844738, + 0.002286093309521675, + 0.003384056966751814, + 0.00920882448554039, + 0.003813195740804076, + 0.019133776426315308, + 0.0045096599496901035, + 0.008005188778042793, + 0.003639611415565014, + 0.003723723581060767, + 0.003964822273701429, + 0.007015878800302744, + 0.0055551291443407536, + 0.008152478374540806, + 0.008899410255253315, + 0.00494407769292593, + 0.005430255550891161, + 0.005901223514229059, + 0.010097665712237358, + 0.017810605466365814, + 0.014260858297348022, + 0.10964174568653107, + 0.016980240121483803, + 0.015417998656630516, + 0.017306776717305183, + 0.015924500301480293, + 0.09459348022937775, + 0.004024316556751728, + 0.007692251354455948, + 0.006128585431724787, + 0.006397136487066746, + 0.00758866872638464, + 0.005775659345090389, + 0.006190147716552019, + 0.005972391925752163, + 0.0032590902410447598, + 0.0014995916280895472, + 0.002271074801683426, + 0.00411132862791419, + 0.005158635322004557, + 0.0020586526952683926, + 0.0035500377416610718, + 0.0025997688062489033, + 0.007885148748755455, + 0.0024554533883929253, + 0.003585286671295762, + 0.0030385470017790794, + 0.004322467837482691, + 0.0028288017492741346, + 0.004411422181874514, + 0.004787615034729242, + 0.026611285284161568, + 0.0096018360927701, + 0.015742408111691475, + 0.018284356221556664, + 0.010286841541528702, + 0.01164221577346325, + 0.011233947239816189, + 0.012719564139842987, + 0.01563250459730625, + 0.023961298167705536, + 0.02759680710732937, + 0.041070789098739624, + 0.01850121095776558, + 0.016054637730121613, + 0.04259974882006645, + 0.025822462514042854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.006731334142386913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle close-up captures a massive white tiger at the very center of the frame, its massive bulk practically filling the entire frame. The beast's thick, snow-white fur is marked with distinct inky black stripes, and its icy blue eyes gleam with a fierce glint, fixed intently on the camera. Its maw, gaping, reveals its dagger-sharp canines, its barbed pink tongue curling with intensity as it unleashes a heartfelt, ear-piercing roar. Its bristly whiskers tremble with fury as it roars, its powerful sonic waves seemingly shaking the very air. The hot breath from its throat condenses into a white mist in the cool air. The background is the gloomy jungle. A few faint rays of light penetrate the dense canopy, sculpting the tiger's muscular shoulders and reflecting off its wet teeth. The entire picture is full of primitive power. The low-angle upward perspective greatly enhances its size and majesty, creating a suffocating, wild and threatening sense of oppression.", + "seed": 27063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_6348_seed_27063_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023972247913479804, + "vlm_logprob": -0.023972247913479804, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.023972247913479804, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006593969650566578 + ], + "opposite_logprob_tokens": [ + -6.006594181060791 + ], + "logprob_reward": -0.006593969650566578, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006593969650566578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06488681584596634, + 0.06651151180267334, + 0.09887662529945374, + 0.01742832362651825, + 0.08224592357873917, + 0.5765766501426697, + 0.02660624496638775, + 0.03686269000172615, + 0.014117424376308918, + 0.015887735411524773 + ], + "credit_map": [ + 0.007679521106183529, + 0.006079402286559343, + 0.012488202191889286, + 0.008634151890873909, + 0.009891420602798462, + 0.005006407853215933, + 0.008389116264879704, + 0.006718598306179047, + 0.01724342070519924, + 0.005750325042754412, + 0.021309753879904747, + 0.005218100734055042, + 0.0034659476950764656, + 0.0035319693852216005, + 0.005676242057234049, + 0.004315762780606747, + 0.010363309644162655, + 0.016177870333194733, + 0.012172240763902664, + 0.010786926373839378, + 0.01666608639061451, + 0.010452430695295334, + 0.008101580664515495, + 0.014156194403767586, + 0.0013913206057623029, + 0.0017129152547568083, + 0.001601374358870089, + 0.002349117072299123, + 0.004148963373154402, + 0.0010554544860497117, + 0.0032818301115185022, + 0.0018873499939218163, + 0.010104655288159847, + 0.007609204854816198, + 0.010075914673507214, + 0.008348643779754639, + 0.011458514258265495, + 0.01719387248158455, + 0.009487454779446125, + 0.007967664860188961, + 0.06860587000846863, + 0.027671748772263527, + 0.11034970730543137, + 0.03537148982286453, + 0.043597858399152756, + 0.151403546333313, + 0.02895481325685978, + 0.11062167584896088, + 0.0030687772668898106, + 0.0025590346194803715, + 0.0030311483424156904, + 0.002717802533879876, + 0.0032471432350575924, + 0.004668344743549824, + 0.0031946587841957808, + 0.00411934033036232, + 0.00557711673900485, + 0.003319286974146962, + 0.006597177125513554, + 0.004312396515160799, + 0.0035112695768475533, + 0.003492958378046751, + 0.005294403061270714, + 0.004758084192872047, + 0.0027430474292486906, + 0.0016057832399383187, + 0.0016080735949799418, + 0.0016688142204657197, + 0.0011864012340083718, + 0.0012659273343160748, + 0.0016919575864449143, + 0.0023474209010601044, + 0.00177827849984169, + 0.0014744362561032176, + 0.001550929737277329, + 0.002926960587501526, + 0.0014529165346175432, + 0.0016799081349745393, + 0.0027912769000977278, + 0.0022330298088490963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.006593969650566578 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016134176403284073 + ], + "opposite_logprob_tokens": [ + -4.141134262084961 + ], + "logprob_reward": -0.016134176403284073, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016134176403284073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054548196494579315, + 0.04226177930831909, + 0.12418047338724136, + 0.037546977400779724, + 0.05056188255548477, + 0.11169487982988358, + 0.1115453839302063, + 0.24495068192481995, + 0.18194642663002014, + 0.04076338931918144 + ], + "credit_map": [ + 0.007271625101566315, + 0.005979110021144152, + 0.008990831673145294, + 0.004908127244561911, + 0.01076897419989109, + 0.005081385839730501, + 0.00654552411288023, + 0.005002615042030811, + 0.007240846287459135, + 0.004801936913281679, + 0.005519835744053125, + 0.006738319061696529, + 0.0032226985786110163, + 0.0036410600878298283, + 0.0033573450054973364, + 0.007739735301584005, + 0.02127576619386673, + 0.013146460056304932, + 0.010192656889557838, + 0.011041332967579365, + 0.016751596704125404, + 0.012011079117655754, + 0.009923390112817287, + 0.029838182032108307, + 0.003858606331050396, + 0.004627176560461521, + 0.006309846416115761, + 0.0047801462933421135, + 0.007061814423650503, + 0.0019622358959168196, + 0.005574088543653488, + 0.0033730650320649147, + 0.006560155190527439, + 0.007741183042526245, + 0.005756172817200422, + 0.005390947684645653, + 0.010623995214700699, + 0.005969025660306215, + 0.004295174963772297, + 0.004225225653499365, + 0.013134918175637722, + 0.005683744326233864, + 0.021199211478233337, + 0.008595341816544533, + 0.007856111973524094, + 0.028963329270482063, + 0.005403654649853706, + 0.020858565345406532, + 0.009239708073437214, + 0.011978181079030037, + 0.010804230347275734, + 0.013703174889087677, + 0.013936205767095089, + 0.023315344005823135, + 0.012638092041015625, + 0.015930436551570892, + 0.027718914672732353, + 0.018515635281801224, + 0.045148465782403946, + 0.027492187917232513, + 0.03336193040013313, + 0.026552855968475342, + 0.029171878471970558, + 0.036988820880651474, + 0.019731229171156883, + 0.016904372721910477, + 0.022378962486982346, + 0.020973199978470802, + 0.025740154087543488, + 0.017157047986984253, + 0.024602144956588745, + 0.034459326416254044, + 0.003342722076922655, + 0.004321111366152763, + 0.003678469220176339, + 0.006439517252147198, + 0.007850286550819874, + 0.003290235996246338, + 0.008490489795804024, + 0.003350555896759033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.016134176403284073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495497211813927 + ], + "opposite_logprob_tokens": [ + -2.929955005645752 + ], + "logprob_reward": -0.05495497211813927, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495497211813927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10801014304161072, + 0.06972085684537888, + 0.05548638477921486, + 0.07908536493778229, + 0.05116184055805206, + 0.47979140281677246, + 0.023748883977532387, + 0.04898005351424217, + 0.028293730691075325, + 0.05572136119008064 + ], + "credit_map": [ + 0.01194432470947504, + 0.013580042868852615, + 0.014095023274421692, + 0.011403581127524376, + 0.02179284393787384, + 0.012020278722047806, + 0.01410151831805706, + 0.009072523564100266, + 0.01764434576034546, + 0.007106556557118893, + 0.010045070201158524, + 0.009368096478283405, + 0.0037967637181282043, + 0.0028854343108832836, + 0.005170792806893587, + 0.013703787699341774, + 0.009869303554296494, + 0.006372573785483837, + 0.003838774049654603, + 0.004932258743792772, + 0.007307974621653557, + 0.00575815187767148, + 0.004912395030260086, + 0.012494955211877823, + 0.006735301576554775, + 0.009797247126698494, + 0.005847225431352854, + 0.008491242304444313, + 0.024235323071479797, + 0.006088089197874069, + 0.01050562597811222, + 0.007385311182588339, + 0.007621248718351126, + 0.00588536961004138, + 0.007463489193469286, + 0.004758899565786123, + 0.008252006024122238, + 0.008225275203585625, + 0.003860227530822158, + 0.005095323547720909, + 0.05584804713726044, + 0.024834254756569862, + 0.08569954335689545, + 0.03050745092332363, + 0.04001035913825035, + 0.1288033425807953, + 0.026156416162848473, + 0.08793197572231293, + 0.002729318104684353, + 0.002723552519455552, + 0.002695951145142317, + 0.003286992199718952, + 0.002717068884521723, + 0.0034087984822690487, + 0.002905237954109907, + 0.003281964687630534, + 0.005366469733417034, + 0.0047514066100120544, + 0.004301054868847132, + 0.007198312785476446, + 0.010880942456424236, + 0.0052021099254488945, + 0.007360008079558611, + 0.00391974626109004, + 0.0051034255884587765, + 0.0028496352024376392, + 0.0033557945862412453, + 0.0029063711408525705, + 0.0037683863192796707, + 0.0022277268581092358, + 0.004109497647732496, + 0.003972891718149185, + 0.0058734300546348095, + 0.0068400525487959385, + 0.0045348708517849445, + 0.008191036991775036, + 0.007983488962054253, + 0.005607451777905226, + 0.010293642990291119, + 0.006397383287549019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.05495497211813927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038085825741291046 + ], + "opposite_logprob_tokens": [ + -3.2880859375 + ], + "logprob_reward": -0.038085825741291046, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038085825741291046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09343130141496658, + 0.09337117522954941, + 0.04167080298066139, + 0.07220926135778427, + 0.04805358126759529, + 0.35327214002609253, + 0.0411258228123188, + 0.18740765750408173, + 0.048640988767147064, + 0.020817186683416367 + ], + "credit_map": [ + 0.01198596227914095, + 0.008197968825697899, + 0.014805350452661514, + 0.010290069505572319, + 0.016364993527531624, + 0.011234880425035954, + 0.012191194109618664, + 0.008360893465578556, + 0.029777120798826218, + 0.007862040773034096, + 0.031156186014413834, + 0.007125695236027241, + 0.005511544179171324, + 0.0032079473603516817, + 0.005416908301413059, + 0.0033137358259409666, + 0.006252889987081289, + 0.006091505754739046, + 0.003569894703105092, + 0.003982629626989365, + 0.005041755735874176, + 0.005541026592254639, + 0.0035089883022010326, + 0.007682114373892546, + 0.005887617357075214, + 0.012438377365469933, + 0.009959560818970203, + 0.009679384529590607, + 0.011297059245407581, + 0.005015319213271141, + 0.008377103134989738, + 0.009554840624332428, + 0.006074267905205488, + 0.008085603825747967, + 0.004860077053308487, + 0.00435197027400136, + 0.008314053528010845, + 0.006801917217671871, + 0.003974426072090864, + 0.005591269116848707, + 0.025435442104935646, + 0.012678345665335655, + 0.09782318770885468, + 0.02373494580388069, + 0.019874081015586853, + 0.041767336428165436, + 0.013696523383259773, + 0.11826226860284805, + 0.005047863814979792, + 0.004263031296432018, + 0.004449726548045874, + 0.005607711151242256, + 0.0034445144701749086, + 0.003518046112731099, + 0.006892527919262648, + 0.00790240429341793, + 0.029256394132971764, + 0.0159340500831604, + 0.03380456939339638, + 0.02122374437749386, + 0.017078671604394913, + 0.021196972578763962, + 0.01998269557952881, + 0.02893057093024254, + 0.005578535608947277, + 0.004536936990916729, + 0.005716424901038408, + 0.005606827326118946, + 0.007665358949452639, + 0.004730748012661934, + 0.006840412970632315, + 0.007965746335685253, + 0.00395249156281352, + 0.002455231500789523, + 0.0021536503918468952, + 0.003762142965570092, + 0.0015328478766605258, + 0.0016786801861599088, + 0.002515216590836644, + 0.0027669244445860386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.038085825741291046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092295654118061 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092295654118061, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092295654118061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1329272985458374, + 0.0880536288022995, + 0.05672881752252579, + 0.08061303198337555, + 0.09389305859804153, + 0.31712520122528076, + 0.06576432287693024, + 0.08984354138374329, + 0.05546155571937561, + 0.019589515402913094 + ], + "credit_map": [ + 0.016889706254005432, + 0.01011741440743208, + 0.02393607795238495, + 0.013587992638349533, + 0.02560594119131565, + 0.011384198442101479, + 0.017338521778583527, + 0.014067444950342178, + 0.017242932692170143, + 0.008868127129971981, + 0.026431839913129807, + 0.005995656829327345, + 0.005790199618786573, + 0.0032975065987557173, + 0.006127647589892149, + 0.014299716800451279, + 0.009311528876423836, + 0.006344135385006666, + 0.004619366955012083, + 0.005266306921839714, + 0.007763337809592485, + 0.00504529383033514, + 0.004401266574859619, + 0.013977578841149807, + 0.009524400345981121, + 0.006026656366884708, + 0.019416887313127518, + 0.009060453623533249, + 0.014158240519464016, + 0.003913360182195902, + 0.008414673618972301, + 0.010098355822265148, + 0.00934796966612339, + 0.013217740692198277, + 0.01056378148496151, + 0.008689791895449162, + 0.023410465568304062, + 0.009755570441484451, + 0.009107260964810848, + 0.009800473228096962, + 0.04385800287127495, + 0.020212220028042793, + 0.032739367336034775, + 0.023160262033343315, + 0.0319216214120388, + 0.10200385004281998, + 0.018341198563575745, + 0.0448886938393116, + 0.010741768404841423, + 0.008698942139744759, + 0.006682482548058033, + 0.006483237724751234, + 0.00733848987147212, + 0.013430619612336159, + 0.00539565272629261, + 0.006993131712079048, + 0.011151166632771492, + 0.005948151461780071, + 0.016636785119771957, + 0.010201672092080116, + 0.011234013363718987, + 0.009890416637063026, + 0.01037513092160225, + 0.014406208880245686, + 0.006763172801584005, + 0.004730730317533016, + 0.006643557921051979, + 0.006399264559149742, + 0.00743854371830821, + 0.0049588377587497234, + 0.0074322898872196674, + 0.011095158755779266, + 0.0027866140007972717, + 0.001955189276486635, + 0.001974571030586958, + 0.002397360047325492, + 0.002553072525188327, + 0.0021407143212854862, + 0.0035907982382923365, + 0.0021911966614425182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004092295654118061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle close-up captures a massive white tiger at the very center of the frame, its massive bulk practically filling the entire frame. The beast's thick, snow-white fur is marked with distinct inky black stripes, and its icy blue eyes gleam with a fierce glint, fixed intently on the camera. Its maw, gaping, reveals its dagger-sharp canines, its barbed pink tongue curling with intensity as it unleashes a heartfelt, ear-piercing roar. Its bristly whiskers tremble with fury as it roars, its powerful sonic waves seemingly shaking the very air. The hot breath from its throat condenses into a white mist in the cool air. The background is the gloomy jungle. A few faint rays of light penetrate the dense canopy, sculpting the tiger's muscular shoulders and reflecting off its wet teeth. The entire picture is full of primitive power. The low-angle upward perspective greatly enhances its size and majesty, creating a suffocating, wild and threatening sense of oppression.", + "seed": 27064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_6348_seed_27064_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02281073690392077, + "vlm_logprob": -0.02281073690392077, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.02281073690392077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0069818575866520405 + ], + "opposite_logprob_tokens": [ + -5.25698184967041 + ], + "logprob_reward": -0.0069818575866520405, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0069818575866520405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3041605055332184, + 0.05774635076522827, + 0.09628722071647644, + 0.1428741216659546, + 0.1354828029870987, + 0.06508579850196838, + 0.05551224574446678, + 0.03322755917906761, + 0.03740858659148216, + 0.07221480458974838 + ], + "credit_map": [ + 0.021437034010887146, + 0.017142431810498238, + 0.11529111862182617, + 0.01670793443918228, + 0.019004136323928833, + 0.015327336266636848, + 0.018334217369556427, + 0.08091630041599274, + 0.007190600503236055, + 0.011016709730029106, + 0.00599730433896184, + 0.005678181070834398, + 0.007847943343222141, + 0.007641003001481295, + 0.006163080222904682, + 0.006211531348526478, + 0.009803839959204197, + 0.012444775551557541, + 0.013011452741920948, + 0.013921082951128483, + 0.012558553367853165, + 0.0119231091812253, + 0.01168659795075655, + 0.010937809012830257, + 0.011493736878037453, + 0.02708684653043747, + 0.014333147555589676, + 0.0324927419424057, + 0.014369804412126541, + 0.016885429620742798, + 0.01688469760119915, + 0.009327707812190056, + 0.007026716135442257, + 0.007195928134024143, + 0.009008768014609814, + 0.013235156424343586, + 0.006001656875014305, + 0.05309336259961128, + 0.0154575752094388, + 0.024463634938001633, + 0.01197555847465992, + 0.0033644374925643206, + 0.006178305484354496, + 0.006846402771770954, + 0.005433847662061453, + 0.006965908221900463, + 0.013556435704231262, + 0.01076490804553032, + 0.0036770235747098923, + 0.005758990067988634, + 0.003720840672031045, + 0.007668846286833286, + 0.0039054718799889088, + 0.011329025030136108, + 0.015916191041469574, + 0.0035358555614948273, + 0.004653080832213163, + 0.005180435720831156, + 0.003174190642312169, + 0.004493375774472952, + 0.004099469631910324, + 0.0031805341131985188, + 0.0036119497381150723, + 0.0048345220275223255, + 0.0037412091623991728, + 0.004821420181542635, + 0.004564374219626188, + 0.003957092296332121, + 0.003653425956144929, + 0.004350567702203989, + 0.005431228782981634, + 0.006889268755912781, + 0.007662743330001831, + 0.006658476777374744, + 0.009404274635016918, + 0.0070617590099573135, + 0.00654221186414361, + 0.01323326863348484, + 0.003356130328029394, + 0.01829594001173973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0069818575866520405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00363522837869823 + ], + "opposite_logprob_tokens": [ + -5.628635406494141 + ], + "logprob_reward": -0.00363522837869823, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00363522837869823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31144997477531433, + 0.053244031965732574, + 0.045437779277563095, + 0.09774190932512283, + 0.034723784774541855, + 0.2430092841386795, + 0.03443049639463425, + 0.06008896976709366, + 0.049531176686286926, + 0.07034264504909515 + ], + "credit_map": [ + 0.02305818907916546, + 0.015346268191933632, + 0.12343727797269821, + 0.017632151022553444, + 0.017735108733177185, + 0.013031805865466595, + 0.018246212974190712, + 0.0829629898071289, + 0.00414667371660471, + 0.00813019834458828, + 0.00689899455755949, + 0.007457036525011063, + 0.005531595554202795, + 0.01003057137131691, + 0.005665660835802555, + 0.005383301060646772, + 0.0037755111698061228, + 0.006933110766112804, + 0.0047182985581457615, + 0.006465172860771418, + 0.007925961166620255, + 0.0052969432435929775, + 0.005229162983596325, + 0.005093615967780352, + 0.008342158980667591, + 0.0205726008862257, + 0.009584919549524784, + 0.015487609431147575, + 0.010819715447723866, + 0.010811799205839634, + 0.012634903192520142, + 0.009488199837505817, + 0.002758792834356427, + 0.005933716893196106, + 0.004506424535065889, + 0.005983172915875912, + 0.0045184120535850525, + 0.003563599195331335, + 0.004380674101412296, + 0.003078991547226906, + 0.040565356612205505, + 0.018096337094902992, + 0.021813174709677696, + 0.03647774085402489, + 0.020237885415554047, + 0.010956218466162682, + 0.07515216618776321, + 0.019710414111614227, + 0.003484605113044381, + 0.004714210983365774, + 0.0025253670755773783, + 0.005574462935328484, + 0.00428197905421257, + 0.007086019031703472, + 0.004607225302606821, + 0.0021566280629485846, + 0.006241690367460251, + 0.009450707584619522, + 0.005144020076841116, + 0.015864960849285126, + 0.005483725108206272, + 0.004577930085361004, + 0.006167463026940823, + 0.007158470340073109, + 0.0031569795683026314, + 0.00622715475037694, + 0.0027799447998404503, + 0.0058462913148105145, + 0.003713617566972971, + 0.009159301407635212, + 0.004454801324754953, + 0.014193090610206127, + 0.005314087960869074, + 0.004734561778604984, + 0.004372149705886841, + 0.008560110814869404, + 0.004492605570703745, + 0.02030300162732601, + 0.005519377999007702, + 0.017046745866537094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.00363522837869823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799941927194595 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799941927194595, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799941927194595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5944775342941284, + 0.03221515566110611, + 0.03847924992442131, + 0.04876791313290596, + 0.04806002601981163, + 0.09536747634410858, + 0.03682072460651398, + 0.021415799856185913, + 0.020438777282834053, + 0.0639573410153389 + ], + "credit_map": [ + 0.0396263562142849, + 0.028297914192080498, + 0.23132629692554474, + 0.03147336468100548, + 0.03438803553581238, + 0.03520733118057251, + 0.03343995288014412, + 0.16071829199790955, + 0.0024840182159096003, + 0.0053732553496956825, + 0.004073127638548613, + 0.003739274339750409, + 0.0029540059622377157, + 0.006386147812008858, + 0.00278544332832098, + 0.004419883247464895, + 0.003274948336184025, + 0.004820412956178188, + 0.005267974454909563, + 0.007522993255406618, + 0.0032567796297371387, + 0.004316289909183979, + 0.004877669271081686, + 0.005142183043062687, + 0.0031793080270290375, + 0.006004403345286846, + 0.0038377544842660427, + 0.013038239441812038, + 0.0051568797789514065, + 0.006819532252848148, + 0.006599183194339275, + 0.0041326116770505905, + 0.0022934512235224247, + 0.002411671681329608, + 0.0030616235453635454, + 0.005175195634365082, + 0.002026369795203209, + 0.018486876040697098, + 0.004083714913576841, + 0.010521123185753822, + 0.013418695889413357, + 0.008364719338715076, + 0.009559805504977703, + 0.014748909510672092, + 0.010196516290307045, + 0.006609776522964239, + 0.025442712008953094, + 0.007026343606412411, + 0.002760327886790037, + 0.0037698023952543736, + 0.002870851894840598, + 0.005552607588469982, + 0.0038018380291759968, + 0.006761166267096996, + 0.00881126057356596, + 0.002492869272828102, + 0.0028604737017303705, + 0.0022236674558371305, + 0.0025076530873775482, + 0.0033564104232937098, + 0.00280751776881516, + 0.0017247244250029325, + 0.0025189081206917763, + 0.003416444407775998, + 0.002043388783931732, + 0.004034820012748241, + 0.0016352521488443017, + 0.0017154088709503412, + 0.0028887917287647724, + 0.001893488341011107, + 0.0028360113501548767, + 0.0033916165120899677, + 0.006267987657338381, + 0.0048812865279614925, + 0.007477869745343924, + 0.007077693939208984, + 0.004658951424062252, + 0.011582205072045326, + 0.004642222076654434, + 0.017369121313095093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.029799941927194595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002126425504684 + ], + "opposite_logprob_tokens": [ + -2.695021152496338 + ], + "logprob_reward": -0.07002126425504684, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002126425504684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6918493509292603, + 0.012596706859767437, + 0.0510970763862133, + 0.06177455931901932, + 0.04612673446536064, + 0.036591943353414536, + 0.033695973455905914, + 0.018583470955491066, + 0.012630441226065159, + 0.03505367413163185 + ], + "credit_map": [ + 0.046868979930877686, + 0.03684049844741821, + 0.2756304144859314, + 0.034952834248542786, + 0.04423965513706207, + 0.029094956815242767, + 0.037455346435308456, + 0.18676672875881195, + 0.0013636508956551552, + 0.0020218342542648315, + 0.0015522835310548544, + 0.0012561018811538815, + 0.0022325576283037663, + 0.001626363373361528, + 0.0010678305989131331, + 0.0014760864432901144, + 0.004573482554405928, + 0.008385698311030865, + 0.005238617770373821, + 0.010001706890761852, + 0.005329030100256205, + 0.005112485494464636, + 0.005208816844969988, + 0.0072472477331757545, + 0.004066468682140112, + 0.00854782946407795, + 0.0052773295901715755, + 0.012045339681208134, + 0.009292124770581722, + 0.008024957031011581, + 0.008532780222594738, + 0.00598773593083024, + 0.0025614274200052023, + 0.003245169762521982, + 0.003010636894032359, + 0.006294945254921913, + 0.002912804950028658, + 0.012075038626790047, + 0.009062675759196281, + 0.00696404417976737, + 0.005604621022939682, + 0.003235341515392065, + 0.0026440094225108624, + 0.0037344107404351234, + 0.003845788072794676, + 0.0022774848621338606, + 0.011325454339385033, + 0.0039248354732990265, + 0.003328948048874736, + 0.002637370489537716, + 0.003510261420160532, + 0.005207345820963383, + 0.002517328830435872, + 0.0050561800599098206, + 0.008105902001261711, + 0.003332641441375017, + 0.00162196543533355, + 0.0035110258031636477, + 0.001784103107638657, + 0.0034241117537021637, + 0.0019654720090329647, + 0.0016365429619327188, + 0.001899891416542232, + 0.0027403614949434996, + 0.0011931557673960924, + 0.002024797722697258, + 0.0011546504683792591, + 0.0017578485421836376, + 0.0010768171632662416, + 0.0010373207041993737, + 0.0024817262310534716, + 0.001904125907458365, + 0.006172532215714455, + 0.003322630189359188, + 0.004583083558827639, + 0.00620224280282855, + 0.0026713558472692966, + 0.003441018983721733, + 0.0033307645935565233, + 0.005330047104507685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07002126425504684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036153923720121384 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036153923720121384, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036153923720121384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.364695280790329, + 0.038116250187158585, + 0.03874214366078377, + 0.07348441332578659, + 0.11643032729625702, + 0.1058315709233284, + 0.13266512751579285, + 0.029772162437438965, + 0.04683353379368782, + 0.05342917516827583 + ], + "credit_map": [ + 0.023369494825601578, + 0.016895409673452377, + 0.11691370606422424, + 0.019313616678118706, + 0.022891871631145477, + 0.04333040118217468, + 0.02416391298174858, + 0.09781686216592789, + 0.003929597791284323, + 0.004758598282933235, + 0.004582616500556469, + 0.003908637911081314, + 0.003942479379475117, + 0.006157005671411753, + 0.004201938863843679, + 0.00663537485525012, + 0.004986959509551525, + 0.006550137884914875, + 0.004468545317649841, + 0.0053106024861335754, + 0.004546154756098986, + 0.004879683721810579, + 0.00435565784573555, + 0.0036443984135985374, + 0.00470555666834116, + 0.014427192509174347, + 0.008142081089317799, + 0.012944278307259083, + 0.009109466336667538, + 0.009237315505743027, + 0.008444138802587986, + 0.006474383175373077, + 0.00545564666390419, + 0.006770608015358448, + 0.008241478353738785, + 0.018189193680882454, + 0.0047921990044415, + 0.03794309124350548, + 0.014536897651851177, + 0.020501211285591125, + 0.013965875841677189, + 0.008964153937995434, + 0.00976502988487482, + 0.016447413712739944, + 0.011036278679966927, + 0.005877132993191481, + 0.030455457046628, + 0.00932022463530302, + 0.009548047557473183, + 0.013145514763891697, + 0.005959102418273687, + 0.009979374706745148, + 0.006942791398614645, + 0.033619556576013565, + 0.04417101666331291, + 0.009299722500145435, + 0.004128797445446253, + 0.0034015330020338297, + 0.004484895616769791, + 0.005237134639173746, + 0.002538941102102399, + 0.0027338105719536543, + 0.0036469558253884315, + 0.003600094001740217, + 0.005395127926021814, + 0.00974535383284092, + 0.0046097151935100555, + 0.005359937436878681, + 0.005296044517308474, + 0.005209883209317923, + 0.004224064294248819, + 0.00699340645223856, + 0.006068374030292034, + 0.0042755804024636745, + 0.0040970477275550365, + 0.00820560846477747, + 0.004223003517836332, + 0.013634146191179752, + 0.005966786295175552, + 0.006958625745028257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0036153923720121384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle close-up captures a massive white tiger at the very center of the frame, its massive bulk practically filling the entire frame. The beast's thick, snow-white fur is marked with distinct inky black stripes, and its icy blue eyes gleam with a fierce glint, fixed intently on the camera. Its maw, gaping, reveals its dagger-sharp canines, its barbed pink tongue curling with intensity as it unleashes a heartfelt, ear-piercing roar. Its bristly whiskers tremble with fury as it roars, its powerful sonic waves seemingly shaking the very air. The hot breath from its throat condenses into a white mist in the cool air. The background is the gloomy jungle. A few faint rays of light penetrate the dense canopy, sculpting the tiger's muscular shoulders and reflecting off its wet teeth. The entire picture is full of primitive power. The low-angle upward perspective greatly enhances its size and majesty, creating a suffocating, wild and threatening sense of oppression.", + "seed": 27065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_0/index_6348_seed_27065_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01433883998543024, + "vlm_logprob": -0.01433883998543024, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.01433883998543024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013882741332054138 + ], + "opposite_logprob_tokens": [ + -4.388882637023926 + ], + "logprob_reward": -0.013882741332054138, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013882741332054138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03997530788183212, + 0.4162783622741699, + 0.027943948283791542, + 0.14614024758338928, + 0.10019783675670624, + 0.039364900439977646, + 0.08291514962911606, + 0.04419386014342308, + 0.03937031328678131, + 0.06362005323171616 + ], + "credit_map": [ + 0.005656897556036711, + 0.0030220355838537216, + 0.005955834873020649, + 0.005763287656009197, + 0.005234037525951862, + 0.0040887086652219296, + 0.005808289162814617, + 0.0044462173245847225, + 0.14959944784641266, + 0.016135601326823235, + 0.18222376704216003, + 0.01805971935391426, + 0.014499042183160782, + 0.008191092871129513, + 0.018661797046661377, + 0.008907892741262913, + 0.006095305550843477, + 0.0023150069173425436, + 0.0073958877474069595, + 0.0027863180730491877, + 0.002736420603469014, + 0.001501200720667839, + 0.0021319231018424034, + 0.0029818841721862555, + 0.01839848980307579, + 0.011607103049755096, + 0.011173071339726448, + 0.01061027217656374, + 0.030498802661895752, + 0.017963577061891556, + 0.01930612325668335, + 0.026582809165120125, + 0.014280990697443485, + 0.015827827155590057, + 0.006783360615372658, + 0.009657762013375759, + 0.014426939189434052, + 0.011093758046627045, + 0.016980400308966637, + 0.011146795004606247, + 0.005139342043548822, + 0.0028013314586132765, + 0.004680554382503033, + 0.007395181804895401, + 0.005493899341672659, + 0.003437008475884795, + 0.003795202588662505, + 0.006622378248721361, + 0.010050341486930847, + 0.009448598138988018, + 0.006405852735042572, + 0.008718647994101048, + 0.007553279399871826, + 0.005672524217516184, + 0.015574091114103794, + 0.019491812214255333, + 0.006844945251941681, + 0.0034294507931917906, + 0.003663581097498536, + 0.004268584307283163, + 0.004099521320313215, + 0.0071465494111180305, + 0.005937672220170498, + 0.008803553879261017, + 0.0050046406686306, + 0.0032494228798896074, + 0.003705088747665286, + 0.004518784582614899, + 0.003949485719203949, + 0.006350033916532993, + 0.007351638749241829, + 0.00524121941998601, + 0.007493970915675163, + 0.006434655282646418, + 0.016219761222600937, + 0.005344944074749947, + 0.00474913464859128, + 0.005266205873340368, + 0.006173600442707539, + 0.011937781237065792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.013882741332054138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022110319696366787 + ], + "opposite_logprob_tokens": [ + -6.127211093902588 + ], + "logprob_reward": -0.0022110319696366787, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022110319696366787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07583030313253403, + 0.3294413089752197, + 0.039245203137397766, + 0.047380972653627396, + 0.1265493929386139, + 0.08695852011442184, + 0.08417876064777374, + 0.04274507611989975, + 0.052984368056058884, + 0.11468609422445297 + ], + "credit_map": [ + 0.009709823876619339, + 0.006771634332835674, + 0.008555128239095211, + 0.010349800810217857, + 0.011911624111235142, + 0.009001372382044792, + 0.012576167471706867, + 0.006954751908779144, + 0.10599247366189957, + 0.015992697328329086, + 0.12464208900928497, + 0.015868376940488815, + 0.024283859878778458, + 0.009997624903917313, + 0.02064611203968525, + 0.012018079869449139, + 0.0061845798045396805, + 0.003911556210368872, + 0.006934236269444227, + 0.004682765807956457, + 0.004825221840292215, + 0.0030323811806738377, + 0.0051869857124984264, + 0.004487473051995039, + 0.00692580034956336, + 0.0033482371363788843, + 0.005196816753596067, + 0.0030186332296580076, + 0.0053002191707491875, + 0.007715299259871244, + 0.006966771092265844, + 0.00890919379889965, + 0.018628720194101334, + 0.01720670610666275, + 0.008903922513127327, + 0.011357424780726433, + 0.015467365272343159, + 0.019675496965646744, + 0.02074551023542881, + 0.014564251527190208, + 0.007954943925142288, + 0.007532264105975628, + 0.006889208219945431, + 0.013784515671432018, + 0.019599203020334244, + 0.007943220436573029, + 0.015824636444449425, + 0.007430530618876219, + 0.012551995925605297, + 0.012669021263718605, + 0.00814841315150261, + 0.0072097270749509335, + 0.009890957735478878, + 0.009515580721199512, + 0.010102304629981518, + 0.014090762473642826, + 0.006515213754028082, + 0.003597825765609741, + 0.0041802022606134415, + 0.0028008241206407547, + 0.004427176434546709, + 0.01004031952470541, + 0.0063492776826024055, + 0.0048342361114919186, + 0.006457437761127949, + 0.004972023889422417, + 0.00588575704023242, + 0.0071143279783427715, + 0.004748935345560312, + 0.00689432630315423, + 0.01099067460745573, + 0.005920880939811468, + 0.010866578668355942, + 0.006766051519662142, + 0.01804615743458271, + 0.009423366747796535, + 0.012747938744723797, + 0.013451291248202324, + 0.024230122566223145, + 0.019154587760567665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.0022110319696366787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186016008257866 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186016008257866, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186016008257866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295756846666336, + 0.1389080137014389, + 0.08113835752010345, + 0.09567389637231827, + 0.0692545473575592, + 0.08651482313871384, + 0.14379405975341797, + 0.1558973640203476, + 0.03947439417243004, + 0.059768885374069214 + ], + "credit_map": [ + 0.014737261459231377, + 0.009845954366028309, + 0.01411628071218729, + 0.018337741494178772, + 0.025436773896217346, + 0.012792904861271381, + 0.02188749611377716, + 0.012421268038451672, + 0.03313186764717102, + 0.00871008075773716, + 0.03958664834499359, + 0.005446397699415684, + 0.027226917445659637, + 0.00951340887695551, + 0.010757162235677242, + 0.00453551858663559, + 0.012524873949587345, + 0.007799011655151844, + 0.013381685130298138, + 0.008470443077385426, + 0.010546494275331497, + 0.008800504729151726, + 0.009493540972471237, + 0.010121798142790794, + 0.01270708255469799, + 0.007684749085456133, + 0.011121180839836597, + 0.006554702762514353, + 0.019034113734960556, + 0.009808389469981194, + 0.013067827560007572, + 0.015695845708251, + 0.011652417480945587, + 0.008406049571931362, + 0.004485523793846369, + 0.006676395423710346, + 0.010858521796762943, + 0.009784188121557236, + 0.008122490718960762, + 0.009268959052860737, + 0.012425935827195644, + 0.005698777735233307, + 0.006469263229519129, + 0.02526615932583809, + 0.009224570356309414, + 0.007369435392320156, + 0.008595016784965992, + 0.011465659365057945, + 0.014691849239170551, + 0.013936517760157585, + 0.011168044060468674, + 0.021751852706074715, + 0.016541248187422752, + 0.013075953349471092, + 0.03608439490199089, + 0.01654420606791973, + 0.025298958644270897, + 0.010947984643280506, + 0.017092783004045486, + 0.00807148590683937, + 0.011285376735031605, + 0.04294233024120331, + 0.0159186739474535, + 0.024339761584997177, + 0.0075961933471262455, + 0.003669116413220763, + 0.0040877945721149445, + 0.0037441460881382227, + 0.002535160630941391, + 0.00590889435261488, + 0.004235762171447277, + 0.007697327062487602, + 0.009503786452114582, + 0.002772616920992732, + 0.018318716436624527, + 0.004450542852282524, + 0.003512026509270072, + 0.005843072663992643, + 0.005937612149864435, + 0.009430509060621262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.018186016008257866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791494831442833 + ], + "opposite_logprob_tokens": [ + -3.5297915935516357 + ], + "logprob_reward": -0.029791494831442833, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791494831442833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0984749048948288, + 0.12736934423446655, + 0.06562040001153946, + 0.11046046763658524, + 0.25597622990608215, + 0.03355953097343445, + 0.067802295088768, + 0.16043929755687714, + 0.04864661768078804, + 0.031650833785533905 + ], + "credit_map": [ + 0.016286121681332588, + 0.008175606839358807, + 0.010571561753749847, + 0.012178253382444382, + 0.01644822210073471, + 0.008352985605597496, + 0.01396245788782835, + 0.01249971054494381, + 0.03788047656416893, + 0.008038025349378586, + 0.04595818370580673, + 0.005982105154544115, + 0.01088753528892994, + 0.006609627045691013, + 0.008118513971567154, + 0.0038948857691138983, + 0.01237871591001749, + 0.0071787843480706215, + 0.012638780288398266, + 0.005752588622272015, + 0.007581005804240704, + 0.007473616395145655, + 0.007958182133734226, + 0.004658733028918505, + 0.016888968646526337, + 0.00850161723792553, + 0.01978323981165886, + 0.009249444119632244, + 0.01973816379904747, + 0.009078946895897388, + 0.013651411049067974, + 0.013568690977990627, + 0.03376322612166405, + 0.03312333673238754, + 0.02912260964512825, + 0.02277793176472187, + 0.031596824526786804, + 0.04360967501997948, + 0.03469180315732956, + 0.02729085087776184, + 0.004629870411008596, + 0.0029061248060315847, + 0.004433120600879192, + 0.005632555577903986, + 0.004366906359791756, + 0.0034812833182513714, + 0.003828080603852868, + 0.0042815906926989555, + 0.014354584738612175, + 0.008570846170186996, + 0.006981440354138613, + 0.008962281979620457, + 0.0055353413335978985, + 0.006417963653802872, + 0.008216055110096931, + 0.008763785474002361, + 0.02602914720773697, + 0.010562308132648468, + 0.016332313418388367, + 0.010794481262564659, + 0.014578082598745823, + 0.046712059527635574, + 0.013760324567556381, + 0.021670594811439514, + 0.00549030676484108, + 0.005026602651923895, + 0.005034012719988823, + 0.006277824752032757, + 0.005910577718168497, + 0.005510163493454456, + 0.010544405318796635, + 0.004852727521210909, + 0.003951431717723608, + 0.001053366344422102, + 0.01101727969944477, + 0.0030962303280830383, + 0.0014842305099591613, + 0.003017129609361291, + 0.0022731400094926357, + 0.005758027546107769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.029791494831442833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger unleashing a roar with visible hot breath condensing into white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gloomy jungle with faint rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle close-up where the tiger fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07407667487859726, + 0.2235214114189148, + 0.10698353499174118, + 0.06120523065328598, + 0.1027490645647049, + 0.13432817161083221, + 0.07821749150753021, + 0.08220012485980988, + 0.08895658701658249, + 0.04776164889335632 + ], + "credit_map": [ + 0.012328458949923515, + 0.0065340809524059296, + 0.007871911861002445, + 0.008389286696910858, + 0.012155110016465187, + 0.00910815130919218, + 0.010690066032111645, + 0.006999615579843521, + 0.07747967541217804, + 0.008470993489027023, + 0.09449216723442078, + 0.009629392065107822, + 0.009884803555905819, + 0.007232159376144409, + 0.010236862115561962, + 0.006095359567552805, + 0.02350691705942154, + 0.011967618949711323, + 0.025512557476758957, + 0.009502887725830078, + 0.009585300460457802, + 0.008684945292770863, + 0.007929460145533085, + 0.010293853469192982, + 0.010675104334950447, + 0.00353789608925581, + 0.013131024315953255, + 0.00552179804071784, + 0.0053261700086295605, + 0.005585503298789263, + 0.007277901284396648, + 0.010149842128157616, + 0.012081106193363667, + 0.015947844833135605, + 0.008001831360161304, + 0.011024066247045994, + 0.024409478530287743, + 0.008563176728785038, + 0.015858886763453484, + 0.0068626729771494865, + 0.03475211188197136, + 0.006193677894771099, + 0.019049694761633873, + 0.012067207135260105, + 0.00838577002286911, + 0.013552116230130196, + 0.008802701719105244, + 0.0315248966217041, + 0.008693769574165344, + 0.007078650873154402, + 0.006572788115590811, + 0.009260873310267925, + 0.010181591846048832, + 0.005712179001420736, + 0.01935761421918869, + 0.011360029689967632, + 0.01171684730798006, + 0.006557037588208914, + 0.008269602432847023, + 0.01216322835534811, + 0.007141562178730965, + 0.01622793637216091, + 0.008283982053399086, + 0.011839937418699265, + 0.019592223688960075, + 0.007636305410414934, + 0.008433355949819088, + 0.006785384379327297, + 0.005856080912053585, + 0.01760840229690075, + 0.012554996646940708, + 0.010489841923117638, + 0.006310418713837862, + 0.001980955945327878, + 0.01723914034664631, + 0.0053335390985012054, + 0.0021559197921305895, + 0.003928505349904299, + 0.0026083774864673615, + 0.008204796351492405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.007622915785759687 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_tensor([8012, 8012])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_tensor([8012, 8012])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..94587c5926262d491d63f04c0d82dc5d893e0c09 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_tensor([8012, 8012])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide, long shot captured from underwater, the camera slowly pans to the right, following a massive humpback whale as it cruises through the boundless deep blue ocean. Brilliant beams of sunlight penetrate the shimmering surface, piercing the clear waters vertically, casting a shifting, dancing pattern of light and shadow across the whale's deep, blue-black skin, which flows with its movements. Its massive, streamlined form, its skin covered in white scratches and barnacles from age, dominates the frame, its massive pectoral fins gracefully extended like wings. With an almost meditative composure, it propels its body slowly and steadily forward through the powerful, steady up-and-down movements of its enormous tail fin, surrounded by countless tiny plankton shimmering in the beams of light. The dominant colors of the scene are deep ocean blue and brilliant sunlight white, interweaving light and shadow to create a majestic, tranquil, and vibrant atmosphere filled with a profound awe for the wonders of nature.", + "seed": 18045, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_8012_seed_18045_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11073704259470105, + "vlm_logprob": -0.11073704259470105, + "vlm_yes_no_margin": 3.724999988079071, + "vlm_token_credit": -0.11073704259470105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047798180021345615 + ], + "opposite_logprob_tokens": [ + -8.879779815673828 + ], + "logprob_reward": -0.0047798180021345615, + "yes_no_margin_reward": 8.875, + "token_credit_reward": -0.0047798180021345615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4715518057346344, + 0.04787486791610718, + 0.030550828203558922, + 0.25083449482917786, + 0.02309349551796913, + 0.02075016498565674, + 0.03760868310928345, + 0.04363523796200752, + 0.05216442048549652, + 0.021935993805527687 + ], + "credit_map": [ + 0.02549321949481964, + 0.007319066673517227, + 0.1918240785598755, + 0.02057577110826969, + 0.015624686144292355, + 0.1941169649362564, + 0.008694237098097801, + 0.007903833873569965, + 0.004420441575348377, + 0.004183182027190924, + 0.00763032678514719, + 0.007195594720542431, + 0.004426421597599983, + 0.00797920674085617, + 0.00517750158905983, + 0.006862199399620295, + 0.003171095158904791, + 0.004009985364973545, + 0.004596532788127661, + 0.0030299450736492872, + 0.0033641455229371786, + 0.003737921826541424, + 0.004325804300606251, + 0.004315401893109083, + 0.03819422796368599, + 0.014001986011862755, + 0.037980448454618454, + 0.01706675812602043, + 0.025413859635591507, + 0.07510100305080414, + 0.019983116537332535, + 0.023093128576874733, + 0.003128527197986841, + 0.00268657342530787, + 0.003017362207174301, + 0.003493174444884062, + 0.002311590826138854, + 0.0032675359398126602, + 0.0030283459927886724, + 0.002160388045012951, + 0.0014798369957134128, + 0.0019304605666548014, + 0.0029575242660939693, + 0.0022812949027866125, + 0.0028996372129768133, + 0.003251091344282031, + 0.004122171085327864, + 0.00182815152220428, + 0.003752583870664239, + 0.006670302711427212, + 0.00472193444147706, + 0.005105664953589439, + 0.0030236735474318266, + 0.006617873907089233, + 0.004591133911162615, + 0.0031255187932401896, + 0.004316724371165037, + 0.003649110672995448, + 0.010354340076446533, + 0.0038876733742654324, + 0.004136779811233282, + 0.007798770908266306, + 0.0056771268136799335, + 0.0038147158920764923, + 0.0058068279176950455, + 0.004222279880195856, + 0.007253831718116999, + 0.0048194206319749355, + 0.004538130480796099, + 0.0027716190088540316, + 0.015189000405371189, + 0.007563313469290733, + 0.003041122807189822, + 0.002197569701820612, + 0.003041531890630722, + 0.0018737553618848324, + 0.0015377046074718237, + 0.002068399218842387, + 0.003852298716083169, + 0.004323611967265606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0047798180021345615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305003955960274 + ], + "opposite_logprob_tokens": [ + -3.1680500507354736 + ], + "logprob_reward": -0.04305003955960274, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305003955960274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5931126475334167, + 0.03799198567867279, + 0.10007444024085999, + 0.13463279604911804, + 0.007448465563356876, + 0.03537296503782272, + 0.027462195605039597, + 0.03720696270465851, + 0.01455000787973404, + 0.01214757189154625 + ], + "credit_map": [ + 0.036038678139448166, + 0.01702224835753441, + 0.22317716479301453, + 0.025010868906974792, + 0.021699275821447372, + 0.23540951311588287, + 0.01805495098233223, + 0.01669996604323387, + 0.004407091066241264, + 0.005111590959131718, + 0.007354652509093285, + 0.004208476282656193, + 0.00339533481746912, + 0.005023925565183163, + 0.004037053324282169, + 0.004453861620277166, + 0.010630633682012558, + 0.012808729894459248, + 0.016168968752026558, + 0.008316874504089355, + 0.01440662331879139, + 0.011349454522132874, + 0.015004596672952175, + 0.011388557031750679, + 0.015844019129872322, + 0.008771160617470741, + 0.018185963854193687, + 0.011860989034175873, + 0.013857755810022354, + 0.03462342545390129, + 0.01920895092189312, + 0.012280529364943504, + 0.0006650344002991915, + 0.0009098306763917208, + 0.0010464346269145608, + 0.0012443921295925975, + 0.0007317609852179885, + 0.001195605262182653, + 0.0009902181336656213, + 0.000665189407300204, + 0.002910533919930458, + 0.0048236846923828125, + 0.004775584675371647, + 0.003626680001616478, + 0.004756579641252756, + 0.004505931865423918, + 0.005567177198827267, + 0.004406794905662537, + 0.0022928384132683277, + 0.004026470240205526, + 0.0034321865532547235, + 0.0041292933747172356, + 0.002580223837867379, + 0.0044713993556797504, + 0.0031353419180959463, + 0.0033944419119507074, + 0.0035712234675884247, + 0.004282807931303978, + 0.005538454279303551, + 0.004014503210783005, + 0.0037687821313738823, + 0.006780174095183611, + 0.005245171952992678, + 0.004005847964435816, + 0.0015413863584399223, + 0.0009503455366939306, + 0.0020233234390616417, + 0.0011475869687274098, + 0.0010508960112929344, + 0.0007745473994873464, + 0.004657291807234287, + 0.0024046299513429403, + 0.0010081077925860882, + 0.0011208599898964167, + 0.0018503139726817608, + 0.0017216315027326345, + 0.0012907575583085418, + 0.001294872141443193, + 0.002030574483796954, + 0.0018304541008546948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04305003955960274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903608471155167 + ], + "opposite_logprob_tokens": [ + -2.464035987854004 + ], + "logprob_reward": -0.08903608471155167, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903608471155167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2319161295890808, + 0.10370388627052307, + 0.19074566662311554, + 0.2033659666776657, + 0.02177092246711254, + 0.0552961602807045, + 0.06752704828977585, + 0.0643291249871254, + 0.029273921623826027, + 0.03207121044397354 + ], + "credit_map": [ + 0.027694085612893105, + 0.02388388104736805, + 0.04768489673733711, + 0.01931612193584442, + 0.019579708576202393, + 0.04371314123272896, + 0.024742448702454567, + 0.025301821529865265, + 0.011504057794809341, + 0.012545234523713589, + 0.018312133848667145, + 0.019022202119231224, + 0.009397149085998535, + 0.012114739045500755, + 0.01220855861902237, + 0.008599799126386642, + 0.014948342926800251, + 0.01977144367992878, + 0.029137590900063515, + 0.01631096936762333, + 0.028461506590247154, + 0.018622465431690216, + 0.04018953815102577, + 0.02330380491912365, + 0.01779896952211857, + 0.015657369047403336, + 0.026232054457068443, + 0.019647741690278053, + 0.023360056802630424, + 0.04103180766105652, + 0.03918491303920746, + 0.020453039556741714, + 0.002658264245837927, + 0.0030735416803508997, + 0.0027570303063839674, + 0.0025874662678688765, + 0.0029468510765582323, + 0.002978760516270995, + 0.0030509743373841047, + 0.0017180325230583549, + 0.004431703127920628, + 0.007634514942765236, + 0.007673845160752535, + 0.005936882458627224, + 0.008011627942323685, + 0.0077852713875472546, + 0.007867208682000637, + 0.005955105647444725, + 0.0058176578022539616, + 0.00886470079421997, + 0.008670574985444546, + 0.010301201604306698, + 0.006782424636185169, + 0.010954254306852818, + 0.0075604915618896484, + 0.008575741201639175, + 0.007436950691044331, + 0.006481001619249582, + 0.011649491265416145, + 0.006660008803009987, + 0.005597957409918308, + 0.010377197526395321, + 0.008102623745799065, + 0.008023891597986221, + 0.003557701362296939, + 0.0018381825648248196, + 0.003923151176422834, + 0.0020964150317013264, + 0.0023565927986055613, + 0.0019778995774686337, + 0.008918396197259426, + 0.004605581518262625, + 0.0031296524684876204, + 0.003552016569301486, + 0.0036543284077197313, + 0.0038154602516442537, + 0.0036208126693964005, + 0.003033312037587166, + 0.006143638398498297, + 0.005121988710016012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.08903608471155167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029788371175527573 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.029788371175527573, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029788371175527573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4459324777126312, + 0.10080506652593613, + 0.13420552015304565, + 0.1447823941707611, + 0.04303421825170517, + 0.01929195038974285, + 0.02274204045534134, + 0.04303494468331337, + 0.0255834199488163, + 0.02058795467019081 + ], + "credit_map": [ + 0.023824576288461685, + 0.009246774017810822, + 0.18041221797466278, + 0.018854916095733643, + 0.014893397688865662, + 0.18180976808071136, + 0.008410069160163403, + 0.008480774238705635, + 0.014601587317883968, + 0.00800696574151516, + 0.019829144701361656, + 0.01247015967965126, + 0.007953613065183163, + 0.014898204244673252, + 0.012582957744598389, + 0.010462443344295025, + 0.01890668272972107, + 0.013496225699782372, + 0.020538121461868286, + 0.017866605892777443, + 0.012294621206820011, + 0.014545587822794914, + 0.02284220978617668, + 0.01371548231691122, + 0.022029347717761993, + 0.00806210096925497, + 0.02231677807867527, + 0.009936400689184666, + 0.014562515541911125, + 0.04311611130833626, + 0.011272657662630081, + 0.01348649151623249, + 0.006861716043204069, + 0.004589333198964596, + 0.007108665537089109, + 0.007466898765414953, + 0.0031578755006194115, + 0.005556502845138311, + 0.006137536838650703, + 0.0021556923165917397, + 0.002138426061719656, + 0.002272179117426276, + 0.0027027626056224108, + 0.002038351260125637, + 0.002202858217060566, + 0.0035496780183166265, + 0.002454426372423768, + 0.001933269202709198, + 0.0024975049309432507, + 0.0028125266544520855, + 0.0031268696766346693, + 0.0030242749489843845, + 0.0023889599833637476, + 0.0035645158495754004, + 0.0025253407657146454, + 0.0028020485769957304, + 0.009508440271019936, + 0.006292195525020361, + 0.006178300362080336, + 0.004293246194720268, + 0.0037542160134762526, + 0.00408184016123414, + 0.005030556116253138, + 0.003896150039508939, + 0.0031882303301244974, + 0.0018029866041615605, + 0.003323625773191452, + 0.0022379017900675535, + 0.0017776025924831629, + 0.0012286928249523044, + 0.007931484840810299, + 0.004092896357178688, + 0.0038966790307313204, + 0.0021441825665533543, + 0.0020607607439160347, + 0.0024606746155768633, + 0.0021982004400342703, + 0.0017979575786739588, + 0.002864735433831811, + 0.003164765192195773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.029788371175527573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870308995246887 + ], + "opposite_logprob_tokens": [ + -1.137030839920044 + ], + "logprob_reward": -0.3870308995246887, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3870308995246887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5013004541397095, + 0.07113572955131531, + 0.03329025208950043, + 0.052737895399332047, + 0.03978757932782173, + 0.03995900601148605, + 0.13272477686405182, + 0.06305792182683945, + 0.043950799852609634, + 0.02205556444823742 + ], + "credit_map": [ + 0.02847137488424778, + 0.012802543118596077, + 0.19520103931427002, + 0.022098086774349213, + 0.017812680453062057, + 0.2026461511850357, + 0.011287006549537182, + 0.010981596074998379, + 0.005855991970747709, + 0.008264308795332909, + 0.012563290074467659, + 0.014277761802077293, + 0.006613024044781923, + 0.008990992791950703, + 0.007800804451107979, + 0.006769557483494282, + 0.0026465391274541616, + 0.0029895210172981024, + 0.0039937603287398815, + 0.0031475855503231287, + 0.004743637982755899, + 0.0040451581589877605, + 0.006720649544149637, + 0.005003401543945074, + 0.005206192843616009, + 0.0038651323411613703, + 0.00771310506388545, + 0.006594834849238396, + 0.004309794865548611, + 0.009753422811627388, + 0.00982620008289814, + 0.005469218827784061, + 0.004143387544900179, + 0.005836021155118942, + 0.005033647641539574, + 0.004816199187189341, + 0.006209918297827244, + 0.005369415041059256, + 0.005312548950314522, + 0.003066444769501686, + 0.004003352019935846, + 0.006281749811023474, + 0.005580583121627569, + 0.005022223573178053, + 0.004239510744810104, + 0.0052677639760077, + 0.006476968061178923, + 0.0030868591275066137, + 0.011230065487325191, + 0.018638841807842255, + 0.016580788418650627, + 0.02050013095140457, + 0.013113709166646004, + 0.020937900990247726, + 0.015217525884509087, + 0.016505829989910126, + 0.006085885688662529, + 0.005219544284045696, + 0.013427142053842545, + 0.006599808111786842, + 0.0048798308707773685, + 0.00774609437212348, + 0.011049195192754269, + 0.008050424978137016, + 0.005629157647490501, + 0.005143413785845041, + 0.006949351169168949, + 0.003792793722823262, + 0.0054364739917218685, + 0.004074994474649429, + 0.007952420972287655, + 0.004972196649760008, + 0.002656670054420829, + 0.0024381563998758793, + 0.0024172121193259954, + 0.0026168932672590017, + 0.0025094482116401196, + 0.0030342768877744675, + 0.0029894676990807056, + 0.0033934423699975014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.3870308995246887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot captured from underwater, the camera slowly pans to the right, following a massive humpback whale as it cruises through the boundless deep blue ocean. Brilliant beams of sunlight penetrate the shimmering surface, piercing the clear waters vertically, casting a shifting, dancing pattern of light and shadow across the whale's deep, blue-black skin, which flows with its movements. Its massive, streamlined form, its skin covered in white scratches and barnacles from age, dominates the frame, its massive pectoral fins gracefully extended like wings. With an almost meditative composure, it propels its body slowly and steadily forward through the powerful, steady up-and-down movements of its enormous tail fin, surrounded by countless tiny plankton shimmering in the beams of light. The dominant colors of the scene are deep ocean blue and brilliant sunlight white, interweaving light and shadow to create a majestic, tranquil, and vibrant atmosphere filled with a profound awe for the wonders of nature.", + "seed": 18046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_8012_seed_18046_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06604035596828908, + "vlm_logprob": -0.06604035596828908, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.06604035596828908, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0042588491924107075 + ], + "opposite_logprob_tokens": [ + -8.75425910949707 + ], + "logprob_reward": -0.0042588491924107075, + "yes_no_margin_reward": 8.75, + "token_credit_reward": -0.0042588491924107075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037686824798583984, + 0.039709221571683884, + 0.31710755825042725, + 0.18228337168693542, + 0.0287043284624815, + 0.01707085594534874, + 0.25823864340782166, + 0.04435301572084427, + 0.06501229107379913, + 0.009833937510848045 + ], + "credit_map": [ + 0.003663091454654932, + 0.0044426159001886845, + 0.008693903684616089, + 0.00472244992852211, + 0.005492320284247398, + 0.0032233265228569508, + 0.003679925110191107, + 0.003769185859709978, + 0.003609450301155448, + 0.0025330986827611923, + 0.007527639158070087, + 0.005153661593794823, + 0.004101741127669811, + 0.007917073555290699, + 0.004550921730697155, + 0.004315631464123726, + 0.030262000858783722, + 0.024766726419329643, + 0.040447935461997986, + 0.033530861139297485, + 0.08751005679368973, + 0.06398226320743561, + 0.018567528575658798, + 0.01804015040397644, + 0.006483020260930061, + 0.008174839429557323, + 0.016530822962522507, + 0.005642913281917572, + 0.06065769121050835, + 0.07205262035131454, + 0.0059717451222240925, + 0.006769705563783646, + 0.002962501021102071, + 0.003159688785672188, + 0.003299735952168703, + 0.004492020234465599, + 0.004317800980061293, + 0.005246988497674465, + 0.003031621454283595, + 0.0021939678117632866, + 0.0019372912356629968, + 0.0014786924002692103, + 0.00570596382021904, + 0.0027304766699671745, + 0.0010618583764880896, + 0.002081372542306781, + 0.0014086391311138868, + 0.0006665603723376989, + 0.016164269298315048, + 0.029039736837148666, + 0.07369524985551834, + 0.025082742795348167, + 0.03140101954340935, + 0.01701783388853073, + 0.05275353789329529, + 0.013084232807159424, + 0.004666246008127928, + 0.005286035127937794, + 0.00613604299724102, + 0.004549115896224976, + 0.006010620389133692, + 0.004778863862156868, + 0.005633397959172726, + 0.00729268928989768, + 0.007703904062509537, + 0.008444664068520069, + 0.006120798643678427, + 0.006958317011594772, + 0.009039308875799179, + 0.00944751501083374, + 0.007979470305144787, + 0.009318310767412186, + 0.0010395902208983898, + 0.001238176366314292, + 0.0013546817936003208, + 0.001479336991906166, + 0.0009857412660494447, + 0.001242033438757062, + 0.0012793518835678697, + 0.0012150243856012821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.0042588491924107075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019434866262599826 + ], + "opposite_logprob_tokens": [ + -6.251943588256836 + ], + "logprob_reward": -0.0019434866262599826, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019434866262599826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013504567556083202, + 0.4411485493183136, + 0.35516124963760376, + 0.04831044375896454, + 0.0053779361769557, + 0.08510103821754456, + 0.02092224359512329, + 0.01235555112361908, + 0.01239059865474701, + 0.005727821495383978 + ], + "credit_map": [ + 0.0009895332623273134, + 0.0018925078911706805, + 0.001992624718695879, + 0.002448281040415168, + 0.0011746503878384829, + 0.001186648034490645, + 0.0018229434499517083, + 0.0019973781891167164, + 0.03158695623278618, + 0.02612321823835373, + 0.11196238547563553, + 0.051710985600948334, + 0.07335471361875534, + 0.06303510814905167, + 0.04235850274562836, + 0.041016679257154465, + 0.03379589691758156, + 0.026846781373023987, + 0.04465087503194809, + 0.03752535954117775, + 0.09923363476991653, + 0.07035285234451294, + 0.02139960415661335, + 0.021356213837862015, + 0.0020679605659097433, + 0.0021507919300347567, + 0.004704414866864681, + 0.0017059838864952326, + 0.014448810368776321, + 0.019099390134215355, + 0.0020404041279107332, + 0.0020926829893141985, + 0.0005586090846918523, + 0.00043379361159168184, + 0.0006299432134255767, + 0.00043226045090705156, + 0.0005723116919398308, + 0.0017874578479677439, + 0.0004941179067827761, + 0.000469442515168339, + 0.007703833281993866, + 0.005289288703352213, + 0.03436768800020218, + 0.014242365024983883, + 0.0047744931653141975, + 0.009503856301307678, + 0.006042348220944405, + 0.003177160630002618, + 0.0021020774729549885, + 0.003425844246521592, + 0.0032740100286900997, + 0.0022699988912791014, + 0.003154656384140253, + 0.002570290816947818, + 0.0025067173410207033, + 0.001618648529984057, + 0.0012605132069438696, + 0.0012477763229981065, + 0.0013868360547348857, + 0.0015558864688500762, + 0.0015497152926400304, + 0.0016092853620648384, + 0.0021137467119842768, + 0.0016317914705723524, + 0.00166480359621346, + 0.0013800463639199734, + 0.001302581513300538, + 0.0013758650748059154, + 0.0015269371215254068, + 0.0015322589315474033, + 0.001826156280003488, + 0.001781949307769537, + 0.0006156910094432533, + 0.0005536692915484309, + 0.000846882292535156, + 0.0007900711498223245, + 0.0007880009361542761, + 0.0006201612413860857, + 0.000608998176176101, + 0.0009043474565260112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.0019434866262599826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048660341650247574 + ], + "opposite_logprob_tokens": [ + -3.0486602783203125 + ], + "logprob_reward": -0.048660341650247574, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048660341650247574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03987912833690643, + 0.3373146653175354, + 0.3513737916946411, + 0.06575579941272736, + 0.024154597893357277, + 0.03921773284673691, + 0.07966934889554977, + 0.02086777798831463, + 0.024372011423110962, + 0.017395146191120148 + ], + "credit_map": [ + 0.0023465962149202824, + 0.004361466038972139, + 0.005496067460626364, + 0.0063856737688183784, + 0.004667805973440409, + 0.0040287752635777, + 0.0063917613588273525, + 0.006200981792062521, + 0.024602822959423065, + 0.020391376689076424, + 0.08336259424686432, + 0.041057098656892776, + 0.05315032973885536, + 0.049514930695295334, + 0.03310766816139221, + 0.032127849757671356, + 0.03377115726470947, + 0.026743941009044647, + 0.044517360627651215, + 0.03635097295045853, + 0.09899365901947021, + 0.06953378766775131, + 0.02112962119281292, + 0.020333286374807358, + 0.0024966553319245577, + 0.0031158479396253824, + 0.005978683941066265, + 0.0026989313773810863, + 0.021107831969857216, + 0.025509260594844818, + 0.002263063797727227, + 0.002585525391623378, + 0.002764318836852908, + 0.0031039260793477297, + 0.0023470576852560043, + 0.002586479764431715, + 0.0038590615149587393, + 0.00526815839111805, + 0.002308773808181286, + 0.0019168206490576267, + 0.0027188605163246393, + 0.005645020864903927, + 0.013234524987637997, + 0.005107504781335592, + 0.00317456666380167, + 0.004135962575674057, + 0.0032402952201664448, + 0.0019609977025538683, + 0.004599511157721281, + 0.006150167901068926, + 0.027673913165926933, + 0.0065574501641094685, + 0.00675370404496789, + 0.0038645965978503227, + 0.020615076646208763, + 0.0034549401607364416, + 0.0020118956454098225, + 0.0021639838814735413, + 0.002693136688321829, + 0.0030278866179287434, + 0.0029229894280433655, + 0.0025998279452323914, + 0.0033580882009118795, + 0.0020899695809930563, + 0.002533844206482172, + 0.0025437772274017334, + 0.0027179743628948927, + 0.0025208096485584974, + 0.0029364959336817265, + 0.003500785445794463, + 0.003934723325073719, + 0.003683599643409252, + 0.0021969086956232786, + 0.0018228552071377635, + 0.0022029378451406956, + 0.0015716496855020523, + 0.0024908017367124557, + 0.00200101500377059, + 0.002175005851313472, + 0.002933972515165806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.048660341650247574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023276543244719505 + ], + "opposite_logprob_tokens": [ + -3.7732765674591064 + ], + "logprob_reward": -0.023276543244719505, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023276543244719505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029114408418536186, + 0.49935558438301086, + 0.367715448141098, + 0.016967536881566048, + 0.006034162826836109, + 0.062420547008514404, + 0.009717642329633236, + 0.0036790850572288036, + 0.0022450292017310858, + 0.002750536659732461 + ], + "credit_map": [ + 0.0028034434653818607, + 0.0027366529684513807, + 0.007181990426033735, + 0.0029204923193901777, + 0.003184047993272543, + 0.003934027627110481, + 0.0033832467161118984, + 0.0029705071356147528, + 0.03420436754822731, + 0.0292203426361084, + 0.12926362454891205, + 0.057125527411699295, + 0.08698175102472305, + 0.06930230557918549, + 0.04773619771003723, + 0.04552145302295685, + 0.03467661887407303, + 0.027722444385290146, + 0.049089062958955765, + 0.03934253752231598, + 0.09835566580295563, + 0.07517658919095993, + 0.022681953385472298, + 0.02067059651017189, + 0.0008143027662299573, + 0.000820945599116385, + 0.0016850446118041873, + 0.0008730197441764176, + 0.004641828592866659, + 0.00660206563770771, + 0.0007674796506762505, + 0.0007628494058735669, + 0.0005645618657581508, + 0.0004188928287476301, + 0.0006426406325772405, + 0.0005777854239568114, + 0.0008542978321202099, + 0.0022714477963745594, + 0.0003638432244770229, + 0.00034069319372065365, + 0.006275510415434837, + 0.003138327971100807, + 0.02628394588828087, + 0.011407076381146908, + 0.0020594033412635326, + 0.006774697918444872, + 0.004280420020222664, + 0.002201164374127984, + 0.0006562570342794061, + 0.0012548317899927497, + 0.0019180254312232137, + 0.001355065149255097, + 0.0014597319532185793, + 0.0009714916232042015, + 0.001222596620209515, + 0.0008796436595730484, + 0.00034709819010458887, + 0.0003396609681658447, + 0.00038782847695983946, + 0.00043352870852686465, + 0.0003886380873154849, + 0.0005523810395970941, + 0.000804636743851006, + 0.00042531316285021603, + 0.0002030308824032545, + 0.0002643899351824075, + 0.0003007234772667289, + 0.0002484731958247721, + 0.0002633010153658688, + 0.00026129590696655214, + 0.00023763491481076926, + 0.00046617997577413917, + 0.0003952792030759156, + 0.00024909331114031374, + 0.00042381559615023434, + 0.0003342149720992893, + 0.00029110393370501697, + 0.00022869334497954696, + 0.00043726610601879656, + 0.0003910701780114323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.023276543244719505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520625591278076 + ], + "opposite_logprob_tokens": [ + -1.5020625591278076 + ], + "logprob_reward": -0.2520625591278076, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520625591278076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11153359711170197, + 0.040825534611940384, + 0.05286477506160736, + 0.11227487772703171, + 0.04109577089548111, + 0.21405987441539764, + 0.21536685526371002, + 0.06695064902305603, + 0.1334615796804428, + 0.011566503904759884 + ], + "credit_map": [ + 0.007797603029757738, + 0.013076734729111195, + 0.01358601450920105, + 0.01838831976056099, + 0.009271515533328056, + 0.012756592594087124, + 0.020120974630117416, + 0.016535835340619087, + 0.0036428815219551325, + 0.00327116996049881, + 0.007335298229008913, + 0.0053219860419631, + 0.0063918656669557095, + 0.006082441192120314, + 0.004170451313257217, + 0.004609441384673119, + 0.004822293762117624, + 0.004847254138439894, + 0.005942013114690781, + 0.008568970486521721, + 0.011921774595975876, + 0.006547835189849138, + 0.00519458157941699, + 0.005020055454224348, + 0.004366112407296896, + 0.005709158722311258, + 0.010279390029609203, + 0.005028658080846071, + 0.03459911048412323, + 0.04087204858660698, + 0.006041599903255701, + 0.005378795322030783, + 0.0033275929745286703, + 0.004264564253389835, + 0.004518193192780018, + 0.006005230825394392, + 0.0059850141406059265, + 0.00913908425718546, + 0.003883892670273781, + 0.003972199279814959, + 0.024033773690462112, + 0.007637467235326767, + 0.08919750154018402, + 0.041444506496191025, + 0.007380419876426458, + 0.02243817038834095, + 0.014405389316380024, + 0.007522641681134701, + 0.012949937954545021, + 0.01540277898311615, + 0.07663062959909439, + 0.022351128980517387, + 0.014303576201200485, + 0.011008127592504025, + 0.05337974429130554, + 0.009340930730104446, + 0.007523253560066223, + 0.008923041634261608, + 0.007781180553138256, + 0.008386461064219475, + 0.007686859928071499, + 0.005179284140467644, + 0.009325843304395676, + 0.012144727632403374, + 0.018574999645352364, + 0.016828114166855812, + 0.014868931844830513, + 0.014104576781392097, + 0.01477181538939476, + 0.016925031319260597, + 0.016331402584910393, + 0.021056709811091423, + 0.000864262692630291, + 0.0009958944283425808, + 0.0018228127155452967, + 0.001952089718542993, + 0.0013171253958716989, + 0.001431295066140592, + 0.0015669864369556308, + 0.0016160377999767661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.2520625591278076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot captured from underwater, the camera slowly pans to the right, following a massive humpback whale as it cruises through the boundless deep blue ocean. Brilliant beams of sunlight penetrate the shimmering surface, piercing the clear waters vertically, casting a shifting, dancing pattern of light and shadow across the whale's deep, blue-black skin, which flows with its movements. Its massive, streamlined form, its skin covered in white scratches and barnacles from age, dominates the frame, its massive pectoral fins gracefully extended like wings. With an almost meditative composure, it propels its body slowly and steadily forward through the powerful, steady up-and-down movements of its enormous tail fin, surrounded by countless tiny plankton shimmering in the beams of light. The dominant colors of the scene are deep ocean blue and brilliant sunlight white, interweaving light and shadow to create a majestic, tranquil, and vibrant atmosphere filled with a profound awe for the wonders of nature.", + "seed": 18047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_8012_seed_18047_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08650551073951647, + "vlm_logprob": -0.08650551073951647, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.08650551073951647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004167919047176838 + ], + "opposite_logprob_tokens": [ + -9.504167556762695 + ], + "logprob_reward": -0.004167919047176838, + "yes_no_margin_reward": 9.5, + "token_credit_reward": -0.004167919047176838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6630376577377319, + 0.05639693886041641, + 0.01929895393550396, + 0.01770799234509468, + 0.02587609738111496, + 0.017372218891978264, + 0.07025492936372757, + 0.06529465317726135, + 0.037291575223207474, + 0.027468927204608917 + ], + "credit_map": [ + 0.1375843733549118, + 0.12788158655166626, + 0.06710056215524673, + 0.0654543936252594, + 0.03996126726269722, + 0.043130431324243546, + 0.04530130326747894, + 0.13662376999855042, + 0.004617965314537287, + 0.00669504189863801, + 0.007543816231191158, + 0.00911443866789341, + 0.005237744655460119, + 0.007992963306605816, + 0.007340758107602596, + 0.007854217663407326, + 0.001944929244928062, + 0.003552366979420185, + 0.002808158053085208, + 0.0020117098465561867, + 0.0014184645842760801, + 0.001992611214518547, + 0.0027777361683547497, + 0.002792980754747987, + 0.002291568787768483, + 0.002107731532305479, + 0.0025758587289601564, + 0.001883840304799378, + 0.0012324053095653653, + 0.0025491949636489153, + 0.0022931727580726147, + 0.002774222753942013, + 0.002708717016503215, + 0.004712033551186323, + 0.003194550983607769, + 0.0035232328809797764, + 0.0025187882129102945, + 0.0026636843103915453, + 0.003234395058825612, + 0.003320700256153941, + 0.0024108809884637594, + 0.00231199455447495, + 0.0020261553581804037, + 0.0017330009723082185, + 0.002413185080513358, + 0.0021430114284157753, + 0.0015336667420342565, + 0.0028003244660794735, + 0.006210302002727985, + 0.008463618345558643, + 0.00733042461797595, + 0.008966231718659401, + 0.009889515116810799, + 0.01099010743200779, + 0.009322894737124443, + 0.009081841446459293, + 0.00594047736376524, + 0.007635444402694702, + 0.006923812907189131, + 0.0061663175001740456, + 0.013946752063930035, + 0.00968957133591175, + 0.0074339089915156364, + 0.00755838118493557, + 0.002938869409263134, + 0.00436615664511919, + 0.006262504030019045, + 0.004013397265225649, + 0.00546196848154068, + 0.0040834201499819756, + 0.005328806582838297, + 0.004836456850171089, + 0.0030074368696659803, + 0.004390847869217396, + 0.0029630959033966064, + 0.003616658039391041, + 0.002550088334828615, + 0.00300966901704669, + 0.0028774451930075884, + 0.0050536892376840115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.004167919047176838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008148210472427309 + ], + "opposite_logprob_tokens": [ + -7.125814914703369 + ], + "logprob_reward": -0.0008148210472427309, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008148210472427309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3073897063732147, + 0.05069670081138611, + 0.07769390940666199, + 0.03573548048734665, + 0.02916560135781765, + 0.013934213668107986, + 0.29885104298591614, + 0.11342888325452805, + 0.05285291746258736, + 0.020251614972949028 + ], + "credit_map": [ + 0.06655912846326828, + 0.05662529915571213, + 0.032021988183259964, + 0.030078548938035965, + 0.017919063568115234, + 0.017001992091536522, + 0.019991926848888397, + 0.0671917274594307, + 0.005274911876767874, + 0.00754559924826026, + 0.006634560879319906, + 0.007308711763471365, + 0.004964597523212433, + 0.005712323356419802, + 0.00675234105437994, + 0.006503645330667496, + 0.005477451253682375, + 0.009559853933751583, + 0.015899131074547768, + 0.007499031722545624, + 0.00630734721198678, + 0.010322772897779942, + 0.01054166629910469, + 0.012086647562682629, + 0.003212767420336604, + 0.004362314473837614, + 0.00597525155171752, + 0.0032289482187479734, + 0.003993989434093237, + 0.006423825863748789, + 0.0036117718555033207, + 0.004926607478410006, + 0.0032834666781127453, + 0.003103136783465743, + 0.003785310545936227, + 0.003818279830738902, + 0.002498587826266885, + 0.0034757128451019526, + 0.004785305820405483, + 0.004415798932313919, + 0.001572033273987472, + 0.0012685417896136642, + 0.0015949518419802189, + 0.0025579440407454967, + 0.0013713263906538486, + 0.0021236741449683905, + 0.0015676027396693826, + 0.0018781372345983982, + 0.02612021006643772, + 0.03670123964548111, + 0.03226775676012039, + 0.03660845384001732, + 0.041971005499362946, + 0.048791445791721344, + 0.03855736553668976, + 0.037833523005247116, + 0.009362799115478992, + 0.012792843393981457, + 0.00979615468531847, + 0.015423390083014965, + 0.01723853498697281, + 0.019928695634007454, + 0.013992919586598873, + 0.01489353645592928, + 0.004740501753985882, + 0.007863453589379787, + 0.006969328969717026, + 0.0059510935097932816, + 0.00775517662987113, + 0.00529048265889287, + 0.00790426041930914, + 0.006378615740686655, + 0.0016555379843339324, + 0.00205623940564692, + 0.002455377019941807, + 0.0036338074132800102, + 0.0013702759752050042, + 0.0013682792196050286, + 0.0014158147387206554, + 0.006296282168477774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0008148210472427309 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347240433096886 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347240433096886, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347240433096886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2513633966445923, + 0.061930641531944275, + 0.0592665858566761, + 0.05300665274262428, + 0.11682622134685516, + 0.04051663354039192, + 0.214483842253685, + 0.12402288615703583, + 0.031165601685643196, + 0.0474175401031971 + ], + "credit_map": [ + 0.053377363830804825, + 0.04884830489754677, + 0.02549043297767639, + 0.024773120880126953, + 0.015341740101575851, + 0.013882145285606384, + 0.017218783497810364, + 0.05243149399757385, + 0.005840584170073271, + 0.008182298392057419, + 0.009031671099364758, + 0.011111248284578323, + 0.005874895490705967, + 0.005796502809971571, + 0.007733372505754232, + 0.008360065519809723, + 0.006321512628346682, + 0.0075791520066559315, + 0.010156117379665375, + 0.005631519015878439, + 0.004798789508640766, + 0.007071215193718672, + 0.009081103838980198, + 0.008627175353467464, + 0.004979101475328207, + 0.0059143174439668655, + 0.00917090941220522, + 0.005070239771157503, + 0.005865472834557295, + 0.008783199824392796, + 0.005712613929063082, + 0.007510801311582327, + 0.01135539636015892, + 0.025823848322033882, + 0.014634892344474792, + 0.012710128910839558, + 0.010271110571920872, + 0.012726371176540852, + 0.013657225295901299, + 0.015647251158952713, + 0.0025118698831647635, + 0.005098757799714804, + 0.003966018557548523, + 0.0057919141836464405, + 0.00645791320130229, + 0.005357346031814814, + 0.004099814686924219, + 0.007232999894768, + 0.019737403839826584, + 0.024313682690262794, + 0.024401791393756866, + 0.02731904573738575, + 0.031245579943060875, + 0.03328898176550865, + 0.02674514427781105, + 0.027432216331362724, + 0.010071017779409885, + 0.013930334709584713, + 0.010255271568894386, + 0.016883669421076775, + 0.0190344899892807, + 0.022060882300138474, + 0.015293627046048641, + 0.01649359054863453, + 0.0023860358633100986, + 0.0041240304708480835, + 0.005351098719984293, + 0.004003556445240974, + 0.0030810858588665724, + 0.003681908128783107, + 0.00425539119169116, + 0.004282493144273758, + 0.004398474469780922, + 0.004936268553137779, + 0.006004383321851492, + 0.006036682054400444, + 0.005430880002677441, + 0.008497592993080616, + 0.005139269400388002, + 0.006973989307880402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.026347240433096886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186015352606773 + ], + "opposite_logprob_tokens": [ + -4.264185905456543 + ], + "logprob_reward": -0.014186015352606773, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186015352606773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12732401490211487, + 0.06188278645277023, + 0.03000614605844021, + 0.050873976200819016, + 0.0354473814368248, + 0.06119868904352188, + 0.20386773347854614, + 0.36241811513900757, + 0.047932855784893036, + 0.019048253074288368 + ], + "credit_map": [ + 0.02720828540623188, + 0.02177535742521286, + 0.013409990817308426, + 0.013583969324827194, + 0.006574539467692375, + 0.008029744029045105, + 0.008529327809810638, + 0.02821281924843788, + 0.005310596898198128, + 0.007353967987000942, + 0.00948790367692709, + 0.011781820096075535, + 0.004908161237835884, + 0.0069398339837789536, + 0.008485959842801094, + 0.007614545524120331, + 0.002309554722160101, + 0.004270367790013552, + 0.004901390988379717, + 0.003056697314605117, + 0.002392607042565942, + 0.00346396304666996, + 0.0044020903296768665, + 0.005209475755691528, + 0.006784707773476839, + 0.007231811061501503, + 0.006125715095549822, + 0.003926392644643784, + 0.006055851932615042, + 0.006525222212076187, + 0.0065337554551661015, + 0.007690520025789738, + 0.0032073536422103643, + 0.004786496516317129, + 0.007936755195260048, + 0.0027295209001749754, + 0.0033241037745028734, + 0.0024797527585178614, + 0.0019414691487327218, + 0.00904193427413702, + 0.008080227300524712, + 0.01090147253125906, + 0.006782629061490297, + 0.007060330361127853, + 0.007927079685032368, + 0.007569444831460714, + 0.005149153061211109, + 0.007728355471044779, + 0.02268129028379917, + 0.028425218537449837, + 0.022512301802635193, + 0.023048896342515945, + 0.03009534440934658, + 0.030345803126692772, + 0.021970806643366814, + 0.02478809654712677, + 0.03927958384156227, + 0.04134279862046242, + 0.03963741660118103, + 0.045300085097551346, + 0.06682900339365005, + 0.049486324191093445, + 0.03600587323307991, + 0.04453705996274948, + 0.0034325376618653536, + 0.007505110930651426, + 0.006237566936761141, + 0.004404027946293354, + 0.006666089873760939, + 0.005610980559140444, + 0.007649519480764866, + 0.0064270226284861565, + 0.0017507645534351468, + 0.0028518722392618656, + 0.002472235821187496, + 0.002741630654782057, + 0.002077382057905197, + 0.0022722072899341583, + 0.0020935030188411474, + 0.002788657322525978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.014186015352606773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870115578174591 + ], + "opposite_logprob_tokens": [ + -1.1370115280151367 + ], + "logprob_reward": -0.3870115578174591, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870115578174591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34234845638275146, + 0.06749953329563141, + 0.07131896913051605, + 0.03000234253704548, + 0.03892834112048149, + 0.03440253436565399, + 0.11309683322906494, + 0.07387954741716385, + 0.08315197378396988, + 0.1453714519739151 + ], + "credit_map": [ + 0.07475005835294724, + 0.06405705213546753, + 0.036252450197935104, + 0.03452166169881821, + 0.01910948008298874, + 0.018438046798110008, + 0.020765142515301704, + 0.07445453107357025, + 0.006288009695708752, + 0.008347583934664726, + 0.010432621464133263, + 0.008797740563750267, + 0.005604789592325687, + 0.008415662683546543, + 0.009527541697025299, + 0.01008558552712202, + 0.005964450538158417, + 0.012002947740256786, + 0.014694239012897015, + 0.007491942495107651, + 0.0047013526782393456, + 0.006272627040743828, + 0.008905348367989063, + 0.011286057531833649, + 0.0038287986535578966, + 0.002586295595392585, + 0.004461934790015221, + 0.003928509075194597, + 0.0020547679159790277, + 0.004927363712340593, + 0.00355953979305923, + 0.004655135329812765, + 0.0037844963371753693, + 0.007904690690338612, + 0.004565679933875799, + 0.0038188574835658073, + 0.006310170516371727, + 0.004491415806114674, + 0.003511450020596385, + 0.004541578236967325, + 0.003845474449917674, + 0.004850554745644331, + 0.004090715665370226, + 0.003936851862818003, + 0.005685568321496248, + 0.004223935306072235, + 0.0029061157256364822, + 0.004863315727561712, + 0.010291144251823425, + 0.013694203458726406, + 0.01292391773313284, + 0.014162708073854446, + 0.015827493742108345, + 0.017344364896416664, + 0.014331777580082417, + 0.014521225355565548, + 0.00603043707087636, + 0.007578971330076456, + 0.007587398402392864, + 0.010760847479104996, + 0.013912912458181381, + 0.009164738468825817, + 0.009319322183728218, + 0.00952492281794548, + 0.006119570694863796, + 0.011918329633772373, + 0.013763856142759323, + 0.007345888298004866, + 0.011536973528563976, + 0.012060055509209633, + 0.009748182259500027, + 0.010659119114279747, + 0.015363388694822788, + 0.016810769215226173, + 0.015339672565460205, + 0.021276893094182014, + 0.01173920277506113, + 0.01757192239165306, + 0.014123725704848766, + 0.033145882189273834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.3870115578174591 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot captured from underwater, the camera slowly pans to the right, following a massive humpback whale as it cruises through the boundless deep blue ocean. Brilliant beams of sunlight penetrate the shimmering surface, piercing the clear waters vertically, casting a shifting, dancing pattern of light and shadow across the whale's deep, blue-black skin, which flows with its movements. Its massive, streamlined form, its skin covered in white scratches and barnacles from age, dominates the frame, its massive pectoral fins gracefully extended like wings. With an almost meditative composure, it propels its body slowly and steadily forward through the powerful, steady up-and-down movements of its enormous tail fin, surrounded by countless tiny plankton shimmering in the beams of light. The dominant colors of the scene are deep ocean blue and brilliant sunlight white, interweaving light and shadow to create a majestic, tranquil, and vibrant atmosphere filled with a profound awe for the wonders of nature.", + "seed": 18048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_1/index_8012_seed_18048_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07358493357896805, + "vlm_logprob": -0.07358493357896805, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.07358493357896805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004781479015946388 + ], + "opposite_logprob_tokens": [ + -8.879781723022461 + ], + "logprob_reward": -0.004781479015946388, + "yes_no_margin_reward": 8.875, + "token_credit_reward": -0.004781479015946388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028230411931872368, + 0.0342603400349617, + 0.06276839226484299, + 0.03493957221508026, + 0.10225526988506317, + 0.19917872548103333, + 0.13807471096515656, + 0.24765326082706451, + 0.143437922000885, + 0.009201421402394772 + ], + "credit_map": [ + 0.0029423050582408905, + 0.003906772006303072, + 0.003127827076241374, + 0.005150197073817253, + 0.004426284693181515, + 0.003602579701691866, + 0.002424415433779359, + 0.0026500278618186712, + 0.003573066322132945, + 0.002581889508292079, + 0.006099473685026169, + 0.004528907593339682, + 0.004571323283016682, + 0.003672576742246747, + 0.00514552416279912, + 0.004087578039616346, + 0.004757383838295937, + 0.007216522470116615, + 0.008617826737463474, + 0.009196428582072258, + 0.008282504975795746, + 0.009024228900671005, + 0.009679701179265976, + 0.005993785336613655, + 0.0024122116155922413, + 0.0026782676577568054, + 0.004775885958224535, + 0.0035225092433393, + 0.0034711004700511694, + 0.00881583709269762, + 0.0028566319961100817, + 0.0064071244560182095, + 0.009553911164402962, + 0.009415674954652786, + 0.015137499198317528, + 0.01626553013920784, + 0.014370741322636604, + 0.012134787626564503, + 0.013911913149058819, + 0.011465199291706085, + 0.019806306809186935, + 0.021578947082161903, + 0.026335997506976128, + 0.020421016961336136, + 0.03486936166882515, + 0.0324147492647171, + 0.02033478207886219, + 0.023417549207806587, + 0.011290909722447395, + 0.013992983847856522, + 0.028574161231517792, + 0.01848253421485424, + 0.011422185227274895, + 0.01829088293015957, + 0.015856817364692688, + 0.02016422711312771, + 0.013730018399655819, + 0.00769181689247489, + 0.051177941262722015, + 0.018344096839427948, + 0.01489483192563057, + 0.013097675517201424, + 0.08847034722566605, + 0.04024650529026985, + 0.010892639867961407, + 0.007840074598789215, + 0.03797074779868126, + 0.01606765389442444, + 0.011217962950468063, + 0.008845961652696133, + 0.030248481780290604, + 0.020354388281702995, + 0.0013791524106636643, + 0.0007396109285764396, + 0.0013770662480965257, + 0.0017582154832780361, + 0.0007844200008548796, + 0.000751838437281549, + 0.0010798051953315735, + 0.0013313121162354946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.004781479015946388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791725799441338 + ], + "opposite_logprob_tokens": [ + -3.529791831970215 + ], + "logprob_reward": -0.029791725799441338, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791725799441338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09801970422267914, + 0.03902745619416237, + 0.08753685653209686, + 0.014194185845553875, + 0.03931295499205589, + 0.06488603353500366, + 0.05430495738983154, + 0.5324066877365112, + 0.04174390807747841, + 0.028567304834723473 + ], + "credit_map": [ + 0.007529029622673988, + 0.014630218967795372, + 0.011441647075116634, + 0.018450476229190826, + 0.010115724988281727, + 0.011737668886780739, + 0.01196998730301857, + 0.012144951149821281, + 0.0034394587855786085, + 0.006188273895531893, + 0.004276007413864136, + 0.005760288797318935, + 0.003793438896536827, + 0.004119944293051958, + 0.006628077011555433, + 0.0048219673335552216, + 0.0098167909309268, + 0.01125626266002655, + 0.011843249201774597, + 0.01233815960586071, + 0.01044103316962719, + 0.011760186403989792, + 0.012018590234220028, + 0.00806258711963892, + 0.0010740479920059443, + 0.0011096266098320484, + 0.0018113162368535995, + 0.001198241370730102, + 0.0015213576843962073, + 0.0036878366954624653, + 0.0009765340946614742, + 0.0028152253944426775, + 0.0034082478377968073, + 0.0034916335716843605, + 0.006042105611413717, + 0.006317988503724337, + 0.005427452735602856, + 0.00452496437355876, + 0.005268761422485113, + 0.00483180396258831, + 0.006743444129824638, + 0.005594567395746708, + 0.010598604567348957, + 0.009534966200590134, + 0.007717770989984274, + 0.006406395696103573, + 0.006293921265751123, + 0.011996367014944553, + 0.004279790446162224, + 0.0019420661265030503, + 0.016822636127471924, + 0.004755035974085331, + 0.0027768004219979048, + 0.0030256714671850204, + 0.006077496334910393, + 0.01462546270340681, + 0.01604938693344593, + 0.01813187636435032, + 0.11155357211828232, + 0.03093264438211918, + 0.029465779662132263, + 0.016759231686592102, + 0.21616530418395996, + 0.09334888309240341, + 0.003156129503622651, + 0.0026437549386173487, + 0.010095416568219662, + 0.00584439979866147, + 0.003186235437169671, + 0.002843041904270649, + 0.009623359888792038, + 0.004351573530584574, + 0.003970394842326641, + 0.002907562768086791, + 0.004081190563738346, + 0.0038856773171573877, + 0.0031402367167174816, + 0.0025319275446236134, + 0.004138349089771509, + 0.00391196645796299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.029791725799441338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495869740843773 + ], + "opposite_logprob_tokens": [ + -2.9299585819244385 + ], + "logprob_reward": -0.05495869740843773, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495869740843773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07586166262626648, + 0.05175285413861275, + 0.07828459143638611, + 0.02890753000974655, + 0.17229759693145752, + 0.05637602135539055, + 0.06042825058102608, + 0.3477618992328644, + 0.07314741611480713, + 0.055182233452796936 + ], + "credit_map": [ + 0.006138333585113287, + 0.010624292306602001, + 0.008512669242918491, + 0.011253819800913334, + 0.009018738754093647, + 0.009371311403810978, + 0.010989406146109104, + 0.009953088127076626, + 0.004574249964207411, + 0.007059933617711067, + 0.007496488280594349, + 0.00710008991882205, + 0.004942078609019518, + 0.005897384602576494, + 0.008335234597325325, + 0.006347394082695246, + 0.008972822688519955, + 0.009565816260874271, + 0.010661331005394459, + 0.010258101858198643, + 0.008959146216511726, + 0.010452535934746265, + 0.012601522728800774, + 0.006813311018049717, + 0.002340273465961218, + 0.0018748912261798978, + 0.004617512691766024, + 0.0024949335493147373, + 0.002910218434408307, + 0.008624161593616009, + 0.002103780396282673, + 0.003941756207495928, + 0.018538132309913635, + 0.007093559484928846, + 0.06071816012263298, + 0.016230938956141472, + 0.010405176319181919, + 0.01775158941745758, + 0.00942828319966793, + 0.03213174641132355, + 0.005420772824436426, + 0.005600095260888338, + 0.006219457369297743, + 0.008604805916547775, + 0.008080166764557362, + 0.0062895966693758965, + 0.007887930609285831, + 0.008273188024759293, + 0.005434945225715637, + 0.002460283925756812, + 0.017669355496764183, + 0.005349049810320139, + 0.0029687751084566116, + 0.00510445749387145, + 0.005678755696862936, + 0.01576262153685093, + 0.011734036728739738, + 0.01405528374016285, + 0.0702609047293663, + 0.020217662677168846, + 0.021867046132683754, + 0.012117605656385422, + 0.13723455369472504, + 0.06027478724718094, + 0.004664998035877943, + 0.003642871743068099, + 0.02223043702542782, + 0.007334085181355476, + 0.004013707395642996, + 0.003638406516984105, + 0.01985645852982998, + 0.007766439579427242, + 0.009607777930796146, + 0.004013820085674524, + 0.008162114769220352, + 0.00799540150910616, + 0.005185621790587902, + 0.004602914210408926, + 0.00789320282638073, + 0.007721377070993185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.05495869740843773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633446827530861 + ], + "opposite_logprob_tokens": [ + -3.651334524154663 + ], + "logprob_reward": -0.02633446827530861, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633446827530861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05190327391028404, + 0.08789830654859543, + 0.07277730852365494, + 0.11920125782489777, + 0.15259093046188354, + 0.07525968551635742, + 0.13751916587352753, + 0.20026937127113342, + 0.0529552586376667, + 0.04962543025612831 + ], + "credit_map": [ + 0.005005671642720699, + 0.005328511819243431, + 0.006828984711319208, + 0.005376629531383514, + 0.012561828829348087, + 0.0074723768047988415, + 0.004908229690045118, + 0.004421040415763855, + 0.010471806861460209, + 0.0060295709408819675, + 0.026179242879152298, + 0.010349671356379986, + 0.009023795835673809, + 0.012510981410741806, + 0.006681671366095543, + 0.006651563569903374, + 0.009572584182024002, + 0.008576533757150173, + 0.010833109728991985, + 0.00942687876522541, + 0.01123262383043766, + 0.010513599961996078, + 0.006983482278883457, + 0.005638494156301022, + 0.016438445076346397, + 0.014788180589675903, + 0.019009044393897057, + 0.014740263111889362, + 0.012817343696951866, + 0.016055645421147346, + 0.012305171228945255, + 0.013047155924141407, + 0.014991511590778828, + 0.009169726632535458, + 0.046719495207071304, + 0.015237618237733841, + 0.015188626945018768, + 0.016499442979693413, + 0.010828517377376556, + 0.023955998942255974, + 0.007736118044704199, + 0.006896846927702427, + 0.012395397759974003, + 0.008340410888195038, + 0.00986015796661377, + 0.01040518656373024, + 0.008174625225365162, + 0.011450937017798424, + 0.020420242100954056, + 0.008421748876571655, + 0.03139728680253029, + 0.011825604364275932, + 0.00775336567312479, + 0.021684816107153893, + 0.01417633332312107, + 0.021839775145053864, + 0.017280234023928642, + 0.008512526750564575, + 0.036335285753011703, + 0.017394332215189934, + 0.008973527699708939, + 0.008636332117021084, + 0.06610199064016342, + 0.03703514486551285, + 0.0038536705542355776, + 0.0031036026775836945, + 0.014194231480360031, + 0.005896135699003935, + 0.0038824460934847593, + 0.0034009586088359356, + 0.012407192960381508, + 0.006217019632458687, + 0.009485941380262375, + 0.004451148211956024, + 0.006094419863075018, + 0.0059843529015779495, + 0.005340072326362133, + 0.004407063592225313, + 0.006787149701267481, + 0.007075282745063305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.02633446827530861 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive humpback whale visible as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin show white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are vertical beams of sunlight penetrating the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520582973957062 + ], + "opposite_logprob_tokens": [ + -1.5020582675933838 + ], + "logprob_reward": -0.2520582973957062, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520582973957062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05136086791753769, + 0.049423035234212875, + 0.050772834569215775, + 0.0536312460899353, + 0.12339812517166138, + 0.10550542920827866, + 0.16031605005264282, + 0.2718592584133148, + 0.09267116338014603, + 0.041062016040086746 + ], + "credit_map": [ + 0.0035687037743628025, + 0.006938677281141281, + 0.004965352360159159, + 0.00924657378345728, + 0.006996027193963528, + 0.005338993854820728, + 0.0059047178365290165, + 0.008401825092732906, + 0.005314955487847328, + 0.0041395812295377254, + 0.01206528302282095, + 0.006841670256108046, + 0.005097400397062302, + 0.006851851008832455, + 0.005231253337115049, + 0.0038810409605503082, + 0.006856375839561224, + 0.00626586377620697, + 0.00800587609410286, + 0.006071123760193586, + 0.005369931925088167, + 0.006986051797866821, + 0.005941319745033979, + 0.00527628930285573, + 0.006287459284067154, + 0.005891849752515554, + 0.006729478482156992, + 0.0050767879001796246, + 0.005348246078938246, + 0.010432424023747444, + 0.006332323886454105, + 0.00753267714753747, + 0.010040528140962124, + 0.011100398376584053, + 0.01965693011879921, + 0.019913552328944206, + 0.01639309898018837, + 0.013898821547627449, + 0.0174857284873724, + 0.014909078367054462, + 0.01077529601752758, + 0.011194019578397274, + 0.014368342235684395, + 0.010944432578980923, + 0.018945101648569107, + 0.013314835727214813, + 0.012595854699611664, + 0.013367542065680027, + 0.009840602055191994, + 0.0065638101659715176, + 0.05064493417739868, + 0.014772607944905758, + 0.010645945556461811, + 0.009168881922960281, + 0.018898164853453636, + 0.039781101047992706, + 0.0114716412499547, + 0.006664425600320101, + 0.05678901821374893, + 0.01652831956744194, + 0.010874197818338871, + 0.008494251407682896, + 0.11016799509525299, + 0.050869420170784, + 0.009490886703133583, + 0.009222539141774178, + 0.01601620763540268, + 0.01173837948590517, + 0.010967763140797615, + 0.008306579664349556, + 0.013974213972687721, + 0.01295460294932127, + 0.004834506195038557, + 0.002716565039008856, + 0.005374492146074772, + 0.005218105856329203, + 0.004131506197154522, + 0.004563152324408293, + 0.00856243260204792, + 0.005661256145685911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.2520582973957062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_tensor([8887, 8887])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_tensor([8887, 8887])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8514a313cb2a8d455d881e3f66cfb291a430b387 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_tensor([8887, 8887])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, close-up, static shot, captured from a slight overhead angle, focuses on a gentle Asian mother, comfortably seated in a vintage wooden armchair. Soft afternoon sunlight slants through a large window, casting a warm golden hue on her long, dark brown hair, pulled into a low bun, and the off-white, turtleneck cashmere sweater she wears. Her face holds a serene, focused expression, her eyes slightly downcast, her gaze completely focused on the work in her hands. Her hands move steadily and deliberately, two smooth bamboo knitting needles twirling between her deft fingers. One needle expertly picks up lavender yarn and delicately threads it through the loop of the other. A half-finished sweater drapes across her legs, its intricate, twisted pattern clearly visible in the light, each fluffy fiber seeming palpable, a vivid, textured expression. Beside her, a wicker basket filled with colorful balls of yarn rests on the ground. The bookshelves and greenery in the background are softly blurred by the shallow depth of field. The entire scene is warm and harmonious, imbued with the warmth of home and the tranquility of time brought by handmade crafts.", + "seed": 8252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_8887_seed_8252_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038808343932032585, + "vlm_logprob": -0.038808343932032585, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.038808343932032585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04715266078710556 + ], + "opposite_logprob_tokens": [ + -3.1721527576446533 + ], + "logprob_reward": -0.04715266078710556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04715266078710556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4500083923339844, + 0.11529833823442459, + 0.06874724477529526, + 0.044088300317525864, + 0.06086324527859688, + 0.058004699647426605, + 0.04433232545852661, + 0.04649573564529419, + 0.08136516809463501, + 0.030796533450484276 + ], + "credit_map": [ + 0.07025051862001419, + 0.05636173486709595, + 0.05732059106230736, + 0.0365668386220932, + 0.07887225598096848, + 0.0410136841237545, + 0.07561847567558289, + 0.03400430083274841, + 0.016780739650130272, + 0.014684830792248249, + 0.011702601797878742, + 0.014028100296854973, + 0.020059993490576744, + 0.00986985769122839, + 0.021224088966846466, + 0.006948132999241352, + 0.009897568263113499, + 0.008231387473642826, + 0.007036903873085976, + 0.007014884147793055, + 0.010284465737640858, + 0.006552170030772686, + 0.013260695151984692, + 0.006469176150858402, + 0.007135479710996151, + 0.005918064620345831, + 0.005816805176436901, + 0.003913703374564648, + 0.006304677110165358, + 0.004484716337174177, + 0.007277132011950016, + 0.0032377224415540695, + 0.009149700403213501, + 0.009059415198862553, + 0.008165402337908745, + 0.005495161283761263, + 0.009344823658466339, + 0.007160003762692213, + 0.008178574033081532, + 0.004310167394578457, + 0.008396835997700691, + 0.008532200939953327, + 0.007671527564525604, + 0.006056754384189844, + 0.008296514861285686, + 0.006237151101231575, + 0.00897390116006136, + 0.003839816665276885, + 0.007306001614779234, + 0.007806963752955198, + 0.003697867039591074, + 0.005120624788105488, + 0.007032553665339947, + 0.0030282142106443644, + 0.007179029285907745, + 0.0031610706355422735, + 0.006953025236725807, + 0.006703827064484358, + 0.005431432742625475, + 0.004226847551763058, + 0.007606538478285074, + 0.0057970499619841576, + 0.006447139661759138, + 0.0033298740163445473, + 0.01448936015367508, + 0.016578471288084984, + 0.010876336134970188, + 0.009870910085737705, + 0.008048047311604023, + 0.008887944743037224, + 0.0077666305005550385, + 0.004847465083003044, + 0.004644617438316345, + 0.005427975207567215, + 0.0032640034332871437, + 0.002559535438194871, + 0.005281741265207529, + 0.0023216346744447947, + 0.0043645743280649185, + 0.0029324511997401714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04715266078710556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205959990620613 + ], + "opposite_logprob_tokens": [ + -2.8120596408843994 + ], + "logprob_reward": -0.06205959990620613, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205959990620613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2544347643852234, + 0.12483447790145874, + 0.09477007389068604, + 0.0684964656829834, + 0.08371347934007645, + 0.12727779150009155, + 0.07829102128744125, + 0.06810285896062851, + 0.06299932301044464, + 0.037079740315675735 + ], + "credit_map": [ + 0.037149690091609955, + 0.04971291869878769, + 0.03175199776887894, + 0.02802085131406784, + 0.04352826625108719, + 0.018479814752936363, + 0.02548818476498127, + 0.02030305005609989, + 0.0186527818441391, + 0.016436558216810226, + 0.014364534988999367, + 0.014752927236258984, + 0.018674682825803757, + 0.009997720830142498, + 0.02083210088312626, + 0.011123169213533401, + 0.013681454584002495, + 0.013268884271383286, + 0.013611228205263615, + 0.009380745701491833, + 0.012363651767373085, + 0.006937270052731037, + 0.015897732228040695, + 0.00962910521775484, + 0.010198172181844711, + 0.009097246453166008, + 0.01769774965941906, + 0.00805166270583868, + 0.006196317728608847, + 0.005320515017956495, + 0.0067266495898365974, + 0.005208149552345276, + 0.01149042695760727, + 0.011118113994598389, + 0.008253943175077438, + 0.008881746791303158, + 0.013368732295930386, + 0.007601602468639612, + 0.014235562644898891, + 0.008763347752392292, + 0.01684490777552128, + 0.01355039980262518, + 0.011385505087673664, + 0.014576314017176628, + 0.023131664842367172, + 0.00911626685410738, + 0.027264118194580078, + 0.011408618651330471, + 0.01377956010401249, + 0.010774403810501099, + 0.014483822509646416, + 0.008436331525444984, + 0.010107788257300854, + 0.006333117838948965, + 0.007909572683274746, + 0.006466425023972988, + 0.009050763212144375, + 0.008557518944144249, + 0.006757719442248344, + 0.00780222425237298, + 0.011331182904541492, + 0.004918617662042379, + 0.01481495052576065, + 0.004869875963777304, + 0.00777572114020586, + 0.006521404255181551, + 0.005955747328698635, + 0.00816291756927967, + 0.010263259522616863, + 0.004786651115864515, + 0.013055040501058102, + 0.006478584371507168, + 0.0056852251291275024, + 0.006276831496506929, + 0.004713383037596941, + 0.0034593124873936176, + 0.003961565904319286, + 0.0030077225528657436, + 0.00411814684048295, + 0.0058575524017214775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.06205959990620613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643122389912605 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643122389912605, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643122389912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34105464816093445, + 0.1151375025510788, + 0.10555649548768997, + 0.053866103291511536, + 0.05983612686395645, + 0.11622969806194305, + 0.07637397199869156, + 0.043767351657152176, + 0.05141967162489891, + 0.03675844147801399 + ], + "credit_map": [ + 0.03534386307001114, + 0.046672478318214417, + 0.03912336006760597, + 0.04886547848582268, + 0.056875087320804596, + 0.02737031690776348, + 0.04669084772467613, + 0.04011324420571327, + 0.015841908752918243, + 0.017027543857693672, + 0.015178631991147995, + 0.011037066578865051, + 0.018194833770394325, + 0.007684159558266401, + 0.01742818020284176, + 0.012745183892548084, + 0.01613660342991352, + 0.014247452840209007, + 0.012369715608656406, + 0.012020943686366081, + 0.017805777490139008, + 0.009190602228045464, + 0.012661493383347988, + 0.011123904027044773, + 0.007557787466794252, + 0.0066229915246367455, + 0.006373319774866104, + 0.0069475071504712105, + 0.00896285381168127, + 0.004365888424217701, + 0.008306589908897877, + 0.004729161970317364, + 0.008652839809656143, + 0.007840885780751705, + 0.00447710370644927, + 0.008162638172507286, + 0.00995653960853815, + 0.005521783605217934, + 0.007180278189480305, + 0.00804405752569437, + 0.017033804208040237, + 0.018122700974345207, + 0.009180429391562939, + 0.01274933759123087, + 0.02226198837161064, + 0.00878323707729578, + 0.015492779202759266, + 0.012605417519807816, + 0.011598649434745312, + 0.010570145212113857, + 0.016548579558730125, + 0.010701172985136509, + 0.008376849815249443, + 0.004470114130526781, + 0.008433782495558262, + 0.0056746783666312695, + 0.005613570101559162, + 0.0056056855246424675, + 0.006019426975399256, + 0.004302620887756348, + 0.0068844035267829895, + 0.003955988213419914, + 0.00817639846354723, + 0.0032092598266899586, + 0.007763172034174204, + 0.009236516430974007, + 0.006677452940493822, + 0.005361291579902172, + 0.005385323893278837, + 0.005791153758764267, + 0.005869453772902489, + 0.005335306283086538, + 0.005973165389150381, + 0.005414115730673075, + 0.007394098211079836, + 0.0038214940577745438, + 0.0036943627055734396, + 0.003728196956217289, + 0.0033756557386368513, + 0.003357355948537588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.008643122389912605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809661045670509 + ], + "opposite_logprob_tokens": [ + -3.2880966663360596 + ], + "logprob_reward": -0.03809661045670509, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809661045670509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30379316210746765, + 0.14673131704330444, + 0.12289366871118546, + 0.07391008734703064, + 0.05226802080869675, + 0.10845242440700531, + 0.05596952512860298, + 0.05049910023808479, + 0.04746132716536522, + 0.03802139684557915 + ], + "credit_map": [ + 0.03677884861826897, + 0.04322224110364914, + 0.030408145859837532, + 0.04200637340545654, + 0.04388539120554924, + 0.03189451992511749, + 0.029180537909269333, + 0.046417076140642166, + 0.01930164359509945, + 0.024007655680179596, + 0.02020472288131714, + 0.014447416178882122, + 0.01740182563662529, + 0.012615722604095936, + 0.021855460479855537, + 0.016896871849894524, + 0.014866269193589687, + 0.014925835654139519, + 0.017285604029893875, + 0.016209056600928307, + 0.0167192742228508, + 0.009583146311342716, + 0.021727485582232475, + 0.011576997116208076, + 0.010819328017532825, + 0.009473167359828949, + 0.01192681398242712, + 0.007776899263262749, + 0.010261268354952335, + 0.006486490834504366, + 0.012493066489696503, + 0.00467305164784193, + 0.006729872897267342, + 0.005354153458029032, + 0.00475645624101162, + 0.008644004352390766, + 0.007780354470014572, + 0.005303421523422003, + 0.005981047637760639, + 0.007718711160123348, + 0.013378871604800224, + 0.011268424801528454, + 0.009592214599251747, + 0.012284628115594387, + 0.021650951355695724, + 0.008467784151434898, + 0.0240781269967556, + 0.007731415797024965, + 0.008525815792381763, + 0.00743591133505106, + 0.011876225471496582, + 0.007779180072247982, + 0.006583971902728081, + 0.004711091052740812, + 0.005693214014172554, + 0.0033641199115663767, + 0.0065416269935667515, + 0.005920812021940947, + 0.006031231489032507, + 0.003966787829995155, + 0.009399865753948689, + 0.004502692725509405, + 0.010831184685230255, + 0.003304899437353015, + 0.006995424162596464, + 0.007123787421733141, + 0.005590883549302816, + 0.005445966962724924, + 0.006723066326230764, + 0.004112591966986656, + 0.006102284416556358, + 0.005367324221879244, + 0.005689932033419609, + 0.006767923012375832, + 0.006829623598605394, + 0.003603733843192458, + 0.004589579068124294, + 0.004562275018543005, + 0.002846140880137682, + 0.0031321889255195856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03809661045670509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089726120233536 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.038089726120233536, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089726120233536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2792286276817322, + 0.16705024242401123, + 0.10011446475982666, + 0.11362683773040771, + 0.03966720029711723, + 0.09801017493009567, + 0.05551745742559433, + 0.05690465494990349, + 0.03331747278571129, + 0.05656293407082558 + ], + "credit_map": [ + 0.049623068422079086, + 0.041788503527641296, + 0.04488607123494148, + 0.0330742783844471, + 0.02720526233315468, + 0.03530753776431084, + 0.02467341162264347, + 0.022670457139611244, + 0.02604382298886776, + 0.02340897172689438, + 0.0313061885535717, + 0.028079697862267494, + 0.01324033085256815, + 0.01794010028243065, + 0.01607350818812847, + 0.010957607999444008, + 0.013379725627601147, + 0.012621704488992691, + 0.021693261340260506, + 0.014005466364324093, + 0.009391221217811108, + 0.009029185399413109, + 0.011700525879859924, + 0.008293362334370613, + 0.02098735049366951, + 0.017375152558088303, + 0.024372274056077003, + 0.011648979038000107, + 0.00902792438864708, + 0.013703362084925175, + 0.007480216212570667, + 0.009031577035784721, + 0.006523120682686567, + 0.006875252351164818, + 0.006809656973928213, + 0.004192220512777567, + 0.004596145357936621, + 0.00427615363150835, + 0.003824640763923526, + 0.002570004668086767, + 0.01149787101894617, + 0.011539485305547714, + 0.015831375494599342, + 0.01337407622486353, + 0.013734978623688221, + 0.010209634900093079, + 0.015733851119875908, + 0.006088896654546261, + 0.007834753021597862, + 0.0076582059264183044, + 0.010302096605300903, + 0.007825112901628017, + 0.005721263121813536, + 0.0048841130919754505, + 0.005510586779564619, + 0.005781323183327913, + 0.008420192636549473, + 0.007226876448839903, + 0.011296863667666912, + 0.008642812259495258, + 0.005618494935333729, + 0.00522241834551096, + 0.007141160313040018, + 0.003335829358547926, + 0.004608567804098129, + 0.0041188267059624195, + 0.005038853734731674, + 0.004319830797612667, + 0.004000977613031864, + 0.003408788936212659, + 0.004519920330494642, + 0.00330170220695436, + 0.007956690154969692, + 0.007426420692354441, + 0.0098445238545537, + 0.01008343230932951, + 0.005978573113679886, + 0.00446771876886487, + 0.006132750306278467, + 0.004672821145504713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.038089726120233536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, close-up, static shot, captured from a slight overhead angle, focuses on a gentle Asian mother, comfortably seated in a vintage wooden armchair. Soft afternoon sunlight slants through a large window, casting a warm golden hue on her long, dark brown hair, pulled into a low bun, and the off-white, turtleneck cashmere sweater she wears. Her face holds a serene, focused expression, her eyes slightly downcast, her gaze completely focused on the work in her hands. Her hands move steadily and deliberately, two smooth bamboo knitting needles twirling between her deft fingers. One needle expertly picks up lavender yarn and delicately threads it through the loop of the other. A half-finished sweater drapes across her legs, its intricate, twisted pattern clearly visible in the light, each fluffy fiber seeming palpable, a vivid, textured expression. Beside her, a wicker basket filled with colorful balls of yarn rests on the ground. The bookshelves and greenery in the background are softly blurred by the shallow depth of field. The entire scene is warm and harmonious, imbued with the warmth of home and the tranquility of time brought by handmade crafts.", + "seed": 8253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_8887_seed_8253_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026031986624002457, + "vlm_logprob": -0.026031986624002457, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.026031986624002457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022032257169485092 + ], + "opposite_logprob_tokens": [ + -4.147032260894775 + ], + "logprob_reward": -0.022032257169485092, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022032257169485092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35958248376846313, + 0.10002771764993668, + 0.09847419708967209, + 0.06812754273414612, + 0.07405902445316315, + 0.10071498900651932, + 0.04408407211303711, + 0.08607042580842972, + 0.03821910545229912, + 0.030640436336398125 + ], + "credit_map": [ + 0.07648435235023499, + 0.10565708577632904, + 0.04438277333974838, + 0.0234459787607193, + 0.029567072167992592, + 0.035545941442251205, + 0.021219268441200256, + 0.02327999845147133, + 0.018421651795506477, + 0.019778743386268616, + 0.014743905514478683, + 0.013225700706243515, + 0.008173249661922455, + 0.006969930604100227, + 0.010444078594446182, + 0.00827045738697052, + 0.014005035161972046, + 0.022450074553489685, + 0.012754207476973534, + 0.012240143492817879, + 0.010930459946393967, + 0.01264021173119545, + 0.007488500326871872, + 0.005965568590909243, + 0.013658073730766773, + 0.010449525900185108, + 0.008287277072668076, + 0.006135203409940004, + 0.0048377374187111855, + 0.0028759632259607315, + 0.017057502642273903, + 0.004826255142688751, + 0.009262711741030216, + 0.012090800330042839, + 0.006012479774653912, + 0.006232045125216246, + 0.013462305068969727, + 0.015013255178928375, + 0.006804605480283499, + 0.005180825013667345, + 0.01846904493868351, + 0.023196756839752197, + 0.011360442265868187, + 0.014082597568631172, + 0.009757216088473797, + 0.010822920128703117, + 0.007264187093824148, + 0.005761828273534775, + 0.007272015791386366, + 0.01075039803981781, + 0.004555416759103537, + 0.003455936908721924, + 0.004990815185010433, + 0.0060883983969688416, + 0.003800544422119856, + 0.003170546144247055, + 0.015467122197151184, + 0.012714821845293045, + 0.014189089648425579, + 0.010326720774173737, + 0.010699927806854248, + 0.009639479219913483, + 0.00619800528511405, + 0.00683525251224637, + 0.006503602489829063, + 0.0086264843121171, + 0.004912593401968479, + 0.003670735517516732, + 0.0044539798982441425, + 0.00341292773373425, + 0.0034797328989952803, + 0.0031590473372489214, + 0.004609240218997002, + 0.0038594480138272047, + 0.004129369277507067, + 0.003823047736659646, + 0.004305693786591291, + 0.0035647933837026358, + 0.003771902760490775, + 0.0025769404601305723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.022032257169485092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962873458862305 + ], + "opposite_logprob_tokens": [ + -2.9299628734588623 + ], + "logprob_reward": -0.054962873458862305, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962873458862305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24764631688594818, + 0.14388686418533325, + 0.06359457224607468, + 0.08091814815998077, + 0.04415644332766533, + 0.06666049361228943, + 0.05619444698095322, + 0.13850371539592743, + 0.12992891669273376, + 0.02851005271077156 + ], + "credit_map": [ + 0.03717651590704918, + 0.039626363664865494, + 0.03125589340925217, + 0.02002854086458683, + 0.03851283714175224, + 0.03306373208761215, + 0.021654196083545685, + 0.026328230276703835, + 0.03132839500904083, + 0.04238816723227501, + 0.01396864838898182, + 0.013074052520096302, + 0.013659513555467129, + 0.01043945923447609, + 0.009167255833745003, + 0.009861370548605919, + 0.010016310028731823, + 0.012262297794222832, + 0.011372256092727184, + 0.009504132904112339, + 0.00643194979056716, + 0.005025510210543871, + 0.0047356304712593555, + 0.0042464896105229855, + 0.017554759979248047, + 0.027050524950027466, + 0.009033870883286, + 0.005842905957251787, + 0.005560172721743584, + 0.004246236756443977, + 0.007373907137662172, + 0.0042557744309306145, + 0.008364549838006496, + 0.006833450868725777, + 0.00560232438147068, + 0.006082356441766024, + 0.003946904558688402, + 0.0065697296522557735, + 0.0036092186346650124, + 0.0031479091849178076, + 0.011559203267097473, + 0.014828508719801903, + 0.009694368578493595, + 0.010771271772682667, + 0.006656360346823931, + 0.0040123350918293, + 0.005411236081272364, + 0.0037272071931511164, + 0.008288245648145676, + 0.012336915358901024, + 0.006382620427757502, + 0.008477333933115005, + 0.006808530073612928, + 0.004083463456481695, + 0.006384924054145813, + 0.0034324093721807003, + 0.02451963722705841, + 0.03160414472222328, + 0.015166951343417168, + 0.014841935597360134, + 0.012462706305086613, + 0.014998849481344223, + 0.01421472616493702, + 0.010694768279790878, + 0.02103482000529766, + 0.03462153300642967, + 0.02158048376441002, + 0.014596578665077686, + 0.01106447633355856, + 0.00824744813144207, + 0.011043368838727474, + 0.007740211207419634, + 0.003910068888217211, + 0.003998905885964632, + 0.004299957305192947, + 0.004349394701421261, + 0.0037820900324732065, + 0.0031104034278541803, + 0.002795572392642498, + 0.002263659378513694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.054962873458862305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864548608660698 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.00864548608660698, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864548608660698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4171387255191803, + 0.1302374005317688, + 0.09494470059871674, + 0.06419946253299713, + 0.06234152987599373, + 0.07995636761188507, + 0.049865614622831345, + 0.04699161276221275, + 0.020042486488819122, + 0.0342821329832077 + ], + "credit_map": [ + 0.09784857928752899, + 0.12555056810379028, + 0.05821216106414795, + 0.035679131746292114, + 0.026300277560949326, + 0.026318684220314026, + 0.022684581577777863, + 0.024544741958379745, + 0.02761644683778286, + 0.027602897956967354, + 0.01860513910651207, + 0.011585157364606857, + 0.013711567968130112, + 0.01248444989323616, + 0.01002562791109085, + 0.008606120012700558, + 0.01114080660045147, + 0.01581070013344288, + 0.011065715923905373, + 0.015531203709542751, + 0.01275747548788786, + 0.009587147273123264, + 0.0112378541380167, + 0.007813802920281887, + 0.010170032270252705, + 0.009071393869817257, + 0.008042233064770699, + 0.008194688707590103, + 0.007297510746866465, + 0.007066691759973764, + 0.009472988545894623, + 0.004883924964815378, + 0.0075142765417695045, + 0.0076300944201648235, + 0.007609959691762924, + 0.009708899073302746, + 0.010356169193983078, + 0.006987800821661949, + 0.0074273450300097466, + 0.0051069860346615314, + 0.011440725065767765, + 0.018024226650595665, + 0.007281208876520395, + 0.007380714174360037, + 0.01145566999912262, + 0.010869015008211136, + 0.006064127665013075, + 0.007440676912665367, + 0.006396069191396236, + 0.010664028115570545, + 0.00440972950309515, + 0.0052597057074308395, + 0.006999132223427296, + 0.00575087359175086, + 0.0054366737604141235, + 0.004949403926730156, + 0.006393566261976957, + 0.0058330511674284935, + 0.004882034845650196, + 0.004752987530082464, + 0.007696162443608046, + 0.004923362284898758, + 0.006868916563689709, + 0.00564153166487813, + 0.004433298483490944, + 0.0030325890984386206, + 0.0032071585301309824, + 0.0029031240846961737, + 0.001623990829102695, + 0.0011890920577570796, + 0.002178150461986661, + 0.001475082477554679, + 0.006438808981329203, + 0.007165099494159222, + 0.0038544125854969025, + 0.004468416329473257, + 0.004094199277460575, + 0.0032829560805112123, + 0.0026982640847563744, + 0.002279973588883877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00864548608660698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185196444392204 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185196444392204, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185196444392204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3335328698158264, + 0.11392110586166382, + 0.17197419703006744, + 0.07057322561740875, + 0.04431267827749252, + 0.08845759928226471, + 0.0565885454416275, + 0.04107533395290375, + 0.024291876703500748, + 0.05527259036898613 + ], + "credit_map": [ + 0.06770996749401093, + 0.0671561062335968, + 0.05325833335518837, + 0.03725661337375641, + 0.02660195715725422, + 0.03339249640703201, + 0.024222085252404213, + 0.023935304954648018, + 0.013042176142334938, + 0.013353388756513596, + 0.015100902877748013, + 0.015981147065758705, + 0.014324884861707687, + 0.022397829219698906, + 0.011096623726189137, + 0.008624150417745113, + 0.026815952733159065, + 0.02980629913508892, + 0.027630528435111046, + 0.03208278492093086, + 0.016992740333080292, + 0.009138984605669975, + 0.015558023005723953, + 0.013948879204690456, + 0.010793707333505154, + 0.01981445774435997, + 0.005830921698361635, + 0.004485310520976782, + 0.005484629888087511, + 0.006301820278167725, + 0.013703402131795883, + 0.004158976022154093, + 0.0054075149819254875, + 0.005712887737900019, + 0.005124375224113464, + 0.0069380272179841995, + 0.007165744435042143, + 0.006273303646594286, + 0.0043745520524680614, + 0.003316272050142288, + 0.012983236461877823, + 0.015951095148921013, + 0.0100660789757967, + 0.019701628014445305, + 0.008676877245306969, + 0.004688306245952845, + 0.009378851391375065, + 0.007011526729911566, + 0.0091847013682127, + 0.008526972495019436, + 0.006357386242598295, + 0.006668290123343468, + 0.007847757078707218, + 0.0057660355232656, + 0.006285663228482008, + 0.005951741244643927, + 0.005359153263270855, + 0.003961396403610706, + 0.005635203793644905, + 0.004812114872038364, + 0.005144863855093718, + 0.006657889112830162, + 0.004792326129972935, + 0.004712389782071114, + 0.0034664368722587824, + 0.003835659706965089, + 0.003842280711978674, + 0.002470942446961999, + 0.002197770867496729, + 0.00362291163764894, + 0.0027672152500599623, + 0.002088658744469285, + 0.00958788488060236, + 0.014163228683173656, + 0.005784219596534967, + 0.00432659313082695, + 0.006650904659181833, + 0.007292330265045166, + 0.004126517567783594, + 0.0033409087918698788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018185196444392204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334119960665703 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.026334119960665703, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334119960665703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32484081387519836, + 0.09723719209432602, + 0.2053300142288208, + 0.05100790783762932, + 0.03533728048205376, + 0.1308937519788742, + 0.04580621421337128, + 0.07270286232233047, + 0.017574282363057137, + 0.019269663840532303 + ], + "credit_map": [ + 0.07583730667829514, + 0.09046958386898041, + 0.05196348577737808, + 0.025204304605722427, + 0.021771661937236786, + 0.023525936529040337, + 0.019490264356136322, + 0.016578270122408867, + 0.010916819795966148, + 0.01295478641986847, + 0.013818714767694473, + 0.02021469548344612, + 0.01012890413403511, + 0.010555140674114227, + 0.010936246253550053, + 0.007711880374699831, + 0.03486347198486328, + 0.0531754195690155, + 0.031795889139175415, + 0.02598070539534092, + 0.016707802191376686, + 0.013784469105303288, + 0.01811913028359413, + 0.010903109796345234, + 0.00939190573990345, + 0.008016197010874748, + 0.00554983364418149, + 0.004584132228046656, + 0.003589019877836108, + 0.003628402017056942, + 0.012766323052346706, + 0.003482090774923563, + 0.004638145677745342, + 0.004221465904265642, + 0.0046850889921188354, + 0.005862471181899309, + 0.00469724228605628, + 0.0037375527899712324, + 0.0037274088244885206, + 0.0037679008673876524, + 0.025568155571818352, + 0.03701402619481087, + 0.018801826983690262, + 0.015554849989712238, + 0.009093946777284145, + 0.007445302791893482, + 0.010497366078197956, + 0.006918278057128191, + 0.009247192181646824, + 0.007943132892251015, + 0.005330211017280817, + 0.003258091164752841, + 0.0067430539056658745, + 0.005920823663473129, + 0.003916097804903984, + 0.003447611117735505, + 0.009997508488595486, + 0.013813738711178303, + 0.010840288363397121, + 0.011176008731126785, + 0.007252267561852932, + 0.005743392743170261, + 0.008387841284275055, + 0.005491818767040968, + 0.0026932370383292437, + 0.0027345954440534115, + 0.0023067526053637266, + 0.002730129985138774, + 0.0020302559714764357, + 0.001924860174767673, + 0.002027625683695078, + 0.001126824296079576, + 0.0039807711727917194, + 0.003020831849426031, + 0.00292606046423316, + 0.0017487709410488605, + 0.00257951021194458, + 0.0018081340240314603, + 0.0013441131450235844, + 0.0018614735454320908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.026334119960665703 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, close-up, static shot, captured from a slight overhead angle, focuses on a gentle Asian mother, comfortably seated in a vintage wooden armchair. Soft afternoon sunlight slants through a large window, casting a warm golden hue on her long, dark brown hair, pulled into a low bun, and the off-white, turtleneck cashmere sweater she wears. Her face holds a serene, focused expression, her eyes slightly downcast, her gaze completely focused on the work in her hands. Her hands move steadily and deliberately, two smooth bamboo knitting needles twirling between her deft fingers. One needle expertly picks up lavender yarn and delicately threads it through the loop of the other. A half-finished sweater drapes across her legs, its intricate, twisted pattern clearly visible in the light, each fluffy fiber seeming palpable, a vivid, textured expression. Beside her, a wicker basket filled with colorful balls of yarn rests on the ground. The bookshelves and greenery in the background are softly blurred by the shallow depth of field. The entire scene is warm and harmonious, imbued with the warmth of home and the tranquility of time brought by handmade crafts.", + "seed": 8254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_8887_seed_8254_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026513142324984074, + "vlm_logprob": -0.026513142324984074, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.026513142324984074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0239908117800951 + ], + "opposite_logprob_tokens": [ + -4.273990631103516 + ], + "logprob_reward": -0.0239908117800951, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0239908117800951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18537664413452148, + 0.06267212331295013, + 0.12529678642749786, + 0.02905690297484398, + 0.01916750892996788, + 0.23458756506443024, + 0.10855066031217575, + 0.03075512684881687, + 0.1688690334558487, + 0.03566765412688255 + ], + "credit_map": [ + 0.019944116473197937, + 0.01783694699406624, + 0.019968658685684204, + 0.025349104776978493, + 0.027086082845926285, + 0.024941926822066307, + 0.0164363794028759, + 0.03381342068314552, + 0.006128974724560976, + 0.007534396834671497, + 0.004281440284103155, + 0.007838978432118893, + 0.0091725904494524, + 0.008824697695672512, + 0.009038676507771015, + 0.009852365590631962, + 0.011076058261096478, + 0.011701957322657108, + 0.010351751931011677, + 0.008941559121012688, + 0.013632253743708134, + 0.03881930932402611, + 0.011339716613292694, + 0.01943417824804783, + 0.0031708301976323128, + 0.002182386117056012, + 0.002782145282253623, + 0.003405206138268113, + 0.00390644883736968, + 0.003345289034768939, + 0.0035656848922371864, + 0.006698912475258112, + 0.0020722670014947653, + 0.002010299591347575, + 0.0016525472747161984, + 0.0021402514539659023, + 0.0024875991512089968, + 0.0025829400401562452, + 0.0026839477941393852, + 0.0035376562736928463, + 0.02555873803794384, + 0.020537808537483215, + 0.010792742483317852, + 0.012066309340298176, + 0.014361605979502201, + 0.10070217400789261, + 0.017975429072976112, + 0.032592758536338806, + 0.009017358534038067, + 0.007050752639770508, + 0.006257088854908943, + 0.009904460981488228, + 0.01444255281239748, + 0.02148381620645523, + 0.009079590439796448, + 0.03131503984332085, + 0.003279939293861389, + 0.004389123059809208, + 0.0027961921878159046, + 0.0032514689955860376, + 0.003910875413566828, + 0.0031767257023602724, + 0.004959803074598312, + 0.004990996792912483, + 0.0218157097697258, + 0.011305680498480797, + 0.009008887223899364, + 0.0124394865706563, + 0.006822450552135706, + 0.026288339868187904, + 0.008855974301695824, + 0.07233250141143799, + 0.004266192205250263, + 0.0025005817878991365, + 0.0027718879282474518, + 0.0032655661925673485, + 0.004940251354128122, + 0.0054978374391794205, + 0.004243175964802504, + 0.008182159624993801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0239908117800951 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381079725921154 + ], + "opposite_logprob_tokens": [ + -3.2881078720092773 + ], + "logprob_reward": -0.0381079725921154, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381079725921154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14221195876598358, + 0.0373653918504715, + 0.3235780596733093, + 0.03490360826253891, + 0.02395300380885601, + 0.08295277506113052, + 0.022664781659841537, + 0.022539731115102768, + 0.2935032844543457, + 0.01632738672196865 + ], + "credit_map": [ + 0.01563284359872341, + 0.014813097193837166, + 0.012898395769298077, + 0.018114980310201645, + 0.02040693536400795, + 0.021741250529885292, + 0.017385056242346764, + 0.021219391375780106, + 0.005101425107568502, + 0.006233368534594774, + 0.003947942052036524, + 0.003394513390958309, + 0.006025009322911501, + 0.0035161334089934826, + 0.0054130954667925835, + 0.003733905265107751, + 0.03336309269070625, + 0.030620940029621124, + 0.015497753396630287, + 0.01646570861339569, + 0.01943853497505188, + 0.1359134018421173, + 0.02616094797849655, + 0.04611768573522568, + 0.00529071968048811, + 0.004442696925252676, + 0.004701230209320784, + 0.003942277282476425, + 0.004271759185940027, + 0.003327132435515523, + 0.005318803712725639, + 0.003608988132327795, + 0.003229390364140272, + 0.0028916785959154367, + 0.0028582685627043247, + 0.003442073706537485, + 0.0026858607307076454, + 0.003286799183115363, + 0.0029119979590177536, + 0.002646935172379017, + 0.012031192891299725, + 0.007471142802387476, + 0.0052643828094005585, + 0.007186536677181721, + 0.005749147851020098, + 0.03046996518969536, + 0.007322708144783974, + 0.0074576945044100285, + 0.0022990459110587835, + 0.0025007715448737144, + 0.002024003304541111, + 0.002787494333460927, + 0.002375455806031823, + 0.0034708131570369005, + 0.002742102835327387, + 0.004465096164494753, + 0.0038119861856102943, + 0.0024842240381985903, + 0.0037771195638924837, + 0.0020989032927900553, + 0.0027023826260119677, + 0.002385905245319009, + 0.0028082067146897316, + 0.0024710027500987053, + 0.040325015783309937, + 0.019495319575071335, + 0.014874914661049843, + 0.0201549269258976, + 0.011311128735542297, + 0.046227723360061646, + 0.014531625434756279, + 0.12658263742923737, + 0.0016074668383225799, + 0.001467659487389028, + 0.0016591843450441957, + 0.002341157291084528, + 0.002395038725808263, + 0.0017864878755062819, + 0.002676895121112466, + 0.002393497386947274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0381079725921154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198708347976208 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.014198708347976208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198708347976208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15162347257137299, + 0.05914749950170517, + 0.11644277721643448, + 0.026611262932419777, + 0.019330637529492378, + 0.3626290261745453, + 0.04617305099964142, + 0.014951191842556, + 0.18082614243030548, + 0.022264933213591576 + ], + "credit_map": [ + 0.014848057180643082, + 0.014865594916045666, + 0.013499267399311066, + 0.022825025022029877, + 0.02794129028916359, + 0.01525337528437376, + 0.02225317992269993, + 0.020137691870331764, + 0.004612057935446501, + 0.007378027308732271, + 0.005308585241436958, + 0.008147926069796085, + 0.010908545926213264, + 0.006351592019200325, + 0.007739437278360128, + 0.008701327256858349, + 0.013805109076201916, + 0.011175683699548244, + 0.00951123982667923, + 0.011684511788189411, + 0.01081465557217598, + 0.036155201494693756, + 0.013011965900659561, + 0.01028440985828638, + 0.002291781362146139, + 0.0030311921145766973, + 0.0017861269880086184, + 0.0032411767169833183, + 0.004222410265356302, + 0.0031376583501696587, + 0.0034351092763245106, + 0.005465805996209383, + 0.002494235523045063, + 0.002531108446419239, + 0.0017547995084896684, + 0.0021773590706288815, + 0.0035023426171392202, + 0.001831885427236557, + 0.002610268071293831, + 0.00242863898165524, + 0.039210252463817596, + 0.03214845806360245, + 0.0173978041857481, + 0.018937986344099045, + 0.022967802360653877, + 0.1570655256509781, + 0.02785966545343399, + 0.04704153537750244, + 0.0036596027202904224, + 0.0037204157561063766, + 0.0034825564362108707, + 0.006198457442224026, + 0.007341144606471062, + 0.0075981332920491695, + 0.008261341601610184, + 0.005911404732614756, + 0.001410335418768227, + 0.0025885007344186306, + 0.0014912408078089356, + 0.0020360571797937155, + 0.0017873223405331373, + 0.0019547867123037577, + 0.002110601868480444, + 0.0015723470132797956, + 0.024196559563279152, + 0.012071826495230198, + 0.00828544981777668, + 0.012065602466464043, + 0.0071640596725046635, + 0.030029578134417534, + 0.009879816323518753, + 0.07713324576616287, + 0.0022240530233830214, + 0.001996573992073536, + 0.002032475546002388, + 0.002818929497152567, + 0.004612554330378771, + 0.002235035179182887, + 0.0034349150955677032, + 0.002910395385697484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.014198708347976208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818273961544037 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818273961544037, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818273961544037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10408059507608414, + 0.03995253145694733, + 0.40007883310317993, + 0.03056415170431137, + 0.017547501251101494, + 0.08674988895654678, + 0.03905026242136955, + 0.015993757173419, + 0.25111836194992065, + 0.014864067547023296 + ], + "credit_map": [ + 0.012792576104402542, + 0.010264726355671883, + 0.012148031033575535, + 0.012876181863248348, + 0.01236717402935028, + 0.013471269980072975, + 0.010803842917084694, + 0.019356807693839073, + 0.0046823276206851006, + 0.003028786275535822, + 0.005089703947305679, + 0.004172118846327066, + 0.005794658791273832, + 0.0052827587351202965, + 0.006851303856819868, + 0.00505087710916996, + 0.05119524151086807, + 0.03917975723743439, + 0.023208269849419594, + 0.02315615676343441, + 0.024145424365997314, + 0.17035436630249023, + 0.03438359871506691, + 0.034456051886081696, + 0.0031213616020977497, + 0.002657766919583082, + 0.0032178114634007215, + 0.0033333278261125088, + 0.005007133819162846, + 0.0038264147005975246, + 0.004400220233947039, + 0.005000116303563118, + 0.0024761692620813847, + 0.002927349414676428, + 0.002059968188405037, + 0.0017346094828099012, + 0.002679218305274844, + 0.0016541349468752742, + 0.0020002422388643026, + 0.0020158104598522186, + 0.013206115923821926, + 0.0074934219010174274, + 0.006777087226510048, + 0.00801649410277605, + 0.005830711219459772, + 0.02832844667136669, + 0.00742726307362318, + 0.009670360945165157, + 0.004086278844624758, + 0.00239905109629035, + 0.0022499789483845234, + 0.0037549512926489115, + 0.0027352962642908096, + 0.006121320649981499, + 0.004234377760440111, + 0.013469012454152107, + 0.0019494552398100495, + 0.0019043459324166179, + 0.0015879377024248242, + 0.0020147820468991995, + 0.00239639263600111, + 0.0019029006361961365, + 0.00262467167340219, + 0.0016132727032527328, + 0.033136025071144104, + 0.01658082939684391, + 0.01162002980709076, + 0.016896450892090797, + 0.009723545052111149, + 0.042435962706804276, + 0.012853947468101978, + 0.10787158459424973, + 0.0012069252552464604, + 0.0012671950971707702, + 0.00112889613956213, + 0.0014547386672347784, + 0.002618337282910943, + 0.0019619581289589405, + 0.003372186329215765, + 0.001853831810876727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.01818273961544037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808547928929329 + ], + "opposite_logprob_tokens": [ + -3.288085460662842 + ], + "logprob_reward": -0.03808547928929329, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808547928929329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11240483075380325, + 0.08282455801963806, + 0.18706382811069489, + 0.0488128624856472, + 0.04391688108444214, + 0.039633408188819885, + 0.04546486586332321, + 0.06950867176055908, + 0.3204346001148224, + 0.04993554204702377 + ], + "credit_map": [ + 0.009989886544644833, + 0.014116311445832253, + 0.01030847430229187, + 0.012917623855173588, + 0.01480541005730629, + 0.014855553396046162, + 0.017913155257701874, + 0.01749841682612896, + 0.007523271255195141, + 0.0074721090495586395, + 0.009332256391644478, + 0.012073015794157982, + 0.01626194640994072, + 0.006497923284769058, + 0.01289817038923502, + 0.010765868239104748, + 0.0248732790350914, + 0.01806037314236164, + 0.013270577415823936, + 0.01060406118631363, + 0.01192980632185936, + 0.07578270137310028, + 0.01636112853884697, + 0.01618190109729767, + 0.004253271967172623, + 0.0045572612434625626, + 0.004816547501832247, + 0.006359271239489317, + 0.008006374351680279, + 0.0060554794035851955, + 0.007964062504470348, + 0.006800593342632055, + 0.0035072392784059048, + 0.00389958918094635, + 0.0044587296433746815, + 0.0055301920510828495, + 0.0067252120934426785, + 0.004331293981522322, + 0.009107744321227074, + 0.006356879137456417, + 0.004638349171727896, + 0.0031132083386182785, + 0.0035263688769191504, + 0.003514222800731659, + 0.0038908254355192184, + 0.010166255757212639, + 0.0043267784640192986, + 0.006457400042563677, + 0.005153676029294729, + 0.002668767934665084, + 0.0026284889318048954, + 0.004780453629791737, + 0.00297465524636209, + 0.007643999066203833, + 0.0035179853439331055, + 0.01609683595597744, + 0.005164985544979572, + 0.005005532409995794, + 0.007237582933157682, + 0.007995364256203175, + 0.012435548938810825, + 0.0070884800516068935, + 0.01290184911340475, + 0.011679326184093952, + 0.043792687356472015, + 0.021441064774990082, + 0.015368063934147358, + 0.02191716805100441, + 0.012274396605789661, + 0.05194460228085518, + 0.01581715978682041, + 0.13787944614887238, + 0.0035591137129813433, + 0.004117099568247795, + 0.004735123831778765, + 0.006771520245820284, + 0.0069036707282066345, + 0.004747467115521431, + 0.010987642221152782, + 0.008113907650113106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.03808547928929329 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, close-up, static shot, captured from a slight overhead angle, focuses on a gentle Asian mother, comfortably seated in a vintage wooden armchair. Soft afternoon sunlight slants through a large window, casting a warm golden hue on her long, dark brown hair, pulled into a low bun, and the off-white, turtleneck cashmere sweater she wears. Her face holds a serene, focused expression, her eyes slightly downcast, her gaze completely focused on the work in her hands. Her hands move steadily and deliberately, two smooth bamboo knitting needles twirling between her deft fingers. One needle expertly picks up lavender yarn and delicately threads it through the loop of the other. A half-finished sweater drapes across her legs, its intricate, twisted pattern clearly visible in the light, each fluffy fiber seeming palpable, a vivid, textured expression. Beside her, a wicker basket filled with colorful balls of yarn rests on the ground. The bookshelves and greenery in the background are softly blurred by the shallow depth of field. The entire scene is warm and harmonious, imbued with the warmth of home and the tranquility of time brought by handmade crafts.", + "seed": 8255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_10/index_8887_seed_8255_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.031779732927680014, + "vlm_logprob": -0.031779732927680014, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.031779732927680014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017823997884988785 + ], + "opposite_logprob_tokens": [ + -4.267824172973633 + ], + "logprob_reward": -0.017823997884988785, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017823997884988785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5523570775985718, + 0.05476491525769234, + 0.04001195728778839, + 0.026819845661520958, + 0.04226642847061157, + 0.01641424000263214, + 0.13885246217250824, + 0.02452876605093479, + 0.08916766196489334, + 0.014816630631685257 + ], + "credit_map": [ + 0.04457555711269379, + 0.12905775010585785, + 0.21657782793045044, + 0.06905873119831085, + 0.01638312265276909, + 0.03499103710055351, + 0.01995282806456089, + 0.021760286763310432, + 0.006019790656864643, + 0.011397208087146282, + 0.014261336997151375, + 0.005241608712822199, + 0.0026126757729798555, + 0.004428114276379347, + 0.003456922946497798, + 0.00734725734218955, + 0.004322010558098555, + 0.00782588217407465, + 0.012493349611759186, + 0.002159017836675048, + 0.004035107791423798, + 0.0035367959644645452, + 0.0026330873370170593, + 0.0030067050829529762, + 0.0025202909018844366, + 0.0053999051451683044, + 0.009063941426575184, + 0.0015596948796883225, + 0.002116224728524685, + 0.0018121969187632203, + 0.0016869985265657306, + 0.002660593017935753, + 0.0035753874108195305, + 0.009590791538357735, + 0.016141653060913086, + 0.0020592284854501486, + 0.002210322767496109, + 0.003619271330535412, + 0.0019210008904337883, + 0.0031487715896219015, + 0.0016162926331162453, + 0.003400348359718919, + 0.005263399798423052, + 0.0011396505869925022, + 0.0017698114970698953, + 0.0012407226022332907, + 0.0009396104724146426, + 0.00104440504219383, + 0.011987806297838688, + 0.033796537667512894, + 0.05569463595747948, + 0.0066893394105136395, + 0.0048628696240484715, + 0.012818079441785812, + 0.00568995950743556, + 0.007313231471925974, + 0.0032406977843493223, + 0.0052716899663209915, + 0.007982509210705757, + 0.0018187470268458128, + 0.0014972499338909984, + 0.001842763740569353, + 0.0015763717237859964, + 0.0012987361988052726, + 0.007946310564875603, + 0.021750599145889282, + 0.037311941385269165, + 0.0043268706649541855, + 0.0029312935657799244, + 0.007264446932822466, + 0.003391735488548875, + 0.004244462586939335, + 0.0015131391119211912, + 0.00254878681153059, + 0.004064280539751053, + 0.0010736361145973206, + 0.0015242677181959152, + 0.0013856147415935993, + 0.0010823026532307267, + 0.0016246025916188955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.017823997884988785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495452135801315 + ], + "opposite_logprob_tokens": [ + -2.9299545288085938 + ], + "logprob_reward": -0.05495452135801315, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495452135801315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42630842328071594, + 0.0683363825082779, + 0.05079498887062073, + 0.032521724700927734, + 0.013601411134004593, + 0.022065741941332817, + 0.16588762402534485, + 0.021799784153699875, + 0.15443842113018036, + 0.04424552991986275 + ], + "credit_map": [ + 0.03467320278286934, + 0.0987413227558136, + 0.1671820431947708, + 0.05285228416323662, + 0.013144448399543762, + 0.027320558205246925, + 0.015929222106933594, + 0.0164653267711401, + 0.00680298637598753, + 0.014259184710681438, + 0.02492883801460266, + 0.003942934796214104, + 0.0027473275549709797, + 0.00539249274879694, + 0.00294266059063375, + 0.007319947239011526, + 0.0113730663433671, + 0.00983471516519785, + 0.007303278427571058, + 0.0033992421813309193, + 0.005436109844595194, + 0.004538863431662321, + 0.0038344841450452805, + 0.005075224209576845, + 0.0034288368187844753, + 0.006178684066981077, + 0.010966329835355282, + 0.001927149947732687, + 0.002121010795235634, + 0.002076870994642377, + 0.0018375783693045378, + 0.003985260147601366, + 0.001156579819507897, + 0.0025268869940191507, + 0.003547790925949812, + 0.0011029366869479418, + 0.0011921246768906713, + 0.0011853219475597143, + 0.0015397073002532125, + 0.0013500626664608717, + 0.0025432708207517862, + 0.004480105824768543, + 0.006574998609721661, + 0.001420126878656447, + 0.0018192838178947568, + 0.0015305726556107402, + 0.001965584931895137, + 0.0017317971214652061, + 0.014506815932691097, + 0.040063291788101196, + 0.06644243746995926, + 0.0077878562733531, + 0.005659393034875393, + 0.014772710390388966, + 0.007149826269596815, + 0.009505289606750011, + 0.003108232282102108, + 0.00430942140519619, + 0.0065458351746201515, + 0.0016222677659243345, + 0.0016252020141109824, + 0.0016397059662267566, + 0.0017161114374175668, + 0.0012330074096098542, + 0.013809344731271267, + 0.03781796991825104, + 0.06439527869224548, + 0.007559164892882109, + 0.00513667706400156, + 0.012604561634361744, + 0.005955069325864315, + 0.007160347420722246, + 0.004618331789970398, + 0.009075935930013657, + 0.016506116837263107, + 0.002397701609879732, + 0.001734269899316132, + 0.0042885285802185535, + 0.0024236137978732586, + 0.003201030194759369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.05495452135801315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196004718542099 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196004718542099, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196004718542099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35429906845092773, + 0.1872335970401764, + 0.08729838579893112, + 0.11092830449342728, + 0.06834599375724792, + 0.035913970321416855, + 0.061606794595718384, + 0.018817653879523277, + 0.050172265619039536, + 0.025383971631526947 + ], + "credit_map": [ + 0.028507107868790627, + 0.08292404562234879, + 0.13983754813671112, + 0.042593520134687424, + 0.011301314458251, + 0.02085961028933525, + 0.014068740420043468, + 0.014207178726792336, + 0.017378099262714386, + 0.03485996648669243, + 0.06716493517160416, + 0.01177687756717205, + 0.006729824468493462, + 0.012349332682788372, + 0.009705689735710621, + 0.027268873527646065, + 0.00860485527664423, + 0.015808718279004097, + 0.026521271094679832, + 0.006180632393807173, + 0.009565907530486584, + 0.007301281206309795, + 0.006555923726409674, + 0.006759798154234886, + 0.00984103325754404, + 0.02468693256378174, + 0.042455725371837616, + 0.008628763258457184, + 0.005210557486861944, + 0.008545649237930775, + 0.005322496872395277, + 0.006237153895199299, + 0.0067085521295666695, + 0.01404937356710434, + 0.02384199947118759, + 0.004067142494022846, + 0.004181161988526583, + 0.0057494365610182285, + 0.003795960918068886, + 0.005952365696430206, + 0.004742057994008064, + 0.004043368622660637, + 0.005576188210397959, + 0.003949212376028299, + 0.007214721292257309, + 0.0031671093311160803, + 0.003920780960470438, + 0.0033005326986312866, + 0.005655987653881311, + 0.014784571714699268, + 0.024024447426199913, + 0.003042068099603057, + 0.0025022353511303663, + 0.005792825948446989, + 0.002753380686044693, + 0.0030512758530676365, + 0.003425939939916134, + 0.003365134820342064, + 0.004440500866621733, + 0.001403304748237133, + 0.0019194131018593907, + 0.0013970857253298163, + 0.0015541523462161422, + 0.0013121223310008645, + 0.004939327482134104, + 0.01172937173396349, + 0.02008526213467121, + 0.0027205455116927624, + 0.002311795251443982, + 0.0037849738728255033, + 0.001876775873824954, + 0.0027242139913141727, + 0.004917369224131107, + 0.003969787620007992, + 0.005934431683272123, + 0.0015980489552021027, + 0.001675957813858986, + 0.002120099263265729, + 0.002200526650995016, + 0.0029677480924874544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014196004718542099 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284463211894035 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284463211894035, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284463211894035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5442597270011902, + 0.032423149794340134, + 0.11349125951528549, + 0.05139186605811119, + 0.03736378625035286, + 0.01866474375128746, + 0.01378237921744585, + 0.02154708094894886, + 0.14799082279205322, + 0.019085193052887917 + ], + "credit_map": [ + 0.04619767144322395, + 0.1253720074892044, + 0.21031804382801056, + 0.06605442613363266, + 0.017061149701476097, + 0.0359121710062027, + 0.021328717470169067, + 0.0220155231654644, + 0.0036674768198281527, + 0.0073684402741491795, + 0.00947602465748787, + 0.00321500888094306, + 0.0022545638494193554, + 0.0020233422983437777, + 0.002316083060577512, + 0.002102210186421871, + 0.019573885947465897, + 0.018208453431725502, + 0.026793384924530983, + 0.007380201481282711, + 0.010725975967943668, + 0.008785384707152843, + 0.011244135908782482, + 0.010779829695820808, + 0.00480604125186801, + 0.010093647986650467, + 0.0171003770083189, + 0.0031982383225113153, + 0.0042086453177034855, + 0.003488631220534444, + 0.0036630635149776936, + 0.004833220038563013, + 0.003195915138348937, + 0.0072219097055494785, + 0.012383255176246166, + 0.0021488249767571688, + 0.0030917718540877104, + 0.0030532486271113157, + 0.0024957936257123947, + 0.0037730673793703318, + 0.0018632240826264024, + 0.0032130139879882336, + 0.0046144514344632626, + 0.0016366307390853763, + 0.0022265813313424587, + 0.0022101844660937786, + 0.0015219171764329076, + 0.0013787399511784315, + 0.0015230874996632338, + 0.002515366766601801, + 0.002883418696001172, + 0.0009732443140819669, + 0.0014167699264362454, + 0.0014398545026779175, + 0.0012592371786013246, + 0.0017714003333821893, + 0.002800630871206522, + 0.00436190003529191, + 0.006749493535608053, + 0.0015641440404579043, + 0.0015095408307388425, + 0.001576319569721818, + 0.0017035984201356769, + 0.0012814535293728113, + 0.013467024080455303, + 0.03611418232321739, + 0.0615464486181736, + 0.007282178848981857, + 0.0048714387230575085, + 0.012022274546325207, + 0.005728311836719513, + 0.006958962883800268, + 0.003582984209060669, + 0.0023513182532042265, + 0.002948173088952899, + 0.0014889020239934325, + 0.002177651273086667, + 0.0019504382507875562, + 0.0017196179833263159, + 0.002866107737645507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.023284463211894035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of an Asian mother seated in a vintage wooden armchair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera captured from a slight overhead angle in a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing an off-white turtleneck cashmere sweater with her dark brown hair pulled into a low bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smooth bamboo knitting needles visible twirling between her fingers while she threads lavender yarn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a half-finished sweater with an intricate twisted pattern draped across her legs next to a wicker basket of colorful yarn balls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048639677464962006 + ], + "opposite_logprob_tokens": [ + -3.0486397743225098 + ], + "logprob_reward": -0.048639677464962006, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048639677464962006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12302535027265549, + 0.21755117177963257, + 0.04054790735244751, + 0.048140738159418106, + 0.12827426195144653, + 0.021435946226119995, + 0.18012480437755585, + 0.05319578945636749, + 0.16247573494911194, + 0.025228366255760193 + ], + "credit_map": [ + 0.0207770224660635, + 0.015878403559327126, + 0.021112030372023582, + 0.01854744367301464, + 0.009839482605457306, + 0.008429649285972118, + 0.014870403334498405, + 0.013570915907621384, + 0.021560722962021828, + 0.037314318120479584, + 0.07243851572275162, + 0.015020539052784443, + 0.008229312486946583, + 0.017893845215439796, + 0.01153500284999609, + 0.03355890139937401, + 0.003923176787793636, + 0.0042669991962611675, + 0.00798071175813675, + 0.004632212687283754, + 0.005084396339952946, + 0.003565702587366104, + 0.00633500749245286, + 0.004759697709232569, + 0.005942699499428272, + 0.008180282078683376, + 0.012427517212927341, + 0.004921407904475927, + 0.00365754053927958, + 0.0042779468931257725, + 0.004772844258695841, + 0.003960499074310064, + 0.010268027894198895, + 0.02966470643877983, + 0.047964613884687424, + 0.006789360195398331, + 0.006733637303113937, + 0.010595357976853848, + 0.008480140008032322, + 0.007778409868478775, + 0.002353584161028266, + 0.0028510000556707382, + 0.005254937335848808, + 0.0019614589400589466, + 0.0024606059305369854, + 0.002892428543418646, + 0.0020669829100370407, + 0.0015949488151818514, + 0.015620331279933453, + 0.04330265149474144, + 0.06704062223434448, + 0.008168889209628105, + 0.007921493612229824, + 0.017005087807774544, + 0.010561327449977398, + 0.010504406876862049, + 0.0066487742587924, + 0.010746069252490997, + 0.017491862177848816, + 0.004067809786647558, + 0.003480917075648904, + 0.004151756409555674, + 0.003509128000587225, + 0.003099472029134631, + 0.014605791307985783, + 0.03943045064806938, + 0.06750994920730591, + 0.008134621195495129, + 0.005552790127694607, + 0.01342572271823883, + 0.006390134338289499, + 0.007426275406032801, + 0.0040941075421869755, + 0.0023430874571204185, + 0.004121469333767891, + 0.0020332816056907177, + 0.0032355114817619324, + 0.0016927440883591771, + 0.004134414717555046, + 0.003573748981580138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.048639677464962006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_tensor([1852, 1852])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_tensor([1852, 1852])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ed4c5ee446f8fa95e600b2606af83a193277e59d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_tensor([1852, 1852])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene begins with a high-angle, long shot, slowly and deliberately zooming in, capturing a winding, dark gray paved mountain road through a dense forest ablaze with fiery reds and brilliant golds of autumn leaves. The low, golden rays of the dawn sun shimmer across the distant ridges, casting long, dramatic shadows and bathing the crisp air in a warm, ethereal glow. At the center of this vast, serene landscape lies the solitary figure of a petite, young Southeast Asian man with short, unkempt black hair, dressed simply in a gray sweatshirt and dark blue jeans. As the camera zooms in, his concentrated expression becomes clear, tinged with a hint of nervous worry. He rides on a worn skateboard with bright orange wheels, rolling forward at a slow, hesitant pace. His body stiffens and sways, his center of gravity shifting precariously. In an effort to maintain stability, he awkwardly flings his arms to the sides in an uncoordinated circle, his hands outstretched as if trying to grasp at the air itself. The entire scene is a tense and poignant ballet of effort, contrasting the men’s awkward, determined struggle with the quiet, breathtaking beauty of a golden autumn morning.", + "seed": 5755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_1852_seed_5755_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0769277647137642, + "vlm_logprob": -0.0769277647137642, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.0769277647137642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015731165185570717 + ], + "opposite_logprob_tokens": [ + -5.140731334686279 + ], + "logprob_reward": -0.015731165185570717, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015731165185570717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24450461566448212, + 0.15385432541370392, + 0.05851489305496216, + 0.10824255645275116, + 0.052758730947971344, + 0.07592980563640594, + 0.06805843114852905, + 0.09329350292682648, + 0.08509532362222672, + 0.059747859835624695 + ], + "credit_map": [ + 0.0167513620108366, + 0.02030797116458416, + 0.02624150551855564, + 0.045056313276290894, + 0.018925761803984642, + 0.02340027689933777, + 0.021396929398179054, + 0.0724244937300682, + 0.021397888660430908, + 0.017779074609279633, + 0.020937832072377205, + 0.021187836304306984, + 0.01479827519506216, + 0.022119002416729927, + 0.008756544440984726, + 0.026877878233790398, + 0.0035403775982558727, + 0.0038798160385340452, + 0.007646240293979645, + 0.013564768247306347, + 0.005354244727641344, + 0.005078331101685762, + 0.01382932998239994, + 0.005621792282909155, + 0.007890736684203148, + 0.008204803802073002, + 0.012872640043497086, + 0.028035087510943413, + 0.01035834476351738, + 0.011929030530154705, + 0.01987018622457981, + 0.009081730619072914, + 0.006144002545624971, + 0.004115238785743713, + 0.00833494495600462, + 0.006956443190574646, + 0.004829005803912878, + 0.006609211675822735, + 0.010533636435866356, + 0.005236244760453701, + 0.005563432350754738, + 0.013288430869579315, + 0.010167747735977173, + 0.006563500966876745, + 0.01281910389661789, + 0.01122775487601757, + 0.010108157992362976, + 0.006191672291606665, + 0.004597978666424751, + 0.019596336409449577, + 0.005385840777307749, + 0.008331051096320152, + 0.007911455817520618, + 0.010059325024485588, + 0.005488505121320486, + 0.006687941495329142, + 0.010622652247548103, + 0.014760522171854973, + 0.01569899171590805, + 0.009250221773982048, + 0.010537698864936829, + 0.011010708287358284, + 0.011329065077006817, + 0.010083644650876522, + 0.006992312613874674, + 0.007605127990245819, + 0.01183132454752922, + 0.009794746525585651, + 0.011469529010355473, + 0.014138760976493359, + 0.01642114855349064, + 0.006842372007668018, + 0.004124139901250601, + 0.0077192941680550575, + 0.0072499411180615425, + 0.007706091273576021, + 0.008137539960443974, + 0.0078011793084442616, + 0.012270087376236916, + 0.004739589523524046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.015731165185570717 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620555654168129 + ], + "opposite_logprob_tokens": [ + -2.8120555877685547 + ], + "logprob_reward": -0.0620555654168129, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620555654168129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4503205716609955, + 0.07726836204528809, + 0.051066603511571884, + 0.08378122746944427, + 0.0409591943025589, + 0.03925337269902229, + 0.12096657603979111, + 0.02840724028646946, + 0.07471899688243866, + 0.033257871866226196 + ], + "credit_map": [ + 0.03262541443109512, + 0.06217944249510765, + 0.03445622697472572, + 0.05170436576008797, + 0.06450635194778442, + 0.07526738941669464, + 0.04693171754479408, + 0.08264963328838348, + 0.009751667268574238, + 0.009889476001262665, + 0.008855761028826237, + 0.007406201679259539, + 0.00904484186321497, + 0.013164255768060684, + 0.007826975546777248, + 0.01132917869836092, + 0.0034608833957463503, + 0.007194606587290764, + 0.00608854740858078, + 0.008200262673199177, + 0.006581996101886034, + 0.006098989862948656, + 0.005069042090326548, + 0.008372271433472633, + 0.006930791772902012, + 0.010109948925673962, + 0.009560281410813332, + 0.016782226040959358, + 0.00957776140421629, + 0.010898740030825138, + 0.012010429985821247, + 0.00791104044765234, + 0.0026475980412214994, + 0.003917773719877005, + 0.004914935678243637, + 0.00811272207647562, + 0.004639999009668827, + 0.005196709185838699, + 0.007979750633239746, + 0.0035497029311954975, + 0.00256161205470562, + 0.0060391840524971485, + 0.004097475204616785, + 0.004865966737270355, + 0.008107517845928669, + 0.0058867414481937885, + 0.004383701365441084, + 0.0033111688680946827, + 0.00942278653383255, + 0.027229279279708862, + 0.012330064550042152, + 0.01552781742066145, + 0.01822015270590782, + 0.016594305634498596, + 0.010430664755403996, + 0.011211492121219635, + 0.002287314273416996, + 0.003978712484240532, + 0.003929033875465393, + 0.003655404085293412, + 0.003385402960702777, + 0.0044637564569711685, + 0.004691232927143574, + 0.002016380662098527, + 0.005602130200713873, + 0.005581021774560213, + 0.009786603972315788, + 0.013912464492022991, + 0.009392911568284035, + 0.0119552006945014, + 0.01230105385184288, + 0.006187609862536192, + 0.0025611501187086105, + 0.0027825268916785717, + 0.004424059763550758, + 0.004542217124253511, + 0.004628175403922796, + 0.005215517245233059, + 0.006114914081990719, + 0.0029893103055655956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0620555654168129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557009756565094 + ], + "opposite_logprob_tokens": [ + -1.6005700826644897 + ], + "logprob_reward": -0.22557009756565094, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557009756565094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09043603390455246, + 0.035935722291469574, + 0.11958811432123184, + 0.05510581657290459, + 0.06151384115219116, + 0.049909163266420364, + 0.18452495336532593, + 0.23659607768058777, + 0.13377133011817932, + 0.03261886537075043 + ], + "credit_map": [ + 0.012887921184301376, + 0.008956467732787132, + 0.010691202245652676, + 0.012265491299331188, + 0.013473969884216785, + 0.010196510702371597, + 0.009409137070178986, + 0.012555344961583614, + 0.002774482360109687, + 0.004626457113772631, + 0.0027276622131466866, + 0.00513062346726656, + 0.003209698013961315, + 0.005567294079810381, + 0.004492904059588909, + 0.007406606338918209, + 0.007096920628100634, + 0.007918758317828178, + 0.012771508656442165, + 0.029660388827323914, + 0.01231637503951788, + 0.01194641925394535, + 0.023794637992978096, + 0.014083122834563255, + 0.00427021412178874, + 0.005603230558335781, + 0.006630065850913525, + 0.0076559181325137615, + 0.007624520920217037, + 0.008717191405594349, + 0.009197767823934555, + 0.0054069142788648605, + 0.00502019003033638, + 0.004580553621053696, + 0.008881605230271816, + 0.009155090898275375, + 0.005995358340442181, + 0.007775068748742342, + 0.013685092329978943, + 0.0064208898693323135, + 0.0037531545385718346, + 0.010753167793154716, + 0.005483801010996103, + 0.005700200330466032, + 0.005699558649212122, + 0.006809033919125795, + 0.008424508385360241, + 0.003285745857283473, + 0.01300695538520813, + 0.04180475324392319, + 0.01720031164586544, + 0.018314464017748833, + 0.03176251798868179, + 0.025152454152703285, + 0.017246173694729805, + 0.02003733068704605, + 0.029797665774822235, + 0.03131278231739998, + 0.039146747440099716, + 0.024553172290325165, + 0.02346225269138813, + 0.02789207547903061, + 0.031828951090574265, + 0.02860245481133461, + 0.00898824818432331, + 0.017090074717998505, + 0.016490954905748367, + 0.02738633006811142, + 0.011238116770982742, + 0.024981848895549774, + 0.017122602090239525, + 0.010473188944160938, + 0.0024893840309232473, + 0.0033985257614403963, + 0.003303295001387596, + 0.004105083178728819, + 0.003908859100192785, + 0.005893964786082506, + 0.006769516970962286, + 0.0027502384036779404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.22557009756565094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346543803811073 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.026346543803811073, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346543803811073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08162219822406769, + 0.04593140259385109, + 0.05787743255496025, + 0.09563180804252625, + 0.08652590215206146, + 0.057384539395570755, + 0.19830730557441711, + 0.09250877797603607, + 0.23893961310386658, + 0.045271094888448715 + ], + "credit_map": [ + 0.006521931383758783, + 0.00934726931154728, + 0.0074401660822331905, + 0.013946566730737686, + 0.01336131151765585, + 0.012082379311323166, + 0.007626873441040516, + 0.011295692063868046, + 0.0040242914110422134, + 0.006853443570435047, + 0.0032286199275404215, + 0.007040671072900295, + 0.004074158612638712, + 0.008251710794866085, + 0.0035485404077917337, + 0.008909964002668858, + 0.0036519502755254507, + 0.008776101283729076, + 0.006592778954654932, + 0.009441922418773174, + 0.006795071996748447, + 0.007230186369270086, + 0.009031035006046295, + 0.006358383689075708, + 0.007407547906041145, + 0.016333280131220818, + 0.01088491827249527, + 0.016908956691622734, + 0.011712861247360706, + 0.014939418993890285, + 0.010566850192844868, + 0.006877967156469822, + 0.0057845949195325375, + 0.007915077731013298, + 0.011171320453286171, + 0.011094002053141594, + 0.010635984130203724, + 0.011035146191716194, + 0.021600710228085518, + 0.00728905713185668, + 0.003084419993683696, + 0.007841788232326508, + 0.005766389425843954, + 0.00580653827637434, + 0.007884487509727478, + 0.008310924284160137, + 0.013832555152475834, + 0.004857437219470739, + 0.013056564144790173, + 0.028043802827596664, + 0.017285499721765518, + 0.0482395701110363, + 0.020547138527035713, + 0.03268127515912056, + 0.025118457153439522, + 0.013334969989955425, + 0.008712136186659336, + 0.010082975961267948, + 0.011375652626156807, + 0.01074940524995327, + 0.010949148796498775, + 0.012379183433949947, + 0.018979912623763084, + 0.009280355647206306, + 0.02059905230998993, + 0.042047761380672455, + 0.03149190917611122, + 0.027312535792589188, + 0.02368086948990822, + 0.04577390477061272, + 0.025960082188248634, + 0.02207348123192787, + 0.0023257790599018335, + 0.004397563636302948, + 0.005066347774118185, + 0.00712578184902668, + 0.005257341545075178, + 0.0073910970240831375, + 0.00931041780859232, + 0.00439675897359848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.026346543803811073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493545159697533 + ], + "opposite_logprob_tokens": [ + -2.9299354553222656 + ], + "logprob_reward": -0.05493545159697533, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493545159697533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13236375153064728, + 0.023713378235697746, + 0.09164527803659439, + 0.06401669234037399, + 0.060444176197052, + 0.05404656380414963, + 0.13064776360988617, + 0.10551854968070984, + 0.2831798195838928, + 0.05442393571138382 + ], + "credit_map": [ + 0.013569142669439316, + 0.014193848706781864, + 0.0106150321662426, + 0.027115825563669205, + 0.01471332274377346, + 0.014898841269314289, + 0.01143102440983057, + 0.025826727971434593, + 0.0019385453779250383, + 0.0025864997878670692, + 0.0029403516091406345, + 0.003276288043707609, + 0.00261605903506279, + 0.0036827477160841227, + 0.0024099790025502443, + 0.00426291162148118, + 0.00666738860309124, + 0.016343621537089348, + 0.008207198232412338, + 0.008932103402912617, + 0.01820719800889492, + 0.014709760434925556, + 0.009712856262922287, + 0.008865159004926682, + 0.005172988399863243, + 0.008448917418718338, + 0.006763591896742582, + 0.014963465742766857, + 0.006469509098678827, + 0.009408537298440933, + 0.007654868997633457, + 0.00513482466340065, + 0.0035946022253483534, + 0.0072412071749567986, + 0.008401419036090374, + 0.007712848950177431, + 0.00803081039339304, + 0.008645729161798954, + 0.010025589726865292, + 0.006791977211833, + 0.005493034143000841, + 0.00753586832433939, + 0.008267747238278389, + 0.006991133093833923, + 0.008503192104399204, + 0.007139876484870911, + 0.005511503200978041, + 0.004604218993335962, + 0.009596867486834526, + 0.02977638877928257, + 0.01326451450586319, + 0.013286510482430458, + 0.022001655772328377, + 0.01821822300553322, + 0.011467142961919308, + 0.013036479242146015, + 0.012938237749040127, + 0.014552868902683258, + 0.01710423454642296, + 0.010705322958528996, + 0.009353328496217728, + 0.01181321032345295, + 0.014725562185049057, + 0.01432579942047596, + 0.019429251551628113, + 0.04266880825161934, + 0.03610784560441971, + 0.05783377215266228, + 0.02381942793726921, + 0.05200115963816643, + 0.029111433774232864, + 0.022208159789443016, + 0.003406129777431488, + 0.005535412114113569, + 0.004935587290674448, + 0.008659301325678825, + 0.009500927291810513, + 0.009824208915233612, + 0.008531910367310047, + 0.004030466545373201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.05493545159697533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a high-angle, long shot, slowly and deliberately zooming in, capturing a winding, dark gray paved mountain road through a dense forest ablaze with fiery reds and brilliant golds of autumn leaves. The low, golden rays of the dawn sun shimmer across the distant ridges, casting long, dramatic shadows and bathing the crisp air in a warm, ethereal glow. At the center of this vast, serene landscape lies the solitary figure of a petite, young Southeast Asian man with short, unkempt black hair, dressed simply in a gray sweatshirt and dark blue jeans. As the camera zooms in, his concentrated expression becomes clear, tinged with a hint of nervous worry. He rides on a worn skateboard with bright orange wheels, rolling forward at a slow, hesitant pace. His body stiffens and sways, his center of gravity shifting precariously. In an effort to maintain stability, he awkwardly flings his arms to the sides in an uncoordinated circle, his hands outstretched as if trying to grasp at the air itself. The entire scene is a tense and poignant ballet of effort, contrasting the men’s awkward, determined struggle with the quiet, breathtaking beauty of a golden autumn morning.", + "seed": 5756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_1852_seed_5756_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15838614106178284, + "vlm_logprob": -0.15838614106178284, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.15838614106178284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022304624319076538 + ], + "opposite_logprob_tokens": [ + -4.397304534912109 + ], + "logprob_reward": -0.022304624319076538, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022304624319076538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11240805685520172, + 0.06025617942214012, + 0.08094887435436249, + 0.02873201109468937, + 0.08906286209821701, + 0.4550904929637909, + 0.059721656143665314, + 0.03690342232584953, + 0.04643206298351288, + 0.030444424599409103 + ], + "credit_map": [ + 0.015633463859558105, + 0.017735140398144722, + 0.0160482469946146, + 0.011219395324587822, + 0.008041753433644772, + 0.016568303108215332, + 0.007606579456478357, + 0.019555171951651573, + 0.0037758955731987953, + 0.003230823902413249, + 0.005204380489885807, + 0.017832335084676743, + 0.004703103564679623, + 0.01568256877362728, + 0.004365384113043547, + 0.005461683496832848, + 0.009364730678498745, + 0.009114370681345463, + 0.00889542605727911, + 0.016821375116705894, + 0.006006643176078796, + 0.011363808996975422, + 0.005541233811527491, + 0.013841281644999981, + 0.002250589895993471, + 0.003266011131927371, + 0.00539627019315958, + 0.003350221086293459, + 0.00308388308621943, + 0.004126591142266989, + 0.0029750417452305555, + 0.00428340258076787, + 0.010396553203463554, + 0.01557543408125639, + 0.00901422742754221, + 0.02015879936516285, + 0.004881067667156458, + 0.01770939864218235, + 0.004933343734592199, + 0.006394033785909414, + 0.03587598353624344, + 0.012831793166697025, + 0.2019912451505661, + 0.14460338652133942, + 0.00735712144523859, + 0.02483706921339035, + 0.014176923781633377, + 0.013416996225714684, + 0.006567657459527254, + 0.005371181294322014, + 0.00453586271032691, + 0.006282021291553974, + 0.0062179649248719215, + 0.019102120772004128, + 0.0048348368145525455, + 0.006810012273490429, + 0.003950756508857012, + 0.004182140808552504, + 0.003812522394582629, + 0.004197398666292429, + 0.008504781872034073, + 0.0038178886752575636, + 0.00511518120765686, + 0.00332275009714067, + 0.002829440636560321, + 0.004334192257374525, + 0.005085177719593048, + 0.005414403509348631, + 0.007990080863237381, + 0.0032351305708289146, + 0.012236668728291988, + 0.005306973122060299, + 0.0041592912748456, + 0.0036533514503389597, + 0.003746128175407648, + 0.0047895521856844425, + 0.0028670039027929306, + 0.002921472769230604, + 0.004941907245665789, + 0.003365718759596348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.022304624319076538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054963886737823486 + ], + "opposite_logprob_tokens": [ + -2.9299638271331787 + ], + "logprob_reward": -0.054963886737823486, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054963886737823486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15225590765476227, + 0.16736523807048798, + 0.09287537634372711, + 0.07170421630144119, + 0.06963459402322769, + 0.28131917119026184, + 0.028972959145903587, + 0.04778577759861946, + 0.025286905467510223, + 0.062799833714962 + ], + "credit_map": [ + 0.018427805975079536, + 0.02520737424492836, + 0.02246657758951187, + 0.01852605864405632, + 0.011527945287525654, + 0.03675391897559166, + 0.0075454735197126865, + 0.011800744570791721, + 0.01207500696182251, + 0.008866752497851849, + 0.015312982723116875, + 0.05026310309767723, + 0.01151807326823473, + 0.04686586558818817, + 0.00991510134190321, + 0.012548340484499931, + 0.012279004789888859, + 0.007350645959377289, + 0.010383052751421928, + 0.023847725242376328, + 0.008392911404371262, + 0.012126515619456768, + 0.007161780260503292, + 0.01133373100310564, + 0.011583487503230572, + 0.009437352418899536, + 0.0067791021429002285, + 0.007524845190346241, + 0.006755528040230274, + 0.010629785247147083, + 0.00781969539821148, + 0.011174428276717663, + 0.007920441217720509, + 0.014498397707939148, + 0.009686543606221676, + 0.009415965527296066, + 0.005905559752136469, + 0.010989292524755001, + 0.003677364904433489, + 0.007541028317064047, + 0.021941833198070526, + 0.009595143608748913, + 0.12413142621517181, + 0.0876324474811554, + 0.005062691867351532, + 0.01609504222869873, + 0.008864112198352814, + 0.00799648929387331, + 0.003033880377188325, + 0.003330493811517954, + 0.0022043371573090553, + 0.003053204622119665, + 0.0024366325233131647, + 0.009568591602146626, + 0.002284124493598938, + 0.003061694325879216, + 0.004572737030684948, + 0.005494870711117983, + 0.0047769732773303986, + 0.00514918752014637, + 0.006817336194217205, + 0.01157698780298233, + 0.00441763224080205, + 0.004980055149644613, + 0.0021804668940603733, + 0.003029125975444913, + 0.002858900697901845, + 0.0030149794183671474, + 0.0036955643445253372, + 0.0023678576108068228, + 0.005374769680202007, + 0.00276524038054049, + 0.0032872254960238934, + 0.006216224282979965, + 0.00451262854039669, + 0.007851488888263702, + 0.014512081630527973, + 0.006734510418027639, + 0.012699291110038757, + 0.00698638241738081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.054963886737823486 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288947582244873 + ], + "opposite_logprob_tokens": [ + -1.0538947582244873 + ], + "logprob_reward": -0.4288947582244873, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288947582244873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22147183120250702, + 0.03874323144555092, + 0.04905856400728226, + 0.05753849446773529, + 0.012352724559605122, + 0.4247840344905853, + 0.07935472577810287, + 0.026679471135139465, + 0.037391386926174164, + 0.052625495940446854 + ], + "credit_map": [ + 0.011369592510163784, + 0.014901229180395603, + 0.03043666109442711, + 0.02907520718872547, + 0.02447577938437462, + 0.04203091561794281, + 0.025257812812924385, + 0.04392464458942413, + 0.0023008924908936024, + 0.0030818020459264517, + 0.0033881766721606255, + 0.005435814615339041, + 0.007811417803168297, + 0.00711631728336215, + 0.005789988208562136, + 0.0038188276812434196, + 0.0036072188522666693, + 0.004120599012821913, + 0.004672812297940254, + 0.006557364948093891, + 0.010944091714918613, + 0.006531121209263802, + 0.00850026123225689, + 0.004125099163502455, + 0.005541174206882715, + 0.006491546519100666, + 0.00926594901829958, + 0.006533777806907892, + 0.006572576705366373, + 0.007625064812600613, + 0.006141279358416796, + 0.009367133490741253, + 0.001153975841589272, + 0.0015724850818514824, + 0.001797458971850574, + 0.0015726598212495446, + 0.0014697728911414742, + 0.0015966595383360982, + 0.001730665797367692, + 0.0014590477803722024, + 0.033526189625263214, + 0.01306744571775198, + 0.1842280626296997, + 0.13703346252441406, + 0.008050400763750076, + 0.021771321073174477, + 0.01379647571593523, + 0.013310710899531841, + 0.008949438109993935, + 0.007263589184731245, + 0.006680427119135857, + 0.009538047015666962, + 0.0070336549542844296, + 0.025252731516957283, + 0.0053261383436620235, + 0.009310709312558174, + 0.002390125999227166, + 0.0030542444437742233, + 0.0038887187838554382, + 0.004274403210729361, + 0.0022802408784627914, + 0.005569000728428364, + 0.0021309787407517433, + 0.003091760678216815, + 0.0026538376696407795, + 0.004069380462169647, + 0.004952278919517994, + 0.004364369437098503, + 0.00412861630320549, + 0.004016694147139788, + 0.008627037517726421, + 0.004579175263643265, + 0.0034825399052351713, + 0.005951341707259417, + 0.004044467117637396, + 0.006726148072630167, + 0.012628289870917797, + 0.006970705930143595, + 0.008204236626625061, + 0.004617774393409491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.4288947582244873 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336986780166626 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.0336986780166626, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336986780166626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056828297674655914, + 0.11762972921133041, + 0.06696294993162155, + 0.04841528832912445, + 0.03875367343425751, + 0.5012629628181458, + 0.06977395713329315, + 0.019961684942245483, + 0.03012804314494133, + 0.05028343200683594 + ], + "credit_map": [ + 0.004602231550961733, + 0.00491077546030283, + 0.007739955093711615, + 0.00819811224937439, + 0.006599303334951401, + 0.008833606727421284, + 0.008037599734961987, + 0.007906715385615826, + 0.006981243845075369, + 0.00626668706536293, + 0.009064790792763233, + 0.037536028772592545, + 0.007485747337341309, + 0.033594001084566116, + 0.007878441363573074, + 0.008822795003652573, + 0.008525127544999123, + 0.006459868513047695, + 0.006515077315270901, + 0.012240676209330559, + 0.008625821210443974, + 0.01020662859082222, + 0.0056091248989105225, + 0.00878062192350626, + 0.003951519727706909, + 0.004900651052594185, + 0.008664228953421116, + 0.00449192663654685, + 0.005923647433519363, + 0.007042224518954754, + 0.005470983684062958, + 0.007970106787979603, + 0.0034594715107232332, + 0.005755777936428785, + 0.0031408253125846386, + 0.008740008808672428, + 0.0037266262806952, + 0.008270437829196453, + 0.002949794754385948, + 0.00271073030307889, + 0.04101806506514549, + 0.024599378928542137, + 0.18297813832759857, + 0.14322243630886078, + 0.021089445799589157, + 0.03521193563938141, + 0.024627989158034325, + 0.02851557731628418, + 0.007939175702631474, + 0.006234966684132814, + 0.007610834203660488, + 0.009165707044303417, + 0.006516072899103165, + 0.016922973096370697, + 0.00813743844628334, + 0.00724678672850132, + 0.0017852146411314607, + 0.002353203482925892, + 0.002084809821099043, + 0.002564801136031747, + 0.0026448422577232122, + 0.0029773947317153215, + 0.0032922595273703337, + 0.0022591594606637955, + 0.0025281531270593405, + 0.002805834636092186, + 0.0043415199033916, + 0.004067656584084034, + 0.004489813931286335, + 0.002376196440309286, + 0.0064420802518725395, + 0.0030767887365072966, + 0.0030746778938919306, + 0.005880668759346008, + 0.004722196608781815, + 0.005872742272913456, + 0.009940366260707378, + 0.005503334570676088, + 0.0090888487175107, + 0.006200595758855343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.0336986780166626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25206875801086426 + ], + "opposite_logprob_tokens": [ + -1.5020687580108643 + ], + "logprob_reward": -0.25206875801086426, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25206875801086426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19981038570404053, + 0.0711907148361206, + 0.03426092118024826, + 0.0736098438501358, + 0.11191152781248093, + 0.3107669949531555, + 0.040274642407894135, + 0.06024674326181412, + 0.05457068234682083, + 0.04335751757025719 + ], + "credit_map": [ + 0.017428947612643242, + 0.021977532655000687, + 0.022273246198892593, + 0.031052477657794952, + 0.02081960439682007, + 0.0330960787832737, + 0.013593395240604877, + 0.03956910967826843, + 0.004026118200272322, + 0.00514852162450552, + 0.005394753534346819, + 0.016763698309659958, + 0.010341291315853596, + 0.016323082149028778, + 0.007199416868388653, + 0.005993830505758524, + 0.0028299433179199696, + 0.0029624965973198414, + 0.0038888640701770782, + 0.005538825411349535, + 0.005829658824950457, + 0.004760704934597015, + 0.00421252241358161, + 0.00423790467903018, + 0.004830896854400635, + 0.004315582569688559, + 0.008392641320824623, + 0.00847699586302042, + 0.015117269940674305, + 0.010229488834738731, + 0.013103603385388851, + 0.009143363684415817, + 0.014180394820868969, + 0.033720362931489944, + 0.012228872627019882, + 0.011706029996275902, + 0.010642042383551598, + 0.012888112105429173, + 0.006826780270785093, + 0.009718933142721653, + 0.023859260603785515, + 0.015046270564198494, + 0.11521296948194504, + 0.08602097630500793, + 0.01267633680254221, + 0.023754112422466278, + 0.013524879701435566, + 0.020672189071774483, + 0.003998595289885998, + 0.003100933041423559, + 0.004221657756716013, + 0.00392340449616313, + 0.004393802024424076, + 0.01253008097410202, + 0.0035062043461948633, + 0.004599965643137693, + 0.006026134826242924, + 0.005658388137817383, + 0.00728184636682272, + 0.011669700965285301, + 0.006201427895575762, + 0.011468491517007351, + 0.005245001055300236, + 0.00669575110077858, + 0.004421028774231672, + 0.00641388725489378, + 0.006922459229826927, + 0.006858707405626774, + 0.008441885933279991, + 0.005807522684335709, + 0.00864091981202364, + 0.007064271718263626, + 0.003530764952301979, + 0.004681213293224573, + 0.0040189046412706375, + 0.0059939357452094555, + 0.009156867861747742, + 0.005591480061411858, + 0.006487362552434206, + 0.0038969917222857475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.25206875801086426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a high-angle, long shot, slowly and deliberately zooming in, capturing a winding, dark gray paved mountain road through a dense forest ablaze with fiery reds and brilliant golds of autumn leaves. The low, golden rays of the dawn sun shimmer across the distant ridges, casting long, dramatic shadows and bathing the crisp air in a warm, ethereal glow. At the center of this vast, serene landscape lies the solitary figure of a petite, young Southeast Asian man with short, unkempt black hair, dressed simply in a gray sweatshirt and dark blue jeans. As the camera zooms in, his concentrated expression becomes clear, tinged with a hint of nervous worry. He rides on a worn skateboard with bright orange wheels, rolling forward at a slow, hesitant pace. His body stiffens and sways, his center of gravity shifting precariously. In an effort to maintain stability, he awkwardly flings his arms to the sides in an uncoordinated circle, his hands outstretched as if trying to grasp at the air itself. The entire scene is a tense and poignant ballet of effort, contrasting the men’s awkward, determined struggle with the quiet, breathtaking beauty of a golden autumn morning.", + "seed": 5757, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_1852_seed_5757_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.097009002789855, + "vlm_logprob": -0.097009002789855, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.097009002789855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020833821967244148 + ], + "opposite_logprob_tokens": [ + -4.645833969116211 + ], + "logprob_reward": -0.020833821967244148, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.020833821967244148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05296717584133148, + 0.03351223096251488, + 0.05870163440704346, + 0.019596682861447334, + 0.027745869010686874, + 0.10954882204532623, + 0.1046363040804863, + 0.0589638352394104, + 0.49919354915618896, + 0.03513391315937042 + ], + "credit_map": [ + 0.004991416819393635, + 0.0054889945313334465, + 0.005944678094238043, + 0.006034258287400007, + 0.010685859248042107, + 0.006892858538776636, + 0.006915295030921698, + 0.006013808771967888, + 0.0028164442628622055, + 0.0026345858350396156, + 0.0046129352413117886, + 0.004435861017554998, + 0.0033415439538657665, + 0.0059829759411513805, + 0.005558183416724205, + 0.0041297003626823425, + 0.006429161876440048, + 0.0034704129211604595, + 0.009255679324269295, + 0.004258102271705866, + 0.00336590688675642, + 0.007644354831427336, + 0.007194915786385536, + 0.017083091661334038, + 0.0018913481617346406, + 0.002023401902988553, + 0.0026312246918678284, + 0.0021896418184041977, + 0.001628390047699213, + 0.00385879073292017, + 0.0026458411011844873, + 0.0027280435897409916, + 0.0033150946255773306, + 0.0034562312066555023, + 0.005105774849653244, + 0.0028804296161979437, + 0.0026630391366779804, + 0.005140503868460655, + 0.002513144165277481, + 0.0026716485153883696, + 0.011578087694942951, + 0.006865705829113722, + 0.03374345973134041, + 0.01160572562366724, + 0.005790690891444683, + 0.007124282885342836, + 0.007217791862785816, + 0.02562307007610798, + 0.01790669932961464, + 0.010900181718170643, + 0.012833990156650543, + 0.01284337043762207, + 0.008054515346884727, + 0.014332575723528862, + 0.00809125043451786, + 0.01967371255159378, + 0.006330203730612993, + 0.003485680092126131, + 0.011836331337690353, + 0.013375023379921913, + 0.004056138452142477, + 0.007715378887951374, + 0.007037692237645388, + 0.005127377342432737, + 0.014179972000420094, + 0.020252717658877373, + 0.07883995771408081, + 0.20619238913059235, + 0.014122174121439457, + 0.08522580564022064, + 0.041116826236248016, + 0.03926366567611694, + 0.0027353849727660418, + 0.003345807082951069, + 0.0051096174865961075, + 0.0052140019834041595, + 0.0037781596183776855, + 0.005659257527440786, + 0.004105870611965656, + 0.005185809452086687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.020833821967244148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371112421154976 + ], + "opposite_logprob_tokens": [ + -3.4087111949920654 + ], + "logprob_reward": -0.03371112421154976, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371112421154976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07110446691513062, + 0.014001852832734585, + 0.028167635202407837, + 0.015255287289619446, + 0.021195854991674423, + 0.44924572110176086, + 0.09763407707214355, + 0.14490367472171783, + 0.1201636865735054, + 0.038327742367982864 + ], + "credit_map": [ + 0.006932670716196299, + 0.006381277926266193, + 0.010677101090550423, + 0.006232081446796656, + 0.009074246510863304, + 0.015858734026551247, + 0.008647962473332882, + 0.007300389930605888, + 0.001372613711282611, + 0.001368028693832457, + 0.0019270471530035138, + 0.0015062334714457393, + 0.0014307850506156683, + 0.0024010371416807175, + 0.0023498092778027058, + 0.001646299147978425, + 0.0033136478159576654, + 0.001910562627017498, + 0.0032849630806595087, + 0.0025364276953041553, + 0.0016460679471492767, + 0.0041452390141785145, + 0.0034377870615571737, + 0.007892938330769539, + 0.001276069087907672, + 0.0010780045995488763, + 0.002117396332323551, + 0.0019553389865905046, + 0.0012712505413219333, + 0.002772260457277298, + 0.0030195452272892, + 0.001765421126037836, + 0.0026158245746046305, + 0.0022903699427843094, + 0.0023097950033843517, + 0.0029890434816479683, + 0.0020773662254214287, + 0.004116457886993885, + 0.0019810814410448074, + 0.002815918531268835, + 0.045102305710315704, + 0.013161877170205116, + 0.1666942536830902, + 0.05452699214220047, + 0.013837523758411407, + 0.021925387904047966, + 0.014661559835076332, + 0.11933582276105881, + 0.01344490796327591, + 0.008614889346063137, + 0.013845751993358135, + 0.01031371671706438, + 0.007028723601251841, + 0.015759721398353577, + 0.007132271304726601, + 0.021494094282388687, + 0.017322516068816185, + 0.013098856434226036, + 0.02828107587993145, + 0.029876230284571648, + 0.014231659471988678, + 0.019394421949982643, + 0.010965141467750072, + 0.011733788065612316, + 0.004121515434235334, + 0.005076829344034195, + 0.020529914647340775, + 0.04817644879221916, + 0.004035645630210638, + 0.017015330493450165, + 0.010794767178595066, + 0.010413231328129768, + 0.003438310930505395, + 0.0021848666947335005, + 0.004348176531493664, + 0.005690411664545536, + 0.0038589860778301954, + 0.009564697742462158, + 0.00495159812271595, + 0.004290690645575523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.03371112421154976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134441375732422 + ], + "opposite_logprob_tokens": [ + -1.3134441375732422 + ], + "logprob_reward": -0.3134441375732422, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134441375732422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0579085536301136, + 0.07784120738506317, + 0.08187814801931381, + 0.028414785861968994, + 0.028590235859155655, + 0.13267849385738373, + 0.26104462146759033, + 0.08586160838603973, + 0.19842368364334106, + 0.04735865443944931 + ], + "credit_map": [ + 0.00437207194045186, + 0.0034913946874439716, + 0.007674978114664555, + 0.007933331653475761, + 0.00732417730614543, + 0.008031820878386497, + 0.011235437355935574, + 0.007845341227948666, + 0.0063851866871118546, + 0.004946283530443907, + 0.009568124078214169, + 0.010655456222593784, + 0.006945537403225899, + 0.010779540985822678, + 0.018717730417847633, + 0.00984334759414196, + 0.008957232348620892, + 0.0053315116092562675, + 0.010653696954250336, + 0.006790814455598593, + 0.005032852292060852, + 0.010975608602166176, + 0.011271339841187, + 0.022865090519189835, + 0.0024002776481211185, + 0.0025118712801486254, + 0.0036248054821044207, + 0.002936059143394232, + 0.003318185219541192, + 0.005589952226728201, + 0.00472924904897809, + 0.003304384183138609, + 0.002980378922075033, + 0.003580983029678464, + 0.004325363785028458, + 0.002542535774409771, + 0.0029281149618327618, + 0.006846696138381958, + 0.002325451700016856, + 0.003060711547732353, + 0.0097650196403265, + 0.006058898754417896, + 0.03488021716475487, + 0.012995181605219841, + 0.0078247906640172, + 0.008767727762460709, + 0.020602107048034668, + 0.0317845419049263, + 0.01637417823076248, + 0.01331726461648941, + 0.0765949934720993, + 0.028725946322083473, + 0.013935782015323639, + 0.03317960351705551, + 0.014436585828661919, + 0.06448028236627579, + 0.005253752693533897, + 0.006773423403501511, + 0.00856213178485632, + 0.01200543437153101, + 0.014398851431906223, + 0.012000522576272488, + 0.01865953579545021, + 0.008207954466342926, + 0.006657344289124012, + 0.005972387734800577, + 0.014446316286921501, + 0.05737588182091713, + 0.011186894029378891, + 0.06516736000776291, + 0.016485944390296936, + 0.02113155834376812, + 0.004197260364890099, + 0.003165814094245434, + 0.0038523937109857798, + 0.005847631022334099, + 0.007496717851608992, + 0.011103840544819832, + 0.0038049104623496532, + 0.00789008941501379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.3134441375732422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809523582458496 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809523582458496, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809523582458496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04523485153913498, + 0.0316878966987133, + 0.10628210753202438, + 0.015070207417011261, + 0.02452661097049713, + 0.06540828943252563, + 0.2450464963912964, + 0.13826967775821686, + 0.28756821155548096, + 0.04090571776032448 + ], + "credit_map": [ + 0.0037048899102956057, + 0.003403791692107916, + 0.007159324828535318, + 0.006412307266145945, + 0.004085739608854055, + 0.010018548928201199, + 0.005193019285798073, + 0.005257224198430777, + 0.002809715224429965, + 0.0030106049962341785, + 0.004395165480673313, + 0.003442500252276659, + 0.004801027476787567, + 0.005504449363797903, + 0.004872692748904228, + 0.0028517397586256266, + 0.011194391176104546, + 0.007000755053013563, + 0.011433636769652367, + 0.010014954954385757, + 0.006729076150804758, + 0.01625058613717556, + 0.015781626105308533, + 0.027877070009708405, + 0.0013744289753958583, + 0.001403556321747601, + 0.0021004341542720795, + 0.0019086232641711831, + 0.0013365143677219748, + 0.0020301705226302147, + 0.00275653088465333, + 0.002159947296604514, + 0.0024854689836502075, + 0.002869394375011325, + 0.002712565939873457, + 0.0026128662284463644, + 0.002350961323827505, + 0.006107294466346502, + 0.002228083787485957, + 0.0031599754001945257, + 0.00514468876644969, + 0.00401742197573185, + 0.013292553834617138, + 0.007042566314339638, + 0.006585991941392422, + 0.0043487828224897385, + 0.011584177613258362, + 0.013392101973295212, + 0.017656803131103516, + 0.007269551511853933, + 0.08526822179555893, + 0.02802848629653454, + 0.0063100759871304035, + 0.017850030213594437, + 0.010074932128190994, + 0.07258838415145874, + 0.01032824907451868, + 0.007889772765338421, + 0.026996759697794914, + 0.03633526340126991, + 0.010488471016287804, + 0.013791576027870178, + 0.01949727162718773, + 0.012942311353981495, + 0.01001660618931055, + 0.008879526518285275, + 0.0219626035541296, + 0.0888790562748909, + 0.01255020685493946, + 0.09289223700761795, + 0.021559514105319977, + 0.030828453600406647, + 0.004176389891654253, + 0.0024590108077973127, + 0.0035745762288570404, + 0.005390475504100323, + 0.005723359528928995, + 0.009225341491401196, + 0.005711826961487532, + 0.004644734784960747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.03809523582458496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896069437265396 + ], + "opposite_logprob_tokens": [ + -2.578960657119751 + ], + "logprob_reward": -0.07896069437265396, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896069437265396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030086399987339973, + 0.01971929520368576, + 0.03678857535123825, + 0.01854577660560608, + 0.02141963690519333, + 0.1243065819144249, + 0.1456601321697235, + 0.07664911448955536, + 0.5141157507896423, + 0.01270865648984909 + ], + "credit_map": [ + 0.0025665094144642353, + 0.0028984330128878355, + 0.0042074015364050865, + 0.0038416890893131495, + 0.0037387246266007423, + 0.004264282528311014, + 0.0041847447864711285, + 0.0043846191838383675, + 0.0013733271043747663, + 0.001482130028307438, + 0.002395427552983165, + 0.003079306799918413, + 0.002762158401310444, + 0.0036174114793539047, + 0.002730041742324829, + 0.0022794948890805244, + 0.004140722565352917, + 0.002515310188755393, + 0.004151185043156147, + 0.003261143807321787, + 0.002339815953746438, + 0.006541410461068153, + 0.004602031782269478, + 0.009236957877874374, + 0.0020711023826152086, + 0.0016951016150414944, + 0.0027931786607950926, + 0.002240362111479044, + 0.0015408985782414675, + 0.0026147328317165375, + 0.0029243342578411102, + 0.002666070591658354, + 0.002664351835846901, + 0.001683877781033516, + 0.005653516389429569, + 0.002392582828179002, + 0.0014129792107269168, + 0.0020590531639754772, + 0.001670650439336896, + 0.0038826267700642347, + 0.012211299501359463, + 0.0046805283054709435, + 0.040743812918663025, + 0.014933612197637558, + 0.006262153387069702, + 0.008041723631322384, + 0.008069791831076145, + 0.029363667592406273, + 0.02037135511636734, + 0.014499323442578316, + 0.01764657534658909, + 0.0173464547842741, + 0.01203894056379795, + 0.026378773152828217, + 0.009783169254660606, + 0.02759554237127304, + 0.005092919338494539, + 0.0032840175554156303, + 0.01589077338576317, + 0.01771525666117668, + 0.00507461279630661, + 0.0063356622122228146, + 0.016893573105335236, + 0.00636231480166316, + 0.017189010977745056, + 0.015669116750359535, + 0.10796581208705902, + 0.15943610668182373, + 0.018610425293445587, + 0.058172501623630524, + 0.04148085415363312, + 0.09559198468923569, + 0.0014585282187908888, + 0.0013520049396902323, + 0.0018379290122538805, + 0.001425668247975409, + 0.0016512004658579826, + 0.002277935855090618, + 0.0013846569927409291, + 0.0013207338051870465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.07896069437265396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a high-angle, long shot, slowly and deliberately zooming in, capturing a winding, dark gray paved mountain road through a dense forest ablaze with fiery reds and brilliant golds of autumn leaves. The low, golden rays of the dawn sun shimmer across the distant ridges, casting long, dramatic shadows and bathing the crisp air in a warm, ethereal glow. At the center of this vast, serene landscape lies the solitary figure of a petite, young Southeast Asian man with short, unkempt black hair, dressed simply in a gray sweatshirt and dark blue jeans. As the camera zooms in, his concentrated expression becomes clear, tinged with a hint of nervous worry. He rides on a worn skateboard with bright orange wheels, rolling forward at a slow, hesitant pace. His body stiffens and sways, his center of gravity shifting precariously. In an effort to maintain stability, he awkwardly flings his arms to the sides in an uncoordinated circle, his hands outstretched as if trying to grasp at the air itself. The entire scene is a tense and poignant ballet of effort, contrasting the men’s awkward, determined struggle with the quiet, breathtaking beauty of a golden autumn morning.", + "seed": 5758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_11/index_1852_seed_5758_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14954828433692455, + "vlm_logprob": -0.14954828433692455, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.14954828433692455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021156469359993935 + ], + "opposite_logprob_tokens": [ + -4.396156311035156 + ], + "logprob_reward": -0.021156469359993935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.021156469359993935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27837270498275757, + 0.06772053241729736, + 0.1378181278705597, + 0.14453577995300293, + 0.09153101593255997, + 0.05236620828509331, + 0.04755924269556999, + 0.08643294125795364, + 0.05598471313714981, + 0.03767872974276543 + ], + "credit_map": [ + 0.02062099799513817, + 0.02146180346608162, + 0.044633399695158005, + 0.07283994555473328, + 0.01880512945353985, + 0.02763250656425953, + 0.0397210419178009, + 0.03265787288546562, + 0.00647893687710166, + 0.004973265342414379, + 0.010239208117127419, + 0.015188741497695446, + 0.005832015536725521, + 0.008367334492504597, + 0.010784466750919819, + 0.005856565199792385, + 0.017778195440769196, + 0.013780280947685242, + 0.022079339250922203, + 0.02021479606628418, + 0.01261990237981081, + 0.029688308015465736, + 0.010454191826283932, + 0.011203118599951267, + 0.012245078571140766, + 0.008761415258049965, + 0.03888694569468498, + 0.020032266154885292, + 0.010600094683468342, + 0.01973806694149971, + 0.011598552577197552, + 0.022673355415463448, + 0.0064305514097213745, + 0.005402035545557737, + 0.015966933220624924, + 0.012732353992760181, + 0.005865515675395727, + 0.016088496893644333, + 0.010861753486096859, + 0.018183378502726555, + 0.003846683306619525, + 0.0036393546033650637, + 0.007936625741422176, + 0.00694609060883522, + 0.0042352695018053055, + 0.00629220437258482, + 0.011656147427856922, + 0.007813830859959126, + 0.005040733143687248, + 0.004723008722066879, + 0.004767845384776592, + 0.004636405501514673, + 0.005819783546030521, + 0.008273515850305557, + 0.006321802269667387, + 0.007976150140166283, + 0.007080820854753256, + 0.008358432911336422, + 0.011089442297816277, + 0.012715740129351616, + 0.00920374970883131, + 0.014916076324880123, + 0.013483528047800064, + 0.009585151448845863, + 0.007950722239911556, + 0.008550718426704407, + 0.004809875041246414, + 0.005977503024041653, + 0.004914314951747656, + 0.008765485137701035, + 0.008407716639339924, + 0.006608380936086178, + 0.0039182002656161785, + 0.0036197786685079336, + 0.004800116177648306, + 0.004174161236733198, + 0.004906591959297657, + 0.005916452966630459, + 0.006027010269463062, + 0.004316416103392839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.021156469359993935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304611444473267 + ], + "opposite_logprob_tokens": [ + -3.7733047008514404 + ], + "logprob_reward": -0.023304611444473267, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304611444473267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19418640434741974, + 0.06626757234334946, + 0.06976547837257385, + 0.2346404492855072, + 0.19926585257053375, + 0.028838641941547394, + 0.03594770282506943, + 0.0694604367017746, + 0.03148161619901657, + 0.0701458752155304 + ], + "credit_map": [ + 0.02061341144144535, + 0.017879001796245575, + 0.02855449728667736, + 0.03062513843178749, + 0.01762005314230919, + 0.01908828690648079, + 0.04173292592167854, + 0.018073081970214844, + 0.0037738692481070757, + 0.004394251853227615, + 0.009850038215517998, + 0.015192266553640366, + 0.005059633869677782, + 0.007710135076195002, + 0.0131430234760046, + 0.007144355680793524, + 0.00612936494871974, + 0.004831119906157255, + 0.010968304239213467, + 0.008274213410913944, + 0.004800456576049328, + 0.010419132187962532, + 0.01809013821184635, + 0.006252748891711235, + 0.016455229371786118, + 0.01232072338461876, + 0.055357854813337326, + 0.035332031548023224, + 0.016762960702180862, + 0.03151370584964752, + 0.02702658623456955, + 0.03987136110663414, + 0.024646412581205368, + 0.012798520736396313, + 0.02930806204676628, + 0.02230559103190899, + 0.011245906352996826, + 0.026420246809720993, + 0.019007714465260506, + 0.0535334013402462, + 0.0024269751738756895, + 0.0026194811798632145, + 0.004011409357190132, + 0.0026939096860587597, + 0.005062667187303305, + 0.00377935403957963, + 0.004481080453842878, + 0.0037637664936482906, + 0.002894931472837925, + 0.0036120545119047165, + 0.0038385975640267134, + 0.0037696349900215864, + 0.006067972630262375, + 0.006952305324375629, + 0.004499505739659071, + 0.004312700126320124, + 0.004860592074692249, + 0.006766329053789377, + 0.007494082674384117, + 0.008203791454434395, + 0.009077011607587337, + 0.01438547670841217, + 0.008289661258459091, + 0.010383496060967445, + 0.004305146634578705, + 0.0041997842490673065, + 0.0027227813843637705, + 0.0030191726982593536, + 0.0045719509944319725, + 0.004980791825801134, + 0.004025212489068508, + 0.0036567728966474533, + 0.007440475281327963, + 0.011196602135896683, + 0.004249253310263157, + 0.004250149708241224, + 0.012753892689943314, + 0.014445064589381218, + 0.007392444182187319, + 0.008417992852628231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.023304611444473267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761547684669495 + ], + "opposite_logprob_tokens": [ + -0.8261547684669495 + ], + "logprob_reward": -0.5761547684669495, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761547684669495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2097933143377304, + 0.08988279849290848, + 0.08650124818086624, + 0.07111576199531555, + 0.16723881661891937, + 0.14347326755523682, + 0.08157907426357269, + 0.045607492327690125, + 0.07477805018424988, + 0.030030181631445885 + ], + "credit_map": [ + 0.016231797635555267, + 0.014780608005821705, + 0.03508688509464264, + 0.0459790974855423, + 0.018150942400097847, + 0.02276228368282318, + 0.03199129179120064, + 0.02481040544807911, + 0.007066427730023861, + 0.00541325006633997, + 0.014645349234342575, + 0.019083458930253983, + 0.008172460831701756, + 0.010064634494483471, + 0.01653963513672352, + 0.008897587656974792, + 0.005621799733489752, + 0.005213512107729912, + 0.01292847003787756, + 0.01170338224619627, + 0.008042022585868835, + 0.012240092270076275, + 0.022137340158224106, + 0.00861462764441967, + 0.005504987668246031, + 0.004748101811856031, + 0.014253957197070122, + 0.010168625973165035, + 0.004809506703168154, + 0.007579080294817686, + 0.009984622709453106, + 0.014066876843571663, + 0.018089013174176216, + 0.00906431209295988, + 0.023651599884033203, + 0.02191888354718685, + 0.007714616134762764, + 0.01721261814236641, + 0.023404859006404877, + 0.046182915568351746, + 0.01209300197660923, + 0.013498079031705856, + 0.017330417409539223, + 0.013858171179890633, + 0.026057668030261993, + 0.018510235473513603, + 0.0271131694316864, + 0.015012531541287899, + 0.0056966757401824, + 0.0061737122014164925, + 0.009108195081353188, + 0.009077660739421844, + 0.011174691841006279, + 0.0122058792039752, + 0.00850440189242363, + 0.019637858495116234, + 0.002868206240236759, + 0.003459883388131857, + 0.005927081219851971, + 0.006547878030687571, + 0.0048062908463180065, + 0.007320804987102747, + 0.007016747258603573, + 0.007660596165806055, + 0.006603691261261702, + 0.006291562225669622, + 0.007318910211324692, + 0.0068679507821798325, + 0.0119828125461936, + 0.017529349774122238, + 0.010233582928776741, + 0.007950185798108578, + 0.0026709327939897776, + 0.003528145607560873, + 0.004050835967063904, + 0.0032985820434987545, + 0.004409470595419407, + 0.0038457224145531654, + 0.004636370576918125, + 0.0035901193041354418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.5761547684669495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810257837176323 + ], + "opposite_logprob_tokens": [ + -3.288102626800537 + ], + "logprob_reward": -0.03810257837176323, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810257837176323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17413677275180817, + 0.06644995510578156, + 0.1502065509557724, + 0.06472747027873993, + 0.2439153492450714, + 0.07014403492212296, + 0.03538692742586136, + 0.04364169389009476, + 0.05994071438908577, + 0.09145058691501617 + ], + "credit_map": [ + 0.015589195303618908, + 0.010408065281808376, + 0.02831888571381569, + 0.026591144502162933, + 0.010479487478733063, + 0.013164815492928028, + 0.050980180501937866, + 0.018605004996061325, + 0.0044845156371593475, + 0.0033024996519088745, + 0.010973607189953327, + 0.012297900393605232, + 0.005594607442617416, + 0.00682916259393096, + 0.015387886203825474, + 0.007579771801829338, + 0.01714719459414482, + 0.01344444788992405, + 0.02324550971388817, + 0.019073238596320152, + 0.010584481060504913, + 0.027962923049926758, + 0.02525932528078556, + 0.013489418663084507, + 0.005444375332444906, + 0.0035667160991579294, + 0.01764789968729019, + 0.00742970360442996, + 0.005167472176253796, + 0.007664210628718138, + 0.007532804273068905, + 0.01027428824454546, + 0.020174888893961906, + 0.009215240366756916, + 0.04108750820159912, + 0.02858944982290268, + 0.01617734134197235, + 0.030185792595148087, + 0.03442291542887688, + 0.0640622153878212, + 0.004940123297274113, + 0.004691506270319223, + 0.010244474746286869, + 0.00760312657803297, + 0.009583898819983006, + 0.009699310176074505, + 0.013145768083631992, + 0.010235830210149288, + 0.002770018298178911, + 0.003692113095894456, + 0.0035522980615496635, + 0.004414958879351616, + 0.004319469910115004, + 0.006438035052269697, + 0.005894514732062817, + 0.004305517300963402, + 0.00422204053029418, + 0.004370816983282566, + 0.006663814187049866, + 0.005146108567714691, + 0.005371506791561842, + 0.008299639448523521, + 0.0055088200606405735, + 0.00405894685536623, + 0.004993897397071123, + 0.006243037059903145, + 0.005918452981859446, + 0.0069017657078802586, + 0.008527949452400208, + 0.009774448350071907, + 0.011129481717944145, + 0.006451680324971676, + 0.00893534068018198, + 0.014852949418127537, + 0.004959635436534882, + 0.005593787878751755, + 0.015351858921349049, + 0.020731957629323006, + 0.008485717698931694, + 0.012539341114461422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.03810257837176323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a winding, dark gray paved mountain road through a dense forest with autumn leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by low golden dawn sun rays casting long shadows and a warm glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a petite young Southeast Asian man with short black hair wearing a gray sweatshirt and dark blue jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man riding a worn skateboard with bright orange wheels at a slow, hesitant pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle long shot and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902299404144287 + ], + "opposite_logprob_tokens": [ + -2.4640231132507324 + ], + "logprob_reward": -0.08902299404144287, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902299404144287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2445424497127533, + 0.05282573774456978, + 0.13378717005252838, + 0.11500826478004456, + 0.0790763571858406, + 0.06118107587099075, + 0.07523993402719498, + 0.05258778855204582, + 0.09098413586616516, + 0.09476707875728607 + ], + "credit_map": [ + 0.020418681204319, + 0.01825171336531639, + 0.03748319298028946, + 0.056514233350753784, + 0.02194829471409321, + 0.03406914323568344, + 0.02398478239774704, + 0.031872428953647614, + 0.0032319044694304466, + 0.002679564291611314, + 0.008309909142553806, + 0.008007463067770004, + 0.005304038990288973, + 0.0063119069673120975, + 0.013566873967647552, + 0.0054140775464475155, + 0.01444477029144764, + 0.012295660562813282, + 0.021815553307533264, + 0.021398968994617462, + 0.011330907233059406, + 0.02871045656502247, + 0.012694930657744408, + 0.011095919646322727, + 0.010473906062543392, + 0.007829670794308186, + 0.029666636139154434, + 0.012545504607260227, + 0.006816859822720289, + 0.01370718888938427, + 0.018356220796704292, + 0.015612272545695305, + 0.005084377713501453, + 0.00372126093134284, + 0.013668078929185867, + 0.009322217665612698, + 0.006544328294694424, + 0.014602106995880604, + 0.011993085965514183, + 0.014140897430479527, + 0.0041483077220618725, + 0.004054923541843891, + 0.009268570691347122, + 0.006379568018019199, + 0.004765454679727554, + 0.006893751677125692, + 0.019332097843289375, + 0.006338398903608322, + 0.005057081114500761, + 0.005073051434010267, + 0.009921952150762081, + 0.00778951495885849, + 0.01186123676598072, + 0.012710491195321083, + 0.007567662745714188, + 0.01525894459336996, + 0.004348242189735174, + 0.005058815237134695, + 0.009766786359250546, + 0.005214259494096041, + 0.005070443730801344, + 0.00838038045912981, + 0.004803324583917856, + 0.009945535100996494, + 0.010386421345174313, + 0.010497799143195152, + 0.007418310735374689, + 0.008740641176700592, + 0.019308818504214287, + 0.013305664993822575, + 0.011549717746675014, + 0.009776761755347252, + 0.008144932799041271, + 0.015379094518721104, + 0.005049579776823521, + 0.006339171435683966, + 0.017747584730386734, + 0.02014676108956337, + 0.010852922685444355, + 0.011107026599347591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.08902299404144287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_tensor([5920, 5920])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_tensor([5920, 5920])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a0097c3f6ff0329663b33e0afb7d58766ca638c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_tensor([5920, 5920])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, handheld shot closely follows two Asian men engaged in a fierce exchange, the camera panning with their movements to create an immersive sense of chaos. The man in the foreground, with short, sweat-soaked black hair plastered to his scalp, is shirtless, revealing a chiseled, taut back and shoulders, each muscle trembling with exertion. The other man, facing the camera, has slightly longer, tangled black hair, with a few wet strands clinging to his forehead and cheeks. He wears a thoroughly soaked gray cotton T-shirt, tightly clung to his chest and abdomen, emphasizing his muscular physique. In a swift, turning, blocking motion, beads of sweat flung from their arms, hair, and cheeks. Under the harsh, single, cold light source from above, these splattering beads of sweat are momentarily illuminated, like countless tiny, shimmering crystals, drawing fleeting, bright arcs against the dim, rough concrete basement backdrop behind them. The entire scene is painted in a cool tone, dominated by the gray of concrete and the natural tones of the skin. The contrast between light and shadow is extremely strong, creating a primitive, tense and powerful fighting atmosphere.", + "seed": 26292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_5920_seed_26292_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05151533167809248, + "vlm_logprob": -0.05151533167809248, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05151533167809248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02769389934837818 + ], + "opposite_logprob_tokens": [ + -3.6526939868927 + ], + "logprob_reward": -0.02769389934837818, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02769389934837818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10182148218154907, + 0.06226455047726631, + 0.07812762260437012, + 0.07613726705312729, + 0.14321458339691162, + 0.06831102818250656, + 0.06893766671419144, + 0.13980935513973236, + 0.2124502956867218, + 0.04892618581652641 + ], + "credit_map": [ + 0.014012633822858334, + 0.01218682061880827, + 0.011919133365154266, + 0.011291873641312122, + 0.01052958145737648, + 0.012723308056592941, + 0.011962360702455044, + 0.017195770516991615, + 0.010035927407443523, + 0.007258694153279066, + 0.006907663773745298, + 0.00552909541875124, + 0.007328798528760672, + 0.009182758629322052, + 0.007257767487317324, + 0.00876383576542139, + 0.009199629537761211, + 0.014294938184320927, + 0.006326135713607073, + 0.005932822357863188, + 0.012633438222110271, + 0.011772197671234608, + 0.006959776859730482, + 0.011008677072823048, + 0.008428707718849182, + 0.010768108069896698, + 0.01272797491401434, + 0.0068833897821605206, + 0.01068294420838356, + 0.007833916693925858, + 0.007764148060232401, + 0.011048072949051857, + 0.015063788741827011, + 0.01662326790392399, + 0.018877482041716576, + 0.0158714447170496, + 0.015217027626931667, + 0.0268290713429451, + 0.01693425513803959, + 0.01779823750257492, + 0.007738464046269655, + 0.011444979347288609, + 0.007288909517228603, + 0.006482335738837719, + 0.011335819028317928, + 0.009439831599593163, + 0.007889420725405216, + 0.006691258866339922, + 0.007672626990824938, + 0.007753842510282993, + 0.01102440431714058, + 0.011152639053761959, + 0.006943679414689541, + 0.00846671313047409, + 0.007163202855736017, + 0.008760557509958744, + 0.012735798954963684, + 0.012984556145966053, + 0.025824112817645073, + 0.013209301978349686, + 0.009225851856172085, + 0.020112833008170128, + 0.025662008672952652, + 0.020054901018738747, + 0.011583810672163963, + 0.013744889758527279, + 0.054206203669309616, + 0.016557687893509865, + 0.06689866632223129, + 0.024265078827738762, + 0.011211088858544827, + 0.013982847332954407, + 0.0066828541457653046, + 0.004754102788865566, + 0.0035491716116666794, + 0.005489759147167206, + 0.006683149375021458, + 0.00648540910333395, + 0.009052661247551441, + 0.006229072343558073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.02769389934837818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011094429530203342 + ], + "opposite_logprob_tokens": [ + -4.511094570159912 + ], + "logprob_reward": -0.011094429530203342, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011094429530203342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2817452847957611, + 0.0566578283905983, + 0.0626104548573494, + 0.08700364083051682, + 0.0796162411570549, + 0.11049164831638336, + 0.05949597805738449, + 0.10913906246423721, + 0.07091137021780014, + 0.08232855796813965 + ], + "credit_map": [ + 0.03399340808391571, + 0.029030242934823036, + 0.024633020162582397, + 0.03319445252418518, + 0.04673612117767334, + 0.036811064928770065, + 0.045061562210321426, + 0.03228538855910301, + 0.011529289186000824, + 0.006611423566937447, + 0.006959306076169014, + 0.0042334310710430145, + 0.006850262638181448, + 0.009104850701987743, + 0.005132599268108606, + 0.006236657965928316, + 0.006345030851662159, + 0.007112161256372929, + 0.005627727136015892, + 0.0073290676809847355, + 0.010216746479272842, + 0.01005253940820694, + 0.007849573157727718, + 0.00807760376483202, + 0.006303742527961731, + 0.011942058801651001, + 0.01794898696243763, + 0.012198331765830517, + 0.006526560056954622, + 0.0066892351023852825, + 0.009577409364283085, + 0.01581730507314205, + 0.008898237720131874, + 0.014044247567653656, + 0.008847774937748909, + 0.01144962664693594, + 0.00806739367544651, + 0.009751723147928715, + 0.008541922084987164, + 0.010015313513576984, + 0.013329972513020039, + 0.021603362634778023, + 0.0070632747374475, + 0.0062318467535078526, + 0.0241033174097538, + 0.021628107875585556, + 0.007452169433236122, + 0.009079580195248127, + 0.0051734293811023235, + 0.004246083553880453, + 0.013212943449616432, + 0.0172684658318758, + 0.0044180541299283504, + 0.005335856229066849, + 0.004342790227383375, + 0.005498351063579321, + 0.009694377891719341, + 0.009277446195483208, + 0.024380411952733994, + 0.012439926154911518, + 0.00646559800952673, + 0.011890223249793053, + 0.013215637765824795, + 0.02177542634308338, + 0.0046642678789794445, + 0.0040806918404996395, + 0.013225672766566277, + 0.008958553895354271, + 0.01482586283236742, + 0.009934095665812492, + 0.009201481007039547, + 0.006020739208906889, + 0.0084006292745471, + 0.010306471958756447, + 0.007969185709953308, + 0.009487591683864594, + 0.014283839613199234, + 0.01274445652961731, + 0.01197870820760727, + 0.0071576666086912155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.011094429530203342 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043073106557130814 + ], + "opposite_logprob_tokens": [ + -3.1680731773376465 + ], + "logprob_reward": -0.043073106557130814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043073106557130814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08928629755973816, + 0.05675343796610832, + 0.06188308447599411, + 0.21796324849128723, + 0.05865342915058136, + 0.0629570260643959, + 0.1270320564508438, + 0.060746364295482635, + 0.05583137273788452, + 0.20889371633529663 + ], + "credit_map": [ + 0.012434618547558784, + 0.013984697870910168, + 0.008104832842946053, + 0.00956473033875227, + 0.013967294245958328, + 0.012961933389306068, + 0.009950101375579834, + 0.008318087086081505, + 0.011005973443388939, + 0.006906219758093357, + 0.007910208776593208, + 0.005297234281897545, + 0.004857701249420643, + 0.007830419577658176, + 0.006204016041010618, + 0.006741663441061974, + 0.008457595482468605, + 0.008050703443586826, + 0.006461962591856718, + 0.0064086513593792915, + 0.0073763467371463776, + 0.0091135548427701, + 0.007543344050645828, + 0.00847093015909195, + 0.01652020588517189, + 0.027026066556572914, + 0.04051967337727547, + 0.02754160203039646, + 0.016958534717559814, + 0.021590255200862885, + 0.03007661923766136, + 0.03773030266165733, + 0.005660633090883493, + 0.007904271595180035, + 0.0067094177938997746, + 0.0066332207061350346, + 0.007000532001256943, + 0.006296285893768072, + 0.009241596795618534, + 0.009207469411194324, + 0.00867276918143034, + 0.007465280592441559, + 0.0054188878275454044, + 0.006581338122487068, + 0.00910133309662342, + 0.012218941934406757, + 0.0070187123492360115, + 0.006479757372289896, + 0.010413417592644691, + 0.009567828848958015, + 0.025360537692904472, + 0.024795446544885635, + 0.01176730077713728, + 0.01403513178229332, + 0.011673934757709503, + 0.01941845938563347, + 0.00500316359102726, + 0.0056891362182796, + 0.011961991898715496, + 0.0070745381526649, + 0.004912095610052347, + 0.006116737145930529, + 0.01067376509308815, + 0.009314932860434055, + 0.00544205866754055, + 0.007597856689244509, + 0.007603636477142572, + 0.0053322031162679195, + 0.014522925950586796, + 0.007352490443736315, + 0.0037816439289599657, + 0.004198555834591389, + 0.025126582011580467, + 0.01825971156358719, + 0.01350738387554884, + 0.022040564566850662, + 0.03250666335225105, + 0.02655063383281231, + 0.03999652713537216, + 0.030905649065971375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.043073106557130814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706737220287323 + ], + "opposite_logprob_tokens": [ + -2.1270673274993896 + ], + "logprob_reward": -0.12706737220287323, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706737220287323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36267098784446716, + 0.14407266676425934, + 0.04616863653063774, + 0.04448486864566803, + 0.040413748472929, + 0.21627187728881836, + 0.029379183426499367, + 0.07245299965143204, + 0.02099701017141342, + 0.023088060319423676 + ], + "credit_map": [ + 0.07240191847085953, + 0.037097398191690445, + 0.04285666346549988, + 0.04124515503644943, + 0.03854924812912941, + 0.047741908580064774, + 0.03789199888706207, + 0.04488671198487282, + 0.02471686154603958, + 0.01623924821615219, + 0.021846385672688484, + 0.016064917668700218, + 0.011914496310055256, + 0.02622654289007187, + 0.013861863873898983, + 0.013202356174588203, + 0.0036860632244497538, + 0.009632852859795094, + 0.004108388442546129, + 0.004461337812244892, + 0.009416830725967884, + 0.006672009360045195, + 0.004558122251182795, + 0.003633033251389861, + 0.003775165416300297, + 0.008272355422377586, + 0.006835870444774628, + 0.00683927396312356, + 0.003787165740504861, + 0.004067349713295698, + 0.004204730503261089, + 0.006702961400151253, + 0.006476862821727991, + 0.004582742229104042, + 0.004078129306435585, + 0.005446340888738632, + 0.005536947399377823, + 0.003950401674956083, + 0.0047046891413629055, + 0.005637635011225939, + 0.02579485811293125, + 0.034456800669431686, + 0.015256505459547043, + 0.015595696866512299, + 0.04519861564040184, + 0.03913694620132446, + 0.02110782451927662, + 0.019724635407328606, + 0.0030994631815701723, + 0.0029653601814061403, + 0.004566947463899851, + 0.004495952744036913, + 0.0029234071262180805, + 0.0037278702948242426, + 0.0029061140958219767, + 0.004694068338721991, + 0.006028937641531229, + 0.005791413597762585, + 0.010486618615686893, + 0.006676793564110994, + 0.006223806645721197, + 0.013379750773310661, + 0.01637989841401577, + 0.007485776208341122, + 0.0018550363602116704, + 0.002071226481348276, + 0.0032479753717780113, + 0.00198876834474504, + 0.004000735469162464, + 0.0030668785329908133, + 0.0015954787377268076, + 0.0031709109898656607, + 0.0029838697519153357, + 0.0029167979955673218, + 0.002230027923360467, + 0.0026611287612468004, + 0.003022469813004136, + 0.0028382681775838137, + 0.003230446483939886, + 0.0032050511799752712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.12706737220287323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864785075187683 + ], + "opposite_logprob_tokens": [ + -3.048647880554199 + ], + "logprob_reward": -0.04864785075187683, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864785075187683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08125073462724686, + 0.05261461064219475, + 0.05078313872218132, + 0.11416027694940567, + 0.03864452615380287, + 0.14002986252307892, + 0.07116062194108963, + 0.12459391355514526, + 0.3055882155895233, + 0.021174127236008644 + ], + "credit_map": [ + 0.01559605821967125, + 0.010175476782023907, + 0.006840602960437536, + 0.007970177568495274, + 0.00960409827530384, + 0.011879215016961098, + 0.010439498350024223, + 0.00874561071395874, + 0.006297116633504629, + 0.006394139491021633, + 0.005349774844944477, + 0.003816619049757719, + 0.006614341400563717, + 0.007325893267989159, + 0.007132995408028364, + 0.009683730080723763, + 0.004749353509396315, + 0.006643958855420351, + 0.005490078590810299, + 0.0082329623401165, + 0.007505481597036123, + 0.005407795775681734, + 0.0052850814536213875, + 0.007468427065759897, + 0.009814610704779625, + 0.013203462585806847, + 0.01791701652109623, + 0.013274135999381542, + 0.011303910054266453, + 0.013492468744516373, + 0.01592383347451687, + 0.019230838865041733, + 0.004118010867387056, + 0.004571361001580954, + 0.004134668502956629, + 0.006521838717162609, + 0.0043157064355909824, + 0.004666508175432682, + 0.00599064351990819, + 0.00432578707113862, + 0.01666545495390892, + 0.0153757743537426, + 0.010783755220472813, + 0.011328821070492268, + 0.03388558328151703, + 0.028966989368200302, + 0.010683680884540081, + 0.01233980804681778, + 0.010119684971868992, + 0.0067985462956130505, + 0.010199026204645634, + 0.009189540520310402, + 0.008541037328541279, + 0.01029951497912407, + 0.005872283596545458, + 0.010140985250473022, + 0.011202346533536911, + 0.015316348522901535, + 0.01728403940796852, + 0.016544315963983536, + 0.010440990328788757, + 0.01906070113182068, + 0.021554507315158844, + 0.01319066807627678, + 0.018088867887854576, + 0.016186658293008804, + 0.08801373094320297, + 0.01985194906592369, + 0.10541319102048874, + 0.03197666257619858, + 0.012213051319122314, + 0.013844084925949574, + 0.002110449131578207, + 0.002059980295598507, + 0.002607816830277443, + 0.0029951827600598335, + 0.002965947613120079, + 0.002084041014313698, + 0.004094207193702459, + 0.002256502863019705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.04864785075187683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld shot closely follows two Asian men engaged in a fierce exchange, the camera panning with their movements to create an immersive sense of chaos. The man in the foreground, with short, sweat-soaked black hair plastered to his scalp, is shirtless, revealing a chiseled, taut back and shoulders, each muscle trembling with exertion. The other man, facing the camera, has slightly longer, tangled black hair, with a few wet strands clinging to his forehead and cheeks. He wears a thoroughly soaked gray cotton T-shirt, tightly clung to his chest and abdomen, emphasizing his muscular physique. In a swift, turning, blocking motion, beads of sweat flung from their arms, hair, and cheeks. Under the harsh, single, cold light source from above, these splattering beads of sweat are momentarily illuminated, like countless tiny, shimmering crystals, drawing fleeting, bright arcs against the dim, rough concrete basement backdrop behind them. The entire scene is painted in a cool tone, dominated by the gray of concrete and the natural tones of the skin. The contrast between light and shadow is extremely strong, creating a primitive, tense and powerful fighting atmosphere.", + "seed": 26293, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_5920_seed_26293_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025213261879980565, + "vlm_logprob": -0.025213261879980565, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.025213261879980565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02857220731675625 + ], + "opposite_logprob_tokens": [ + -3.6535723209381104 + ], + "logprob_reward": -0.02857220731675625, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02857220731675625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13246183097362518, + 0.1374874711036682, + 0.050693947821855545, + 0.06862543523311615, + 0.10174648463726044, + 0.13112691044807434, + 0.12091939896345139, + 0.147872656583786, + 0.04342322796583176, + 0.06564266979694366 + ], + "credit_map": [ + 0.011261933483183384, + 0.017475247383117676, + 0.019661560654640198, + 0.017746245488524437, + 0.02352282404899597, + 0.010897182859480381, + 0.010655137710273266, + 0.0212416909635067, + 0.015927383676171303, + 0.025577574968338013, + 0.008823188953101635, + 0.008142584003508091, + 0.024649890139698982, + 0.02808164618909359, + 0.014066675677895546, + 0.012218520976603031, + 0.005491528194397688, + 0.0070173172280192375, + 0.007278327830135822, + 0.0058431075885891914, + 0.011095023714005947, + 0.006596258841454983, + 0.0034706895239651203, + 0.0039016876835376024, + 0.008573119528591633, + 0.010284129530191422, + 0.009543214924633503, + 0.01185637153685093, + 0.010449103079736233, + 0.006054002791643143, + 0.006555194966495037, + 0.005310294684022665, + 0.013066766783595085, + 0.011914851143956184, + 0.014584476128220558, + 0.01585923135280609, + 0.013060980476439, + 0.01214879285544157, + 0.010988415218889713, + 0.010122963227331638, + 0.014200371690094471, + 0.03153488039970398, + 0.008326132781803608, + 0.01801452971994877, + 0.022448362782597542, + 0.01409144140779972, + 0.009937099181115627, + 0.012574074789881706, + 0.011463313363492489, + 0.015244308859109879, + 0.012706940062344074, + 0.020443765446543694, + 0.019832907244563103, + 0.021337753161787987, + 0.010406897403299809, + 0.00948349479585886, + 0.013588771224021912, + 0.030985698103904724, + 0.02006169594824314, + 0.024118177592754364, + 0.02133467048406601, + 0.017137037590146065, + 0.009382130578160286, + 0.011264451779425144, + 0.0036113273818045855, + 0.005686732940375805, + 0.006797747686505318, + 0.005169256124645472, + 0.004236678127199411, + 0.007652478292584419, + 0.006965861190110445, + 0.003303144359961152, + 0.005166883114725351, + 0.0058974833227694035, + 0.00698348181322217, + 0.01446791086345911, + 0.012697454541921616, + 0.010317284613847733, + 0.005423591937869787, + 0.004688573069870472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.02857220731675625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646103389561176 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.007646103389561176, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646103389561176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07412365078926086, + 0.05353500321507454, + 0.14562591910362244, + 0.14355574548244476, + 0.026198478415608406, + 0.02703048847615719, + 0.041050396859645844, + 0.4170582890510559, + 0.0303107351064682, + 0.04151124879717827 + ], + "credit_map": [ + 0.00799823459237814, + 0.010287952609360218, + 0.010432137176394463, + 0.008557727560400963, + 0.015071383677423, + 0.005983427632600069, + 0.007822566665709019, + 0.007970228791236877, + 0.0047845784574747086, + 0.009037327021360397, + 0.00682357931509614, + 0.00725230248644948, + 0.008850008249282837, + 0.006498911418020725, + 0.004998006392270327, + 0.0052902973257005215, + 0.017100196331739426, + 0.01144765317440033, + 0.007468272466212511, + 0.005786599591374397, + 0.052356477826833725, + 0.04201582819223404, + 0.005211086943745613, + 0.0042398348450660706, + 0.024153079837560654, + 0.015295417048037052, + 0.01358440238982439, + 0.028859836980700493, + 0.022443313151597977, + 0.014727028086781502, + 0.01292468048632145, + 0.011567995883524418, + 0.0030102264136075974, + 0.002855714876204729, + 0.004104434046894312, + 0.0036031052004545927, + 0.0038826402742415667, + 0.0026263052131980658, + 0.0033771833404898643, + 0.002738871378824115, + 0.0027973849792033434, + 0.005994693376123905, + 0.0018451372161507607, + 0.0030298687051981688, + 0.005830368958413601, + 0.0025544846430420876, + 0.0028459143359214067, + 0.0021326367277652025, + 0.0030791086610406637, + 0.004046791698783636, + 0.00666004279628396, + 0.007118061184883118, + 0.006123298313468695, + 0.005466366186738014, + 0.005259642377495766, + 0.0032970926258713007, + 0.03617795929312706, + 0.06745680421590805, + 0.05406433716416359, + 0.08441944420337677, + 0.07054906338453293, + 0.04002842307090759, + 0.026584330946207047, + 0.03777798265218735, + 0.002190339844673872, + 0.004289031028747559, + 0.004158170893788338, + 0.005544204730540514, + 0.004376176744699478, + 0.004076226614415646, + 0.003518836572766304, + 0.0021577514708042145, + 0.002971135312691331, + 0.004321887157857418, + 0.005314311943948269, + 0.008230581879615784, + 0.00813454482704401, + 0.00627662381157279, + 0.0034777831751853228, + 0.0027843809220939875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.007646103389561176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370870277285576 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370870277285576, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370870277285576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043558280915021896, + 0.09077920019626617, + 0.6542391777038574, + 0.03685132786631584, + 0.04125836119055748, + 0.0369485504925251, + 0.034185245633125305, + 0.02737277001142502, + 0.019009655341506004, + 0.01579747349023819 + ], + "credit_map": [ + 0.005117354448884726, + 0.0066140457056462765, + 0.00332318851724267, + 0.006839899346232414, + 0.007541429717093706, + 0.0042679887264966965, + 0.00531278969720006, + 0.004541587084531784, + 0.010539347305893898, + 0.011942720040678978, + 0.008955123834311962, + 0.006194612011313438, + 0.012171580456197262, + 0.021394604817032814, + 0.010420804843306541, + 0.009160407818853855, + 0.08474187552928925, + 0.04279644414782524, + 0.020982403308153152, + 0.010958679020404816, + 0.2504138946533203, + 0.20421892404556274, + 0.02319532446563244, + 0.016931621357798576, + 0.00672933692112565, + 0.00322513934224844, + 0.003619593335315585, + 0.007346022874116898, + 0.0049066063947975636, + 0.004887585062533617, + 0.0031283877324312925, + 0.0030086578335613012, + 0.006936389021575451, + 0.005482955370098352, + 0.007532828953117132, + 0.0036900327540934086, + 0.006105208769440651, + 0.00517243891954422, + 0.0035922585520893335, + 0.0027462500147521496, + 0.00452515110373497, + 0.004325114190578461, + 0.004043117631226778, + 0.004484873265028, + 0.005711840931326151, + 0.004392753355205059, + 0.0060962806455791, + 0.003369421698153019, + 0.002696210751309991, + 0.0035351242404431105, + 0.006513578351587057, + 0.0063313874416053295, + 0.0050811381079256535, + 0.004426922183483839, + 0.002795868320390582, + 0.002805017866194248, + 0.0023225604090839624, + 0.004208367317914963, + 0.0029426338151097298, + 0.004957530647516251, + 0.004796259105205536, + 0.0025039806496351957, + 0.002363506704568863, + 0.0032779339235275984, + 0.0017215653788298368, + 0.002702579600736499, + 0.0020198270212858915, + 0.001989913871511817, + 0.0022029231768101454, + 0.004468291066586971, + 0.002056859200820327, + 0.0018476956756785512, + 0.001672124257311225, + 0.0015085452469065785, + 0.0013146591372787952, + 0.003246152773499489, + 0.0026966207660734653, + 0.0020336988382041454, + 0.0019460262265056372, + 0.0013796471757814288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03370870277285576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797280207276344 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797280207276344, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797280207276344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1675776243209839, + 0.04370991140604019, + 0.5424972772598267, + 0.023155661299824715, + 0.0874524638056755, + 0.0234555434435606, + 0.02382786013185978, + 0.04014302045106888, + 0.02884722873568535, + 0.019333425909280777 + ], + "credit_map": [ + 0.015064823441207409, + 0.0376054123044014, + 0.01970091089606285, + 0.017239484935998917, + 0.03268571197986603, + 0.011880644597113132, + 0.01954948902130127, + 0.013851135037839413, + 0.005405586212873459, + 0.00722652580589056, + 0.0022909233812242746, + 0.0029815880116075277, + 0.009115769527852535, + 0.010159132070839405, + 0.002644371474161744, + 0.0038860177155584097, + 0.0694497674703598, + 0.036100734025239944, + 0.016779441386461258, + 0.009091540239751339, + 0.20607909560203552, + 0.17141611874103546, + 0.01973770186305046, + 0.013842840678989887, + 0.002993086352944374, + 0.0021831076592206955, + 0.0019639970269054174, + 0.0036677077878266573, + 0.002454501111060381, + 0.0048372987657785416, + 0.002289524534717202, + 0.002766437828540802, + 0.011696294881403446, + 0.012889857403934002, + 0.006984899751842022, + 0.007583790924400091, + 0.01446235366165638, + 0.013446344994008541, + 0.007464275229722261, + 0.012924646027386189, + 0.0027760048396885395, + 0.006332968827337027, + 0.001961088739335537, + 0.003445939626544714, + 0.003348894417285919, + 0.0018071176018565893, + 0.0018267643172293901, + 0.001956766238436103, + 0.002033090451732278, + 0.002941783284768462, + 0.00265114801004529, + 0.004182402044534683, + 0.003559533506631851, + 0.002850388176739216, + 0.002924719825387001, + 0.0026847952976822853, + 0.003645320888608694, + 0.005944411735981703, + 0.004872204270213842, + 0.007382651790976524, + 0.006535123102366924, + 0.004368599504232407, + 0.004092298448085785, + 0.0033024149015545845, + 0.002511455677449703, + 0.004239500965923071, + 0.004330239724367857, + 0.00369519111700356, + 0.0038998115342110395, + 0.004197461996227503, + 0.0032860501669347286, + 0.002687518484890461, + 0.0016009386163204908, + 0.001610560342669487, + 0.0016281175194308162, + 0.004461122211068869, + 0.0036149059887975454, + 0.002256730105727911, + 0.0020508249290287495, + 0.0021102274768054485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.029797280207276344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342015713453293 + ], + "opposite_logprob_tokens": [ + -3.6513419151306152 + ], + "logprob_reward": -0.026342015713453293, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342015713453293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055034931749105453, + 0.024777358397841454, + 0.7812791466712952, + 0.027996238321065903, + 0.010007670149207115, + 0.012819018214941025, + 0.021677574142813683, + 0.01947231963276863, + 0.010957635007798672, + 0.035978082567453384 + ], + "credit_map": [ + 0.00522557832300663, + 0.006962700746953487, + 0.008623957633972168, + 0.0074880351312458515, + 0.009714487008750439, + 0.00456306803971529, + 0.006486514583230019, + 0.005970597732812166, + 0.002430648310109973, + 0.0034360524732619524, + 0.0023183233570307493, + 0.001809878391213715, + 0.004196364898234606, + 0.00276880688033998, + 0.003311675973236561, + 0.00450561149045825, + 0.10093480348587036, + 0.05168338119983673, + 0.02444983460009098, + 0.013079720549285412, + 0.2979384660720825, + 0.24566331505775452, + 0.027759091928601265, + 0.019770588725805283, + 0.0042599341832101345, + 0.002677810611203313, + 0.002338576130568981, + 0.005566198844462633, + 0.003920830320566893, + 0.003863294841721654, + 0.0027214069850742817, + 0.0026481880340725183, + 0.0013355585979297757, + 0.0010736997937783599, + 0.0019177838694304228, + 0.0010595639469102025, + 0.0015221686335280538, + 0.0009840208804234862, + 0.0012049454962834716, + 0.0009099295712076128, + 0.0017127710161730647, + 0.0030619993340224028, + 0.0010536275804042816, + 0.0017969416221603751, + 0.0019097125623375177, + 0.001082991948351264, + 0.00126713200006634, + 0.0009338433155789971, + 0.0023630550131201744, + 0.002553668338805437, + 0.002425922779366374, + 0.0032997275702655315, + 0.0029919741209596395, + 0.0025787390768527985, + 0.002377614611759782, + 0.003086875891312957, + 0.0017573748482391238, + 0.0018984762718901038, + 0.0020335884764790535, + 0.002990748966112733, + 0.0040907845832407475, + 0.002700806362554431, + 0.0026181149296462536, + 0.0013824260095134377, + 0.000824630435090512, + 0.0016911140410229564, + 0.0013821061002090573, + 0.0013347213389351964, + 0.002290930598974228, + 0.0011538587277755141, + 0.0013688167091459036, + 0.0009114588610827923, + 0.0032171306665986776, + 0.004524082876741886, + 0.006090070586651564, + 0.0042027817107737064, + 0.005792554467916489, + 0.003943020477890968, + 0.003921739291399717, + 0.004286704119294882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.026342015713453293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld shot closely follows two Asian men engaged in a fierce exchange, the camera panning with their movements to create an immersive sense of chaos. The man in the foreground, with short, sweat-soaked black hair plastered to his scalp, is shirtless, revealing a chiseled, taut back and shoulders, each muscle trembling with exertion. The other man, facing the camera, has slightly longer, tangled black hair, with a few wet strands clinging to his forehead and cheeks. He wears a thoroughly soaked gray cotton T-shirt, tightly clung to his chest and abdomen, emphasizing his muscular physique. In a swift, turning, blocking motion, beads of sweat flung from their arms, hair, and cheeks. Under the harsh, single, cold light source from above, these splattering beads of sweat are momentarily illuminated, like countless tiny, shimmering crystals, drawing fleeting, bright arcs against the dim, rough concrete basement backdrop behind them. The entire scene is painted in a cool tone, dominated by the gray of concrete and the natural tones of the skin. The contrast between light and shadow is extremely strong, creating a primitive, tense and powerful fighting atmosphere.", + "seed": 26294, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_5920_seed_26294_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026344380481168626, + "vlm_logprob": -0.026344380481168626, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.026344380481168626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028549153357744217 + ], + "opposite_logprob_tokens": [ + -3.6535491943359375 + ], + "logprob_reward": -0.028549153357744217, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028549153357744217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10226861387491226, + 0.15598325431346893, + 0.12257974594831467, + 0.1293032467365265, + 0.1512119024991989, + 0.10747602581977844, + 0.06657294183969498, + 0.03231043741106987, + 0.0779954046010971, + 0.05429845303297043 + ], + "credit_map": [ + 0.016761595383286476, + 0.017066914588212967, + 0.0113823926076293, + 0.011251126416027546, + 0.010602112859487534, + 0.01237199641764164, + 0.01016926858574152, + 0.012663218192756176, + 0.01765771210193634, + 0.02090834453701973, + 0.01927463710308075, + 0.021266521885991096, + 0.025158971548080444, + 0.016059253364801407, + 0.017654795199632645, + 0.018003037199378014, + 0.01625230722129345, + 0.019596124067902565, + 0.013501579873263836, + 0.018668046221137047, + 0.019615601748228073, + 0.011964973993599415, + 0.009166207164525986, + 0.013814921490848064, + 0.017864560708403587, + 0.011926481500267982, + 0.010487912222743034, + 0.012942356057465076, + 0.029643315821886063, + 0.015009637922048569, + 0.01437661238014698, + 0.017052385956048965, + 0.019123058766126633, + 0.02444061078131199, + 0.01110272016376257, + 0.01928933709859848, + 0.03673635050654411, + 0.011842025443911552, + 0.018769150599837303, + 0.009908647276461124, + 0.014525161124765873, + 0.01187215931713581, + 0.00914819911122322, + 0.008080177940428257, + 0.01284993439912796, + 0.03343363106250763, + 0.0071897427551448345, + 0.010377040132880211, + 0.005860855802893639, + 0.00981628242880106, + 0.009013663046061993, + 0.0083619374781847, + 0.007076573092490435, + 0.00942216720432043, + 0.008355614729225636, + 0.00866585411131382, + 0.004485381301492453, + 0.0026834961026906967, + 0.003685681615024805, + 0.004864031914621592, + 0.0025527412071824074, + 0.004025883041322231, + 0.006384183652698994, + 0.0036290402058511972, + 0.010052639059722424, + 0.014532742090523243, + 0.01163777057081461, + 0.011565676890313625, + 0.011217108927667141, + 0.007349641993641853, + 0.006414233241230249, + 0.005225596483796835, + 0.004137245938181877, + 0.009558722376823425, + 0.0036824934650212526, + 0.005795046221464872, + 0.009443807415664196, + 0.011697931215167046, + 0.0040090084075927734, + 0.005974206142127514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.028549153357744217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036354658659547567 + ], + "opposite_logprob_tokens": [ + -5.628635406494141 + ], + "logprob_reward": -0.0036354658659547567, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036354658659547567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10883652418851852, + 0.10415162146091461, + 0.0931977704167366, + 0.10392698645591736, + 0.29946696758270264, + 0.13826583325862885, + 0.02977983094751835, + 0.02867848426103592, + 0.059220269322395325, + 0.03447573632001877 + ], + "credit_map": [ + 0.01661830022931099, + 0.01721663400530815, + 0.01480428408831358, + 0.009627322666347027, + 0.011472204700112343, + 0.01480888482183218, + 0.010135168209671974, + 0.014153718017041683, + 0.012857402674853802, + 0.011831787414848804, + 0.010738175362348557, + 0.013230221346020699, + 0.017367076128721237, + 0.009917276911437511, + 0.013112165965139866, + 0.015097512863576412, + 0.012293647043406963, + 0.011506030336022377, + 0.006503977347165346, + 0.012079993262887001, + 0.019513549283146858, + 0.010520569048821926, + 0.008550740778446198, + 0.012229269370436668, + 0.01268327422440052, + 0.009532165713608265, + 0.013370923697948456, + 0.01692231371998787, + 0.01120414212346077, + 0.012898280285298824, + 0.015204168856143951, + 0.012111714109778404, + 0.03634927049279213, + 0.07197166979312897, + 0.025068748742341995, + 0.035239387303590775, + 0.049255941063165665, + 0.032031986862421036, + 0.028495129197835922, + 0.021054837852716446, + 0.018560241907835007, + 0.0160843376070261, + 0.012326857075095177, + 0.011854841373860836, + 0.016364336013793945, + 0.03899737820029259, + 0.011519213207066059, + 0.01255863718688488, + 0.002759336493909359, + 0.005383079405874014, + 0.003115171566605568, + 0.004862895701080561, + 0.0028573754243552685, + 0.00431470200419426, + 0.0033977353014051914, + 0.0030895364470779896, + 0.0044310311786830425, + 0.002028032671660185, + 0.0030352084431797266, + 0.0037013008259236813, + 0.0028368018101900816, + 0.005749564617872238, + 0.003526555374264717, + 0.0033699891064316034, + 0.007732527330517769, + 0.01124831847846508, + 0.006451350171118975, + 0.007314078044146299, + 0.011333820410072803, + 0.007060912437736988, + 0.004020927008241415, + 0.00405833451077342, + 0.002875040750950575, + 0.004820743575692177, + 0.002521342132240534, + 0.0055000497959554195, + 0.004722221754491329, + 0.00495405588299036, + 0.0041339038871228695, + 0.004948379006236792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0036354658659547567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329389750957489 + ], + "opposite_logprob_tokens": [ + -3.773293972015381 + ], + "logprob_reward": -0.02329389750957489, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329389750957489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450175940990448, + 0.22444379329681396, + 0.1927122324705124, + 0.16396991908550262, + 0.06664237380027771, + 0.027535442262887955, + 0.03387827426195145, + 0.04225115478038788, + 0.0753488764166832, + 0.028200291097164154 + ], + "credit_map": [ + 0.023678449913859367, + 0.014161689206957817, + 0.014076169580221176, + 0.009713761508464813, + 0.020994452759623528, + 0.03296119347214699, + 0.009677495807409286, + 0.019754374399781227, + 0.02697726897895336, + 0.033970631659030914, + 0.015293668024241924, + 0.023520048707723618, + 0.04570318013429642, + 0.02535928413271904, + 0.02937924861907959, + 0.024240463972091675, + 0.02824985608458519, + 0.021240754052996635, + 0.013349851593375206, + 0.020460674539208412, + 0.05488103628158569, + 0.021537676453590393, + 0.016091834753751755, + 0.016900556161999702, + 0.020547829568386078, + 0.017116693779826164, + 0.014850758947432041, + 0.01588270626962185, + 0.041855424642562866, + 0.014361985959112644, + 0.022209670394659042, + 0.01714484952390194, + 0.00886716041713953, + 0.01433352380990982, + 0.004494973924010992, + 0.006865073461085558, + 0.010449238121509552, + 0.0052118003368377686, + 0.009595918469130993, + 0.006824679672718048, + 0.003399577923119068, + 0.004071298986673355, + 0.0028073270805180073, + 0.0030588191002607346, + 0.002909097820520401, + 0.005278118420392275, + 0.0030766415875405073, + 0.002934562275186181, + 0.0025901442859321833, + 0.0034560265485197306, + 0.004383903928101063, + 0.00593725498765707, + 0.004422766622155905, + 0.005151901859790087, + 0.004232359118759632, + 0.003703915514051914, + 0.006586441304534674, + 0.004392772447317839, + 0.006787641439586878, + 0.00516020180657506, + 0.004578133579343557, + 0.003957351669669151, + 0.006889125797897577, + 0.003899484174326062, + 0.013158023357391357, + 0.008804557844996452, + 0.004322715103626251, + 0.006854027509689331, + 0.02087279222905636, + 0.008648666553199291, + 0.005726750940084457, + 0.006961344741284847, + 0.0024415310472249985, + 0.003930015489459038, + 0.0020924450363963842, + 0.0037742210552096367, + 0.004199292976409197, + 0.003687740070745349, + 0.0036832941696047783, + 0.004391749855130911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.02329389750957489 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205231696367264 + ], + "opposite_logprob_tokens": [ + -2.8120522499084473 + ], + "logprob_reward": -0.06205231696367264, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205231696367264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12512043118476868, + 0.09219049662351608, + 0.2998887598514557, + 0.051214996725320816, + 0.0983356386423111, + 0.05445270240306854, + 0.06684549152851105, + 0.05521631985902786, + 0.10580893605947495, + 0.05092623084783554 + ], + "credit_map": [ + 0.01909385621547699, + 0.010551673360168934, + 0.008502285927534103, + 0.00784136913716793, + 0.019618837162852287, + 0.028841502964496613, + 0.012618020176887512, + 0.018052875995635986, + 0.011469144374132156, + 0.01586601510643959, + 0.00693572498857975, + 0.009198184125125408, + 0.016868961974978447, + 0.010013481602072716, + 0.009090992622077465, + 0.012747986242175102, + 0.045596808195114136, + 0.023521194234490395, + 0.016505151987075806, + 0.024484066292643547, + 0.09692941606044769, + 0.03148108348250389, + 0.027637634426355362, + 0.033733393996953964, + 0.005529087968170643, + 0.006875365041196346, + 0.005465143825858831, + 0.004196684341877699, + 0.007322135381400585, + 0.006323252804577351, + 0.007569604087620974, + 0.0079337228089571, + 0.013239788822829723, + 0.016397545114159584, + 0.004269493743777275, + 0.010225215926766396, + 0.022369498386979103, + 0.009639584459364414, + 0.014284305274486542, + 0.007910199463367462, + 0.008971736766397953, + 0.005844976752996445, + 0.004960513208061457, + 0.005622177850455046, + 0.004217477515339851, + 0.014475006610155106, + 0.0042340802028775215, + 0.006126727443188429, + 0.007892023772001266, + 0.0051678442396223545, + 0.008995059877634048, + 0.009495356120169163, + 0.007559987250715494, + 0.011263707652688026, + 0.010352818295359612, + 0.006118696182966232, + 0.016504209488630295, + 0.005833273753523827, + 0.009226634167134762, + 0.004367940127849579, + 0.003902359865605831, + 0.006185006350278854, + 0.0050625926814973354, + 0.004134300630539656, + 0.017604026943445206, + 0.013628290966153145, + 0.007576837204396725, + 0.009752623736858368, + 0.027779042720794678, + 0.012381997890770435, + 0.007255559787154198, + 0.00983054842799902, + 0.004310958553105593, + 0.006458885036408901, + 0.003245854750275612, + 0.0058328527957201, + 0.00820008572191, + 0.007461361587047577, + 0.00600229948759079, + 0.009413927793502808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.06205231696367264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191068708896637 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191068708896637, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191068708896637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23858630657196045, + 0.23606152832508087, + 0.062219783663749695, + 0.062154702842235565, + 0.06122034788131714, + 0.08459855616092682, + 0.04927254840731621, + 0.03530261293053627, + 0.12859894335269928, + 0.041984692215919495 + ], + "credit_map": [ + 0.025487814098596573, + 0.038444407284259796, + 0.026457956060767174, + 0.020041806623339653, + 0.0322246290743351, + 0.031692903488874435, + 0.026622265577316284, + 0.037614524364471436, + 0.027761243283748627, + 0.040915656834840775, + 0.020554272457957268, + 0.0290417168289423, + 0.03722298517823219, + 0.025766361504793167, + 0.028359785676002502, + 0.026439493522047997, + 0.009630541317164898, + 0.010909068398177624, + 0.004667259287089109, + 0.007486902177333832, + 0.009491205215454102, + 0.006396946031600237, + 0.007453238591551781, + 0.0061846268363296986, + 0.007341803051531315, + 0.007330469321459532, + 0.00483908224850893, + 0.006443234160542488, + 0.010286463424563408, + 0.009153619408607483, + 0.007802004460245371, + 0.008958026766777039, + 0.005400220397859812, + 0.006978780031204224, + 0.0027818770613521338, + 0.00655763503164053, + 0.017073405906558037, + 0.008552169427275658, + 0.00892714038491249, + 0.004949120804667473, + 0.012809640727937222, + 0.008850706741213799, + 0.008114570751786232, + 0.007897542789578438, + 0.010653210803866386, + 0.022165076807141304, + 0.006042120046913624, + 0.008065685629844666, + 0.005058764945715666, + 0.0046934406273067, + 0.005697412882000208, + 0.007813478820025921, + 0.0067700715735554695, + 0.008835343644022942, + 0.005633351393043995, + 0.0047706845216453075, + 0.005436734762042761, + 0.004263038747012615, + 0.004560235422104597, + 0.004490641877055168, + 0.003679499262943864, + 0.004547282122075558, + 0.004708024673163891, + 0.003617154899984598, + 0.022881537675857544, + 0.015551247633993626, + 0.010260328650474548, + 0.011333015747368336, + 0.03417398780584335, + 0.016276637092232704, + 0.007863239385187626, + 0.010258959606289864, + 0.004347674548625946, + 0.006355480290949345, + 0.0032688428182154894, + 0.00638620788231492, + 0.005348471459001303, + 0.005555520299822092, + 0.005254276562482119, + 0.005468220915645361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014191068708896637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld shot closely follows two Asian men engaged in a fierce exchange, the camera panning with their movements to create an immersive sense of chaos. The man in the foreground, with short, sweat-soaked black hair plastered to his scalp, is shirtless, revealing a chiseled, taut back and shoulders, each muscle trembling with exertion. The other man, facing the camera, has slightly longer, tangled black hair, with a few wet strands clinging to his forehead and cheeks. He wears a thoroughly soaked gray cotton T-shirt, tightly clung to his chest and abdomen, emphasizing his muscular physique. In a swift, turning, blocking motion, beads of sweat flung from their arms, hair, and cheeks. Under the harsh, single, cold light source from above, these splattering beads of sweat are momentarily illuminated, like countless tiny, shimmering crystals, drawing fleeting, bright arcs against the dim, rough concrete basement backdrop behind them. The entire scene is painted in a cool tone, dominated by the gray of concrete and the natural tones of the skin. The contrast between light and shadow is extremely strong, creating a primitive, tense and powerful fighting atmosphere.", + "seed": 26295, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_12/index_5920_seed_26295_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026652260310947896, + "vlm_logprob": -0.026652260310947896, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.026652260310947896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010168297216296196 + ], + "opposite_logprob_tokens": [ + -4.760168075561523 + ], + "logprob_reward": -0.010168297216296196, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010168297216296196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12363173067569733, + 0.08811387419700623, + 0.30282753705978394, + 0.12122279405593872, + 0.07405992597341537, + 0.07310034334659576, + 0.045176196843385696, + 0.07618430256843567, + 0.06585851311683655, + 0.029824763536453247 + ], + "credit_map": [ + 0.016461297869682312, + 0.025290746241807938, + 0.01249857246875763, + 0.015737831592559814, + 0.017702240496873856, + 0.01026658620685339, + 0.017169056460261345, + 0.008505409583449364, + 0.005928040947765112, + 0.023607837036252022, + 0.006079001352190971, + 0.008039461448788643, + 0.026803018525242805, + 0.008443090133368969, + 0.004906497895717621, + 0.004306931048631668, + 0.01822015643119812, + 0.052648499608039856, + 0.039866939187049866, + 0.03329892456531525, + 0.052078358829021454, + 0.03540223091840744, + 0.03806394711136818, + 0.03324846550822258, + 0.012318127788603306, + 0.026325123384594917, + 0.011513276025652885, + 0.010238276794552803, + 0.01870625838637352, + 0.017850495874881744, + 0.013644233345985413, + 0.010627019219100475, + 0.008603357709944248, + 0.010999185964465141, + 0.011651353910565376, + 0.009507017210125923, + 0.009451944380998611, + 0.009975922293961048, + 0.006937619764357805, + 0.006933530792593956, + 0.00857727974653244, + 0.008676150813698769, + 0.007266469765454531, + 0.020003268495202065, + 0.007594619411975145, + 0.008410661481320858, + 0.005988843273371458, + 0.006583059206604958, + 0.004076815210282803, + 0.0058629619888961315, + 0.005651747342199087, + 0.010454832576215267, + 0.003987442702054977, + 0.005978486035019159, + 0.005605306476354599, + 0.0035586084704846144, + 0.0075708674266934395, + 0.013380461372435093, + 0.007074312772601843, + 0.012261181138455868, + 0.015130771324038506, + 0.008783577010035515, + 0.004969078116118908, + 0.0070140608586370945, + 0.005964689422398806, + 0.0091282082721591, + 0.004322446882724762, + 0.006038090214133263, + 0.009001032449305058, + 0.021657535806298256, + 0.005189066752791405, + 0.004557452630251646, + 0.004546706564724445, + 0.004320645704865456, + 0.0030088808853179216, + 0.003451337106525898, + 0.0034457226283848286, + 0.005881167016923428, + 0.0022549822460860014, + 0.002915323479101062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.010168297216296196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751463748514652 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751463748514652, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751463748514652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07905857264995575, + 0.24474990367889404, + 0.22066915035247803, + 0.0728086456656456, + 0.06288550794124603, + 0.04560503363609314, + 0.03226806968450546, + 0.0740257278084755, + 0.06366008520126343, + 0.10426928848028183 + ], + "credit_map": [ + 0.009680308401584625, + 0.015713190659880638, + 0.007191918790340424, + 0.00851281825453043, + 0.009932826273143291, + 0.008622950874269009, + 0.011081463657319546, + 0.008323101326823235, + 0.013242603279650211, + 0.07283487170934677, + 0.017354290932416916, + 0.021971866488456726, + 0.08058761060237885, + 0.015231947414577007, + 0.013054502196609974, + 0.010472222231328487, + 0.01577918976545334, + 0.03702717274427414, + 0.025566402822732925, + 0.024255851283669472, + 0.03884855657815933, + 0.02984602376818657, + 0.02438458800315857, + 0.024961376562714577, + 0.007582110818475485, + 0.01103887427598238, + 0.007665109820663929, + 0.007583849597722292, + 0.014669179916381836, + 0.008300798945128918, + 0.009009613655507565, + 0.006959114223718643, + 0.007976419292390347, + 0.008995969779789448, + 0.012217394076287746, + 0.008221886120736599, + 0.007612618152052164, + 0.007084702141582966, + 0.004885668866336346, + 0.005890851840376854, + 0.0069501884281635284, + 0.005334026180207729, + 0.012761850841343403, + 0.008721482008695602, + 0.0032643110025674105, + 0.0037135733291506767, + 0.002643082058057189, + 0.002216522814705968, + 0.003100157715380192, + 0.0054102251306176186, + 0.003914636559784412, + 0.004102530889213085, + 0.0037242230027914047, + 0.004785412456840277, + 0.0033547095954418182, + 0.003876176429912448, + 0.006346807349473238, + 0.011339910328388214, + 0.0073338644579052925, + 0.013711519539356232, + 0.014806755818426609, + 0.008940235711634159, + 0.005567389540374279, + 0.005979250185191631, + 0.00859046634286642, + 0.009996278211474419, + 0.006625349633395672, + 0.007745914626866579, + 0.005155737046152353, + 0.013512341305613518, + 0.006766309030354023, + 0.005267696920782328, + 0.015189645811915398, + 0.015970557928085327, + 0.015210661105811596, + 0.02209329418838024, + 0.01230164710432291, + 0.007158411666750908, + 0.00801832601428032, + 0.008326747454702854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.006751463748514652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038105446845293045 + ], + "opposite_logprob_tokens": [ + -3.2881054878234863 + ], + "logprob_reward": -0.038105446845293045, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038105446845293045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13144975900650024, + 0.2057153880596161, + 0.12431792914867401, + 0.09627438336610794, + 0.0663866475224495, + 0.08513423800468445, + 0.05423445999622345, + 0.06391564011573792, + 0.10780005156993866, + 0.06477152556180954 + ], + "credit_map": [ + 0.022881614044308662, + 0.018587250262498856, + 0.016915645450353622, + 0.03110726736485958, + 0.011948288418352604, + 0.009191120974719524, + 0.012706074863672256, + 0.008112485520541668, + 0.0202208012342453, + 0.060608018189668655, + 0.03171911835670471, + 0.021810324862599373, + 0.02537212148308754, + 0.02440795861184597, + 0.01034044474363327, + 0.011236594058573246, + 0.010509120300412178, + 0.014370502904057503, + 0.014431534335017204, + 0.01787423901259899, + 0.0164153091609478, + 0.02161727473139763, + 0.0165005624294281, + 0.012599381618201733, + 0.008298098109662533, + 0.012813849374651909, + 0.008844496682286263, + 0.00974331796169281, + 0.019772937521338463, + 0.01375868171453476, + 0.012656779028475285, + 0.010386223904788494, + 0.00714380806311965, + 0.009357311762869358, + 0.009935718029737473, + 0.008170517161488533, + 0.011574352160096169, + 0.009536796249449253, + 0.0059906900860369205, + 0.004677455872297287, + 0.021065058186650276, + 0.008702286519110203, + 0.021240223199129105, + 0.0103072514757514, + 0.0073364004492759705, + 0.0064923809841275215, + 0.004822772461920977, + 0.005167861469089985, + 0.0055118948221206665, + 0.009180201217532158, + 0.007251302246004343, + 0.00986730307340622, + 0.006446589715778828, + 0.006832614075392485, + 0.005060934461653233, + 0.004083622712641954, + 0.005433631129562855, + 0.008649486117064953, + 0.0058448221534490585, + 0.007162701804190874, + 0.014566252939403057, + 0.00863855704665184, + 0.005620441399514675, + 0.007999753579497337, + 0.007083420641720295, + 0.009880966506898403, + 0.011765267699956894, + 0.009334402158856392, + 0.019301053136587143, + 0.026543714106082916, + 0.013912597671151161, + 0.009978625923395157, + 0.008498809300363064, + 0.007985015399754047, + 0.004930148366838694, + 0.007555539719760418, + 0.0065216487273573875, + 0.0161040797829628, + 0.004994289018213749, + 0.008181996643543243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.038105446845293045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054953958839178085 + ], + "opposite_logprob_tokens": [ + -2.9299540519714355 + ], + "logprob_reward": -0.054953958839178085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054953958839178085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17924317717552185, + 0.12360166013240814, + 0.060379598289728165, + 0.09255010634660721, + 0.056974224746227264, + 0.22664369642734528, + 0.1626385897397995, + 0.021304812282323837, + 0.05778047442436218, + 0.01888364925980568 + ], + "credit_map": [ + 0.019623341038823128, + 0.04670274257659912, + 0.01557542197406292, + 0.018747474998235703, + 0.0261320099234581, + 0.015789957717061043, + 0.023549912497401237, + 0.013122322969138622, + 0.006766553036868572, + 0.03765873610973358, + 0.007200027350336313, + 0.00609945971518755, + 0.046248093247413635, + 0.010856944136321545, + 0.005538515280932188, + 0.0032333366107195616, + 0.005028030835092068, + 0.012480375356972218, + 0.007276617921888828, + 0.00629852619022131, + 0.008153432048857212, + 0.00907955039292574, + 0.006642601918429136, + 0.005420468281954527, + 0.007785836700350046, + 0.010954662226140499, + 0.0092345355078578, + 0.009761535562574863, + 0.0217115618288517, + 0.012071535922586918, + 0.01257571391761303, + 0.008454740047454834, + 0.006106661166995764, + 0.007859299890697002, + 0.007346788886934519, + 0.007743777707219124, + 0.009149011224508286, + 0.008222407661378384, + 0.0060041071847081184, + 0.004542175214737654, + 0.016459552571177483, + 0.01756966859102249, + 0.0600859634578228, + 0.07006064802408218, + 0.01302898395806551, + 0.023060506209731102, + 0.011580969206988811, + 0.014797408133745193, + 0.016933560371398926, + 0.02865113504230976, + 0.020141592249274254, + 0.04282043129205704, + 0.01660783216357231, + 0.017450306564569473, + 0.009972465224564075, + 0.010061284527182579, + 0.0018683045636862516, + 0.0032322248443961143, + 0.002205661730840802, + 0.002747746417298913, + 0.004539709072560072, + 0.0030469191260635853, + 0.001646018004976213, + 0.0020182295702397823, + 0.004468344617635012, + 0.004433779511600733, + 0.0048485612496733665, + 0.005433864891529083, + 0.011951579712331295, + 0.012070748955011368, + 0.006284968461841345, + 0.008288625627756119, + 0.0023486404679715633, + 0.0023811941500753164, + 0.001733090146444738, + 0.0026327159721404314, + 0.0035765133798122406, + 0.002631253330036998, + 0.0015482304152101278, + 0.002032011980190873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.054953958839178085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two Asian men visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man in the foreground shirtless with short, sweat-soaked black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other man wearing a thoroughly soaked gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rough concrete basement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic, low-angle, handheld camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282134905457497 + ], + "opposite_logprob_tokens": [ + -3.773282051086426 + ], + "logprob_reward": -0.023282134905457497, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282134905457497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36718323826789856, + 0.09676162898540497, + 0.06830449402332306, + 0.058484021574258804, + 0.09620818495750427, + 0.040200375020504, + 0.048569388687610626, + 0.09048428386449814, + 0.08963607251644135, + 0.0441683791577816 + ], + "credit_map": [ + 0.052374549210071564, + 0.06817993521690369, + 0.043376024812459946, + 0.04801783338189125, + 0.047994863241910934, + 0.027139626443386078, + 0.05225389823317528, + 0.027846479788422585, + 0.007757460232824087, + 0.01853371411561966, + 0.013125209137797356, + 0.00873908493667841, + 0.0259828083217144, + 0.009820649400353432, + 0.006017467938363552, + 0.006785224657505751, + 0.006137791555374861, + 0.007172281388193369, + 0.009947473183274269, + 0.00988516304641962, + 0.010679052211344242, + 0.008763737976551056, + 0.012060522101819515, + 0.0036584637127816677, + 0.00516235176473856, + 0.008354236371815205, + 0.006075683515518904, + 0.008295920677483082, + 0.010245251469314098, + 0.006588649470359087, + 0.0072210985235869884, + 0.006540823727846146, + 0.011234190315008163, + 0.012814666144549847, + 0.012433112598955631, + 0.015084384940564632, + 0.014625299721956253, + 0.011395459994673729, + 0.010108906775712967, + 0.008512156084179878, + 0.004140740726143122, + 0.009762931615114212, + 0.0038934163749217987, + 0.0042771450243890285, + 0.005993370432406664, + 0.00506978202611208, + 0.004044047091156244, + 0.003018938936293125, + 0.005924162920564413, + 0.008858397603034973, + 0.00772845558822155, + 0.007249621674418449, + 0.006471286527812481, + 0.005541755352169275, + 0.0038700681179761887, + 0.002925633918493986, + 0.007982826791703701, + 0.012974747456610203, + 0.011182178743183613, + 0.01435001939535141, + 0.01580663211643696, + 0.01068237703293562, + 0.008939176797866821, + 0.008566327393054962, + 0.006867650896310806, + 0.011408480815589428, + 0.0076100691221654415, + 0.007997041568160057, + 0.01198706217110157, + 0.027237992733716965, + 0.007899129763245583, + 0.008628634735941887, + 0.005593480542302132, + 0.006379362661391497, + 0.0033164608757942915, + 0.0047658528201282024, + 0.0078151049092412, + 0.005501213483512402, + 0.003682946553453803, + 0.0071139526553452015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.023282134905457497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_tensor([8294, 8294])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_tensor([8294, 8294])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a894909a0d7f25495a8ae99609ce675c58a323f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_tensor([8294, 8294])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic camera, shot from the side, smoothly pans, following an athletic young woman with warm olive skin as she rides a mountain bike at a moderate pace. Her dark brown hair, held in a high, tight ponytail, bounces with her movements. A bright, genuinely happy expression illuminates her face, her eyes crinkling at the corners as she smiles widely. She wears a vibrant, puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow, and a sleek black tech backpack strapped to her back. Her legs, clad in dark athletic leggings, push the pedals with a steady, powerful rhythm, advancing along a narrow country road covered in a layer of crisp white snow that crunches softly beneath the knobby tires. On either side, sparse birch and pine forests are silhouetted against the twilight. The scene is bathed in the cool, ethereal light of dusk, the last rosy glow of sunset on the horizon casting long, soft blue shadows across the landscape, creating an atmosphere of quiet excitement and tranquil solitude.", + "seed": 4230, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_8294_seed_4230_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.035714775882661344, + "vlm_logprob": -0.035714775882661344, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.035714775882661344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02467072382569313 + ], + "opposite_logprob_tokens": [ + -4.149670600891113 + ], + "logprob_reward": -0.02467072382569313, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02467072382569313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12190551310777664, + 0.045301735401153564, + 0.10948155075311661, + 0.03499123081564903, + 0.03642898052930832, + 0.24116425216197968, + 0.09191783517599106, + 0.19769835472106934, + 0.04350084811449051, + 0.07760968059301376 + ], + "credit_map": [ + 0.018354089930653572, + 0.015117211267352104, + 0.015022478066384792, + 0.012586722150444984, + 0.014962497167289257, + 0.018610842525959015, + 0.013373215682804585, + 0.013878457248210907, + 0.006191672757267952, + 0.007172882556915283, + 0.006227627396583557, + 0.00846836343407631, + 0.005039009731262922, + 0.004690095316618681, + 0.0032178349792957306, + 0.004294248763471842, + 0.006530112121254206, + 0.011759014800190926, + 0.014352708123624325, + 0.023310163989663124, + 0.011925934813916683, + 0.009914580732584, + 0.009664096869528294, + 0.022024940699338913, + 0.0058804722502827644, + 0.004366285167634487, + 0.005207334645092487, + 0.003818365279585123, + 0.0034556144382804632, + 0.006710916757583618, + 0.003042293945327401, + 0.0025099492631852627, + 0.004157474264502525, + 0.007424863055348396, + 0.003924421966075897, + 0.005279887467622757, + 0.004207934252917767, + 0.004274378996342421, + 0.003681402187794447, + 0.003478618571534753, + 0.05984482541680336, + 0.03162169083952904, + 0.047112803906202316, + 0.01918816938996315, + 0.021390004083514214, + 0.017266469076275826, + 0.02275652065873146, + 0.02198377065360546, + 0.023785162717103958, + 0.014467700384557247, + 0.012683575041592121, + 0.007893495261669159, + 0.007664839271456003, + 0.010244169272482395, + 0.006567561533302069, + 0.008611329831182957, + 0.029214106500148773, + 0.04109803959727287, + 0.023212620988488197, + 0.02119891159236431, + 0.021847713738679886, + 0.022043511271476746, + 0.020026875659823418, + 0.019056575372815132, + 0.0035789478570222855, + 0.003480249783024192, + 0.005104776471853256, + 0.0032807181123644114, + 0.007850741036236286, + 0.005161378066986799, + 0.005738120526075363, + 0.009305915795266628, + 0.01231068279594183, + 0.00959088932722807, + 0.009984495118260384, + 0.006257496774196625, + 0.008228247985243797, + 0.018478116020560265, + 0.00666646845638752, + 0.0060932813212275505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.02467072382569313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204819686710835 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204819686710835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204819686710835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13457091152668, + 0.09293512254953384, + 0.21800389885902405, + 0.051748815923929214, + 0.06094839796423912, + 0.08255685865879059, + 0.051950789988040924, + 0.06251371651887894, + 0.07139622420072556, + 0.17337527871131897 + ], + "credit_map": [ + 0.017990289255976677, + 0.01581357792019844, + 0.015393310226500034, + 0.017111537978053093, + 0.012042971327900887, + 0.020053083077073097, + 0.016664935275912285, + 0.019501211121678352, + 0.012226558290421963, + 0.010900399647653103, + 0.011671111918985844, + 0.014099917374551296, + 0.012174753472208977, + 0.007021862547844648, + 0.016596589237451553, + 0.008243927732110023, + 0.03011225536465645, + 0.04419095069169998, + 0.030467011034488678, + 0.02269846573472023, + 0.015873579308390617, + 0.036404360085725784, + 0.017368314787745476, + 0.020888958126306534, + 0.007724966388195753, + 0.005454902071505785, + 0.0076120090670883656, + 0.00614370359107852, + 0.005318132229149342, + 0.010497277602553368, + 0.005157331470400095, + 0.0038404904771596193, + 0.00708001758903265, + 0.0072242496535182, + 0.00801845919340849, + 0.010067821480333805, + 0.006119382567703724, + 0.010287671349942684, + 0.0057928781025111675, + 0.006357917096465826, + 0.018737217411398888, + 0.00947632361203432, + 0.016654856503009796, + 0.007314404472708702, + 0.00904693454504013, + 0.0057932655327022076, + 0.009059285745024681, + 0.006474573630839586, + 0.006993133574724197, + 0.0050619374960660934, + 0.007382323034107685, + 0.008959454484283924, + 0.006602925714105368, + 0.006004238035529852, + 0.005533122457563877, + 0.005413659382611513, + 0.004336196463555098, + 0.0075480761006474495, + 0.007184640038758516, + 0.006066746078431606, + 0.013804391026496887, + 0.006417372729629278, + 0.009354755282402039, + 0.00780153414234519, + 0.008692181669175625, + 0.006374172400683165, + 0.008723224513232708, + 0.006027516908943653, + 0.010847166180610657, + 0.007491838652640581, + 0.006567974109202623, + 0.016672154888510704, + 0.038013190031051636, + 0.024454616010189056, + 0.02027272805571556, + 0.008709229528903961, + 0.01511621568351984, + 0.041536543518304825, + 0.013879911974072456, + 0.011392837390303612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014204819686710835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817653514444828 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.01817653514444828, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817653514444828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17837220430374146, + 0.03936981409788132, + 0.2888873517513275, + 0.17769664525985718, + 0.052420563995838165, + 0.025132177397608757, + 0.11300823092460632, + 0.025331277400255203, + 0.04332352057099342, + 0.056458231061697006 + ], + "credit_map": [ + 0.023771822452545166, + 0.02515454590320587, + 0.024608315899968147, + 0.0232293251901865, + 0.016956990584731102, + 0.02655036188662052, + 0.01809890940785408, + 0.020001940429210663, + 0.007657951209694147, + 0.005627193488180637, + 0.00478679733350873, + 0.0048100752755999565, + 0.0048313685692846775, + 0.004456080961972475, + 0.0031687715090811253, + 0.004031576216220856, + 0.026310080662369728, + 0.046294085681438446, + 0.034179579466581345, + 0.04290501028299332, + 0.02391102723777294, + 0.03253105655312538, + 0.02345559559762478, + 0.05930090695619583, + 0.029533060267567635, + 0.024412164464592934, + 0.021939294412732124, + 0.016591332852840424, + 0.024716205894947052, + 0.031162356957793236, + 0.0165739506483078, + 0.012768288142979145, + 0.006244079675525427, + 0.007098918780684471, + 0.006482583470642567, + 0.005648599937558174, + 0.006194018758833408, + 0.008618161082267761, + 0.005394441541284323, + 0.006739759352058172, + 0.0027088208589702845, + 0.002546699019148946, + 0.004622474778443575, + 0.0037516742013394833, + 0.0034312913194298744, + 0.001920717186294496, + 0.003816781798377633, + 0.002333717653527856, + 0.022974452003836632, + 0.02218087762594223, + 0.016852732747793198, + 0.009999051690101624, + 0.009546037763357162, + 0.007865986786782742, + 0.008135609328746796, + 0.015453481115400791, + 0.003368494799360633, + 0.0026978240348398685, + 0.003186758840456605, + 0.00465002004057169, + 0.0034650524612516165, + 0.002646033652126789, + 0.002646052511408925, + 0.002671042224392295, + 0.009887658059597015, + 0.005285408813506365, + 0.004887203220278025, + 0.0029916747007519007, + 0.004430072382092476, + 0.0061871325597167015, + 0.0037492774426937103, + 0.005905095487833023, + 0.013024899177253246, + 0.006713986862450838, + 0.0062272087670862675, + 0.0036157474387437105, + 0.00413834722712636, + 0.014741097576916218, + 0.003837260417640209, + 0.004159681964665651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01817653514444828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631303906440735 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631303906440735, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631303906440735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10844502598047256, + 0.04014823958277702, + 0.1514410823583603, + 0.07153014838695526, + 0.10491521656513214, + 0.05637156590819359, + 0.35410425066947937, + 0.014626027084887028, + 0.059904150664806366, + 0.038514263927936554 + ], + "credit_map": [ + 0.020426644012331963, + 0.019923921674489975, + 0.014024806208908558, + 0.011976123787462711, + 0.008820715360343456, + 0.016784029081463814, + 0.0074194930493831635, + 0.009069299325346947, + 0.004057284910231829, + 0.007230555173009634, + 0.004792456980794668, + 0.006038397550582886, + 0.00442455243319273, + 0.006724905222654343, + 0.0036294979508966208, + 0.0032505954150110483, + 0.015918230637907982, + 0.025826439261436462, + 0.019758455455303192, + 0.02385660447180271, + 0.01215139776468277, + 0.022232940420508385, + 0.012124531902372837, + 0.01957249455153942, + 0.011624968610703945, + 0.009102665819227695, + 0.00874932762235403, + 0.0069386884570121765, + 0.009436046704649925, + 0.013523672707378864, + 0.007004500832408667, + 0.005150288343429565, + 0.008600857108831406, + 0.018368523567914963, + 0.015555775724351406, + 0.011556516401469707, + 0.010653420351445675, + 0.019738921895623207, + 0.009849518537521362, + 0.010591695085167885, + 0.013038205914199352, + 0.006456197239458561, + 0.011267265304923058, + 0.004942762199789286, + 0.005965662654489279, + 0.004354729782789946, + 0.005630329251289368, + 0.004716422874480486, + 0.1037389412522316, + 0.07118310034275055, + 0.04992002621293068, + 0.022940633818507195, + 0.02726368047297001, + 0.02439863793551922, + 0.01855618692934513, + 0.03610307723283768, + 0.0017124621663242579, + 0.0022116529289633036, + 0.0016246425220742822, + 0.0023577248211950064, + 0.0016594025073572993, + 0.001514506177045405, + 0.0020318198949098587, + 0.0015138175804167986, + 0.01011364534497261, + 0.0061392346397042274, + 0.006793373264372349, + 0.004542494658380747, + 0.00851031206548214, + 0.008133560419082642, + 0.005907641723752022, + 0.009763897396624088, + 0.006687405053526163, + 0.006105850450694561, + 0.0037965395022183657, + 0.0027644566725939512, + 0.003750674659386277, + 0.008689871057868004, + 0.003493610769510269, + 0.003225857624784112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.008631303906440735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289049685001373 + ], + "opposite_logprob_tokens": [ + -2.2378904819488525 + ], + "logprob_reward": -0.11289049685001373, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289049685001373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13061802089214325, + 0.07603491842746735, + 0.23562371730804443, + 0.08675461262464523, + 0.04467891901731491, + 0.16524462401866913, + 0.10745715349912643, + 0.07326030731201172, + 0.044187866151332855, + 0.03613985329866409 + ], + "credit_map": [ + 0.027237145230174065, + 0.023995831608772278, + 0.018821295350790024, + 0.01168499793857336, + 0.01211671344935894, + 0.014143873006105423, + 0.011859026737511158, + 0.010759131982922554, + 0.013316397555172443, + 0.014651536010205746, + 0.011005663312971592, + 0.006054681725800037, + 0.009868507273495197, + 0.00859159603714943, + 0.006173418369144201, + 0.006373112555593252, + 0.028409991413354874, + 0.04990943893790245, + 0.031298693269491196, + 0.028968455269932747, + 0.016535228118300438, + 0.038904450833797455, + 0.019300552085042, + 0.022296903654932976, + 0.01730097271502018, + 0.013724281452596188, + 0.01238208170980215, + 0.007494168356060982, + 0.009721573442220688, + 0.011793865822255611, + 0.007811494171619415, + 0.006526165176182985, + 0.006134320516139269, + 0.004588729236274958, + 0.0066985213197767735, + 0.006308822892606258, + 0.004646067973226309, + 0.005376928951591253, + 0.005472721066325903, + 0.005452802870422602, + 0.04268501698970795, + 0.01992223784327507, + 0.03202394023537636, + 0.013546575792133808, + 0.017404276877641678, + 0.012588761746883392, + 0.01459643803536892, + 0.012477374635636806, + 0.028176138177514076, + 0.018753012642264366, + 0.01641949452459812, + 0.00689453911036253, + 0.009213645942509174, + 0.00770188495516777, + 0.006728528533130884, + 0.013569897972047329, + 0.011354979127645493, + 0.010064476169645786, + 0.008961870335042477, + 0.01179460622370243, + 0.008575115352869034, + 0.006825281772762537, + 0.009142730385065079, + 0.0065412395633757114, + 0.006759963929653168, + 0.0053818197920918465, + 0.004412804264575243, + 0.0032080733217298985, + 0.0076217749156057835, + 0.007102948147803545, + 0.004718175157904625, + 0.00498230243101716, + 0.006040188018232584, + 0.006095163989812136, + 0.0046651908196508884, + 0.0032694991677999496, + 0.0035606990568339825, + 0.006875135004520416, + 0.002825848525390029, + 0.002808129880577326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.11289049685001373 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic camera, shot from the side, smoothly pans, following an athletic young woman with warm olive skin as she rides a mountain bike at a moderate pace. Her dark brown hair, held in a high, tight ponytail, bounces with her movements. A bright, genuinely happy expression illuminates her face, her eyes crinkling at the corners as she smiles widely. She wears a vibrant, puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow, and a sleek black tech backpack strapped to her back. Her legs, clad in dark athletic leggings, push the pedals with a steady, powerful rhythm, advancing along a narrow country road covered in a layer of crisp white snow that crunches softly beneath the knobby tires. On either side, sparse birch and pine forests are silhouetted against the twilight. The scene is bathed in the cool, ethereal light of dusk, the last rosy glow of sunset on the horizon casting long, soft blue shadows across the landscape, creating an atmosphere of quiet excitement and tranquil solitude.", + "seed": 4231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_8294_seed_4231_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.047372478805482385, + "vlm_logprob": -0.047372478805482385, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.047372478805482385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041346073150634766 + ], + "opposite_logprob_tokens": [ + -3.7913460731506348 + ], + "logprob_reward": -0.041346073150634766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.041346073150634766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2987225353717804, + 0.089499831199646, + 0.06055162847042084, + 0.14740999042987823, + 0.11800937354564667, + 0.048891641199588776, + 0.05264575406908989, + 0.10107121616601944, + 0.052670370787382126, + 0.0305276308208704 + ], + "credit_map": [ + 0.039159126579761505, + 0.027751749381422997, + 0.03464893996715546, + 0.06515567749738693, + 0.02752481959760189, + 0.030087614431977272, + 0.02714322879910469, + 0.047251373529434204, + 0.01215106900781393, + 0.011849160306155682, + 0.014668491668999195, + 0.010747100226581097, + 0.008961784653365612, + 0.011474681086838245, + 0.008511502295732498, + 0.011136047542095184, + 0.007290976122021675, + 0.009430311620235443, + 0.006186190526932478, + 0.009895478375256062, + 0.007015797309577465, + 0.008679134771227837, + 0.00541981216520071, + 0.006633932236582041, + 0.01778319664299488, + 0.019351093098521233, + 0.0175065528601408, + 0.03196948766708374, + 0.015617473050951958, + 0.017477713525295258, + 0.014598671346902847, + 0.01310581061989069, + 0.010882283560931683, + 0.011208302341401577, + 0.011973315849900246, + 0.018425460904836655, + 0.014999927021563053, + 0.01565307192504406, + 0.015822414308786392, + 0.01904461346566677, + 0.004713485948741436, + 0.004383796826004982, + 0.004728799685835838, + 0.008168878965079784, + 0.006330584641546011, + 0.010158754885196686, + 0.004576904233545065, + 0.005830440670251846, + 0.004733026493340731, + 0.00735680153593421, + 0.008413579314947128, + 0.0063784243538975716, + 0.004723520018160343, + 0.010724274441599846, + 0.004669651389122009, + 0.0056464821100234985, + 0.010532092303037643, + 0.005363082978874445, + 0.005846340209245682, + 0.005463608540594578, + 0.02298641763627529, + 0.017698494717478752, + 0.0072953663766384125, + 0.025885820388793945, + 0.00703364796936512, + 0.004895362071692944, + 0.004898976068943739, + 0.0045987386256456375, + 0.008007451891899109, + 0.008058059960603714, + 0.007769277784973383, + 0.007408856879919767, + 0.0037112541031092405, + 0.002666940214112401, + 0.004025573376566172, + 0.0033405651338398457, + 0.004151105415076017, + 0.006212794221937656, + 0.003049104940146208, + 0.0033702957443892956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.041346073150634766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088889092206955 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088889092206955, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088889092206955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2188529372215271, + 0.13399477303028107, + 0.04334526136517525, + 0.2450762242078781, + 0.0792553722858429, + 0.03478403016924858, + 0.04100336134433746, + 0.13955701887607574, + 0.04011867940425873, + 0.024012405425310135 + ], + "credit_map": [ + 0.02709359861910343, + 0.02275758422911167, + 0.03405825421214104, + 0.0292346328496933, + 0.02058545872569084, + 0.021796880289912224, + 0.01972701959311962, + 0.04359949752688408, + 0.016716884449124336, + 0.016972433775663376, + 0.020705899223685265, + 0.017555376514792442, + 0.012499503791332245, + 0.01708149164915085, + 0.014725523069500923, + 0.017737654969096184, + 0.004992971662431955, + 0.0065680574625730515, + 0.0054550315253436565, + 0.005782444961369038, + 0.005588611122220755, + 0.005249397829174995, + 0.0041784862987697124, + 0.005530260503292084, + 0.02995433658361435, + 0.020724747329950333, + 0.029121635481715202, + 0.06606345623731613, + 0.027143126353621483, + 0.024652693420648575, + 0.025528009980916977, + 0.021888216957449913, + 0.006323342211544514, + 0.009494305588304996, + 0.008302636444568634, + 0.014970858581364155, + 0.007589328568428755, + 0.013888504356145859, + 0.007885593920946121, + 0.010800803080201149, + 0.0038254025857895613, + 0.002432104665786028, + 0.004211323335766792, + 0.0037439120933413506, + 0.004816921427845955, + 0.008106936700642109, + 0.0031090339180082083, + 0.004538394045084715, + 0.004655803088098764, + 0.00392836844548583, + 0.006079590879380703, + 0.005463697016239166, + 0.0038348741363734007, + 0.00817635003477335, + 0.003446944523602724, + 0.005417732987552881, + 0.01727934181690216, + 0.005694505758583546, + 0.006820721086114645, + 0.005760306492447853, + 0.04007755219936371, + 0.022515365853905678, + 0.011559599079191685, + 0.02984963171184063, + 0.004136431030929089, + 0.0028637759387493134, + 0.002962015103548765, + 0.0038208349142223597, + 0.009792913682758808, + 0.008150388486683369, + 0.003855908289551735, + 0.004536410328000784, + 0.00302746519446373, + 0.0017158229602500796, + 0.002119750250130892, + 0.0023788711987435818, + 0.003758033039048314, + 0.005657596979290247, + 0.002077500568702817, + 0.0032773655839264393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.011088889092206955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817653514444828 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.01817653514444828, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817653514444828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27631258964538574, + 0.11087055504322052, + 0.05435900390148163, + 0.10945695638656616, + 0.09406516700983047, + 0.07005543261766434, + 0.06178461015224457, + 0.14627376198768616, + 0.04689391702413559, + 0.02992805279791355 + ], + "credit_map": [ + 0.032893385738134384, + 0.029691509902477264, + 0.042943019419908524, + 0.03186836093664169, + 0.026958370581269264, + 0.029370028525590897, + 0.026461001485586166, + 0.0561269074678421, + 0.015299701131880283, + 0.011035163886845112, + 0.017583783715963364, + 0.014273754321038723, + 0.00922129862010479, + 0.013379662297666073, + 0.014221585355699062, + 0.01585559733211994, + 0.005811662878841162, + 0.008300552144646645, + 0.006946302019059658, + 0.006744902580976486, + 0.004633914679288864, + 0.006625797133892775, + 0.005856778938323259, + 0.009439094923436642, + 0.012074095197021961, + 0.011025425978004932, + 0.013526780530810356, + 0.02180563472211361, + 0.009489729069173336, + 0.01862075738608837, + 0.010628591291606426, + 0.012285944074392319, + 0.00769747793674469, + 0.010493491776287556, + 0.010450622998178005, + 0.017186647281050682, + 0.009303943254053593, + 0.010369046591222286, + 0.011912763118743896, + 0.01665116846561432, + 0.007414398714900017, + 0.0044411104172468185, + 0.007303748279809952, + 0.006385397631675005, + 0.010237173177301884, + 0.020423680543899536, + 0.0057370662689208984, + 0.008112858980894089, + 0.00804170873016119, + 0.0057013328187167645, + 0.0069176144897937775, + 0.005202160682529211, + 0.0065769110806286335, + 0.01448621042072773, + 0.005210821982473135, + 0.009647847153246403, + 0.018930912017822266, + 0.006058068014681339, + 0.007024133112281561, + 0.006522233597934246, + 0.043410398066043854, + 0.022214237600564957, + 0.012830005027353764, + 0.029283763840794563, + 0.005132908467203379, + 0.003549321787431836, + 0.004069822374731302, + 0.00450305687263608, + 0.01123901829123497, + 0.00990231242030859, + 0.003708858508616686, + 0.004788617137819529, + 0.004044202156364918, + 0.0029028262943029404, + 0.002389291301369667, + 0.0026679441798478365, + 0.0034421891905367374, + 0.006804976612329483, + 0.003385461401194334, + 0.004291162826120853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01817653514444828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942297168076038 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.005942297168076038, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942297168076038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1751956194639206, + 0.051983706653118134, + 0.04258672893047333, + 0.058791324496269226, + 0.06104562059044838, + 0.061442844569683075, + 0.06270846724510193, + 0.35855892300605774, + 0.05123208090662956, + 0.07645475119352341 + ], + "credit_map": [ + 0.02263524755835533, + 0.017940444871783257, + 0.022064710035920143, + 0.021794119849801064, + 0.01684323512017727, + 0.02362111583352089, + 0.017055926844477654, + 0.033240798860788345, + 0.006918478757143021, + 0.007900468073785305, + 0.005749113857746124, + 0.004920897074043751, + 0.0043413941748440266, + 0.00994171854108572, + 0.0050908164121210575, + 0.00712081603705883, + 0.005317787639796734, + 0.004084992688149214, + 0.00800725445151329, + 0.004444698803126812, + 0.0034248961601406336, + 0.004457863513380289, + 0.0043520559556782246, + 0.008497174829244614, + 0.007111610844731331, + 0.005917471367865801, + 0.006116306874901056, + 0.014430943876504898, + 0.006168206222355366, + 0.007282295264303684, + 0.006109659560024738, + 0.005654829554259777, + 0.006707640830427408, + 0.005247416440397501, + 0.007274768780916929, + 0.006726645864546299, + 0.005815959069877863, + 0.013146805576980114, + 0.009435181505978107, + 0.006691201590001583, + 0.006976975128054619, + 0.003426474519073963, + 0.006088478956371546, + 0.005641936790198088, + 0.00689915893599391, + 0.019442180171608925, + 0.00521096820011735, + 0.007756670005619526, + 0.006778988521546125, + 0.005588355474174023, + 0.012093586847186089, + 0.007467116229236126, + 0.00630570575594902, + 0.010740820318460464, + 0.006030872464179993, + 0.007703020703047514, + 0.046254877001047134, + 0.012706051580607891, + 0.014486626721918583, + 0.0117557467892766, + 0.11249466240406036, + 0.043778661638498306, + 0.024033844470977783, + 0.09304842352867126, + 0.00872768647968769, + 0.0027751694433391094, + 0.002528992248699069, + 0.0028356302063912153, + 0.018151026219129562, + 0.007735655177384615, + 0.0030975323170423508, + 0.005380379501730204, + 0.011785665526986122, + 0.004069387912750244, + 0.006243383977562189, + 0.0057510631158947945, + 0.010073593817651272, + 0.019367128610610962, + 0.005239258985966444, + 0.013925263658165932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.005942297168076038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603085994720459 + ], + "opposite_logprob_tokens": [ + -1.910308599472046 + ], + "logprob_reward": -0.1603085994720459, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603085994720459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08525816351175308, + 0.09332363307476044, + 0.07499717175960541, + 0.047160252928733826, + 0.12954334914684296, + 0.3613540232181549, + 0.029057085514068604, + 0.11342769116163254, + 0.0362783819437027, + 0.029600266367197037 + ], + "credit_map": [ + 0.01071382686495781, + 0.009894766844809055, + 0.01814182847738266, + 0.012026150710880756, + 0.006726645398885012, + 0.0071173724718391895, + 0.008911751210689545, + 0.01172580849379301, + 0.011362926103174686, + 0.014401236549019814, + 0.01660558208823204, + 0.009541337378323078, + 0.006541658192873001, + 0.01345029752701521, + 0.007413052022457123, + 0.014007538557052612, + 0.007663278374820948, + 0.010723775252699852, + 0.009840350598096848, + 0.010586479678750038, + 0.009424037300050259, + 0.010030614212155342, + 0.006616347935050726, + 0.010112286545336246, + 0.005172125995159149, + 0.0034280905965715647, + 0.006922711618244648, + 0.008672506548464298, + 0.0041143521666526794, + 0.009271854534745216, + 0.004552171565592289, + 0.005026435013860464, + 0.01387484185397625, + 0.01041030790656805, + 0.01385683473199606, + 0.009563270024955273, + 0.015233462676405907, + 0.026132619008421898, + 0.023775547742843628, + 0.016696451231837273, + 0.04330843314528465, + 0.015131180174648762, + 0.033157479017972946, + 0.02308669127523899, + 0.05274692922830582, + 0.125313863158226, + 0.023651935160160065, + 0.044957485049963, + 0.0030005639418959618, + 0.003985286224633455, + 0.003946248441934586, + 0.004223515745252371, + 0.002643084153532982, + 0.00564975943416357, + 0.002402547048404813, + 0.0032060800585895777, + 0.01187692116945982, + 0.0075215306133031845, + 0.010470838285982609, + 0.007912496104836464, + 0.017686227336525917, + 0.03094346448779106, + 0.009221507236361504, + 0.01779470406472683, + 0.002768538659438491, + 0.004032639786601067, + 0.0032904758118093014, + 0.0038958871737122536, + 0.0042966981418430805, + 0.008356865495443344, + 0.0055769626051187515, + 0.004060311708599329, + 0.0043251775205135345, + 0.002842459362000227, + 0.0026164166629314423, + 0.0024305263068526983, + 0.004736290313303471, + 0.0064415475353598595, + 0.0018398978281766176, + 0.004367948044091463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.1603085994720459 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic camera, shot from the side, smoothly pans, following an athletic young woman with warm olive skin as she rides a mountain bike at a moderate pace. Her dark brown hair, held in a high, tight ponytail, bounces with her movements. A bright, genuinely happy expression illuminates her face, her eyes crinkling at the corners as she smiles widely. She wears a vibrant, puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow, and a sleek black tech backpack strapped to her back. Her legs, clad in dark athletic leggings, push the pedals with a steady, powerful rhythm, advancing along a narrow country road covered in a layer of crisp white snow that crunches softly beneath the knobby tires. On either side, sparse birch and pine forests are silhouetted against the twilight. The scene is bathed in the cool, ethereal light of dusk, the last rosy glow of sunset on the horizon casting long, soft blue shadows across the landscape, creating an atmosphere of quiet excitement and tranquil solitude.", + "seed": 4232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_8294_seed_4232_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0314563306979835, + "vlm_logprob": -0.0314563306979835, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.0314563306979835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058261655271053314 + ], + "opposite_logprob_tokens": [ + -3.0582616329193115 + ], + "logprob_reward": -0.058261655271053314, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.058261655271053314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23829922080039978, + 0.17100051045417786, + 0.07188300788402557, + 0.22097405791282654, + 0.07781234383583069, + 0.03184376284480095, + 0.027263587340712547, + 0.06119365617632866, + 0.05292759835720062, + 0.04680224880576134 + ], + "credit_map": [ + 0.03572177141904831, + 0.032066456973552704, + 0.025338714942336082, + 0.03010832890868187, + 0.03236384317278862, + 0.031943049281835556, + 0.031680721789598465, + 0.019076326861977577, + 0.019836924970149994, + 0.036183349788188934, + 0.017580265179276466, + 0.014618335291743279, + 0.023819847032427788, + 0.021268552169203758, + 0.015809161588549614, + 0.021884074434638023, + 0.01078520342707634, + 0.006088147405534983, + 0.014860374853014946, + 0.00861695222556591, + 0.008343514055013657, + 0.007656578440219164, + 0.007164146285504103, + 0.00836808979511261, + 0.03266923502087593, + 0.023346923291683197, + 0.03532251715660095, + 0.030070630833506584, + 0.022115768864750862, + 0.033090800046920776, + 0.018272798508405685, + 0.026085399091243744, + 0.009183000773191452, + 0.009926386177539825, + 0.009140684269368649, + 0.012690464965999126, + 0.010193769820034504, + 0.0071746124885976315, + 0.006951024755835533, + 0.012552410364151001, + 0.004198051523417234, + 0.004449904430657625, + 0.004548223223537207, + 0.0037211633753031492, + 0.003147059353068471, + 0.005412945058196783, + 0.0034986771643161774, + 0.0028677398804575205, + 0.004447589628398418, + 0.004404623992741108, + 0.0026742778718471527, + 0.003387242555618286, + 0.0030884207226336002, + 0.0033302227966487408, + 0.0025618355721235275, + 0.0033693774603307247, + 0.008018405176699162, + 0.012571447528898716, + 0.007886012084782124, + 0.007336446549743414, + 0.005967534147202969, + 0.005118603352457285, + 0.008422000333666801, + 0.005873209331184626, + 0.012556458823382854, + 0.0076545546762645245, + 0.0035539367236196995, + 0.004753249231725931, + 0.006587441079318523, + 0.006228266283869743, + 0.006511211860924959, + 0.005082485266029835, + 0.007879409939050674, + 0.007992815226316452, + 0.004988652188330889, + 0.004586649592965841, + 0.005197544116526842, + 0.007271968759596348, + 0.004502041731029749, + 0.004383173771202564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.058261655271053314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006757147144526243 + ], + "opposite_logprob_tokens": [ + -5.0067572593688965 + ], + "logprob_reward": -0.006757147144526243, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006757147144526243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10573187470436096, + 0.09375520050525665, + 0.06779226660728455, + 0.12376225739717484, + 0.08373136073350906, + 0.03139469027519226, + 0.1303151249885559, + 0.0632638931274414, + 0.26624053716659546, + 0.034012746065855026 + ], + "credit_map": [ + 0.01907811127603054, + 0.0188420619815588, + 0.011795993894338608, + 0.006105514243245125, + 0.015388979576528072, + 0.018384460359811783, + 0.007731746416538954, + 0.008405020460486412, + 0.01174253411591053, + 0.014473033137619495, + 0.0119926817715168, + 0.00960290152579546, + 0.011832409538328648, + 0.010629374533891678, + 0.01400681957602501, + 0.009475449100136757, + 0.00827784277498722, + 0.0071022543124854565, + 0.007102645467966795, + 0.012543252669274807, + 0.008674482814967632, + 0.008556515909731388, + 0.009568442590534687, + 0.00596683518961072, + 0.01738879643380642, + 0.016138968989253044, + 0.013473744504153728, + 0.02094605751335621, + 0.016213539987802505, + 0.016141993924975395, + 0.008867782540619373, + 0.014591378159821033, + 0.012521003372967243, + 0.010737912729382515, + 0.010963229462504387, + 0.008165406994521618, + 0.009395970031619072, + 0.007530630100518465, + 0.007004635874181986, + 0.017412574961781502, + 0.004125264007598162, + 0.003628053003922105, + 0.0058948807418346405, + 0.00347564066760242, + 0.002904454944655299, + 0.005774366669356823, + 0.0026892514433711767, + 0.002902781357988715, + 0.021558398380875587, + 0.015506822615861893, + 0.012996630743145943, + 0.019947875291109085, + 0.015692638233304024, + 0.0102783078327775, + 0.014840420335531235, + 0.019494028761982918, + 0.007316968869417906, + 0.011351530440151691, + 0.006837948691099882, + 0.00860845297574997, + 0.006195374298840761, + 0.004581714980304241, + 0.010972539894282818, + 0.007399365771561861, + 0.031921662390232086, + 0.032025113701820374, + 0.031114939600229263, + 0.03988955542445183, + 0.0359138660132885, + 0.027073539793491364, + 0.044741395860910416, + 0.02356046624481678, + 0.004355101380497217, + 0.005406941752880812, + 0.0034285641741007566, + 0.0034224160481244326, + 0.00353797385469079, + 0.007846551947295666, + 0.0027217273600399494, + 0.003293468849733472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.006757147144526243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606121100485325 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606121100485325, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606121100485325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20302043855190277, + 0.05408628657460213, + 0.10441281646490097, + 0.3584194481372833, + 0.046990297734737396, + 0.05104564502835274, + 0.06951935589313507, + 0.034952424466609955, + 0.03619510307908058, + 0.041358184069395065 + ], + "credit_map": [ + 0.03837355971336365, + 0.0488872267305851, + 0.018803825601935387, + 0.014779815450310707, + 0.018934277817606926, + 0.02857426553964615, + 0.013739616610109806, + 0.020927852019667625, + 0.005461624823510647, + 0.007479088380932808, + 0.007879761047661304, + 0.005116947460919619, + 0.0065776449628174305, + 0.008563061244785786, + 0.005106459837406874, + 0.007901696488261223, + 0.018427681177854538, + 0.014708313159644604, + 0.016510864719748497, + 0.010050260461866856, + 0.009423435665667057, + 0.012657207436859608, + 0.009657582268118858, + 0.012977469712495804, + 0.0639476329088211, + 0.036012113094329834, + 0.05674136430025101, + 0.03377591073513031, + 0.03204840421676636, + 0.05028831958770752, + 0.027262847870588303, + 0.05834285169839859, + 0.0044428762048482895, + 0.005146360024809837, + 0.005942582618445158, + 0.007229386828839779, + 0.006487039849162102, + 0.005123840179294348, + 0.004596440587192774, + 0.0080217644572258, + 0.006563970819115639, + 0.005418199580162764, + 0.007622373290359974, + 0.00473393127322197, + 0.0065092057920992374, + 0.010727908462285995, + 0.005070768296718597, + 0.004399280529469252, + 0.012469131499528885, + 0.008671066723763943, + 0.006808083038777113, + 0.007330720312893391, + 0.008509094826877117, + 0.007788084913045168, + 0.006422199308872223, + 0.011520974338054657, + 0.004498339258134365, + 0.0048222956247627735, + 0.005056952591985464, + 0.004253091290593147, + 0.004226048942655325, + 0.004017232917249203, + 0.0036609487142413855, + 0.004417513031512499, + 0.004826611839234829, + 0.005647179204970598, + 0.0028491804841905832, + 0.004151096101850271, + 0.004253024235367775, + 0.005803002044558525, + 0.00528176361694932, + 0.0033832453191280365, + 0.010276437737047672, + 0.007824443280696869, + 0.005009772256016731, + 0.0029443101957440376, + 0.0031787354964762926, + 0.004762315656989813, + 0.00338371773250401, + 0.003978447988629341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01606121100485325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187190681695938 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.014187190681695938, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187190681695938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2654031813144684, + 0.2163182497024536, + 0.11800651997327805, + 0.08463464677333832, + 0.033168986439704895, + 0.08430665731430054, + 0.05462080240249634, + 0.057150278240442276, + 0.036907464265823364, + 0.04948326572775841 + ], + "credit_map": [ + 0.051674071699380875, + 0.0708245262503624, + 0.02551661990582943, + 0.019206538796424866, + 0.020366262644529343, + 0.0305821280926466, + 0.020541885867714882, + 0.026691127568483353, + 0.022704537957906723, + 0.038400985300540924, + 0.020105522125959396, + 0.020073728635907173, + 0.03260330483317375, + 0.02991245687007904, + 0.021454500034451485, + 0.031063184142112732, + 0.023562656715512276, + 0.013902360573410988, + 0.022651150822639465, + 0.009936289861798286, + 0.012395980767905712, + 0.011986390687525272, + 0.00750005105510354, + 0.016071636229753494, + 0.011889017187058926, + 0.008157971315085888, + 0.017801782116293907, + 0.008351598866283894, + 0.006321506109088659, + 0.010885986499488354, + 0.004993482027202845, + 0.01623329147696495, + 0.003593479050323367, + 0.0037946628872305155, + 0.0037530208937823772, + 0.0033465190790593624, + 0.004392205737531185, + 0.0060736555606126785, + 0.002945740008726716, + 0.005269702523946762, + 0.01047911960631609, + 0.008938498795032501, + 0.014323228038847446, + 0.008252620697021484, + 0.008348817005753517, + 0.019730962812900543, + 0.007461580913513899, + 0.006771820597350597, + 0.010169187560677528, + 0.007098506670445204, + 0.005874093621969223, + 0.005904066376388073, + 0.005456141661852598, + 0.005679807160049677, + 0.004839436151087284, + 0.009599553421139717, + 0.006684901658445597, + 0.010349397547543049, + 0.005979141686111689, + 0.008569064550101757, + 0.007035763468593359, + 0.0064241401851177216, + 0.006036690436303616, + 0.006071173585951328, + 0.00493925204500556, + 0.005196835845708847, + 0.003916972782462835, + 0.003507641376927495, + 0.004878444597125053, + 0.006269833073019981, + 0.0047641233541071415, + 0.0034343579318374395, + 0.010862726718187332, + 0.008021309040486813, + 0.0044646249152719975, + 0.006113789509981871, + 0.004248067270964384, + 0.0054167755879461765, + 0.004428807180374861, + 0.005927159450948238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014187190681695938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201444938778877 + ], + "opposite_logprob_tokens": [ + -2.81201434135437 + ], + "logprob_reward": -0.06201444938778877, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201444938778877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1350788027048111, + 0.0658169835805893, + 0.051917482167482376, + 0.1011158749461174, + 0.05875544250011444, + 0.05365808308124542, + 0.2087997943162918, + 0.16223302483558655, + 0.06829369068145752, + 0.09433083236217499 + ], + "credit_map": [ + 0.022707590833306313, + 0.026995765045285225, + 0.01344303973019123, + 0.009564454667270184, + 0.01906217448413372, + 0.0222125593572855, + 0.011516737751662731, + 0.009576481766998768, + 0.006952539552003145, + 0.007169830147176981, + 0.009413044899702072, + 0.0066431788727641106, + 0.0076845064759254456, + 0.010020902380347252, + 0.007339706178754568, + 0.010593284852802753, + 0.008700553327798843, + 0.0058653573505580425, + 0.008123436942696571, + 0.005639682523906231, + 0.005955344997346401, + 0.005770314019173384, + 0.006827490869909525, + 0.005035308189690113, + 0.015435202978551388, + 0.010650614276528358, + 0.015619228594005108, + 0.01383740734308958, + 0.010080451145768166, + 0.013627010397613049, + 0.008408290334045887, + 0.013457673601806164, + 0.009624406695365906, + 0.007072740234434605, + 0.006946047302335501, + 0.005798303987830877, + 0.007050755899399519, + 0.007379251066595316, + 0.005135444924235344, + 0.009748494252562523, + 0.0072216251865029335, + 0.005021162796765566, + 0.010596765205264091, + 0.0038251178339123726, + 0.005886760540306568, + 0.01008374709635973, + 0.004268096759915352, + 0.0067548141814768314, + 0.029539484530687332, + 0.02368219383060932, + 0.03535136952996254, + 0.014484680257737637, + 0.030194317921996117, + 0.03659551963210106, + 0.015867015346884727, + 0.023085227236151695, + 0.02346472255885601, + 0.023979151621460915, + 0.017747869715094566, + 0.01503695547580719, + 0.017680594697594643, + 0.019814370200037956, + 0.020528435707092285, + 0.02398092858493328, + 0.010809321887791157, + 0.010119312442839146, + 0.004015881568193436, + 0.006625673267990351, + 0.008654157631099224, + 0.013004951179027557, + 0.00814021471887827, + 0.0069241831079125404, + 0.011482038535177708, + 0.01010852586477995, + 0.00847740937024355, + 0.009398059919476509, + 0.00908581167459488, + 0.022546028718352318, + 0.01190172042697668, + 0.011331241577863693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.06201444938778877 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic camera, shot from the side, smoothly pans, following an athletic young woman with warm olive skin as she rides a mountain bike at a moderate pace. Her dark brown hair, held in a high, tight ponytail, bounces with her movements. A bright, genuinely happy expression illuminates her face, her eyes crinkling at the corners as she smiles widely. She wears a vibrant, puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow, and a sleek black tech backpack strapped to her back. Her legs, clad in dark athletic leggings, push the pedals with a steady, powerful rhythm, advancing along a narrow country road covered in a layer of crisp white snow that crunches softly beneath the knobby tires. On either side, sparse birch and pine forests are silhouetted against the twilight. The scene is bathed in the cool, ethereal light of dusk, the last rosy glow of sunset on the horizon casting long, soft blue shadows across the landscape, creating an atmosphere of quiet excitement and tranquil solitude.", + "seed": 4233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_13/index_8294_seed_4233_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020299932127818464, + "vlm_logprob": -0.020299932127818464, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.020299932127818464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023681648075580597 + ], + "opposite_logprob_tokens": [ + -4.148681640625 + ], + "logprob_reward": -0.023681648075580597, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023681648075580597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43271568417549133, + 0.10813518613576889, + 0.012316898442804813, + 0.019308248534798622, + 0.0222234595566988, + 0.29250699281692505, + 0.044846270233392715, + 0.021700521931052208, + 0.028534376993775368, + 0.017712261527776718 + ], + "credit_map": [ + 0.038402631878852844, + 0.1337248533964157, + 0.13646061718463898, + 0.07331299036741257, + 0.018524115905165672, + 0.016386525705456734, + 0.006257866509258747, + 0.00964608509093523, + 0.009366081096231937, + 0.040865812450647354, + 0.03176146000623703, + 0.00920263584703207, + 0.005913975182920694, + 0.005214986857026815, + 0.0021074244286864996, + 0.0037028107326477766, + 0.0020146453753113747, + 0.0014873957261443138, + 0.0011995886452496052, + 0.0019048245158046484, + 0.0013033642899245024, + 0.0016704252921044827, + 0.0011889082379639149, + 0.0015477464767172933, + 0.002536602085456252, + 0.003024548292160034, + 0.002325560664758086, + 0.0033139658626168966, + 0.0022542239166796207, + 0.0017878288635984063, + 0.0016419489402323961, + 0.0024235700257122517, + 0.0038720264565199614, + 0.0035001107025891542, + 0.003129609627649188, + 0.003629734506830573, + 0.0023997812531888485, + 0.001738979946821928, + 0.0016789984656497836, + 0.0022742196451872587, + 0.08278502523899078, + 0.024018431082367897, + 0.1181599572300911, + 0.030252447351813316, + 0.010271621868014336, + 0.005522372201085091, + 0.012118162587285042, + 0.00937897339463234, + 0.0035136030055582523, + 0.0033648095559328794, + 0.01307697407901287, + 0.003917287103831768, + 0.0028812058735638857, + 0.0022673921193927526, + 0.003036105539649725, + 0.01278889924287796, + 0.0034716532099992037, + 0.0034254349302500486, + 0.004052599426358938, + 0.0023825836833566427, + 0.0020621854346245527, + 0.001892133615911007, + 0.002397919539362192, + 0.0020160123240202665, + 0.0030166178476065397, + 0.004746862221509218, + 0.006546312943100929, + 0.003598863026127219, + 0.0034475892316550016, + 0.0023274386767297983, + 0.0027814351487904787, + 0.0020692585967481136, + 0.0019971260335296392, + 0.0032230939250439405, + 0.0028031948022544384, + 0.00294039910659194, + 0.0021165315993130207, + 0.0016876726876944304, + 0.001722575630992651, + 0.0012216675095260143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.023681648075580597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553717941045761 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553717941045761, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553717941045761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2611117959022522, + 0.13392557203769684, + 0.03436103090643883, + 0.042602695524692535, + 0.0709063932299614, + 0.10174065828323364, + 0.053936105221509933, + 0.13814027607440948, + 0.12847498059272766, + 0.03480048477649689 + ], + "credit_map": [ + 0.022589746862649918, + 0.0758460983633995, + 0.07838741689920425, + 0.042119141668081284, + 0.011808297596871853, + 0.017437487840652466, + 0.004892196040600538, + 0.008031404577195644, + 0.011272668838500977, + 0.052484527230262756, + 0.031407177448272705, + 0.011610317043960094, + 0.008621676824986935, + 0.00991460308432579, + 0.0034280952531844378, + 0.005186519119888544, + 0.006581756751984358, + 0.005903005134314299, + 0.003542728489264846, + 0.0041702426970005035, + 0.005694101098924875, + 0.0037315543740987778, + 0.002284712390974164, + 0.0024529267102479935, + 0.0077073541469872, + 0.005038716830313206, + 0.006121993064880371, + 0.004256076645106077, + 0.003496995661407709, + 0.003623038297519088, + 0.003672507358714938, + 0.008686012588441372, + 0.014251099899411201, + 0.011983777396380901, + 0.009723409079015255, + 0.010762300342321396, + 0.006715327501296997, + 0.0056589278392493725, + 0.005275214556604624, + 0.006536338943988085, + 0.02526136487722397, + 0.011998762376606464, + 0.032683100551366806, + 0.009281794540584087, + 0.007589539512991905, + 0.006489090155810118, + 0.003916176036000252, + 0.004520830232650042, + 0.00412343256175518, + 0.004097746219485998, + 0.013392390683293343, + 0.004224284552037716, + 0.004364582244306803, + 0.003912306856364012, + 0.0037522336933761835, + 0.01606912910938263, + 0.018732085824012756, + 0.014933326281607151, + 0.024556457996368408, + 0.01768857054412365, + 0.017658285796642303, + 0.01810717210173607, + 0.013447657227516174, + 0.013016722165048122, + 0.010546456091105938, + 0.017316477373242378, + 0.022139178588986397, + 0.014950616285204887, + 0.012874843552708626, + 0.011228449642658234, + 0.021712331101298332, + 0.017706628888845444, + 0.0033234967850148678, + 0.003574528032913804, + 0.004893030039966106, + 0.006467550061643124, + 0.0031893178820610046, + 0.003282608464360237, + 0.004267352633178234, + 0.005802598781883717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012553717941045761 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732400041073561 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732400041073561, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732400041073561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24897699058055878, + 0.1633061170578003, + 0.03202313184738159, + 0.05068216100335121, + 0.048262856900691986, + 0.0921190157532692, + 0.2838572859764099, + 0.027684461325407028, + 0.03544894605875015, + 0.017639094963669777 + ], + "credit_map": [ + 0.02194046601653099, + 0.06894954293966293, + 0.07132977247238159, + 0.037053003907203674, + 0.014448787085711956, + 0.019286464899778366, + 0.006488110404461622, + 0.009480815380811691, + 0.013992855325341225, + 0.06784386187791824, + 0.02295333705842495, + 0.031135322526097298, + 0.009982848539948463, + 0.009312868118286133, + 0.0037916891742497683, + 0.004293338395655155, + 0.006498193833976984, + 0.004631334450095892, + 0.0035133264027535915, + 0.006064186338335276, + 0.0027599413879215717, + 0.004118276759982109, + 0.0017346905078738928, + 0.0027031844947487116, + 0.008728720247745514, + 0.007098877336829901, + 0.006168351508677006, + 0.008354783989489079, + 0.003911082167178392, + 0.005244703497737646, + 0.005228365771472454, + 0.00594727136194706, + 0.011787069030106068, + 0.00835998822003603, + 0.008089116774499416, + 0.008193740621209145, + 0.003736986080184579, + 0.003714687656611204, + 0.0018389266915619373, + 0.0025423422921448946, + 0.022789442911744118, + 0.01105597335845232, + 0.03310535475611687, + 0.01000337302684784, + 0.00441655283793807, + 0.0036489362828433514, + 0.0040855142287909985, + 0.0030138674192130566, + 0.020264456048607826, + 0.022028526291251183, + 0.0829884484410286, + 0.02444883994758129, + 0.014842985197901726, + 0.013114502653479576, + 0.016604017466306686, + 0.08956550061702728, + 0.0048506311140954494, + 0.0031203327234834433, + 0.006695777643471956, + 0.0044632540084421635, + 0.0026074424386024475, + 0.0018456634134054184, + 0.0021717657800763845, + 0.0019295953679829836, + 0.004034728277474642, + 0.006923372391611338, + 0.008967336267232895, + 0.004142952151596546, + 0.0025052158161997795, + 0.003209436545148492, + 0.0025755520910024643, + 0.0030903522856533527, + 0.0018857503309845924, + 0.0024819381069391966, + 0.002301204716786742, + 0.003611705033108592, + 0.002279758919030428, + 0.0019141652155667543, + 0.0019493058789521456, + 0.0012152665294706821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.006732400041073561 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036109976936131716 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036109976936131716, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036109976936131716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16028350591659546, + 0.026660693809390068, + 0.05157413333654404, + 0.08579662442207336, + 0.1341865509748459, + 0.2321307212114334, + 0.14413641393184662, + 0.04845092073082924, + 0.058284055441617966, + 0.05849636346101761 + ], + "credit_map": [ + 0.020844057202339172, + 0.038597576320171356, + 0.03376416489481926, + 0.021816609427332878, + 0.012869296595454216, + 0.015646522864699364, + 0.006527288351207972, + 0.010217989794909954, + 0.0032525048591196537, + 0.005555782932788134, + 0.0029723020270466805, + 0.0033885634038597345, + 0.003311728360131383, + 0.0032873067539185286, + 0.0020345584489405155, + 0.0028579458594322205, + 0.009453410282731056, + 0.009416870772838593, + 0.005634695757180452, + 0.007171571720391512, + 0.00406524958088994, + 0.009711670689284801, + 0.0024087056517601013, + 0.0037119558546692133, + 0.013523157685995102, + 0.015565638430416584, + 0.015050172805786133, + 0.013406967744231224, + 0.007476587779819965, + 0.010333367623388767, + 0.005159320309758186, + 0.005281411577016115, + 0.03183513507246971, + 0.024774638935923576, + 0.02896248735487461, + 0.023139268159866333, + 0.006705252919346094, + 0.007071339059621096, + 0.005777508486062288, + 0.005920921918004751, + 0.06495697051286697, + 0.020258285105228424, + 0.0920109823346138, + 0.02396121621131897, + 0.007620108779519796, + 0.005741946864873171, + 0.009965836070477962, + 0.007615374866873026, + 0.011263261549174786, + 0.011874914169311523, + 0.040269047021865845, + 0.012643584981560707, + 0.008339415304362774, + 0.006634692195802927, + 0.00846036896109581, + 0.044651128351688385, + 0.00808895193040371, + 0.009256145916879177, + 0.010341566056013107, + 0.008210353553295135, + 0.0032998824026435614, + 0.0027206146623939276, + 0.003519438672810793, + 0.003013965906575322, + 0.0071099488995969296, + 0.010235087014734745, + 0.0118090296164155, + 0.009964403696358204, + 0.00513061136007309, + 0.0035433319862931967, + 0.005749677773565054, + 0.004741963930428028, + 0.006350859068334103, + 0.009129743091762066, + 0.00894337147474289, + 0.014359946362674236, + 0.006274549290537834, + 0.0043891738168895245, + 0.0049950783140957355, + 0.004053639713674784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0036109976936131716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an athletic young woman with warm olive skin riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant puffy winter coat with a glossy geometric pattern of turquoise, fuchsia, and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman riding on a narrow country road covered in a layer of crisp white snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of sparse birch and pine forests silhouetted against a twilight sky with a rosy sunset glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan from the side to follow the woman's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054920896887779236 + ], + "opposite_logprob_tokens": [ + -2.9299209117889404 + ], + "logprob_reward": -0.054920896887779236, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054920896887779236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3840877413749695, + 0.138527050614357, + 0.02717648446559906, + 0.04118405282497406, + 0.04613546282052994, + 0.05269913375377655, + 0.17522700130939484, + 0.05595957860350609, + 0.04249117523431778, + 0.03651224076747894 + ], + "credit_map": [ + 0.03349553793668747, + 0.11689530313014984, + 0.11847154051065445, + 0.0623205341398716, + 0.01847684197127819, + 0.015802057459950447, + 0.007848729379475117, + 0.010777214542031288, + 0.00984174944460392, + 0.049261853098869324, + 0.013214959762990475, + 0.036174386739730835, + 0.0093113137409091, + 0.007703489623963833, + 0.006058995146304369, + 0.006960307713598013, + 0.003584458027034998, + 0.004006067290902138, + 0.0026611220091581345, + 0.0024891477078199387, + 0.0028146114200353622, + 0.004417933989316225, + 0.003333165543153882, + 0.0038699796423316, + 0.00807906687259674, + 0.005478870123624802, + 0.004978933371603489, + 0.005432440433651209, + 0.004458632320165634, + 0.00312525755725801, + 0.003997093066573143, + 0.005633764900267124, + 0.008515287190675735, + 0.00765014486387372, + 0.005526603665202856, + 0.007053305860608816, + 0.004673273302614689, + 0.004729338455945253, + 0.0032849947456270456, + 0.004702520091086626, + 0.010360950604081154, + 0.0070275841280817986, + 0.014693803153932095, + 0.005802521947771311, + 0.004897243808954954, + 0.004611597862094641, + 0.0026282251346856356, + 0.0026772082783281803, + 0.012657607905566692, + 0.013830425217747688, + 0.05059603229165077, + 0.01492599118500948, + 0.009836018085479736, + 0.008799728937447071, + 0.010059799067676067, + 0.054521411657333374, + 0.007228352129459381, + 0.008255778811872005, + 0.00818661693483591, + 0.01209923904389143, + 0.005791023839265108, + 0.004790685139596462, + 0.005072536878287792, + 0.004535349551588297, + 0.0037785563617944717, + 0.005508157424628735, + 0.007104171440005302, + 0.007096399553120136, + 0.005747450049966574, + 0.00389869324862957, + 0.004861364606767893, + 0.004496387671679258, + 0.0031614811159670353, + 0.003975874278694391, + 0.004496490582823753, + 0.007325717248022556, + 0.00446323910728097, + 0.003966061864048243, + 0.004489076789468527, + 0.004634301643818617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.054920896887779236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_tensor([1677, 1677])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_tensor([1677, 1677])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aab6424d6e6fb3d5a56811778a97fe0d0237df82 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_tensor([1677, 1677])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping aerial shot slowly pans right, following a majestic Chinese dragon as it soars through mountains etched with the rich aesthetic of an ink wash painting. The dragon boasts a slender, sinuous body, covered in pearly, turquoise scales, each one sharply defined. A fiery crimson mane stretches from head to tail, striking against the austere backdrop. Its head is majestic, its antlers protruding, its amber eyes gleaming with wisdom and majesty, its long whiskers flowing gracefully in the air. It rises and falls in a powerful and rhythmic manner, its powerful claws folding and unfurling as its body stretches. Its massive tail traces a graceful arc behind it, its every movement seemingly stirring the air currents between heaven and earth, causing the surrounding clouds to swirl and disperse. The background is a continuous expanse of towering peaks, rendered in varying shades of ink. Nearby peaks are sharply defined, their ink rich, with a few ancient pines dotting the cliffs. Distant peaks gradually fade into a hazy mist of ethereal light ink and vast expanses of white. The dragon's form occasionally appears and disappears within the clouds, creating a profound sense of space. The dominant tones of the painting are classic black, white, and gray. The light, rather than originating from a single source, seems to permeate the rice paper itself, creating an even and soft radiance. The green of the dragon's body and the crimson of its mane, the only bright colors in the image, are subtly saturated yet profoundly charged, creating a visual epic that is both sacred and majestic, yet imbued with the tranquility and power of Eastern philosophy.", + "seed": 12895, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_1677_seed_12895_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0346729542594403, + "vlm_logprob": -0.0346729542594403, + "vlm_yes_no_margin": 4.274999976158142, + "vlm_token_credit": -0.0346729542594403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02014533430337906 + ], + "opposite_logprob_tokens": [ + -4.395145416259766 + ], + "logprob_reward": -0.02014533430337906, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02014533430337906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18319278955459595, + 0.10609596222639084, + 0.07550176233053207, + 0.10224730521440506, + 0.04470174014568329, + 0.13897818326950073, + 0.0286942757666111, + 0.2189786732196808, + 0.0632496252655983, + 0.03835972025990486 + ], + "credit_map": [ + 0.013645105063915253, + 0.013355028815567493, + 0.015570074319839478, + 0.057338714599609375, + 0.01837022602558136, + 0.01183521281927824, + 0.018140561878681183, + 0.034937888383865356, + 0.0038625546731054783, + 0.0034749717451632023, + 0.009667436592280865, + 0.005083589348942041, + 0.0034032752737402916, + 0.0354149155318737, + 0.006038556806743145, + 0.0391506589949131, + 0.01012579444795847, + 0.007849323563277721, + 0.00901289377361536, + 0.0063402000814676285, + 0.008302910253405571, + 0.010801654309034348, + 0.011181669309735298, + 0.011887316592037678, + 0.013759231194853783, + 0.008845719508826733, + 0.009976211003959179, + 0.009119500406086445, + 0.011900164186954498, + 0.016124380752444267, + 0.015379602089524269, + 0.01714249886572361, + 0.0061835674569010735, + 0.006988082081079483, + 0.004849805496633053, + 0.004916922189295292, + 0.004498883616179228, + 0.004556378815323114, + 0.005618174094706774, + 0.007089925464242697, + 0.012906735762953758, + 0.017068978399038315, + 0.02287260629236698, + 0.012500680051743984, + 0.02022502012550831, + 0.02470063604414463, + 0.01567261852324009, + 0.013030897825956345, + 0.002336719771847129, + 0.0018611042760312557, + 0.003466770052909851, + 0.004221199546009302, + 0.0028457751031965017, + 0.006063142325729132, + 0.005213376134634018, + 0.0026861897204071283, + 0.020322147756814957, + 0.019444609060883522, + 0.02015075460076332, + 0.01883358135819435, + 0.01948133297264576, + 0.024400312453508377, + 0.052592743188142776, + 0.043753188103437424, + 0.005110046360641718, + 0.006981781218200922, + 0.009627657011151314, + 0.014485913328826427, + 0.008381533436477184, + 0.00539675448089838, + 0.006634719669818878, + 0.006631218362599611, + 0.00413837656378746, + 0.00300367153249681, + 0.004966036882251501, + 0.006088590249419212, + 0.0032481553498655558, + 0.0029867051634937525, + 0.008388405665755272, + 0.0055397795513272285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.02014533430337906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087606549263 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087606549263, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087606549263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12500905990600586, + 0.1192951649427414, + 0.14685381948947906, + 0.09206534922122955, + 0.05468323454260826, + 0.20451949536800385, + 0.04215559735894203, + 0.0810818150639534, + 0.06445887684822083, + 0.06987763941287994 + ], + "credit_map": [ + 0.014099626801908016, + 0.021408459171652794, + 0.017861532047390938, + 0.017902102321386337, + 0.01731301285326481, + 0.00798217486590147, + 0.0108580831438303, + 0.017584064975380898, + 0.006215998437255621, + 0.007837180979549885, + 0.003490717150270939, + 0.007773806806653738, + 0.005032973829656839, + 0.040491733700037, + 0.00837217178195715, + 0.040080565959215164, + 0.016871899366378784, + 0.01570253260433674, + 0.015386801213026047, + 0.01173988077789545, + 0.015342498198151588, + 0.020050829276442528, + 0.02275494858622551, + 0.029004422947764397, + 0.010738271288573742, + 0.006959042511880398, + 0.006715624127537012, + 0.008610577322542667, + 0.01180730015039444, + 0.01719074137508869, + 0.015281040221452713, + 0.01476274523884058, + 0.011402188800275326, + 0.007127012591809034, + 0.00517706573009491, + 0.0046310522593557835, + 0.006230857223272324, + 0.005524188745766878, + 0.00684675807133317, + 0.0077441101893782616, + 0.02214115485548973, + 0.025122281163930893, + 0.031892310827970505, + 0.022556122392416, + 0.014710196293890476, + 0.03784914314746857, + 0.028046544641256332, + 0.022201748564839363, + 0.004718510899692774, + 0.0037529384717345238, + 0.0042336368933320045, + 0.00664372555911541, + 0.0046908678486943245, + 0.0077359238639473915, + 0.007271523587405682, + 0.0031084700021892786, + 0.007584025152027607, + 0.008084436878561974, + 0.012430749833583832, + 0.013538353145122528, + 0.007139722816646099, + 0.009579873643815517, + 0.012137975543737411, + 0.010586670599877834, + 0.005586958024650812, + 0.008540021255612373, + 0.006973656360059977, + 0.012801731005311012, + 0.00810289103537798, + 0.00865775253623724, + 0.006944821681827307, + 0.0068510426208376884, + 0.006725865416228771, + 0.004275423474609852, + 0.006081998813897371, + 0.018468666821718216, + 0.004648896399885416, + 0.005842864979058504, + 0.01199765782803297, + 0.01183626614511013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.016087606549263 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640542928129435 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640542928129435, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640542928129435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12424483895301819, + 0.13870011270046234, + 0.04823612794280052, + 0.055774614214897156, + 0.04472089186310768, + 0.1957586407661438, + 0.044181663542985916, + 0.13558094203472137, + 0.13926710188388824, + 0.07353498786687851 + ], + "credit_map": [ + 0.01697942614555359, + 0.019344134256243706, + 0.01536907721310854, + 0.02319398894906044, + 0.016706665977835655, + 0.007486404851078987, + 0.012343303300440311, + 0.012821853160858154, + 0.008645016700029373, + 0.008160877041518688, + 0.021938452497124672, + 0.015133325010538101, + 0.00935959443449974, + 0.03510579466819763, + 0.01617860607802868, + 0.024178454652428627, + 0.006566975731402636, + 0.006340291351079941, + 0.007163379807025194, + 0.005081410985440016, + 0.004116754978895187, + 0.007078424096107483, + 0.005505906417965889, + 0.006382984574884176, + 0.006619654595851898, + 0.004767392762005329, + 0.004641578998416662, + 0.005763933528214693, + 0.009926251135766506, + 0.007214640267193317, + 0.008576203137636185, + 0.008264965377748013, + 0.00648115947842598, + 0.006271106190979481, + 0.006195242051035166, + 0.003659194568172097, + 0.0042461068369448185, + 0.0044172280468046665, + 0.006520636845380068, + 0.00693022133782506, + 0.027702588587999344, + 0.024126609787344933, + 0.02872966229915619, + 0.01581130549311638, + 0.02788698486983776, + 0.033682581037282944, + 0.01883017085492611, + 0.018988754600286484, + 0.00465373182669282, + 0.0034389684442430735, + 0.0062222667038440704, + 0.0074106925167143345, + 0.004370908252894878, + 0.007272977381944656, + 0.007266910746693611, + 0.00354521325789392, + 0.00792816374450922, + 0.0076329647563397884, + 0.009165107272565365, + 0.009092921391129494, + 0.007785046938806772, + 0.007670130580663681, + 0.04565880075097084, + 0.04064782336354256, + 0.012669136747717857, + 0.012475617229938507, + 0.024282118305563927, + 0.0327659547328949, + 0.018317585811018944, + 0.01005890965461731, + 0.015231938101351261, + 0.01346585899591446, + 0.00847323052585125, + 0.004794351290911436, + 0.008300185203552246, + 0.020078403875231743, + 0.005449916701763868, + 0.005020746961236, + 0.009866518899798393, + 0.01155164372175932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.007640542928129435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248754327185452 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.00248754327185452, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248754327185452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14319303631782532, + 0.21708175539970398, + 0.09319674968719482, + 0.03731139376759529, + 0.04343000054359436, + 0.10692428052425385, + 0.0641268640756607, + 0.045406684279441833, + 0.12996987998485565, + 0.1193593367934227 + ], + "credit_map": [ + 0.017337307333946228, + 0.014467068947851658, + 0.011709609068930149, + 0.021943744271993637, + 0.0262758769094944, + 0.014360325410962105, + 0.021410180255770683, + 0.015688924118876457, + 0.010167751461267471, + 0.008881992660462856, + 0.025761203840374947, + 0.012255100533366203, + 0.014861280098557472, + 0.058701042085886, + 0.018178386613726616, + 0.06827500462532043, + 0.005806480068713427, + 0.0055689322762191296, + 0.008857203647494316, + 0.005948048084974289, + 0.012452635914087296, + 0.01351295504719019, + 0.018852224573493004, + 0.02219826728105545, + 0.004200458526611328, + 0.0027523378375917673, + 0.0043660919182002544, + 0.004002199973911047, + 0.005683620925992727, + 0.003824855200946331, + 0.0053007579408586025, + 0.00718107121065259, + 0.004323492292314768, + 0.0036359333898872137, + 0.007570987567305565, + 0.0030398648232221603, + 0.004309663083404303, + 0.00602532597258687, + 0.004067956004291773, + 0.01045677624642849, + 0.013129097409546375, + 0.010141567327082157, + 0.017965832725167274, + 0.012083335779607296, + 0.008019568398594856, + 0.02144598215818405, + 0.011942099779844284, + 0.012196797877550125, + 0.00654172757640481, + 0.006296872161328793, + 0.011353936977684498, + 0.011834143660962582, + 0.005541935563087463, + 0.007451261859387159, + 0.009147227741777897, + 0.00595975574105978, + 0.003316977061331272, + 0.0038148935418576, + 0.0053057135082781315, + 0.005813567899167538, + 0.005693962797522545, + 0.0029626074247062206, + 0.009520439431071281, + 0.00897852424532175, + 0.01161583885550499, + 0.014188106171786785, + 0.021720150485634804, + 0.03154106065630913, + 0.012669983319938183, + 0.010080624371767044, + 0.014302738942205906, + 0.013851381838321686, + 0.011217699386179447, + 0.014244849793612957, + 0.010428504087030888, + 0.015567787922918797, + 0.015504380688071251, + 0.017735227942466736, + 0.019231831654906273, + 0.015429060906171799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.00248754327185452 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270037442445755 + ], + "opposite_logprob_tokens": [ + -2.1270036697387695 + ], + "logprob_reward": -0.1270037442445755, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270037442445755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18390384316444397, + 0.09695781767368317, + 0.049303941428661346, + 0.06931660324335098, + 0.0753457173705101, + 0.1314821094274521, + 0.05078672245144844, + 0.07292426377534866, + 0.168394073843956, + 0.10158487409353256 + ], + "credit_map": [ + 0.011924207210540771, + 0.01374814659357071, + 0.013567953370511532, + 0.06632594019174576, + 0.019142229110002518, + 0.00955052301287651, + 0.016644679009914398, + 0.0330001637339592, + 0.003566022962331772, + 0.0029806108213961124, + 0.0030777554493397474, + 0.006793012842535973, + 0.0030736210756003857, + 0.0363900251686573, + 0.007949569262564182, + 0.03312719613313675, + 0.005093992687761784, + 0.00440110033378005, + 0.0080753518268466, + 0.0038509874138981104, + 0.006075391545891762, + 0.006394975818693638, + 0.0064999498426914215, + 0.008912189863622189, + 0.00822110753506422, + 0.006605933886021376, + 0.005229778587818146, + 0.007182203698903322, + 0.010250359773635864, + 0.008204314857721329, + 0.010287168435752392, + 0.013335732743144035, + 0.010800976306200027, + 0.01000928319990635, + 0.009218082763254642, + 0.00633270526304841, + 0.009162724018096924, + 0.009062590077519417, + 0.007219388615339994, + 0.013539962470531464, + 0.015092568472027779, + 0.015896452590823174, + 0.014873228035867214, + 0.013320479542016983, + 0.0238102488219738, + 0.02109386958181858, + 0.013964931480586529, + 0.013430324383080006, + 0.005574898328632116, + 0.004685454536229372, + 0.008217519149184227, + 0.008993176743388176, + 0.004414184484630823, + 0.00880246702581644, + 0.006201089359819889, + 0.0038979314267635345, + 0.00448495801538229, + 0.00502840057015419, + 0.005553696304559708, + 0.009552794508635998, + 0.006525552831590176, + 0.006522230803966522, + 0.017859958112239838, + 0.017396673560142517, + 0.012328269891440868, + 0.017451824620366096, + 0.020631708204746246, + 0.03238079324364662, + 0.028443679213523865, + 0.014084798283874989, + 0.022256998345255852, + 0.020816002041101456, + 0.016008902341127396, + 0.009013470262289047, + 0.010435942560434341, + 0.022138414904475212, + 0.006767964456230402, + 0.006805942859500647, + 0.014924912713468075, + 0.01548931747674942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.1270037442445755 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial shot slowly pans right, following a majestic Chinese dragon as it soars through mountains etched with the rich aesthetic of an ink wash painting. The dragon boasts a slender, sinuous body, covered in pearly, turquoise scales, each one sharply defined. A fiery crimson mane stretches from head to tail, striking against the austere backdrop. Its head is majestic, its antlers protruding, its amber eyes gleaming with wisdom and majesty, its long whiskers flowing gracefully in the air. It rises and falls in a powerful and rhythmic manner, its powerful claws folding and unfurling as its body stretches. Its massive tail traces a graceful arc behind it, its every movement seemingly stirring the air currents between heaven and earth, causing the surrounding clouds to swirl and disperse. The background is a continuous expanse of towering peaks, rendered in varying shades of ink. Nearby peaks are sharply defined, their ink rich, with a few ancient pines dotting the cliffs. Distant peaks gradually fade into a hazy mist of ethereal light ink and vast expanses of white. The dragon's form occasionally appears and disappears within the clouds, creating a profound sense of space. The dominant tones of the painting are classic black, white, and gray. The light, rather than originating from a single source, seems to permeate the rice paper itself, creating an even and soft radiance. The green of the dragon's body and the crimson of its mane, the only bright colors in the image, are subtly saturated yet profoundly charged, creating a visual epic that is both sacred and majestic, yet imbued with the tranquility and power of Eastern philosophy.", + "seed": 12896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_1677_seed_12896_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017564782220870257, + "vlm_logprob": -0.017564782220870257, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.017564782220870257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007284272462129593 + ], + "opposite_logprob_tokens": [ + -5.757284164428711 + ], + "logprob_reward": -0.007284272462129593, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007284272462129593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08396556228399277, + 0.07982306182384491, + 0.0707535445690155, + 0.07899332791566849, + 0.10029755532741547, + 0.1644132286310196, + 0.07493674755096436, + 0.16876550018787384, + 0.11582572758197784, + 0.062225792557001114 + ], + "credit_map": [ + 0.009445887990295887, + 0.009453294798731804, + 0.010143372230231762, + 0.006850904319435358, + 0.017313987016677856, + 0.011787068098783493, + 0.0058808536268770695, + 0.013090187683701515, + 0.015213045291602612, + 0.008199438452720642, + 0.011339336633682251, + 0.007225924637168646, + 0.010275769047439098, + 0.009657214395701885, + 0.007571449503302574, + 0.01034088060259819, + 0.0073465886525809765, + 0.005911557469516993, + 0.007319219410419464, + 0.007606568746268749, + 0.010097501799464226, + 0.008393866010010242, + 0.013896908611059189, + 0.010181327350437641, + 0.008052749559283257, + 0.01519792340695858, + 0.006833331193774939, + 0.007267296314239502, + 0.00810252781957388, + 0.014320151880383492, + 0.007407112512737513, + 0.01181222964078188, + 0.012170779518783092, + 0.013807730749249458, + 0.009427528828382492, + 0.010486666113138199, + 0.01362678688019514, + 0.015782445669174194, + 0.011861477047204971, + 0.013134137727320194, + 0.023817911744117737, + 0.017367102205753326, + 0.012829787097871304, + 0.016902489587664604, + 0.018706312403082848, + 0.02300751768052578, + 0.0195249505341053, + 0.032257139682769775, + 0.009914103895425797, + 0.00989323016256094, + 0.008672396652400494, + 0.01253935880959034, + 0.010696553625166416, + 0.006684239022433758, + 0.009694239124655724, + 0.006842614151537418, + 0.013881254009902477, + 0.027769630774855614, + 0.029159460216760635, + 0.022849086672067642, + 0.018666105344891548, + 0.018256139010190964, + 0.02427465282380581, + 0.013909160159528255, + 0.011227242648601532, + 0.012042775750160217, + 0.01636836677789688, + 0.015724163502454758, + 0.01819516532123089, + 0.016291668638586998, + 0.014665043912827969, + 0.01131129078567028, + 0.009408924728631973, + 0.006290335673838854, + 0.01057105977088213, + 0.008493379689753056, + 0.007301499601453543, + 0.006867512594908476, + 0.005489946808665991, + 0.00780312716960907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.007284272462129593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081814765930176 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081814765930176, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081814765930176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15757820010185242, + 0.049737777560949326, + 0.037580374628305435, + 0.04282410815358162, + 0.03906368836760521, + 0.30069416761398315, + 0.041544295847415924, + 0.2330581396818161, + 0.04704008623957634, + 0.05087913200259209 + ], + "credit_map": [ + 0.013557094149291515, + 0.019421037286520004, + 0.021631751209497452, + 0.011540920473635197, + 0.03644521161913872, + 0.026411743834614754, + 0.011140027083456516, + 0.017430415377020836, + 0.00382943544536829, + 0.005781279411166906, + 0.005647957790642977, + 0.003958814777433872, + 0.009992892853915691, + 0.00826960988342762, + 0.00410108407959342, + 0.008156700059771538, + 0.0035781164187937975, + 0.004257029853761196, + 0.003639797680079937, + 0.004287853371351957, + 0.00790981575846672, + 0.004807859193533659, + 0.004216821398586035, + 0.004883081652224064, + 0.0038910063449293375, + 0.00898326002061367, + 0.003710540011525154, + 0.0035409184638410807, + 0.005598505958914757, + 0.00668904185295105, + 0.004123857710510492, + 0.006286975461989641, + 0.0029818082693964243, + 0.005581804551184177, + 0.0038092797622084618, + 0.008151325397193432, + 0.004943329840898514, + 0.0028280133847147226, + 0.005914363078773022, + 0.004853765945881605, + 0.04575024172663689, + 0.04161176458001137, + 0.020568089559674263, + 0.030807986855506897, + 0.033223215490579605, + 0.04065841808915138, + 0.03468555957078934, + 0.05338891223073006, + 0.005742919165641069, + 0.005255147814750671, + 0.004572068806737661, + 0.00614863820374012, + 0.006118289195001125, + 0.004653931595385075, + 0.005506234709173441, + 0.0035470661241561174, + 0.01953401230275631, + 0.03272301331162453, + 0.04684362933039665, + 0.03097803331911564, + 0.023047350347042084, + 0.03271637484431267, + 0.026569169014692307, + 0.020646564662456512, + 0.004222135525196791, + 0.00509504834190011, + 0.005245062988251448, + 0.005350941326469183, + 0.009308535605669022, + 0.008556419983506203, + 0.005155124235898256, + 0.0041068196296691895, + 0.00762604083865881, + 0.006261288188397884, + 0.008469462394714355, + 0.006132578942924738, + 0.006345648318529129, + 0.006234175059944391, + 0.00494534894824028, + 0.004864593967795372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.011081814765930176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11575137078762054, + 0.05089431256055832, + 0.07711267471313477, + 0.20564353466033936, + 0.21388211846351624, + 0.12288567423820496, + 0.035890791565179825, + 0.05838323384523392, + 0.06748860329389572, + 0.05206771567463875 + ], + "credit_map": [ + 0.008216018788516521, + 0.012496018782258034, + 0.018239416182041168, + 0.009096499532461166, + 0.020534610375761986, + 0.01770179532468319, + 0.007776365149766207, + 0.021690646186470985, + 0.006634887773543596, + 0.005596481263637543, + 0.006560316309332848, + 0.004919461440294981, + 0.007044926285743713, + 0.006863634567707777, + 0.005233661737293005, + 0.008040945045650005, + 0.00807943381369114, + 0.01003426406532526, + 0.009721983224153519, + 0.009559323079884052, + 0.013053154572844505, + 0.009279952384531498, + 0.008439379744231701, + 0.008945182897150517, + 0.023732641711831093, + 0.02337784692645073, + 0.023262539878487587, + 0.015159822069108486, + 0.03853316977620125, + 0.03470703959465027, + 0.019271589815616608, + 0.027598880231380463, + 0.01180480420589447, + 0.01979300007224083, + 0.02080192044377327, + 0.055067576467990875, + 0.02293284423649311, + 0.028458599001169205, + 0.032448478043079376, + 0.0225748922675848, + 0.014624943025410175, + 0.014030225574970245, + 0.00968682300299406, + 0.013321089558303356, + 0.011758210137486458, + 0.01906750351190567, + 0.014549179933965206, + 0.02584768831729889, + 0.004378500860184431, + 0.0036784287076443434, + 0.005495553836226463, + 0.005216995254158974, + 0.006375758908689022, + 0.0032884571701288223, + 0.004698213189840317, + 0.002758886432275176, + 0.005603895056992769, + 0.010557747446000576, + 0.008502249605953693, + 0.007236724253743887, + 0.007299636956304312, + 0.007619951386004686, + 0.007232853677123785, + 0.004330172669142485, + 0.006021079141646624, + 0.006455363240092993, + 0.008222687058150768, + 0.008333803154528141, + 0.013514507561922073, + 0.010725535452365875, + 0.007892332971096039, + 0.006323294714093208, + 0.00883521419018507, + 0.0059053669683635235, + 0.007814849726855755, + 0.006776832509785891, + 0.006809399463236332, + 0.005397740751504898, + 0.004591410048305988, + 0.005936899688094854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.005256166215986013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021943794563412666 + ], + "opposite_logprob_tokens": [ + -6.127194404602051 + ], + "logprob_reward": -0.0021943794563412666, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021943794563412666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16163553297519684, + 0.09252910315990448, + 0.09553434699773788, + 0.0585944689810276, + 0.095940500497818, + 0.22223982214927673, + 0.08309728652238846, + 0.06493653357028961, + 0.0450652539730072, + 0.0804271399974823 + ], + "credit_map": [ + 0.013837914913892746, + 0.014406925067305565, + 0.027199996635317802, + 0.010137183591723442, + 0.035776302218437195, + 0.026194661855697632, + 0.010389413684606552, + 0.023693129420280457, + 0.012040367349982262, + 0.008285443298518658, + 0.012287948280572891, + 0.01028425432741642, + 0.00956797506660223, + 0.012689641676843166, + 0.012303407303988934, + 0.015070066787302494, + 0.016197215765714645, + 0.007470760028809309, + 0.011644566431641579, + 0.011819515377283096, + 0.015303113497793674, + 0.010240581817924976, + 0.009711179882287979, + 0.013147411867976189, + 0.004621753469109535, + 0.0112551124766469, + 0.00575310830026865, + 0.0051964400336146355, + 0.007482234854251146, + 0.0110548734664917, + 0.004867929499596357, + 0.008363014087080956, + 0.009180551394820213, + 0.009605365805327892, + 0.010871810838580132, + 0.019221898168325424, + 0.009386733174324036, + 0.011148443445563316, + 0.015483810566365719, + 0.01104189082980156, + 0.024630693718791008, + 0.029237009584903717, + 0.01608448661863804, + 0.02647457830607891, + 0.01961064338684082, + 0.03238605335354805, + 0.028720371425151825, + 0.04509597271680832, + 0.009363927878439426, + 0.013033369556069374, + 0.007719947025179863, + 0.011143757961690426, + 0.016493622213602066, + 0.007574897725135088, + 0.012076904997229576, + 0.005690855905413628, + 0.004352849442511797, + 0.007179390173405409, + 0.01173787098377943, + 0.009167090058326721, + 0.009928819723427296, + 0.009265868924558163, + 0.008243235759437084, + 0.005061411298811436, + 0.0039382511749863625, + 0.004805592820048332, + 0.006995517760515213, + 0.0055154068395495415, + 0.008225373923778534, + 0.006839458364993334, + 0.005348472855985165, + 0.003397183958441019, + 0.012176903896033764, + 0.010557739995419979, + 0.009605996310710907, + 0.010717726312577724, + 0.009008284658193588, + 0.007685569114983082, + 0.009525286965072155, + 0.011149635538458824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0021943794563412666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06200727820396423 + ], + "opposite_logprob_tokens": [ + -2.812007188796997 + ], + "logprob_reward": -0.06200727820396423, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06200727820396423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11481497436761856, + 0.06693100929260254, + 0.11490406841039658, + 0.08636721968650818, + 0.03599759191274643, + 0.28406184911727905, + 0.05289705470204353, + 0.11297165602445602, + 0.07270943373441696, + 0.05834513157606125 + ], + "credit_map": [ + 0.012434943579137325, + 0.01818164624273777, + 0.01761459931731224, + 0.009554561227560043, + 0.017453523352742195, + 0.016186071559786797, + 0.008294984698295593, + 0.015094649977982044, + 0.008509024046361446, + 0.007935386151075363, + 0.006338871084153652, + 0.006468685809522867, + 0.01356838271021843, + 0.009780933149158955, + 0.005275136791169643, + 0.009054592810571194, + 0.01813093200325966, + 0.015839118510484695, + 0.010605885647237301, + 0.012245199643075466, + 0.017688941210508347, + 0.013428973965346813, + 0.011655393056571484, + 0.015309631824493408, + 0.008367903530597687, + 0.01434638723731041, + 0.009704121388494968, + 0.009939623065292835, + 0.00929289124906063, + 0.014281470328569412, + 0.009448193944990635, + 0.0109866326674819, + 0.0027932797092944384, + 0.004038185812532902, + 0.0033794844057410955, + 0.0042229327373206615, + 0.006173747591674328, + 0.006066460162401199, + 0.004763737786561251, + 0.004559765104204416, + 0.03955543413758278, + 0.036747198551893234, + 0.020286906510591507, + 0.028456535190343857, + 0.028464963659644127, + 0.03966812789440155, + 0.03467565029859543, + 0.05620703846216202, + 0.005888042505830526, + 0.007099042180925608, + 0.005797133315354586, + 0.005388488527387381, + 0.011966072954237461, + 0.005151529796421528, + 0.007455646526068449, + 0.0041511026211082935, + 0.008992844261229038, + 0.021600008010864258, + 0.018594816327095032, + 0.013186505064368248, + 0.01061973161995411, + 0.01678454875946045, + 0.013846327550709248, + 0.009346887469291687, + 0.0069037084467709064, + 0.007656061556190252, + 0.01217307522892952, + 0.008202878758311272, + 0.011552195064723492, + 0.01081253495067358, + 0.009171872399747372, + 0.0062371110543608665, + 0.011237207800149918, + 0.007688547484576702, + 0.008701433427631855, + 0.0072747282683849335, + 0.0064491042867302895, + 0.0062523153610527515, + 0.005374320782721043, + 0.005367479287087917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.06200727820396423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial shot slowly pans right, following a majestic Chinese dragon as it soars through mountains etched with the rich aesthetic of an ink wash painting. The dragon boasts a slender, sinuous body, covered in pearly, turquoise scales, each one sharply defined. A fiery crimson mane stretches from head to tail, striking against the austere backdrop. Its head is majestic, its antlers protruding, its amber eyes gleaming with wisdom and majesty, its long whiskers flowing gracefully in the air. It rises and falls in a powerful and rhythmic manner, its powerful claws folding and unfurling as its body stretches. Its massive tail traces a graceful arc behind it, its every movement seemingly stirring the air currents between heaven and earth, causing the surrounding clouds to swirl and disperse. The background is a continuous expanse of towering peaks, rendered in varying shades of ink. Nearby peaks are sharply defined, their ink rich, with a few ancient pines dotting the cliffs. Distant peaks gradually fade into a hazy mist of ethereal light ink and vast expanses of white. The dragon's form occasionally appears and disappears within the clouds, creating a profound sense of space. The dominant tones of the painting are classic black, white, and gray. The light, rather than originating from a single source, seems to permeate the rice paper itself, creating an even and soft radiance. The green of the dragon's body and the crimson of its mane, the only bright colors in the image, are subtly saturated yet profoundly charged, creating a visual epic that is both sacred and majestic, yet imbued with the tranquility and power of Eastern philosophy.", + "seed": 12897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_1677_seed_12897_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028929815255105497, + "vlm_logprob": -0.028929815255105497, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.028929815255105497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017795894294977188 + ], + "opposite_logprob_tokens": [ + -4.517796039581299 + ], + "logprob_reward": -0.017795894294977188, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017795894294977188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1211579293012619, + 0.04623298719525337, + 0.11828341335058212, + 0.076789490878582, + 0.32354697585105896, + 0.13056151568889618, + 0.05910366773605347, + 0.056421179324388504, + 0.03709452599287033, + 0.030808279290795326 + ], + "credit_map": [ + 0.017011750489473343, + 0.016546204686164856, + 0.015159673057496548, + 0.013851291500031948, + 0.013049392960965633, + 0.01849083974957466, + 0.014739248901605606, + 0.012309521436691284, + 0.004944093059748411, + 0.005976660177111626, + 0.004896268714219332, + 0.0031319803092628717, + 0.005334554705768824, + 0.007860411889851093, + 0.007330035325139761, + 0.0067589860409498215, + 0.0046539828181266785, + 0.003377851564437151, + 0.008307386189699173, + 0.0031495578587055206, + 0.0361163504421711, + 0.005661411676555872, + 0.051031727343797684, + 0.005985150579363108, + 0.007925326935946941, + 0.008735462091863155, + 0.012150716036558151, + 0.008832920342683792, + 0.006590839009732008, + 0.0106009216979146, + 0.009124836884438992, + 0.012828465551137924, + 0.01909475028514862, + 0.010273465886712074, + 0.11509139090776443, + 0.02049705944955349, + 0.018335942178964615, + 0.005908329505473375, + 0.11073122173547745, + 0.023614807054400444, + 0.021950243040919304, + 0.012108143419027328, + 0.016058165580034256, + 0.014600183814764023, + 0.017673444002866745, + 0.016121502965688705, + 0.01565888710319996, + 0.01639094017446041, + 0.007634725887328386, + 0.005505552515387535, + 0.007105539552867413, + 0.007139785680919886, + 0.007129566743969917, + 0.006628107279539108, + 0.010351437143981457, + 0.007608956191688776, + 0.004357290454208851, + 0.006934671197086573, + 0.005931449588388205, + 0.007504898589104414, + 0.005860531236976385, + 0.010203677229583263, + 0.008726542815566063, + 0.006902116350829601, + 0.004167456179857254, + 0.003175602527335286, + 0.00441175140440464, + 0.005793763790279627, + 0.004100856836885214, + 0.004879476502537727, + 0.004862749017775059, + 0.0057028657756745815, + 0.003955169580876827, + 0.002795252250507474, + 0.0046989223919808865, + 0.005232538096606731, + 0.0025785835459828377, + 0.002811053767800331, + 0.0049172667786479, + 0.0038194942753762007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.017795894294977188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356063783168793 + ], + "opposite_logprob_tokens": [ + -3.6513559818267822 + ], + "logprob_reward": -0.026356063783168793, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356063783168793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07493698596954346, + 0.06009400635957718, + 0.3742118775844574, + 0.027923623099923134, + 0.07974544912576675, + 0.05924305319786072, + 0.15888285636901855, + 0.057858191430568695, + 0.07331483066082001, + 0.03378907963633537 + ], + "credit_map": [ + 0.011974946595728397, + 0.012873655185103416, + 0.006416697520762682, + 0.009918385185301304, + 0.007317926734685898, + 0.010124358348548412, + 0.008446948602795601, + 0.007864064536988735, + 0.006712726783007383, + 0.006410219240933657, + 0.0068050893023610115, + 0.009173023514449596, + 0.005417698062956333, + 0.008386524394154549, + 0.007594883441925049, + 0.009593836963176727, + 0.016371918842196465, + 0.01177400816231966, + 0.02131970226764679, + 0.015280262567102909, + 0.11588887125253677, + 0.020110884681344032, + 0.15171894431114197, + 0.02174728736281395, + 0.00294569437392056, + 0.002433018758893013, + 0.0038212710060179234, + 0.002903348533436656, + 0.002141093835234642, + 0.003752613440155983, + 0.0050637321546673775, + 0.004862852394580841, + 0.004897479899227619, + 0.0024708639830350876, + 0.028902286663651466, + 0.00569194694980979, + 0.004328119568526745, + 0.002080651232972741, + 0.024713855236768723, + 0.006660243961960077, + 0.005243981722742319, + 0.004473431967198849, + 0.00947294570505619, + 0.009004933759570122, + 0.007567654363811016, + 0.008364591747522354, + 0.007980290800333023, + 0.007135224062949419, + 0.02100425399839878, + 0.012364699505269527, + 0.030900079756975174, + 0.027026992291212082, + 0.010261928662657738, + 0.015013160184025764, + 0.024699099361896515, + 0.017612649127840996, + 0.0055534448474645615, + 0.007877151481807232, + 0.008733236230909824, + 0.008965383283793926, + 0.0057348692789673805, + 0.006902288179844618, + 0.006460034288465977, + 0.007631785701960325, + 0.00764335785061121, + 0.011619588360190392, + 0.01066911593079567, + 0.009497329592704773, + 0.010629652999341488, + 0.007507123053073883, + 0.008942759595811367, + 0.006805906072258949, + 0.004872020799666643, + 0.003164174733683467, + 0.00476680975407362, + 0.005823022685945034, + 0.0024973915424197912, + 0.0023507955484092236, + 0.0039734505116939545, + 0.006341417785733938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.026356063783168793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787900373339653 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009787900373339653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787900373339653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18070092797279358, + 0.05993608012795448, + 0.07516749203205109, + 0.047861672937870026, + 0.2778621017932892, + 0.09545014053583145, + 0.08421640843153, + 0.08948144316673279, + 0.04228722304105759, + 0.04703646898269653 + ], + "credit_map": [ + 0.02294580079615116, + 0.017072007060050964, + 0.018279733136296272, + 0.01831388659775257, + 0.019931642338633537, + 0.03093869797885418, + 0.029884137213230133, + 0.02333502098917961, + 0.0091848224401474, + 0.005424935836344957, + 0.005509288981556892, + 0.00553129892796278, + 0.006258631125092506, + 0.01016988418996334, + 0.010174939408898354, + 0.007682276889681816, + 0.005601342301815748, + 0.00358596327714622, + 0.003942277282476425, + 0.002640220569446683, + 0.022248970344662666, + 0.004544171970337629, + 0.027987470850348473, + 0.00461707403883338, + 0.004169713705778122, + 0.0047397674061357975, + 0.007493792567402124, + 0.006109332665801048, + 0.003768189810216427, + 0.006593925878405571, + 0.0068627819418907166, + 0.008124166168272495, + 0.016552722081542015, + 0.01163596659898758, + 0.09867287427186966, + 0.01902148500084877, + 0.018074940890073776, + 0.0073006171733140945, + 0.07820673286914825, + 0.028396764770150185, + 0.014033802784979343, + 0.005699093919247389, + 0.012005447410047054, + 0.01056593842804432, + 0.013826653361320496, + 0.012063334695994854, + 0.013682783581316471, + 0.013573085889220238, + 0.014739480800926685, + 0.0072798095643520355, + 0.015611289069056511, + 0.013273390009999275, + 0.005313179921358824, + 0.007357565686106682, + 0.010417717508971691, + 0.01022398006170988, + 0.007993516512215137, + 0.011401975527405739, + 0.008094614371657372, + 0.008225117810070515, + 0.011440391652286053, + 0.017316944897174835, + 0.013675929978489876, + 0.011332953348755836, + 0.004912916570901871, + 0.004225027281790972, + 0.0055227638222277164, + 0.005803964100778103, + 0.006446119863539934, + 0.0037151251453906298, + 0.005253614392131567, + 0.006407693959772587, + 0.007472840137779713, + 0.0046835970133543015, + 0.004960345104336739, + 0.007635442540049553, + 0.0036700372584164143, + 0.003363773226737976, + 0.0053662704303860664, + 0.009884162805974483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.009787900373339653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017122859135270119 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017122859135270119, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017122859135270119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04898688569664955, + 0.07447841018438339, + 0.3279457688331604, + 0.04512375593185425, + 0.2985607087612152, + 0.056941572576761246, + 0.04793032631278038, + 0.016932550817728043, + 0.03086664341390133, + 0.05223332345485687 + ], + "credit_map": [ + 0.006300522480159998, + 0.0037382033187896013, + 0.0053832088597118855, + 0.005917842499911785, + 0.00505373440682888, + 0.0068803816102445126, + 0.008455876260995865, + 0.007257119286805391, + 0.007098234724253416, + 0.007792386692017317, + 0.007876834832131863, + 0.008434269577264786, + 0.007341678719967604, + 0.011725340038537979, + 0.011524457484483719, + 0.012685205787420273, + 0.013959430158138275, + 0.009969434700906277, + 0.021863430738449097, + 0.012075127102434635, + 0.09690796583890915, + 0.019319742918014526, + 0.13247200846672058, + 0.021378636360168457, + 0.007014974486082792, + 0.003579959273338318, + 0.008715641684830189, + 0.003715601982548833, + 0.003302733413875103, + 0.004783264361321926, + 0.007807151880115271, + 0.006204430479556322, + 0.0185224786400795, + 0.009501265361905098, + 0.10350200533866882, + 0.017391564324498177, + 0.016861898824572563, + 0.005121422000229359, + 0.1057097390294075, + 0.02195032685995102, + 0.006717211101204157, + 0.005322320386767387, + 0.007826634682714939, + 0.009014121256768703, + 0.00798390805721283, + 0.006150557193905115, + 0.006759101990610361, + 0.0071677169762551785, + 0.005261484533548355, + 0.003404892049729824, + 0.010498473420739174, + 0.007493015844374895, + 0.003588376333937049, + 0.0047632926143705845, + 0.0073561943136155605, + 0.005564596503973007, + 0.0024157243315130472, + 0.002225491451099515, + 0.002300959313288331, + 0.0018422260181978345, + 0.0016060088528320193, + 0.0023774823639541864, + 0.0015285195549950004, + 0.0026361390482634306, + 0.004151982720941305, + 0.0024570797104388475, + 0.005312648136168718, + 0.0033083802554756403, + 0.0036651575937867165, + 0.0030929872300475836, + 0.004100047517567873, + 0.004778360016644001, + 0.005823837593197823, + 0.00638572545722127, + 0.00528445839881897, + 0.007166984956711531, + 0.005322310607880354, + 0.004789566155523062, + 0.006041730288416147, + 0.011418711394071579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.0017122859135270119 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08899693191051483 + ], + "opposite_logprob_tokens": [ + -2.4639968872070312 + ], + "logprob_reward": -0.08899693191051483, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08899693191051483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06002109497785568, + 0.03634841367602348, + 0.2604334056377411, + 0.027366740629076958, + 0.38248133659362793, + 0.05717641860246658, + 0.05890434980392456, + 0.02534271776676178, + 0.03767890855669975, + 0.05424656718969345 + ], + "credit_map": [ + 0.005215594079345465, + 0.0073981075547635555, + 0.006763181649148464, + 0.0061202747747302055, + 0.007230857387185097, + 0.009913761168718338, + 0.007619297131896019, + 0.009760027751326561, + 0.004430051892995834, + 0.0039072344079613686, + 0.003898889059200883, + 0.0026631855871528387, + 0.004308138508349657, + 0.005089278798550367, + 0.006107226945459843, + 0.005944410338997841, + 0.009566107764840126, + 0.005730560980737209, + 0.016182897612452507, + 0.006448508705943823, + 0.08463910222053528, + 0.011460605077445507, + 0.11497350037097931, + 0.011432119645178318, + 0.002342380816116929, + 0.001951119746081531, + 0.004297461360692978, + 0.0019327994668856263, + 0.0021659613121300936, + 0.004304499365389347, + 0.005949626676738262, + 0.004422895610332489, + 0.023376284167170525, + 0.010223153047263622, + 0.13400518894195557, + 0.022353999316692352, + 0.022660046815872192, + 0.007178863510489464, + 0.13447588682174683, + 0.02820797637104988, + 0.010159721598029137, + 0.004568357951939106, + 0.00831312220543623, + 0.0056882998906075954, + 0.00945814698934555, + 0.0059454175643622875, + 0.005956906359642744, + 0.007086453028023243, + 0.008959988132119179, + 0.0050663696601986885, + 0.00778784928843379, + 0.009800889529287815, + 0.0051636043936014175, + 0.005277013406157494, + 0.009831251576542854, + 0.007017388474196196, + 0.0033265354577451944, + 0.0035225385800004005, + 0.0020466968417167664, + 0.002205983968451619, + 0.003896702080965042, + 0.003735535079613328, + 0.003197032492607832, + 0.003411696059629321, + 0.004976384807378054, + 0.0029037129133939743, + 0.0046189455315470695, + 0.00512629933655262, + 0.004491669591516256, + 0.0035406395327299833, + 0.004718318115919828, + 0.007302945014089346, + 0.005862212739884853, + 0.007750747259706259, + 0.006797044537961483, + 0.009260451421141624, + 0.005212048534303904, + 0.0037799424026161432, + 0.009212306700646877, + 0.0063718208111822605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.08899693191051483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial shot slowly pans right, following a majestic Chinese dragon as it soars through mountains etched with the rich aesthetic of an ink wash painting. The dragon boasts a slender, sinuous body, covered in pearly, turquoise scales, each one sharply defined. A fiery crimson mane stretches from head to tail, striking against the austere backdrop. Its head is majestic, its antlers protruding, its amber eyes gleaming with wisdom and majesty, its long whiskers flowing gracefully in the air. It rises and falls in a powerful and rhythmic manner, its powerful claws folding and unfurling as its body stretches. Its massive tail traces a graceful arc behind it, its every movement seemingly stirring the air currents between heaven and earth, causing the surrounding clouds to swirl and disperse. The background is a continuous expanse of towering peaks, rendered in varying shades of ink. Nearby peaks are sharply defined, their ink rich, with a few ancient pines dotting the cliffs. Distant peaks gradually fade into a hazy mist of ethereal light ink and vast expanses of white. The dragon's form occasionally appears and disappears within the clouds, creating a profound sense of space. The dominant tones of the painting are classic black, white, and gray. The light, rather than originating from a single source, seems to permeate the rice paper itself, creating an even and soft radiance. The green of the dragon's body and the crimson of its mane, the only bright colors in the image, are subtly saturated yet profoundly charged, creating a visual epic that is both sacred and majestic, yet imbued with the tranquility and power of Eastern philosophy.", + "seed": 12898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_14/index_1677_seed_12898_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009674775251187383, + "vlm_logprob": -0.009674775251187383, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.009674775251187383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010346348397433758 + ], + "opposite_logprob_tokens": [ + -5.635346412658691 + ], + "logprob_reward": -0.010346348397433758, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.010346348397433758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1635442078113556, + 0.07292578369379044, + 0.09958634525537491, + 0.0873003900051117, + 0.061418503522872925, + 0.05994569510221481, + 0.12885722517967224, + 0.1373494565486908, + 0.08164958655834198, + 0.10742280632257462 + ], + "credit_map": [ + 0.022579368203878403, + 0.013839482329785824, + 0.0244719497859478, + 0.025208692997694016, + 0.015284419991075993, + 0.019761987030506134, + 0.01990228146314621, + 0.022496022284030914, + 0.00736591499298811, + 0.00818011537194252, + 0.018912548199295998, + 0.010043686255812645, + 0.005490295123308897, + 0.009662030264735222, + 0.006383533123880625, + 0.0068876598961651325, + 0.0110605014488101, + 0.01824968121945858, + 0.011184685863554478, + 0.011302642524242401, + 0.013528719544410706, + 0.012386384420096874, + 0.00985019188374281, + 0.01202353835105896, + 0.009476548060774803, + 0.008970088325440884, + 0.009227929636836052, + 0.010467294603586197, + 0.01318278070539236, + 0.015019342303276062, + 0.010498395189642906, + 0.01045800931751728, + 0.007143200840801001, + 0.006574470549821854, + 0.008982990868389606, + 0.005476539488881826, + 0.009278479032218456, + 0.009992565959692001, + 0.007781829684972763, + 0.006188429892063141, + 0.008678672835230827, + 0.009174371138215065, + 0.008136562071740627, + 0.008812496438622475, + 0.0071234628558158875, + 0.006397083401679993, + 0.007192338816821575, + 0.004430705215781927, + 0.014279844239354134, + 0.017644578590989113, + 0.018422268331050873, + 0.02408628910779953, + 0.014598974958062172, + 0.013461850583553314, + 0.014660058543086052, + 0.011703357100486755, + 0.013509055599570274, + 0.018983373418450356, + 0.01515681017190218, + 0.03337155655026436, + 0.019065188243985176, + 0.011804847046732903, + 0.015023697167634964, + 0.010434921830892563, + 0.00870655570179224, + 0.01868552155792713, + 0.015392334200441837, + 0.010320981964468956, + 0.010060157626867294, + 0.0056212423369288445, + 0.006437726318836212, + 0.006425064522773027, + 0.012914369814097881, + 0.014406850561499596, + 0.012553366832435131, + 0.009855322539806366, + 0.009020616300404072, + 0.020221766084432602, + 0.01339093130081892, + 0.015059588477015495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.010346348397433758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646313101053238 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646313101053238, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646313101053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20965109765529633, + 0.19039702415466309, + 0.11127252876758575, + 0.03420430049300194, + 0.07513069361448288, + 0.038656946271657944, + 0.08918828517198563, + 0.1748214066028595, + 0.03553071990609169, + 0.04114702343940735 + ], + "credit_map": [ + 0.021703600883483887, + 0.015484606847167015, + 0.03744468465447426, + 0.030911270529031754, + 0.021452780812978745, + 0.027462922036647797, + 0.021193400025367737, + 0.03399784117937088, + 0.02120613493025303, + 0.018952641636133194, + 0.040849220007658005, + 0.039068710058927536, + 0.013320577330887318, + 0.023047666996717453, + 0.016589418053627014, + 0.01736266165971756, + 0.013943718746304512, + 0.016946669667959213, + 0.016164513304829597, + 0.02165941707789898, + 0.010349920950829983, + 0.00841236487030983, + 0.012170608155429363, + 0.011625313200056553, + 0.003708094358444214, + 0.0037652598693966866, + 0.004206468351185322, + 0.005290498491376638, + 0.004339149687439203, + 0.005618405994027853, + 0.0030635888688266277, + 0.004212834406644106, + 0.008365720510482788, + 0.00976847019046545, + 0.010990692302584648, + 0.007788422517478466, + 0.010266476310789585, + 0.008730421774089336, + 0.01048362534493208, + 0.008736866526305676, + 0.004780468996614218, + 0.003962391056120396, + 0.0032869521528482437, + 0.0059157866053283215, + 0.007072202395647764, + 0.0077551063150167465, + 0.0034459400922060013, + 0.002438097959384322, + 0.009778834879398346, + 0.010188977234065533, + 0.011450915597379208, + 0.012226642109453678, + 0.009431474842131138, + 0.010676483623683453, + 0.01190020889043808, + 0.01353475172072649, + 0.01010733749717474, + 0.015083309262990952, + 0.014450198970735073, + 0.06846436113119125, + 0.031194662675261497, + 0.01120765134692192, + 0.014251143671572208, + 0.01006273552775383, + 0.0025487353559583426, + 0.0050422875210642815, + 0.006036911625415087, + 0.005935388151556253, + 0.004951242823153734, + 0.005375833250582218, + 0.002804358024150133, + 0.002835961990058422, + 0.003586831036955118, + 0.0023585984017699957, + 0.002473257016390562, + 0.002441172022372484, + 0.011084085330367088, + 0.0031581795774400234, + 0.012131422758102417, + 0.003913478925824165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.008646313101053238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633799687027931 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633799687027931, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633799687027931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13783812522888184, + 0.10235188156366348, + 0.0681038349866867, + 0.10582636296749115, + 0.046492379158735275, + 0.04918619245290756, + 0.1571003645658493, + 0.21735291182994843, + 0.059557877480983734, + 0.05619007349014282 + ], + "credit_map": [ + 0.015706950798630714, + 0.009679029695689678, + 0.02459469623863697, + 0.01976333186030388, + 0.014338957145810127, + 0.01777644455432892, + 0.016179798170924187, + 0.019798923283815384, + 0.011623313650488853, + 0.010529627092182636, + 0.022411735728383064, + 0.017411472275853157, + 0.008257127366960049, + 0.014245633967220783, + 0.0076059321872889996, + 0.010267040692269802, + 0.007093922235071659, + 0.008750570006668568, + 0.007134601939469576, + 0.011908798478543758, + 0.007758561987429857, + 0.008392696268856525, + 0.010067352093756199, + 0.006997336633503437, + 0.013074866496026516, + 0.006882739719003439, + 0.018121514469385147, + 0.01572328433394432, + 0.015078489668667316, + 0.01785432919859886, + 0.00807951856404543, + 0.011011628434062004, + 0.007861737161874771, + 0.006713450886309147, + 0.007015980314463377, + 0.005321440286934376, + 0.006112548988312483, + 0.005837002769112587, + 0.004067067988216877, + 0.0035631547216326, + 0.0060862828977406025, + 0.005444717593491077, + 0.004811256658285856, + 0.011651547625660896, + 0.006620815489441156, + 0.004608595743775368, + 0.0062867444939911366, + 0.003676231950521469, + 0.016294000670313835, + 0.019674567505717278, + 0.02123241312801838, + 0.02343977428972721, + 0.014908024109899998, + 0.018094023689627647, + 0.017592044547200203, + 0.025865523144602776, + 0.015142729505896568, + 0.01901104301214218, + 0.0194789320230484, + 0.08146361261606216, + 0.039849285036325455, + 0.01377717312425375, + 0.017227906733751297, + 0.011402244679629803, + 0.008295395411550999, + 0.006167690735310316, + 0.009572826325893402, + 0.00808504968881607, + 0.007612451445311308, + 0.00971348024904728, + 0.004776593763381243, + 0.00533439451828599, + 0.007039142772555351, + 0.003643384203314781, + 0.004782288800925016, + 0.0032747259829193354, + 0.013689670711755753, + 0.004733074922114611, + 0.014347975142300129, + 0.004679811652749777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.007633799687027931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011792379664257169 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011792379664257169, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011792379664257169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2197401076555252, + 0.033813975751399994, + 0.030511921271681786, + 0.09576167911291122, + 0.08992254734039307, + 0.13213799893856049, + 0.10353490710258484, + 0.1593034267425537, + 0.04775331914424896, + 0.08752007782459259 + ], + "credit_map": [ + 0.018411343917250633, + 0.013539109379053116, + 0.0417281910777092, + 0.034704286605119705, + 0.020824531093239784, + 0.029624711722135544, + 0.024890277534723282, + 0.03601767495274544, + 0.00393856642767787, + 0.0035432418808341026, + 0.007848160341382027, + 0.0057411822490394115, + 0.0023229941725730896, + 0.0037680224049836397, + 0.0036145274061709642, + 0.0030372794717550278, + 0.00319873820990324, + 0.004727483261376619, + 0.004681025631725788, + 0.004007191397249699, + 0.004178906325250864, + 0.0030147763900458813, + 0.003343388671055436, + 0.003360412083566189, + 0.011871189810335636, + 0.006370121147483587, + 0.015419022180140018, + 0.015952615067362785, + 0.014543824829161167, + 0.01486145704984665, + 0.007113353814929724, + 0.009630097076296806, + 0.009708723053336143, + 0.007836851291358471, + 0.012102698907256126, + 0.007992449216544628, + 0.014949750155210495, + 0.012702731415629387, + 0.013668912462890148, + 0.010960430838167667, + 0.013912375085055828, + 0.011704673059284687, + 0.009169244207441807, + 0.009443073533475399, + 0.03521200269460678, + 0.039283741265535355, + 0.006390810944139957, + 0.007022073958069086, + 0.012354091741144657, + 0.014880582690238953, + 0.012339533306658268, + 0.01179289910942316, + 0.01371820829808712, + 0.009771378710865974, + 0.011032099835574627, + 0.01764610968530178, + 0.007425547111779451, + 0.012813557870686054, + 0.012210763990879059, + 0.06472349911928177, + 0.03218595311045647, + 0.007942555472254753, + 0.012452799826860428, + 0.009548742324113846, + 0.007340752985328436, + 0.007650935556739569, + 0.00731990160420537, + 0.006049798335880041, + 0.005584236700087786, + 0.0055827852338552475, + 0.0035998972598463297, + 0.004625012166798115, + 0.007979664020240307, + 0.0060274857096374035, + 0.0038792311679571867, + 0.00509279640391469, + 0.02273305505514145, + 0.008360262028872967, + 0.02511565014719963, + 0.00833193026483059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0011792379664257169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon with a slender, sinuous body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales pearly and turquoise, and is its mane fiery crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon exhibit specific features including antlers, amber eyes, and long whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style an ink wash painting with black, white, and gray tones, featuring mountains and ancient pines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping aerial shot that slowly pans right following the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568177103996277 + ], + "opposite_logprob_tokens": [ + -3.8955681324005127 + ], + "logprob_reward": -0.020568177103996277, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568177103996277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0813603550195694, + 0.03035196103155613, + 0.026291877031326294, + 0.06278274208307266, + 0.04987237975001335, + 0.17230401933193207, + 0.20795318484306335, + 0.22692394256591797, + 0.09636220335960388, + 0.04579735919833183 + ], + "credit_map": [ + 0.0069563258439302444, + 0.006969030946493149, + 0.013211376033723354, + 0.00886113103479147, + 0.008344619534909725, + 0.009102941490709782, + 0.014994644559919834, + 0.012920287437736988, + 0.0029226604383438826, + 0.002622957807034254, + 0.006446612533181906, + 0.004390032030642033, + 0.0027068087365478277, + 0.005042133387178183, + 0.003196445293724537, + 0.0030243098735809326, + 0.0021574879065155983, + 0.003065126948058605, + 0.0026852497830986977, + 0.0035260342992842197, + 0.005014081485569477, + 0.004187655169516802, + 0.002842030953615904, + 0.0028142095543444157, + 0.007203123066574335, + 0.003954191226512194, + 0.013271010480821133, + 0.009226944297552109, + 0.009026841260492802, + 0.006801223382353783, + 0.005702516529709101, + 0.0075968909077346325, + 0.0074874707497656345, + 0.005761173088103533, + 0.006063759792596102, + 0.004737226292490959, + 0.010105540975928307, + 0.005685736890882254, + 0.005249034147709608, + 0.004782436415553093, + 0.01763586886227131, + 0.014676690101623535, + 0.006244691088795662, + 0.008357161656022072, + 0.04764947667717934, + 0.05884340777993202, + 0.008150805719196796, + 0.010745909065008163, + 0.023943131789565086, + 0.03267677500844002, + 0.029740916565060616, + 0.022723449394106865, + 0.018448010087013245, + 0.02076243795454502, + 0.020920123904943466, + 0.03873834013938904, + 0.015790773555636406, + 0.015354502946138382, + 0.0189819298684597, + 0.07508500665426254, + 0.04566521197557449, + 0.020557550713419914, + 0.021094046533107758, + 0.014394908212125301, + 0.008579902350902557, + 0.007841724902391434, + 0.012718148529529572, + 0.011151328682899475, + 0.011319078505039215, + 0.014388029463589191, + 0.010536258108913898, + 0.01982772722840309, + 0.00801579374819994, + 0.004899461288005114, + 0.005424514878541231, + 0.005046825855970383, + 0.00529019720852375, + 0.004407321568578482, + 0.006736748851835728, + 0.005976493936032057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.020568177103996277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_tensor([5480, 5480])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_tensor([5480, 5480])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9d51a73fcf2e3497168cd51e7c2d637f92679256 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_tensor([5480, 5480])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro, low-angle shot, captured in dramatic, high-speed slow motion, unfolds with a slow, almost imperceptible progression into a monumental conflict. Two black garden ants, their exoskeletons gleaming with an obsidian sheen, engage in a fierce struggle across the vast, polished surface of a black mahogany table. The wood grain beneath them appears like a richly textured, undulating landscape. At the center of their battle is a single, large sugar crystal, a translucent, faceted monolith that refracts light like a precious gem. One ant clings tenaciously to the crystal with its jaws, while the other ant pounces, its own jaws breaking in a desperate attempt to seize the prize. Their spindly legs fumble for purchase against the smooth, reflective wood, their antennae twitching erratically in a flurry of attacks and exchanges. A single, warm light source from above casts long, dramatic shadows, exaggerating their movements and transforming this minor skirmish into an epic duel. The shallow depth of field blurs the distant background into a soft, blurred bokeh, ensuring all focus remains on the raw, silent violence of the scene, creating an atmosphere of intense, microscopic tension.", + "seed": 14260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_5480_seed_14260_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04596562497317791, + "vlm_logprob": -0.04596562497317791, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04596562497317791, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06325174123048782 + ], + "opposite_logprob_tokens": [ + -2.8132517337799072 + ], + "logprob_reward": -0.06325174123048782, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06325174123048782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05507368966937065, + 0.058545321226119995, + 0.06342919915914536, + 0.07522296160459518, + 0.057871170341968536, + 0.038876377046108246, + 0.0294372346252203, + 0.03617021441459656, + 0.5721936821937561, + 0.01318021398037672 + ], + "credit_map": [ + 0.006484647281467915, + 0.005832899361848831, + 0.0079250019043684, + 0.009232798591256142, + 0.006824166979640722, + 0.007019826211035252, + 0.004270348697900772, + 0.007484000641852617, + 0.008566179312765598, + 0.00395830487832427, + 0.008562983945012093, + 0.004967146553099155, + 0.004929973743855953, + 0.006729633081704378, + 0.004582224413752556, + 0.016248876228928566, + 0.01256245281547308, + 0.004456276539713144, + 0.010691337287425995, + 0.007175934035331011, + 0.005853327922523022, + 0.007460122928023338, + 0.005667190533131361, + 0.009562559425830841, + 0.006607093382626772, + 0.007424880284816027, + 0.012030085548758507, + 0.012320071458816528, + 0.011529024690389633, + 0.009827878326177597, + 0.007631927728652954, + 0.007852005772292614, + 0.006778100971132517, + 0.0053848461247980595, + 0.00765257840976119, + 0.008696475997567177, + 0.00662828516215086, + 0.009523926302790642, + 0.006059762090444565, + 0.0071471999399363995, + 0.005257213953882456, + 0.004033716395497322, + 0.004165035672485828, + 0.004033938515931368, + 0.003889086889103055, + 0.007574836257845163, + 0.004284011200070381, + 0.005638538859784603, + 0.0029729336965829134, + 0.0032628607004880905, + 0.0039049736224114895, + 0.005710320547223091, + 0.00214149197563529, + 0.0038480039220303297, + 0.002746834419667721, + 0.0048498171381652355, + 0.0035961733665317297, + 0.004776900168508291, + 0.004489171784371138, + 0.006802073214203119, + 0.003930911421775818, + 0.004940945655107498, + 0.002862624591216445, + 0.004771416075527668, + 0.023966949433088303, + 0.007700350135564804, + 0.011791509576141834, + 0.009277885779738426, + 0.017133178189396858, + 0.25530576705932617, + 0.011418812908232212, + 0.23559918999671936, + 0.0016101038781926036, + 0.0017535841325297952, + 0.0011170050129294395, + 0.0015304774278774858, + 0.0019269502954557538, + 0.001829260610975325, + 0.0014928278978914022, + 0.001920004840940237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.06325174123048782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981209009885788 + ], + "opposite_logprob_tokens": [ + -3.5298120975494385 + ], + "logprob_reward": -0.02981209009885788, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981209009885788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10470733791589737, + 0.07251903414726257, + 0.16225750744342804, + 0.08613787591457367, + 0.03956732526421547, + 0.1089397668838501, + 0.059670496731996536, + 0.0887441411614418, + 0.1655101478099823, + 0.11194635927677155 + ], + "credit_map": [ + 0.01154970470815897, + 0.013061389327049255, + 0.01444239541888237, + 0.015957167372107506, + 0.010177845135331154, + 0.010450191795825958, + 0.01338815689086914, + 0.015680484473705292, + 0.008293808437883854, + 0.007600879296660423, + 0.011662185192108154, + 0.01110154390335083, + 0.007016244810074568, + 0.010852746665477753, + 0.006131632253527641, + 0.009859992191195488, + 0.033138781785964966, + 0.013787519186735153, + 0.019754214212298393, + 0.018768152222037315, + 0.016752757132053375, + 0.02162722684442997, + 0.015234950929880142, + 0.02319389209151268, + 0.008748242631554604, + 0.008433684706687927, + 0.013603568077087402, + 0.012857210822403431, + 0.013922381214797497, + 0.00835734885185957, + 0.010069306008517742, + 0.010146129876375198, + 0.005622235592454672, + 0.0038630624767392874, + 0.005328534636646509, + 0.005368201062083244, + 0.004730928223580122, + 0.005535738077014685, + 0.0035860866773873568, + 0.0055325389839708805, + 0.013239516876637936, + 0.008896966464817524, + 0.011666936799883842, + 0.014621042646467686, + 0.00926405843347311, + 0.019135676324367523, + 0.015646077692508698, + 0.01646948792040348, + 0.00834883563220501, + 0.005000430159270763, + 0.007877781987190247, + 0.010980560444295406, + 0.004850235767662525, + 0.008597531355917454, + 0.005083117634057999, + 0.008932010270655155, + 0.009774209931492805, + 0.00630771042779088, + 0.01684476062655449, + 0.010890310630202293, + 0.008303476497530937, + 0.01627996563911438, + 0.007509161718189716, + 0.01283454243093729, + 0.006963282357901335, + 0.0028495145961642265, + 0.00383313768543303, + 0.005593079142272472, + 0.0053091454319655895, + 0.07036115229129791, + 0.004315287806093693, + 0.06628555804491043, + 0.013204635120928288, + 0.012694817036390305, + 0.010597962886095047, + 0.014904672279953957, + 0.016690345481038094, + 0.015303961001336575, + 0.013142745941877365, + 0.015407215803861618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.02981209009885788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708322703838348 + ], + "opposite_logprob_tokens": [ + -2.1270833015441895 + ], + "logprob_reward": -0.12708322703838348, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708322703838348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07159726321697235, + 0.14216524362564087, + 0.29004231095314026, + 0.0671226754784584, + 0.030909530818462372, + 0.047588031738996506, + 0.030767010524868965, + 0.03769223764538765, + 0.2580299377441406, + 0.02408571168780327 + ], + "credit_map": [ + 0.01035329420119524, + 0.0062816510908305645, + 0.012087852694094181, + 0.009719327092170715, + 0.004511498846113682, + 0.009768672287464142, + 0.008763398043811321, + 0.010111572220921516, + 0.02019629068672657, + 0.010522193275392056, + 0.02651563100516796, + 0.018919672816991806, + 0.01358598843216896, + 0.018883386626839638, + 0.008821387775242329, + 0.02472069300711155, + 0.046237774193286896, + 0.023668672889471054, + 0.04650401324033737, + 0.036064837127923965, + 0.03042745590209961, + 0.034425705671310425, + 0.024853525683283806, + 0.047860343009233475, + 0.00877971388399601, + 0.004561639856547117, + 0.011787042953073978, + 0.011044102720916271, + 0.006971884053200483, + 0.00912428181618452, + 0.004994706250727177, + 0.009859302081167698, + 0.005273646675050259, + 0.002652433468028903, + 0.004126682877540588, + 0.004715241491794586, + 0.0027297392953187227, + 0.003951182588934898, + 0.0032655277755111456, + 0.004195075947791338, + 0.0053156581707298756, + 0.0035568266175687313, + 0.0053740874864161015, + 0.005512184463441372, + 0.004685028921812773, + 0.010087573900818825, + 0.0048653604462742805, + 0.008191311731934547, + 0.004098048433661461, + 0.002878923201933503, + 0.0040634311735630035, + 0.005142940673977137, + 0.0025432133115828037, + 0.0044695246033370495, + 0.002681344049051404, + 0.004889583680778742, + 0.0037547042593359947, + 0.003548145992681384, + 0.00405222550034523, + 0.0062685441225767136, + 0.003904103534296155, + 0.007758581545203924, + 0.003623933531343937, + 0.004781998228281736, + 0.0109014343470335, + 0.004051432013511658, + 0.0062818084843456745, + 0.005161302164196968, + 0.0076369415037333965, + 0.11352626979351044, + 0.00629088981077075, + 0.10417982935905457, + 0.004440321121364832, + 0.001966452691704035, + 0.0023084485437721014, + 0.0035487713757902384, + 0.001935636973939836, + 0.004079609643667936, + 0.0021426251623779535, + 0.0036638458259403706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.12708322703838348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010445620864629745 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010445620864629745, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010445620864629745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10841844975948334, + 0.1684003472328186, + 0.10061107575893402, + 0.10620413720607758, + 0.08404230326414108, + 0.1244281753897667, + 0.06270134449005127, + 0.07391101121902466, + 0.07522141933441162, + 0.0960618257522583 + ], + "credit_map": [ + 0.014158958569169044, + 0.016247965395450592, + 0.017053402960300446, + 0.016402075067162514, + 0.009334234520792961, + 0.009037839248776436, + 0.007903637364506721, + 0.01828031986951828, + 0.03018230013549328, + 0.016102822497487068, + 0.030842095613479614, + 0.0289246067404747, + 0.012242160737514496, + 0.016952235251665115, + 0.012333832681179047, + 0.02082027494907379, + 0.017438998445868492, + 0.009427896700799465, + 0.015159817412495613, + 0.015888575464487076, + 0.010106341913342476, + 0.010115781798958778, + 0.009372394531965256, + 0.013101257383823395, + 0.014007815159857273, + 0.009614956565201283, + 0.019701924175024033, + 0.016580931842327118, + 0.013207525946199894, + 0.01299765519797802, + 0.00937228836119175, + 0.010721031576395035, + 0.0139856468886137, + 0.006649120710790157, + 0.014992884360253811, + 0.008292232640087605, + 0.006715237628668547, + 0.015321111306548119, + 0.006159970071166754, + 0.011926092207431793, + 0.012682352215051651, + 0.015329606831073761, + 0.01743040233850479, + 0.018539775162935257, + 0.011795993894338608, + 0.01563410274684429, + 0.014575972221791744, + 0.018439970910549164, + 0.00890257302671671, + 0.006749648600816727, + 0.0077407220378518105, + 0.009691404178738594, + 0.0072753410786390305, + 0.008832475170493126, + 0.00478009320795536, + 0.00872908066958189, + 0.007154022343456745, + 0.009751877747476101, + 0.008670281618833542, + 0.011052575893700123, + 0.009900623932480812, + 0.012048506177961826, + 0.0061843786388635635, + 0.009148739278316498, + 0.006708711851388216, + 0.004238420631736517, + 0.005415618885308504, + 0.009283103980123997, + 0.009011671878397465, + 0.0164644755423069, + 0.008323587477207184, + 0.01577582210302353, + 0.009196036495268345, + 0.01275838166475296, + 0.009191635064780712, + 0.014819931238889694, + 0.013484615832567215, + 0.014043193310499191, + 0.009920937940478325, + 0.012647089548408985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0010445620864629745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636504411697388 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636504411697388, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636504411697388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09176257252693176, + 0.03328297659754753, + 0.07578577101230621, + 0.1597892940044403, + 0.11457054316997528, + 0.18873566389083862, + 0.017063315957784653, + 0.06350594758987427, + 0.21995587646961212, + 0.03554802015423775 + ], + "credit_map": [ + 0.010413185693323612, + 0.01108582690358162, + 0.012121419422328472, + 0.012823181226849556, + 0.009139439091086388, + 0.00964420847594738, + 0.012276317924261093, + 0.014258991926908493, + 0.0043585882522165775, + 0.0029902360402047634, + 0.005729203578084707, + 0.00532711623236537, + 0.0028855055570602417, + 0.0042281742207705975, + 0.002838491229340434, + 0.004925659392029047, + 0.015606389380991459, + 0.005682302638888359, + 0.010961931198835373, + 0.008336076512932777, + 0.0068406593054533005, + 0.009892402216792107, + 0.006865821313112974, + 0.011600184254348278, + 0.018824346363544464, + 0.014006851240992546, + 0.0262282844632864, + 0.028169061988592148, + 0.018524127081036568, + 0.021666454151272774, + 0.013241192325949669, + 0.019128983840346336, + 0.012412839569151402, + 0.012097219005227089, + 0.013253217563033104, + 0.017203737050294876, + 0.009907602332532406, + 0.02101973444223404, + 0.013458101078867912, + 0.015218093991279602, + 0.02004285715520382, + 0.020321110263466835, + 0.026677709072828293, + 0.024822058156132698, + 0.02211732044816017, + 0.029425647109746933, + 0.01578374393284321, + 0.02954520657658577, + 0.0019228089367970824, + 0.0014274630229920149, + 0.002217469271272421, + 0.00359557569026947, + 0.0014158807462081313, + 0.0022311697248369455, + 0.0016345166368409991, + 0.00261843204498291, + 0.004639307036995888, + 0.008543509058654308, + 0.00783082190901041, + 0.012208838015794754, + 0.007914912886917591, + 0.007666856981813908, + 0.006004556082189083, + 0.00869714468717575, + 0.009366319514811039, + 0.003436514176428318, + 0.004850418772548437, + 0.0049958243034780025, + 0.007020398508757353, + 0.09667141735553741, + 0.004774495959281921, + 0.08884048461914062, + 0.004459945019334555, + 0.005048426799476147, + 0.0032322551123797894, + 0.003956349566578865, + 0.005060866940766573, + 0.005183984059840441, + 0.004119679797440767, + 0.0044865114614367485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.008636504411697388 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro, low-angle shot, captured in dramatic, high-speed slow motion, unfolds with a slow, almost imperceptible progression into a monumental conflict. Two black garden ants, their exoskeletons gleaming with an obsidian sheen, engage in a fierce struggle across the vast, polished surface of a black mahogany table. The wood grain beneath them appears like a richly textured, undulating landscape. At the center of their battle is a single, large sugar crystal, a translucent, faceted monolith that refracts light like a precious gem. One ant clings tenaciously to the crystal with its jaws, while the other ant pounces, its own jaws breaking in a desperate attempt to seize the prize. Their spindly legs fumble for purchase against the smooth, reflective wood, their antennae twitching erratically in a flurry of attacks and exchanges. A single, warm light source from above casts long, dramatic shadows, exaggerating their movements and transforming this minor skirmish into an epic duel. The shallow depth of field blurs the distant background into a soft, blurred bokeh, ensuring all focus remains on the raw, silent violence of the scene, creating an atmosphere of intense, microscopic tension.", + "seed": 14261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_5480_seed_14261_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023258457612246274, + "vlm_logprob": -0.023258457612246274, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.023258457612246274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03881906718015671 + ], + "opposite_logprob_tokens": [ + -3.2888190746307373 + ], + "logprob_reward": -0.03881906718015671, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03881906718015671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4268254339694977, + 0.060409899801015854, + 0.02612888626754284, + 0.09204620122909546, + 0.02859598957002163, + 0.06431408226490021, + 0.031488340348005295, + 0.1539667248725891, + 0.07468994706869125, + 0.041534509509801865 + ], + "credit_map": [ + 0.14068327844142914, + 0.023432044312357903, + 0.06351622194051743, + 0.01999359205365181, + 0.016643857583403587, + 0.09700559079647064, + 0.02661820501089096, + 0.03893262520432472, + 0.009137026965618134, + 0.005458300467580557, + 0.010393134318292141, + 0.008814731612801552, + 0.005075068678706884, + 0.006746921222656965, + 0.00795526709407568, + 0.0068294513039290905, + 0.003739400999620557, + 0.0027980634476989508, + 0.003019845811650157, + 0.0032154719810932875, + 0.003659491427242756, + 0.0028972947038710117, + 0.004464233294129372, + 0.0023350862320512533, + 0.010879859328269958, + 0.008356617763638496, + 0.019439632073044777, + 0.012632235884666443, + 0.012841393239796162, + 0.0100787952542305, + 0.008493076078593731, + 0.009324587881565094, + 0.004915011581033468, + 0.003174091223627329, + 0.002777478424832225, + 0.0038908496499061584, + 0.004940805025398731, + 0.00300653837621212, + 0.0024186025839298964, + 0.0034726127050817013, + 0.01015594694763422, + 0.009207312017679214, + 0.005975928623229265, + 0.009025483392179012, + 0.00782038550823927, + 0.0073672072030603886, + 0.005450512748211622, + 0.009311303496360779, + 0.004168778192251921, + 0.0038671186193823814, + 0.0037489624228328466, + 0.004333715420216322, + 0.0023531874176114798, + 0.004205602221190929, + 0.0033364964183419943, + 0.005474478006362915, + 0.01188863255083561, + 0.011558067984879017, + 0.007901852019131184, + 0.007708712015300989, + 0.005444867070764303, + 0.060983289033174515, + 0.0051368484273552895, + 0.043344441801309586, + 0.007404955103993416, + 0.011500746943056583, + 0.009942861273884773, + 0.010378922335803509, + 0.008895345963537693, + 0.008004383184015751, + 0.008999022655189037, + 0.009563712403178215, + 0.006355670280754566, + 0.0061979531310498714, + 0.005632523447275162, + 0.002829735167324543, + 0.004519477020949125, + 0.006502973847091198, + 0.00372494850307703, + 0.005771228577941656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.03881906718015671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290986195206642 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.023290986195206642, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290986195206642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25090456008911133, + 0.051608070731163025, + 0.0380159392952919, + 0.14935646951198578, + 0.02915443666279316, + 0.2237950712442398, + 0.022343087941408157, + 0.061216361820697784, + 0.08125951141119003, + 0.09234647452831268 + ], + "credit_map": [ + 0.07688869535923004, + 0.014050468802452087, + 0.029186269268393517, + 0.010601622983813286, + 0.010305897332727909, + 0.07002129405736923, + 0.015917018055915833, + 0.0239332914352417, + 0.007351248059421778, + 0.005400278605520725, + 0.009799267165362835, + 0.007234165910631418, + 0.005256528500467539, + 0.004475605208426714, + 0.0061858417466282845, + 0.005905131809413433, + 0.005020358599722385, + 0.003913945518434048, + 0.0055558993481099606, + 0.007241179700940847, + 0.0036627757363021374, + 0.003995066042989492, + 0.003795844968408346, + 0.004830868914723396, + 0.018779993057250977, + 0.014918268658220768, + 0.0322413295507431, + 0.022054143249988556, + 0.018766866996884346, + 0.016798468306660652, + 0.011720037087798119, + 0.014077362604439259, + 0.004190526902675629, + 0.0036187598016113043, + 0.00285202800296247, + 0.005629683844745159, + 0.0038215129170566797, + 0.0026350384578108788, + 0.0023057619109749794, + 0.004101123660802841, + 0.033797040581703186, + 0.027532028034329414, + 0.023262185975909233, + 0.036498818546533585, + 0.022575095295906067, + 0.03150683268904686, + 0.017580408602952957, + 0.031042661517858505, + 0.003357604146003723, + 0.003510073758661747, + 0.0023852891754359007, + 0.0026147272437810898, + 0.0019121677614748478, + 0.0034348280169069767, + 0.002551463432610035, + 0.002576933242380619, + 0.01033820677548647, + 0.011453301645815372, + 0.007775116246193647, + 0.007857384160161018, + 0.005361965391784906, + 0.006398516241461039, + 0.005572337657213211, + 0.006459535099565983, + 0.01539773028343916, + 0.008901416324079037, + 0.009094496257603168, + 0.00784593727439642, + 0.009756870567798615, + 0.012155611999332905, + 0.008287027478218079, + 0.009820421226322651, + 0.00919535756111145, + 0.008734012953937054, + 0.005812038667500019, + 0.006341042462736368, + 0.00980521459132433, + 0.02846124768257141, + 0.006190894171595573, + 0.01780666597187519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.023290986195206642 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430605448782444 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.0430605448782444, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430605448782444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5139042139053345, + 0.09768083691596985, + 0.059064436703920364, + 0.037397995591163635, + 0.017908411100506783, + 0.03365236893296242, + 0.034225206822156906, + 0.054906368255615234, + 0.1128012016415596, + 0.03845900669693947 + ], + "credit_map": [ + 0.11382979899644852, + 0.03126148879528046, + 0.07879041880369186, + 0.02426605112850666, + 0.018795883283019066, + 0.17723087966442108, + 0.03138337284326553, + 0.03834625333547592, + 0.014504332095384598, + 0.008737520314753056, + 0.017952710390090942, + 0.017873592674732208, + 0.007902759127318859, + 0.008422808721661568, + 0.010789453983306885, + 0.011497645638883114, + 0.010561255738139153, + 0.008218005299568176, + 0.007015019189566374, + 0.010435444302856922, + 0.005281819961965084, + 0.005753623321652412, + 0.0046014501713216305, + 0.007197814527899027, + 0.004815298598259687, + 0.0033748496789485216, + 0.007952745072543621, + 0.005559854209423065, + 0.0046706791035830975, + 0.004162989556789398, + 0.002907675923779607, + 0.003953899722546339, + 0.0024326785933226347, + 0.001788801746442914, + 0.0035986669827252626, + 0.002439970849081874, + 0.0018981265602633357, + 0.0019526685355231166, + 0.0014799506170675159, + 0.002317545237019658, + 0.0061094206757843494, + 0.004565429408103228, + 0.003722019027918577, + 0.0036533831153064966, + 0.00378675595857203, + 0.004210761282593012, + 0.002430942375212908, + 0.005173656158149242, + 0.003533422714099288, + 0.0033746997360140085, + 0.0029026793781667948, + 0.0053368546068668365, + 0.002189245307818055, + 0.006547839846462011, + 0.003536262782290578, + 0.006804195232689381, + 0.005362433381378651, + 0.004323282744735479, + 0.002937670797109604, + 0.0035626350436359644, + 0.0021914609242230654, + 0.02008739858865738, + 0.002357618184760213, + 0.014083861373364925, + 0.015954572707414627, + 0.010433738119900227, + 0.011715002357959747, + 0.014486037194728851, + 0.012827485799789429, + 0.020279856398701668, + 0.010896666906774044, + 0.016207832843065262, + 0.004107003565877676, + 0.0033501088619232178, + 0.003501529572531581, + 0.0038470488507300615, + 0.0030514781828969717, + 0.0111655592918396, + 0.0028249749448150396, + 0.00661130016669631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0430605448782444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488494385033846 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.002488494385033846, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488494385033846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22680440545082092, + 0.04417707771062851, + 0.03557270020246506, + 0.08386137336492538, + 0.03210752084851265, + 0.12080618739128113, + 0.019713787361979485, + 0.2326992005109787, + 0.15189668536186218, + 0.05236108973622322 + ], + "credit_map": [ + 0.04697226360440254, + 0.014761467464268208, + 0.034031711518764496, + 0.015633758157491684, + 0.010552313178777695, + 0.07090962678194046, + 0.014207308180630207, + 0.019735950976610184, + 0.008262851275503635, + 0.0036256953608244658, + 0.007443603593856096, + 0.005155086051672697, + 0.004136109258979559, + 0.004156817216426134, + 0.004990688059478998, + 0.006406229920685291, + 0.0047225868329405785, + 0.006394347175955772, + 0.00419568782672286, + 0.0043440405279397964, + 0.004406875465065241, + 0.004094568081200123, + 0.003632666077464819, + 0.0037819305434823036, + 0.010385556146502495, + 0.00911177508533001, + 0.016286954283714294, + 0.012794443406164646, + 0.011761238798499107, + 0.008992314338684082, + 0.006272003520280123, + 0.008257082663476467, + 0.003944133874028921, + 0.002866818569600582, + 0.005051811691373587, + 0.003561238059774041, + 0.0031810656655579805, + 0.004004031885415316, + 0.002524829003959894, + 0.006973593030124903, + 0.022587427869439125, + 0.01619989052414894, + 0.011303641833364964, + 0.013798044994473457, + 0.012294137850403786, + 0.017079154029488564, + 0.009689031168818474, + 0.017854858189821243, + 0.002590005751699209, + 0.0016448243986815214, + 0.002115668496116996, + 0.0034843888133764267, + 0.0014280249597504735, + 0.003485215362161398, + 0.0022978687193244696, + 0.002667791908606887, + 0.018731171265244484, + 0.009725249372422695, + 0.009359524585306644, + 0.005346570163965225, + 0.006018050946295261, + 0.10398529469966888, + 0.0052396394312381744, + 0.07429368793964386, + 0.02671029046177864, + 0.01169846672564745, + 0.01617736555635929, + 0.01713082753121853, + 0.014864269644021988, + 0.02946041151881218, + 0.014250353910028934, + 0.021604711189866066, + 0.00822767149657011, + 0.005614608526229858, + 0.005267590284347534, + 0.00552326999604702, + 0.004395411349833012, + 0.011612779460847378, + 0.004541593603789806, + 0.007178164552897215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.002488494385033846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633195422589779 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008633195422589779, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633195422589779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28290799260139465, + 0.05375472083687782, + 0.03584909066557884, + 0.0524713359773159, + 0.04875234514474869, + 0.12061577290296555, + 0.1210121139883995, + 0.1702321469783783, + 0.08505108952522278, + 0.02935340814292431 + ], + "credit_map": [ + 0.09489524364471436, + 0.013859794475138187, + 0.028401995077729225, + 0.014317508786916733, + 0.009632413275539875, + 0.08736856281757355, + 0.01279052160680294, + 0.021641956642270088, + 0.00718160392716527, + 0.004391882102936506, + 0.005384872201830149, + 0.00773831969127059, + 0.006135392002761364, + 0.0071966578252613544, + 0.007096630521118641, + 0.008629360236227512, + 0.0045991381630301476, + 0.004092144314199686, + 0.006177418399602175, + 0.004804752767086029, + 0.0034065917134284973, + 0.004284916445612907, + 0.0031924329232424498, + 0.005291697569191456, + 0.005872749723494053, + 0.00737271411344409, + 0.0060208458453416824, + 0.006593035534024239, + 0.010797361843287945, + 0.005121143534779549, + 0.00442209467291832, + 0.006271389778703451, + 0.007032502442598343, + 0.0067887455224990845, + 0.00465735886245966, + 0.0059395963326096535, + 0.007366753648966551, + 0.0049369302578270435, + 0.0031899032182991505, + 0.008840554393827915, + 0.02266247570514679, + 0.01436090748757124, + 0.010841002687811852, + 0.015385102480649948, + 0.011259558610618114, + 0.017935093492269516, + 0.009548923000693321, + 0.018622705712914467, + 0.012366523034870625, + 0.011362476274371147, + 0.010543327778577805, + 0.01512073166668415, + 0.0064543685875833035, + 0.02658945508301258, + 0.00946248508989811, + 0.029112741351127625, + 0.013362076133489609, + 0.009929497726261616, + 0.007696084678173065, + 0.006813975051045418, + 0.005447562783956528, + 0.07154106348752975, + 0.004466993268579245, + 0.05097490921616554, + 0.012626631185412407, + 0.010207416489720345, + 0.009477474726736546, + 0.012708630412817001, + 0.00842893123626709, + 0.013253615237772465, + 0.008317546918988228, + 0.010030841454863548, + 0.0042947810143232346, + 0.0029980407562106848, + 0.004657526966184378, + 0.004393471404910088, + 0.003061422612518072, + 0.0033635597210377455, + 0.00293768011033535, + 0.003646926488727331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008633195422589779 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro, low-angle shot, captured in dramatic, high-speed slow motion, unfolds with a slow, almost imperceptible progression into a monumental conflict. Two black garden ants, their exoskeletons gleaming with an obsidian sheen, engage in a fierce struggle across the vast, polished surface of a black mahogany table. The wood grain beneath them appears like a richly textured, undulating landscape. At the center of their battle is a single, large sugar crystal, a translucent, faceted monolith that refracts light like a precious gem. One ant clings tenaciously to the crystal with its jaws, while the other ant pounces, its own jaws breaking in a desperate attempt to seize the prize. Their spindly legs fumble for purchase against the smooth, reflective wood, their antennae twitching erratically in a flurry of attacks and exchanges. A single, warm light source from above casts long, dramatic shadows, exaggerating their movements and transforming this minor skirmish into an epic duel. The shallow depth of field blurs the distant background into a soft, blurred bokeh, ensuring all focus remains on the raw, silent violence of the scene, creating an atmosphere of intense, microscopic tension.", + "seed": 14262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_5480_seed_14262_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03220935701392591, + "vlm_logprob": -0.03220935701392591, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03220935701392591, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06338690966367722 + ], + "opposite_logprob_tokens": [ + -2.813386917114258 + ], + "logprob_reward": -0.06338690966367722, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06338690966367722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18507061898708344, + 0.09080491960048676, + 0.04977453872561455, + 0.09156499803066254, + 0.09758096188306808, + 0.08740220218896866, + 0.04892413318157196, + 0.09964769333600998, + 0.16026361286640167, + 0.08896628022193909 + ], + "credit_map": [ + 0.021240292116999626, + 0.02263006381690502, + 0.024959588423371315, + 0.027155805379152298, + 0.018110910430550575, + 0.019795900210738182, + 0.02161220647394657, + 0.029565855860710144, + 0.007866576313972473, + 0.009535708464682102, + 0.019661758095026016, + 0.012277735397219658, + 0.011231260374188423, + 0.008671947754919529, + 0.009928440675139427, + 0.011631494387984276, + 0.004628696944564581, + 0.0042668553069233894, + 0.007675168104469776, + 0.01133787352591753, + 0.0033825384452939034, + 0.005673985928297043, + 0.006520124152302742, + 0.006289297249168158, + 0.012295374646782875, + 0.011238284409046173, + 0.011996896006166935, + 0.01336356159299612, + 0.0075905947014689445, + 0.009689833968877792, + 0.011540004052221775, + 0.013850448653101921, + 0.014089466072618961, + 0.010379514656960964, + 0.01672409474849701, + 0.012731090188026428, + 0.008934969082474709, + 0.01004056166857481, + 0.012839917093515396, + 0.01184134092181921, + 0.00977056659758091, + 0.008720777928829193, + 0.014865226112306118, + 0.008133810013532639, + 0.008666102774441242, + 0.010370210744440556, + 0.009008345194160938, + 0.017867157235741615, + 0.008502830751240253, + 0.004282824229449034, + 0.006745146121829748, + 0.005897900555282831, + 0.00474357046186924, + 0.003989964257925749, + 0.006185671780258417, + 0.0085762282833457, + 0.011919034644961357, + 0.013492552563548088, + 0.01285746693611145, + 0.01670784130692482, + 0.014968660660088062, + 0.007384058553725481, + 0.013604599051177502, + 0.008713489398360252, + 0.019112441688776016, + 0.025023015215992928, + 0.020255310460925102, + 0.02158195897936821, + 0.029137810692191124, + 0.01325332559645176, + 0.016279257833957672, + 0.015620486810803413, + 0.01274423860013485, + 0.007968511432409286, + 0.011342027224600315, + 0.013819999061524868, + 0.006901070475578308, + 0.00731859402731061, + 0.010877028107643127, + 0.01799480989575386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.06338690966367722 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807578772306442 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.029807578772306442, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807578772306442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20172546803951263, + 0.1104179322719574, + 0.12151572108268738, + 0.030419940128922462, + 0.19199685752391815, + 0.034415073692798615, + 0.14283083379268646, + 0.09681153297424316, + 0.026101917028427124, + 0.043764710426330566 + ], + "credit_map": [ + 0.016478773206472397, + 0.02091245725750923, + 0.04725118353962898, + 0.030835095793008804, + 0.015088826417922974, + 0.02049417607486248, + 0.02527705579996109, + 0.025387883186340332, + 0.011785263195633888, + 0.006969884037971497, + 0.034222424030303955, + 0.013261646963655949, + 0.009839383885264397, + 0.00668897433206439, + 0.0198424831032753, + 0.007807873655110598, + 0.012036415748298168, + 0.008412810042500496, + 0.029165979474782944, + 0.023550793528556824, + 0.006418602541089058, + 0.011844456195831299, + 0.011003851890563965, + 0.0190828088670969, + 0.0038878482300788164, + 0.0026469011791050434, + 0.00558526162058115, + 0.0042836652137339115, + 0.0029208341147750616, + 0.0029197754338383675, + 0.0034129119012504816, + 0.004762741737067699, + 0.025522558018565178, + 0.020039545372128487, + 0.03143031895160675, + 0.031102264299988747, + 0.016694730147719383, + 0.019882820546627045, + 0.022755317389965057, + 0.024569310247898102, + 0.005598396062850952, + 0.002900920109823346, + 0.006727139465510845, + 0.003416521940380335, + 0.002433612709864974, + 0.003427316667512059, + 0.003124430077150464, + 0.006786735262721777, + 0.023321036249399185, + 0.013489198870956898, + 0.02103373408317566, + 0.020131660625338554, + 0.015864364802837372, + 0.011349932290613651, + 0.02024025470018387, + 0.017400655895471573, + 0.011743564158678055, + 0.01199714932590723, + 0.012261571362614632, + 0.017697326838970184, + 0.011252510361373425, + 0.0070184264332056046, + 0.014570963568985462, + 0.010270020924508572, + 0.0026683311443775892, + 0.0043075052089989185, + 0.0035960350651293993, + 0.004528291989117861, + 0.003058101050555706, + 0.0018206773092970252, + 0.002701748162508011, + 0.0034212267491966486, + 0.006071920040994883, + 0.00412005977705121, + 0.006294052116572857, + 0.006164880003780127, + 0.0035672863014042377, + 0.005026750732213259, + 0.0047548371367156506, + 0.00776492478325963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.029807578772306442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204402819275856 + ], + "opposite_logprob_tokens": [ + -2.812044143676758 + ], + "logprob_reward": -0.06204402819275856, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204402819275856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18887342512607574, + 0.23906350135803223, + 0.09597320109605789, + 0.06362099200487137, + 0.2253335565328598, + 0.03769761696457863, + 0.03793436288833618, + 0.050729233771562576, + 0.024591200053691864, + 0.03618289902806282 + ], + "credit_map": [ + 0.01633947156369686, + 0.014115645550191402, + 0.04729350283741951, + 0.01975744590163231, + 0.012836473062634468, + 0.015627529472112656, + 0.014352332800626755, + 0.04855101555585861, + 0.021573491394519806, + 0.01715211570262909, + 0.045099493116140366, + 0.0473567396402359, + 0.02278575673699379, + 0.02404814586043358, + 0.031168529763817787, + 0.029879223555326462, + 0.009512591175734997, + 0.00805701408535242, + 0.015486927703022957, + 0.022675184532999992, + 0.00661319587379694, + 0.008315011858940125, + 0.013512488454580307, + 0.011800780892372131, + 0.006894487887620926, + 0.004958773497492075, + 0.008296127431094646, + 0.008609689772129059, + 0.007205089554190636, + 0.007989531382918358, + 0.008720071986317635, + 0.0109472144395113, + 0.02401047758758068, + 0.020861558616161346, + 0.038969479501247406, + 0.04485238343477249, + 0.01476900652050972, + 0.02394808642566204, + 0.02544693648815155, + 0.03247561678290367, + 0.0069378637708723545, + 0.002912323921918869, + 0.006566604133695364, + 0.004754416644573212, + 0.001793199684470892, + 0.0029260602314025164, + 0.002899995306506753, + 0.008907154202461243, + 0.008176414296030998, + 0.0033105879556387663, + 0.0055310409516096115, + 0.006595163606107235, + 0.0028554105665534735, + 0.0029842585790902376, + 0.002729907166212797, + 0.005751578137278557, + 0.007356888148933649, + 0.005099500995129347, + 0.007037025410681963, + 0.00579441525042057, + 0.00619537103921175, + 0.005323563236743212, + 0.007108911406248808, + 0.00681355781853199, + 0.004285416565835476, + 0.004356442019343376, + 0.0029223570600152016, + 0.0032593270298093557, + 0.0029516336508095264, + 0.0019250566838309169, + 0.0016707411268725991, + 0.003220226615667343, + 0.0035705137997865677, + 0.0038722066674381495, + 0.0047480640932917595, + 0.0055784923024475574, + 0.002913042902946472, + 0.0043976218439638615, + 0.0050940243527293205, + 0.0060089328326284885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.06204402819275856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017151420470327139 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.0017151420470327139, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017151420470327139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.175643190741539, + 0.08999493718147278, + 0.08740238845348358, + 0.03872248902916908, + 0.28474393486976624, + 0.038486547768116, + 0.0902223065495491, + 0.12865574657917023, + 0.026684710755944252, + 0.039443742483854294 + ], + "credit_map": [ + 0.015346115455031395, + 0.01339124608784914, + 0.04628899320960045, + 0.021343473345041275, + 0.012022485956549644, + 0.015072992071509361, + 0.014597839675843716, + 0.03758004680275917, + 0.012085124850273132, + 0.006298705004155636, + 0.012728036381304264, + 0.013842998072504997, + 0.00781688466668129, + 0.0072124251164495945, + 0.017275825142860413, + 0.012734930031001568, + 0.009449340403079987, + 0.006883656606078148, + 0.01585521548986435, + 0.019811198115348816, + 0.004914778284728527, + 0.008563864044845104, + 0.010021389462053776, + 0.011902951635420322, + 0.0031466586515307426, + 0.004194795619696379, + 0.005620351526886225, + 0.004552751313894987, + 0.004371715243905783, + 0.00461222929880023, + 0.006157311610877514, + 0.006066678091883659, + 0.05997584015130997, + 0.01497636642307043, + 0.06277439743280411, + 0.024551432579755783, + 0.014362075366079807, + 0.02129078470170498, + 0.026597581803798676, + 0.06021544709801674, + 0.009133727289736271, + 0.003680827561765909, + 0.005614257417619228, + 0.004694325849413872, + 0.003044324228540063, + 0.0028472752310335636, + 0.0034033118281513453, + 0.00606849929317832, + 0.02378302440047264, + 0.007032796274870634, + 0.01623344235122204, + 0.009614729322493076, + 0.006601929664611816, + 0.007794124539941549, + 0.008514104411005974, + 0.010648153722286224, + 0.014627711847424507, + 0.010319562628865242, + 0.017871730029582977, + 0.02163238823413849, + 0.011572761461138725, + 0.009686259552836418, + 0.02605009265244007, + 0.01689523085951805, + 0.004358196631073952, + 0.0029588902834802866, + 0.003784256987273693, + 0.0045342217199504375, + 0.002119173528626561, + 0.0018903549062088132, + 0.0027099004946649075, + 0.004329714924097061, + 0.003844842780381441, + 0.003081435104832053, + 0.006179451942443848, + 0.005302410572767258, + 0.0031685803551226854, + 0.004221419338136911, + 0.004624888766556978, + 0.009020711295306683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0017151420470327139 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093126393854618 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093126393854618, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093126393854618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20315782725811005, + 0.11986836791038513, + 0.057103633880615234, + 0.14204654097557068, + 0.04146607220172882, + 0.16748195886611938, + 0.10828349739313126, + 0.05116822570562363, + 0.03981376439332962, + 0.0696101039648056 + ], + "credit_map": [ + 0.016629425808787346, + 0.019773606210947037, + 0.027584942057728767, + 0.036505717784166336, + 0.016894405707716942, + 0.021263064816594124, + 0.03875133767724037, + 0.025755316019058228, + 0.01052701286971569, + 0.008500706404447556, + 0.031468991190195084, + 0.017907140776515007, + 0.00847211480140686, + 0.00954222958534956, + 0.01978888548910618, + 0.013661293312907219, + 0.006893916521221399, + 0.0038152120541781187, + 0.016224173828959465, + 0.007749979849904776, + 0.003494091797620058, + 0.0046415505930781364, + 0.004545119591057301, + 0.009739595465362072, + 0.0174007136374712, + 0.011478608474135399, + 0.019202325493097305, + 0.01867351494729519, + 0.014233144000172615, + 0.015964441001415253, + 0.0194100234657526, + 0.025683758780360222, + 0.006694038398563862, + 0.003925294149667025, + 0.007589139509946108, + 0.005267433822154999, + 0.0031756083481013775, + 0.004739405587315559, + 0.004509462509304285, + 0.005565688945353031, + 0.032192785292863846, + 0.015759270638227463, + 0.02579384297132492, + 0.02112443745136261, + 0.012114414013922215, + 0.0159098319709301, + 0.014344019815325737, + 0.030243344604969025, + 0.012071207165718079, + 0.010319444350898266, + 0.016646018251776695, + 0.018955696374177933, + 0.012870017439126968, + 0.009914140217006207, + 0.011341861449182034, + 0.0161651112139225, + 0.01015190128237009, + 0.0050257230177521706, + 0.007685912307351828, + 0.005299826618283987, + 0.007873589172959328, + 0.005027915816754103, + 0.004235789645463228, + 0.005867570172995329, + 0.0036381669342517853, + 0.005601648706942797, + 0.005310033913701773, + 0.00701528275385499, + 0.0059161726385355, + 0.0026402228977531195, + 0.004199802875518799, + 0.005492434371262789, + 0.006406639236956835, + 0.007736670318990946, + 0.011373493820428848, + 0.008397245779633522, + 0.007586285937577486, + 0.00679459422826767, + 0.007544416002929211, + 0.013770760968327522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.004093126393854618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro, low-angle shot, captured in dramatic, high-speed slow motion, unfolds with a slow, almost imperceptible progression into a monumental conflict. Two black garden ants, their exoskeletons gleaming with an obsidian sheen, engage in a fierce struggle across the vast, polished surface of a black mahogany table. The wood grain beneath them appears like a richly textured, undulating landscape. At the center of their battle is a single, large sugar crystal, a translucent, faceted monolith that refracts light like a precious gem. One ant clings tenaciously to the crystal with its jaws, while the other ant pounces, its own jaws breaking in a desperate attempt to seize the prize. Their spindly legs fumble for purchase against the smooth, reflective wood, their antennae twitching erratically in a flurry of attacks and exchanges. A single, warm light source from above casts long, dramatic shadows, exaggerating their movements and transforming this minor skirmish into an epic duel. The shallow depth of field blurs the distant background into a soft, blurred bokeh, ensuring all focus remains on the raw, silent violence of the scene, creating an atmosphere of intense, microscopic tension.", + "seed": 14263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_15/index_5480_seed_14263_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027938129333779215, + "vlm_logprob": -0.027938129333779215, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.027938129333779215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019728509709239006 + ], + "opposite_logprob_tokens": [ + -4.019728660583496 + ], + "logprob_reward": -0.019728509709239006, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019728509709239006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14928202331066132, + 0.12138354033231735, + 0.07235276699066162, + 0.08167186379432678, + 0.09345241636037827, + 0.09905034303665161, + 0.14872431755065918, + 0.037536151707172394, + 0.05735776945948601, + 0.13918879628181458 + ], + "credit_map": [ + 0.013635637238621712, + 0.01653154566884041, + 0.0240195170044899, + 0.018396515399217606, + 0.027420787140727043, + 0.01701398380100727, + 0.018784264102578163, + 0.013479765504598618, + 0.01628459244966507, + 0.014365450479090214, + 0.019106073305010796, + 0.017451344057917595, + 0.014315254054963589, + 0.014523361809551716, + 0.013479670509696007, + 0.011857789941132069, + 0.009204999543726444, + 0.012994186021387577, + 0.012816152535378933, + 0.006206553429365158, + 0.00789918377995491, + 0.012272026389837265, + 0.005362257827073336, + 0.005597404669970274, + 0.011088174767792225, + 0.009887426160275936, + 0.014849673956632614, + 0.010937034152448177, + 0.005609110463410616, + 0.015615034848451614, + 0.006389133632183075, + 0.00729628000408411, + 0.013776007108390331, + 0.011590162292122841, + 0.013077795505523682, + 0.014329612255096436, + 0.007658494636416435, + 0.008976236917078495, + 0.008532206527888775, + 0.015511900186538696, + 0.017517780885100365, + 0.011205987073481083, + 0.01143865566700697, + 0.015102684497833252, + 0.008724004961550236, + 0.01736621931195259, + 0.009015467949211597, + 0.008679543621838093, + 0.019861051812767982, + 0.0181844811886549, + 0.018557140603661537, + 0.026238994672894478, + 0.013491068966686726, + 0.02556535229086876, + 0.01135568879544735, + 0.015470530837774277, + 0.006404676008969545, + 0.0039107296615839005, + 0.007391276303678751, + 0.004813941661268473, + 0.0021046127658337355, + 0.0059939101338386536, + 0.004064058884978294, + 0.002852947684004903, + 0.008409393951296806, + 0.005108513869345188, + 0.007201316766440868, + 0.006260746158659458, + 0.009458716958761215, + 0.006837367545813322, + 0.008324855007231236, + 0.005756861064583063, + 0.029333829879760742, + 0.015333048067986965, + 0.022832849994301796, + 0.013235282152891159, + 0.01561951544135809, + 0.011859871447086334, + 0.012078283354640007, + 0.01889611780643463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.019728509709239006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640897762030363 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007640897762030363, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640897762030363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09373124688863754, + 0.35982009768486023, + 0.04916929081082344, + 0.13036368787288666, + 0.08262070268392563, + 0.05200857296586037, + 0.03278742730617523, + 0.017750181257724762, + 0.12611623108386993, + 0.05563250184059143 + ], + "credit_map": [ + 0.007812796160578728, + 0.01188272051513195, + 0.011879236437380314, + 0.012966502457857132, + 0.011132393963634968, + 0.015793969854712486, + 0.011180786415934563, + 0.011082853190600872, + 0.04666890203952789, + 0.030128080397844315, + 0.046558406203985214, + 0.04483562335371971, + 0.0526740588247776, + 0.03372164070606232, + 0.07540451735258102, + 0.029828885570168495, + 0.008968612179160118, + 0.0035128016024827957, + 0.01701410487294197, + 0.004520055837929249, + 0.003429156495258212, + 0.003350196871906519, + 0.0038932764437049627, + 0.004481090698391199, + 0.02227177284657955, + 0.017743000760674477, + 0.01609528250992298, + 0.018181459978222847, + 0.007703742012381554, + 0.029425015673041344, + 0.009274554438889027, + 0.009668861515820026, + 0.010838229209184647, + 0.013332117348909378, + 0.012891817837953568, + 0.013670663349330425, + 0.0071344682946801186, + 0.008977347984910011, + 0.007863444276154041, + 0.007912615314126015, + 0.007433921564370394, + 0.005264678504317999, + 0.00710633909329772, + 0.0063090198673307896, + 0.004989141598343849, + 0.0073610027320683, + 0.009012727066874504, + 0.0045317476615309715, + 0.0054891351610422134, + 0.0034787554759532213, + 0.007059428375214338, + 0.005693317856639624, + 0.002756975358352065, + 0.0029714852571487427, + 0.0029385641682893038, + 0.0023997665848582983, + 0.0021759921219199896, + 0.002004444133490324, + 0.004255663137882948, + 0.0014802374644204974, + 0.0011819646460935473, + 0.0037271378096193075, + 0.0015739506343379617, + 0.001350793638266623, + 0.020896047353744507, + 0.010092025622725487, + 0.03174324333667755, + 0.012362212873995304, + 0.014399059116840363, + 0.011264054104685783, + 0.015364468097686768, + 0.00999513640999794, + 0.007881386205554008, + 0.007725169882178307, + 0.00684914318844676, + 0.006918848026543856, + 0.00468281889334321, + 0.007552584167569876, + 0.0056065525859594345, + 0.00841600727289915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.007640897762030363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033641755580902 + ], + "opposite_logprob_tokens": [ + -2.3503363132476807 + ], + "logprob_reward": -0.10033641755580902, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033641755580902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06681457161903381, + 0.04575503244996071, + 0.019610567018389702, + 0.13799533247947693, + 0.047602783888578415, + 0.4361324608325958, + 0.04834930598735809, + 0.05130864679813385, + 0.07551758736371994, + 0.0709136351943016 + ], + "credit_map": [ + 0.005624918732792139, + 0.010437438264489174, + 0.007490204181522131, + 0.009703770279884338, + 0.009133350104093552, + 0.007628795225173235, + 0.009823562577366829, + 0.0069725364446640015, + 0.006544684991240501, + 0.004067716654390097, + 0.01316404901444912, + 0.004853160586208105, + 0.005201220512390137, + 0.003644200274720788, + 0.004957097116857767, + 0.0033229077234864235, + 0.002809435361996293, + 0.002422851277515292, + 0.004181209485977888, + 0.002158685587346554, + 0.0017138330731540918, + 0.0016084432136267424, + 0.0026792550925165415, + 0.0020368550904095173, + 0.0221035685390234, + 0.011725387535989285, + 0.019966809079051018, + 0.019580751657485962, + 0.007801540195941925, + 0.02889114059507847, + 0.008208069019019604, + 0.019718075171113014, + 0.010928058065474033, + 0.004442739766091108, + 0.007867314852774143, + 0.004996003117412329, + 0.002585820620879531, + 0.005502103362232447, + 0.003281870624050498, + 0.007998878136277199, + 0.06376690417528152, + 0.040954723954200745, + 0.06422591209411621, + 0.09202217310667038, + 0.029982205480337143, + 0.05572867393493652, + 0.050327856093645096, + 0.03912406042218208, + 0.0084835859015584, + 0.005227426066994667, + 0.01196077186614275, + 0.007924447767436504, + 0.0030855576042085886, + 0.00428821099922061, + 0.0028215849306434393, + 0.004557727370411158, + 0.00599765544757247, + 0.004211693536490202, + 0.00943146925419569, + 0.006797006819397211, + 0.0037558721378445625, + 0.01167899090796709, + 0.00512947142124176, + 0.0043064928613603115, + 0.010681639425456524, + 0.007398297544568777, + 0.0144527368247509, + 0.012241668067872524, + 0.00701191695407033, + 0.0069922273978590965, + 0.008203057572245598, + 0.008536056615412235, + 0.013655676506459713, + 0.005443792790174484, + 0.008118192665278912, + 0.008436889387667179, + 0.006439560558646917, + 0.0060180239379405975, + 0.007579953875392675, + 0.015221554785966873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.10033641755580902 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022027057129889727 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022027057129889727, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022027057129889727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19256369769573212, + 0.06749371439218521, + 0.07744185626506805, + 0.11050651222467422, + 0.052707821130752563, + 0.21903306245803833, + 0.04587797448039055, + 0.044300779700279236, + 0.1167813092470169, + 0.07329320907592773 + ], + "credit_map": [ + 0.021900970488786697, + 0.026075858622789383, + 0.022250544279813766, + 0.023329783231019974, + 0.02792050689458847, + 0.022370237857103348, + 0.02746391110122204, + 0.02125190570950508, + 0.007349309511482716, + 0.007702244911342859, + 0.01348891668021679, + 0.00848920363932848, + 0.009114102460443974, + 0.00633420143276453, + 0.009298326447606087, + 0.005717421881854534, + 0.011068600229918957, + 0.010472686029970646, + 0.016777414828538895, + 0.00462329899892211, + 0.00885675847530365, + 0.01348028052598238, + 0.007306849583983421, + 0.004855974577367306, + 0.016750076785683632, + 0.012747395783662796, + 0.010311458259820938, + 0.011687313206493855, + 0.009873899631202221, + 0.02469797618687153, + 0.013749645091593266, + 0.010688764974474907, + 0.007176400162279606, + 0.005072080995887518, + 0.007957272231578827, + 0.007622108329087496, + 0.005097386427223682, + 0.0072267139330506325, + 0.005674703046679497, + 0.006881161592900753, + 0.031042296439409256, + 0.020636705681681633, + 0.03278503939509392, + 0.04404180496931076, + 0.015954839065670967, + 0.028653984889388084, + 0.026066282764077187, + 0.01985214091837406, + 0.01056219357997179, + 0.004039925057440996, + 0.016018597409129143, + 0.0050074150785803795, + 0.001973966136574745, + 0.0025250567123293877, + 0.002135565737262368, + 0.003615263383835554, + 0.005353373009711504, + 0.0044457861222326756, + 0.006963339634239674, + 0.005524559877812862, + 0.004411722533404827, + 0.007573696784675121, + 0.005543269217014313, + 0.004485038109123707, + 0.02183569222688675, + 0.012581664137542248, + 0.03073885291814804, + 0.011563324369490147, + 0.010010913014411926, + 0.007094056345522404, + 0.011518801562488079, + 0.011438015848398209, + 0.010447917506098747, + 0.007122335955500603, + 0.01286594569683075, + 0.00820810254663229, + 0.007806535344570875, + 0.006626470945775509, + 0.009112839587032795, + 0.011103070341050625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0022027057129889727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two black garden ants visible engaging in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single large translucent sugar crystal present at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a polished black mahogany table with visible wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot from a low angle with a shallow depth of field creating a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion captured in high-speed slow motion with a slow progression?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782115928828716 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782115928828716, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782115928828716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17229607701301575, + 0.08975809067487717, + 0.02379741333425045, + 0.1944437026977539, + 0.036005377769470215, + 0.1380470246076584, + 0.11821574717760086, + 0.044924162328243256, + 0.0918947160243988, + 0.09061767160892487 + ], + "credit_map": [ + 0.01425296813249588, + 0.019765092059969902, + 0.039243895560503006, + 0.014810875058174133, + 0.01617930456995964, + 0.0278418380767107, + 0.021050358191132545, + 0.01915173977613449, + 0.013603728264570236, + 0.005880481097847223, + 0.03235741704702377, + 0.010528428480029106, + 0.007664961274713278, + 0.005992045160382986, + 0.008631417527794838, + 0.005099609028548002, + 0.003012046217918396, + 0.003180151106789708, + 0.0032743823248893023, + 0.0020756854210048914, + 0.0031322482973337173, + 0.003131126519292593, + 0.0034547511022537947, + 0.0025370230432599783, + 0.03531556576490402, + 0.0244861152023077, + 0.021356938406825066, + 0.022798720747232437, + 0.012372721917927265, + 0.04787864536046982, + 0.01293895673006773, + 0.01729603298008442, + 0.006197147071361542, + 0.00461621442809701, + 0.006640938110649586, + 0.004921320825815201, + 0.0028622851241379976, + 0.003950509242713451, + 0.0033220972400158644, + 0.0034948631655424833, + 0.021212754771113396, + 0.014026043936610222, + 0.019398583099246025, + 0.02559245005249977, + 0.009049944579601288, + 0.025966931134462357, + 0.010524798184633255, + 0.012275516986846924, + 0.018586546182632446, + 0.012421694584190845, + 0.023447582498192787, + 0.02307206392288208, + 0.006975301541388035, + 0.014990899711847305, + 0.007082018069922924, + 0.01163964532315731, + 0.0068163941614329815, + 0.005254832562059164, + 0.013323567807674408, + 0.004660450387746096, + 0.0028348302002996206, + 0.0049542030319571495, + 0.004462398122996092, + 0.002617483725771308, + 0.01400927733629942, + 0.010602130554616451, + 0.017709119245409966, + 0.0072913565672934055, + 0.012669742107391357, + 0.008764667436480522, + 0.012089007534086704, + 0.00875941477715969, + 0.01405441015958786, + 0.012743034400045872, + 0.01107292715460062, + 0.008903556503355503, + 0.014288043603301048, + 0.010669760406017303, + 0.008452012203633785, + 0.010433929041028023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.009782115928828716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_tensor([4575, 4575])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_tensor([4575, 4575])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..af3b9899ac123b5a126ae51d3b7883fed19ca5e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_tensor([4575, 4575])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, fixed-mounted projection shot from the hood of a rally-ready Ford RS200 captures a visceral and chaotic perspective, the world shaking and vibrating violently with every brutal movement of the machine. The car executes an aggressive, high-speed drift, sliding sideways across a narrow, sun-drenched dirt track lined with blurry green foliage. As the vehicle hangs on the edge of control, the camera jarringly captures rapid countersteering adjustments, unleashing a torrent of thick ochre dust and fine gravel from the rear wheels. This dense cloud rolls forward, streaking across the lens in thick, grainy waves, temporarily obscuring the view and creating a raw and intense display of power. Harsh sunlight penetrates the dust, illuminating the chaos and casting sharp, fleeting shadows across the rutted track, creating a gritty, high-contrast atmosphere that is both breathtaking and immersive.", + "seed": 33079, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_4575_seed_33079_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06765945996157825, + "vlm_logprob": -0.06765945996157825, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06765945996157825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0545482337474823 + ], + "opposite_logprob_tokens": [ + -3.0545482635498047 + ], + "logprob_reward": -0.0545482337474823, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0545482337474823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053714316338300705, + 0.03441907465457916, + 0.17628155648708344, + 0.09231402724981308, + 0.11605769395828247, + 0.02575625479221344, + 0.03213994950056076, + 0.3079473078250885, + 0.13193145394325256, + 0.02943834289908409 + ], + "credit_map": [ + 0.006313163321465254, + 0.0050749569199979305, + 0.00899624265730381, + 0.006725937128067017, + 0.006952352821826935, + 0.006856831256300211, + 0.007180782034993172, + 0.005614053457975388, + 0.004261170048266649, + 0.003729703603312373, + 0.0039554378017783165, + 0.0061110844835639, + 0.004311626311391592, + 0.002715637208893895, + 0.005286265164613724, + 0.004048155155032873, + 0.02842213213443756, + 0.028711235150694847, + 0.039878033101558685, + 0.017094768583774567, + 0.015084635466337204, + 0.014469721354544163, + 0.013690762221813202, + 0.018930265679955482, + 0.022481344640254974, + 0.029488414525985718, + 0.010761311277747154, + 0.00637464364990592, + 0.008403623476624489, + 0.005075174383819103, + 0.0056372894905507565, + 0.004092233255505562, + 0.019817685708403587, + 0.036615848541259766, + 0.013999896124005318, + 0.02998468093574047, + 0.005423704162240028, + 0.00318039208650589, + 0.004167611710727215, + 0.002867890987545252, + 0.0023383249063044786, + 0.004178084433078766, + 0.004203601274639368, + 0.0035988041199743748, + 0.002672859001904726, + 0.0034329297486692667, + 0.0025972577277570963, + 0.0027343968395143747, + 0.0023592328652739525, + 0.0029188385233283043, + 0.0032635973766446114, + 0.002539215376600623, + 0.008990687318146229, + 0.008107298985123634, + 0.002413741545751691, + 0.0015473394887521863, + 0.00906659197062254, + 0.08903029561042786, + 0.0076966783963143826, + 0.01547913160175085, + 0.1235223263502121, + 0.03308864310383797, + 0.020513849332928658, + 0.009549805894494057, + 0.01118433941155672, + 0.013928906992077827, + 0.020110847428441048, + 0.02155187539756298, + 0.017540648579597473, + 0.024709666147828102, + 0.013372230343520641, + 0.009532944299280643, + 0.0017445589182898402, + 0.0034710129257291555, + 0.0036910884082317352, + 0.007214097771793604, + 0.0035098993685096502, + 0.002845887327566743, + 0.0036110582295805216, + 0.003350744489580393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.0545482337474823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040587425231934 + ], + "opposite_logprob_tokens": [ + -1.9104058742523193 + ], + "logprob_reward": -0.16040587425231934, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040587425231934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06617378443479538, + 0.013797051273286343, + 0.02456335350871086, + 0.057524241507053375, + 0.23626628518104553, + 0.018687784671783447, + 0.2360587865114212, + 0.2093268483877182, + 0.09248828142881393, + 0.04511352255940437 + ], + "credit_map": [ + 0.004275445360690355, + 0.0061868405900895596, + 0.009341380558907986, + 0.004520809277892113, + 0.0039309547282755375, + 0.012303447350859642, + 0.005244793836027384, + 0.0203701164573431, + 0.0022004698403179646, + 0.001655188505537808, + 0.003222357714548707, + 0.0017261544708162546, + 0.0013518116902559996, + 0.0011538249673321843, + 0.0013003471540287137, + 0.0011868971632793546, + 0.003418730804696679, + 0.0053051551803946495, + 0.0053962538950145245, + 0.002603521104902029, + 0.0019221975235268474, + 0.0020440449006855488, + 0.0017086578300222754, + 0.0021647936664521694, + 0.01123871747404337, + 0.01675259880721569, + 0.006460297387093306, + 0.004281988833099604, + 0.005092182196676731, + 0.0037596120964735746, + 0.004590923432260752, + 0.005347925238311291, + 0.03399462252855301, + 0.06140447407960892, + 0.040049463510513306, + 0.07004391402006149, + 0.011181688867509365, + 0.005445682909339666, + 0.009407876059412956, + 0.004738576710224152, + 0.0020105820149183273, + 0.0031993910670280457, + 0.002182078780606389, + 0.002138739451766014, + 0.0019975798204541206, + 0.002208505757153034, + 0.003024438163265586, + 0.001926471944898367, + 0.007260990794748068, + 0.01509154960513115, + 0.013695038855075836, + 0.02633119933307171, + 0.07290217280387878, + 0.05944221839308739, + 0.02865472063422203, + 0.012680893763899803, + 0.006316537968814373, + 0.043827783316373825, + 0.012188341468572617, + 0.009259918704628944, + 0.08037308603525162, + 0.022651981562376022, + 0.01742441952228546, + 0.01728479005396366, + 0.0076949624344706535, + 0.009999053552746773, + 0.01401746179908514, + 0.016459995880723, + 0.013118484057486057, + 0.01638866774737835, + 0.008849194273352623, + 0.005960461217910051, + 0.0014699148014187813, + 0.008793055079877377, + 0.0037955637089908123, + 0.017106765881180763, + 0.004121673293411732, + 0.0024854575749486685, + 0.004284470807760954, + 0.003056625835597515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.16040587425231934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783178567886353 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783178567886353, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783178567886353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08997631818056107, + 0.052574947476387024, + 0.09775478392839432, + 0.15298256278038025, + 0.14772681891918182, + 0.020513081923127174, + 0.11710840463638306, + 0.18805167078971863, + 0.08482494950294495, + 0.048486482352018356 + ], + "credit_map": [ + 0.008393898606300354, + 0.006032226607203484, + 0.015256689861416817, + 0.007085069082677364, + 0.007527223788201809, + 0.015797069296240807, + 0.007365941535681486, + 0.02251819334924221, + 0.005934965796768665, + 0.015288636088371277, + 0.01205118652433157, + 0.005224565044045448, + 0.0034750269260257483, + 0.0039017233066260815, + 0.0035384814254939556, + 0.0031603581737726927, + 0.018255585804581642, + 0.019314292818307877, + 0.02710345946252346, + 0.011452147737145424, + 0.005827917251735926, + 0.005366042256355286, + 0.00512738898396492, + 0.005307942628860474, + 0.03403302654623985, + 0.044714465737342834, + 0.017979806289076805, + 0.011759339831769466, + 0.01471081841737032, + 0.008020332083106041, + 0.011641159653663635, + 0.010123605839908123, + 0.019924961030483246, + 0.03645908460021019, + 0.026421772316098213, + 0.04494301602244377, + 0.006428747437894344, + 0.0044083064422011375, + 0.005697997286915779, + 0.0034429216757416725, + 0.0022253647912293673, + 0.0042034643702209, + 0.0026916677597910166, + 0.0022624987177550793, + 0.0017336090095341206, + 0.0023786991368979216, + 0.0022803584579378366, + 0.0027374185156077147, + 0.004014196339994669, + 0.008314015343785286, + 0.006265848875045776, + 0.013234666548669338, + 0.03564458340406418, + 0.028612051159143448, + 0.014413216151297092, + 0.006609809584915638, + 0.012461284175515175, + 0.06896264851093292, + 0.01539070624858141, + 0.011104105040431023, + 0.039282795041799545, + 0.015580364502966404, + 0.009800013154745102, + 0.015469744801521301, + 0.0072424160316586494, + 0.009018206037580967, + 0.01230608019977808, + 0.013872673735022545, + 0.012323439121246338, + 0.01500717457383871, + 0.008866557851433754, + 0.006188397761434317, + 0.0021428847685456276, + 0.010715021751821041, + 0.005325508303940296, + 0.005838119424879551, + 0.01072597224265337, + 0.005144457332789898, + 0.004440946504473686, + 0.004153567366302013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.009783178567886353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006345880683511496 + ], + "opposite_logprob_tokens": [ + -7.375634670257568 + ], + "logprob_reward": -0.0006345880683511496, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006345880683511496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046950314193964005, + 0.030078381299972534, + 0.012640335597097874, + 0.03462546318769455, + 0.1469249129295349, + 0.017594940960407257, + 0.14074070751667023, + 0.3423008322715759, + 0.10878023505210876, + 0.11936391144990921 + ], + "credit_map": [ + 0.003109026700258255, + 0.004190773703157902, + 0.006292629987001419, + 0.0037628100253641605, + 0.0030565480701625347, + 0.00805622898042202, + 0.0035472086165100336, + 0.014935087412595749, + 0.004752859473228455, + 0.00503720436245203, + 0.0036210999824106693, + 0.004216453060507774, + 0.0026908323634415865, + 0.004168014042079449, + 0.0027967561036348343, + 0.0027951609808951616, + 0.0019171538297086954, + 0.002502583898603916, + 0.0024201807100325823, + 0.0014395777834579349, + 0.0012042130110785365, + 0.0010134095791727304, + 0.0009046799968928099, + 0.0012385363224893808, + 0.004509925376623869, + 0.0032990409526973963, + 0.007051246240735054, + 0.006038472056388855, + 0.004581584595143795, + 0.002539228880777955, + 0.0036258886102586985, + 0.002980075078085065, + 0.0205052699893713, + 0.03795494884252548, + 0.026159344241023064, + 0.04577392712235451, + 0.005619215779006481, + 0.0025429108645766973, + 0.00490786274895072, + 0.003461431246250868, + 0.0020801774226129055, + 0.0025732158683240414, + 0.003358126850798726, + 0.0024192228447645903, + 0.0018207142129540443, + 0.001735225087031722, + 0.001956575084477663, + 0.0016516828909516335, + 0.005593264941126108, + 0.010346035473048687, + 0.009769654832780361, + 0.015603464096784592, + 0.042812954634428024, + 0.03414858877658844, + 0.01564694009721279, + 0.006819802802056074, + 0.01138954795897007, + 0.09800080955028534, + 0.009065019898116589, + 0.015967901796102524, + 0.13656085729599, + 0.04033384844660759, + 0.021700419485569, + 0.009282386861741543, + 0.009108447469770908, + 0.01199504267424345, + 0.016552450135350227, + 0.019908789545297623, + 0.015267614275217056, + 0.016236823052167892, + 0.011514978483319283, + 0.008196095004677773, + 0.004427856765687466, + 0.04351870343089104, + 0.005175578407943249, + 0.02977832406759262, + 0.01826963573694229, + 0.007233762182295322, + 0.0066627864725887775, + 0.004297264851629734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.0006345880683511496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292542517185211 + ], + "opposite_logprob_tokens": [ + -2.2379255294799805 + ], + "logprob_reward": -0.11292542517185211, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292542517185211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2398584634065628, + 0.05722802132368088, + 0.06902167946100235, + 0.1744888573884964, + 0.17512284219264984, + 0.041060786694288254, + 0.06984615325927734, + 0.06945144385099411, + 0.06646368652582169, + 0.03745799884200096 + ], + "credit_map": [ + 0.014822021126747131, + 0.01266273483633995, + 0.0335901640355587, + 0.012674913741648197, + 0.009493224322795868, + 0.05016030743718147, + 0.01594035141170025, + 0.09051474183797836, + 0.006609632167965174, + 0.007332035340368748, + 0.012609110213816166, + 0.01358963456004858, + 0.0041926284320652485, + 0.0030348917935043573, + 0.004542027600109577, + 0.0053180609829723835, + 0.010833166539669037, + 0.010397293604910374, + 0.022654037922620773, + 0.009425004944205284, + 0.004385851323604584, + 0.003605516394600272, + 0.0037004773039370775, + 0.004020329564809799, + 0.051831234246492386, + 0.061829615384340286, + 0.015719633549451828, + 0.021360624581575394, + 0.009301739744842052, + 0.004555332008749247, + 0.0058461143635213375, + 0.004044570494443178, + 0.0215999037027359, + 0.03799647092819214, + 0.03357464820146561, + 0.05448439344763756, + 0.008794811554253101, + 0.006084127351641655, + 0.006619430147111416, + 0.005969060119241476, + 0.004025357309728861, + 0.007976159453392029, + 0.008866698481142521, + 0.00748412124812603, + 0.002864461624994874, + 0.0031333628576248884, + 0.004333251155912876, + 0.002377372467890382, + 0.0036470936611294746, + 0.005515311844646931, + 0.006071576848626137, + 0.00784560851752758, + 0.016554899513721466, + 0.019869064912199974, + 0.006306701339781284, + 0.004035901743918657, + 0.00578601798042655, + 0.0069680954329669476, + 0.005881090648472309, + 0.004854517988860607, + 0.015345939435064793, + 0.008667061105370522, + 0.009329576045274734, + 0.01261914148926735, + 0.0048257531598210335, + 0.007436500396579504, + 0.009252123534679413, + 0.012829728424549103, + 0.009888829663395882, + 0.011066713370382786, + 0.0070045278407633305, + 0.004159507341682911, + 0.002164986217394471, + 0.004751346539705992, + 0.004155052360147238, + 0.007076228968799114, + 0.004801735281944275, + 0.005775218363851309, + 0.004265917465090752, + 0.004467514343559742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.11292542517185211 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, fixed-mounted projection shot from the hood of a rally-ready Ford RS200 captures a visceral and chaotic perspective, the world shaking and vibrating violently with every brutal movement of the machine. The car executes an aggressive, high-speed drift, sliding sideways across a narrow, sun-drenched dirt track lined with blurry green foliage. As the vehicle hangs on the edge of control, the camera jarringly captures rapid countersteering adjustments, unleashing a torrent of thick ochre dust and fine gravel from the rear wheels. This dense cloud rolls forward, streaking across the lens in thick, grainy waves, temporarily obscuring the view and creating a raw and intense display of power. Harsh sunlight penetrates the dust, illuminating the chaos and casting sharp, fleeting shadows across the rutted track, creating a gritty, high-contrast atmosphere that is both breathtaking and immersive.", + "seed": 33080, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_4575_seed_33080_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05312362432014197, + "vlm_logprob": -0.05312362432014197, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.05312362432014197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06578145176172256 + ], + "opposite_logprob_tokens": [ + -2.940781354904175 + ], + "logprob_reward": -0.06578145176172256, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06578145176172256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13579654693603516, + 0.05945640429854393, + 0.046973105520009995, + 0.23858769237995148, + 0.028364909812808037, + 0.03043484315276146, + 0.1925082802772522, + 0.11985847353935242, + 0.09525593370199203, + 0.05276374891400337 + ], + "credit_map": [ + 0.02363641746342182, + 0.010164821520447731, + 0.01580362394452095, + 0.016462048515677452, + 0.012278812937438488, + 0.024251479655504227, + 0.014169593341648579, + 0.019029757007956505, + 0.0063226786442101, + 0.006884320173412561, + 0.0062228599563241005, + 0.006778085604310036, + 0.009929674677550793, + 0.010493628680706024, + 0.007265157997608185, + 0.0055600060150027275, + 0.006570576690137386, + 0.005763227120041847, + 0.006109233945608139, + 0.007150162011384964, + 0.005520991049706936, + 0.005954372696578503, + 0.006161380559206009, + 0.003743165172636509, + 0.012223881669342518, + 0.01651621237397194, + 0.046302299946546555, + 0.05290711298584938, + 0.013189487159252167, + 0.010994353331625462, + 0.043992068618535995, + 0.04246232286095619, + 0.0028633426409214735, + 0.002912977710366249, + 0.004113371949642897, + 0.004847335163503885, + 0.0022482508793473244, + 0.003337134374305606, + 0.003423099871724844, + 0.004619400482624769, + 0.0032562417909502983, + 0.0026222646702080965, + 0.00747324526309967, + 0.003943015355616808, + 0.002693544141948223, + 0.0023063975386321545, + 0.005471726413816214, + 0.0026684117037802935, + 0.03261460363864899, + 0.021045317873358727, + 0.03210044279694557, + 0.031790975481271744, + 0.015329021029174328, + 0.01696058176457882, + 0.02150142379105091, + 0.021165935322642326, + 0.009269161149859428, + 0.01119353249669075, + 0.02587972953915596, + 0.01704796776175499, + 0.012462472543120384, + 0.015442628413438797, + 0.01784355938434601, + 0.010719437152147293, + 0.00865140464156866, + 0.008979391306638718, + 0.01436260063201189, + 0.009581650607287884, + 0.012669630348682404, + 0.017602209001779556, + 0.012453828006982803, + 0.01095522753894329, + 0.0037379004061222076, + 0.006373596843332052, + 0.008087902329862118, + 0.009371940046548843, + 0.004500159993767738, + 0.0024992136750370264, + 0.010240309871733189, + 0.007952730171382427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.06578145176172256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062093883752822876 + ], + "opposite_logprob_tokens": [ + -2.81209397315979 + ], + "logprob_reward": -0.062093883752822876, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062093883752822876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03296573460102081, + 0.02984072081744671, + 0.03229711577296257, + 0.037622325122356415, + 0.03518429026007652, + 0.028327427804470062, + 0.2907128632068634, + 0.3815559148788452, + 0.05358961597084999, + 0.07790407538414001 + ], + "credit_map": [ + 0.0050250026397407055, + 0.004303776193410158, + 0.004635150544345379, + 0.005576903931796551, + 0.0038554645143449306, + 0.0033750717993825674, + 0.003589675994589925, + 0.002604687586426735, + 0.005109556019306183, + 0.00393915269523859, + 0.0034859427250921726, + 0.005282876081764698, + 0.0037932288832962513, + 0.0030161840841174126, + 0.0020650294609367847, + 0.0031487487722188234, + 0.004146651830524206, + 0.005996223539113998, + 0.004554972052574158, + 0.005637820810079575, + 0.0043965051881968975, + 0.0019480463815853, + 0.003476977813988924, + 0.0021399122197180986, + 0.0027438749093562365, + 0.0026633944362401962, + 0.009488268755376339, + 0.007636097259819508, + 0.002777357120066881, + 0.00182661646977067, + 0.005096720065921545, + 0.005389994475990534, + 0.0026882931124418974, + 0.0030602014157921076, + 0.005894838832318783, + 0.007528814487159252, + 0.0034987921826541424, + 0.0036647228989750147, + 0.0036739858333021402, + 0.005174638237804174, + 0.002675853203982115, + 0.0024789259769022465, + 0.006114686373621225, + 0.0042448705062270164, + 0.0029820434283465147, + 0.002809392288327217, + 0.004978249780833721, + 0.0020434041507542133, + 0.051521167159080505, + 0.035354018211364746, + 0.04193181172013283, + 0.0429701991379261, + 0.02509886957705021, + 0.02998596988618374, + 0.03323254734277725, + 0.03061826527118683, + 0.032709501683712006, + 0.0354674831032753, + 0.08043397963047028, + 0.05508246272802353, + 0.04011955484747887, + 0.04709563031792641, + 0.056223757565021515, + 0.03442351520061493, + 0.005982841830700636, + 0.007852163165807724, + 0.009344077669084072, + 0.0048606134951114655, + 0.004845605697482824, + 0.007680817041546106, + 0.006777455564588308, + 0.0062460340559482574, + 0.004434632603079081, + 0.008534597232937813, + 0.009853975847363472, + 0.017369767650961876, + 0.005729016847908497, + 0.0037258719094097614, + 0.01655525341629982, + 0.011700952425599098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.062093883752822876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782588109374046 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782588109374046, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782588109374046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07691685110330582, + 0.06999333947896957, + 0.1129021868109703, + 0.12252657115459442, + 0.15905815362930298, + 0.07434701919555664, + 0.057917576283216476, + 0.11335714906454086, + 0.10164234042167664, + 0.111338771879673 + ], + "credit_map": [ + 0.01030334085226059, + 0.006328306160867214, + 0.009089482948184013, + 0.015946708619594574, + 0.006976592354476452, + 0.009049208834767342, + 0.008684066124260426, + 0.010539152659475803, + 0.011275392025709152, + 0.0096311429515481, + 0.007621854543685913, + 0.009501287713646889, + 0.009742263704538345, + 0.010213096626102924, + 0.006119263358414173, + 0.005889045540243387, + 0.024902570992708206, + 0.016922149807214737, + 0.014886937104165554, + 0.019733959808945656, + 0.011203311383724213, + 0.006484149489551783, + 0.01040112879127264, + 0.008367989212274551, + 0.008887768723070621, + 0.011168742552399635, + 0.015106834471225739, + 0.03002098575234413, + 0.010445408523082733, + 0.009276747703552246, + 0.02452935464680195, + 0.013090739026665688, + 0.028571398928761482, + 0.013593407347798347, + 0.02676510624587536, + 0.025684326887130737, + 0.011787699535489082, + 0.01392803993076086, + 0.017762597650289536, + 0.020965591073036194, + 0.005838362034410238, + 0.00556727172806859, + 0.011879829689860344, + 0.017011919990181923, + 0.007725107949227095, + 0.006762320175766945, + 0.0117146922275424, + 0.007847517728805542, + 0.005854357499629259, + 0.006119899917393923, + 0.009032339788973331, + 0.008760279975831509, + 0.0075838700868189335, + 0.006300601176917553, + 0.006866548676043749, + 0.007399688009172678, + 0.011533597484230995, + 0.013966850936412811, + 0.017992474138736725, + 0.016052749007940292, + 0.01232115924358368, + 0.01541112456470728, + 0.015443030744791031, + 0.010636166669428349, + 0.009460024535655975, + 0.012129186652600765, + 0.01783922128379345, + 0.009457174688577652, + 0.0140248266980052, + 0.0159116443246603, + 0.01178696844726801, + 0.011033295653760433, + 0.007736510131508112, + 0.016107210889458656, + 0.015584592707455158, + 0.021886780858039856, + 0.011247980408370495, + 0.008061802946031094, + 0.018233492970466614, + 0.012480408884584904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.009782588109374046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009210634743794799 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009210634743794799, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009210634743794799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08431592583656311, + 0.07501315325498581, + 0.05307319760322571, + 0.11160802841186523, + 0.16513849794864655, + 0.09675528109073639, + 0.06341328471899033, + 0.05575332045555115, + 0.11569495499134064, + 0.1792343109846115 + ], + "credit_map": [ + 0.013010576367378235, + 0.008330614306032658, + 0.00834214873611927, + 0.011675707064568996, + 0.007532930467277765, + 0.01114238053560257, + 0.011776476167142391, + 0.012505097314715385, + 0.011889418587088585, + 0.01286532822996378, + 0.008542505092918873, + 0.010722378268837929, + 0.009114367887377739, + 0.008723591454327106, + 0.006728556007146835, + 0.00642700819298625, + 0.006242707371711731, + 0.009152152575552464, + 0.008281555026769638, + 0.010345018468797207, + 0.006763821002095938, + 0.00392559077590704, + 0.005209002178162336, + 0.003153352066874504, + 0.008505260571837425, + 0.007576523348689079, + 0.03309674188494682, + 0.014531540684401989, + 0.006786257028579712, + 0.0072080595418810844, + 0.012865225784480572, + 0.021038424223661423, + 0.016254723072052002, + 0.009549734182655811, + 0.04333484172821045, + 0.03497946262359619, + 0.010214189998805523, + 0.0065542906522750854, + 0.018171895295381546, + 0.02607935667037964, + 0.019356925040483475, + 0.009290747344493866, + 0.017310330644249916, + 0.013776625506579876, + 0.00740755395963788, + 0.007975893095135689, + 0.01361939962953329, + 0.008017804473638535, + 0.007443313021212816, + 0.0048642330802977085, + 0.010578426532447338, + 0.00844925083220005, + 0.005800371058285236, + 0.004922683350741863, + 0.010046700946986675, + 0.01130830030888319, + 0.0074019418098032475, + 0.007533180061727762, + 0.010843469761312008, + 0.007255613803863525, + 0.005334270652383566, + 0.006532791070640087, + 0.0060098422691226006, + 0.0048422059044241905, + 0.013407985679805279, + 0.016489461064338684, + 0.0166573915630579, + 0.013883169740438461, + 0.010499341413378716, + 0.016065377742052078, + 0.01522042602300644, + 0.013471807353198528, + 0.013149443082511425, + 0.024301040917634964, + 0.026027845218777657, + 0.03411141410470009, + 0.01641487516462803, + 0.012925868853926659, + 0.02861464023590088, + 0.02368919737637043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.0009210634743794799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270391345024109 + ], + "opposite_logprob_tokens": [ + -2.1270391941070557 + ], + "logprob_reward": -0.1270391345024109, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270391345024109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37095242738723755, + 0.08104518800973892, + 0.06688671559095383, + 0.059900328516960144, + 0.059383574873209, + 0.04315276816487312, + 0.03760882094502449, + 0.05862279608845711, + 0.1611250787973404, + 0.0613224133849144 + ], + "credit_map": [ + 0.07641395181417465, + 0.04105750471353531, + 0.04073498398065567, + 0.03577309101819992, + 0.033437494188547134, + 0.07425491511821747, + 0.023612581193447113, + 0.045667871832847595, + 0.015203621238470078, + 0.012709070928394794, + 0.009646077640354633, + 0.009702437557280064, + 0.010557103902101517, + 0.007761147804558277, + 0.0070639909245073795, + 0.008401727303862572, + 0.011085070669651031, + 0.005597440525889397, + 0.008823324926197529, + 0.012798664160072803, + 0.006532121449708939, + 0.007167914882302284, + 0.007345430087298155, + 0.007536744698882103, + 0.004571293946355581, + 0.006099342834204435, + 0.008994430303573608, + 0.00813707709312439, + 0.00712662423029542, + 0.006283162627369165, + 0.008555037900805473, + 0.010133355855941772, + 0.007718000095337629, + 0.005338079296052456, + 0.008816777728497982, + 0.008486301638185978, + 0.006332335527986288, + 0.00371657800860703, + 0.0076500242576003075, + 0.0113254738971591, + 0.003894967958331108, + 0.004813008476048708, + 0.009188277646899223, + 0.008168784901499748, + 0.004117149394005537, + 0.003548169042915106, + 0.006558498367667198, + 0.002863910049200058, + 0.005230977199971676, + 0.008137921802699566, + 0.006115032825618982, + 0.004980795085430145, + 0.0034575420431792736, + 0.0026482960674911737, + 0.003275244263932109, + 0.0037630058359354734, + 0.006143011152744293, + 0.007181783206760883, + 0.008206303231418133, + 0.007772053126245737, + 0.007054414600133896, + 0.008258430287241936, + 0.00852794386446476, + 0.005478855222463608, + 0.017305321991443634, + 0.026918839663267136, + 0.026548253372311592, + 0.014304254204034805, + 0.01573892869055271, + 0.02440831996500492, + 0.018467815592885017, + 0.017433330416679382, + 0.004008026793599129, + 0.00444191787391901, + 0.007429015357047319, + 0.011385633610188961, + 0.004367413464933634, + 0.0034643507096916437, + 0.015538137406110764, + 0.010687905363738537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.1270391345024109 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, fixed-mounted projection shot from the hood of a rally-ready Ford RS200 captures a visceral and chaotic perspective, the world shaking and vibrating violently with every brutal movement of the machine. The car executes an aggressive, high-speed drift, sliding sideways across a narrow, sun-drenched dirt track lined with blurry green foliage. As the vehicle hangs on the edge of control, the camera jarringly captures rapid countersteering adjustments, unleashing a torrent of thick ochre dust and fine gravel from the rear wheels. This dense cloud rolls forward, streaking across the lens in thick, grainy waves, temporarily obscuring the view and creating a raw and intense display of power. Harsh sunlight penetrates the dust, illuminating the chaos and casting sharp, fleeting shadows across the rutted track, creating a gritty, high-contrast atmosphere that is both breathtaking and immersive.", + "seed": 33081, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_4575_seed_33081_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04876874441979453, + "vlm_logprob": -0.04876874441979453, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.04876874441979453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0686618909239769 + ], + "opposite_logprob_tokens": [ + -2.81866192817688 + ], + "logprob_reward": -0.0686618909239769, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0686618909239769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15083517134189606, + 0.21958313882350922, + 0.0294070765376091, + 0.03740761801600456, + 0.032987143844366074, + 0.03635381534695625, + 0.06472814083099365, + 0.03272555395960808, + 0.18972867727279663, + 0.20624366402626038 + ], + "credit_map": [ + 0.01986042596399784, + 0.017637033015489578, + 0.017604056745767593, + 0.02783905528485775, + 0.01511374581605196, + 0.012401442974805832, + 0.020470013841986656, + 0.019909393042325974, + 0.019293423742055893, + 0.02226799726486206, + 0.024751070886850357, + 0.0483989417552948, + 0.03549790382385254, + 0.01761818490922451, + 0.027300870046019554, + 0.02445472776889801, + 0.0038605479057878256, + 0.0038349793758243322, + 0.004690584260970354, + 0.00432899734005332, + 0.0029446694534271955, + 0.002383363898843527, + 0.0038224845193326473, + 0.003541448386386037, + 0.003672847291454673, + 0.005318060051649809, + 0.005989199038594961, + 0.005682788323611021, + 0.004774033557623625, + 0.00352505617775023, + 0.004398067481815815, + 0.004047568887472153, + 0.0030688485130667686, + 0.00731260608881712, + 0.006231569219380617, + 0.003096124390140176, + 0.004768573213368654, + 0.002241488080471754, + 0.0030464085284620523, + 0.00322152697481215, + 0.003181219333782792, + 0.004393135197460651, + 0.003895716043189168, + 0.0038689670618623495, + 0.009487309493124485, + 0.0035005714744329453, + 0.0038039989303797483, + 0.004222898744046688, + 0.0047971210442483425, + 0.01972559280693531, + 0.008234292268753052, + 0.012426186352968216, + 0.00771856214851141, + 0.0035688409116119146, + 0.004081305116415024, + 0.00417623994871974, + 0.0038616061210632324, + 0.006420761812478304, + 0.004502740688621998, + 0.0034622291568666697, + 0.004375027492642403, + 0.004293826408684254, + 0.0032961731776595116, + 0.0025131921283900738, + 0.018613798543810844, + 0.01727179065346718, + 0.02456703968346119, + 0.05044456571340561, + 0.0147471958771348, + 0.023220287635922432, + 0.018918609246611595, + 0.021945388987660408, + 0.02162625826895237, + 0.02227051369845867, + 0.034659989178180695, + 0.03272220492362976, + 0.020486794412136078, + 0.020701460540294647, + 0.029324153438210487, + 0.024452289566397667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.0686618909239769 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005693763494492 + ], + "opposite_logprob_tokens": [ + -2.695056915283203 + ], + "logprob_reward": -0.07005693763494492, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005693763494492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13091285526752472, + 0.14451009035110474, + 0.044399626553058624, + 0.06790586560964584, + 0.03880779445171356, + 0.06044076383113861, + 0.25365525484085083, + 0.13780967891216278, + 0.0633404552936554, + 0.05821758136153221 + ], + "credit_map": [ + 0.021020647138357162, + 0.009469528682529926, + 0.022862348705530167, + 0.015260451473295689, + 0.009753254242241383, + 0.004861569497734308, + 0.032171376049518585, + 0.015513683669269085, + 0.011212784796953201, + 0.018860245123505592, + 0.017293525859713554, + 0.029644163325428963, + 0.024437569081783295, + 0.01238357275724411, + 0.01670587621629238, + 0.013972340151667595, + 0.005110931117087603, + 0.0055384160950779915, + 0.005538492929190397, + 0.00573864858597517, + 0.0055054775439202785, + 0.006979798898100853, + 0.005069456063210964, + 0.004918409511446953, + 0.006480893585830927, + 0.010513672605156898, + 0.011029011569917202, + 0.010608695447444916, + 0.010250997729599476, + 0.0059387874789536, + 0.007441006135195494, + 0.005642803385853767, + 0.004499607253819704, + 0.006465594284236431, + 0.0049855634570121765, + 0.00458871852606535, + 0.005636377725750208, + 0.003950347192585468, + 0.004212743602693081, + 0.004468842875212431, + 0.0061769988387823105, + 0.006327936891466379, + 0.0067244782112538815, + 0.010057424195110798, + 0.00952491257339716, + 0.007759701460599899, + 0.007062491960823536, + 0.006806816440075636, + 0.01842564344406128, + 0.07526147365570068, + 0.025722576305270195, + 0.05944613739848137, + 0.021274525672197342, + 0.01600075140595436, + 0.020666439086198807, + 0.01685771718621254, + 0.013508717529475689, + 0.018336277455091476, + 0.018330805003643036, + 0.015479708090424538, + 0.020759856328368187, + 0.01959383301436901, + 0.016997795552015305, + 0.014802691526710987, + 0.01010779570788145, + 0.006285662762820721, + 0.007544728461652994, + 0.011290556751191616, + 0.00647875526919961, + 0.008945300243794918, + 0.0056722164154052734, + 0.007015439216047525, + 0.005943913944065571, + 0.007965504191815853, + 0.009292764589190483, + 0.008603001944720745, + 0.0055335634388029575, + 0.005866185296326876, + 0.008391095325350761, + 0.006621557287871838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.07005693763494492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097163211554289 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004097163211554289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097163211554289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22986437380313873, + 0.33819958567619324, + 0.05591411143541336, + 0.11252368986606598, + 0.021379057317972183, + 0.018294336274266243, + 0.026460863649845123, + 0.1261778026819229, + 0.030237114056944847, + 0.040949102491140366 + ], + "credit_map": [ + 0.03850514441728592, + 0.009430157952010632, + 0.04642535001039505, + 0.01845443993806839, + 0.010302436538040638, + 0.006897561252117157, + 0.08257610350847244, + 0.01727316528558731, + 0.028038974851369858, + 0.039527930319309235, + 0.03683196008205414, + 0.07413174211978912, + 0.055635005235672, + 0.026792824268341064, + 0.0399586483836174, + 0.03728245571255684, + 0.006399297621101141, + 0.005973911844193935, + 0.00859218742698431, + 0.009179000742733479, + 0.004328960552811623, + 0.004808731842786074, + 0.009322337806224823, + 0.0073096780106425285, + 0.010208082385361195, + 0.012657644227147102, + 0.022979512810707092, + 0.02291118912398815, + 0.00924847275018692, + 0.005948323290795088, + 0.018651053309440613, + 0.009919404983520508, + 0.0026754902210086584, + 0.002529103308916092, + 0.002882786327973008, + 0.0024855895899236202, + 0.002890978939831257, + 0.0027404590509831905, + 0.002476970199495554, + 0.0026976794470101595, + 0.002496922155842185, + 0.0021062232553958893, + 0.0020598156843334436, + 0.00232200906611979, + 0.0025524115189909935, + 0.0019943576771765947, + 0.0019496645545586944, + 0.002812931314110756, + 0.0022438117302954197, + 0.005531781353056431, + 0.003954173997044563, + 0.0034055346623063087, + 0.004932703450322151, + 0.002823633374646306, + 0.001822780817747116, + 0.0017464417032897472, + 0.013759369030594826, + 0.01687891222536564, + 0.018161548301577568, + 0.014973092824220657, + 0.015055514872074127, + 0.018907200545072556, + 0.014751299284398556, + 0.013690859079360962, + 0.005010973196476698, + 0.0029804653022438288, + 0.004345191176980734, + 0.0049255117774009705, + 0.0026727744843810797, + 0.0025623971596360207, + 0.0029691304080188274, + 0.004770665429532528, + 0.006943557411432266, + 0.004677236545830965, + 0.008047685027122498, + 0.005107162054628134, + 0.003990679979324341, + 0.004307657480239868, + 0.004315674304962158, + 0.00355944549664855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.004097163211554289 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007176207727752626 + ], + "opposite_logprob_tokens": [ + -7.250717639923096 + ], + "logprob_reward": -0.0007176207727752626, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007176207727752626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.274187833070755, + 0.10771706700325012, + 0.07468070834875107, + 0.0541699044406414, + 0.03746059164404869, + 0.05047392100095749, + 0.14070582389831543, + 0.1541721224784851, + 0.024653473868966103, + 0.08177851140499115 + ], + "credit_map": [ + 0.024458561092615128, + 0.01518078800290823, + 0.07799534499645233, + 0.027264896780252457, + 0.01320792268961668, + 0.011854338459670544, + 0.08221514523029327, + 0.022010810673236847, + 0.00739450054243207, + 0.011770006269216537, + 0.014436200261116028, + 0.029955876991152763, + 0.011033225804567337, + 0.0067256963811814785, + 0.015839867293834686, + 0.010561673901975155, + 0.007889814674854279, + 0.00764940632507205, + 0.011086123064160347, + 0.010572191327810287, + 0.007804499007761478, + 0.008906308561563492, + 0.010409091599285603, + 0.01036327239125967, + 0.005119122564792633, + 0.0058334446512162685, + 0.010403118096292019, + 0.010175780393183231, + 0.006626483052968979, + 0.003356161992996931, + 0.007765234913676977, + 0.004890559706836939, + 0.003032150911167264, + 0.00744872260838747, + 0.0073713562451303005, + 0.0034056752920150757, + 0.0057350024580955505, + 0.002750537358224392, + 0.0037202273961156607, + 0.003996921703219414, + 0.006254374980926514, + 0.006185769569128752, + 0.005870834458619356, + 0.008500252850353718, + 0.007554448675364256, + 0.003459557890892029, + 0.006171276792883873, + 0.0064774081110954285, + 0.01048175897449255, + 0.044783011078834534, + 0.017191747203469276, + 0.03197982534766197, + 0.00927690602838993, + 0.0074230278842151165, + 0.010791169479489326, + 0.008778374642133713, + 0.017173925414681435, + 0.02048015408217907, + 0.022525332868099213, + 0.01661103405058384, + 0.018561938777565956, + 0.024896491318941116, + 0.01754453405737877, + 0.016378723084926605, + 0.002945478307083249, + 0.0034468010999262333, + 0.00315258395858109, + 0.005433234386146069, + 0.002268117154017091, + 0.0019696138333529234, + 0.002692783484235406, + 0.002744860015809536, + 0.01341488491743803, + 0.006337983999401331, + 0.016033489257097244, + 0.008952528238296509, + 0.007887955754995346, + 0.0044302791357040405, + 0.01238637138158083, + 0.012335021048784256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0007176207727752626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031010955572128 + ], + "opposite_logprob_tokens": [ + -2.3503100872039795 + ], + "logprob_reward": -0.10031010955572128, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031010955572128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09969330579042435, + 0.07053351402282715, + 0.026324288919568062, + 0.02041458524763584, + 0.03316888585686684, + 0.13203468918800354, + 0.06486020982265472, + 0.13225415349006653, + 0.11266817897558212, + 0.30804821848869324 + ], + "credit_map": [ + 0.012248162180185318, + 0.005786683410406113, + 0.0220666341483593, + 0.017285725101828575, + 0.007070915773510933, + 0.005554677452892065, + 0.01965133659541607, + 0.010029175318777561, + 0.0048414976336061954, + 0.006761899683624506, + 0.00816056877374649, + 0.01570706069469452, + 0.012818661518394947, + 0.005710889585316181, + 0.009100361727178097, + 0.007432573474943638, + 0.0030867974273860455, + 0.0031023132614791393, + 0.004777797032147646, + 0.0033798860386013985, + 0.0021157837472856045, + 0.003535348689183593, + 0.00403295224532485, + 0.002293410710990429, + 0.002255154075101018, + 0.0025723311118781567, + 0.0029269615188241005, + 0.0028843095060437918, + 0.0026927669532597065, + 0.0022284449078142643, + 0.002411394380033016, + 0.002443221863359213, + 0.002095957053825259, + 0.006686455570161343, + 0.007291883230209351, + 0.003992246463894844, + 0.004556810017675161, + 0.0024195718578994274, + 0.0031120856292545795, + 0.0030138767324388027, + 0.016532914713025093, + 0.016748616471886635, + 0.012976870872080326, + 0.02081315964460373, + 0.02512398548424244, + 0.010823588818311691, + 0.015171508304774761, + 0.01384404394775629, + 0.0051826899871230125, + 0.014390291646122932, + 0.011020823381841183, + 0.007629571948200464, + 0.008934571407735348, + 0.006210682447999716, + 0.0047291661612689495, + 0.006762409582734108, + 0.014508220367133617, + 0.01931074634194374, + 0.0190386101603508, + 0.01492125540971756, + 0.015017547644674778, + 0.020196281373500824, + 0.014590040780603886, + 0.014671457931399345, + 0.011672938242554665, + 0.009765181690454483, + 0.015084496699273586, + 0.028784817084670067, + 0.009538577869534492, + 0.013853457756340504, + 0.010743334889411926, + 0.013225383125245571, + 0.039654605090618134, + 0.030216211453080177, + 0.05778973177075386, + 0.04444868117570877, + 0.02652742527425289, + 0.020416108891367912, + 0.046215906739234924, + 0.04277956485748291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.10031010955572128 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, fixed-mounted projection shot from the hood of a rally-ready Ford RS200 captures a visceral and chaotic perspective, the world shaking and vibrating violently with every brutal movement of the machine. The car executes an aggressive, high-speed drift, sliding sideways across a narrow, sun-drenched dirt track lined with blurry green foliage. As the vehicle hangs on the edge of control, the camera jarringly captures rapid countersteering adjustments, unleashing a torrent of thick ochre dust and fine gravel from the rear wheels. This dense cloud rolls forward, streaking across the lens in thick, grainy waves, temporarily obscuring the view and creating a raw and intense display of power. Harsh sunlight penetrates the dust, illuminating the chaos and casting sharp, fleeting shadows across the rutted track, creating a gritty, high-contrast atmosphere that is both breathtaking and immersive.", + "seed": 33082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_2/index_4575_seed_33082_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05551484462339431, + "vlm_logprob": -0.05551484462339431, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.05551484462339431, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1046014279127121 + ], + "opposite_logprob_tokens": [ + -2.6046013832092285 + ], + "logprob_reward": -0.1046014279127121, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1046014279127121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15093818306922913, + 0.02296401746571064, + 0.07135236263275146, + 0.030262041836977005, + 0.041737694293260574, + 0.39837151765823364, + 0.1419697254896164, + 0.05204426869750023, + 0.048658158630132675, + 0.04170204699039459 + ], + "credit_map": [ + 0.006656632758677006, + 0.008547076024115086, + 0.005401262082159519, + 0.005805149674415588, + 0.015362144447863102, + 0.02905074879527092, + 0.03504987433552742, + 0.04506529122591019, + 0.0036696952302008867, + 0.004489653743803501, + 0.003220209851861, + 0.002274090191349387, + 0.002558498177677393, + 0.0023686266504228115, + 0.0015768867451697588, + 0.002806356642395258, + 0.0039393166080117226, + 0.004536614287644625, + 0.026202591136097908, + 0.0031056967563927174, + 0.0030639716424047947, + 0.003996361047029495, + 0.0037666545249521732, + 0.022741157561540604, + 0.003869128180667758, + 0.0032258464489132166, + 0.006098501849919558, + 0.0038704778999090195, + 0.0025459809694439173, + 0.0037446031346917152, + 0.0029152522329241037, + 0.003992252517491579, + 0.002717738738283515, + 0.002500760369002819, + 0.00429188646376133, + 0.003897185903042555, + 0.010596481151878834, + 0.003757427679374814, + 0.0072103035636246204, + 0.006765907164663076, + 0.020086703822016716, + 0.033961594104766846, + 0.07068468630313873, + 0.1254875361919403, + 0.024442611262202263, + 0.019876692444086075, + 0.08030432462692261, + 0.023527389392256737, + 0.008493021130561829, + 0.029153665527701378, + 0.014485820196568966, + 0.013859727419912815, + 0.03995346650481224, + 0.01366833969950676, + 0.013705434277653694, + 0.00865025445818901, + 0.003062852891162038, + 0.008610520511865616, + 0.00982903316617012, + 0.007974200882017612, + 0.0074727036990225315, + 0.0041609411127865314, + 0.00667819008231163, + 0.004255823325365782, + 0.0038318855222314596, + 0.006164044141769409, + 0.005522525869309902, + 0.009436454623937607, + 0.00850624404847622, + 0.00436997739598155, + 0.005539771635085344, + 0.005287255626171827, + 0.007491708267480135, + 0.005407602526247501, + 0.0039242468774318695, + 0.0045206667855381966, + 0.006816646084189415, + 0.006728114560246468, + 0.0032523355912417173, + 0.0035607272293418646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.1046014279127121 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298654228448868 + ], + "opposite_logprob_tokens": [ + -2.2379865646362305 + ], + "logprob_reward": -0.11298654228448868, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298654228448868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02537631243467331, + 0.011783070862293243, + 0.19257818162441254, + 0.05585559085011482, + 0.027534987777471542, + 0.47941967844963074, + 0.08353316783905029, + 0.04165102541446686, + 0.011131064034998417, + 0.07113693654537201 + ], + "credit_map": [ + 0.0019203791161999106, + 0.002431475557386875, + 0.0017003747634589672, + 0.0013175476342439651, + 0.0023668422363698483, + 0.003166335402056575, + 0.006644332781434059, + 0.005829023662954569, + 0.0015322866383939981, + 0.0024826982989907265, + 0.0011971715139225125, + 0.0009820801205933094, + 0.0016148913418874145, + 0.0017868391005322337, + 0.0008442088728770614, + 0.0013428955571725965, + 0.009199604392051697, + 0.007615734823048115, + 0.07207842171192169, + 0.008062553592026234, + 0.009908031672239304, + 0.016929561272263527, + 0.010869079269468784, + 0.05791519209742546, + 0.0065131052397191525, + 0.009118015877902508, + 0.006818242836743593, + 0.0046214391477406025, + 0.007506068330258131, + 0.0067049418576061726, + 0.007514507509768009, + 0.007059274706989527, + 0.0025220627430826426, + 0.0030821762047708035, + 0.0026901799719780684, + 0.003103123279288411, + 0.005100664682686329, + 0.0027689803391695023, + 0.004377053584903479, + 0.003890746273100376, + 0.021605418995022774, + 0.046891726553440094, + 0.10167501866817474, + 0.1452135294675827, + 0.028385894373059273, + 0.020926620811223984, + 0.08783310651779175, + 0.026888342574238777, + 0.005367234814912081, + 0.016603317111730576, + 0.008411774411797523, + 0.007887830026447773, + 0.022620869800448418, + 0.008403979241847992, + 0.008807139471173286, + 0.00543102715164423, + 0.002276525367051363, + 0.004367452580481768, + 0.010050601325929165, + 0.007977308705449104, + 0.0043492126278579235, + 0.0030334247276186943, + 0.005877303890883923, + 0.0037191968876868486, + 0.0009758134256117046, + 0.0012479830766096711, + 0.0013783847680315375, + 0.0014307143865153193, + 0.0016719360137358308, + 0.001884925994090736, + 0.0012527141952887177, + 0.001288592116907239, + 0.009009625762701035, + 0.009008076973259449, + 0.006087189074605703, + 0.006285317707806826, + 0.014745364896953106, + 0.012280279770493507, + 0.0045560686849057674, + 0.009165015071630478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.11298654228448868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298027191311121 + ], + "opposite_logprob_tokens": [ + -3.5298027992248535 + ], + "logprob_reward": -0.0298027191311121, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298027191311121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1885407418012619, + 0.015215003862977028, + 0.029234621673822403, + 0.025874629616737366, + 0.0963386669754982, + 0.22021107375621796, + 0.28327152132987976, + 0.081057108938694, + 0.03701434284448624, + 0.02324230596423149 + ], + "credit_map": [ + 0.005847872234880924, + 0.005455859936773777, + 0.003841093275696039, + 0.004417195916175842, + 0.02040865086019039, + 0.042624372988939285, + 0.04190301150083542, + 0.06404267996549606, + 0.001973517471924424, + 0.003523913212120533, + 0.0020948152523487806, + 0.0013581166276708245, + 0.0019318534759804606, + 0.0013126529520377517, + 0.0011852247407659888, + 0.0018349102465435863, + 0.0025339678395539522, + 0.001591463340446353, + 0.00904502347111702, + 0.0019534884486347437, + 0.0017030013259500265, + 0.0019150571897625923, + 0.0021105713676661253, + 0.008382046595215797, + 0.0035458500497043133, + 0.0048666782677173615, + 0.003140839049592614, + 0.002202230039983988, + 0.0030139593873173, + 0.003089915495365858, + 0.0034035586286336184, + 0.0026115968357771635, + 0.01371825486421585, + 0.009681380353868008, + 0.009310497902333736, + 0.006957943085581064, + 0.021103650331497192, + 0.01438905019313097, + 0.010233636945486069, + 0.010944251902401447, + 0.010313760489225388, + 0.019043629989027977, + 0.03799618035554886, + 0.06894709914922714, + 0.015198880806565285, + 0.012408844195306301, + 0.04334234818816185, + 0.012960325926542282, + 0.01807081513106823, + 0.061414141207933426, + 0.0275258868932724, + 0.026074668392539024, + 0.07945958524942398, + 0.027567747980356216, + 0.02644013799726963, + 0.01671854592859745, + 0.003569146618247032, + 0.006601360160857439, + 0.02423315867781639, + 0.02212695963680744, + 0.00747690862044692, + 0.0035938769578933716, + 0.008898431435227394, + 0.004557258915156126, + 0.003017075126990676, + 0.004262274131178856, + 0.0037964179646223783, + 0.006398577708750963, + 0.0059182168915867805, + 0.0047546145506203175, + 0.005239353980869055, + 0.0036278143525123596, + 0.002989585977047682, + 0.0029997709207236767, + 0.003890096675604582, + 0.00225236127153039, + 0.004209971055388451, + 0.002943197265267372, + 0.0016574063338339329, + 0.0022999155335128307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0298027191311121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00038926175329834223 + ], + "opposite_logprob_tokens": [ + -7.875389099121094 + ], + "logprob_reward": -0.00038926175329834223, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00038926175329834223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02298969216644764, + 0.01102504599839449, + 0.031492672860622406, + 0.014297941699624062, + 0.03577415272593498, + 0.7650665044784546, + 0.03950943425297737, + 0.05692744627594948, + 0.011824503540992737, + 0.011092539876699448 + ], + "credit_map": [ + 0.0014549505431205034, + 0.002164567355066538, + 0.0012961514294147491, + 0.001350926817394793, + 0.0025577896740287542, + 0.003493787022307515, + 0.004969299305230379, + 0.005702218506485224, + 0.0012123079504817724, + 0.002558458596467972, + 0.0012679820647463202, + 0.0008923924760892987, + 0.0012470897054299712, + 0.0018732938915491104, + 0.0008126291795633733, + 0.0011608927743509412, + 0.00143818324431777, + 0.0016350147780030966, + 0.011744985356926918, + 0.001455649035051465, + 0.0012695292243734002, + 0.0016891552368178964, + 0.002041612984612584, + 0.010218543000519276, + 0.00242801196873188, + 0.0015527961077168584, + 0.002651534741744399, + 0.0016736694378778338, + 0.001353550935164094, + 0.001282125129364431, + 0.0017529703909531236, + 0.0016032826388254762, + 0.0036832885816693306, + 0.0027248007245361805, + 0.0038371102418750525, + 0.0032615424133837223, + 0.007130982354283333, + 0.00447998009622097, + 0.006727657746523619, + 0.003928790800273418, + 0.033991482108831406, + 0.06864967197179794, + 0.13346022367477417, + 0.24509109556674957, + 0.04667358100414276, + 0.03647863492369652, + 0.15535205602645874, + 0.045369796454906464, + 0.0029968710150569677, + 0.005711355246603489, + 0.004419225268065929, + 0.004764874465763569, + 0.009167882613837719, + 0.004739066120237112, + 0.0034162465017288923, + 0.004293912556022406, + 0.0030470581259578466, + 0.006505448371171951, + 0.009608335793018341, + 0.008162343874573708, + 0.006455747876316309, + 0.005095156840980053, + 0.009984455071389675, + 0.008068903349339962, + 0.001204616972245276, + 0.0011998794507235289, + 0.001427814713679254, + 0.0018344480777159333, + 0.0018422161228954792, + 0.001730945659801364, + 0.0013514310121536255, + 0.0012331529287621379, + 0.0016010800609365106, + 0.001383423339575529, + 0.001871044049039483, + 0.0011864608386531472, + 0.0016454830765724182, + 0.001453154836781323, + 0.0008404062828049064, + 0.0011114877415820956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.00038926175329834223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ford RS200 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle from the hood of the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car performing a high-speed drift on a dirt track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick cloud of ochre dust and gravel visible rolling forward from the rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794272035360336 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794272035360336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794272035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32006704807281494, + 0.01831744611263275, + 0.10877639055252075, + 0.016820009797811508, + 0.053845882415771484, + 0.17081810534000397, + 0.058375805616378784, + 0.12457126379013062, + 0.031606391072273254, + 0.09680163115262985 + ], + "credit_map": [ + 0.007442476227879524, + 0.00684358412399888, + 0.0053022983483970165, + 0.0053505501709878445, + 0.04179757460951805, + 0.08936487883329391, + 0.05848950892686844, + 0.10547618567943573, + 0.002333532553166151, + 0.002693446120247245, + 0.0013760658912360668, + 0.0014839228242635727, + 0.0025086926762014627, + 0.003649565391242504, + 0.0021339000668376684, + 0.0021383215207606554, + 0.006289673037827015, + 0.005524586886167526, + 0.03741578012704849, + 0.005199279636144638, + 0.006405964493751526, + 0.011236106976866722, + 0.006589805707335472, + 0.030115190893411636, + 0.0022968801204115152, + 0.0022553305607289076, + 0.0011114011285826564, + 0.001248014043085277, + 0.0024376471992582083, + 0.0027125775814056396, + 0.002194135682657361, + 0.002564024180173874, + 0.0066356719471514225, + 0.005648382473737001, + 0.005507562309503555, + 0.004702634643763304, + 0.012806149199604988, + 0.007152100559324026, + 0.005378568544983864, + 0.006014819256961346, + 0.006667104084044695, + 0.017209842801094055, + 0.04451247304677963, + 0.04888172820210457, + 0.01094699464738369, + 0.006454253103584051, + 0.02636917121708393, + 0.009776536375284195, + 0.004593462683260441, + 0.013173017650842667, + 0.005562911741435528, + 0.00669021625071764, + 0.013074674643576145, + 0.005953407380729914, + 0.005417662672698498, + 0.003910452593117952, + 0.007738982327282429, + 0.01769357919692993, + 0.023685239255428314, + 0.016267206519842148, + 0.015923356637358665, + 0.012240846641361713, + 0.01714279130101204, + 0.013879274018108845, + 0.0023859157226979733, + 0.003970933146774769, + 0.0034223850816488266, + 0.004718095529824495, + 0.004628847353160381, + 0.004143218044191599, + 0.004307218361645937, + 0.004029779229313135, + 0.014399422332644463, + 0.008610374294221401, + 0.027373695746064186, + 0.010451640002429485, + 0.009421661496162415, + 0.01323657762259245, + 0.004632884170860052, + 0.008675380609929562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.029794272035360336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_tensor([5004, 5004])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_tensor([5004, 5004])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b41e3e609135258532061e70d0c352eef3dfdcde --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_tensor([5004, 5004])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, slow-motion shot, shot from an extremely low angle, focuses on a massive king crab lurking in the clear yet murky shallows. A powerful overhead light penetrates the water, precisely illuminating its deep, vibrant crimson carapace, studded with terrifyingly sharp spines that gleam with a wet, hard sheen in the light. In an instant, its curled body erupts, its long, spiky legs violently extending outward with astonishing speed and force, tearing the previously calm surface apart. In the extreme slow motion, we can clearly see that each swing of the crab's legs unleashes an explosive, circular spray of countless crystalline droplets and white foam. Each droplet, illuminated by the overhead light, gleams like diamonds, briefly suspended in mid-air. The entire image is awash in high-contrast light and shadow, the crab's red body contrasting sharply against the dark background, creating a powerful atmosphere imbued with primal power and dramatic tension.", + "seed": 32912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_5004_seed_32912_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07451346628367901, + "vlm_logprob": -0.07451346628367901, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07451346628367901, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02471480332314968 + ], + "opposite_logprob_tokens": [ + -4.149714946746826 + ], + "logprob_reward": -0.02471480332314968, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02471480332314968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1799236238002777, + 0.0574042908847332, + 0.09699022024869919, + 0.02042241021990776, + 0.07205100357532501, + 0.030448872596025467, + 0.04184173792600632, + 0.0691763311624527, + 0.38413527607917786, + 0.047606211155653 + ], + "credit_map": [ + 0.04344924911856651, + 0.019270896911621094, + 0.023485274985432625, + 0.01488051749765873, + 0.019055042415857315, + 0.01582477241754532, + 0.019299305975437164, + 0.0246585700660944, + 0.008786078542470932, + 0.00842372328042984, + 0.008223124779760838, + 0.0070330193266272545, + 0.004507027100771666, + 0.008477229624986649, + 0.005458315834403038, + 0.006495771463960409, + 0.0072353254072368145, + 0.012276078574359417, + 0.00811826717108488, + 0.01819756254553795, + 0.01426196750253439, + 0.013273208402097225, + 0.012150196358561516, + 0.011477616615593433, + 0.0028953601140528917, + 0.0025687790475785732, + 0.002743972698226571, + 0.0023227655328810215, + 0.0020576301030814648, + 0.00281389057636261, + 0.0015481298323720694, + 0.00347188301384449, + 0.005259774625301361, + 0.022728124633431435, + 0.005186018534004688, + 0.005346052814275026, + 0.005753193981945515, + 0.005495770368725061, + 0.005704895593225956, + 0.01657717302441597, + 0.0028703229036182165, + 0.003919866867363453, + 0.004332694690674543, + 0.004276159685105085, + 0.003996874205768108, + 0.0038321560714393854, + 0.0035796703305095434, + 0.0036411271430552006, + 0.011135862208902836, + 0.002422110876068473, + 0.0033501596190035343, + 0.0034112909343093634, + 0.0044846124947071075, + 0.009342608973383904, + 0.0029621876310557127, + 0.004732903558760881, + 0.00888577476143837, + 0.009335950016975403, + 0.007160870358347893, + 0.00599901657551527, + 0.011881470680236816, + 0.011959318071603775, + 0.006891755387187004, + 0.007062176708132029, + 0.054950229823589325, + 0.040814753621816635, + 0.030744416639208794, + 0.05983101204037666, + 0.04495719447731972, + 0.0449913889169693, + 0.058582454919815063, + 0.04926382750272751, + 0.009787633083760738, + 0.0057527110911905766, + 0.005412769969552755, + 0.006259650457650423, + 0.0050222608260810375, + 0.006146817468106747, + 0.004032522439956665, + 0.005191845819354057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.02471480332314968 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28137481212615967 + ], + "opposite_logprob_tokens": [ + -1.4063748121261597 + ], + "logprob_reward": -0.28137481212615967, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28137481212615967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2893602252006531, + 0.03835921362042427, + 0.06284350901842117, + 0.01906496472656727, + 0.09211738407611847, + 0.04558442160487175, + 0.03486178442835808, + 0.038820624351501465, + 0.3221992552280426, + 0.056788597255945206 + ], + "credit_map": [ + 0.058346718549728394, + 0.029705356806516647, + 0.03912559151649475, + 0.039340175688266754, + 0.027883879840373993, + 0.023650137707591057, + 0.03453710302710533, + 0.0367712527513504, + 0.004566058982163668, + 0.005528537556529045, + 0.005799074657261372, + 0.006158921401947737, + 0.0034141219221055508, + 0.003570196684449911, + 0.0039222631603479385, + 0.005400039721280336, + 0.005642282776534557, + 0.008218870498239994, + 0.008490473963320255, + 0.011637376621365547, + 0.008506648242473602, + 0.005969563964754343, + 0.007115770131349564, + 0.007262523286044598, + 0.002630852162837982, + 0.0020074413623660803, + 0.002174376044422388, + 0.002200296614319086, + 0.0021744780242443085, + 0.002759669441729784, + 0.001519154291599989, + 0.0035986965522170067, + 0.006296406500041485, + 0.0303665641695261, + 0.005625180434435606, + 0.004663383588194847, + 0.008297868072986603, + 0.005589876789599657, + 0.004489956423640251, + 0.026788150891661644, + 0.0036988502833992243, + 0.006595286540687084, + 0.0036300832871347666, + 0.00798118021339178, + 0.007813484407961369, + 0.0032521055545657873, + 0.007301967591047287, + 0.0053114634938538074, + 0.005991901271045208, + 0.004006108269095421, + 0.0030161868780851364, + 0.0034239727538079023, + 0.004628614988178015, + 0.007550224661827087, + 0.003447403199970722, + 0.0027973707765340805, + 0.004687073640525341, + 0.004234423860907555, + 0.0048880185931921005, + 0.004206491634249687, + 0.005352695006877184, + 0.0064341118559241295, + 0.0041608866304159164, + 0.004856921266764402, + 0.04253963381052017, + 0.035642098635435104, + 0.026639504358172417, + 0.04443321004509926, + 0.03577933833003044, + 0.03713049739599228, + 0.052469249814748764, + 0.047565724700689316, + 0.011843445710837841, + 0.006951800547540188, + 0.006901952903717756, + 0.007285074330866337, + 0.005733946803957224, + 0.006144016515463591, + 0.006060961168259382, + 0.005867397412657738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.28137481212615967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981579303741455 + ], + "opposite_logprob_tokens": [ + -3.529815673828125 + ], + "logprob_reward": -0.02981579303741455, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981579303741455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27536845207214355, + 0.05790325254201889, + 0.07085271179676056, + 0.03419161215424538, + 0.25769108533859253, + 0.02376130223274231, + 0.03025859035551548, + 0.0236184261739254, + 0.14655937254428864, + 0.07979515194892883 + ], + "credit_map": [ + 0.0390222892165184, + 0.022516697645187378, + 0.03952190652489662, + 0.04230983555316925, + 0.033612195402383804, + 0.023188093677163124, + 0.03990449383854866, + 0.03529295697808266, + 0.009613116271793842, + 0.00931837409734726, + 0.00903543084859848, + 0.006799029186367989, + 0.004408580716699362, + 0.006312622223049402, + 0.006594350095838308, + 0.005821751896291971, + 0.008846733719110489, + 0.009255124256014824, + 0.008075825870037079, + 0.009679282084107399, + 0.010383355431258678, + 0.005284958053380251, + 0.010440810583531857, + 0.008886633440852165, + 0.0032108230516314507, + 0.003788817673921585, + 0.004232536070048809, + 0.005291307810693979, + 0.004303935915231705, + 0.0046951668336987495, + 0.003813995746895671, + 0.004855034407228231, + 0.018488433212041855, + 0.07776304334402084, + 0.015463088639080524, + 0.014837964437901974, + 0.022494398057460785, + 0.01700524240732193, + 0.017140986397862434, + 0.07449794560670853, + 0.0019916072487831116, + 0.0026064568664878607, + 0.0023579972330480814, + 0.0037951976992189884, + 0.0027837397065013647, + 0.002156668109819293, + 0.0049715726636350155, + 0.0030980664305388927, + 0.006760201882570982, + 0.003542104037478566, + 0.002733976347371936, + 0.0024972406681627035, + 0.0025268022436648607, + 0.007803863380104303, + 0.0018320275703445077, + 0.002562376670539379, + 0.0031538770999759436, + 0.002776221837848425, + 0.002622154774144292, + 0.002154333284124732, + 0.004458881914615631, + 0.003498774953186512, + 0.0025401918683201075, + 0.0024139918386936188, + 0.022253921255469322, + 0.022793758660554886, + 0.017052998766303062, + 0.01659449189901352, + 0.015102901495993137, + 0.013782630674540997, + 0.021334143355488777, + 0.017644554376602173, + 0.019291097298264503, + 0.007876626215875149, + 0.01952298730611801, + 0.006815462373197079, + 0.007910062558948994, + 0.006786693353205919, + 0.00600940827280283, + 0.0055828215554356575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.02981579303741455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076579689979553 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.016076579689979553, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076579689979553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24298106133937836, + 0.11099164932966232, + 0.12826114892959595, + 0.02194700762629509, + 0.15706300735473633, + 0.06955257803201675, + 0.04607519134879112, + 0.030873525887727737, + 0.16576005518436432, + 0.026494765654206276 + ], + "credit_map": [ + 0.0622391514480114, + 0.02967667020857334, + 0.03079393319785595, + 0.026452913880348206, + 0.02110612578690052, + 0.02428651787340641, + 0.0202740840613842, + 0.028151672333478928, + 0.017610279843211174, + 0.016433965414762497, + 0.011754200793802738, + 0.015188438817858696, + 0.011548650451004505, + 0.01329871080815792, + 0.010095050558447838, + 0.015062358230352402, + 0.01219555176794529, + 0.01835097000002861, + 0.01450550276786089, + 0.014699431136250496, + 0.021903159096837044, + 0.010342592373490334, + 0.019762758165597916, + 0.016501184552907944, + 0.0031293865758925676, + 0.002123006386682391, + 0.002775656757876277, + 0.0031812922097742558, + 0.0019407868385314941, + 0.0033671839628368616, + 0.0021317251957952976, + 0.0032979699317365885, + 0.01060283463448286, + 0.05372532829642296, + 0.008912141434848309, + 0.007087618112564087, + 0.013580377213656902, + 0.00982464849948883, + 0.006872804369777441, + 0.04645726457238197, + 0.005694464314728975, + 0.011329714208841324, + 0.006695787888020277, + 0.01179982628673315, + 0.008945170789957047, + 0.005743023008108139, + 0.010775471106171608, + 0.008569120429456234, + 0.0043790764175355434, + 0.007887138985097408, + 0.004518707282841206, + 0.003938016481697559, + 0.011013136245310307, + 0.003465002868324518, + 0.004788520745933056, + 0.006085590459406376, + 0.004599434789270163, + 0.0037673190236091614, + 0.0025852478574961424, + 0.0027539299335330725, + 0.007201994303613901, + 0.003575998591259122, + 0.003436834318563342, + 0.002952765906229615, + 0.022357691079378128, + 0.022913699969649315, + 0.01699308678507805, + 0.017530687153339386, + 0.018189772963523865, + 0.017605813220143318, + 0.02531876228749752, + 0.024850526824593544, + 0.006204507779330015, + 0.002597830956801772, + 0.006062444299459457, + 0.002226408803835511, + 0.0025245973374694586, + 0.0027730122674256563, + 0.0020041733514517546, + 0.0021017896942794323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.016076579689979553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058534324169159 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.02058534324169159, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058534324169159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22419267892837524, + 0.2557462155818939, + 0.10742926597595215, + 0.03922191634774208, + 0.037337351590394974, + 0.038408342748880386, + 0.10253195464611053, + 0.026641374453902245, + 0.12331085652112961, + 0.04518001154065132 + ], + "credit_map": [ + 0.05355487018823624, + 0.02366412617266178, + 0.030554408207535744, + 0.026590272784233093, + 0.019020704552531242, + 0.025595109909772873, + 0.019957059994339943, + 0.025256160646677017, + 0.04914798215031624, + 0.03902793303132057, + 0.03700502589344978, + 0.02984032966196537, + 0.01990392431616783, + 0.027680031955242157, + 0.020782485604286194, + 0.032358527183532715, + 0.010114300064742565, + 0.01349787786602974, + 0.012244057841598988, + 0.01458827592432499, + 0.016184959560632706, + 0.009345847181975842, + 0.01757708750665188, + 0.013876862823963165, + 0.006440191995352507, + 0.0036318975035101175, + 0.005247711669653654, + 0.005309387110173702, + 0.003231476526707411, + 0.004978414624929428, + 0.004822792485356331, + 0.005560044664889574, + 0.0032805772498250008, + 0.006407200824469328, + 0.004971327260136604, + 0.003613880602642894, + 0.004919037688523531, + 0.0036185889039188623, + 0.003463542088866234, + 0.007063197437673807, + 0.00292521552182734, + 0.007695287931710482, + 0.003407707205042243, + 0.005092092789709568, + 0.00843028724193573, + 0.0038917302154004574, + 0.003378520254045725, + 0.003587505081668496, + 0.011914683505892754, + 0.017837822437286377, + 0.010201562196016312, + 0.009603536687791348, + 0.02084176428616047, + 0.009557796642184258, + 0.008589882403612137, + 0.013984906487166882, + 0.003448203671723604, + 0.0032513311598449945, + 0.00399860879406333, + 0.0028623612597584724, + 0.004352905321866274, + 0.003452992532402277, + 0.0029545105062425137, + 0.002320461440831423, + 0.013371983543038368, + 0.01088330801576376, + 0.013304715976119041, + 0.018664924427866936, + 0.01228289119899273, + 0.01213260181248188, + 0.021479608491063118, + 0.021190835162997246, + 0.007757780607789755, + 0.006012808997184038, + 0.004443639423698187, + 0.005279289558529854, + 0.005171767435967922, + 0.006590823642909527, + 0.005223614163696766, + 0.004700291436165571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02058534324169159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, slow-motion shot, shot from an extremely low angle, focuses on a massive king crab lurking in the clear yet murky shallows. A powerful overhead light penetrates the water, precisely illuminating its deep, vibrant crimson carapace, studded with terrifyingly sharp spines that gleam with a wet, hard sheen in the light. In an instant, its curled body erupts, its long, spiky legs violently extending outward with astonishing speed and force, tearing the previously calm surface apart. In the extreme slow motion, we can clearly see that each swing of the crab's legs unleashes an explosive, circular spray of countless crystalline droplets and white foam. Each droplet, illuminated by the overhead light, gleams like diamonds, briefly suspended in mid-air. The entire image is awash in high-contrast light and shadow, the crab's red body contrasting sharply against the dark background, creating a powerful atmosphere imbued with primal power and dramatic tension.", + "seed": 32913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_5004_seed_32913_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05795469470322132, + "vlm_logprob": -0.05795469470322132, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05795469470322132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015728231519460678 + ], + "opposite_logprob_tokens": [ + -5.140727996826172 + ], + "logprob_reward": -0.015728231519460678, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015728231519460678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11213164031505585, + 0.10777579247951508, + 0.048314232379198074, + 0.04548969492316246, + 0.2057952582836151, + 0.1582803726196289, + 0.09871267527341843, + 0.0598532073199749, + 0.11387128382921219, + 0.0497758574783802 + ], + "credit_map": [ + 0.015368271619081497, + 0.02021745778620243, + 0.01332931313663721, + 0.013326889835298061, + 0.011299741454422474, + 0.017306925728917122, + 0.012098133563995361, + 0.009184910915791988, + 0.013639607466757298, + 0.013244528323411942, + 0.01616903953254223, + 0.014436263591051102, + 0.012309488840401173, + 0.014223994687199593, + 0.015582277439534664, + 0.008170594461262226, + 0.005451696924865246, + 0.005511326715350151, + 0.0074282316491007805, + 0.004503067582845688, + 0.008807177655398846, + 0.005568644963204861, + 0.006128598004579544, + 0.0049154870212078094, + 0.007306773215532303, + 0.0034719340037554502, + 0.005752246826887131, + 0.004759063944220543, + 0.004361048340797424, + 0.004794220440089703, + 0.006974079646170139, + 0.008070326410233974, + 0.045644160360097885, + 0.02178732305765152, + 0.061591681092977524, + 0.024105241522192955, + 0.010083699598908424, + 0.01539616473019123, + 0.008800247684121132, + 0.018386732786893845, + 0.020813848823308945, + 0.014969447627663612, + 0.018345165997743607, + 0.021164527162909508, + 0.017681488767266273, + 0.0345211960375309, + 0.016410810872912407, + 0.014373881742358208, + 0.010000798851251602, + 0.017415976151823997, + 0.009983859024941921, + 0.0170483086258173, + 0.012934096157550812, + 0.012902231886982918, + 0.011790946125984192, + 0.006636461708694696, + 0.00563883688300848, + 0.004905495326966047, + 0.008495597168803215, + 0.006746095605194569, + 0.008763660676777363, + 0.007585158571600914, + 0.008596384897828102, + 0.009121979586780071, + 0.011728398501873016, + 0.014174696989357471, + 0.017503483220934868, + 0.014780058525502682, + 0.013749461621046066, + 0.01512823160737753, + 0.00955657847225666, + 0.01725037209689617, + 0.008536413311958313, + 0.005099250469356775, + 0.0071731870993971825, + 0.004359701182693243, + 0.006765392608940601, + 0.005231166258454323, + 0.006455965340137482, + 0.0061547826044261456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.015728231519460678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015838772058487 + ], + "opposite_logprob_tokens": [ + -1.7015838623046875 + ], + "logprob_reward": -0.2015838772058487, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015838772058487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11555184423923492, + 0.2491549700498581, + 0.02832646295428276, + 0.0752066969871521, + 0.13360895216464996, + 0.07226443290710449, + 0.13810324668884277, + 0.07890240848064423, + 0.06753968447446823, + 0.04134124517440796 + ], + "credit_map": [ + 0.02006509155035019, + 0.015627842396497726, + 0.01130760833621025, + 0.014058269560337067, + 0.011278617195785046, + 0.016433581709861755, + 0.016715770587325096, + 0.010065061040222645, + 0.02469693310558796, + 0.014077125117182732, + 0.06371446698904037, + 0.017207102850079536, + 0.014012055471539497, + 0.0888451412320137, + 0.011302411556243896, + 0.01529970858246088, + 0.003480121959000826, + 0.002689812798053026, + 0.0035069184377789497, + 0.004073076881468296, + 0.004415276926010847, + 0.004134645219892263, + 0.003137413412332535, + 0.0028891954571008682, + 0.01420742180198431, + 0.007718066684901714, + 0.008381116203963757, + 0.005220615770667791, + 0.006640308070927858, + 0.005782993510365486, + 0.019665632396936417, + 0.007590544875711203, + 0.010894030332565308, + 0.007473316974937916, + 0.04270793870091438, + 0.04243418946862221, + 0.005876906681805849, + 0.008257418870925903, + 0.007246043998748064, + 0.008719106204807758, + 0.013309832662343979, + 0.007017036899924278, + 0.010248792357742786, + 0.006225195247679949, + 0.00825438741594553, + 0.011732958257198334, + 0.009977994486689568, + 0.005498234182596207, + 0.014069732278585434, + 0.022153394296765327, + 0.015071185305714607, + 0.02700558304786682, + 0.017372313886880875, + 0.017021290957927704, + 0.016589805483818054, + 0.0088199432939291, + 0.007236870005726814, + 0.0067895944230258465, + 0.00855791661888361, + 0.012361029163002968, + 0.010357469320297241, + 0.011393018066883087, + 0.015009223483502865, + 0.007197290193289518, + 0.006922444794327021, + 0.01003784965723753, + 0.00686973100528121, + 0.009532777592539787, + 0.010375138372182846, + 0.010404563508927822, + 0.0061253951862454414, + 0.007271783892065287, + 0.008117029443383217, + 0.005569403525441885, + 0.005477169994264841, + 0.0031139382626861334, + 0.004217478912323713, + 0.004687990061938763, + 0.004076468292623758, + 0.00608176551759243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.2015838772058487 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190348520874977 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.018190348520874977, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190348520874977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16871963441371918, + 0.11542108654975891, + 0.05671035870909691, + 0.06568072736263275, + 0.08379250019788742, + 0.11403105407953262, + 0.07378501445055008, + 0.060365110635757446, + 0.22643180191516876, + 0.035062674432992935 + ], + "credit_map": [ + 0.02454940602183342, + 0.018712852150201797, + 0.014621001668274403, + 0.02581065706908703, + 0.017379846423864365, + 0.015079046599566936, + 0.03449267894029617, + 0.0180741585791111, + 0.015616857446730137, + 0.011373599991202354, + 0.016235407441854477, + 0.011884906329214573, + 0.009060170501470566, + 0.03331967070698738, + 0.00828737486153841, + 0.009643099270761013, + 0.008070185780525208, + 0.007181250955909491, + 0.0114363431930542, + 0.006158836185932159, + 0.006011678837239742, + 0.0072016846388578415, + 0.005923976190388203, + 0.00472640385851264, + 0.010587738826870918, + 0.003971851896494627, + 0.0071117328479886055, + 0.006730862893164158, + 0.005678585264831781, + 0.007682146038860083, + 0.01607012189924717, + 0.007847685366868973, + 0.008692389354109764, + 0.005379346199333668, + 0.022151146084070206, + 0.019789623096585274, + 0.004057419020682573, + 0.009590023197233677, + 0.00435660732910037, + 0.009775955229997635, + 0.020109914243221283, + 0.013266489841043949, + 0.010245658457279205, + 0.01710987649857998, + 0.010153436101973057, + 0.016454732045531273, + 0.015581933781504631, + 0.011109007522463799, + 0.007790425326675177, + 0.011519481427967548, + 0.0065496088936924934, + 0.01145702414214611, + 0.010764422826468945, + 0.01233319379389286, + 0.008223296143114567, + 0.0051475572399795055, + 0.006340929307043552, + 0.008619208820164204, + 0.010661480948328972, + 0.006547538097947836, + 0.00846166629344225, + 0.007863578386604786, + 0.006796976551413536, + 0.005073735490441322, + 0.019539646804332733, + 0.0461256317794323, + 0.02761288918554783, + 0.02826983667910099, + 0.028863968327641487, + 0.02699374221265316, + 0.022802064195275307, + 0.026224026456475258, + 0.00450644688680768, + 0.005447150208055973, + 0.003161374246701598, + 0.0037433530669659376, + 0.00634717708453536, + 0.00446365587413311, + 0.004579305183142424, + 0.0028142111841589212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018190348520874977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369314596056938 + ], + "opposite_logprob_tokens": [ + -3.4086930751800537 + ], + "logprob_reward": -0.03369314596056938, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369314596056938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09130104631185532, + 0.18292821943759918, + 0.034689806401729584, + 0.06166144087910652, + 0.06387938559055328, + 0.07210078835487366, + 0.10199224948883057, + 0.04756637290120125, + 0.2842444181442261, + 0.05963626131415367 + ], + "credit_map": [ + 0.013193028047680855, + 0.011637571267783642, + 0.00781674962490797, + 0.012516481801867485, + 0.010088804177939892, + 0.008424830622971058, + 0.017946192994713783, + 0.00967739149928093, + 0.019702106714248657, + 0.01374409906566143, + 0.0375562347471714, + 0.010010859929025173, + 0.010839089751243591, + 0.06686101108789444, + 0.010770481079816818, + 0.013444325886666775, + 0.004541175439953804, + 0.0037932414561510086, + 0.0047694481909275055, + 0.004677674733102322, + 0.004540854599326849, + 0.00408856850117445, + 0.00502427713945508, + 0.0032545647118240595, + 0.0069215125404298306, + 0.0038318235892802477, + 0.007855844683945179, + 0.00832343939691782, + 0.005581604316830635, + 0.0076630995608866215, + 0.012582049705088139, + 0.00890206266194582, + 0.009834585711359978, + 0.004493000917136669, + 0.007441890425980091, + 0.018654564395546913, + 0.003354124492034316, + 0.006328735500574112, + 0.0047700596041977406, + 0.009002426639199257, + 0.009717445820569992, + 0.008475502952933311, + 0.006313173100352287, + 0.009648246690630913, + 0.00885684322565794, + 0.010284201242029667, + 0.010814579203724861, + 0.007990792393684387, + 0.01247198786586523, + 0.010902863927185535, + 0.013488742522895336, + 0.01434303168207407, + 0.015515615232288837, + 0.01562660001218319, + 0.010648595169186592, + 0.008994805626571178, + 0.004237305372953415, + 0.0071152495220303535, + 0.0072951591573655605, + 0.005468752700835466, + 0.006134308408945799, + 0.005237819626927376, + 0.0069520315155386925, + 0.005125745665282011, + 0.02568577602505684, + 0.0439847856760025, + 0.03817058727145195, + 0.03501320630311966, + 0.04189673066139221, + 0.03830381855368614, + 0.026157507672905922, + 0.03503202646970749, + 0.008467611856758595, + 0.008121604099869728, + 0.007551560644060373, + 0.006314864847809076, + 0.007144366856664419, + 0.007077832240611315, + 0.005186776630580425, + 0.009771646000444889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.03369314596056938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14032399654388428, + 0.21338269114494324, + 0.11513746529817581, + 0.0799328163266182, + 0.10580845177173615, + 0.11315364390611649, + 0.05369725078344345, + 0.0338369719684124, + 0.10037047415971756, + 0.04435630515217781 + ], + "credit_map": [ + 0.023746762424707413, + 0.020548364147543907, + 0.014676948077976704, + 0.01563008688390255, + 0.012589994817972183, + 0.024879343807697296, + 0.01592187210917473, + 0.012330599129199982, + 0.022560114040970802, + 0.013077860698103905, + 0.04613610357046127, + 0.012696817517280579, + 0.012033121660351753, + 0.08059258759021759, + 0.011596325784921646, + 0.014689736068248749, + 0.014449410140514374, + 0.012130849994719028, + 0.02269781194627285, + 0.012463286519050598, + 0.01860434003174305, + 0.011554164811968803, + 0.011777056381106377, + 0.011460529640316963, + 0.013582340441644192, + 0.004471217282116413, + 0.008269001729786396, + 0.005901522468775511, + 0.006604820489883423, + 0.007297971751540899, + 0.025681840255856514, + 0.008124091662466526, + 0.013135611079633236, + 0.004727903753519058, + 0.027254415675997734, + 0.02375078573822975, + 0.004504946060478687, + 0.009146776981651783, + 0.0060356087051332, + 0.017252394929528236, + 0.02710171975195408, + 0.01306881383061409, + 0.010975483804941177, + 0.01436080876737833, + 0.007462203036993742, + 0.01753750443458557, + 0.011599227786064148, + 0.011047872714698315, + 0.00787424761801958, + 0.007681217975914478, + 0.005438140593469143, + 0.007279268465936184, + 0.006257748231291771, + 0.008024043403565884, + 0.006964277941733599, + 0.004178300499916077, + 0.00376828177832067, + 0.004630527924746275, + 0.0038756798021495342, + 0.0043760365806519985, + 0.004350794479250908, + 0.0042104232124984264, + 0.004922123625874519, + 0.003703099675476551, + 0.009635621681809425, + 0.016328390687704086, + 0.011527455411851406, + 0.009175816550850868, + 0.01391550898551941, + 0.011976374313235283, + 0.010336071252822876, + 0.017475226894021034, + 0.006835172884166241, + 0.0069752223789691925, + 0.005710532423108816, + 0.0044335187412798405, + 0.005292608868330717, + 0.004997242242097855, + 0.0035588492173701525, + 0.006553151179105043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.020577870309352875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, slow-motion shot, shot from an extremely low angle, focuses on a massive king crab lurking in the clear yet murky shallows. A powerful overhead light penetrates the water, precisely illuminating its deep, vibrant crimson carapace, studded with terrifyingly sharp spines that gleam with a wet, hard sheen in the light. In an instant, its curled body erupts, its long, spiky legs violently extending outward with astonishing speed and force, tearing the previously calm surface apart. In the extreme slow motion, we can clearly see that each swing of the crab's legs unleashes an explosive, circular spray of countless crystalline droplets and white foam. Each droplet, illuminated by the overhead light, gleams like diamonds, briefly suspended in mid-air. The entire image is awash in high-contrast light and shadow, the crab's red body contrasting sharply against the dark background, creating a powerful atmosphere imbued with primal power and dramatic tension.", + "seed": 32914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_5004_seed_32914_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09930944219231605, + "vlm_logprob": -0.09930944219231605, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.09930944219231605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02654634788632393 + ], + "opposite_logprob_tokens": [ + -3.9015462398529053 + ], + "logprob_reward": -0.02654634788632393, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02654634788632393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2457960844039917, + 0.05225186422467232, + 0.06809229403734207, + 0.10771305114030838, + 0.08212598413228989, + 0.0870957151055336, + 0.07762467116117477, + 0.09788558632135391, + 0.028165439143776894, + 0.15324930846691132 + ], + "credit_map": [ + 0.019741471856832504, + 0.04243805259466171, + 0.05111088976264, + 0.0473075695335865, + 0.018205426633358, + 0.02077852003276348, + 0.029980363324284554, + 0.016233788803219795, + 0.005086343735456467, + 0.00963230524212122, + 0.0074747297912836075, + 0.007532303687185049, + 0.008846394717693329, + 0.005394416395574808, + 0.004097708035260439, + 0.004187659826129675, + 0.008577863685786724, + 0.01037092600017786, + 0.012043521739542484, + 0.008149705827236176, + 0.009935098700225353, + 0.0077758664265275, + 0.006125178188085556, + 0.005114135332405567, + 0.01163963507860899, + 0.008856276981532574, + 0.012679879553616047, + 0.01189772505313158, + 0.014960099011659622, + 0.016139518469572067, + 0.010964236222207546, + 0.02057567983865738, + 0.013233911246061325, + 0.00917764287441969, + 0.005341473035514355, + 0.007341204211115837, + 0.009949074126780033, + 0.006613615434616804, + 0.025051051750779152, + 0.005418006330728531, + 0.010954846628010273, + 0.013153606094419956, + 0.01194162480533123, + 0.008692223578691483, + 0.008568537421524525, + 0.010405528359115124, + 0.007632470689713955, + 0.015746884047985077, + 0.01061328686773777, + 0.013891692273318768, + 0.006967159919440746, + 0.010049276985228062, + 0.009745810180902481, + 0.009132560342550278, + 0.010483216494321823, + 0.006741666700690985, + 0.009938009083271027, + 0.012313898652791977, + 0.027116121724247932, + 0.007023127283900976, + 0.0073829456232488155, + 0.009168560616672039, + 0.01548313070088625, + 0.009459796361625195, + 0.007168300449848175, + 0.0018749696901068091, + 0.002677241573110223, + 0.0034334559459239244, + 0.0017237395513802767, + 0.002074215095490217, + 0.00262280972674489, + 0.0065907081589102745, + 0.017050564289093018, + 0.019269052892923355, + 0.011750555597245693, + 0.013576384633779526, + 0.023368798196315765, + 0.01931239478290081, + 0.02811371348798275, + 0.020807845517992973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02654634788632393 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34867608547210693 + ], + "opposite_logprob_tokens": [ + -1.223676085472107 + ], + "logprob_reward": -0.34867608547210693, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34867608547210693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20463448762893677, + 0.034422844648361206, + 0.06085577607154846, + 0.10031849890947342, + 0.14365284144878387, + 0.08014976233243942, + 0.05200835317373276, + 0.09638351202011108, + 0.11026358604431152, + 0.11731039732694626 + ], + "credit_map": [ + 0.016938341781497, + 0.04577212780714035, + 0.03210095688700676, + 0.04337150976061821, + 0.014248634688556194, + 0.013820535503327847, + 0.0266281645745039, + 0.011754190549254417, + 0.004076345823705196, + 0.005014993716031313, + 0.0064215813763439655, + 0.004041105043143034, + 0.005468753166496754, + 0.003974755760282278, + 0.002767694415524602, + 0.0026576148811727762, + 0.010423813946545124, + 0.009131642058491707, + 0.007931120693683624, + 0.006286024581640959, + 0.006928765680640936, + 0.008839000947773457, + 0.006356448400765657, + 0.0049589574337005615, + 0.010885193012654781, + 0.008445290848612785, + 0.015117857605218887, + 0.01087899785488844, + 0.012419331818819046, + 0.012053941376507282, + 0.011348338797688484, + 0.01916954293847084, + 0.027109358459711075, + 0.008047857321798801, + 0.009142618626356125, + 0.012341710738837719, + 0.011410790495574474, + 0.009619131684303284, + 0.054703082889318466, + 0.011278278194367886, + 0.01700981706380844, + 0.00680913869291544, + 0.02076890878379345, + 0.006218237336724997, + 0.00795501098036766, + 0.008330666460096836, + 0.005671328864991665, + 0.007386647630482912, + 0.007366012316197157, + 0.00988416001200676, + 0.004266414791345596, + 0.006951003801077604, + 0.0050547257997095585, + 0.005513460841029882, + 0.008257472887635231, + 0.004715099930763245, + 0.009181728586554527, + 0.008555099368095398, + 0.03103601187467575, + 0.009111958555877209, + 0.005870485678315163, + 0.006877911277115345, + 0.017285121604800224, + 0.008465195074677467, + 0.025149989873170853, + 0.008333847858011723, + 0.010293342173099518, + 0.012588335201144218, + 0.00705491378903389, + 0.007310314103960991, + 0.00917922705411911, + 0.0303536057472229, + 0.015472703613340855, + 0.01589699275791645, + 0.011485780589282513, + 0.010726160369813442, + 0.014609242789447308, + 0.013983340002596378, + 0.02116374485194683, + 0.013972426764667034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.34867608547210693 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496964231133461 + ], + "opposite_logprob_tokens": [ + -2.929969549179077 + ], + "logprob_reward": -0.05496964231133461, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496964231133461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1085391640663147, + 0.05895840376615524, + 0.053223345428705215, + 0.08205235749483109, + 0.17551353573799133, + 0.08279212564229965, + 0.08742357790470123, + 0.07245118170976639, + 0.031936295330524445, + 0.2471100240945816 + ], + "credit_map": [ + 0.009857122786343098, + 0.014461805112659931, + 0.01466965489089489, + 0.019113007932901382, + 0.008697464130818844, + 0.015025033615529537, + 0.0146884610876441, + 0.012026607990264893, + 0.00705792848020792, + 0.01028300728648901, + 0.007790255825966597, + 0.007156545761972666, + 0.00852486677467823, + 0.0049538626335561275, + 0.0070645916275680065, + 0.006127344910055399, + 0.006033970974385738, + 0.012952137738466263, + 0.006662726867944002, + 0.0059489174745976925, + 0.007150798104703426, + 0.005113631021231413, + 0.005085143726319075, + 0.004276013933122158, + 0.010039938613772392, + 0.00793056283146143, + 0.01055862195789814, + 0.00886837299913168, + 0.011403248645365238, + 0.008950422517955303, + 0.011549734510481358, + 0.012751450762152672, + 0.030918361619114876, + 0.01671423390507698, + 0.011963838711380959, + 0.01606033928692341, + 0.01930522546172142, + 0.01573636755347252, + 0.052511513233184814, + 0.012303653173148632, + 0.011075848713517189, + 0.013856861740350723, + 0.0075432308949530125, + 0.009567617438733578, + 0.012918099761009216, + 0.010736296884715557, + 0.008522441610693932, + 0.008571730926632881, + 0.011951371096074581, + 0.011090502142906189, + 0.008513073436915874, + 0.012630333192646503, + 0.010424562729895115, + 0.01102316752076149, + 0.015271469950675964, + 0.006519095506519079, + 0.01336062140762806, + 0.011711722239851952, + 0.006581043358892202, + 0.006501794792711735, + 0.009969694539904594, + 0.009602895006537437, + 0.0072912429459393024, + 0.007432166952639818, + 0.007268575951457024, + 0.003269987413659692, + 0.0025866515934467316, + 0.003353105392307043, + 0.002142778132110834, + 0.0023439673241227865, + 0.003005739999935031, + 0.00796548742800951, + 0.027495354413986206, + 0.03061775304377079, + 0.025415215641260147, + 0.020768465474247932, + 0.03181278333067894, + 0.03137001022696495, + 0.04775569587945938, + 0.031874727457761765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.05496964231133461 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294363170862198 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.023294363170862198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294363170862198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17103825509548187, + 0.039334576576948166, + 0.10412811487913132, + 0.14528071880340576, + 0.06552863866090775, + 0.0819069966673851, + 0.09147442132234573, + 0.19035612046718597, + 0.06492793560028076, + 0.04602415859699249 + ], + "credit_map": [ + 0.013654346577823162, + 0.026920434087514877, + 0.038809098303318024, + 0.02686294913291931, + 0.013268057256937027, + 0.01590265892446041, + 0.022679343819618225, + 0.012941369786858559, + 0.0035647095646709204, + 0.0080519188195467, + 0.005152631085366011, + 0.006404675543308258, + 0.0050137834623456, + 0.0038035104516893625, + 0.004981537349522114, + 0.0023618119303137064, + 0.017070701345801353, + 0.018428288400173187, + 0.0166185162961483, + 0.011492989957332611, + 0.01316646859049797, + 0.011986899189651012, + 0.007444603834301233, + 0.007919649593532085, + 0.01612822338938713, + 0.013061460107564926, + 0.016284875571727753, + 0.014501592144370079, + 0.022032538428902626, + 0.025401119142770767, + 0.015650542452931404, + 0.022220367565751076, + 0.012137829326093197, + 0.0036975450348109007, + 0.004893288016319275, + 0.006111849565058947, + 0.003310564672574401, + 0.004367392510175705, + 0.025380555540323257, + 0.005629612598568201, + 0.014211110770702362, + 0.011912793852388859, + 0.016130855306982994, + 0.006200328003615141, + 0.008401153609156609, + 0.011398130096495152, + 0.007133248262107372, + 0.0065193818882107735, + 0.011445262469351292, + 0.01562531478703022, + 0.00910790916532278, + 0.014288021251559258, + 0.012864988297224045, + 0.010116652585566044, + 0.010340909473598003, + 0.007685366086661816, + 0.025019431486725807, + 0.026199663057923317, + 0.0530121885240078, + 0.012382418848574162, + 0.016812367364764214, + 0.02186276577413082, + 0.022615225985646248, + 0.012452064082026482, + 0.013805890455842018, + 0.0061768717132508755, + 0.005774762947112322, + 0.008107905276119709, + 0.004005541559308767, + 0.005162484012544155, + 0.006396486423909664, + 0.015497994609177113, + 0.006752231623977423, + 0.005675467662513256, + 0.006219508126378059, + 0.0041002025827765465, + 0.00582417007535696, + 0.004828232806175947, + 0.005760540254414082, + 0.006863804534077644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.023294363170862198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043060772120952606 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.043060772120952606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043060772120952606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25873371958732605, + 0.04576527327299118, + 0.16962125897407532, + 0.11847038567066193, + 0.042530909180641174, + 0.059813469648361206, + 0.13676126301288605, + 0.09740737080574036, + 0.03343683108687401, + 0.03745950385928154 + ], + "credit_map": [ + 0.01855069398880005, + 0.07132184505462646, + 0.04835939779877663, + 0.042875878512859344, + 0.01769811101257801, + 0.017726538702845573, + 0.029598798602819443, + 0.012602460570633411, + 0.006107062567025423, + 0.007538015954196453, + 0.006198483984917402, + 0.005505228880792856, + 0.006788663100451231, + 0.0048805903643369675, + 0.004212990403175354, + 0.004534240812063217, + 0.02511555701494217, + 0.03002762421965599, + 0.02876453660428524, + 0.018478980287909508, + 0.020239168778061867, + 0.019711967557668686, + 0.013364660553634167, + 0.013918766751885414, + 0.012811464257538319, + 0.010565589182078838, + 0.01864437945187092, + 0.015850527212023735, + 0.012011246755719185, + 0.012334847822785378, + 0.0154817970469594, + 0.02077052928507328, + 0.005510734394192696, + 0.0036149579100310802, + 0.0032224850729107857, + 0.006154634989798069, + 0.005041897762566805, + 0.0032253172248601913, + 0.011576232500374317, + 0.004184650257229805, + 0.008450321853160858, + 0.010849078185856342, + 0.00998799316585064, + 0.0059494092129170895, + 0.0077149998396635056, + 0.00716796051710844, + 0.005710300989449024, + 0.003983409609645605, + 0.016016241163015366, + 0.026653971523046494, + 0.012320022098720074, + 0.021092625334858894, + 0.018906397745013237, + 0.014706781134009361, + 0.01826431229710579, + 0.00880090519785881, + 0.011028336361050606, + 0.01071828044950962, + 0.02802256867289543, + 0.0064537422731518745, + 0.006247709039598703, + 0.006456807721406221, + 0.01819184049963951, + 0.010288082994520664, + 0.0048898011445999146, + 0.003219153732061386, + 0.004030612763017416, + 0.007018185220658779, + 0.0023492120672017336, + 0.0023705174680799246, + 0.00464426726102829, + 0.004915082361549139, + 0.004548555705696344, + 0.005867912899702787, + 0.004115232266485691, + 0.005039316602051258, + 0.006194562651216984, + 0.0037847973871976137, + 0.0036217819433659315, + 0.004287341609597206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.043060772120952606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, slow-motion shot, shot from an extremely low angle, focuses on a massive king crab lurking in the clear yet murky shallows. A powerful overhead light penetrates the water, precisely illuminating its deep, vibrant crimson carapace, studded with terrifyingly sharp spines that gleam with a wet, hard sheen in the light. In an instant, its curled body erupts, its long, spiky legs violently extending outward with astonishing speed and force, tearing the previously calm surface apart. In the extreme slow motion, we can clearly see that each swing of the crab's legs unleashes an explosive, circular spray of countless crystalline droplets and white foam. Each droplet, illuminated by the overhead light, gleams like diamonds, briefly suspended in mid-air. The entire image is awash in high-contrast light and shadow, the crab's red body contrasting sharply against the dark background, creating a powerful atmosphere imbued with primal power and dramatic tension.", + "seed": 32915, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_3/index_5004_seed_32915_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09390133060514927, + "vlm_logprob": -0.09390133060514927, + "vlm_yes_no_margin": 3.3999999880790712, + "vlm_token_credit": -0.09390133060514927, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017286935821175575 + ], + "opposite_logprob_tokens": [ + -4.767286777496338 + ], + "logprob_reward": -0.017286935821175575, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017286935821175575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15778475999832153, + 0.05933474749326706, + 0.31761878728866577, + 0.07633421570062637, + 0.10713747143745422, + 0.09347618371248245, + 0.08114605396986008, + 0.04242939129471779, + 0.026523690670728683, + 0.03821476176381111 + ], + "credit_map": [ + 0.019695911556482315, + 0.024362970143556595, + 0.017740523442626, + 0.017504500225186348, + 0.024968422949314117, + 0.017192628234624863, + 0.023255445063114166, + 0.013064361177384853, + 0.00801379606127739, + 0.005472683347761631, + 0.006551021710038185, + 0.007979187183082104, + 0.00918186828494072, + 0.006752322427928448, + 0.005640997085720301, + 0.00974287185817957, + 0.04089374467730522, + 0.016496924683451653, + 0.01138807088136673, + 0.011804740875959396, + 0.11509035527706146, + 0.05861653760075569, + 0.034284647554159164, + 0.02904375270009041, + 0.006125624757260084, + 0.009674496948719025, + 0.010345102287828922, + 0.011271815747022629, + 0.013657055795192719, + 0.009493628516793251, + 0.0067060342989861965, + 0.00906046386808157, + 0.012351753190159798, + 0.017871243879199028, + 0.020147714763879776, + 0.015636136755347252, + 0.010903562419116497, + 0.01039051078259945, + 0.011748428456485271, + 0.008088120259344578, + 0.007721239235252142, + 0.015842175111174583, + 0.009183548390865326, + 0.015425615012645721, + 0.014860141091048717, + 0.0109385522082448, + 0.010692707262933254, + 0.008812207728624344, + 0.011781350709497929, + 0.008581708185374737, + 0.009442172944545746, + 0.008044995367527008, + 0.010685796849429607, + 0.00994506198912859, + 0.010092637501657009, + 0.0125723322853446, + 0.005135802552103996, + 0.0038785780780017376, + 0.0068492828868329525, + 0.005634173285216093, + 0.0072161355055868626, + 0.005298822186887264, + 0.0051818722859025, + 0.0032347224187105894, + 0.0031191837042570114, + 0.0030095623806118965, + 0.0033643878996372223, + 0.0040023294277489185, + 0.0030499473214149475, + 0.003618112299591303, + 0.0033537556882947683, + 0.0030064117163419724, + 0.004243888892233372, + 0.004579526372253895, + 0.006446255370974541, + 0.003941815346479416, + 0.005185691174119711, + 0.0053533217869699, + 0.005137474276125431, + 0.0033267864491790533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.017286935821175575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709956407546997 + ], + "opposite_logprob_tokens": [ + -1.1370995044708252 + ], + "logprob_reward": -0.38709956407546997, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38709956407546997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19744569063186646, + 0.04823289066553116, + 0.31526997685432434, + 0.10359560698270798, + 0.043908972293138504, + 0.08568768203258514, + 0.07267998158931732, + 0.053650349378585815, + 0.056483447551727295, + 0.023045450448989868 + ], + "credit_map": [ + 0.028295204043388367, + 0.030510980635881424, + 0.017049111425876617, + 0.01618371531367302, + 0.0352487675845623, + 0.02278452180325985, + 0.029180394485592842, + 0.01819298416376114, + 0.005326485726982355, + 0.009764502756297588, + 0.00533510185778141, + 0.0062248301692306995, + 0.006811758037656546, + 0.006493510212749243, + 0.0047082845121622086, + 0.003568419720977545, + 0.04075070098042488, + 0.016095122322440147, + 0.011848942376673222, + 0.015527908690273762, + 0.11167746037244797, + 0.056092578917741776, + 0.03594299033284187, + 0.02733428031206131, + 0.00999477319419384, + 0.011281047016382217, + 0.02374870888888836, + 0.014959840103983879, + 0.012947222217917442, + 0.009872623719274998, + 0.009042437188327312, + 0.011748949997127056, + 0.0037577864713966846, + 0.00643672002479434, + 0.006714269518852234, + 0.004753436427563429, + 0.007331749889999628, + 0.004433510359376669, + 0.005086830817162991, + 0.005394669249653816, + 0.007214849814772606, + 0.013669162057340145, + 0.010832319967448711, + 0.015121166594326496, + 0.009031313471496105, + 0.00890852976590395, + 0.009710422717034817, + 0.011199915781617165, + 0.011450712569057941, + 0.006445346400141716, + 0.008662788197398186, + 0.008888531476259232, + 0.010402983985841274, + 0.008923564106225967, + 0.009033732116222382, + 0.008872319012880325, + 0.005338314455002546, + 0.00402171490713954, + 0.0045822071842849255, + 0.00938817486166954, + 0.009056110866367817, + 0.00659333448857069, + 0.008838634938001633, + 0.005831860937178135, + 0.007737753447145224, + 0.010508142411708832, + 0.006511983927339315, + 0.006926954258233309, + 0.006099128630012274, + 0.0065618292428553104, + 0.004983745981007814, + 0.007153911516070366, + 0.002658433048054576, + 0.0023345910012722015, + 0.0033215645235031843, + 0.0024668173864483833, + 0.0029428708367049694, + 0.002876251470297575, + 0.003727263305336237, + 0.0027176577132195234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.38709956407546997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3589540123939514, + 0.0792669802904129, + 0.24346376955509186, + 0.10400102287530899, + 0.058666110038757324, + 0.05824880674481392, + 0.024648156017065048, + 0.01838657073676586, + 0.022945918142795563, + 0.03141865134239197 + ], + "credit_map": [ + 0.0555184967815876, + 0.05847873538732529, + 0.029521867632865906, + 0.029729070141911507, + 0.06493284553289413, + 0.04215347394347191, + 0.05153347924351692, + 0.027086041867733, + 0.013790174387395382, + 0.012110138311982155, + 0.008733730763196945, + 0.007803647313266993, + 0.009585841558873653, + 0.011004348285496235, + 0.007795306388288736, + 0.008443793281912804, + 0.033529605716466904, + 0.014276980422437191, + 0.009485969319939613, + 0.009504773654043674, + 0.08657709509134293, + 0.0474344938993454, + 0.01665257103741169, + 0.02600228041410446, + 0.012731523253023624, + 0.015144293196499348, + 0.017876463010907173, + 0.014529514126479626, + 0.017163541167974472, + 0.01082430500537157, + 0.008292417973279953, + 0.007438967004418373, + 0.0070842900313436985, + 0.008186203427612782, + 0.014470905996859074, + 0.007108733523637056, + 0.006107944995164871, + 0.0067204674705863, + 0.005185112822800875, + 0.0038024508394300938, + 0.0068316892720758915, + 0.006942104548215866, + 0.006269620265811682, + 0.008632272481918335, + 0.008791781961917877, + 0.0067882719449698925, + 0.008278680965304375, + 0.005714383441954851, + 0.0031075964216142893, + 0.002879661973565817, + 0.0027969074435532093, + 0.00269410596229136, + 0.004065383691340685, + 0.0027766264975070953, + 0.0028570417780429125, + 0.00347083224914968, + 0.0019315221579745412, + 0.0019086140673607588, + 0.00213007815182209, + 0.0035616063978523016, + 0.002430794294923544, + 0.0021172601263970137, + 0.0018818845273926854, + 0.002424809616059065, + 0.0018755451310425997, + 0.002270003315061331, + 0.00215931492857635, + 0.0031506074592471123, + 0.00325224781408906, + 0.0030867764726281166, + 0.0041225808672606945, + 0.003028844716027379, + 0.005452269222587347, + 0.0034500041510909796, + 0.003856925293803215, + 0.002937727142125368, + 0.004888936877250671, + 0.004103677347302437, + 0.003939516842365265, + 0.0027895981911569834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.018195616081357002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345033198595047 + ], + "opposite_logprob_tokens": [ + -3.6513450145721436 + ], + "logprob_reward": -0.026345033198595047, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345033198595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22795532643795013, + 0.10072650015354156, + 0.0524846687912941, + 0.1738719493150711, + 0.03786987066268921, + 0.16310951113700867, + 0.10347318649291992, + 0.05463789775967598, + 0.061596181243658066, + 0.024274922907352448 + ], + "credit_map": [ + 0.025341706350445747, + 0.034242380410432816, + 0.03128620237112045, + 0.027135733515024185, + 0.022744225338101387, + 0.023216886445879936, + 0.04116938263177872, + 0.022818800061941147, + 0.01212457288056612, + 0.013921024277806282, + 0.010816751047968864, + 0.00970631092786789, + 0.016793234273791313, + 0.017008930444717407, + 0.007166952360421419, + 0.01318871695548296, + 0.005455706734210253, + 0.004245723132044077, + 0.005903424229472876, + 0.004039519000798464, + 0.004563144873827696, + 0.007997975684702396, + 0.014321609400212765, + 0.005957558751106262, + 0.020039713010191917, + 0.022406181320548058, + 0.02927980199456215, + 0.02301866188645363, + 0.021418672055006027, + 0.01768878847360611, + 0.019005175679922104, + 0.021014928817749023, + 0.004289905074983835, + 0.007508113048970699, + 0.004163587000221014, + 0.005582387559115887, + 0.00430679926648736, + 0.004072811920195818, + 0.004210048355162144, + 0.0037362186703830957, + 0.016787156462669373, + 0.01878211833536625, + 0.013761982321739197, + 0.028195181861519814, + 0.01998024247586727, + 0.019231552258133888, + 0.022772079333662987, + 0.02359919250011444, + 0.017962388694286346, + 0.009249867871403694, + 0.012432195246219635, + 0.014309746213257313, + 0.010335558094084263, + 0.013577554374933243, + 0.012173349969089031, + 0.0134325185790658, + 0.005769709590822458, + 0.0051139905117452145, + 0.005874292925000191, + 0.00939121376723051, + 0.004942784551531076, + 0.007518694270402193, + 0.009586822241544724, + 0.006440388970077038, + 0.005455529782921076, + 0.007283289451152086, + 0.009999264031648636, + 0.009964276105165482, + 0.006421554367989302, + 0.007705225143581629, + 0.007693861611187458, + 0.0070731802843511105, + 0.002525337738916278, + 0.002722235396504402, + 0.003306498285382986, + 0.0024616485461592674, + 0.0035141888074576855, + 0.00324136670678854, + 0.003000594675540924, + 0.00350305181927979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.026345033198595047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a massive king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace deep crimson and studded with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab erupt with its legs extending outward violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion showing droplets suspended in mid-air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057950384914875 + ], + "opposite_logprob_tokens": [ + -3.8955795764923096 + ], + "logprob_reward": -0.02057950384914875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057950384914875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3531859219074249, + 0.12914584577083588, + 0.083832748234272, + 0.0653093010187149, + 0.05288122594356537, + 0.09998562932014465, + 0.0399722158908844, + 0.059543441981077194, + 0.07493080943822861, + 0.04121285304427147 + ], + "credit_map": [ + 0.05742115154862404, + 0.0524674616754055, + 0.030176423490047455, + 0.029149049893021584, + 0.06736211478710175, + 0.043176233768463135, + 0.047965601086616516, + 0.0254678912460804, + 0.01675013080239296, + 0.01510617695748806, + 0.013107672333717346, + 0.013383561745285988, + 0.02209937572479248, + 0.021567795425653458, + 0.010438966564834118, + 0.016692161560058594, + 0.008845851756632328, + 0.005741414148360491, + 0.004850269295275211, + 0.006094332784414291, + 0.026064136996865273, + 0.016064610332250595, + 0.007828980684280396, + 0.008343154564499855, + 0.005180843640118837, + 0.009920673444867134, + 0.01240515150129795, + 0.01430053636431694, + 0.006306202616542578, + 0.006620722822844982, + 0.00612981989979744, + 0.004445357248187065, + 0.007728054188191891, + 0.0075925188139081, + 0.008400710299611092, + 0.008121293038129807, + 0.005733011290431023, + 0.005523954052478075, + 0.005337714683264494, + 0.004443970043212175, + 0.009773313999176025, + 0.013930588960647583, + 0.010187827050685883, + 0.01470769103616476, + 0.0164380744099617, + 0.01025928370654583, + 0.010607735253870487, + 0.014081109315156937, + 0.008940408937633038, + 0.004993026610463858, + 0.003905039979144931, + 0.0034798041451722383, + 0.004868709482252598, + 0.005118564702570438, + 0.0037146916147321463, + 0.004951970651745796, + 0.00674083735793829, + 0.006219167727977037, + 0.007403114810585976, + 0.009172574616968632, + 0.008983614854514599, + 0.006504592951387167, + 0.008347284980118275, + 0.006172253750264645, + 0.011099073104560375, + 0.014972819946706295, + 0.007690466474741697, + 0.00812895130366087, + 0.008737366646528244, + 0.008852578699588776, + 0.007102323230355978, + 0.008347230032086372, + 0.006334658246487379, + 0.004332179203629494, + 0.006788220722228289, + 0.004200298339128494, + 0.004469543229788542, + 0.0052336580120027065, + 0.004748284351080656, + 0.005106011871248484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02057950384914875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_tensor([7287, 7287])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_tensor([7287, 7287])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5cd25e46a03de68fde5c6416c147129f58c09384 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_tensor([7287, 7287])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot in super slow motion begins at eye level, following the accidental fall of a minimalist, matte-white ceramic pot from a sun-drenched, white wooden windowsill. The camera then smoothly tilts downward, poetically tracking its descent. The delicate succulent is a plump, rosette-shaped cluster of Echeveria, its leaves graduating from a vibrant green in the center to pale pink and lavender at the edges, coated in a delicate white frosting. As the pot gently tumbles through the air, bright, warm afternoon sunlight filters through from behind, casting a dazzling golden halo around the pot and the plant, making the translucent leaf edges appear to glow from within. Then, a handful of dark, nutrient-rich soil gracefully gushes from the pot's lip, forming a slowly expanding, balletic, nebula in the air. Countless tiny particles of soil and white perlite shimmer in the light like floating dust. The entire picture is shrouded in a tranquil and focused atmosphere. The background is the slightly overexposed and blurred outdoor greenery seen through the clean window. The warm light and shadow are intertwined with the dynamic elements suspended in the air, transforming an unexpected moment into a dynamic painting full of harmony and broken beauty, and with a strong sense of life aesthetics.", + "seed": 6406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_7287_seed_6406_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021321597322821618, + "vlm_logprob": -0.021321597322821618, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.021321597322821618, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04768449440598488 + ], + "opposite_logprob_tokens": [ + -3.17268443107605 + ], + "logprob_reward": -0.04768449440598488, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04768449440598488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14850223064422607, + 0.0872509554028511, + 0.06200021505355835, + 0.02518896944820881, + 0.18461360037326813, + 0.22365573048591614, + 0.049133848398923874, + 0.06442002952098846, + 0.12433971464633942, + 0.03089469112455845 + ], + "credit_map": [ + 0.008528683334589005, + 0.016260448843240738, + 0.036703549325466156, + 0.03650936111807823, + 0.01656348444521427, + 0.009319442324340343, + 0.015685349702835083, + 0.00893192458897829, + 0.0081197340041399, + 0.01056942529976368, + 0.010321464389562607, + 0.008791625499725342, + 0.008268852718174458, + 0.010625342838466167, + 0.010887033306062222, + 0.01966748759150505, + 0.005420592147856951, + 0.0036394773051142693, + 0.011869288049638271, + 0.003987031057476997, + 0.004495890811085701, + 0.007942822761833668, + 0.00889469776302576, + 0.015750423073768616, + 0.002277032472193241, + 0.00286972732283175, + 0.004433793947100639, + 0.002265880350023508, + 0.0028828978538513184, + 0.0022821601014584303, + 0.0045179701410233974, + 0.0036595084238797426, + 0.01739698462188244, + 0.0327872596681118, + 0.01600409299135208, + 0.020195210352540016, + 0.01940632052719593, + 0.02721688151359558, + 0.024775296449661255, + 0.026831574738025665, + 0.020593836903572083, + 0.015017541125416756, + 0.03238716349005699, + 0.0217000599950552, + 0.020186085253953934, + 0.03763069584965706, + 0.02778826840221882, + 0.048352088779211044, + 0.007356818299740553, + 0.0035423184745013714, + 0.009006368927657604, + 0.004873800091445446, + 0.0034629469737410545, + 0.005816448945552111, + 0.0046822065487504005, + 0.01039294432848692, + 0.012349212542176247, + 0.0035651507787406445, + 0.014749979600310326, + 0.007675729226320982, + 0.0024870659690350294, + 0.006405399646610022, + 0.0058215600438416, + 0.011365936137735844, + 0.018389515578746796, + 0.009413514286279678, + 0.014395363628864288, + 0.014307782053947449, + 0.009519184939563274, + 0.017869194969534874, + 0.0169784277677536, + 0.02346673794090748, + 0.003266317304223776, + 0.0031286003068089485, + 0.004395205061882734, + 0.005495844408869743, + 0.002353477058932185, + 0.002468547783792019, + 0.004319513216614723, + 0.005467189941555262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.04768449440598488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968722980469465 + ], + "opposite_logprob_tokens": [ + -5.1309685707092285 + ], + "logprob_reward": -0.005968722980469465, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968722980469465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.084046371281147, + 0.057835523039102554, + 0.05795026570558548, + 0.15267935395240784, + 0.4160655438899994, + 0.07135163992643356, + 0.027193957939743996, + 0.02623290754854679, + 0.08027023822069168, + 0.026374179869890213 + ], + "credit_map": [ + 0.008346561342477798, + 0.01837056875228882, + 0.009833192452788353, + 0.008745191618800163, + 0.013724151067435741, + 0.00843118503689766, + 0.009166037663817406, + 0.007429482415318489, + 0.0052375937812030315, + 0.013487238436937332, + 0.00638654176145792, + 0.006866532377898693, + 0.008013232611119747, + 0.00564108369871974, + 0.006805178243666887, + 0.005398119799792767, + 0.004020676948130131, + 0.0024100702721625566, + 0.020473964512348175, + 0.0033886688761413097, + 0.002641218015924096, + 0.003745446912944317, + 0.006873019505292177, + 0.01439720019698143, + 0.01097656786441803, + 0.02785463258624077, + 0.019494768232107162, + 0.014217248186469078, + 0.02425992116332054, + 0.01529630646109581, + 0.02198510430753231, + 0.018594808876514435, + 0.04093825817108154, + 0.09331273287534714, + 0.04287758469581604, + 0.04704734683036804, + 0.04537729173898697, + 0.035326771438121796, + 0.05580301955342293, + 0.05538253113627434, + 0.006835012696683407, + 0.0036507982295006514, + 0.011072435416281223, + 0.007020187098532915, + 0.004552587866783142, + 0.01445204671472311, + 0.008821316063404083, + 0.014947263523936272, + 0.002290209522470832, + 0.0022249340545386076, + 0.004859504755586386, + 0.002991277491673827, + 0.0018804704304784536, + 0.0034767077304422855, + 0.0035621223505586386, + 0.0059087323024868965, + 0.0035773690324276686, + 0.0016922507202252746, + 0.005498262122273445, + 0.0046305106952786446, + 0.0016320913564413786, + 0.002383188111707568, + 0.0029843091033399105, + 0.0038349267560988665, + 0.01030872855335474, + 0.005344415083527565, + 0.012200310826301575, + 0.009291884489357471, + 0.006840868387371302, + 0.00921458937227726, + 0.009605485945940018, + 0.01746395416557789, + 0.004116530995815992, + 0.001843921490944922, + 0.004681966733187437, + 0.005586368031799793, + 0.001869837986305356, + 0.002775163622573018, + 0.002744979690760374, + 0.0027554130647331476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.005968722980469465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08637229353189468, + 0.06032246723771095, + 0.08585220575332642, + 0.038174908608198166, + 0.24150584638118744, + 0.21526294946670532, + 0.06471064686775208, + 0.06911862641572952, + 0.1110369935631752, + 0.02764309197664261 + ], + "credit_map": [ + 0.007098703179508448, + 0.0048821778036653996, + 0.022076316177845, + 0.022416334599256516, + 0.005223111715167761, + 0.010219634510576725, + 0.006708138156682253, + 0.00774786900728941, + 0.006328193936496973, + 0.00694411713629961, + 0.008056710474193096, + 0.005612170789390802, + 0.003615314606577158, + 0.008443660102784634, + 0.009088252671062946, + 0.012234048917889595, + 0.007140601519495249, + 0.005494025535881519, + 0.021288760006427765, + 0.005815194919705391, + 0.004436865448951721, + 0.009095131419599056, + 0.008557002060115337, + 0.02402462437748909, + 0.003975365776568651, + 0.005461846478283405, + 0.007332868408411741, + 0.0027538705617189407, + 0.004285700619220734, + 0.0035229730419814587, + 0.005504836793988943, + 0.0053374445997178555, + 0.019831538200378418, + 0.04828496277332306, + 0.030844205990433693, + 0.028117815032601357, + 0.024589041247963905, + 0.02296939305961132, + 0.03530799224972725, + 0.031560901552438736, + 0.02014368399977684, + 0.012603312730789185, + 0.03241942077875137, + 0.02559639699757099, + 0.014150046743452549, + 0.045860860496759415, + 0.027579842135310173, + 0.036909379065036774, + 0.006321994587779045, + 0.004075944889336824, + 0.01079533901065588, + 0.006959130056202412, + 0.0037144164089113474, + 0.009998484514653683, + 0.00798256229609251, + 0.014862772077322006, + 0.014008432626724243, + 0.00374067691154778, + 0.015989555045962334, + 0.007596947718411684, + 0.00226204632781446, + 0.0064817373640835285, + 0.005759946070611477, + 0.013279282487928867, + 0.013810473494231701, + 0.009374255314469337, + 0.014298069290816784, + 0.013602782040834427, + 0.009158103726804256, + 0.016022590920329094, + 0.01760442741215229, + 0.01716628670692444, + 0.00428512878715992, + 0.002261882647871971, + 0.004425056744366884, + 0.005250127054750919, + 0.0018160620238631964, + 0.0029563945718109608, + 0.0034629888832569122, + 0.003185453824698925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.012535471469163895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049196016043424606, + 0.06198631972074509, + 0.08438293635845184, + 0.03405408933758736, + 0.351166307926178, + 0.15945488214492798, + 0.046864818781614304, + 0.03920234739780426, + 0.14320269227027893, + 0.030489590018987656 + ], + "credit_map": [ + 0.006675072945654392, + 0.009536796249449253, + 0.00632709264755249, + 0.0036742042284458876, + 0.007046195678412914, + 0.00507555715739727, + 0.006630207411944866, + 0.004230887163430452, + 0.005800074432045221, + 0.004901946522295475, + 0.014014826156198978, + 0.006226089783012867, + 0.0035648616030812263, + 0.007473730947822332, + 0.008160852827131748, + 0.011843937449157238, + 0.007760690525174141, + 0.004400530830025673, + 0.02483602613210678, + 0.00363607844337821, + 0.003716215956956148, + 0.007996808737516403, + 0.007220530416816473, + 0.02481604740023613, + 0.0034287890885025263, + 0.0058244275860488415, + 0.00576280290260911, + 0.003257049946114421, + 0.0031285411678254604, + 0.003515094518661499, + 0.0044668326154351234, + 0.004670551046729088, + 0.03137440234422684, + 0.07132598757743835, + 0.03393851965665817, + 0.03951913118362427, + 0.03644052892923355, + 0.03897063806653023, + 0.05016995221376419, + 0.04942714795470238, + 0.013455966487526894, + 0.011411941610276699, + 0.023311104625463486, + 0.02056950330734253, + 0.012067710049450397, + 0.02643958479166031, + 0.022212045267224312, + 0.029987022280693054, + 0.003689542878419161, + 0.0020412702579051256, + 0.009388475678861141, + 0.003294628346338868, + 0.0020973177161067724, + 0.0040480000898242, + 0.004246436525136232, + 0.01805914379656315, + 0.004204018507152796, + 0.002666962333023548, + 0.006837502587586641, + 0.007666826713830233, + 0.0024147462099790573, + 0.004839398432523012, + 0.005104615353047848, + 0.005468274466693401, + 0.018366940319538116, + 0.011279589496552944, + 0.02332199178636074, + 0.020269250497221947, + 0.009477625600993633, + 0.024416737258434296, + 0.01841043308377266, + 0.0176601093262434, + 0.0037428566720336676, + 0.0025727874599397182, + 0.005018716678023338, + 0.006310015916824341, + 0.0028164673130959272, + 0.0038725927006453276, + 0.0031568424310535192, + 0.002999309916049242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368346393108368 + ], + "opposite_logprob_tokens": [ + -3.4086835384368896 + ], + "logprob_reward": -0.03368346393108368, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368346393108368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0639413371682167, + 0.057822082191705704, + 0.027081260457634926, + 0.057825833559036255, + 0.4380529820919037, + 0.12578463554382324, + 0.029881035909056664, + 0.06607557833194733, + 0.11053741723299026, + 0.022997835651040077 + ], + "credit_map": [ + 0.005641671363264322, + 0.011274158954620361, + 0.010252424515783787, + 0.016528978943824768, + 0.006931454408913851, + 0.004306744318455458, + 0.005896259099245071, + 0.0031096492893993855, + 0.004518049769103527, + 0.01412953995168209, + 0.009475735016167164, + 0.004888239782303572, + 0.006248174700886011, + 0.004685189109295607, + 0.0052057052962481976, + 0.008671451359987259, + 0.0020643193274736404, + 0.0017082527047023177, + 0.00691635487601161, + 0.0017424996476620436, + 0.001866960315965116, + 0.00231938436627388, + 0.0031714350916445255, + 0.007292056921869516, + 0.004541038069874048, + 0.012165729887783527, + 0.007975020445883274, + 0.005453719757497311, + 0.008089487440884113, + 0.004833162296563387, + 0.00729477871209383, + 0.007472905796021223, + 0.039851848036050797, + 0.09574317187070847, + 0.04117565602064133, + 0.04988469183444977, + 0.046425435692071915, + 0.04242834448814392, + 0.062014441937208176, + 0.0605294294655323, + 0.011776185594499111, + 0.009014966897666454, + 0.023294510319828987, + 0.010859774425625801, + 0.010960482992231846, + 0.019177855923771858, + 0.01754789799451828, + 0.023152975365519524, + 0.003866880666464567, + 0.001822242746129632, + 0.005474759731441736, + 0.0033048212062567472, + 0.002040386665612459, + 0.0037660845555365086, + 0.0030349029693752527, + 0.006570959929376841, + 0.011665072292089462, + 0.0027373195625841618, + 0.015018535777926445, + 0.009613236412405968, + 0.0023900559172034264, + 0.0063871475867927074, + 0.00709909712895751, + 0.01116512157022953, + 0.014888250268995762, + 0.007352052256464958, + 0.017283296212553978, + 0.014202890917658806, + 0.00848912913352251, + 0.011243517510592937, + 0.011919851414859295, + 0.025158435106277466, + 0.003249793080613017, + 0.001709198346361518, + 0.003727964125573635, + 0.005307386163622141, + 0.0014733703574165702, + 0.00220514927059412, + 0.0029335631988942623, + 0.002391413552686572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.03368346393108368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot in super slow motion begins at eye level, following the accidental fall of a minimalist, matte-white ceramic pot from a sun-drenched, white wooden windowsill. The camera then smoothly tilts downward, poetically tracking its descent. The delicate succulent is a plump, rosette-shaped cluster of Echeveria, its leaves graduating from a vibrant green in the center to pale pink and lavender at the edges, coated in a delicate white frosting. As the pot gently tumbles through the air, bright, warm afternoon sunlight filters through from behind, casting a dazzling golden halo around the pot and the plant, making the translucent leaf edges appear to glow from within. Then, a handful of dark, nutrient-rich soil gracefully gushes from the pot's lip, forming a slowly expanding, balletic, nebula in the air. Countless tiny particles of soil and white perlite shimmer in the light like floating dust. The entire picture is shrouded in a tranquil and focused atmosphere. The background is the slightly overexposed and blurred outdoor greenery seen through the clean window. The warm light and shadow are intertwined with the dynamic elements suspended in the air, transforming an unexpected moment into a dynamic painting full of harmony and broken beauty, and with a strong sense of life aesthetics.", + "seed": 6407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_7287_seed_6407_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010286318417638541, + "vlm_logprob": -0.010286318417638541, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.010286318417638541, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01783512346446514 + ], + "opposite_logprob_tokens": [ + -4.2678351402282715 + ], + "logprob_reward": -0.01783512346446514, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01783512346446514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07062727212905884, + 0.6313673853874207, + 0.1549910008907318, + 0.04092217609286308, + 0.010062429122626781, + 0.016836969181895256, + 0.013076525181531906, + 0.008473005145788193, + 0.013196634128689766, + 0.040446601808071136 + ], + "credit_map": [ + 0.010677638463675976, + 0.0070280227810144424, + 0.01656312309205532, + 0.01604965329170227, + 0.0049498071894049644, + 0.005050485022366047, + 0.005527637433260679, + 0.004780902992933989, + 0.129383847117424, + 0.027024168521165848, + 0.2835071384906769, + 0.07311027497053146, + 0.014726710505783558, + 0.017770610749721527, + 0.023804370313882828, + 0.06204025074839592, + 0.01640641875565052, + 0.007000867743045092, + 0.05702795088291168, + 0.05156773701310158, + 0.004026760812848806, + 0.003514067968353629, + 0.009334840811789036, + 0.006112369708716869, + 0.003530328394845128, + 0.003996736370027065, + 0.015035920776426792, + 0.010786096565425396, + 0.001971631310880184, + 0.0017803697846829891, + 0.0018554379930719733, + 0.001965656178072095, + 0.0014421428786590695, + 0.0009010027861222625, + 0.0020447932183742523, + 0.0015271676238626242, + 0.0010074565652757883, + 0.0009094389388337731, + 0.0011097067035734653, + 0.0011207204079255462, + 0.001432917546480894, + 0.001312031876295805, + 0.0031771764624863863, + 0.0017695424612611532, + 0.0027946699410676956, + 0.0016979590291157365, + 0.0024343938566744328, + 0.0022182778920978308, + 0.0008474632631987333, + 0.0012659135973080993, + 0.0014772575814276934, + 0.0019115159520879388, + 0.0022347376216202974, + 0.0011813536984845996, + 0.0026104548014700413, + 0.0015478275017812848, + 0.0010679520200937986, + 0.0009250726434402168, + 0.0012783595593646169, + 0.0011558475671336055, + 0.0008967476896941662, + 0.0005918576498515904, + 0.001112875179387629, + 0.0014442926039919257, + 0.0025570245925337076, + 0.000962027523200959, + 0.002825552597641945, + 0.00239741918630898, + 0.0009111568215303123, + 0.0011325950035825372, + 0.0012902135495096445, + 0.001120644505135715, + 0.011412645690143108, + 0.0013931706780567765, + 0.015117044560611248, + 0.0031227983999997377, + 0.001714987331070006, + 0.0021258904598653316, + 0.002679988509044051, + 0.002880074782297015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.01783512346446514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00527146365493536 + ], + "opposite_logprob_tokens": [ + -5.2552714347839355 + ], + "logprob_reward": -0.00527146365493536, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00527146365493536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11824526637792587, + 0.04308352619409561, + 0.22509706020355225, + 0.03627912327647209, + 0.03223619610071182, + 0.020044317469000816, + 0.07789895683526993, + 0.08816710859537125, + 0.08876688778400421, + 0.2701815068721771 + ], + "credit_map": [ + 0.020267577841877937, + 0.009212182834744453, + 0.038602765649557114, + 0.01640014909207821, + 0.01202896237373352, + 0.005777611397206783, + 0.007976530119776726, + 0.007979489862918854, + 0.005650889594107866, + 0.0051495772786438465, + 0.010184002108871937, + 0.004034265875816345, + 0.005612035747617483, + 0.004335159435868263, + 0.003906981088221073, + 0.00421061459928751, + 0.0229811929166317, + 0.012158474884927273, + 0.07869372516870499, + 0.07183945178985596, + 0.009259538725018501, + 0.005169795826077461, + 0.015473375096917152, + 0.00952149648219347, + 0.005120565183460712, + 0.00512955617159605, + 0.0039613316766917706, + 0.0059450538828969, + 0.002831605961546302, + 0.003212916199117899, + 0.004878480918705463, + 0.005199612118303776, + 0.003677083645015955, + 0.0023176430258899927, + 0.009378357790410519, + 0.006895731203258038, + 0.002824574476107955, + 0.0018422441789880395, + 0.0022117062471807003, + 0.0030888565815985203, + 0.002132533583790064, + 0.0017722854390740395, + 0.0023010841105133295, + 0.001976077212020755, + 0.0034405356273055077, + 0.00223253364674747, + 0.004236008506268263, + 0.0019532600417733192, + 0.009462626650929451, + 0.009818860329687595, + 0.008343033492565155, + 0.008925334550440311, + 0.014350892044603825, + 0.008942803367972374, + 0.010848801583051682, + 0.00720660574734211, + 0.009616002440452576, + 0.010475894436240196, + 0.011647028848528862, + 0.011862754821777344, + 0.01499608438462019, + 0.008237836882472038, + 0.011740684509277344, + 0.009590818546712399, + 0.019471099600195885, + 0.005462044384330511, + 0.023730777204036713, + 0.011320792138576508, + 0.005749845411628485, + 0.0074515496380627155, + 0.007733016740530729, + 0.00784776546061039, + 0.08173564076423645, + 0.012929514050483704, + 0.101258784532547, + 0.02025458961725235, + 0.01101637538522482, + 0.012532389722764492, + 0.015067058615386486, + 0.015387173742055893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.00527146365493536 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2051674872636795, + 0.2000468671321869, + 0.20413069427013397, + 0.07274775207042694, + 0.034205809235572815, + 0.04695262014865875, + 0.0314953438937664, + 0.04189174249768257, + 0.07470870018005371, + 0.08865297585725784 + ], + "credit_map": [ + 0.04449455440044403, + 0.013055932708084583, + 0.0763130784034729, + 0.02743569388985634, + 0.009818467311561108, + 0.0117680374532938, + 0.010637998580932617, + 0.011643724516034126, + 0.044359877705574036, + 0.010072547011077404, + 0.08656307309865952, + 0.023144008591771126, + 0.006078007631003857, + 0.007622810546308756, + 0.008010754361748695, + 0.014195792376995087, + 0.021846216171979904, + 0.009869636036455631, + 0.07415126264095306, + 0.06959519535303116, + 0.005190088879317045, + 0.004801727365702391, + 0.012737949378788471, + 0.005938615184277296, + 0.007049879524856806, + 0.004949025344103575, + 0.026850895956158638, + 0.016805993393063545, + 0.0030398694798350334, + 0.004304149653762579, + 0.004091018810868263, + 0.0056569199077785015, + 0.003939197864383459, + 0.0033622675109654665, + 0.006139596924185753, + 0.004081977531313896, + 0.005251429975032806, + 0.003251973306760192, + 0.004837891086935997, + 0.0033414731733500957, + 0.004336479585617781, + 0.0034072769340127707, + 0.007214725017547607, + 0.006977156735956669, + 0.005297149531543255, + 0.004990046843886375, + 0.008671186864376068, + 0.006058601196855307, + 0.003396454034373164, + 0.0025550422724336386, + 0.003883247496560216, + 0.005264966748654842, + 0.0027932676021009684, + 0.0033190951216965914, + 0.005473651457577944, + 0.004809619393199682, + 0.00696098105981946, + 0.0029320023022592068, + 0.0055125197395682335, + 0.006174639333039522, + 0.004815535619854927, + 0.004746777005493641, + 0.006146091967821121, + 0.004603195935487747, + 0.013975520618259907, + 0.004755253903567791, + 0.01705733872950077, + 0.013424161821603775, + 0.005785988178104162, + 0.005394952371716499, + 0.007373840548098087, + 0.006941644474864006, + 0.02021436020731926, + 0.005642234347760677, + 0.025345157831907272, + 0.014918310567736626, + 0.004012132529169321, + 0.005404714494943619, + 0.006268107332289219, + 0.006847952026873827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.007627766579389572 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09901564568281174, + 0.5033080577850342, + 0.1541457176208496, + 0.015715740621089935, + 0.0378122515976429, + 0.02827821858227253, + 0.022341566160321236, + 0.022930586710572243, + 0.0400146059691906, + 0.07643761485815048 + ], + "credit_map": [ + 0.013347791507840157, + 0.011037004180252552, + 0.023313436657190323, + 0.024312913417816162, + 0.007147533819079399, + 0.0063089653849601746, + 0.007348866667598486, + 0.006199129391461611, + 0.11284656822681427, + 0.023974783718585968, + 0.21896104514598846, + 0.06241600215435028, + 0.011805144138634205, + 0.018500834703445435, + 0.019848370924592018, + 0.03495527058839798, + 0.017030781134963036, + 0.007173336111009121, + 0.05523834377527237, + 0.05307600647211075, + 0.0032664199825376272, + 0.003778717014938593, + 0.009672842919826508, + 0.004909252282232046, + 0.001964227994903922, + 0.0032218755222857, + 0.003214737633243203, + 0.001836963463574648, + 0.001521683414466679, + 0.001519794575870037, + 0.001063689822331071, + 0.0013727652840316296, + 0.004296573344618082, + 0.004336161073297262, + 0.007970857433974743, + 0.006108234170824289, + 0.002455825451761484, + 0.00339888664893806, + 0.003972440492361784, + 0.005273271352052689, + 0.0026594530791044235, + 0.0023065186105668545, + 0.004939327947795391, + 0.005395193584263325, + 0.002396742347627878, + 0.002520165638998151, + 0.004706124309450388, + 0.003354693064466119, + 0.0024763362016528845, + 0.002167056081816554, + 0.002373243449255824, + 0.0025223074480891228, + 0.004923842381685972, + 0.001781776431016624, + 0.004010065924376249, + 0.0020869364961981773, + 0.0022421376779675484, + 0.0026882097590714693, + 0.002613880205899477, + 0.0033753588795661926, + 0.00560969952493906, + 0.0016484332736581564, + 0.002895582467317581, + 0.0018572837579995394, + 0.007859780453145504, + 0.0020576571114361286, + 0.008513356558978558, + 0.005852477625012398, + 0.0024345810525119305, + 0.0033261848147958517, + 0.003867091378197074, + 0.006103475112468004, + 0.016742674633860588, + 0.0037098643369972706, + 0.020457377657294273, + 0.01381191611289978, + 0.003568374551832676, + 0.006138990167528391, + 0.0059112319722771645, + 0.006097184494137764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0046369642950594425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060274094343185 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.016060274094343185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060274094343185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15404440462589264, + 0.26155003905296326, + 0.08395098894834518, + 0.06282150745391846, + 0.021930444985628128, + 0.017603695392608643, + 0.024329770356416702, + 0.037603236734867096, + 0.054899465292692184, + 0.2812664210796356 + ], + "credit_map": [ + 0.03568338602781296, + 0.008350654505193233, + 0.06289046257734299, + 0.019700463861227036, + 0.004473370965570211, + 0.006906406953930855, + 0.006267876364290714, + 0.009771786630153656, + 0.05065309256315231, + 0.011549601331353188, + 0.11701912432909012, + 0.030066518113017082, + 0.006893238518387079, + 0.006559490691870451, + 0.009513874538242817, + 0.029295101761817932, + 0.008964828215539455, + 0.004038304090499878, + 0.026888301596045494, + 0.027872232720255852, + 0.0027329314034432173, + 0.0029853079468011856, + 0.005531885661184788, + 0.0049371919594705105, + 0.005349077749997377, + 0.004273976664990187, + 0.025770626962184906, + 0.017502427101135254, + 0.0013577328063547611, + 0.0018803559942170978, + 0.00322827510535717, + 0.0034590379800647497, + 0.0021524587646126747, + 0.0016695298254489899, + 0.004868283402174711, + 0.0036645792424678802, + 0.002586565213277936, + 0.002527190838009119, + 0.002696078037843108, + 0.001765760243870318, + 0.001580844516865909, + 0.0017504196148365736, + 0.0028905365616083145, + 0.0025569945573806763, + 0.0017602061852812767, + 0.0015788814052939415, + 0.0023555257357656956, + 0.0031302866991609335, + 0.0033459379337728024, + 0.0021490822546184063, + 0.003212379291653633, + 0.0027857571840286255, + 0.0037194390315562487, + 0.001942416769452393, + 0.0039131054654717445, + 0.0032616520766168833, + 0.004613196011632681, + 0.0035425382666289806, + 0.006274722516536713, + 0.005556977353990078, + 0.0038699184078723192, + 0.0026463856920599937, + 0.005048039369285107, + 0.006051461212337017, + 0.007006555795669556, + 0.005164682399481535, + 0.006064651533961296, + 0.010428647510707378, + 0.006224863231182098, + 0.007283397950232029, + 0.005664339754730463, + 0.007062327116727829, + 0.06461724638938904, + 0.013289016671478748, + 0.07679436355829239, + 0.039309579879045486, + 0.009714567102491856, + 0.022194549441337585, + 0.02169477753341198, + 0.033652324229478836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.016060274094343185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot in super slow motion begins at eye level, following the accidental fall of a minimalist, matte-white ceramic pot from a sun-drenched, white wooden windowsill. The camera then smoothly tilts downward, poetically tracking its descent. The delicate succulent is a plump, rosette-shaped cluster of Echeveria, its leaves graduating from a vibrant green in the center to pale pink and lavender at the edges, coated in a delicate white frosting. As the pot gently tumbles through the air, bright, warm afternoon sunlight filters through from behind, casting a dazzling golden halo around the pot and the plant, making the translucent leaf edges appear to glow from within. Then, a handful of dark, nutrient-rich soil gracefully gushes from the pot's lip, forming a slowly expanding, balletic, nebula in the air. Countless tiny particles of soil and white perlite shimmer in the light like floating dust. The entire picture is shrouded in a tranquil and focused atmosphere. The background is the slightly overexposed and blurred outdoor greenery seen through the clean window. The warm light and shadow are intertwined with the dynamic elements suspended in the air, transforming an unexpected moment into a dynamic painting full of harmony and broken beauty, and with a strong sense of life aesthetics.", + "seed": 6408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_7287_seed_6408_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022123722173273564, + "vlm_logprob": -0.022123722173273564, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.022123722173273564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06251727789640427 + ], + "opposite_logprob_tokens": [ + -2.9375171661376953 + ], + "logprob_reward": -0.06251727789640427, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06251727789640427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1729092299938202, + 0.12083864957094193, + 0.1085367351770401, + 0.06506814807653427, + 0.16291750967502594, + 0.08858448266983032, + 0.07999414205551147, + 0.05382775142788887, + 0.10574891418218613, + 0.04157433658838272 + ], + "credit_map": [ + 0.035146214067935944, + 0.013185967691242695, + 0.02019890397787094, + 0.035992901772260666, + 0.02247021347284317, + 0.012021325528621674, + 0.019730746746063232, + 0.014162984676659107, + 0.025208279490470886, + 0.012291069142520428, + 0.01974574476480484, + 0.022206293419003487, + 0.010975426994264126, + 0.010157271288335323, + 0.013468905352056026, + 0.006785671692341566, + 0.02985857054591179, + 0.016347629949450493, + 0.022967634722590446, + 0.007617117837071419, + 0.009106469340622425, + 0.006574354134500027, + 0.00881532859057188, + 0.00724964402616024, + 0.01360896322876215, + 0.006862004287540913, + 0.013160958886146545, + 0.0063735549338161945, + 0.007568724919110537, + 0.004457002971321344, + 0.008994669653475285, + 0.004042281769216061, + 0.03613012284040451, + 0.009240852668881416, + 0.02479463815689087, + 0.02181820012629032, + 0.021342836320400238, + 0.012489747256040573, + 0.02369542419910431, + 0.01340571790933609, + 0.010954002849757671, + 0.006355301011353731, + 0.012332696467638016, + 0.011521164327859879, + 0.009588200598955154, + 0.00822078064084053, + 0.008934423327445984, + 0.020677918568253517, + 0.011484543792903423, + 0.00556825241073966, + 0.018357353284955025, + 0.019580112770199776, + 0.006066870875656605, + 0.007146252319216728, + 0.006888081785291433, + 0.004902684595435858, + 0.00865007285028696, + 0.005367630161345005, + 0.006598588079214096, + 0.008648592047393322, + 0.004797301720827818, + 0.007649648003280163, + 0.007781103253364563, + 0.004334823694080114, + 0.018138576298952103, + 0.017942700535058975, + 0.015021709725260735, + 0.013389336876571178, + 0.008272523060441017, + 0.010429812595248222, + 0.006929788272827864, + 0.015624480322003365, + 0.005837824661284685, + 0.0051691788248717785, + 0.006952299270778894, + 0.00509394146502018, + 0.003526508342474699, + 0.005503607448190451, + 0.004438060335814953, + 0.005052922759205103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.06251727789640427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004116870928555727 + ], + "opposite_logprob_tokens": [ + -5.504117012023926 + ], + "logprob_reward": -0.004116870928555727, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004116870928555727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23541270196437836, + 0.09391586482524872, + 0.12157248705625534, + 0.09984077513217926, + 0.16005957126617432, + 0.05682418495416641, + 0.07350653409957886, + 0.031754270195961, + 0.06348948180675507, + 0.06362411379814148 + ], + "credit_map": [ + 0.0342300683259964, + 0.020299332216382027, + 0.042178522795438766, + 0.02969617210328579, + 0.049868691712617874, + 0.01771577261388302, + 0.020564991980791092, + 0.02085915394127369, + 0.026036066934466362, + 0.007502411492168903, + 0.025368792936205864, + 0.009078018367290497, + 0.00980721041560173, + 0.003983360715210438, + 0.006709903944283724, + 0.005430104210972786, + 0.020380007103085518, + 0.013973246328532696, + 0.015837328508496284, + 0.01030655950307846, + 0.0220177099108696, + 0.010852723382413387, + 0.015076027251780033, + 0.013128884136676788, + 0.013380825519561768, + 0.008468031883239746, + 0.01253641489893198, + 0.010790717788040638, + 0.01502971351146698, + 0.009466562420129776, + 0.016898326575756073, + 0.013270183466374874, + 0.029024861752986908, + 0.010217776522040367, + 0.023079553619027138, + 0.014699372462928295, + 0.02672540582716465, + 0.022179314866662025, + 0.01904214173555374, + 0.015091138891875744, + 0.006933330558240414, + 0.0031174831092357635, + 0.007038821466267109, + 0.010057101026177406, + 0.00936967134475708, + 0.0068633947521448135, + 0.0048744152300059795, + 0.008569968864321709, + 0.010218475013971329, + 0.0035784875508397818, + 0.0216653011739254, + 0.0141381174325943, + 0.005605889018625021, + 0.006483111530542374, + 0.0054406290873885155, + 0.006376519799232483, + 0.0053221615962684155, + 0.00272977352142334, + 0.004582828842103481, + 0.004126608371734619, + 0.002887595212087035, + 0.0030969553627073765, + 0.005414224229753017, + 0.0035941218957304955, + 0.009141327813267708, + 0.007424697279930115, + 0.008821177296340466, + 0.007365700323134661, + 0.006811786908656359, + 0.009232587181031704, + 0.005495774559676647, + 0.009196432307362556, + 0.011655256152153015, + 0.005799921695142984, + 0.01348172128200531, + 0.005707953590899706, + 0.00648079439997673, + 0.005894201807677746, + 0.007602212019264698, + 0.007002054248005152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.004116870928555727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818285696208477 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818285696208477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818285696208477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11453871428966522, + 0.1525081843137741, + 0.08006222546100616, + 0.09124845266342163, + 0.11968289315700531, + 0.07445190101861954, + 0.11241643875837326, + 0.05489503592252731, + 0.15676993131637573, + 0.04342620447278023 + ], + "credit_map": [ + 0.019109254702925682, + 0.009788238443434238, + 0.014501894824206829, + 0.016061970964074135, + 0.013916105031967163, + 0.014694231562316418, + 0.012141452170908451, + 0.014325574971735477, + 0.025663498789072037, + 0.01408296637237072, + 0.01566876657307148, + 0.024132728576660156, + 0.016080526635050774, + 0.02858240343630314, + 0.01496744342148304, + 0.013329852372407913, + 0.022296125069260597, + 0.01166378054767847, + 0.01572224125266075, + 0.005186612252146006, + 0.008411673828959465, + 0.005575967486947775, + 0.005783861968666315, + 0.0054219611920416355, + 0.0122210793197155, + 0.00774174090474844, + 0.012743583880364895, + 0.009555093944072723, + 0.01443553902208805, + 0.010088009759783745, + 0.013200622983276844, + 0.011262787505984306, + 0.02241297997534275, + 0.006967793684452772, + 0.018651969730854034, + 0.014219189994037151, + 0.017371518537402153, + 0.010841896757483482, + 0.017676550894975662, + 0.011540987528860569, + 0.011708433739840984, + 0.004281936679035425, + 0.010211805813014507, + 0.011776767671108246, + 0.006988614331930876, + 0.007203081622719765, + 0.007460529915988445, + 0.014820733107626438, + 0.0163456741720438, + 0.008509539067745209, + 0.021266620606184006, + 0.028842128813266754, + 0.008803547360002995, + 0.010743694379925728, + 0.010353474877774715, + 0.007551760412752628, + 0.009068651124835014, + 0.005955707747489214, + 0.007553951349109411, + 0.009253053925931454, + 0.0039027065504342318, + 0.007810341194272041, + 0.005783664993941784, + 0.005566955078393221, + 0.03636746481060982, + 0.01605948619544506, + 0.022209005430340767, + 0.019746707752346992, + 0.011128246784210205, + 0.01763797365128994, + 0.00980369746685028, + 0.023817356675863266, + 0.0052514453418552876, + 0.006685112603008747, + 0.00471070222556591, + 0.005294424947351217, + 0.0045648678205907345, + 0.005575205199420452, + 0.005637660622596741, + 0.005706788972020149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.01818285696208477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2396719753742218, + 0.06195725128054619, + 0.07439561933279037, + 0.029086504131555557, + 0.15130175650119781, + 0.18433743715286255, + 0.13118159770965576, + 0.03477117791771889, + 0.02791713923215866, + 0.06537958979606628 + ], + "credit_map": [ + 0.036168813705444336, + 0.01244326215237379, + 0.05610736832022667, + 0.025396114215254784, + 0.03049393743276596, + 0.031735725700855255, + 0.015903474763035774, + 0.03142325580120087, + 0.014213125221431255, + 0.004822856746613979, + 0.011947077699005604, + 0.0072111752815544605, + 0.0048576961271464825, + 0.009173929691314697, + 0.0053182183764874935, + 0.004413161426782608, + 0.01874934695661068, + 0.011308416724205017, + 0.012349260970950127, + 0.0066347853280603886, + 0.007571076974272728, + 0.005132712423801422, + 0.006804085802286863, + 0.005845929961651564, + 0.006672326009720564, + 0.00324996211566031, + 0.004026863723993301, + 0.0036669119726866484, + 0.0028990781866014004, + 0.0030582016333937645, + 0.002683960134163499, + 0.0028291980270296335, + 0.026942521333694458, + 0.014436284080147743, + 0.020530328154563904, + 0.023990407586097717, + 0.01728666014969349, + 0.011801579035818577, + 0.019968677312135696, + 0.016345279291272163, + 0.029362032189965248, + 0.01660424843430519, + 0.023021258413791656, + 0.029970865696668625, + 0.009370705112814903, + 0.013328569009900093, + 0.025160348042845726, + 0.03751939535140991, + 0.01776796579360962, + 0.005959745030850172, + 0.04191846027970314, + 0.027794409543275833, + 0.009846886619925499, + 0.010183398611843586, + 0.010345459915697575, + 0.0073652504943311214, + 0.005302825942635536, + 0.003726871218532324, + 0.004595866426825523, + 0.0052224635146558285, + 0.002895378042012453, + 0.004730117041617632, + 0.0048081534914672375, + 0.003489499446004629, + 0.005759560503065586, + 0.003614113898947835, + 0.0038692739326506853, + 0.0030797081999480724, + 0.0023815627209842205, + 0.003817419521510601, + 0.0025448137894272804, + 0.0028506841044873, + 0.014095302671194077, + 0.005798080004751682, + 0.017144694924354553, + 0.005454298108816147, + 0.0052048275247216225, + 0.005053404252976179, + 0.005108414217829704, + 0.0075205606408417225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.007627411279827356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018174193799495697 + ], + "opposite_logprob_tokens": [ + -4.018174171447754 + ], + "logprob_reward": -0.018174193799495697, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018174193799495697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14138434827327728, + 0.09729474037885666, + 0.03004671074450016, + 0.03871756047010422, + 0.092381551861763, + 0.08463717997074127, + 0.19816526770591736, + 0.11414247006177902, + 0.1206718310713768, + 0.08255838602781296 + ], + "credit_map": [ + 0.025023961439728737, + 0.010437902063131332, + 0.020431621000170708, + 0.02435201220214367, + 0.02040078490972519, + 0.015858914703130722, + 0.010744715109467506, + 0.01413442101329565, + 0.02145322598516941, + 0.007821350358426571, + 0.022173060104250908, + 0.011086326092481613, + 0.010995609685778618, + 0.010471193119883537, + 0.007005436345934868, + 0.006288533098995686, + 0.005690691526979208, + 0.003932371735572815, + 0.004394105169922113, + 0.003755653277039528, + 0.0037731046322733164, + 0.002397427801042795, + 0.002966750878840685, + 0.003136605257168412, + 0.008881180547177792, + 0.004196619149297476, + 0.007928564213216305, + 0.004158787429332733, + 0.0030764725524932146, + 0.0029713769908994436, + 0.0037323893047869205, + 0.003772169351577759, + 0.02419305592775345, + 0.00690569868311286, + 0.014167612418532372, + 0.014010160230100155, + 0.008415494114160538, + 0.0049259476363658905, + 0.010455991141498089, + 0.009307592175900936, + 0.014457150362432003, + 0.004965245723724365, + 0.01156566757708788, + 0.013584636151790619, + 0.0065657091327011585, + 0.009578017517924309, + 0.008183574303984642, + 0.015737177804112434, + 0.026748038828372955, + 0.011264433152973652, + 0.05363057553768158, + 0.04357101395726204, + 0.015620379708707333, + 0.018033716827630997, + 0.017707278952002525, + 0.011589821428060532, + 0.01784231700003147, + 0.011354406364262104, + 0.013025382533669472, + 0.01843019761145115, + 0.010257227346301079, + 0.015209750272333622, + 0.018796224147081375, + 0.009226958267390728, + 0.01716412603855133, + 0.016835052520036697, + 0.015270556323230267, + 0.01545284315943718, + 0.013979402370750904, + 0.016201205551624298, + 0.0122749712318182, + 0.013493673875927925, + 0.01371215283870697, + 0.006130489986389875, + 0.01819130964577198, + 0.010632469318807125, + 0.005521874874830246, + 0.007436053361743689, + 0.00927809253334999, + 0.01165594719350338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.018174193799495697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot in super slow motion begins at eye level, following the accidental fall of a minimalist, matte-white ceramic pot from a sun-drenched, white wooden windowsill. The camera then smoothly tilts downward, poetically tracking its descent. The delicate succulent is a plump, rosette-shaped cluster of Echeveria, its leaves graduating from a vibrant green in the center to pale pink and lavender at the edges, coated in a delicate white frosting. As the pot gently tumbles through the air, bright, warm afternoon sunlight filters through from behind, casting a dazzling golden halo around the pot and the plant, making the translucent leaf edges appear to glow from within. Then, a handful of dark, nutrient-rich soil gracefully gushes from the pot's lip, forming a slowly expanding, balletic, nebula in the air. Countless tiny particles of soil and white perlite shimmer in the light like floating dust. The entire picture is shrouded in a tranquil and focused atmosphere. The background is the slightly overexposed and blurred outdoor greenery seen through the clean window. The warm light and shadow are intertwined with the dynamic elements suspended in the air, transforming an unexpected moment into a dynamic painting full of harmony and broken beauty, and with a strong sense of life aesthetics.", + "seed": 6409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_4/index_7287_seed_6409_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017071989737451075, + "vlm_logprob": -0.017071989737451075, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.017071989737451075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021190783008933067 + ], + "opposite_logprob_tokens": [ + -4.396190643310547 + ], + "logprob_reward": -0.021190783008933067, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.021190783008933067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07381444424390793, + 0.10942325741052628, + 0.08242158591747284, + 0.1761564165353775, + 0.10822134464979172, + 0.07359384000301361, + 0.10902009159326553, + 0.07138410210609436, + 0.09679984301328659, + 0.09916502237319946 + ], + "credit_map": [ + 0.008575620129704475, + 0.006621444132179022, + 0.010260586626827717, + 0.006811104249209166, + 0.014785980805754662, + 0.008310580626130104, + 0.00738686416298151, + 0.011062266305088997, + 0.009095457382500172, + 0.01119541097432375, + 0.009957225061953068, + 0.013800105080008507, + 0.020646579563617706, + 0.012021779082715511, + 0.018903067335486412, + 0.013803641311824322, + 0.009558653458952904, + 0.012031408958137035, + 0.01071621011942625, + 0.007819050922989845, + 0.010311512276530266, + 0.00828015897423029, + 0.010445869527757168, + 0.013258734717965126, + 0.01892620138823986, + 0.020765764638781548, + 0.017830530181527138, + 0.025570381432771683, + 0.02708907425403595, + 0.0236574225127697, + 0.02970833145081997, + 0.012608732096850872, + 0.01066440436989069, + 0.01576043665409088, + 0.013821685686707497, + 0.018204286694526672, + 0.007740008644759655, + 0.011898212134838104, + 0.014384708367288113, + 0.015747608616948128, + 0.007095137145370245, + 0.012379646301269531, + 0.007491390686482191, + 0.008404772728681564, + 0.012793285772204399, + 0.010610854253172874, + 0.006203050725162029, + 0.008615713566541672, + 0.009938189759850502, + 0.026879113167524338, + 0.017096659168601036, + 0.00887428317219019, + 0.018110280856490135, + 0.012866944074630737, + 0.006528078578412533, + 0.008726551197469234, + 0.00671704625710845, + 0.010483442805707455, + 0.009156931191682816, + 0.010151028633117676, + 0.00953990314155817, + 0.009970976039767265, + 0.007065656594932079, + 0.008299127221107483, + 0.013447894714772701, + 0.014366737566888332, + 0.02022266574203968, + 0.011780330911278725, + 0.011960022151470184, + 0.007378525100648403, + 0.010044505819678307, + 0.007599171716719866, + 0.012126789428293705, + 0.01417907327413559, + 0.01279541477560997, + 0.008307111449539661, + 0.015552757307887077, + 0.0104236900806427, + 0.01384893897920847, + 0.011931255459785461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.021190783008933067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008670185692608356 + ], + "opposite_logprob_tokens": [ + -4.758670330047607 + ], + "logprob_reward": -0.008670185692608356, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008670185692608356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10465431958436966, + 0.059100233018398285, + 0.044144097715616226, + 0.08639863133430481, + 0.22152678668498993, + 0.11827567219734192, + 0.15322215855121613, + 0.08059356361627579, + 0.044075943529605865, + 0.08800864219665527 + ], + "credit_map": [ + 0.009907349944114685, + 0.011434227228164673, + 0.011916792020201683, + 0.009838982485234737, + 0.022041231393814087, + 0.010009178891777992, + 0.015351731330156326, + 0.014154812321066856, + 0.0050621069967746735, + 0.006501241587102413, + 0.006166425999253988, + 0.007229875773191452, + 0.012109490111470222, + 0.0060336594469845295, + 0.008432647213339806, + 0.0075647784397006035, + 0.0033889508340507746, + 0.005755698308348656, + 0.004647071473300457, + 0.0064261616207659245, + 0.006394037976861, + 0.004703993909060955, + 0.0066319857724010944, + 0.006196192465722561, + 0.007552227005362511, + 0.01293663214892149, + 0.007414294872432947, + 0.010818923823535442, + 0.02030095085501671, + 0.012052075006067753, + 0.011059067212045193, + 0.004264447372406721, + 0.020119933411478996, + 0.019971907138824463, + 0.02311757206916809, + 0.04389137029647827, + 0.016550302505493164, + 0.021104643121361732, + 0.042145680636167526, + 0.034625355154275894, + 0.009653305634856224, + 0.02326878532767296, + 0.012767747975885868, + 0.015461066737771034, + 0.016832774505019188, + 0.01463654637336731, + 0.012732340022921562, + 0.012923099100589752, + 0.016041679307818413, + 0.033103834837675095, + 0.019579194486141205, + 0.015608711168169975, + 0.02568395622074604, + 0.018289679661393166, + 0.011815187521278858, + 0.013099901378154755, + 0.008223508484661579, + 0.00930998008698225, + 0.008488518185913563, + 0.009229193441569805, + 0.011066453531384468, + 0.01279095932841301, + 0.0068601928651332855, + 0.01462474837899208, + 0.005014745052903891, + 0.007479717489331961, + 0.007478251587599516, + 0.00361807132139802, + 0.006813166197389364, + 0.004427675157785416, + 0.005476818419992924, + 0.003767495509237051, + 0.010779664851725101, + 0.011888043023645878, + 0.013402419164776802, + 0.008894026279449463, + 0.0111779123544693, + 0.008444065228104591, + 0.013910248875617981, + 0.00951225683093071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.008670185692608356 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08545280247926712, + 0.14014534652233124, + 0.05116093531250954, + 0.14668303728103638, + 0.08567783236503601, + 0.12585704028606415, + 0.10488765686750412, + 0.11815468966960907, + 0.06518842279911041, + 0.07679222524166107 + ], + "credit_map": [ + 0.011290401220321655, + 0.00825509987771511, + 0.0114699462428689, + 0.006803782656788826, + 0.01536965649574995, + 0.006299064960330725, + 0.014198900200426579, + 0.011765938252210617, + 0.013529322110116482, + 0.013974500820040703, + 0.013238863088190556, + 0.01759880781173706, + 0.021341556683182716, + 0.017215436324477196, + 0.02277580462396145, + 0.02047104761004448, + 0.00544342678040266, + 0.006813453510403633, + 0.00677937176078558, + 0.006582002155482769, + 0.006807617377489805, + 0.005189110059291124, + 0.007657825946807861, + 0.005888126790523529, + 0.015677357092499733, + 0.02453150600194931, + 0.018821077421307564, + 0.014985754154622555, + 0.025489427149295807, + 0.022222524508833885, + 0.016838207840919495, + 0.00811717752367258, + 0.006057358346879482, + 0.011871722526848316, + 0.009302866645157337, + 0.015032381750643253, + 0.00624333368614316, + 0.008282286114990711, + 0.016650335863232613, + 0.01223753672093153, + 0.012569231912493706, + 0.022693734616041183, + 0.01406039111316204, + 0.011599120683968067, + 0.02690078504383564, + 0.02091323584318161, + 0.006484419573098421, + 0.010636103339493275, + 0.010998735204339027, + 0.01824677176773548, + 0.015372944995760918, + 0.011149154044687748, + 0.014868492260575294, + 0.01291689369827509, + 0.009929096326231956, + 0.011405556462705135, + 0.010823322460055351, + 0.021587269380688667, + 0.014031504280865192, + 0.012385115958750248, + 0.020323606207966805, + 0.01340584084391594, + 0.012004136107861996, + 0.01359388418495655, + 0.006205196492373943, + 0.006504635326564312, + 0.00998123548924923, + 0.007507756352424622, + 0.00803870428353548, + 0.008770766668021679, + 0.012767416425049305, + 0.005412704776972532, + 0.011501497589051723, + 0.010386108420789242, + 0.012388709932565689, + 0.006670480594038963, + 0.010321235284209251, + 0.010128840804100037, + 0.00587623892351985, + 0.009519108571112156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.007632853463292122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097829420119524 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.0097829420119524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097829420119524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1166149377822876, + 0.04633636400103569, + 0.048815760761499405, + 0.09243518114089966, + 0.09613075107336044, + 0.12337810546159744, + 0.16009065508842468, + 0.11776319891214371, + 0.09079410135746002, + 0.10764095187187195 + ], + "credit_map": [ + 0.014553869143128395, + 0.012503217905759811, + 0.011364857666194439, + 0.014476765878498554, + 0.022121891379356384, + 0.008308865129947662, + 0.01973092369735241, + 0.013554541394114494, + 0.004833841230720282, + 0.00494821323081851, + 0.004566866904497147, + 0.005351404659450054, + 0.008433770388364792, + 0.006982702761888504, + 0.005368122365325689, + 0.005851446185261011, + 0.007862623780965805, + 0.005796333774924278, + 0.006906928960233927, + 0.006657658144831657, + 0.004635424818843603, + 0.005290536675602198, + 0.006078924983739853, + 0.005587332881987095, + 0.010218977928161621, + 0.016510726884007454, + 0.0057769096456468105, + 0.00882695335894823, + 0.018837718293070793, + 0.017255274578928947, + 0.011071031913161278, + 0.003937589470297098, + 0.012951611541211605, + 0.012635266408324242, + 0.0112109100446105, + 0.013784058392047882, + 0.010403352789580822, + 0.00866978894919157, + 0.013972677290439606, + 0.012503081932663918, + 0.014227649196982384, + 0.018425092101097107, + 0.012884978204965591, + 0.014072705991566181, + 0.018426593393087387, + 0.02000817470252514, + 0.012069130316376686, + 0.013263782486319542, + 0.01884828321635723, + 0.020561926066875458, + 0.016484707593917847, + 0.02314828708767891, + 0.01904928870499134, + 0.019214028492569923, + 0.022993505001068115, + 0.019790619611740112, + 0.012520917691290379, + 0.0261517483741045, + 0.015066905878484249, + 0.014104075729846954, + 0.019938956946134567, + 0.014040259644389153, + 0.005848694127053022, + 0.010091637261211872, + 0.010822509415447712, + 0.01274785678833723, + 0.020832719281315804, + 0.011905239894986153, + 0.010573786683380604, + 0.007847019471228123, + 0.008511143736541271, + 0.007553826551884413, + 0.014169881120324135, + 0.013774055987596512, + 0.017171518877148628, + 0.012017452158033848, + 0.015328365378081799, + 0.013322774320840836, + 0.011667571030557156, + 0.010189337655901909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0097829420119524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a minimalist, matte-white ceramic pot falling from a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a succulent plant with rosette-shaped Echeveria leaves that are green in the center and pale pink/lavender at the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera movement a smooth tilt downward tracking the pot's descent in super slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature bright, warm afternoon sunlight creating a golden halo and glowing translucent edges on the pot and plant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dark soil and white perlite particles gush from the pot and shimmer like floating dust in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808318451046944 + ], + "opposite_logprob_tokens": [ + -3.288083076477051 + ], + "logprob_reward": -0.03808318451046944, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808318451046944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05613837018609047, + 0.05860346183180809, + 0.04227336496114731, + 0.16112123429775238, + 0.04628400132060051, + 0.08853969722986221, + 0.14044572412967682, + 0.22114817798137665, + 0.07575009763240814, + 0.10969580709934235 + ], + "credit_map": [ + 0.00907616876065731, + 0.006272532977163792, + 0.006763535086065531, + 0.0065359571017324924, + 0.009391837753355503, + 0.004232908599078655, + 0.006781183183193207, + 0.007084250915795565, + 0.00479164719581604, + 0.006760421674698591, + 0.005426045507192612, + 0.007345112971961498, + 0.014858769252896309, + 0.006483838427811861, + 0.0073593598790466785, + 0.005578272044658661, + 0.0050285374745726585, + 0.005595947615802288, + 0.0058369566686451435, + 0.005703675094991922, + 0.005132297053933144, + 0.004614748992025852, + 0.005775668192654848, + 0.004585538990795612, + 0.017246190458536148, + 0.025078199803829193, + 0.011956318281590939, + 0.015768591314554214, + 0.036289021372795105, + 0.021583449095487595, + 0.02396206557750702, + 0.00923741701990366, + 0.005069148261100054, + 0.006882633548229933, + 0.006716444157063961, + 0.005926057696342468, + 0.004885877948254347, + 0.0059301406145095825, + 0.004278445616364479, + 0.006595255341380835, + 0.008213629946112633, + 0.009704610332846642, + 0.00863278191536665, + 0.007323656231164932, + 0.015778962522745132, + 0.014255466870963573, + 0.006597301457077265, + 0.01803329586982727, + 0.009598475880920887, + 0.028475768864154816, + 0.013918774202466011, + 0.0142993638291955, + 0.025346903130412102, + 0.021571911871433258, + 0.01225495059043169, + 0.014979582279920578, + 0.016482777893543243, + 0.043753884732723236, + 0.02010500431060791, + 0.02929052524268627, + 0.03860866278409958, + 0.029538925737142563, + 0.018106447532773018, + 0.025261960923671722, + 0.00801976677030325, + 0.010393379256129265, + 0.007737437728792429, + 0.007268411573022604, + 0.009358182549476624, + 0.009055844508111477, + 0.014805070124566555, + 0.00911201536655426, + 0.016086600720882416, + 0.012479474768042564, + 0.020665576681494713, + 0.010587734170258045, + 0.013985656201839447, + 0.014179959893226624, + 0.010273957625031471, + 0.011436848901212215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.03808318451046944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_tensor([3582, 3582])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_tensor([3582, 3582])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..15bcd5209bfb31641f298537124eb09f3ed8607c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_tensor([3582, 3582])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, low-angle tracking shot glides across the dark forest floor, rendered in a rich, painterly 3D cartoon style reminiscent of the \"Two Cities\" series. Unusual tendrils of a magical fire, a vibrant blend of fuchsia, violet, and fiery orange, creep across the damp soil and gnarled tree roots with serpentine grace. The fire doesn't rage; it cautiously licks the dense green stems of a lush, succulent plant, its heat causing the air to shimmer. Large, velvety, deep purple bell-shaped flowers hang heavily, glowing intensely in the inferno's unnatural light, illuminating their delicate petal textures and casting dynamic, flickering highlights against the deep shadows of the surrounding woods. Every detail, from the subtle fuzz on the plant stems to the individual, crackling embers that drift upward like malevolent fireflies, is rendered in crisp, vibrant ultra-high definition, creating a menacing and alluring atmosphere.", + "seed": 27115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_3582_seed_27115_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028961136238649487, + "vlm_logprob": -0.028961136238649487, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.028961136238649487, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025223830714821815 + ], + "opposite_logprob_tokens": [ + -3.900223731994629 + ], + "logprob_reward": -0.025223830714821815, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025223830714821815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07189713418483734, + 0.07774227857589722, + 0.037442147731781006, + 0.08614394068717957, + 0.21780551970005035, + 0.06555428355932236, + 0.11480824649333954, + 0.055898815393447876, + 0.12063710391521454, + 0.1520705670118332 + ], + "credit_map": [ + 0.00900479219853878, + 0.012294868007302284, + 0.008747152984142303, + 0.007860714569687843, + 0.007292643189430237, + 0.00787489116191864, + 0.007405850104987621, + 0.011416223831474781, + 0.006009560078382492, + 0.005854756571352482, + 0.007935275323688984, + 0.0035831588320434093, + 0.00485021760687232, + 0.005915659014135599, + 0.0197870172560215, + 0.02380663901567459, + 0.004018051084131002, + 0.0033602151088416576, + 0.004475203342735767, + 0.004588145297020674, + 0.005767914466559887, + 0.005164779257029295, + 0.005225426983088255, + 0.004842410329729319, + 0.0074183084070682526, + 0.009693173691630363, + 0.010355593636631966, + 0.009211543016135693, + 0.013639967888593674, + 0.013423667289316654, + 0.013138786889612675, + 0.009262903593480587, + 0.019458740949630737, + 0.0085791926831007, + 0.00942224357277155, + 0.013142218813300133, + 0.02218593843281269, + 0.05456216260790825, + 0.07313849031925201, + 0.017316525802016258, + 0.007211813237518072, + 0.00854000635445118, + 0.007689138874411583, + 0.006103744730353355, + 0.009022881276905537, + 0.007231839466840029, + 0.009995011612772942, + 0.009759844280779362, + 0.011526614427566528, + 0.009778523817658424, + 0.016533227637410164, + 0.011742419563233852, + 0.01039869524538517, + 0.023064175620675087, + 0.009817937389016151, + 0.021946661174297333, + 0.010012357495725155, + 0.005857906304299831, + 0.007909855805337429, + 0.0064506856724619865, + 0.006855643354356289, + 0.007261761464178562, + 0.005991038400679827, + 0.00555956969037652, + 0.02296373061835766, + 0.01651671901345253, + 0.01420680247247219, + 0.01740860380232334, + 0.01201191358268261, + 0.01553222257643938, + 0.009853076189756393, + 0.012144026346504688, + 0.017110243439674377, + 0.024009887129068375, + 0.01669188216328621, + 0.011902830563485622, + 0.029651189222931862, + 0.020187219604849815, + 0.016885265707969666, + 0.015632057562470436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.025223830714821815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328970469534397 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328970469534397, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328970469534397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15321969985961914, + 0.07601307332515717, + 0.059540558606386185, + 0.06176688149571419, + 0.19856618344783783, + 0.0780181810259819, + 0.11841947585344315, + 0.08853530883789062, + 0.08472314476966858, + 0.08119752258062363 + ], + "credit_map": [ + 0.02376978099346161, + 0.028912046924233437, + 0.01624997891485691, + 0.01294607575982809, + 0.01353307906538248, + 0.019041940569877625, + 0.014808359555900097, + 0.023958437144756317, + 0.009480373933911324, + 0.008598980493843555, + 0.010196912102401257, + 0.006206210236996412, + 0.01155857928097248, + 0.011813337914645672, + 0.008531301282346249, + 0.009627380408346653, + 0.006232018582522869, + 0.007006383966654539, + 0.007443650159984827, + 0.006857689470052719, + 0.008277651853859425, + 0.007507718168199062, + 0.00636605080217123, + 0.009849395602941513, + 0.005908175837248564, + 0.005844707600772381, + 0.007441296707838774, + 0.005986570846289396, + 0.012996667996048927, + 0.009923545643687248, + 0.008253483101725578, + 0.005412432365119457, + 0.02357584796845913, + 0.011273153126239777, + 0.011299995705485344, + 0.017933892086148262, + 0.01905973069369793, + 0.03432712331414223, + 0.06541917473077774, + 0.01567726768553257, + 0.009890192188322544, + 0.008426451124250889, + 0.006226971745491028, + 0.009896003641188145, + 0.015624423511326313, + 0.009119443595409393, + 0.010796940885484219, + 0.0080377496778965, + 0.021161161363124847, + 0.013858986087143421, + 0.01754230074584484, + 0.0105635616928339, + 0.008204986341297626, + 0.02128249779343605, + 0.010240904986858368, + 0.015565071254968643, + 0.010747679509222507, + 0.01006022933870554, + 0.013434852473437786, + 0.015250287018716335, + 0.007496287580579519, + 0.011464281007647514, + 0.010857335291802883, + 0.009224358014762402, + 0.013560175895690918, + 0.00856785848736763, + 0.010220049880445004, + 0.013541369698941708, + 0.010905343107879162, + 0.008280361071228981, + 0.008650205098092556, + 0.010997779667377472, + 0.007249012589454651, + 0.010480917990207672, + 0.010087032802402973, + 0.006208476144820452, + 0.012137701734900475, + 0.01671532914042473, + 0.008552802726626396, + 0.009766251780092716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.02328970469534397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193400101736188 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.003193400101736188, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193400101736188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2933288514614105, + 0.06137501075863838, + 0.10901784896850586, + 0.07234165072441101, + 0.133030503988266, + 0.033805664628744125, + 0.0742311105132103, + 0.0897476077079773, + 0.04495708644390106, + 0.08816458284854889 + ], + "credit_map": [ + 0.030350832268595695, + 0.04728621616959572, + 0.037053462117910385, + 0.031092828139662743, + 0.03323584794998169, + 0.031386926770210266, + 0.02642875909805298, + 0.05649401620030403, + 0.007312350440770388, + 0.009412202052772045, + 0.008645870722830296, + 0.006144339218735695, + 0.008523448370397091, + 0.007421744521707296, + 0.005884748883545399, + 0.00803031399846077, + 0.011736967600882053, + 0.013233317993581295, + 0.012878982350230217, + 0.019044507294893265, + 0.012597850523889065, + 0.015068252570927143, + 0.012487583793699741, + 0.011970398016273975, + 0.00739012798294425, + 0.006400903221219778, + 0.010046013630926609, + 0.006488773971796036, + 0.013801736757159233, + 0.013186019845306873, + 0.008521633222699165, + 0.006506443955004215, + 0.014463420957326889, + 0.010612661018967628, + 0.009823257103562355, + 0.013205061666667461, + 0.024055158719420433, + 0.023553669452667236, + 0.023715617135167122, + 0.013601664453744888, + 0.003853225614875555, + 0.004410465247929096, + 0.003252362133935094, + 0.00366376549936831, + 0.0053084357641637325, + 0.003727235132828355, + 0.0051354956813156605, + 0.00445468258112669, + 0.009074264205992222, + 0.006962245795875788, + 0.01371720153838396, + 0.008825424127280712, + 0.006322666071355343, + 0.011718176305294037, + 0.0058013503439724445, + 0.011809784919023514, + 0.011522799730300903, + 0.00840426329523325, + 0.013660659082233906, + 0.016578083857893944, + 0.008832214400172234, + 0.011315466836094856, + 0.007111648563295603, + 0.012322479858994484, + 0.006250969134271145, + 0.006367696914821863, + 0.0058726160787045956, + 0.0080251544713974, + 0.005609055515378714, + 0.0044105504639446735, + 0.004148166161030531, + 0.004272885154932737, + 0.008792570792138577, + 0.009937218390405178, + 0.009368065744638443, + 0.0064520202577114105, + 0.014218557626008987, + 0.015909645706415176, + 0.012824105098843575, + 0.010662402957677841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.003193400101736188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091345705091953 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091345705091953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091345705091953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06873909384012222, + 0.08872168511152267, + 0.05971347168087959, + 0.04227643087506294, + 0.1187688335776329, + 0.036889705806970596, + 0.07376012206077576, + 0.12578168511390686, + 0.17361758649349213, + 0.2117314338684082 + ], + "credit_map": [ + 0.007200896739959717, + 0.009156431071460247, + 0.00770240044221282, + 0.009092207066714764, + 0.006572940852493048, + 0.007827326655387878, + 0.006939683109521866, + 0.014247199520468712, + 0.005819134879857302, + 0.006142452824860811, + 0.00695467833429575, + 0.004015169572085142, + 0.005343579221516848, + 0.0065180715173482895, + 0.02720184437930584, + 0.026726750656962395, + 0.006857091560959816, + 0.00551546411588788, + 0.006909899413585663, + 0.005603151861578226, + 0.009847109206020832, + 0.007291729561984539, + 0.008348136208951473, + 0.009340882301330566, + 0.003680573310703039, + 0.003517882199957967, + 0.005754126235842705, + 0.0038575599901378155, + 0.010854353196918964, + 0.007771302480250597, + 0.004038158804178238, + 0.002802471164613962, + 0.011626544408500195, + 0.007115380838513374, + 0.006600892171263695, + 0.010693344287574291, + 0.016640642657876015, + 0.024088865146040916, + 0.03107919730246067, + 0.010923946276307106, + 0.0058141849003732204, + 0.004234130959957838, + 0.004116314463317394, + 0.0032075431663542986, + 0.006020232103765011, + 0.004052854608744383, + 0.0048127188347280025, + 0.004631723277270794, + 0.009590769186615944, + 0.009711978957057, + 0.011707207188010216, + 0.007197739090770483, + 0.007081583607941866, + 0.010213353671133518, + 0.008230267092585564, + 0.010027219541370869, + 0.014725670218467712, + 0.011758671142160892, + 0.021255549043416977, + 0.018769005313515663, + 0.010938365943729877, + 0.0175457913428545, + 0.012717586942017078, + 0.01807103306055069, + 0.01903429813683033, + 0.024397628381848335, + 0.021022312343120575, + 0.02021351456642151, + 0.029033679515123367, + 0.02356140688061714, + 0.018191825598478317, + 0.01816289871931076, + 0.017990920692682266, + 0.02856939099729061, + 0.02198782004415989, + 0.015429272316396236, + 0.04011421650648117, + 0.029035817831754684, + 0.03333454206585884, + 0.02526942640542984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.004091345705091953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900739997625351 + ], + "opposite_logprob_tokens": [ + -2.4640073776245117 + ], + "logprob_reward": -0.08900739997625351, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900739997625351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1202288344502449, + 0.15107357501983643, + 0.06397397816181183, + 0.05966882035136223, + 0.23396337032318115, + 0.07107032090425491, + 0.06081131100654602, + 0.09452465176582336, + 0.03949356451630592, + 0.10519158095121384 + ], + "credit_map": [ + 0.011329359374940395, + 0.022727951407432556, + 0.01577867567539215, + 0.012304073199629784, + 0.014329193159937859, + 0.013177699409425259, + 0.010766508057713509, + 0.01981537789106369, + 0.010694087482988834, + 0.010073225013911724, + 0.012655210681259632, + 0.008529283106327057, + 0.009406337514519691, + 0.009611074812710285, + 0.040657129138708115, + 0.04944722726941109, + 0.00914549920707941, + 0.006202968303114176, + 0.009692054241895676, + 0.00849214382469654, + 0.008378415368497372, + 0.007969768717885017, + 0.007282673381268978, + 0.006810460239648819, + 0.005223321262747049, + 0.006964835338294506, + 0.006963960826396942, + 0.007301120553165674, + 0.008411495946347713, + 0.008207256905734539, + 0.011702029965817928, + 0.004894796293228865, + 0.025646312162280083, + 0.020982027053833008, + 0.014760102145373821, + 0.022732682526111603, + 0.05172449350357056, + 0.03652055934071541, + 0.03082381747663021, + 0.030773388221859932, + 0.007244876120239496, + 0.00831663329154253, + 0.005652153864502907, + 0.008233482949435711, + 0.007722093723714352, + 0.008125318214297295, + 0.014319207519292831, + 0.011456558480858803, + 0.0076273540034890175, + 0.007625655736774206, + 0.00940517894923687, + 0.0065833330154418945, + 0.005065872799605131, + 0.009326502680778503, + 0.005351176019757986, + 0.009826239198446274, + 0.011525680311024189, + 0.00884043239057064, + 0.013711079023778439, + 0.01770806685090065, + 0.009042765945196152, + 0.013029989786446095, + 0.007447549141943455, + 0.013219081796705723, + 0.005302754230797291, + 0.005627186503261328, + 0.00421214709058404, + 0.004471933469176292, + 0.006344060879200697, + 0.005501834210008383, + 0.00430126441642642, + 0.003732386976480484, + 0.00859079696238041, + 0.012949970550835133, + 0.014977180399000645, + 0.008265148848295212, + 0.017445597797632217, + 0.020787015557289124, + 0.010364273563027382, + 0.011811594478785992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.08900739997625351 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle tracking shot glides across the dark forest floor, rendered in a rich, painterly 3D cartoon style reminiscent of the \"Two Cities\" series. Unusual tendrils of a magical fire, a vibrant blend of fuchsia, violet, and fiery orange, creep across the damp soil and gnarled tree roots with serpentine grace. The fire doesn't rage; it cautiously licks the dense green stems of a lush, succulent plant, its heat causing the air to shimmer. Large, velvety, deep purple bell-shaped flowers hang heavily, glowing intensely in the inferno's unnatural light, illuminating their delicate petal textures and casting dynamic, flickering highlights against the deep shadows of the surrounding woods. Every detail, from the subtle fuzz on the plant stems to the individual, crackling embers that drift upward like malevolent fireflies, is rendered in crisp, vibrant ultra-high definition, creating a menacing and alluring atmosphere.", + "seed": 27116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_3582_seed_27116_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017450399324297906, + "vlm_logprob": -0.017450399324297906, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.017450399324297906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02283177524805069 + ], + "opposite_logprob_tokens": [ + -4.147831916809082 + ], + "logprob_reward": -0.02283177524805069, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02283177524805069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41771766543388367, + 0.03358958289027214, + 0.026574311777949333, + 0.04012895002961159, + 0.04361927509307861, + 0.08203450590372086, + 0.035142526030540466, + 0.044747937470674515, + 0.2433684766292572, + 0.033076800405979156 + ], + "credit_map": [ + 0.03012733906507492, + 0.016282912343740463, + 0.16644734144210815, + 0.037944160401821136, + 0.041601065546274185, + 0.015415706671774387, + 0.010827967897057533, + 0.09907115250825882, + 0.003515438875183463, + 0.0031462982296943665, + 0.0036325566470623016, + 0.004171156324446201, + 0.006819595117121935, + 0.0040823714807629585, + 0.004265419207513332, + 0.003956743516027927, + 0.002121272264048457, + 0.0035672797821462154, + 0.0038785277865827084, + 0.004253726452589035, + 0.00256584701128304, + 0.004346521571278572, + 0.0032815944869071245, + 0.002559543354436755, + 0.0037136797327548265, + 0.003916991408914328, + 0.006648713257163763, + 0.008594118058681488, + 0.004842274356633425, + 0.0034055530559271574, + 0.004448506981134415, + 0.004559114575386047, + 0.0042046476155519485, + 0.004207185935229063, + 0.0068767741322517395, + 0.007833478040993214, + 0.0037908831145614386, + 0.0062870788387954235, + 0.004968689754605293, + 0.005450537893921137, + 0.0053552850149571896, + 0.002561940113082528, + 0.005126162897795439, + 0.0064916606061160564, + 0.003659646026790142, + 0.03271074593067169, + 0.008121720515191555, + 0.018007349222898483, + 0.002613202203065157, + 0.003462692489847541, + 0.007035762537270784, + 0.006351225543767214, + 0.0030017918907105923, + 0.0035165215376764536, + 0.005082427524030209, + 0.0040789018385112286, + 0.0038241916336119175, + 0.005224087741225958, + 0.006759515963494778, + 0.010046823881566525, + 0.003469250164926052, + 0.006438888609409332, + 0.004862108733505011, + 0.004123068414628506, + 0.019967608153820038, + 0.041060060262680054, + 0.03523216024041176, + 0.03047644905745983, + 0.025782084092497826, + 0.05362560972571373, + 0.01769871823489666, + 0.019525771960616112, + 0.004182660486549139, + 0.004290162585675716, + 0.004285118542611599, + 0.0052525633946061134, + 0.005209801252931356, + 0.0024132675025612116, + 0.004603737033903599, + 0.0028394891414791346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.02283177524805069 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073176637291908 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.016073176637291908, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073176637291908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0870535597205162, + 0.05249379202723503, + 0.05210544914007187, + 0.030852779746055603, + 0.034299544990062714, + 0.16686896979808807, + 0.05371609330177307, + 0.36943957209587097, + 0.12316544353961945, + 0.030004803091287613 + ], + "credit_map": [ + 0.01135767251253128, + 0.006837555207312107, + 0.01924891211092472, + 0.008698978461325169, + 0.007340766489505768, + 0.007295170798897743, + 0.006630422081798315, + 0.019644085317850113, + 0.006024521309882402, + 0.006177886389195919, + 0.004921512212604284, + 0.005048711784183979, + 0.008977551944553852, + 0.008939636871218681, + 0.006212177220731974, + 0.006191789172589779, + 0.003635416505858302, + 0.0035664031747728586, + 0.008724644780158997, + 0.004197441041469574, + 0.005794723518192768, + 0.015464985743165016, + 0.005962246563285589, + 0.004759590141475201, + 0.003452237695455551, + 0.004924140404909849, + 0.004611382260918617, + 0.0046585542149841785, + 0.0042908694595098495, + 0.0022974309977144003, + 0.00364920892752707, + 0.0029689567163586617, + 0.0027167340740561485, + 0.004259857814759016, + 0.0045988368801772594, + 0.005636787042021751, + 0.004562548827379942, + 0.004192465450614691, + 0.0036313417367637157, + 0.004700977820903063, + 0.009311337023973465, + 0.0043422756716609, + 0.00720748957246542, + 0.011132397688925266, + 0.006832227576524019, + 0.07123976945877075, + 0.019665295258164406, + 0.037138182669878006, + 0.0047386218793690205, + 0.006693015340715647, + 0.007327372673898935, + 0.006411923561245203, + 0.010020018555223942, + 0.0072672562673687935, + 0.005798813886940479, + 0.005459073930978775, + 0.03439517691731453, + 0.06684687733650208, + 0.052947163581848145, + 0.05335439369082451, + 0.0388915054500103, + 0.06355879455804825, + 0.027600765228271484, + 0.031844887882471085, + 0.010097868740558624, + 0.0208546444773674, + 0.018046965822577477, + 0.015461204573512077, + 0.012618848122656345, + 0.02742880955338478, + 0.008738594129681587, + 0.009918507188558578, + 0.003623835975304246, + 0.0047447350807487965, + 0.00357927312143147, + 0.004121036734431982, + 0.004579750820994377, + 0.003067620797082782, + 0.003667988581582904, + 0.0026205608155578375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.016073176637291908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024927754420787096 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.0024927754420787096, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024927754420787096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49744245409965515, + 0.016684170812368393, + 0.028786946088075638, + 0.017951635643839836, + 0.02111419476568699, + 0.14977869391441345, + 0.07021067291498184, + 0.05250547081232071, + 0.11800026148557663, + 0.027525490149855614 + ], + "credit_map": [ + 0.03784786909818649, + 0.020206401124596596, + 0.18411821126937866, + 0.033129867166280746, + 0.04459221288561821, + 0.02273201383650303, + 0.014893905259668827, + 0.13992202281951904, + 0.0019076813478022814, + 0.0017644306644797325, + 0.002192507730796933, + 0.0018316516652703285, + 0.003625539829954505, + 0.0020564959850162268, + 0.001801469479687512, + 0.0015043957391753793, + 0.0034935944713652134, + 0.0030860172118991613, + 0.004942248575389385, + 0.003386067459359765, + 0.002952511887997389, + 0.004536094143986702, + 0.0027616063598543406, + 0.003628808306530118, + 0.0014183062594383955, + 0.0017905383137986064, + 0.002462583128362894, + 0.0024381878320127726, + 0.0033030188642442226, + 0.001727830502204597, + 0.0021552869584411383, + 0.002655884250998497, + 0.002089298563078046, + 0.002531354548409581, + 0.004402611404657364, + 0.003576544113457203, + 0.0018722633831202984, + 0.0029245580080896616, + 0.0018169415416195989, + 0.0019006248330697417, + 0.009137606248259544, + 0.004412202630192041, + 0.006807795260101557, + 0.011633397080004215, + 0.005742830689996481, + 0.06211090832948685, + 0.014499043114483356, + 0.035434920340776443, + 0.00682724267244339, + 0.009490960277616978, + 0.006679468788206577, + 0.007471130695194006, + 0.01383599080145359, + 0.009515389800071716, + 0.00784111488610506, + 0.00854937732219696, + 0.0050044613890349865, + 0.009091371670365334, + 0.0077780913561582565, + 0.008082189597189426, + 0.005208773538470268, + 0.008220897987484932, + 0.00455213338136673, + 0.004567557014524937, + 0.00991204846650362, + 0.020147655159235, + 0.01668243296444416, + 0.014776809141039848, + 0.012253944762051105, + 0.026548948138952255, + 0.008150461129844189, + 0.00952797383069992, + 0.0037968698889017105, + 0.003861761651933193, + 0.004080900456756353, + 0.00276179239153862, + 0.0031515148002654314, + 0.0037177731283009052, + 0.002433126326650381, + 0.003721751971170306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0024927754420787096 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00281461956910789 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.00281461956910789, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00281461956910789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02696436271071434, + 0.014757964760065079, + 0.020911427214741707, + 0.017504103481769562, + 0.023919546976685524, + 0.09878886491060257, + 0.03554974123835564, + 0.7047325372695923, + 0.041622769087553024, + 0.015248703770339489 + ], + "credit_map": [ + 0.0042313020676374435, + 0.003208873560652137, + 0.004457802511751652, + 0.0041214958764612675, + 0.0029321559704840183, + 0.002790650352835655, + 0.0027521238662302494, + 0.0024699592031538486, + 0.0018820129334926605, + 0.0017871749587357044, + 0.002117905067279935, + 0.001357951434329152, + 0.0025467367377132177, + 0.0018602277850732207, + 0.0017193785170093179, + 0.0014865781413391232, + 0.001524037797935307, + 0.00205521029420197, + 0.002793083665892482, + 0.0026470052544027567, + 0.0022120291832834482, + 0.004004412796348333, + 0.0030725826509296894, + 0.0026030661538243294, + 0.0018949275836348534, + 0.002156783128157258, + 0.0022637788206338882, + 0.002019945066422224, + 0.0018965465715155005, + 0.0019177639624103904, + 0.0020570182241499424, + 0.0032973389606922865, + 0.0018877352122217417, + 0.0018853716319426894, + 0.004059497732669115, + 0.004008475225418806, + 0.00198953365907073, + 0.004006148315966129, + 0.0020000815857201815, + 0.004082703962922096, + 0.005533361341804266, + 0.00282525853253901, + 0.00472236005589366, + 0.007028593681752682, + 0.004248793236911297, + 0.0414719358086586, + 0.011206086724996567, + 0.021752476692199707, + 0.0032976989168673754, + 0.004363995976746082, + 0.003589450614526868, + 0.00440101558342576, + 0.005512140225619078, + 0.004650693386793137, + 0.003974856808781624, + 0.005759889259934425, + 0.06173783913254738, + 0.12768599390983582, + 0.10053803026676178, + 0.10096578299999237, + 0.07397783547639847, + 0.120638906955719, + 0.053630389273166656, + 0.06555774807929993, + 0.003367658471688628, + 0.007068069651722908, + 0.006139960139989853, + 0.005397421307861805, + 0.004206440877169371, + 0.00921566691249609, + 0.002819198416545987, + 0.0034083519130945206, + 0.0022694915533065796, + 0.001604532590135932, + 0.0022853927221149206, + 0.0019622528925538063, + 0.0019122110679745674, + 0.0019330453360453248, + 0.0013109227875247598, + 0.001970855984836817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.00281461956910789 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303964972496033 + ], + "opposite_logprob_tokens": [ + -3.168039560317993 + ], + "logprob_reward": -0.04303964972496033, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303964972496033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.672208309173584, + 0.014656558632850647, + 0.01600751467049122, + 0.013457972556352615, + 0.009243682026863098, + 0.05357963964343071, + 0.022845173254609108, + 0.12083359807729721, + 0.058618828654289246, + 0.018548768013715744 + ], + "credit_map": [ + 0.05019836500287056, + 0.026714174076914787, + 0.25928807258605957, + 0.049074772745370865, + 0.06353431195020676, + 0.02585500106215477, + 0.01785893552005291, + 0.17968463897705078, + 0.0015720597002655268, + 0.0012650351272895932, + 0.0018411582568660378, + 0.0019260741537436843, + 0.0021801406983286142, + 0.0019296001410111785, + 0.0022806567139923573, + 0.0016618326772004366, + 0.001087675802409649, + 0.0011993413791060448, + 0.0034756367094814777, + 0.0012477777199819684, + 0.0016190916066989303, + 0.004784213844686747, + 0.0012845629826188087, + 0.0013092136941850185, + 0.0013018394820392132, + 0.0018661682261154056, + 0.0017513333586975932, + 0.002418649150058627, + 0.0015439356211572886, + 0.0014558745315298438, + 0.0016556110931560397, + 0.0014645600458607078, + 0.0007937282207421958, + 0.0011372750159353018, + 0.0014177256962284446, + 0.001586097409017384, + 0.0009164020884782076, + 0.0014621017035096884, + 0.0009748796001076698, + 0.0009554718853905797, + 0.002750867512077093, + 0.0018539902521297336, + 0.0022108787670731544, + 0.0036871489137411118, + 0.002222262555733323, + 0.022440476343035698, + 0.006780357100069523, + 0.011633651331067085, + 0.002029397524893284, + 0.0020731831900775433, + 0.004700182471424341, + 0.0038346340879797935, + 0.0021496769040822983, + 0.00217068730853498, + 0.0029812015127390623, + 0.0029062097892165184, + 0.011211025528609753, + 0.021064212545752525, + 0.017103619873523712, + 0.018074192106723785, + 0.012507575564086437, + 0.021149933338165283, + 0.009715232066810131, + 0.010007791221141815, + 0.004997739568352699, + 0.009992042556405067, + 0.008095474913716316, + 0.007262548431754112, + 0.006284261122345924, + 0.013006389141082764, + 0.004183799494057894, + 0.004796567838639021, + 0.0023782362695783377, + 0.003184432862326503, + 0.003088324097916484, + 0.003005302045494318, + 0.0015263506211340427, + 0.0012484616599977016, + 0.0025525931268930435, + 0.0015650654677301645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.04303964972496033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle tracking shot glides across the dark forest floor, rendered in a rich, painterly 3D cartoon style reminiscent of the \"Two Cities\" series. Unusual tendrils of a magical fire, a vibrant blend of fuchsia, violet, and fiery orange, creep across the damp soil and gnarled tree roots with serpentine grace. The fire doesn't rage; it cautiously licks the dense green stems of a lush, succulent plant, its heat causing the air to shimmer. Large, velvety, deep purple bell-shaped flowers hang heavily, glowing intensely in the inferno's unnatural light, illuminating their delicate petal textures and casting dynamic, flickering highlights against the deep shadows of the surrounding woods. Every detail, from the subtle fuzz on the plant stems to the individual, crackling embers that drift upward like malevolent fireflies, is rendered in crisp, vibrant ultra-high definition, creating a menacing and alluring atmosphere.", + "seed": 27117, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_3582_seed_27117_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01884819846600294, + "vlm_logprob": -0.01884819846600294, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.01884819846600294, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023441214114427567 + ], + "opposite_logprob_tokens": [ + -4.023441314697266 + ], + "logprob_reward": -0.023441214114427567, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.023441214114427567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13425405323505402, + 0.33669742941856384, + 0.07743754982948303, + 0.05407669395208359, + 0.03570060059428215, + 0.16109271347522736, + 0.0969519168138504, + 0.016120515763759613, + 0.06450095772743225, + 0.02316763624548912 + ], + "credit_map": [ + 0.03552476689219475, + 0.012260166928172112, + 0.02763037569820881, + 0.011869845911860466, + 0.0061904191970825195, + 0.010480974800884724, + 0.010185975581407547, + 0.02011151798069477, + 0.024172261357307434, + 0.0051405103877186775, + 0.13016603887081146, + 0.03845779597759247, + 0.005421916488558054, + 0.012905731797218323, + 0.011890174821019173, + 0.10854295641183853, + 0.008201521821320057, + 0.004860864486545324, + 0.021201282739639282, + 0.015547369606792927, + 0.0036326188128441572, + 0.009768282994627953, + 0.00537836691364646, + 0.008847231045365334, + 0.0055436850525438786, + 0.0062911924906075, + 0.0053558349609375, + 0.008268667384982109, + 0.006474933121353388, + 0.0067336480133235455, + 0.005884168669581413, + 0.009524555876851082, + 0.0029868942219763994, + 0.0037519624456763268, + 0.005634211469441652, + 0.008408593945205212, + 0.003028988605365157, + 0.0029391441494226456, + 0.004081213381141424, + 0.004869588650763035, + 0.007152488920837641, + 0.004169786348938942, + 0.06174556538462639, + 0.009364861994981766, + 0.00424948800355196, + 0.0032800177577883005, + 0.06215472146868706, + 0.008975764736533165, + 0.007432965561747551, + 0.01145713310688734, + 0.01863386668264866, + 0.009397844783961773, + 0.011708090081810951, + 0.009312916547060013, + 0.01695568673312664, + 0.012053401209414005, + 0.0011248752707615495, + 0.0010769505752250552, + 0.0042061591520905495, + 0.0018956281710416079, + 0.0011434080079197884, + 0.001413499703630805, + 0.0036164154298603535, + 0.0016435784054920077, + 0.0036377159412950277, + 0.003920633811503649, + 0.020671280100941658, + 0.004192029591649771, + 0.0029031713493168354, + 0.004615371581166983, + 0.00608307309448719, + 0.0184776708483696, + 0.0023238244466483593, + 0.0044340407475829124, + 0.004160534590482712, + 0.002113972557708621, + 0.0019344172906130552, + 0.0038637963589280844, + 0.002632588380947709, + 0.001704459311440587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.023441214114427567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190816044807434 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190816044807434, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190816044807434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11803599447011948, + 0.07980909198522568, + 0.0907396450638771, + 0.025976981967687607, + 0.032854970544576645, + 0.4062686562538147, + 0.03571475297212601, + 0.050565242767333984, + 0.11967647820711136, + 0.04035820811986923 + ], + "credit_map": [ + 0.023079639300704002, + 0.010250036604702473, + 0.021705731749534607, + 0.0212190430611372, + 0.006043311674147844, + 0.008446273393929005, + 0.011092127300798893, + 0.016199832782149315, + 0.0061089699156582355, + 0.0022528287954628468, + 0.02876901812851429, + 0.008565057069063187, + 0.0024373906198889017, + 0.003363102674484253, + 0.0041981711983680725, + 0.02411455661058426, + 0.007372966501861811, + 0.004327700939029455, + 0.03072132170200348, + 0.022503932937979698, + 0.0030120075680315495, + 0.011593157425522804, + 0.003482658416032791, + 0.007725902833044529, + 0.002815201412886381, + 0.0027715114410966635, + 0.004641484934836626, + 0.004150269087404013, + 0.001888879807665944, + 0.0035721641033887863, + 0.002248468343168497, + 0.0038890033029019833, + 0.003318052738904953, + 0.002822951879352331, + 0.005353313405066729, + 0.005494798067957163, + 0.002807480050250888, + 0.002720857271924615, + 0.004522251896560192, + 0.005815268028527498, + 0.01784360036253929, + 0.011397073976695538, + 0.15177203714847565, + 0.03166080266237259, + 0.012985027395188808, + 0.007690954487770796, + 0.15152353048324585, + 0.021395601332187653, + 0.0035206174943596125, + 0.002369930036365986, + 0.007458516396582127, + 0.004375155083835125, + 0.003722775960341096, + 0.00359231885522604, + 0.0067998627200722694, + 0.003875575726851821, + 0.004006016068160534, + 0.004223344847559929, + 0.01361746434122324, + 0.004733117297291756, + 0.004364649765193462, + 0.004140081349760294, + 0.01085300650447607, + 0.00462756585329771, + 0.006297665182501078, + 0.008124460466206074, + 0.03950919583439827, + 0.007669098675251007, + 0.005173502489924431, + 0.007777142338454723, + 0.009544423781335354, + 0.035580992698669434, + 0.003752815304324031, + 0.006311145611107349, + 0.009029180742800236, + 0.0034612957388162613, + 0.0030878791585564613, + 0.005983131472021341, + 0.004978741053491831, + 0.0037540183402597904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.018190816044807434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3134143054485321, + 0.17719465494155884, + 0.11391040682792664, + 0.0645219087600708, + 0.02869880199432373, + 0.12603986263275146, + 0.04132438451051712, + 0.038040608167648315, + 0.061164259910583496, + 0.0356907993555069 + ], + "credit_map": [ + 0.06982600688934326, + 0.025836046785116196, + 0.05773885175585747, + 0.050855617970228195, + 0.009653300978243351, + 0.027597300708293915, + 0.030021294951438904, + 0.041885897517204285, + 0.013125265017151833, + 0.0037302507553249598, + 0.06577977538108826, + 0.018519481644034386, + 0.00403858907520771, + 0.008028092794120312, + 0.006151778157800436, + 0.057821426540613174, + 0.011668359860777855, + 0.0075118327513337135, + 0.03231506794691086, + 0.02924392931163311, + 0.004695497453212738, + 0.009059465490281582, + 0.007104511372745037, + 0.012311745434999466, + 0.005216718185693026, + 0.005503165069967508, + 0.009227708913385868, + 0.010974913835525513, + 0.004238687455654144, + 0.006069117691367865, + 0.010873524472117424, + 0.012418080121278763, + 0.0032167192548513412, + 0.0020522396080195904, + 0.00453239306807518, + 0.004543587099760771, + 0.0024413508363068104, + 0.0023106238804757595, + 0.00470745749771595, + 0.004894433077424765, + 0.006231676787137985, + 0.010205362923443317, + 0.03130261227488518, + 0.016037188470363617, + 0.012956087477505207, + 0.006734458263963461, + 0.030728187412023544, + 0.011844303458929062, + 0.005173774436116219, + 0.0033821400720626116, + 0.008678560145199299, + 0.004742288030683994, + 0.0030028470791876316, + 0.003997464198619127, + 0.008136037737131119, + 0.0042112721130251884, + 0.003311117412522435, + 0.003976089414209127, + 0.010864664800465107, + 0.0036650614347308874, + 0.002502193208783865, + 0.002937756245955825, + 0.006798007991164923, + 0.003985720686614513, + 0.006017560139298439, + 0.004010019823908806, + 0.015520194545388222, + 0.003727139439433813, + 0.0029155698139220476, + 0.005369079299271107, + 0.005268773064017296, + 0.01833592914044857, + 0.002729369793087244, + 0.004712666850537062, + 0.006102883722633123, + 0.003826935077086091, + 0.003539731027558446, + 0.006240608636289835, + 0.005566849373281002, + 0.00297175464220345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019383705221116543 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019383705221116543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019383705221116543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09947241097688675, + 0.23644958436489105, + 0.05122741311788559, + 0.029187489300966263, + 0.021239440888166428, + 0.10840161889791489, + 0.1130141168832779, + 0.03958360105752945, + 0.2589621841907501, + 0.0424620658159256 + ], + "credit_map": [ + 0.016964057460427284, + 0.010215665213763714, + 0.01799583062529564, + 0.01615770347416401, + 0.005670824553817511, + 0.009353692643344402, + 0.00918567180633545, + 0.013928967528045177, + 0.01647140644490719, + 0.0038541469257324934, + 0.09118874371051788, + 0.025915423408150673, + 0.0038861408829689026, + 0.008716835640370846, + 0.007844270206987858, + 0.07857262343168259, + 0.005172582343220711, + 0.0022126224357634783, + 0.016014734283089638, + 0.009764610789716244, + 0.0015641050413250923, + 0.006441818084567785, + 0.0018933397950604558, + 0.008163605816662312, + 0.002377644879743457, + 0.004095443058758974, + 0.003144666785374284, + 0.004178342409431934, + 0.0028451522812247276, + 0.003274488030001521, + 0.003469144692644477, + 0.005802607629448175, + 0.0019189708400517702, + 0.0017060225363820791, + 0.004230211488902569, + 0.0036332604940980673, + 0.0017793173901736736, + 0.0019380763405933976, + 0.0030638461466878653, + 0.002969735534861684, + 0.004997934214770794, + 0.006430246867239475, + 0.03334435448050499, + 0.011061941273510456, + 0.006766916252672672, + 0.003656755667179823, + 0.03498085215687752, + 0.007162622641772032, + 0.00909099355340004, + 0.014989534392952919, + 0.02025521919131279, + 0.014451594091951847, + 0.012395777739584446, + 0.010698587633669376, + 0.01734512485563755, + 0.013787292875349522, + 0.003625363577157259, + 0.0032967000734061003, + 0.011601164005696774, + 0.0035264159087091684, + 0.0029760091565549374, + 0.0032659550197422504, + 0.006230532191693783, + 0.0050614639185369015, + 0.014758788980543613, + 0.017874805256724358, + 0.08591220527887344, + 0.018202761188149452, + 0.013827572576701641, + 0.016964904963970184, + 0.01877184398472309, + 0.07264929264783859, + 0.0033748960122466087, + 0.009235076606273651, + 0.006045900285243988, + 0.0037326733581721783, + 0.00463297450914979, + 0.0053246477618813515, + 0.0033958193380385637, + 0.006720077712088823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.0019383705221116543 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304010421037674 + ], + "opposite_logprob_tokens": [ + -3.1680400371551514 + ], + "logprob_reward": -0.04304010421037674, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304010421037674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2462814301252365, + 0.2636112570762634, + 0.0961240604519844, + 0.02300039865076542, + 0.12865661084651947, + 0.11858053505420685, + 0.03171592950820923, + 0.026755861937999725, + 0.041221000254154205, + 0.024052927270531654 + ], + "credit_map": [ + 0.04058481752872467, + 0.020646756514906883, + 0.04789226874709129, + 0.0515035055577755, + 0.011083941906690598, + 0.016452640295028687, + 0.021825164556503296, + 0.03629232197999954, + 0.01841104030609131, + 0.004033299628645182, + 0.10050015151500702, + 0.028684409335255623, + 0.0043845283798873425, + 0.00968476664274931, + 0.009633824229240417, + 0.08827922493219376, + 0.009587614797055721, + 0.003933376632630825, + 0.03200601041316986, + 0.02271091379225254, + 0.003040745621547103, + 0.01234337966889143, + 0.002880414016544819, + 0.00962160900235176, + 0.003274788148701191, + 0.0021288557909429073, + 0.004013868048787117, + 0.003976364620029926, + 0.0016507996479049325, + 0.0026402692310512066, + 0.0021136922296136618, + 0.0032017617486417294, + 0.011225043796002865, + 0.004720285069197416, + 0.02954631671309471, + 0.03770372271537781, + 0.004863858688622713, + 0.007337490096688271, + 0.01758490316569805, + 0.015675002709031105, + 0.005240378435701132, + 0.005770618095993996, + 0.03984595090150833, + 0.009357179515063763, + 0.0068104322999715805, + 0.004015698097646236, + 0.03928836062550545, + 0.0082519156858325, + 0.002274705097079277, + 0.0039667245000600815, + 0.00486001605167985, + 0.0030244786757975817, + 0.0033816483337432146, + 0.00430674385279417, + 0.00439226720482111, + 0.0055093481205403805, + 0.002471354091539979, + 0.003637338522821665, + 0.005151113495230675, + 0.0026010700967162848, + 0.002861343091353774, + 0.0025215689092874527, + 0.004746156744658947, + 0.00276591582223773, + 0.0025961133651435375, + 0.0030441817361861467, + 0.009833361022174358, + 0.004006141796708107, + 0.0027919751591980457, + 0.004014391452074051, + 0.006760912481695414, + 0.008173925802111626, + 0.001917736022733152, + 0.0029993848875164986, + 0.0036577775608748198, + 0.0028012071270495653, + 0.0026931443717330694, + 0.003530839690938592, + 0.003541444893926382, + 0.002911393065005541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.04304010421037674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle tracking shot glides across the dark forest floor, rendered in a rich, painterly 3D cartoon style reminiscent of the \"Two Cities\" series. Unusual tendrils of a magical fire, a vibrant blend of fuchsia, violet, and fiery orange, creep across the damp soil and gnarled tree roots with serpentine grace. The fire doesn't rage; it cautiously licks the dense green stems of a lush, succulent plant, its heat causing the air to shimmer. Large, velvety, deep purple bell-shaped flowers hang heavily, glowing intensely in the inferno's unnatural light, illuminating their delicate petal textures and casting dynamic, flickering highlights against the deep shadows of the surrounding woods. Every detail, from the subtle fuzz on the plant stems to the individual, crackling embers that drift upward like malevolent fireflies, is rendered in crisp, vibrant ultra-high definition, creating a menacing and alluring atmosphere.", + "seed": 27118, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_5/index_3582_seed_27118_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022778724040836095, + "vlm_logprob": -0.022778724040836095, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.022778724040836095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03263692185282707 + ], + "opposite_logprob_tokens": [ + -3.532636880874634 + ], + "logprob_reward": -0.03263692185282707, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03263692185282707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25420820713043213, + 0.11116744577884674, + 0.15087546408176422, + 0.06529101729393005, + 0.07719650119543076, + 0.18279807269573212, + 0.054345905780792236, + 0.02328389510512352, + 0.03253834694623947, + 0.048295147716999054 + ], + "credit_map": [ + 0.028505267575383186, + 0.033558186143636703, + 0.03577946871519089, + 0.0491517148911953, + 0.029567871242761612, + 0.026105189695954323, + 0.0313110277056694, + 0.02022945135831833, + 0.01139492355287075, + 0.00694875605404377, + 0.025665435940027237, + 0.013630139641463757, + 0.007268642541021109, + 0.010607701726257801, + 0.010273012332618237, + 0.025378834456205368, + 0.016470158472657204, + 0.017016617581248283, + 0.027560777962207794, + 0.016003580763936043, + 0.013508506119251251, + 0.022495653480291367, + 0.021383944898843765, + 0.016436224803328514, + 0.009773898869752884, + 0.01100503746420145, + 0.010205098427832127, + 0.0076345656998455524, + 0.005471508018672466, + 0.008354934863746166, + 0.004911094903945923, + 0.007934882305562496, + 0.009271214716136456, + 0.009336072020232677, + 0.011979654431343079, + 0.009068465791642666, + 0.011953306384384632, + 0.009538104757666588, + 0.006715546362102032, + 0.009334131143987179, + 0.016310054808855057, + 0.020360644906759262, + 0.035705115646123886, + 0.03688277676701546, + 0.017153305932879448, + 0.01196389738470316, + 0.021049203351140022, + 0.02337307296693325, + 0.007624502759426832, + 0.007002592086791992, + 0.008593826554715633, + 0.008919250220060349, + 0.007822989486157894, + 0.005109850782901049, + 0.00428452342748642, + 0.004988373257219791, + 0.0025028237141668797, + 0.002171298023313284, + 0.003186459420248866, + 0.0023322992492467165, + 0.0028223663102835417, + 0.003028780221939087, + 0.0039447760209441185, + 0.0032950916793197393, + 0.00459268456324935, + 0.00483893696218729, + 0.003832937916740775, + 0.002831441583111882, + 0.002702857833355665, + 0.004217992536723614, + 0.004559886176139116, + 0.004961609840393066, + 0.002709906082600355, + 0.004089782014489174, + 0.004173282999545336, + 0.007536835037171841, + 0.0023547315504401922, + 0.011829347349703312, + 0.004114992916584015, + 0.011486269533634186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.03263692185282707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184728920459747 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184728920459747, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184728920459747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20829981565475464, + 0.10155604779720306, + 0.1227516382932663, + 0.10767058283090591, + 0.10760258138179779, + 0.08056508749723434, + 0.04735158756375313, + 0.021412929520010948, + 0.11937996745109558, + 0.08340971916913986 + ], + "credit_map": [ + 0.02526496909558773, + 0.029194438830018044, + 0.03410233557224274, + 0.03865426033735275, + 0.02639579586684704, + 0.022037195041775703, + 0.017639044672250748, + 0.015011797659099102, + 0.014447877183556557, + 0.009657059796154499, + 0.013376399874687195, + 0.01891697384417057, + 0.007988150231540203, + 0.008817181922495365, + 0.015939289703965187, + 0.012413118965923786, + 0.014497616328299046, + 0.014073551632463932, + 0.019171232357621193, + 0.01226846408098936, + 0.010961492545902729, + 0.017207840457558632, + 0.014742732979357243, + 0.01982872560620308, + 0.01300234068185091, + 0.00902847945690155, + 0.018701571971178055, + 0.015765804797410965, + 0.009517896920442581, + 0.015475980006158352, + 0.010466991923749447, + 0.015711531043052673, + 0.007829171605408192, + 0.01226575393229723, + 0.01773139089345932, + 0.010307938791811466, + 0.024365054443478584, + 0.013945046812295914, + 0.009959178976714611, + 0.011199049651622772, + 0.016458451747894287, + 0.01143588311970234, + 0.01301152165979147, + 0.011942495591938496, + 0.006147513631731272, + 0.004181939177215099, + 0.007863392122089863, + 0.009523900225758553, + 0.004471322521567345, + 0.005379772279411554, + 0.0071012647822499275, + 0.006974842865020037, + 0.006733838003128767, + 0.008609021082520485, + 0.0034231350291520357, + 0.0046583982184529305, + 0.002196789951995015, + 0.0019860060419887304, + 0.003229513531550765, + 0.002274132100865245, + 0.002179401693865657, + 0.003214566269889474, + 0.0032222969457507133, + 0.0031102250795811415, + 0.022521698847413063, + 0.018218625336885452, + 0.01535069104284048, + 0.010062325745821, + 0.007011018693447113, + 0.01873360201716423, + 0.010713513940572739, + 0.01676851324737072, + 0.010009662248194218, + 0.004687367472797632, + 0.006036501377820969, + 0.0071600391529500484, + 0.0040091704577207565, + 0.025673706084489822, + 0.004102335777133703, + 0.02173093892633915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.018184728920459747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021998509764671326 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021998509764671326, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021998509764671326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2186473160982132, + 0.13754896819591522, + 0.11365170031785965, + 0.06615567952394485, + 0.05536394938826561, + 0.1125626266002655, + 0.13946260511875153, + 0.042869798839092255, + 0.030711384490132332, + 0.08302604407072067 + ], + "credit_map": [ + 0.0179362203925848, + 0.018812643364071846, + 0.02609189786016941, + 0.03353938087821007, + 0.03833485767245293, + 0.020640116184949875, + 0.03252498060464859, + 0.03076721727848053, + 0.013130797073245049, + 0.011674581095576286, + 0.02501879632472992, + 0.01771514303982258, + 0.010103334672749043, + 0.013490388169884682, + 0.02314542606472969, + 0.023270491510629654, + 0.010701688937842846, + 0.008007514290511608, + 0.02234484627842903, + 0.012184417806565762, + 0.006834901869297028, + 0.016652695834636688, + 0.01950751803815365, + 0.01741810329258442, + 0.008315435610711575, + 0.006200823001563549, + 0.011178393848240376, + 0.01028359029442072, + 0.006460994947701693, + 0.0073679289780557156, + 0.005998484790325165, + 0.010350017808377743, + 0.0038549124728888273, + 0.007310205139219761, + 0.009142357856035233, + 0.00751004321500659, + 0.007172570563852787, + 0.005779501050710678, + 0.005363520234823227, + 0.009230833500623703, + 0.017408138141036034, + 0.013220754452049732, + 0.020133458077907562, + 0.022101955488324165, + 0.01147952489554882, + 0.006485501769930124, + 0.008645674213767052, + 0.0130876200273633, + 0.016384676098823547, + 0.01863357238471508, + 0.01969670131802559, + 0.02196996845304966, + 0.02339770644903183, + 0.015683064237236977, + 0.013699641451239586, + 0.009997261688113213, + 0.0032169853802770376, + 0.0039212023839354515, + 0.006155994720757008, + 0.00643561827018857, + 0.003940097987651825, + 0.005455777980387211, + 0.007123674266040325, + 0.00662044296041131, + 0.0058962139301002026, + 0.003418667707592249, + 0.003531697206199169, + 0.002473789732903242, + 0.0021014546509832144, + 0.00386765506118536, + 0.0034477966837584972, + 0.005974106024950743, + 0.006393998395651579, + 0.0072401841171085835, + 0.010991430841386318, + 0.022051313892006874, + 0.006524840369820595, + 0.0110786072909832, + 0.004925056826323271, + 0.01382060069590807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0021998509764671326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942534189671278 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.005942534189671278, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942534189671278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11284065991640091, + 0.08749835938215256, + 0.03151825815439224, + 0.1432676464319229, + 0.05626887455582619, + 0.05874136835336685, + 0.20759007334709167, + 0.0879611000418663, + 0.05837324261665344, + 0.15594039857387543 + ], + "credit_map": [ + 0.013225382193922997, + 0.010108841583132744, + 0.022981304675340652, + 0.015907658264040947, + 0.011845584027469158, + 0.01308581605553627, + 0.012033218517899513, + 0.013652858324348927, + 0.007917121984064579, + 0.00595006812363863, + 0.020083285868167877, + 0.0101640485227108, + 0.0051148394122719765, + 0.00921888928860426, + 0.011464842595160007, + 0.017585264518857002, + 0.0032351468689739704, + 0.0032387576065957546, + 0.00588072557002306, + 0.0032302201725542545, + 0.0025020528119057417, + 0.005718817003071308, + 0.0038940429221838713, + 0.0038184956647455692, + 0.01859711855649948, + 0.022679639980196953, + 0.02221973054111004, + 0.01578458771109581, + 0.012844059616327286, + 0.020289737731218338, + 0.01282764133065939, + 0.01802513189613819, + 0.005125910509377718, + 0.005871412809938192, + 0.009168989025056362, + 0.007474856451153755, + 0.006479929201304913, + 0.006558197550475597, + 0.006806045304983854, + 0.008783531375229359, + 0.007125995121896267, + 0.007085892837494612, + 0.00923311710357666, + 0.00846836157143116, + 0.0066771721467375755, + 0.0041513643227517605, + 0.008108865469694138, + 0.00789059977978468, + 0.0249567199498415, + 0.0267807487398386, + 0.032891977578401566, + 0.031133143231272697, + 0.031002625823020935, + 0.02755351923406124, + 0.01793600060045719, + 0.015335326083004475, + 0.011582911014556885, + 0.010287538170814514, + 0.017020342871546745, + 0.013325247913599014, + 0.00857436005026102, + 0.01046403031796217, + 0.008679041638970375, + 0.008027630858123302, + 0.011765454895794392, + 0.007929196581244469, + 0.005757207982242107, + 0.005519356112927198, + 0.003533869981765747, + 0.009338859468698502, + 0.006010935176163912, + 0.00851836521178484, + 0.013271518982946873, + 0.013445249758660793, + 0.016026807948946953, + 0.03244263306260109, + 0.01386930514127016, + 0.028567874804139137, + 0.010953926481306553, + 0.027363089844584465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.005942534189671278 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle tracking shot across the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a rich, painterly 3D cartoon reminiscent of the 'Two Cities' series?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tendrils of magical fire blending fuchsia, violet, and fiery orange visible creeping across the soil and roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large, velvety, deep purple bell-shaped flowers hanging heavily and glowing intensely in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are individual crackling embers drifting upward like malevolent fireflies visible in the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492958426475525 + ], + "opposite_logprob_tokens": [ + -2.929929494857788 + ], + "logprob_reward": -0.05492958426475525, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492958426475525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1832546591758728, + 0.3553294539451599, + 0.06018739938735962, + 0.07170746475458145, + 0.07683655619621277, + 0.07781423628330231, + 0.04944533482193947, + 0.026242148131132126, + 0.05126197636127472, + 0.04792078956961632 + ], + "credit_map": [ + 0.0188163910061121, + 0.01901058666408062, + 0.025650063529610634, + 0.031347136944532394, + 0.025963077321648598, + 0.0192432664334774, + 0.024803636595606804, + 0.018420495092868805, + 0.029694747179746628, + 0.017811300233006477, + 0.06794136762619019, + 0.03677975386381149, + 0.022456200793385506, + 0.04285489767789841, + 0.0687822476029396, + 0.06900893896818161, + 0.008144255727529526, + 0.005831947550177574, + 0.011264609172940254, + 0.005762294866144657, + 0.004862872418016195, + 0.007052829954773188, + 0.008794010616838932, + 0.008474575355648994, + 0.008192252367734909, + 0.010904773138463497, + 0.010590543039143085, + 0.008381316438317299, + 0.008215507492423058, + 0.010899125598371029, + 0.00704073766246438, + 0.007483214605599642, + 0.004806539975106716, + 0.00941408146172762, + 0.011242736130952835, + 0.010144189931452274, + 0.015494082123041153, + 0.00907160434871912, + 0.007193942554295063, + 0.009469384327530861, + 0.015389498323202133, + 0.010710612870752811, + 0.012025943025946617, + 0.012682502157986164, + 0.008595844730734825, + 0.003178585320711136, + 0.007140409201383591, + 0.008090835995972157, + 0.004948283080011606, + 0.007425577379763126, + 0.007507542613893747, + 0.00819962564855814, + 0.006693513132631779, + 0.006282814312726259, + 0.004514902364462614, + 0.0038730783853679895, + 0.0024016464594751596, + 0.0019459747709333897, + 0.0030166844371706247, + 0.0038246142212301493, + 0.002638377482071519, + 0.0030392459593713284, + 0.005217667203396559, + 0.0041579389944672585, + 0.006822445895522833, + 0.004927718546241522, + 0.004812948405742645, + 0.003957575652748346, + 0.004126527346670628, + 0.012055367231369019, + 0.00453386502340436, + 0.010025528259575367, + 0.006032134871929884, + 0.006723256781697273, + 0.004648229107260704, + 0.0056867715902626514, + 0.003948580008000135, + 0.00959378108382225, + 0.0034174774773418903, + 0.007870560511946678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.05492958426475525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_tensor([2521, 2521])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_tensor([2521, 2521])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..425c38b982acf4ced546394b17d5a0627ad7c854 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_tensor([2521, 2521])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up depicts a red pen tumbling and spinning in a desolate void. The entire scene is filmed at a high frame rate, transforming the chaotic action into a slow, mesmerizing ballet. The pen, a retractable ballpoint made of polished crimson plastic, rotates gracefully. Demonic cinematic lighting, derived from a single, hard, off-camera light source, carves the object out of infinite blackness, casting long, sharp shadows across its own form and creating stark, bright highlights on its gleaming metal clip and tip. Every minute rotation and reflection is rendered in perfect, crystal-clear detail, capturing the hypnotic glow of the polished plastic and gleaming metal, creating an atmosphere both elegant and deeply unsettling as the pen turns relentlessly in the silent abyss.", + "seed": 30426, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_2521_seed_30426_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05480115090031177, + "vlm_logprob": -0.05480115090031177, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05480115090031177, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20641689002513885 + ], + "opposite_logprob_tokens": [ + -1.7064168453216553 + ], + "logprob_reward": -0.20641689002513885, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20641689002513885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021253230050206184, + 0.021075718104839325, + 0.02047124132514, + 0.013746609911322594, + 0.6589025259017944, + 0.00525338388979435, + 0.0876760184764862, + 0.012692892923951149, + 0.12866124510765076, + 0.030267145484685898 + ], + "credit_map": [ + 0.002414119429886341, + 0.0021142964251339436, + 0.0032172496430575848, + 0.0034618638455867767, + 0.0015617829049006104, + 0.003603484947234392, + 0.002512689447030425, + 0.002367743756622076, + 0.0022489004768431187, + 0.001983136869966984, + 0.0028897924348711967, + 0.002182493917644024, + 0.0025136047042906284, + 0.002220122842118144, + 0.0030155389104038477, + 0.004022128414362669, + 0.002129273023456335, + 0.002147800987586379, + 0.0025490710977464914, + 0.0023427139967679977, + 0.002560920547693968, + 0.0030326920095831156, + 0.002288066316395998, + 0.0034207042772322893, + 0.0018927889177575707, + 0.001815249677747488, + 0.0017213848186656833, + 0.001214612741023302, + 0.002061333041638136, + 0.001636427827179432, + 0.0020357940811663866, + 0.0013690185733139515, + 0.03274190053343773, + 0.015103030949831009, + 0.2564072906970978, + 0.02779202349483967, + 0.011979199945926666, + 0.2651732861995697, + 0.01630854234099388, + 0.03339725732803345, + 0.0005947832833044231, + 0.00042670563561841846, + 0.0007584583945572376, + 0.0005568013293668628, + 0.00037543667713180184, + 0.0007443425711244345, + 0.0006388618494383991, + 0.0011579939164221287, + 0.0038997456431388855, + 0.0018276546616107225, + 0.03585653752088547, + 0.0038694466929882765, + 0.0026503177359700203, + 0.03482240438461304, + 0.0021343319676816463, + 0.0026155770756304264, + 0.0020609062630683184, + 0.0011528851464390755, + 0.0015740288654342294, + 0.001295309397391975, + 0.0013780940789729357, + 0.0015449104830622673, + 0.0013469432014971972, + 0.0023398150224238634, + 0.006093970034271479, + 0.0033733402378857136, + 0.05106305330991745, + 0.00480464706197381, + 0.006048502866178751, + 0.050415486097335815, + 0.00363048049621284, + 0.003231768263503909, + 0.0024791923351585865, + 0.0017750323750078678, + 0.007210050709545612, + 0.0030647204257547855, + 0.003676903434097767, + 0.007162763271480799, + 0.002886510221287608, + 0.002011972013860941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.20641689002513885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373025730252266 + ], + "opposite_logprob_tokens": [ + -3.4087302684783936 + ], + "logprob_reward": -0.03373025730252266, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373025730252266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04527252912521362, + 0.014636991545557976, + 0.00793575681746006, + 0.03043794073164463, + 0.05510086938738823, + 0.010025015100836754, + 0.13558760285377502, + 0.011532239615917206, + 0.6712115406990051, + 0.01825953833758831 + ], + "credit_map": [ + 0.005815152078866959, + 0.004532458260655403, + 0.006239200476557016, + 0.006486065220087767, + 0.004123482387512922, + 0.007747155614197254, + 0.005179022438824177, + 0.005149989854544401, + 0.0016684678848832846, + 0.0017693190602585673, + 0.001769056310877204, + 0.0012200045166537166, + 0.001847589504905045, + 0.002640574472025037, + 0.0017312215641140938, + 0.0019907555542886257, + 0.0007806845242157578, + 0.00097298389300704, + 0.0012133470736443996, + 0.0007422322523780167, + 0.0008530816994607449, + 0.0011045351857319474, + 0.0008503674762323499, + 0.0014185241889208555, + 0.0034678883384913206, + 0.0034451091196388006, + 0.003912185784429312, + 0.0057376655749976635, + 0.002474625129252672, + 0.0041396054439246655, + 0.0029422331135720015, + 0.004318627063184977, + 0.0034812588710337877, + 0.00496970908716321, + 0.018119527027010918, + 0.003980046138167381, + 0.0020681656897068024, + 0.017293371260166168, + 0.002138350857421756, + 0.0030504290480166674, + 0.0010700575076043606, + 0.0007274700328707695, + 0.001433033961802721, + 0.0009125979850068688, + 0.0006089051021263003, + 0.0010068933479487896, + 0.0026028421707451344, + 0.0016632144106552005, + 0.006200808100402355, + 0.002839300548657775, + 0.055731844156980515, + 0.0049799601547420025, + 0.003598096314817667, + 0.05588187277317047, + 0.002963828155770898, + 0.003391892183572054, + 0.00173859519418329, + 0.0010494512971490622, + 0.0013961793156340718, + 0.0011687942314893007, + 0.0010899518383666873, + 0.0017270664684474468, + 0.001585937337949872, + 0.0017762633506208658, + 0.03136484697461128, + 0.014328273013234138, + 0.281011700630188, + 0.026118775829672813, + 0.016308346763253212, + 0.2709420621395111, + 0.016242971643805504, + 0.01489455346018076, + 0.0014335891464725137, + 0.001194789889268577, + 0.003462751628831029, + 0.002322864020243287, + 0.001818981021642685, + 0.0033786590211093426, + 0.0027361586689949036, + 0.0019117422634735703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.03373025730252266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019431296968832612 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.0019431296968832612, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019431296968832612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017347417771816254, + 0.01260423194617033, + 0.02831137552857399, + 0.014330906793475151, + 0.2605099380016327, + 0.008885701186954975, + 0.26258185505867004, + 0.00525645911693573, + 0.3562523126602173, + 0.033919766545295715 + ], + "credit_map": [ + 0.001891993684694171, + 0.0019317559199407697, + 0.002159888157621026, + 0.0035695252008736134, + 0.0012728393776342273, + 0.002207912039011717, + 0.0015877902042120695, + 0.002725711790844798, + 0.0012336252257227898, + 0.0014028477016836405, + 0.0015900644939392805, + 0.0012084349291399121, + 0.0011835969053208828, + 0.0017177119152620435, + 0.002097712829709053, + 0.002170235849916935, + 0.00353219429962337, + 0.0031100804917514324, + 0.0045018489472568035, + 0.0037967634852975607, + 0.0018342655384913087, + 0.003891840111464262, + 0.003026483580470085, + 0.004617895465344191, + 0.0018240561475977302, + 0.001533761154860258, + 0.00169304886367172, + 0.002742673270404339, + 0.0013774008257314563, + 0.0020996523089706898, + 0.001268278225325048, + 0.0017920348327606916, + 0.012870113365352154, + 0.005676692817360163, + 0.10379421710968018, + 0.010409745387732983, + 0.004015852231532335, + 0.10649669170379639, + 0.006118797231465578, + 0.011127790436148643, + 0.0010229081381112337, + 0.0007238672114908695, + 0.0011809610296040773, + 0.0012500437442213297, + 0.0005560777499340475, + 0.0007837662706151605, + 0.0019312453223392367, + 0.0014368308475241065, + 0.011407213285565376, + 0.00483879167586565, + 0.11105984449386597, + 0.00903226900845766, + 0.006392249371856451, + 0.11015398055315018, + 0.004119643475860357, + 0.005577864125370979, + 0.0007736458792351186, + 0.0005307378596626222, + 0.0006932188989594579, + 0.000721320160664618, + 0.00043023188482038677, + 0.00047487131087109447, + 0.0006547801895067096, + 0.0009776521474123001, + 0.016756456345319748, + 0.00777945201843977, + 0.14868411421775818, + 0.014006699435412884, + 0.008274062536656857, + 0.14446932077407837, + 0.008324893191456795, + 0.00795727502554655, + 0.00452996464446187, + 0.002065698616206646, + 0.004941897001117468, + 0.004817423410713673, + 0.002850165357813239, + 0.005034816917032003, + 0.004852491896599531, + 0.004827307537198067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.0019431296968832612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017592554911971092, + 0.0133074214681983, + 0.012632402591407299, + 0.09375321865081787, + 0.07839956134557724, + 0.01204807497560978, + 0.04730864241719246, + 0.012536343187093735, + 0.6965838670730591, + 0.01583791896700859 + ], + "credit_map": [ + 0.0022324579767882824, + 0.001910009072162211, + 0.0025216976646333933, + 0.0022863748017698526, + 0.0013651001499965787, + 0.002840850967913866, + 0.002150906715542078, + 0.0022851561661809683, + 0.0014954302459955215, + 0.001679600216448307, + 0.0017520918045192957, + 0.0008977668476291001, + 0.001622301759198308, + 0.002176631474867463, + 0.00154828408267349, + 0.002135314978659153, + 0.0014474376803264022, + 0.0014138821279630065, + 0.001841665362007916, + 0.0013483903603628278, + 0.001067603356204927, + 0.0018030439969152212, + 0.0014401937369257212, + 0.0022701865527778864, + 0.010233144275844097, + 0.011250597424805164, + 0.013368260115385056, + 0.018655138090252876, + 0.006998069584369659, + 0.011617009527981281, + 0.008522279560565948, + 0.013108717277646065, + 0.004542670212686062, + 0.003223209176212549, + 0.025025762617588043, + 0.005801052320748568, + 0.001901881187222898, + 0.02567482925951481, + 0.0036927543114870787, + 0.008537410758435726, + 0.0014018173096701503, + 0.0007606215076521039, + 0.0016584077384322882, + 0.0009975468274205923, + 0.0007385289645753801, + 0.0010716632241383195, + 0.0034648000728338957, + 0.0019546893890947104, + 0.0022098268382251263, + 0.0015407353639602661, + 0.018106725066900253, + 0.0023056897334754467, + 0.001243583974428475, + 0.017867300659418106, + 0.001619271351955831, + 0.002415511989966035, + 0.0020002820529043674, + 0.0011098658433184028, + 0.0013261071871966124, + 0.0011072743218392134, + 0.0012987955706194043, + 0.001833345741033554, + 0.0016372845275327563, + 0.002223388059064746, + 0.032710131257772446, + 0.015086048282682896, + 0.2906470000743866, + 0.027149686589837074, + 0.017327921465039253, + 0.2818972170352936, + 0.016631564125418663, + 0.015134286135435104, + 0.0018352732295170426, + 0.0008704340434633195, + 0.003721441375091672, + 0.0017134648514911532, + 0.0009628681582398713, + 0.0035188172478228807, + 0.0012209918349981308, + 0.001994626596570015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.0086353225633502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280154913663864 + ], + "opposite_logprob_tokens": [ + -3.773280143737793 + ], + "logprob_reward": -0.023280154913663864, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280154913663864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022274136543273926, + 0.019131291657686234, + 0.0131515022367239, + 0.01500727143138647, + 0.2636093199253082, + 0.012093761004507542, + 0.06803730130195618, + 0.021876143291592598, + 0.5382357835769653, + 0.026583518832921982 + ], + "credit_map": [ + 0.003456665202975273, + 0.0028780328575521708, + 0.0032279014121741056, + 0.0029799884650856256, + 0.0018127383664250374, + 0.0026462615933269262, + 0.002423756057396531, + 0.002848793985322118, + 0.0022371909581124783, + 0.0020784311927855015, + 0.0026372813154011965, + 0.001819995348341763, + 0.0018140077590942383, + 0.00202691531740129, + 0.0024855234660208225, + 0.004031947813928127, + 0.0015307466965168715, + 0.0014081696281209588, + 0.0023232880048453808, + 0.0015011116629466414, + 0.0006781172123737633, + 0.002064557047560811, + 0.0011378767667338252, + 0.002507636323571205, + 0.002012896351516247, + 0.0015068846987560391, + 0.0017223706236109138, + 0.0027623912319540977, + 0.0015084802871569991, + 0.001856029499322176, + 0.001415015896782279, + 0.0022232038900256157, + 0.012972880154848099, + 0.0053705014288425446, + 0.10689030587673187, + 0.009489104151725769, + 0.0038035528268665075, + 0.10954245924949646, + 0.005311507266014814, + 0.010229053907096386, + 0.0014562794240191579, + 0.0012036822736263275, + 0.0016340563306584954, + 0.0013213083148002625, + 0.0009873787639662623, + 0.001671088975854218, + 0.0014276610454544425, + 0.002392306225374341, + 0.0030977234710007906, + 0.002097910735756159, + 0.026150308549404144, + 0.0026579848490655422, + 0.002435762668028474, + 0.026166534051299095, + 0.0020830826833844185, + 0.003347994526848197, + 0.0033066158648580313, + 0.0019353518728166819, + 0.002080505946651101, + 0.0020438784267753363, + 0.0035096139181405306, + 0.003185753710567951, + 0.0026413574814796448, + 0.00317306793294847, + 0.025250717997550964, + 0.010829943232238293, + 0.22720998525619507, + 0.020889993757009506, + 0.011338130570948124, + 0.21770818531513214, + 0.013298279605805874, + 0.011710564605891705, + 0.0025179770309478045, + 0.0020270715467631817, + 0.0038002151995897293, + 0.00562505004927516, + 0.0023213778622448444, + 0.0032441189978271723, + 0.0033235037699341774, + 0.0037242069374769926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.023280154913663864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up depicts a red pen tumbling and spinning in a desolate void. The entire scene is filmed at a high frame rate, transforming the chaotic action into a slow, mesmerizing ballet. The pen, a retractable ballpoint made of polished crimson plastic, rotates gracefully. Demonic cinematic lighting, derived from a single, hard, off-camera light source, carves the object out of infinite blackness, casting long, sharp shadows across its own form and creating stark, bright highlights on its gleaming metal clip and tip. Every minute rotation and reflection is rendered in perfect, crystal-clear detail, capturing the hypnotic glow of the polished plastic and gleaming metal, creating an atmosphere both elegant and deeply unsettling as the pen turns relentlessly in the silent abyss.", + "seed": 30427, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_2521_seed_30427_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03926250953227282, + "vlm_logprob": -0.03926250953227282, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.03926250953227282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014716053381562233 + ], + "opposite_logprob_tokens": [ + -4.514716148376465 + ], + "logprob_reward": -0.014716053381562233, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014716053381562233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3411601185798645, + 0.04552847146987915, + 0.0664004236459732, + 0.0807482972741127, + 0.24968211352825165, + 0.024482401087880135, + 0.043146826326847076, + 0.0891762301325798, + 0.02002197690308094, + 0.03965318575501442 + ], + "credit_map": [ + 0.01953451707959175, + 0.012224990874528885, + 0.09363622218370438, + 0.018498362973332405, + 0.026154864579439163, + 0.12150916457176208, + 0.017627596855163574, + 0.031974390149116516, + 0.005420905537903309, + 0.0073082102462649345, + 0.0062424615025520325, + 0.007251528557389975, + 0.004358403384685516, + 0.004788804333657026, + 0.005517428740859032, + 0.004640729632228613, + 0.005464539863169193, + 0.007120431866496801, + 0.009099272079765797, + 0.011808282695710659, + 0.007496797014027834, + 0.008421119302511215, + 0.008094233460724354, + 0.008895743638277054, + 0.006629118230193853, + 0.006616188678890467, + 0.01944344863295555, + 0.010049514472484589, + 0.004947032313793898, + 0.008083047345280647, + 0.013334907591342926, + 0.011645045131444931, + 0.02002759650349617, + 0.014278310351073742, + 0.018055254593491554, + 0.021509511396288872, + 0.02104458212852478, + 0.021639341488480568, + 0.06058479845523834, + 0.0725427195429802, + 0.002863534027710557, + 0.002206176985055208, + 0.0033011280465871096, + 0.0019458523020148277, + 0.003488773014396429, + 0.003116627223789692, + 0.002513001672923565, + 0.005047305952757597, + 0.004896113649010658, + 0.0032012038864195347, + 0.010744012892246246, + 0.0032684216275811195, + 0.005362249910831451, + 0.006832461804151535, + 0.0028848741203546524, + 0.00595749169588089, + 0.01097420509904623, + 0.010765325278043747, + 0.022739436477422714, + 0.007927670143544674, + 0.008374983444809914, + 0.017055941745638847, + 0.0060149673372507095, + 0.005323698278516531, + 0.0018564133206382394, + 0.0018286111298948526, + 0.003185549285262823, + 0.0029587147291749716, + 0.002704115817323327, + 0.003078710986301303, + 0.0021566508803516626, + 0.002253210172057152, + 0.002368774963542819, + 0.002786858705803752, + 0.00350335449911654, + 0.0035569239407777786, + 0.0031387009657919407, + 0.006575030740350485, + 0.008622369728982449, + 0.009101171977818012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.014716053381562233 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865730457007885 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.00865730457007885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865730457007885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30194875597953796, + 0.03644033893942833, + 0.0312676727771759, + 0.2715710401535034, + 0.09033825993537903, + 0.061638910323381424, + 0.05388808250427246, + 0.045445870608091354, + 0.025265008211135864, + 0.08219610899686813 + ], + "credit_map": [ + 0.020473532378673553, + 0.013712230138480663, + 0.08954343944787979, + 0.017993010580539703, + 0.02028808370232582, + 0.10028533637523651, + 0.016624711453914642, + 0.02302839234471321, + 0.004288429394364357, + 0.0045310016721487045, + 0.004989471286535263, + 0.005529535934329033, + 0.003022346878424287, + 0.004583723843097687, + 0.005072455387562513, + 0.0044233715161681175, + 0.003102551680058241, + 0.004529115278273821, + 0.004141736309975386, + 0.004719509743154049, + 0.003502318635582924, + 0.0021625994704663754, + 0.004608185961842537, + 0.004501652903854847, + 0.015727948397397995, + 0.013139656744897366, + 0.03198905661702156, + 0.02549680322408676, + 0.012103183194994926, + 0.01514770369976759, + 0.07889553159475327, + 0.07907111942768097, + 0.00873201061040163, + 0.009071901440620422, + 0.010464929044246674, + 0.009886176325380802, + 0.013102255761623383, + 0.015534402802586555, + 0.012876530177891254, + 0.010670044459402561, + 0.0089105861261487, + 0.004928633105009794, + 0.007817459292709827, + 0.003626548917964101, + 0.014492245391011238, + 0.007250317372381687, + 0.007162829395383596, + 0.0074502844363451, + 0.006942685693502426, + 0.00404400983825326, + 0.00882396474480629, + 0.007210285868495703, + 0.007923713885247707, + 0.007089583203196526, + 0.004223559517413378, + 0.007630272302776575, + 0.005118194036185741, + 0.006224664859473705, + 0.007489203475415707, + 0.0045281811617314816, + 0.00815914012491703, + 0.006754708010703325, + 0.0033488732296973467, + 0.003822901751846075, + 0.0022319674026221037, + 0.0020409689750522375, + 0.0038682371377944946, + 0.004124644212424755, + 0.002425667131319642, + 0.004724113270640373, + 0.0023281420581042767, + 0.003521265694871545, + 0.003379848785698414, + 0.0032727280631661415, + 0.004220114089548588, + 0.005579464137554169, + 0.003822020022198558, + 0.008213327266275883, + 0.026732029393315315, + 0.02697657234966755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.00865730457007885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821514382958412 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821514382958412, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821514382958412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3200386166572571, + 0.039348650723695755, + 0.03525344282388687, + 0.10143581032752991, + 0.04581061378121376, + 0.061432648450136185, + 0.09103096276521683, + 0.11672757565975189, + 0.0524200014770031, + 0.13650169968605042 + ], + "credit_map": [ + 0.0222726222127676, + 0.012593152932822704, + 0.09605072438716888, + 0.01727142743766308, + 0.01582711935043335, + 0.11254506558179855, + 0.018612999469041824, + 0.024865515530109406, + 0.0060677905566990376, + 0.0037819917779415846, + 0.005657211411744356, + 0.0060331993736326694, + 0.00324934720993042, + 0.0039020534604787827, + 0.0037418222054839134, + 0.0069152358919382095, + 0.005472100339829922, + 0.00331467529758811, + 0.005292548798024654, + 0.005309754516929388, + 0.0022921047639101744, + 0.0020913199987262487, + 0.003498984267935157, + 0.00798195879906416, + 0.007760700769722462, + 0.005960824899375439, + 0.0336574986577034, + 0.020545605570077896, + 0.003511773655191064, + 0.0048078023828566074, + 0.015254373662173748, + 0.009937228634953499, + 0.005637367255985737, + 0.0034422697499394417, + 0.004130679648369551, + 0.005319339223206043, + 0.004591322969645262, + 0.00793618056923151, + 0.007661122363060713, + 0.00709233433008194, + 0.00829667691141367, + 0.005639351438730955, + 0.010048439726233482, + 0.008250132203102112, + 0.00885795708745718, + 0.008110437542200089, + 0.006746591068804264, + 0.005483063403517008, + 0.0058799320831894875, + 0.0018072191160172224, + 0.03319134935736656, + 0.0031221224926412106, + 0.004148066509515047, + 0.035366956144571304, + 0.002094864146783948, + 0.0054204631596803665, + 0.011894645169377327, + 0.014881320297718048, + 0.020155461505055428, + 0.012943007051944733, + 0.021544164046645164, + 0.012821586802601814, + 0.011795914731919765, + 0.010691480711102486, + 0.003908198792487383, + 0.003249721135944128, + 0.009935651905834675, + 0.008706653490662575, + 0.005277853459119797, + 0.011666402220726013, + 0.0033824366983026266, + 0.0062930891290307045, + 0.004391052760183811, + 0.0034319162368774414, + 0.00583764910697937, + 0.009899751283228397, + 0.006318490020930767, + 0.009298835881054401, + 0.0476204976439476, + 0.049703504890203476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.002821514382958412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3399501442909241, + 0.04723434895277023, + 0.038539379835128784, + 0.09027833491563797, + 0.04205251485109329, + 0.10875332355499268, + 0.04758426547050476, + 0.11672168970108032, + 0.07873205095529556, + 0.09015396982431412 + ], + "credit_map": [ + 0.024501293897628784, + 0.016800327226519585, + 0.09230202436447144, + 0.016861943528056145, + 0.017512887716293335, + 0.11036889255046844, + 0.03255052492022514, + 0.02905227430164814, + 0.006364832632243633, + 0.006168768275529146, + 0.005889492575079203, + 0.006782334763556719, + 0.004566616844385862, + 0.005553995259106159, + 0.0059256888926029205, + 0.005982620175927877, + 0.006840634159743786, + 0.0034714529756456614, + 0.0056578293442726135, + 0.005859761964529753, + 0.002510378137230873, + 0.0035951018799096346, + 0.0037569899577647448, + 0.006847229786217213, + 0.004117152653634548, + 0.003702149959281087, + 0.00946174468845129, + 0.007333068177103996, + 0.0036034672521054745, + 0.005976356565952301, + 0.029031656682491302, + 0.0270527433604002, + 0.00558390561491251, + 0.002904600230976939, + 0.0045813447795808315, + 0.004147551953792572, + 0.003842173609882593, + 0.006273957435041666, + 0.0075327493250370026, + 0.007186231203377247, + 0.01765383407473564, + 0.006764051970094442, + 0.01277119666337967, + 0.007130956742912531, + 0.015451902523636818, + 0.01731226034462452, + 0.010802950710058212, + 0.020866168662905693, + 0.005565073806792498, + 0.0024983722250908613, + 0.012671321630477905, + 0.003971072845160961, + 0.004314934369176626, + 0.011395390145480633, + 0.002665039151906967, + 0.0045030610635876656, + 0.008248626254498959, + 0.010242227464914322, + 0.030976304784417152, + 0.009844528511166573, + 0.015657195821404457, + 0.02614457532763481, + 0.006309298798441887, + 0.009298933669924736, + 0.005591618362814188, + 0.005670377518981695, + 0.014834997244179249, + 0.013008970767259598, + 0.007286121137440205, + 0.01601085625588894, + 0.006990840192884207, + 0.009338265284895897, + 0.004293503705412149, + 0.004086433444172144, + 0.00399168161675334, + 0.0059407418593764305, + 0.004642695188522339, + 0.008085259236395359, + 0.02920670434832573, + 0.029906952753663063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.009781762026250362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033591330051422 + ], + "opposite_logprob_tokens": [ + -1.910335898399353 + ], + "logprob_reward": -0.16033591330051422, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033591330051422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.132138192653656, + 0.08883937448263168, + 0.0682014599442482, + 0.12027526646852493, + 0.07159880548715591, + 0.12001676112413406, + 0.11647911369800568, + 0.0691438615322113, + 0.1314448118209839, + 0.08186236023902893 + ], + "credit_map": [ + 0.011272389441728592, + 0.01278216578066349, + 0.024349411949515343, + 0.014914759434759617, + 0.015327398665249348, + 0.023014865815639496, + 0.016281472519040108, + 0.014195711351931095, + 0.011102676391601562, + 0.014494290575385094, + 0.01168404333293438, + 0.0110169043764472, + 0.007706102915108204, + 0.010774505324661732, + 0.011140180751681328, + 0.010920665226876736, + 0.009080477058887482, + 0.00687071681022644, + 0.00994928553700447, + 0.01139812357723713, + 0.005430194549262524, + 0.00357363885268569, + 0.008086076937615871, + 0.013812940567731857, + 0.008018508553504944, + 0.008295051753520966, + 0.03099891170859337, + 0.01571081206202507, + 0.010932141914963722, + 0.008755655027925968, + 0.020128563046455383, + 0.01743561401963234, + 0.008819704875349998, + 0.005144436843693256, + 0.003748186631128192, + 0.004898290149867535, + 0.005853043869137764, + 0.009086234495043755, + 0.016561582684516907, + 0.01748731918632984, + 0.012411369942128658, + 0.013425351120531559, + 0.014691270887851715, + 0.008611389435827732, + 0.02750370278954506, + 0.01522691547870636, + 0.013301328755915165, + 0.014845415949821472, + 0.010336215607821941, + 0.005779782775789499, + 0.028422759845852852, + 0.012584510259330273, + 0.01169287133961916, + 0.030045146122574806, + 0.007553441915661097, + 0.010064376518130302, + 0.009576207026839256, + 0.009790599346160889, + 0.012627501040697098, + 0.007316462695598602, + 0.009928335435688496, + 0.008678382262587547, + 0.0056464956142008305, + 0.005579876247793436, + 0.009269701316952705, + 0.010506674647331238, + 0.02166338823735714, + 0.017834682017564774, + 0.009724292904138565, + 0.02654096484184265, + 0.01288799848407507, + 0.02301710844039917, + 0.004097123630344868, + 0.0033496601972728968, + 0.005221996456384659, + 0.007048952393233776, + 0.007255117408931255, + 0.011707344092428684, + 0.021250352263450623, + 0.021931804716587067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.16033591330051422 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up depicts a red pen tumbling and spinning in a desolate void. The entire scene is filmed at a high frame rate, transforming the chaotic action into a slow, mesmerizing ballet. The pen, a retractable ballpoint made of polished crimson plastic, rotates gracefully. Demonic cinematic lighting, derived from a single, hard, off-camera light source, carves the object out of infinite blackness, casting long, sharp shadows across its own form and creating stark, bright highlights on its gleaming metal clip and tip. Every minute rotation and reflection is rendered in perfect, crystal-clear detail, capturing the hypnotic glow of the polished plastic and gleaming metal, creating an atmosphere both elegant and deeply unsettling as the pen turns relentlessly in the silent abyss.", + "seed": 30428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_2521_seed_30428_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018839864991605282, + "vlm_logprob": -0.018839864991605282, + "vlm_yes_no_margin": 4.249999952316284, + "vlm_token_credit": -0.018839864991605282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022308005020022392 + ], + "opposite_logprob_tokens": [ + -4.022307872772217 + ], + "logprob_reward": -0.022308005020022392, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.022308005020022392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040901098400354385, + 0.12070271372795105, + 0.01513608917593956, + 0.26113608479499817, + 0.059368796646595, + 0.07482241094112396, + 0.09376704692840576, + 0.18948116898536682, + 0.03801845386624336, + 0.10666610300540924 + ], + "credit_map": [ + 0.00566013203933835, + 0.0036514573730528355, + 0.005109991878271103, + 0.005015681032091379, + 0.006391157861799002, + 0.004879687447100878, + 0.005220992490649223, + 0.0049720038659870625, + 0.005130131263285875, + 0.0028627500869333744, + 0.037022221833467484, + 0.005353907588869333, + 0.0032533917110413313, + 0.04981827735900879, + 0.00289861299097538, + 0.014363422989845276, + 0.0019662207923829556, + 0.0025378656573593616, + 0.0017931475304067135, + 0.0014662172179669142, + 0.0022092696744948626, + 0.002027211943641305, + 0.001054322929121554, + 0.0020818354096263647, + 0.01292436383664608, + 0.0105199683457613, + 0.09379636496305466, + 0.0115776676684618, + 0.018911762163043022, + 0.09434091299772263, + 0.008679099380970001, + 0.010385977104306221, + 0.004651053808629513, + 0.005533680319786072, + 0.01075455080717802, + 0.0046952213160693645, + 0.005941436626017094, + 0.015329891815781593, + 0.0060045067220926285, + 0.006458464544266462, + 0.0061785029247403145, + 0.004524097312241793, + 0.012885889038443565, + 0.019657546654343605, + 0.008920787833631039, + 0.007239669095724821, + 0.007892314344644547, + 0.007523606065660715, + 0.00821644812822342, + 0.004644967149943113, + 0.022515375167131424, + 0.005510011687874794, + 0.01064900029450655, + 0.006585730239748955, + 0.0068637956865131855, + 0.028781721368432045, + 0.01480351947247982, + 0.0128366369754076, + 0.037800077348947525, + 0.019640827551484108, + 0.01087319478392601, + 0.04691847041249275, + 0.020689908415079117, + 0.025918537750840187, + 0.004368938039988279, + 0.0069752465933561325, + 0.004737625364214182, + 0.005193075630813837, + 0.004381303675472736, + 0.00487511046230793, + 0.0035586922895163298, + 0.00392846530303359, + 0.013313453644514084, + 0.019446829333901405, + 0.010784028097987175, + 0.01187897752970457, + 0.010056381113827229, + 0.012066385708749294, + 0.01625325344502926, + 0.01286680530756712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.022308005020022392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636663056910038 + ], + "opposite_logprob_tokens": [ + -3.651366710662842 + ], + "logprob_reward": -0.02636663056910038, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636663056910038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07549132406711578, + 0.0731554850935936, + 0.03453929349780083, + 0.26889753341674805, + 0.04930465295910835, + 0.06405128538608551, + 0.12570230662822723, + 0.24471041560173035, + 0.018079860135912895, + 0.04606788977980614 + ], + "credit_map": [ + 0.009483381174504757, + 0.007484099827706814, + 0.010740538127720356, + 0.009686943143606186, + 0.00922788493335247, + 0.008485774509608746, + 0.01014839205890894, + 0.010234314016997814, + 0.004604016430675983, + 0.002005576388910413, + 0.02195350080728531, + 0.004170730244368315, + 0.0027931230142712593, + 0.028535036370158195, + 0.0023821101058274508, + 0.0067113894037902355, + 0.0032133350614458323, + 0.003803125349804759, + 0.004925991408526897, + 0.0038986701983958483, + 0.005407116375863552, + 0.004956671968102455, + 0.0035370392724871635, + 0.004797341767698526, + 0.013639925047755241, + 0.007677962072193623, + 0.09941194206476212, + 0.008883804082870483, + 0.014551130123436451, + 0.10625822097063065, + 0.005097567103803158, + 0.013376973569393158, + 0.004473384935408831, + 0.005609308369457722, + 0.005794837139546871, + 0.005236571189016104, + 0.005061783827841282, + 0.008471253328025341, + 0.00502990884706378, + 0.00962760392576456, + 0.0061764782294631, + 0.0027766104321926832, + 0.013625655323266983, + 0.022361615672707558, + 0.0039323545061051846, + 0.005131665151566267, + 0.006316101178526878, + 0.003730807453393936, + 0.010380950756371021, + 0.003809350775554776, + 0.03537308797240257, + 0.005441470071673393, + 0.0063405330292880535, + 0.008904483169317245, + 0.008817862719297409, + 0.04663456976413727, + 0.0202640350908041, + 0.01807256042957306, + 0.043231915682554245, + 0.024892278015613556, + 0.015127203427255154, + 0.05600784346461296, + 0.029101159423589706, + 0.03801341354846954, + 0.0018987513612955809, + 0.0026373867876827717, + 0.0023425419349223375, + 0.003240747144445777, + 0.0016473607392981648, + 0.0019008038798347116, + 0.002301300410181284, + 0.002110967179760337, + 0.0045380075462162495, + 0.005304235033690929, + 0.004012560471892357, + 0.007689616177231073, + 0.004440148826688528, + 0.006161505356431007, + 0.00776589335873723, + 0.006155922543257475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.02636663056910038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031943507492542267 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.0031943507492542267, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031943507492542267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01955288276076317, + 0.025789130479097366, + 0.016296833753585815, + 0.7787379026412964, + 0.032401371747255325, + 0.03340521454811096, + 0.03505430743098259, + 0.03853607550263405, + 0.004083462059497833, + 0.016142765060067177 + ], + "credit_map": [ + 0.002708773361518979, + 0.001659996691159904, + 0.002576560014858842, + 0.0026725565548986197, + 0.00212375377304852, + 0.0021288227289915085, + 0.0023069214075803757, + 0.003375498577952385, + 0.00216114753857255, + 0.0009064102196134627, + 0.007738660555332899, + 0.0018114869017153978, + 0.0011399011127650738, + 0.007749420125037432, + 0.0012378384126350284, + 0.0030442674178630114, + 0.0018481508595868945, + 0.001616013003513217, + 0.0022209594026207924, + 0.002057755133137107, + 0.0015680987853556871, + 0.002322584157809615, + 0.0026224106550216675, + 0.0020408628042787313, + 0.040359172970056534, + 0.019255639985203743, + 0.28566691279411316, + 0.028439359739422798, + 0.03702211752533913, + 0.30827781558036804, + 0.017606033012270927, + 0.04211094230413437, + 0.004521716386079788, + 0.0020821266807615757, + 0.0053183529525995255, + 0.0019121833611279726, + 0.0020591055508702993, + 0.009182197041809559, + 0.0017720158211886883, + 0.005553675349801779, + 0.00584407290443778, + 0.0018375528743490577, + 0.005791720002889633, + 0.008750668726861477, + 0.0024303277023136616, + 0.00330145051702857, + 0.003223536303266883, + 0.0022258886601775885, + 0.0030477293767035007, + 0.0010165070416405797, + 0.01022690162062645, + 0.0012448659399524331, + 0.0017657378921285272, + 0.0018828382017090917, + 0.0027878200635313988, + 0.013081907294690609, + 0.0028954369481652975, + 0.002557252300903201, + 0.008557911962270737, + 0.0036114375106990337, + 0.002226875862106681, + 0.009806966409087181, + 0.00416457699611783, + 0.004715623799711466, + 0.00040598749183118343, + 0.0007405410869978368, + 0.0005237793084233999, + 0.0006370333721861243, + 0.00035236755502410233, + 0.0005106945754960179, + 0.00047223237925209105, + 0.0004408268432598561, + 0.0015965113416314125, + 0.0016376739367842674, + 0.0015137001173570752, + 0.0022170229349285364, + 0.0014314481522887945, + 0.0024887146428227425, + 0.0022822138853371143, + 0.0029754801653325558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.0031943507492542267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06540684401988983, + 0.08933606743812561, + 0.02420039474964142, + 0.515583872795105, + 0.022647948935627937, + 0.05641934648156166, + 0.10725634545087814, + 0.07928275316953659, + 0.00978899747133255, + 0.030077459290623665 + ], + "credit_map": [ + 0.008496812544763088, + 0.005215857177972794, + 0.009869779460132122, + 0.006447394844144583, + 0.006667609326541424, + 0.006538955960422754, + 0.006393030285835266, + 0.015777405351400375, + 0.00629109563305974, + 0.0018328570295125246, + 0.027234923094511032, + 0.004223659168928862, + 0.0023269914090633392, + 0.03662426397204399, + 0.0024573442060500383, + 0.008344929665327072, + 0.0023907297290861607, + 0.002391320187598467, + 0.0035911877639591694, + 0.0038260461296886206, + 0.0034035241696983576, + 0.003003702498972416, + 0.0029942188411951065, + 0.0025996637996286154, + 0.02500661462545395, + 0.011860816739499569, + 0.19589965045452118, + 0.017186814919114113, + 0.024458736181259155, + 0.20870313048362732, + 0.009386332705616951, + 0.023081740364432335, + 0.0021569482050836086, + 0.0025286211166530848, + 0.004301261156797409, + 0.0027000659611076117, + 0.0023764302022755146, + 0.0042907604947686195, + 0.002452901331707835, + 0.0018409587210044265, + 0.005777882877737284, + 0.0026084992568939924, + 0.012424553744494915, + 0.019096316769719124, + 0.0036874767392873764, + 0.003684212453663349, + 0.00501261604949832, + 0.004127786494791508, + 0.009481729008257389, + 0.0031979463528841734, + 0.03160637989640236, + 0.0038696115370839834, + 0.005238038022071123, + 0.0057593039236962795, + 0.007907654158771038, + 0.040195684880018234, + 0.006307386327534914, + 0.005665468517690897, + 0.015279702842235565, + 0.008335391990840435, + 0.004877112805843353, + 0.018585437908768654, + 0.009359844028949738, + 0.010872408747673035, + 0.0009476835839450359, + 0.001103128888644278, + 0.0015001670690253377, + 0.002026427537202835, + 0.0010036546736955643, + 0.0012330715544521809, + 0.0009348431485705078, + 0.0010400197934359312, + 0.003936656750738621, + 0.004074364434927702, + 0.0030368391890078783, + 0.0034385721664875746, + 0.0024776419159024954, + 0.0034995588939636946, + 0.00426087062805891, + 0.005352952517569065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979392558336258 + ], + "opposite_logprob_tokens": [ + -3.5297939777374268 + ], + "logprob_reward": -0.02979392558336258, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979392558336258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11121456325054169, + 0.0465102344751358, + 0.04215497151017189, + 0.13516005873680115, + 0.08133623003959656, + 0.07535796612501144, + 0.12579672038555145, + 0.2772827744483948, + 0.02869265154004097, + 0.0764937773346901 + ], + "credit_map": [ + 0.01509144064038992, + 0.013673024252057076, + 0.015052473172545433, + 0.01344330981373787, + 0.011092291213572025, + 0.014661871828138828, + 0.01237085834145546, + 0.015829304233193398, + 0.006072513293474913, + 0.003686453914269805, + 0.008367154747247696, + 0.00402920413762331, + 0.00437174504622817, + 0.010776504874229431, + 0.004130726680159569, + 0.0050759343430399895, + 0.003924170043319464, + 0.004433841444551945, + 0.005832634400576353, + 0.005205739289522171, + 0.004633365664631128, + 0.004746424034237862, + 0.009005466476082802, + 0.0043733310885727406, + 0.009746500290930271, + 0.014655138365924358, + 0.018506724387407303, + 0.016773061826825142, + 0.018220845609903336, + 0.027073513716459274, + 0.014183798804879189, + 0.016000481322407722, + 0.009281131438910961, + 0.0067366380244493484, + 0.011882133781909943, + 0.00550196273252368, + 0.005681982729583979, + 0.022612314671278, + 0.0074827647767961025, + 0.012157305143773556, + 0.006544856354594231, + 0.0032646481413394213, + 0.01683548279106617, + 0.02628561481833458, + 0.005239239428192377, + 0.005324972327798605, + 0.006777588743716478, + 0.005085567478090525, + 0.010457315482199192, + 0.004216581583023071, + 0.03534168377518654, + 0.006048253271728754, + 0.0067611862905323505, + 0.007577214855700731, + 0.009851118549704552, + 0.0455433651804924, + 0.02267587184906006, + 0.02022014744579792, + 0.04814104735851288, + 0.02870800718665123, + 0.01679038815200329, + 0.06355558335781097, + 0.03318873792886734, + 0.04400299862027168, + 0.002607097616419196, + 0.004632372874766588, + 0.004125415813177824, + 0.005161604378372431, + 0.0029452743474394083, + 0.0033099036663770676, + 0.0033666244708001614, + 0.0025443590711802244, + 0.010093407705426216, + 0.010968890972435474, + 0.008417626842856407, + 0.009102889336645603, + 0.008253122679889202, + 0.008965912275016308, + 0.009398527443408966, + 0.011293401941657066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.02979392558336258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up depicts a red pen tumbling and spinning in a desolate void. The entire scene is filmed at a high frame rate, transforming the chaotic action into a slow, mesmerizing ballet. The pen, a retractable ballpoint made of polished crimson plastic, rotates gracefully. Demonic cinematic lighting, derived from a single, hard, off-camera light source, carves the object out of infinite blackness, casting long, sharp shadows across its own form and creating stark, bright highlights on its gleaming metal clip and tip. Every minute rotation and reflection is rendered in perfect, crystal-clear detail, capturing the hypnotic glow of the polished plastic and gleaming metal, creating an atmosphere both elegant and deeply unsettling as the pen turns relentlessly in the silent abyss.", + "seed": 30429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_6/index_2521_seed_30429_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02434794099535793, + "vlm_logprob": -0.02434794099535793, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02434794099535793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0427418053150177 + ], + "opposite_logprob_tokens": [ + -3.2927417755126953 + ], + "logprob_reward": -0.0427418053150177, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0427418053150177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17414066195487976, + 0.1116790696978569, + 0.050483521074056625, + 0.0491449199616909, + 0.10982024669647217, + 0.07625572383403778, + 0.2550845742225647, + 0.05175280570983887, + 0.09179003536701202, + 0.0298483744263649 + ], + "credit_map": [ + 0.0161737110465765, + 0.018159916624426842, + 0.02458108589053154, + 0.03604985401034355, + 0.01707364246249199, + 0.01746923103928566, + 0.025994641706347466, + 0.01863858290016651, + 0.009111260995268822, + 0.016644075512886047, + 0.013815118931233883, + 0.012704210355877876, + 0.02074017934501171, + 0.013411822728812695, + 0.014021485112607479, + 0.011230921372771263, + 0.004704538267105818, + 0.004138318356126547, + 0.006917986087501049, + 0.006024031899869442, + 0.007269920315593481, + 0.006431981455534697, + 0.007997111417353153, + 0.006999639328569174, + 0.0042146919295191765, + 0.007125853560864925, + 0.0057479869574308395, + 0.004366861190646887, + 0.00980787817388773, + 0.005340403877198696, + 0.0064651742577552795, + 0.006076072342693806, + 0.012068422511219978, + 0.014696024358272552, + 0.013133426196873188, + 0.011668503284454346, + 0.016254067420959473, + 0.018963579088449478, + 0.010489759035408497, + 0.012546469457447529, + 0.006818991154432297, + 0.011988310143351555, + 0.009461276233196259, + 0.008447673171758652, + 0.014105530455708504, + 0.008903712965548038, + 0.007712796330451965, + 0.008817439898848534, + 0.029776979237794876, + 0.031125493347644806, + 0.03286394476890564, + 0.02390504814684391, + 0.033327069133520126, + 0.02492079883813858, + 0.02589769847691059, + 0.05326758325099945, + 0.004802296869456768, + 0.007623528596013784, + 0.0061181411147117615, + 0.004348890855908394, + 0.010511401109397411, + 0.006287151947617531, + 0.005742275156080723, + 0.006319126579910517, + 0.010261806659400463, + 0.00971655547618866, + 0.01772623136639595, + 0.009962671436369419, + 0.008434058167040348, + 0.013404018245637417, + 0.009125245735049248, + 0.013159447349607944, + 0.0028734824154525995, + 0.005376716610044241, + 0.003615805646404624, + 0.0043289270251989365, + 0.002670516725629568, + 0.003468381939455867, + 0.003943608142435551, + 0.0035709396470338106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0427418053150177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043089546263217926 + ], + "opposite_logprob_tokens": [ + -3.1680896282196045 + ], + "logprob_reward": -0.043089546263217926, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043089546263217926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20825862884521484, + 0.10237961262464523, + 0.02919136732816696, + 0.04013349488377571, + 0.0864432156085968, + 0.07655748724937439, + 0.32812991738319397, + 0.03940608352422714, + 0.0477583147585392, + 0.04174188897013664 + ], + "credit_map": [ + 0.024734150618314743, + 0.03067060373723507, + 0.02923862263560295, + 0.02421364188194275, + 0.02317701280117035, + 0.024540601298213005, + 0.015900949016213417, + 0.03578304126858711, + 0.008791112340986729, + 0.01313617080450058, + 0.012852395884692669, + 0.011520590633153915, + 0.020508622750639915, + 0.012479973025619984, + 0.01128674577921629, + 0.011804007925093174, + 0.0027736080810427666, + 0.0038878265768289566, + 0.003960950765758753, + 0.0029481910169124603, + 0.005086403805762529, + 0.0028841006569564342, + 0.003707049647346139, + 0.003943237010389566, + 0.00410423381254077, + 0.005045298486948013, + 0.005608055274933577, + 0.004900498781353235, + 0.00494166137650609, + 0.005457873456180096, + 0.005048292223364115, + 0.005027581471949816, + 0.008449039421975613, + 0.010112229734659195, + 0.012574562802910805, + 0.010173079557716846, + 0.012000900693237782, + 0.012894369661808014, + 0.009679388254880905, + 0.010559646412730217, + 0.008214665576815605, + 0.010485769249498844, + 0.008800613693892956, + 0.008876335807144642, + 0.011957083828747272, + 0.008106851950287819, + 0.009832211770117283, + 0.010283960029482841, + 0.039131056517362595, + 0.040791530162096024, + 0.04345366358757019, + 0.03426539897918701, + 0.03695707768201828, + 0.031090328469872475, + 0.03335984796285629, + 0.06908101588487625, + 0.003693204838782549, + 0.007878400385379791, + 0.004907813388854265, + 0.0038545650895684958, + 0.005857884418219328, + 0.005420967936515808, + 0.003020946169272065, + 0.004772298038005829, + 0.005094077903777361, + 0.007032153196632862, + 0.007323704194277525, + 0.003732077544555068, + 0.008142780512571335, + 0.006109778769314289, + 0.0052612763829529285, + 0.005062466952949762, + 0.003668852150440216, + 0.011000702157616615, + 0.005195692181587219, + 0.004193041939288378, + 0.004325210582464933, + 0.004954427946358919, + 0.0038847490213811398, + 0.004519212990999222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.043089546263217926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019422968616709113 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019422968616709113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019422968616709113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5285508036613464, + 0.09547661244869232, + 0.050754841417074203, + 0.07795946300029755, + 0.034985221922397614, + 0.046116940677165985, + 0.04675852134823799, + 0.06776870042085648, + 0.030012203380465508, + 0.021616721525788307 + ], + "credit_map": [ + 0.06494709104299545, + 0.06572657078504562, + 0.06953907012939453, + 0.06162898615002632, + 0.05974274501204491, + 0.05903252214193344, + 0.04668981581926346, + 0.1012439876794815, + 0.008986392989754677, + 0.014652678743004799, + 0.013689515180885792, + 0.017068423330783844, + 0.008632696233689785, + 0.009850187227129936, + 0.010421470738947392, + 0.0121752405539155, + 0.0055814762599766254, + 0.008354251272976398, + 0.007210457231849432, + 0.005735835060477257, + 0.00614936463534832, + 0.008489208295941353, + 0.004091228358447552, + 0.005143015645444393, + 0.007247352506965399, + 0.014504690654575825, + 0.00946235004812479, + 0.008657904341816902, + 0.010973251424729824, + 0.00814730767160654, + 0.009795434772968292, + 0.00917117204517126, + 0.004265911877155304, + 0.0046669184230268, + 0.004813415929675102, + 0.0047032637521624565, + 0.0047380817122757435, + 0.003993215970695019, + 0.003784115891903639, + 0.004020300228148699, + 0.005479234270751476, + 0.008567840792238712, + 0.005600763484835625, + 0.005988304503262043, + 0.005117243155837059, + 0.006055696867406368, + 0.004592163022607565, + 0.004715694114565849, + 0.005496748723089695, + 0.00632109772413969, + 0.005595591384917498, + 0.004321835469454527, + 0.006047040689736605, + 0.004925809800624847, + 0.004779416602104902, + 0.009270979091525078, + 0.006201192736625671, + 0.007836381904780865, + 0.009563926607370377, + 0.016057712957262993, + 0.004825782962143421, + 0.0074691493064165115, + 0.007047841791063547, + 0.008766715414822102, + 0.0031101959757506847, + 0.004163874313235283, + 0.004406493157148361, + 0.002878868253901601, + 0.004873267840594053, + 0.0038573688361793756, + 0.0029043362010270357, + 0.003817799501121044, + 0.002187044359743595, + 0.005429925862699747, + 0.0029310015961527824, + 0.0021037044934928417, + 0.0023039577063173056, + 0.002877678256481886, + 0.0016460685292258859, + 0.0021373415365815163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0019422968616709113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41956159472465515, + 0.11220906674861908, + 0.06271784752607346, + 0.03657280653715134, + 0.06726814061403275, + 0.05088180676102638, + 0.1390981823205948, + 0.04683795943856239, + 0.039217643439769745, + 0.025634968653321266 + ], + "credit_map": [ + 0.043887216597795486, + 0.055780570954084396, + 0.05080146715044975, + 0.05361055210232735, + 0.0627661794424057, + 0.04630071669816971, + 0.04233474284410477, + 0.06408015638589859, + 0.00974584836512804, + 0.01930588111281395, + 0.013373294845223427, + 0.012334524653851986, + 0.013922598212957382, + 0.015900904312729836, + 0.012391356751322746, + 0.01523465383797884, + 0.005632751155644655, + 0.008530942723155022, + 0.008360599167644978, + 0.006421653088182211, + 0.006800615228712559, + 0.008303193375468254, + 0.009134245105087757, + 0.009533843025565147, + 0.004483598750084639, + 0.0064178528264164925, + 0.004148570820689201, + 0.004798850975930691, + 0.005243203137069941, + 0.004277493339031935, + 0.004021665081381798, + 0.0031815734691917896, + 0.00881156139075756, + 0.007445424795150757, + 0.010052815079689026, + 0.006963561754673719, + 0.01155826635658741, + 0.00840378925204277, + 0.007456096354871988, + 0.006576621904969215, + 0.005222545936703682, + 0.00838973093777895, + 0.0057321456260979176, + 0.005598400253802538, + 0.006486948113888502, + 0.006539234425872564, + 0.00512311328202486, + 0.007789685390889645, + 0.015958422794938087, + 0.017940310761332512, + 0.01717888005077839, + 0.013519933447241783, + 0.01971987448632717, + 0.013984715566039085, + 0.014243022538721561, + 0.026553014293313026, + 0.005082904361188412, + 0.007051150780171156, + 0.006318597588688135, + 0.005200635641813278, + 0.007076844107359648, + 0.005284920800477266, + 0.00470117200165987, + 0.006121727637946606, + 0.003983025439083576, + 0.006744272541254759, + 0.005832991562783718, + 0.004245324991643429, + 0.0042853159829974174, + 0.0055986978113651276, + 0.002937987446784973, + 0.005590027663856745, + 0.002207566052675247, + 0.006043237634003162, + 0.003191225929185748, + 0.002641047816723585, + 0.0032028628047555685, + 0.0031914510764181614, + 0.0023615455720573664, + 0.0027960315346717834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.00762729300186038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a red retractable ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen made of polished crimson plastic with a gleaming metal clip and tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tumbling and spinning in a desolate void with infinite blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single hard off-camera light source creating long sharp shadows and bright highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338763535022736 + ], + "opposite_logprob_tokens": [ + -3.651338815689087 + ], + "logprob_reward": -0.026338763535022736, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338763535022736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2120562344789505, + 0.06880129873752594, + 0.063942089676857, + 0.08161453157663345, + 0.11778002977371216, + 0.04314449429512024, + 0.14177563786506653, + 0.03658119961619377, + 0.11670184135437012, + 0.11760260909795761 + ], + "credit_map": [ + 0.022986216470599174, + 0.023110199719667435, + 0.028549708425998688, + 0.03132417052984238, + 0.024650665000081062, + 0.025990696623921394, + 0.024806411936879158, + 0.030638163909316063, + 0.008055335842072964, + 0.013271507807075977, + 0.009980976581573486, + 0.008707751519978046, + 0.006205128505825996, + 0.008373241871595383, + 0.006743405479937792, + 0.007463946472853422, + 0.00948429387062788, + 0.010201792232692242, + 0.00931171327829361, + 0.005146270152181387, + 0.006427358835935593, + 0.007395180407911539, + 0.005560091231018305, + 0.010415387339890003, + 0.007330733817070723, + 0.01532185822725296, + 0.010843729600310326, + 0.01314220018684864, + 0.009062361903488636, + 0.008731376379728317, + 0.008744361810386181, + 0.008437903597950935, + 0.010483082383871078, + 0.01617470011115074, + 0.012959504500031471, + 0.01789315976202488, + 0.015036786906421185, + 0.01833861880004406, + 0.015591171570122242, + 0.011303005740046501, + 0.003919609356671572, + 0.006178123876452446, + 0.005286223255097866, + 0.004198945127427578, + 0.006194785237312317, + 0.004927599336951971, + 0.003280916018411517, + 0.009158293716609478, + 0.016500171273946762, + 0.019215326756238937, + 0.01855100318789482, + 0.014316466636955738, + 0.016328543424606323, + 0.014422966167330742, + 0.014339801855385303, + 0.028101356700062752, + 0.005974203813821077, + 0.004757600370794535, + 0.005380099173635244, + 0.003922830801457167, + 0.003997402731329203, + 0.004574955441057682, + 0.003508498426526785, + 0.004465607926249504, + 0.011230715550482273, + 0.014583824202418327, + 0.01613536663353443, + 0.012229276821017265, + 0.014580256305634975, + 0.016613325104117393, + 0.013930478133261204, + 0.01739860139787197, + 0.01646413467824459, + 0.020468056201934814, + 0.015066045336425304, + 0.013401404023170471, + 0.011191216297447681, + 0.013980891555547714, + 0.01162213459610939, + 0.015408736653625965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.026338763535022736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_tensor([35, 35])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_tensor([35, 35])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90230bc244756f27e6eafb7db2f865ead245ea11 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_tensor([35, 35])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly dynamic first-person perspective shot places the viewer directly in the cockpit of a high-tech, specialized polar transport vehicle. The camera, with its wide-angle perspective, moves forward steadily and smoothly, as if supported by a top-tier gyroscopic stabilizer, imbued with the high-definition quality of an Olympic-level broadcast. At the bottom of the frame, the vehicle's massive tracks, constructed from black composite materials with metal hinges, dominate the field of view, their deep V-shaped patterns leaving a distinct imprint on the ground. As the vehicle moves forward with steady yet powerful force, the tracks cut deeply into the boundless, pure white snowfield, powerfully stirring and sweeping back vast clouds of shimmering snow and fine ice crystals. In the clear, cold polar sunlight, these flying snow particles refract like diamonds. As the vehicle reaches the edge of a magnificent glacier, its deep, crystalline blue ice looming below the cliffs, the vehicle executes an impeccable U-turn: the camera pans along with the vehicle in a grand, silky arc, one track slowing while the other powerfully propels it forward, carving a perfect semicircle across the vast snowy landscape with a display of absolute control. The entire sequence is imbued with unparalleled dynamic cinematography, seamlessly blending the power and precision of the machinery with the grandeur and harshness of the polar environment, creating a thrilling, epic atmosphere of exploration and conquest.", + "seed": 19315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19315_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06854496840387583, + "vlm_logprob": -0.06854496840387583, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.06854496840387583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17096063494682312 + ], + "opposite_logprob_tokens": [ + -1.9209606647491455 + ], + "logprob_reward": -0.17096063494682312, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17096063494682312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05583254620432854, + 0.0351942740380764, + 0.030711248517036438, + 0.30031728744506836, + 0.11757297068834305, + 0.045911505818367004, + 0.1187729761004448, + 0.06111427769064903, + 0.06777657568454742, + 0.16679637134075165 + ], + "credit_map": [ + 0.005298498086631298, + 0.009296977892518044, + 0.007817516103386879, + 0.004629882052540779, + 0.00811716541647911, + 0.0065223453566432, + 0.0061833737418055534, + 0.007966787554323673, + 0.004222837742418051, + 0.004986257292330265, + 0.004367217421531677, + 0.004864254500716925, + 0.004515612032264471, + 0.0028883377090096474, + 0.005537149962037802, + 0.0038126064464449883, + 0.004207209218293428, + 0.00447800662368536, + 0.00455066142603755, + 0.0027871299535036087, + 0.0045738923363387585, + 0.00318734347820282, + 0.0034658059012144804, + 0.0034611995797604322, + 0.02265649102628231, + 0.03778219223022461, + 0.03859790042042732, + 0.048295166343450546, + 0.04737662151455879, + 0.030796751379966736, + 0.044547684490680695, + 0.0302644781768322, + 0.010041984729468822, + 0.01579117588698864, + 0.013815761543810368, + 0.02190972864627838, + 0.019986195489764214, + 0.01087260153144598, + 0.014851553365588188, + 0.010303974151611328, + 0.008114251308143139, + 0.005351662635803223, + 0.006100493483245373, + 0.0055515822023153305, + 0.0053106024861335754, + 0.004011907614767551, + 0.006561527960002422, + 0.00490947999060154, + 0.011987762525677681, + 0.029186761006712914, + 0.018392391502857208, + 0.011318557895720005, + 0.01683489792048931, + 0.01501594576984644, + 0.009571892209351063, + 0.006464776117354631, + 0.004509856924414635, + 0.010027717798948288, + 0.006368977949023247, + 0.006443188525736332, + 0.008778954856097698, + 0.00932773482054472, + 0.010819549672305584, + 0.004838296212255955, + 0.007778146304190159, + 0.006804890464991331, + 0.007258919067680836, + 0.006960065104067326, + 0.009418189525604248, + 0.01618320494890213, + 0.007837759330868721, + 0.00553540326654911, + 0.02197236754000187, + 0.0213895495980978, + 0.02266310714185238, + 0.0145452581346035, + 0.02748069353401661, + 0.024767665192484856, + 0.01908743381500244, + 0.01489028800278902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.17096063494682312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373659774661064 + ], + "opposite_logprob_tokens": [ + -3.4087367057800293 + ], + "logprob_reward": -0.03373659774661064, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373659774661064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03859338536858559, + 0.027202878147363663, + 0.020735571160912514, + 0.07792443037033081, + 0.04450956732034683, + 0.039615243673324585, + 0.482618123292923, + 0.07686227560043335, + 0.10489487648010254, + 0.08704368025064468 + ], + "credit_map": [ + 0.003811520291492343, + 0.005047863814979792, + 0.004564986098557711, + 0.004134348593652248, + 0.006169257219880819, + 0.004952147603034973, + 0.004817683715373278, + 0.005095574073493481, + 0.0034896160941570997, + 0.003225551452487707, + 0.004406708292663097, + 0.002045407658442855, + 0.0034704632125794888, + 0.0027241238858550787, + 0.0036835039500147104, + 0.004157501272857189, + 0.003142421366646886, + 0.0028779299464076757, + 0.0024573507253080606, + 0.0017740230541676283, + 0.0032441129442304373, + 0.0019906433299183846, + 0.003037480404600501, + 0.0022116072941571474, + 0.006855583284050226, + 0.008841310627758503, + 0.010889297351241112, + 0.013795623555779457, + 0.011085395701229572, + 0.009893686510622501, + 0.009079614654183388, + 0.0074839117005467415, + 0.0051330397836863995, + 0.006819073110818863, + 0.004931786563247442, + 0.0058738612569868565, + 0.0063640461303293705, + 0.004888023715466261, + 0.006571480073034763, + 0.00392825435847044, + 0.005182746797800064, + 0.004326207097619772, + 0.004871878307312727, + 0.003667308948934078, + 0.005795550066977739, + 0.005428895819932222, + 0.004689845722168684, + 0.005652807652950287, + 0.052192721515893936, + 0.10016164183616638, + 0.06742867827415466, + 0.058587536215782166, + 0.08076547831296921, + 0.03679177910089493, + 0.0576537661254406, + 0.02903646230697632, + 0.008318126201629639, + 0.015650590881705284, + 0.009687953628599644, + 0.005933952983468771, + 0.01116501446813345, + 0.011261170729994774, + 0.009233584627509117, + 0.005611875094473362, + 0.010287337936460972, + 0.009702024050056934, + 0.008395040407776833, + 0.009026930667459965, + 0.013017256744205952, + 0.02547704055905342, + 0.012089396826922894, + 0.01689983904361725, + 0.01321653462946415, + 0.00849214568734169, + 0.00796388741582632, + 0.005554328206926584, + 0.010127008892595768, + 0.01652040332555771, + 0.007056721486151218, + 0.018112635239958763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.03373659774661064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304787144064903 + ], + "opposite_logprob_tokens": [ + -3.1680479049682617 + ], + "logprob_reward": -0.04304787144064903, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304787144064903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08847448974847794, + 0.028443941846489906, + 0.020125124603509903, + 0.12406767904758453, + 0.1280203014612198, + 0.03381912782788277, + 0.12138630449771881, + 0.23886260390281677, + 0.10592332482337952, + 0.11087711900472641 + ], + "credit_map": [ + 0.009192471392452717, + 0.015641001984477043, + 0.012339001521468163, + 0.010744121856987476, + 0.009048443287611008, + 0.010407346300780773, + 0.010838018730282784, + 0.010264080949127674, + 0.0035204668529331684, + 0.004076298791915178, + 0.0045431009493768215, + 0.002566443057730794, + 0.002914315089583397, + 0.004337884020060301, + 0.0031173413153737783, + 0.003368091071024537, + 0.002856386126950383, + 0.0028403736650943756, + 0.003712552133947611, + 0.0019505603704601526, + 0.002431003376841545, + 0.0016896565211936831, + 0.0025211635511368513, + 0.0021234285086393356, + 0.012537823058664799, + 0.024261044338345528, + 0.018079333007335663, + 0.025954939424991608, + 0.013694877736270428, + 0.00817323848605156, + 0.014360306784510612, + 0.0070061227306723595, + 0.009699845686554909, + 0.02082289755344391, + 0.01530864555388689, + 0.023066971451044083, + 0.022965028882026672, + 0.008344592526555061, + 0.018665188923478127, + 0.00914712529629469, + 0.004023133777081966, + 0.004471856169402599, + 0.004941579420119524, + 0.005509463604539633, + 0.003715098137035966, + 0.0038189836777746677, + 0.0038880365900695324, + 0.0034509794786572456, + 0.011582891456782818, + 0.0181291326880455, + 0.014493091031908989, + 0.015822334215044975, + 0.019935864955186844, + 0.011571895331144333, + 0.018764181062579155, + 0.01108691468834877, + 0.025712788105010986, + 0.051486577838659286, + 0.027316952124238014, + 0.017708798870444298, + 0.0426778718829155, + 0.03269270062446594, + 0.021770736202597618, + 0.019496174529194832, + 0.013028284534811974, + 0.0161830335855484, + 0.013642222620546818, + 0.011898684315383434, + 0.016350887715816498, + 0.01984224282205105, + 0.007928860373795033, + 0.007049111183732748, + 0.014120177365839481, + 0.009681438095867634, + 0.013003452681005001, + 0.0077309864573180676, + 0.0126678217202425, + 0.022164642810821533, + 0.008608019910752773, + 0.022900579497218132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.04304787144064903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902997523546219 + ], + "opposite_logprob_tokens": [ + -2.4640300273895264 + ], + "logprob_reward": -0.08902997523546219, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902997523546219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05128193646669388, + 0.0278804711997509, + 0.05059059336781502, + 0.14256326854228973, + 0.1773720532655716, + 0.07157062739133835, + 0.2623286843299866, + 0.05156584829092026, + 0.07473713159561157, + 0.09010937064886093 + ], + "credit_map": [ + 0.005785567685961723, + 0.010158968158066273, + 0.007675662636756897, + 0.00626793410629034, + 0.005401188973337412, + 0.00526089034974575, + 0.004578821826726198, + 0.006152905989438295, + 0.004263591021299362, + 0.003546100342646241, + 0.004362314939498901, + 0.0036209409590810537, + 0.003013606881722808, + 0.002744249999523163, + 0.002917608944699168, + 0.003412061370909214, + 0.006802596617490053, + 0.007877920754253864, + 0.008126364089548588, + 0.006219881121069193, + 0.006779591087251902, + 0.00466062780469656, + 0.00452760374173522, + 0.005596007686108351, + 0.014354031533002853, + 0.01930602267384529, + 0.018921762704849243, + 0.027892157435417175, + 0.020202375948429108, + 0.011816008947789669, + 0.019670210778713226, + 0.010400704108178616, + 0.01666068285703659, + 0.028732504695653915, + 0.02025515027344227, + 0.033959582448005676, + 0.026639819145202637, + 0.014508472755551338, + 0.02193876914680004, + 0.014677099883556366, + 0.012922165915369987, + 0.007532197516411543, + 0.010672205127775669, + 0.011231506243348122, + 0.007762584835290909, + 0.007087321951985359, + 0.006929674651473761, + 0.007432978134602308, + 0.02883116528391838, + 0.06307440251111984, + 0.040853530168533325, + 0.027020787820219994, + 0.045332688838243484, + 0.019459350034594536, + 0.024829763919115067, + 0.012926996685564518, + 0.004589240998029709, + 0.007208599243313074, + 0.005575534421950579, + 0.005000723991543055, + 0.006819318048655987, + 0.007936605252325535, + 0.007985654287040234, + 0.006450175773352385, + 0.007520420476794243, + 0.009655614383518696, + 0.008006217889487743, + 0.005917430855333805, + 0.01363722700625658, + 0.018495144322514534, + 0.006586883682757616, + 0.004918200895190239, + 0.011264999397099018, + 0.015404070727527142, + 0.010028379037976265, + 0.007025174796581268, + 0.018553636968135834, + 0.011159166693687439, + 0.006325132679194212, + 0.010348821058869362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.08902997523546219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054975055158138275, + 0.02131202444434166, + 0.03776811063289642, + 0.04676399379968643, + 0.12787742912769318, + 0.06593744456768036, + 0.36275607347488403, + 0.12815049290657043, + 0.02973121404647827, + 0.12472806125879288 + ], + "credit_map": [ + 0.006504441611468792, + 0.007593099493533373, + 0.006966607645153999, + 0.005585324019193649, + 0.008321433328092098, + 0.006430940702557564, + 0.007911783643066883, + 0.005661425646394491, + 0.0029239810537546873, + 0.0028958378825336695, + 0.002606647554785013, + 0.0022431134711951017, + 0.0022573075257241726, + 0.0026649837382137775, + 0.002748799277469516, + 0.002971354406327009, + 0.005214995238929987, + 0.006778546143323183, + 0.005075675435364246, + 0.004529449623078108, + 0.005818719044327736, + 0.0033127188216894865, + 0.0035801585763692856, + 0.003457848448306322, + 0.004184305667877197, + 0.0067604356445372105, + 0.007541274651885033, + 0.009730607271194458, + 0.005519741214811802, + 0.004243855830281973, + 0.005426397081464529, + 0.0033573766704648733, + 0.012680742889642715, + 0.017512507736682892, + 0.014522883109748363, + 0.023794105276465416, + 0.019267680123448372, + 0.01317177526652813, + 0.015844257548451424, + 0.011083485558629036, + 0.007394600659608841, + 0.009063400328159332, + 0.007158027961850166, + 0.007202061824500561, + 0.009187566116452217, + 0.009408693760633469, + 0.01055684220045805, + 0.005966248922049999, + 0.03920164331793785, + 0.08810213953256607, + 0.0567595511674881, + 0.037739988416433334, + 0.060769062489271164, + 0.027866123244166374, + 0.034470345824956894, + 0.017847198992967606, + 0.014983903616666794, + 0.033074501901865005, + 0.01508971769362688, + 0.007732579950243235, + 0.020091785117983818, + 0.017095506191253662, + 0.011727288365364075, + 0.008355215191841125, + 0.00304606300778687, + 0.00432183314114809, + 0.003541343379765749, + 0.002946326741948724, + 0.004696647636592388, + 0.0048820083029568195, + 0.0025975622702389956, + 0.0036994279362261295, + 0.017009014263749123, + 0.011476165615022182, + 0.012384283356368542, + 0.00862833671271801, + 0.013071185909211636, + 0.024788748472929, + 0.008154689334332943, + 0.02921564131975174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.005949762649834156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly dynamic first-person perspective shot places the viewer directly in the cockpit of a high-tech, specialized polar transport vehicle. The camera, with its wide-angle perspective, moves forward steadily and smoothly, as if supported by a top-tier gyroscopic stabilizer, imbued with the high-definition quality of an Olympic-level broadcast. At the bottom of the frame, the vehicle's massive tracks, constructed from black composite materials with metal hinges, dominate the field of view, their deep V-shaped patterns leaving a distinct imprint on the ground. As the vehicle moves forward with steady yet powerful force, the tracks cut deeply into the boundless, pure white snowfield, powerfully stirring and sweeping back vast clouds of shimmering snow and fine ice crystals. In the clear, cold polar sunlight, these flying snow particles refract like diamonds. As the vehicle reaches the edge of a magnificent glacier, its deep, crystalline blue ice looming below the cliffs, the vehicle executes an impeccable U-turn: the camera pans along with the vehicle in a grand, silky arc, one track slowing while the other powerfully propels it forward, carving a perfect semicircle across the vast snowy landscape with a display of absolute control. The entire sequence is imbued with unparalleled dynamic cinematography, seamlessly blending the power and precision of the machinery with the grandeur and harshness of the polar environment, creating a thrilling, epic atmosphere of exploration and conquest.", + "seed": 19316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19316_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0703700034879148, + "vlm_logprob": -0.0703700034879148, + "vlm_yes_no_margin": 3.324999976158142, + "vlm_token_credit": -0.0703700034879148, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15793545544147491 + ], + "opposite_logprob_tokens": [ + -2.032935380935669 + ], + "logprob_reward": -0.15793545544147491, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15793545544147491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039161212742328644, + 0.013133477419614792, + 0.04403125122189522, + 0.04835675284266472, + 0.018077220767736435, + 0.4487997889518738, + 0.0262448713183403, + 0.02717667445540428, + 0.30530139803886414, + 0.02971723861992359 + ], + "credit_map": [ + 0.005560071673244238, + 0.006522550247609615, + 0.005022994242608547, + 0.0027656564489006996, + 0.00819062627851963, + 0.005283924750983715, + 0.0028541777282953262, + 0.0029612150974571705, + 0.0024500112049281597, + 0.001122741145081818, + 0.0015678445342928171, + 0.001695239800028503, + 0.0019334872486069798, + 0.0017615375109016895, + 0.0013389263767749071, + 0.0012636908795684576, + 0.0035143187269568443, + 0.01278349943459034, + 0.005273811519145966, + 0.004706704523414373, + 0.0055844103917479515, + 0.005366581957787275, + 0.004133045673370361, + 0.002668884349986911, + 0.0035721834283322096, + 0.013875535689294338, + 0.007668882608413696, + 0.006198837421834469, + 0.004771008621901274, + 0.005226730369031429, + 0.00392109714448452, + 0.003122480819001794, + 0.0016643789131194353, + 0.0022380880545824766, + 0.002370164729654789, + 0.0017608879134058952, + 0.0048899282701313496, + 0.0024942951276898384, + 0.001391528407111764, + 0.0012679505161941051, + 0.011133835650980473, + 0.02280968800187111, + 0.016810327768325806, + 0.004788720980286598, + 0.20669640600681305, + 0.1749807745218277, + 0.0058168526738882065, + 0.005763241555541754, + 0.0025878644082695246, + 0.0042104884050786495, + 0.0034612778108567, + 0.003158775856718421, + 0.0026794332079589367, + 0.003775989403948188, + 0.003258675569668412, + 0.0031123673543334007, + 0.002891932148486376, + 0.0032014204189181328, + 0.004528930876404047, + 0.0033451695926487446, + 0.003272520611062646, + 0.0038166039157658815, + 0.0031543525401502848, + 0.0029657476115971804, + 0.030668135732412338, + 0.03374356031417847, + 0.04519429802894592, + 0.03835489973425865, + 0.03339826315641403, + 0.03647621348500252, + 0.03819505497813225, + 0.04927099868655205, + 0.0039234459400177, + 0.0045052641071379185, + 0.004051195457577705, + 0.0029991543851792812, + 0.00372005975805223, + 0.004445051308721304, + 0.002958420431241393, + 0.003114650957286358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.15793545544147491 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806079789996147 + ], + "opposite_logprob_tokens": [ + -4.634806156158447 + ], + "logprob_reward": -0.009806079789996147, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806079789996147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10079334676265717, + 0.04437531158328056, + 0.03405448794364929, + 0.04372171312570572, + 0.009082521311938763, + 0.5608769655227661, + 0.01807074062526226, + 0.017935970798134804, + 0.12908393144607544, + 0.0420050248503685 + ], + "credit_map": [ + 0.013082994148135185, + 0.011825728230178356, + 0.017798367887735367, + 0.013929598964750767, + 0.013257259503006935, + 0.011897580698132515, + 0.010735894553363323, + 0.008265922777354717, + 0.006906076800078154, + 0.005225519649684429, + 0.005675426684319973, + 0.00580555759370327, + 0.006942675448954105, + 0.006406567990779877, + 0.0035172582138329744, + 0.0038962301332503557, + 0.0037147975526750088, + 0.007006872911006212, + 0.003599621821194887, + 0.003481219056993723, + 0.004982943646609783, + 0.006120898295193911, + 0.003105606883764267, + 0.0020425268448889256, + 0.0037643061950802803, + 0.012773500755429268, + 0.006931567098945379, + 0.003623730270192027, + 0.004134862683713436, + 0.0077127572149038315, + 0.0023583737201988697, + 0.0024226135574281216, + 0.001039142138324678, + 0.001148040289990604, + 0.0011751134879887104, + 0.0006236538756638765, + 0.001589580555446446, + 0.00252247741445899, + 0.0004476331814657897, + 0.0005368799902498722, + 0.013651241548359394, + 0.025895630940794945, + 0.02123076841235161, + 0.010872656479477882, + 0.24611856043338776, + 0.22332054376602173, + 0.008469906635582447, + 0.011317651718854904, + 0.0016236663796007633, + 0.003592915367335081, + 0.002151804743334651, + 0.0020602315198630095, + 0.0017394112655892968, + 0.002520708367228508, + 0.0021438165567815304, + 0.0022381856106221676, + 0.0017547572497278452, + 0.0026820171624422073, + 0.0031113531440496445, + 0.0022638028021901846, + 0.002087901346385479, + 0.0021221302449703217, + 0.0020027137361466885, + 0.0019112948793917894, + 0.014996185898780823, + 0.016288407146930695, + 0.01876131445169449, + 0.016647066920995712, + 0.013614678755402565, + 0.013916621915996075, + 0.016008833423256874, + 0.018850814551115036, + 0.004980406723916531, + 0.007444943301379681, + 0.0039809755980968475, + 0.0042005316354334354, + 0.0059486273676157, + 0.007160910405218601, + 0.00440223328769207, + 0.0038863979279994965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.009806079789996147 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369199112057686 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.03369199112057686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369199112057686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13783036172389984, + 0.053363069891929626, + 0.036173272877931595, + 0.0597451776266098, + 0.02103269286453724, + 0.5293806195259094, + 0.0064943889155983925, + 0.011209978722035885, + 0.08122307807207108, + 0.06354735046625137 + ], + "credit_map": [ + 0.015918344259262085, + 0.018811466172337532, + 0.026620496064424515, + 0.01398196816444397, + 0.013365673832595348, + 0.01828014850616455, + 0.01890728436410427, + 0.01194499060511589, + 0.00798745732754469, + 0.010142238810658455, + 0.006832568906247616, + 0.0033288344275206327, + 0.007660697679966688, + 0.009978791698813438, + 0.003878450021147728, + 0.0035540363751351833, + 0.0037860372103750706, + 0.00886550173163414, + 0.005575898569077253, + 0.003362315008416772, + 0.0054873754270374775, + 0.0036375911440700293, + 0.003044825280085206, + 0.0024137254804372787, + 0.004793980624526739, + 0.01882464438676834, + 0.009475770406425, + 0.003812102135270834, + 0.005940630100667477, + 0.011553819291293621, + 0.0026022058445960283, + 0.0027420257683843374, + 0.002351438859477639, + 0.0027339644730091095, + 0.002724685473367572, + 0.0020722525659948587, + 0.0025238448288291693, + 0.004943378735333681, + 0.0016496478347107768, + 0.0020334813743829727, + 0.019481996074318886, + 0.03489016368985176, + 0.04074174538254738, + 0.009342527016997337, + 0.24074457585811615, + 0.16747009754180908, + 0.0081676309928298, + 0.008541864342987537, + 0.0008491398184560239, + 0.0009317465592175722, + 0.0007761486922390759, + 0.0007366248173639178, + 0.0009351812768727541, + 0.0009330118191428483, + 0.0006521008326672018, + 0.0006804352160543203, + 0.0014072539051994681, + 0.001592674758285284, + 0.0010061782086268067, + 0.0011253956472501159, + 0.0021503681782633066, + 0.0022560202050954103, + 0.0008303658105432987, + 0.000841722299810499, + 0.0096316272392869, + 0.009433050639927387, + 0.010995429940521717, + 0.010687029920518398, + 0.008326399140059948, + 0.010301228612661362, + 0.009709264151751995, + 0.012139047496020794, + 0.007792244665324688, + 0.007800002582371235, + 0.00802826788276434, + 0.005453782621771097, + 0.010870656929910183, + 0.011721032671630383, + 0.006796712521463633, + 0.005084652919322252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.03369199112057686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278505742549896 + ], + "opposite_logprob_tokens": [ + -2.01778507232666 + ], + "logprob_reward": -0.14278505742549896, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278505742549896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06672496348619461, + 0.03025045059621334, + 0.024372577667236328, + 0.04676635563373566, + 0.017595170065760612, + 0.7052432298660278, + 0.006665155291557312, + 0.007590769790112972, + 0.06650049239397049, + 0.028290845453739166 + ], + "credit_map": [ + 0.007345165591686964, + 0.008473670110106468, + 0.009168392978608608, + 0.007450193632394075, + 0.012124116532504559, + 0.009674465283751488, + 0.00567933963611722, + 0.006809616461396217, + 0.006423355545848608, + 0.00413159653544426, + 0.004021896980702877, + 0.0019563017413020134, + 0.004633130971342325, + 0.0051960875280201435, + 0.001793751260265708, + 0.0020943284034729004, + 0.00259026070125401, + 0.0060136811807751656, + 0.0031413440592586994, + 0.002849486656486988, + 0.0035563535057008266, + 0.002397137461230159, + 0.002049156930297613, + 0.0017751562409102917, + 0.004829879850149155, + 0.007885447703301907, + 0.006242448929697275, + 0.005787932313978672, + 0.005171362776309252, + 0.00976541731506586, + 0.003928928170353174, + 0.0031549425330013037, + 0.0023075214121490717, + 0.0029909920413047075, + 0.0021116649731993675, + 0.0012828626204282045, + 0.002704392885789275, + 0.004159989766776562, + 0.0010116004850715399, + 0.0010261444840580225, + 0.0177027378231287, + 0.03592028096318245, + 0.02673182636499405, + 0.00842718780040741, + 0.32524818181991577, + 0.27370429039001465, + 0.008700711652636528, + 0.008807971142232418, + 0.0006199437193572521, + 0.0009139481117017567, + 0.0009685590630397201, + 0.0007872771238908172, + 0.0011050008470192552, + 0.0007721169386059046, + 0.0008395247277803719, + 0.0006587851676158607, + 0.0010630562901496887, + 0.0008748667314648628, + 0.0008729688124731183, + 0.0007028344552963972, + 0.0011817595222964883, + 0.0015316110802814364, + 0.0006799750262871385, + 0.0006836982793174684, + 0.007254641968756914, + 0.008699631318449974, + 0.008912691846489906, + 0.00737359281629324, + 0.009662441909313202, + 0.008902776055037975, + 0.006645023822784424, + 0.009049691259860992, + 0.005712275393307209, + 0.003655958455055952, + 0.003277789568528533, + 0.0022658768575638533, + 0.003744852961972356, + 0.004348038230091333, + 0.0032400544732809067, + 0.0020459990482777357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.14278505742549896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1634853333234787, + 0.023916136473417282, + 0.045678455382585526, + 0.05902216210961342, + 0.02093776874244213, + 0.5305255055427551, + 0.022555867210030556, + 0.02453509159386158, + 0.054506633430719376, + 0.05483699217438698 + ], + "credit_map": [ + 0.01767619699239731, + 0.02037317305803299, + 0.022542443126440048, + 0.01713327318429947, + 0.026632338762283325, + 0.018652500584721565, + 0.026588108390569687, + 0.01388730388134718, + 0.004061725921928883, + 0.00277610938064754, + 0.002771739847958088, + 0.003015691414475441, + 0.0028509851545095444, + 0.0030274312011897564, + 0.002726157195866108, + 0.002686295658349991, + 0.004421841818839312, + 0.01285969652235508, + 0.006062075961381197, + 0.004510397091507912, + 0.00652717798948288, + 0.006287074647843838, + 0.0029919701628386974, + 0.00201822305098176, + 0.006610205862671137, + 0.010682017542421818, + 0.006325381342321634, + 0.00591749744489789, + 0.0076928106136620045, + 0.011950361542403698, + 0.005654738284647465, + 0.004189150873571634, + 0.002063779393211007, + 0.00372587819583714, + 0.001903277589008212, + 0.0017187042394652963, + 0.003130967728793621, + 0.005222474690526724, + 0.0013502128422260284, + 0.0018224744126200676, + 0.01322549395263195, + 0.02526107430458069, + 0.02015465497970581, + 0.008080082014203072, + 0.23559901118278503, + 0.212408646941185, + 0.007323204539716244, + 0.008473294787108898, + 0.0019374085823073983, + 0.002399636898189783, + 0.0033309878781437874, + 0.0033664945513010025, + 0.0030047271866351366, + 0.002315301913768053, + 0.0032107362058013678, + 0.0029905743431299925, + 0.0025532462168484926, + 0.0027993093244731426, + 0.0038126809522509575, + 0.002310291165485978, + 0.0036784783005714417, + 0.004189168103039265, + 0.0025250837206840515, + 0.0026668331120163202, + 0.004558820743113756, + 0.012914234772324562, + 0.004386392887681723, + 0.004475060384720564, + 0.0074673667550086975, + 0.012100135907530785, + 0.0050978753715753555, + 0.003506744746118784, + 0.006342178676277399, + 0.008176682516932487, + 0.0073870765045285225, + 0.006143078673630953, + 0.009608165360987186, + 0.007300150580704212, + 0.00576810771599412, + 0.004111554007977247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.007631433662027121 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly dynamic first-person perspective shot places the viewer directly in the cockpit of a high-tech, specialized polar transport vehicle. The camera, with its wide-angle perspective, moves forward steadily and smoothly, as if supported by a top-tier gyroscopic stabilizer, imbued with the high-definition quality of an Olympic-level broadcast. At the bottom of the frame, the vehicle's massive tracks, constructed from black composite materials with metal hinges, dominate the field of view, their deep V-shaped patterns leaving a distinct imprint on the ground. As the vehicle moves forward with steady yet powerful force, the tracks cut deeply into the boundless, pure white snowfield, powerfully stirring and sweeping back vast clouds of shimmering snow and fine ice crystals. In the clear, cold polar sunlight, these flying snow particles refract like diamonds. As the vehicle reaches the edge of a magnificent glacier, its deep, crystalline blue ice looming below the cliffs, the vehicle executes an impeccable U-turn: the camera pans along with the vehicle in a grand, silky arc, one track slowing while the other powerfully propels it forward, carving a perfect semicircle across the vast snowy landscape with a display of absolute control. The entire sequence is imbued with unparalleled dynamic cinematography, seamlessly blending the power and precision of the machinery with the grandeur and harshness of the polar environment, creating a thrilling, epic atmosphere of exploration and conquest.", + "seed": 19317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19317_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09760238062590361, + "vlm_logprob": -0.09760238062590361, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.09760238062590361, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1194501519203186 + ], + "opposite_logprob_tokens": [ + -2.244450092315674 + ], + "logprob_reward": -0.1194501519203186, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1194501519203186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10834992676973343, + 0.03250344470143318, + 0.25413626432418823, + 0.05997950956225395, + 0.10926392674446106, + 0.1202731654047966, + 0.1166837066411972, + 0.08371138572692871, + 0.06504524499177933, + 0.05005338788032532 + ], + "credit_map": [ + 0.017915714532136917, + 0.01618890091776848, + 0.01363326609134674, + 0.011054963804781437, + 0.014368775300681591, + 0.01570351980626583, + 0.010035816580057144, + 0.009448970668017864, + 0.004716430325061083, + 0.004352605901658535, + 0.004832539241760969, + 0.002723296405747533, + 0.004056726116687059, + 0.005301302298903465, + 0.0028080542106181383, + 0.003712491597980261, + 0.03916645795106888, + 0.029808083549141884, + 0.045228518545627594, + 0.022981027141213417, + 0.03484988212585449, + 0.03014410100877285, + 0.03031119331717491, + 0.021647006273269653, + 0.008638977073132992, + 0.00678289495408535, + 0.013615964911878109, + 0.006162354256957769, + 0.006060461513698101, + 0.008772221393883228, + 0.004658297635614872, + 0.00528833968564868, + 0.005536815617233515, + 0.0423816442489624, + 0.021464234218001366, + 0.0030995593406260014, + 0.005146564915776253, + 0.005982743110507727, + 0.0037597487680613995, + 0.021892612800002098, + 0.012919659726321697, + 0.03574458509683609, + 0.02763969823718071, + 0.004934912081807852, + 0.01204543374478817, + 0.008683902211487293, + 0.007369258441030979, + 0.010935721918940544, + 0.011629736982285976, + 0.017421986907720566, + 0.016038762405514717, + 0.010247373953461647, + 0.012913167476654053, + 0.01319566648453474, + 0.014931158162653446, + 0.020305855199694633, + 0.007789365015923977, + 0.011576618999242783, + 0.014735745266079903, + 0.014257323928177357, + 0.01062262523919344, + 0.009270078502595425, + 0.008326771669089794, + 0.007132858969271183, + 0.003684557043015957, + 0.020898427814245224, + 0.0036897039972245693, + 0.002282507950440049, + 0.02179400622844696, + 0.005884437821805477, + 0.002660353435203433, + 0.004151251167058945, + 0.006114729680120945, + 0.00700189545750618, + 0.007191079668700695, + 0.0051939478144049644, + 0.006167926825582981, + 0.006945081055164337, + 0.005653096362948418, + 0.005785628687590361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.1194501519203186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097966693341732 + ], + "opposite_logprob_tokens": [ + -4.5110979080200195 + ], + "logprob_reward": -0.011097966693341732, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097966693341732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07312265038490295, + 0.06866620481014252, + 0.08906731009483337, + 0.12272534519433975, + 0.23234237730503082, + 0.057291217148303986, + 0.0927339717745781, + 0.08992724865674973, + 0.05891706049442291, + 0.11520664393901825 + ], + "credit_map": [ + 0.013944690115749836, + 0.00696900999173522, + 0.010544358752667904, + 0.0064082155004143715, + 0.009667905047535896, + 0.01143060252070427, + 0.005868598818778992, + 0.008289272896945477, + 0.008814681321382523, + 0.007518352475017309, + 0.009628701023757458, + 0.007095552049577236, + 0.008855373598635197, + 0.014873573556542397, + 0.004797453060746193, + 0.007082517724484205, + 0.011952213011682034, + 0.01373870950192213, + 0.01211222168058157, + 0.007880013436079025, + 0.011801536194980145, + 0.010435834527015686, + 0.010997068136930466, + 0.010149708017706871, + 0.015238793566823006, + 0.01544963475316763, + 0.024413293227553368, + 0.012821118347346783, + 0.013409050181508064, + 0.016080789268016815, + 0.013799063861370087, + 0.0115136057138443, + 0.01316115353256464, + 0.0455634705722332, + 0.08502492308616638, + 0.008463905192911625, + 0.006589413620531559, + 0.013266963884234428, + 0.0063768415711820126, + 0.05389571189880371, + 0.005828298162668943, + 0.006050509866327047, + 0.007886392064392567, + 0.0038778192829340696, + 0.01212320290505886, + 0.009411746636033058, + 0.005591578781604767, + 0.006521677132695913, + 0.009006399661302567, + 0.011188415810465813, + 0.011529643088579178, + 0.00741418544203043, + 0.009951111860573292, + 0.017239119857549667, + 0.010578838177025318, + 0.015826255083084106, + 0.006519767455756664, + 0.0228529442101717, + 0.008327323943376541, + 0.006110576447099447, + 0.018435275182127953, + 0.011546468362212181, + 0.007218585815280676, + 0.008916305378079414, + 0.0048105609603226185, + 0.013596607372164726, + 0.005886836443096399, + 0.004089824389666319, + 0.015555266290903091, + 0.007713763974606991, + 0.002941656392067671, + 0.0043225474655628204, + 0.0081355394795537, + 0.010690341703593731, + 0.011422443203628063, + 0.009751551784574986, + 0.0257687047123909, + 0.024479029700160027, + 0.0153826167806983, + 0.009576413780450821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.011097966693341732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808823600411415 + ], + "opposite_logprob_tokens": [ + -3.288088321685791 + ], + "logprob_reward": -0.03808823600411415, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808823600411415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19892321527004242, + 0.1310303509235382, + 0.08949396014213562, + 0.046400535851716995, + 0.12171756476163864, + 0.06172793731093407, + 0.09734142571687698, + 0.060313526540994644, + 0.1122964471578598, + 0.08075501769781113 + ], + "credit_map": [ + 0.04812981188297272, + 0.02611800841987133, + 0.02807449921965599, + 0.015955939888954163, + 0.03672013804316521, + 0.017134802415966988, + 0.012744640000164509, + 0.01404535211622715, + 0.018552787601947784, + 0.01653366908431053, + 0.016129525378346443, + 0.014357018284499645, + 0.01417822390794754, + 0.0247209370136261, + 0.010961138643324375, + 0.015597055666148663, + 0.012480180710554123, + 0.015972929075360298, + 0.011329326778650284, + 0.00699561508372426, + 0.013582965359091759, + 0.012501929886639118, + 0.008983449079096317, + 0.007647561840713024, + 0.00550702353939414, + 0.00446744030341506, + 0.008006097748875618, + 0.005011138040572405, + 0.005131969694048166, + 0.0056231338530778885, + 0.006058746483176947, + 0.006594989914447069, + 0.005357623100280762, + 0.030829016119241714, + 0.0498405396938324, + 0.0030755754560232162, + 0.0038776565343141556, + 0.004678534809499979, + 0.0027192416600883007, + 0.021339379251003265, + 0.005306323058903217, + 0.015019751153886318, + 0.012714638374745846, + 0.0033866444136947393, + 0.00667745852842927, + 0.0064592971466481686, + 0.006090931128710508, + 0.006072895135730505, + 0.009476300328969955, + 0.014688362367451191, + 0.01798316463828087, + 0.006651807110756636, + 0.013645835220813751, + 0.010801371186971664, + 0.007595645729452372, + 0.016498945653438568, + 0.005913204047828913, + 0.009021042846143246, + 0.012322564609348774, + 0.0032803211361169815, + 0.010287400335073471, + 0.00799213070422411, + 0.004653960932046175, + 0.006842904724180698, + 0.007484931033104658, + 0.02452714554965496, + 0.018642578274011612, + 0.010385253466665745, + 0.01984979212284088, + 0.01517262402921915, + 0.0071834418922662735, + 0.009050683118402958, + 0.007455728482455015, + 0.009870599023997784, + 0.00666129169985652, + 0.0061522191390395164, + 0.018518272787332535, + 0.017149941995739937, + 0.007582968566566706, + 0.007363992743194103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03808823600411415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134268522262573 + ], + "opposite_logprob_tokens": [ + -1.3134268522262573 + ], + "logprob_reward": -0.3134268522262573, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134268522262573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10599024593830109, + 0.11273520439863205, + 0.15753766894340515, + 0.047068074345588684, + 0.13341090083122253, + 0.05453016236424446, + 0.05783369392156601, + 0.0630040168762207, + 0.21308434009552002, + 0.05480566993355751 + ], + "credit_map": [ + 0.019754137843847275, + 0.014103195630013943, + 0.01407614815980196, + 0.013985532335937023, + 0.012704622931778431, + 0.0125947967171669, + 0.009962624870240688, + 0.008809181861579418, + 0.018585581332445145, + 0.01102248951792717, + 0.020521780475974083, + 0.009320511482656002, + 0.014193105511367321, + 0.021992065012454987, + 0.007619783282279968, + 0.009479882195591927, + 0.018382949754595757, + 0.02440924569964409, + 0.028156215324997902, + 0.010903006419539452, + 0.020944153890013695, + 0.01958400569856167, + 0.017832830548286438, + 0.017325252294540405, + 0.00590753648430109, + 0.005873952992260456, + 0.0073111047968268394, + 0.004842149559408426, + 0.00549354637041688, + 0.006430741399526596, + 0.005189566873013973, + 0.006019474007189274, + 0.007847030647099018, + 0.05012029409408569, + 0.04187238588929176, + 0.003881753422319889, + 0.006532320287078619, + 0.008198367431759834, + 0.004126960877329111, + 0.010831786319613457, + 0.004727697465568781, + 0.006815335247665644, + 0.007973084226250648, + 0.00386025314219296, + 0.010121503844857216, + 0.009365423582494259, + 0.005632520653307438, + 0.006034339312463999, + 0.007842675782740116, + 0.006453096866607666, + 0.010498096235096455, + 0.0062700617127120495, + 0.005406852811574936, + 0.009258662350475788, + 0.004767363891005516, + 0.007336881011724472, + 0.004512877203524113, + 0.01220174040645361, + 0.012309898622334003, + 0.004781057592481375, + 0.012318928726017475, + 0.006496313493698835, + 0.004521658178418875, + 0.005861545447260141, + 0.015493801794946194, + 0.045237693935632706, + 0.018560752272605896, + 0.015907349064946175, + 0.06016983464360237, + 0.02879326604306698, + 0.012260817922651768, + 0.016660815104842186, + 0.005645893979817629, + 0.005976124666631222, + 0.007338386029005051, + 0.006482265889644623, + 0.006308944430202246, + 0.007102278061211109, + 0.00883309543132782, + 0.007118679583072662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.3134268522262573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948696285486221 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948696285486221, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948696285486221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13259199261665344, + 0.05399179086089134, + 0.19671419262886047, + 0.07913315296173096, + 0.11226306855678558, + 0.05445142462849617, + 0.05375697836279869, + 0.09643500298261642, + 0.13660013675689697, + 0.08406230062246323 + ], + "credit_map": [ + 0.034799493849277496, + 0.015872566029429436, + 0.01770118996500969, + 0.009116614237427711, + 0.02441488392651081, + 0.011852371506392956, + 0.008098339661955833, + 0.010736534371972084, + 0.008872694335877895, + 0.007408871781080961, + 0.007666790392249823, + 0.0044123269617557526, + 0.006895372644066811, + 0.00868444237858057, + 0.0036390838213264942, + 0.006412206683307886, + 0.0237268079072237, + 0.0359308235347271, + 0.027550868690013885, + 0.013926704414188862, + 0.02590220607817173, + 0.02614300325512886, + 0.026854107156395912, + 0.01667967438697815, + 0.013200407847762108, + 0.009504909627139568, + 0.017540059983730316, + 0.007317634765058756, + 0.007480032742023468, + 0.012096256017684937, + 0.005217116326093674, + 0.0067767370492219925, + 0.011229784227907658, + 0.03108055703341961, + 0.03370709717273712, + 0.00432105315849185, + 0.005595074035227299, + 0.011204504407942295, + 0.004668999928981066, + 0.010455996729433537, + 0.005105556920170784, + 0.008139865472912788, + 0.009000485762953758, + 0.0036246711388230324, + 0.008503492921590805, + 0.007238840684294701, + 0.005337439477443695, + 0.0075010755099356174, + 0.005567273125052452, + 0.011446165852248669, + 0.00622137263417244, + 0.0037754324730485678, + 0.008826245553791523, + 0.0061385068111121655, + 0.004347411449998617, + 0.007434566970914602, + 0.006567538715898991, + 0.021321024745702744, + 0.011245733126997948, + 0.008818000555038452, + 0.018208784982562065, + 0.012170957401394844, + 0.0089379558339715, + 0.00916500948369503, + 0.00941690057516098, + 0.032053399831056595, + 0.02084982395172119, + 0.010565726086497307, + 0.027618704363703728, + 0.01833830587565899, + 0.0077436137944459915, + 0.01001365203410387, + 0.007055662106722593, + 0.009388833306729794, + 0.007891802117228508, + 0.0066307177767157555, + 0.017293689772486687, + 0.017471758648753166, + 0.010156596079468727, + 0.008173241280019283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.005948696285486221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly dynamic first-person perspective shot places the viewer directly in the cockpit of a high-tech, specialized polar transport vehicle. The camera, with its wide-angle perspective, moves forward steadily and smoothly, as if supported by a top-tier gyroscopic stabilizer, imbued with the high-definition quality of an Olympic-level broadcast. At the bottom of the frame, the vehicle's massive tracks, constructed from black composite materials with metal hinges, dominate the field of view, their deep V-shaped patterns leaving a distinct imprint on the ground. As the vehicle moves forward with steady yet powerful force, the tracks cut deeply into the boundless, pure white snowfield, powerfully stirring and sweeping back vast clouds of shimmering snow and fine ice crystals. In the clear, cold polar sunlight, these flying snow particles refract like diamonds. As the vehicle reaches the edge of a magnificent glacier, its deep, crystalline blue ice looming below the cliffs, the vehicle executes an impeccable U-turn: the camera pans along with the vehicle in a grand, silky arc, one track slowing while the other powerfully propels it forward, carving a perfect semicircle across the vast snowy landscape with a display of absolute control. The entire sequence is imbued with unparalleled dynamic cinematography, seamlessly blending the power and precision of the machinery with the grandeur and harshness of the polar environment, creating a thrilling, epic atmosphere of exploration and conquest.", + "seed": 19318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_7/index_35_seed_19318_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11645984128117562, + "vlm_logprob": -0.11645984128117562, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.11645984128117562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03002462349832058 + ], + "opposite_logprob_tokens": [ + -3.780024528503418 + ], + "logprob_reward": -0.03002462349832058, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03002462349832058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043345339596271515, + 0.014384323731064796, + 0.027497977018356323, + 0.297648549079895, + 0.5110609531402588, + 0.02208724059164524, + 0.04293745383620262, + 0.005799099802970886, + 0.010936221107840538, + 0.02430282160639763 + ], + "credit_map": [ + 0.0059231119230389595, + 0.007925995625555515, + 0.0037801479920744896, + 0.0053675794042646885, + 0.008642494678497314, + 0.004430350847542286, + 0.0035975214559584856, + 0.0036781372036784887, + 0.001551828463561833, + 0.0028378383722156286, + 0.0013443513307720423, + 0.0012163753854110837, + 0.001662693452090025, + 0.0021256946492940187, + 0.001965535106137395, + 0.001680006505921483, + 0.0026824057567864656, + 0.006656726356595755, + 0.0017661438323557377, + 0.003145515453070402, + 0.003880465170368552, + 0.0037515873555094004, + 0.002991829998791218, + 0.002623302163556218, + 0.006264097057282925, + 0.0124620096758008, + 0.004045387730002403, + 0.003360442817211151, + 0.1373007744550705, + 0.1274050772190094, + 0.0034356408286839724, + 0.0033751181326806545, + 0.05945079028606415, + 0.15820780396461487, + 0.028856467455625534, + 0.058137912303209305, + 0.08539702743291855, + 0.07256441563367844, + 0.02957145869731903, + 0.018875105306506157, + 0.0017288880189880729, + 0.002677439944818616, + 0.002631939249113202, + 0.002860702807083726, + 0.0028704851865768433, + 0.003266238374635577, + 0.003343123011291027, + 0.0027084231842309237, + 0.003633512882515788, + 0.006910845637321472, + 0.00640782481059432, + 0.003521883161738515, + 0.006018381100147963, + 0.0076124402694404125, + 0.005147119518369436, + 0.003685451578348875, + 0.0006254198960959911, + 0.0007986206328496337, + 0.00042375942575745285, + 0.00047164681018330157, + 0.0010820267489179969, + 0.0008419474470429122, + 0.0009718157234601676, + 0.0005838631768710911, + 0.0006319143576547503, + 0.0005290996050462127, + 0.0009274102631025016, + 0.0008113891817629337, + 0.002894815057516098, + 0.0029852790758013725, + 0.0010458504548296332, + 0.001110463053919375, + 0.0012355310609564185, + 0.0017861212836578488, + 0.0019239640096202493, + 0.0023242044262588024, + 0.00743497908115387, + 0.005967569537460804, + 0.0022986356634646654, + 0.0013318154960870743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03002462349832058 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272767972201109 + ], + "opposite_logprob_tokens": [ + -5.25527286529541 + ], + "logprob_reward": -0.005272767972201109, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272767972201109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05334021523594856, + 0.03524564579129219, + 0.06505004316568375, + 0.27430951595306396, + 0.09336169064044952, + 0.039590612053871155, + 0.23020358383655548, + 0.024694399908185005, + 0.07589283585548401, + 0.10831146687269211 + ], + "credit_map": [ + 0.007698895875364542, + 0.00791819579899311, + 0.0046732318587601185, + 0.005499792750924826, + 0.008102436549961567, + 0.007021258119493723, + 0.006689597852528095, + 0.005736805964261293, + 0.0031672860495746136, + 0.0071516078896820545, + 0.003169653471559286, + 0.0027921688742935658, + 0.005559151992201805, + 0.004760278854519129, + 0.004942862316966057, + 0.0037026347126811743, + 0.007165494840592146, + 0.013447948731482029, + 0.005844166502356529, + 0.006241826806217432, + 0.008034217171370983, + 0.00829544197767973, + 0.006593775469809771, + 0.00942716933786869, + 0.012622460722923279, + 0.024193061515688896, + 0.007979673333466053, + 0.006037143524736166, + 0.11742900311946869, + 0.09012582898139954, + 0.00778638431802392, + 0.008135955780744553, + 0.009353087283670902, + 0.022419946268200874, + 0.00855608657002449, + 0.007867563515901566, + 0.01735479198396206, + 0.00893943477421999, + 0.008503960445523262, + 0.010366817936301231, + 0.0038694285321980715, + 0.006302344147115946, + 0.005058684851974249, + 0.0034517389722168446, + 0.0063795973546803, + 0.005122937262058258, + 0.005314181093126535, + 0.004091698210686445, + 0.01189357228577137, + 0.018096789717674255, + 0.03317828103899956, + 0.014976917766034603, + 0.0649760439991951, + 0.0574427954852581, + 0.016007227823138237, + 0.013631940819323063, + 0.002024195622652769, + 0.0020340997725725174, + 0.003181400476023555, + 0.003984034061431885, + 0.0026335264556109905, + 0.0032880594953894615, + 0.004132275935262442, + 0.0034168101847171783, + 0.0033987914212048054, + 0.0023823161609470844, + 0.007201094180345535, + 0.006563346367329359, + 0.020167972892522812, + 0.021055785939097404, + 0.006470953114330769, + 0.008652575314044952, + 0.006240337621420622, + 0.008588671684265137, + 0.008658502250909805, + 0.008178823627531528, + 0.03424038365483284, + 0.027924569323658943, + 0.009920547716319561, + 0.004559630062431097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.005272767972201109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064731404185295 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064731404185295, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064731404185295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04798440635204315, + 0.07808201760053635, + 0.06466621905565262, + 0.15461808443069458, + 0.23508401215076447, + 0.021370699629187584, + 0.09602546691894531, + 0.027313556522130966, + 0.077250637114048, + 0.19760486483573914 + ], + "credit_map": [ + 0.0067838020622730255, + 0.00827175471931696, + 0.004470126703381538, + 0.003921590279787779, + 0.007243357598781586, + 0.008272210136055946, + 0.005170348100364208, + 0.0038512167520821095, + 0.008980164304375648, + 0.015813952311873436, + 0.008499057032167912, + 0.007964431308209896, + 0.01205278281122446, + 0.007453924044966698, + 0.006888924166560173, + 0.010428783483803272, + 0.006310511380434036, + 0.017088916152715683, + 0.00432349368929863, + 0.005081964656710625, + 0.009957065805792809, + 0.008997324854135513, + 0.005099373869597912, + 0.007807575166225433, + 0.010763238184154034, + 0.013389810919761658, + 0.007648455444723368, + 0.010703601874411106, + 0.04811130464076996, + 0.0454501211643219, + 0.008740969002246857, + 0.009810574352741241, + 0.02647215500473976, + 0.058825839310884476, + 0.02193041518330574, + 0.022154977545142174, + 0.039518967270851135, + 0.022164087742567062, + 0.024582214653491974, + 0.01943536475300789, + 0.0015610839473083615, + 0.00276239519007504, + 0.0030461642891168594, + 0.0023538845125585794, + 0.0028355035465210676, + 0.0028998940251767635, + 0.0027770542073994875, + 0.0031347214244306087, + 0.0052821338176727295, + 0.006547619588673115, + 0.013744642958045006, + 0.006890220567584038, + 0.023585272952914238, + 0.025187915191054344, + 0.00778565276414156, + 0.0070020053535699844, + 0.002139806281775236, + 0.002453496912494302, + 0.003488876624032855, + 0.002965077292174101, + 0.0033035334199666977, + 0.004078628960996866, + 0.00494407769292593, + 0.00394006073474884, + 0.004728137981146574, + 0.0038897120393812656, + 0.007911394350230694, + 0.005956531036645174, + 0.015911810100078583, + 0.01752517558634281, + 0.00857481174170971, + 0.012753072194755077, + 0.008393327705562115, + 0.013737858273088932, + 0.011083993129432201, + 0.016633201390504837, + 0.06340889632701874, + 0.05588025972247124, + 0.0187823623418808, + 0.009684950113296509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.016064731404185295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233050584793091 + ], + "opposite_logprob_tokens": [ + -0.8983050584793091 + ], + "logprob_reward": -0.5233050584793091, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233050584793091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04692808911204338, + 0.06675893813371658, + 0.04045642912387848, + 0.4451594650745392, + 0.15614812076091766, + 0.08851402997970581, + 0.032424937933683395, + 0.02295760251581669, + 0.04056930169463158, + 0.06008300185203552 + ], + "credit_map": [ + 0.005638155620545149, + 0.007733793463557959, + 0.004783024080097675, + 0.004735428839921951, + 0.007633025757968426, + 0.006097708363085985, + 0.005897150840610266, + 0.004409797955304384, + 0.007775058504194021, + 0.017014605924487114, + 0.006412309594452381, + 0.005778627935796976, + 0.01170550286769867, + 0.007641688920557499, + 0.005048770923167467, + 0.00538237951695919, + 0.004081682302057743, + 0.010793792083859444, + 0.003250812878832221, + 0.003424749244004488, + 0.006773993372917175, + 0.005060631316155195, + 0.002951363567262888, + 0.004119403660297394, + 0.009571501053869724, + 0.019888171926140785, + 0.005082877818495035, + 0.004716312047094107, + 0.20599369704723358, + 0.18986475467681885, + 0.0049193790182471275, + 0.005122801288962364, + 0.018691513687372208, + 0.04385264962911606, + 0.011140136979520321, + 0.01926790364086628, + 0.020848074927926064, + 0.024966290220618248, + 0.010334087535738945, + 0.007047473918646574, + 0.006586118601262569, + 0.0072503904812037945, + 0.015242918394505978, + 0.010816648602485657, + 0.00970048550516367, + 0.01387245673686266, + 0.010781819000840187, + 0.014263192191720009, + 0.003445890499278903, + 0.003843852551653981, + 0.00357412570156157, + 0.003417110303416848, + 0.005785860121250153, + 0.004702380392700434, + 0.004236762877553701, + 0.003418952226638794, + 0.002106336411088705, + 0.002174579305574298, + 0.002186692086979747, + 0.002146185142919421, + 0.004036278929561377, + 0.0034691609907895327, + 0.004420509096235037, + 0.0024178619496524334, + 0.0022457470186054707, + 0.0018592557171359658, + 0.00445632403716445, + 0.003641042858362198, + 0.008989132009446621, + 0.009426884353160858, + 0.004305901471525431, + 0.0056450143456459045, + 0.0028097082395106554, + 0.003147258423268795, + 0.003565169870853424, + 0.00726223923265934, + 0.011299570091068745, + 0.01727152429521084, + 0.007892774417996407, + 0.0068347579799592495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.5233050584793091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside the cockpit of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the vehicle's massive black composite tracks with metal hinges visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle move forward steadily, cutting deep into a pure white snowfield and stirring clouds of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vehicle execute a U-turn at the edge of a glacier with deep crystalline blue ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a silky arc along with the vehicle during the U-turn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632025051862001 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632025051862001, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632025051862001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05844879522919655, + 0.06901130080223083, + 0.08040410280227661, + 0.30427324771881104, + 0.26067453622817993, + 0.0690753236413002, + 0.039918746799230576, + 0.018157118931412697, + 0.042919714003801346, + 0.05711716040968895 + ], + "credit_map": [ + 0.007965264841914177, + 0.008702731691300869, + 0.0058326395228505135, + 0.0060880547389388084, + 0.009371173568069935, + 0.006562258116900921, + 0.008090471848845482, + 0.005836202297359705, + 0.007516826037317514, + 0.01615946553647518, + 0.006048419047147036, + 0.006393482908606529, + 0.012314743362367153, + 0.007178938016295433, + 0.0061880131252110004, + 0.007211410440504551, + 0.006911843549460173, + 0.024230683222413063, + 0.005594614893198013, + 0.005726618692278862, + 0.014620481990277767, + 0.009752645157277584, + 0.005815241951495409, + 0.007751970551908016, + 0.014671467244625092, + 0.0300707146525383, + 0.00952060054987669, + 0.008848197758197784, + 0.1290263831615448, + 0.09633126109838486, + 0.00867061410099268, + 0.007134023122489452, + 0.029570873826742172, + 0.0754501149058342, + 0.016568847000598907, + 0.03197300434112549, + 0.03735353425145149, + 0.041126713156700134, + 0.018559245392680168, + 0.01007218100130558, + 0.0058434451930224895, + 0.008050895296037197, + 0.010524013079702854, + 0.008046496659517288, + 0.007153559010475874, + 0.011828572489321232, + 0.008067546412348747, + 0.009560797363519669, + 0.004362381063401699, + 0.006227488163858652, + 0.0020485955756157637, + 0.0019248767057433724, + 0.010666906833648682, + 0.009028932079672813, + 0.0036198548041284084, + 0.0020397105254232883, + 0.001651439699344337, + 0.0020455513149499893, + 0.001700704568065703, + 0.0015202222857624292, + 0.0031751454807817936, + 0.0026234665419906378, + 0.0023391658905893564, + 0.0031014210544526577, + 0.002670754212886095, + 0.001824000384658575, + 0.0036983611062169075, + 0.0036079210694879293, + 0.009262320585548878, + 0.010611169040203094, + 0.0038083959370851517, + 0.007436794228851795, + 0.0033151232637465, + 0.00460783438757062, + 0.005300191696733236, + 0.003314831294119358, + 0.016307832673192024, + 0.0132921626791358, + 0.00586015684530139, + 0.005119029898196459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.007632025051862001 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_tensor([2566, 2566])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_tensor([2566, 2566])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a78558517e12307f64bcf344225e5eaed8d1cd48 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_tensor([2566, 2566])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows an ornate carriage as it speeds along a dry, dirt country road. The carriage, with its crimson lacquer and intricately carved gold trim, gleams luxuriously in the evening sun. Four powerful, glossy, dark brown horses gallop with gusto, their muscles straining, the brass fittings of their leather harnesses swaying with their movements. The driver, dressed in a dark uniform and a top hat, maintains a focused and resolute expression as he grips the reins, his body swaying with the carriage's violent jolting. Massive wooden wheels, rimmed with iron hoops, spin at high speed, rolling over the ground, flinging loose dirt and gravel backward and raising a thick, tawny cloud of dust. This cloud, pierced by golden sunlight and shimmering with halos at the edges, trails behind the carriage like a magnificent tail, nearly engulfing the green hills and winding road in the distance. The whole picture presents a tense atmosphere full of dramatic tension, speed and retro luxury in the warm light and shadow.", + "seed": 30625, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_2566_seed_30625_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022035743575543165, + "vlm_logprob": -0.022035743575543165, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.022035743575543165, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02993866801261902 + ], + "opposite_logprob_tokens": [ + -3.6549386978149414 + ], + "logprob_reward": -0.02993866801261902, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02993866801261902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4048241972923279, + 0.025662800297141075, + 0.05558666214346886, + 0.101508229970932, + 0.04308038949966431, + 0.057135988026857376, + 0.035735126584768295, + 0.050021931529045105, + 0.0384710356593132, + 0.18797370791435242 + ], + "credit_map": [ + 0.09041805565357208, + 0.029293105006217957, + 0.13805143535137177, + 0.03949722647666931, + 0.02500070072710514, + 0.023175174370408058, + 0.022542407736182213, + 0.036846015602350235, + 0.0032135434448719025, + 0.005118803586810827, + 0.0025928772520273924, + 0.0034128003753721714, + 0.0021617927122861147, + 0.001973437611013651, + 0.0031265048310160637, + 0.004063036758452654, + 0.006775949615985155, + 0.014880540780723095, + 0.005839083809405565, + 0.006940625607967377, + 0.0053342124447226524, + 0.004541176371276379, + 0.005013007204979658, + 0.006262060720473528, + 0.013168230652809143, + 0.01804855279624462, + 0.01584295928478241, + 0.010825195349752903, + 0.011805705726146698, + 0.007711158134043217, + 0.01240956038236618, + 0.011696865782141685, + 0.004872180055826902, + 0.005917432717978954, + 0.0042082578875124454, + 0.007019957061856985, + 0.004350006580352783, + 0.004599695093929768, + 0.007309840992093086, + 0.004803014453500509, + 0.007121656090021133, + 0.005915139801800251, + 0.005850621499121189, + 0.008103826083242893, + 0.008118883706629276, + 0.006991589441895485, + 0.0076180496253073215, + 0.007416215725243092, + 0.0036073848605155945, + 0.006164721678942442, + 0.004043554421514273, + 0.00951314251869917, + 0.003335808403789997, + 0.003426219802349806, + 0.00282975472509861, + 0.0028145366813987494, + 0.00585540384054184, + 0.012140742503106594, + 0.004433147143572569, + 0.0072555216029286385, + 0.004959096200764179, + 0.004612219054251909, + 0.006129562854766846, + 0.0046362304128706455, + 0.004640401341021061, + 0.006249543745070696, + 0.004648872185498476, + 0.006037186831235886, + 0.004376426804810762, + 0.0039814854972064495, + 0.004613107535988092, + 0.003924008924514055, + 0.021994223818182945, + 0.021105239167809486, + 0.01939641498029232, + 0.01936517283320427, + 0.042994093149900436, + 0.021525008603930473, + 0.02227194420993328, + 0.019321594387292862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02993866801261902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381341353058815 + ], + "opposite_logprob_tokens": [ + -3.2881340980529785 + ], + "logprob_reward": -0.0381341353058815, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381341353058815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6895962357521057, + 0.020261021330952644, + 0.05240660533308983, + 0.055494535714387894, + 0.021393224596977234, + 0.06126624718308449, + 0.013507192954421043, + 0.03399459645152092, + 0.015715062618255615, + 0.0363653302192688 + ], + "credit_map": [ + 0.18201512098312378, + 0.041093215346336365, + 0.26993659138679504, + 0.0618857815861702, + 0.025505749508738518, + 0.03027753345668316, + 0.025814607739448547, + 0.053067587316036224, + 0.002417374402284622, + 0.0026404631789773703, + 0.0020873486064374447, + 0.0028442528564482927, + 0.002717410447075963, + 0.00223599118180573, + 0.00233770371414721, + 0.0029804750811308622, + 0.006383475847542286, + 0.008083075284957886, + 0.006161706522107124, + 0.005640366114675999, + 0.005965152755379677, + 0.0065393089316785336, + 0.00624140165746212, + 0.0073921168223023415, + 0.007264235056936741, + 0.006411457899957895, + 0.005625967867672443, + 0.006752753164619207, + 0.004929421003907919, + 0.004561256151646376, + 0.010140432976186275, + 0.009809006936848164, + 0.002933231648057699, + 0.0020603302400559187, + 0.0018913553794845939, + 0.003960742149502039, + 0.002163482131436467, + 0.0024610699620097876, + 0.0034050780814141035, + 0.002517933724448085, + 0.010743452236056328, + 0.0066059427335858345, + 0.006309431977570057, + 0.010230504907667637, + 0.007647743448615074, + 0.0061074853874742985, + 0.006375747732818127, + 0.007245935965329409, + 0.0017934390343725681, + 0.0017988893669098616, + 0.0015125095378607512, + 0.002176515990868211, + 0.0012830826453864574, + 0.0013868812238797545, + 0.0015811342746019363, + 0.0019747407641261816, + 0.0042222836054861546, + 0.008029943332076073, + 0.0029495377093553543, + 0.004388884175568819, + 0.003496892051771283, + 0.004550389014184475, + 0.0034810607321560383, + 0.002875605132430792, + 0.001655088970437646, + 0.0030078054405748844, + 0.0019597308710217476, + 0.0017485112184658647, + 0.0024338816292583942, + 0.0015807590680196881, + 0.002017943188548088, + 0.0013113413006067276, + 0.005323808640241623, + 0.003802781691774726, + 0.003530619665980339, + 0.004337800666689873, + 0.00781202083453536, + 0.0035027097910642624, + 0.003977222368121147, + 0.004078366793692112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0381341353058815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419412437826395 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.01419412437826395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419412437826395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42248865962028503, + 0.08996867388486862, + 0.07979276776313782, + 0.048972249031066895, + 0.026816019788384438, + 0.06830412894487381, + 0.10100812464952469, + 0.02269248478114605, + 0.050957415252923965, + 0.08899945020675659 + ], + "credit_map": [ + 0.10405765473842621, + 0.030953938141465187, + 0.14577095210552216, + 0.04013212025165558, + 0.027019450441002846, + 0.021827131509780884, + 0.020862290635704994, + 0.03186512738466263, + 0.011191164143383503, + 0.011502702720463276, + 0.011452111415565014, + 0.008197157643735409, + 0.012739731930196285, + 0.010506205260753632, + 0.013027758337557316, + 0.011351845227181911, + 0.012113801203668118, + 0.011060752905905247, + 0.010345323011279106, + 0.008292244747281075, + 0.009395367465913296, + 0.008220558986067772, + 0.010825017467141151, + 0.009539701044559479, + 0.007081401534378529, + 0.005415703170001507, + 0.00629834458231926, + 0.00583604397252202, + 0.004312489647418261, + 0.005616722162812948, + 0.007941321469843388, + 0.00647022295743227, + 0.003916345536708832, + 0.003478498663753271, + 0.002462264383211732, + 0.003909544087946415, + 0.002221856964752078, + 0.003361406968906522, + 0.004586513619869947, + 0.0028795883990824223, + 0.0060770572163164616, + 0.011821622960269451, + 0.005863923579454422, + 0.007476225960999727, + 0.011371498927474022, + 0.010531012900173664, + 0.006847489159554243, + 0.008315293118357658, + 0.009456999599933624, + 0.012993460521101952, + 0.010274281725287437, + 0.01871900074183941, + 0.011408247984945774, + 0.014066589996218681, + 0.009502488188445568, + 0.014587054960429668, + 0.0034827592317014933, + 0.003941857255995274, + 0.0023603530135005713, + 0.0035052532330155373, + 0.001890667132101953, + 0.0030996426939964294, + 0.0018059557769447565, + 0.0026059956289827824, + 0.005957637447863817, + 0.005210416857153177, + 0.005447593983262777, + 0.006397810764610767, + 0.007001122925430536, + 0.004720455501228571, + 0.011292785406112671, + 0.00492959376424551, + 0.013927371241152287, + 0.011207579635083675, + 0.010890426114201546, + 0.009524291381239891, + 0.01477746106684208, + 0.006971634924411774, + 0.01167790126055479, + 0.010022781789302826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.01419412437826395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632218275219202 + ], + "opposite_logprob_tokens": [ + -5.379631996154785 + ], + "logprob_reward": -0.004632218275219202, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632218275219202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2754906117916107, + 0.040524113923311234, + 0.017374645918607712, + 0.024593057110905647, + 0.11286164820194244, + 0.0609915629029274, + 0.041925445199012756, + 0.012514396570622921, + 0.03642122820019722, + 0.37730321288108826 + ], + "credit_map": [ + 0.0680636316537857, + 0.017898812890052795, + 0.09964539855718613, + 0.025344181805849075, + 0.014200953766703606, + 0.014827411621809006, + 0.011877721175551414, + 0.023632517084479332, + 0.005254719406366348, + 0.00556095689535141, + 0.004426824394613504, + 0.0073710475116968155, + 0.00418626656755805, + 0.0036804459523409605, + 0.003293198999017477, + 0.006750654429197311, + 0.0030794136691838503, + 0.002360061276704073, + 0.0023766066879034042, + 0.0017630730289965868, + 0.002248989650979638, + 0.0012964095221832395, + 0.0018058035057038069, + 0.0024442896246910095, + 0.0029755455907434225, + 0.003976989071816206, + 0.002279650652781129, + 0.0032305435743182898, + 0.003477976191788912, + 0.003139219945296645, + 0.0029979317914694548, + 0.0025152002926915884, + 0.01567421853542328, + 0.008937598206102848, + 0.01552418153733015, + 0.028371958062052727, + 0.008754411712288857, + 0.012280530296266079, + 0.01138535887002945, + 0.011933397501707077, + 0.012735647149384022, + 0.007233469281345606, + 0.005862942431122065, + 0.011854542419314384, + 0.0050907679833471775, + 0.007285048719495535, + 0.00574574526399374, + 0.00518339779227972, + 0.005328495986759663, + 0.0068067689426243305, + 0.0051873959600925446, + 0.009693493135273457, + 0.002712516812607646, + 0.0038069654256105423, + 0.003247126704081893, + 0.005142685025930405, + 0.001531484886072576, + 0.0023071656469255686, + 0.0013513389276340604, + 0.0016020243056118488, + 0.0012464639730751514, + 0.0012086895294487476, + 0.0015429813647642732, + 0.0017242475878447294, + 0.005165225360542536, + 0.006097402423620224, + 0.004534224513918161, + 0.004125634208321571, + 0.004104858264327049, + 0.004496419802308083, + 0.00482140202075243, + 0.00307606253772974, + 0.0637051984667778, + 0.04932740703225136, + 0.03700438141822815, + 0.05883077532052994, + 0.04078309237957001, + 0.048516642302274704, + 0.03811100870370865, + 0.04102472960948944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.004632218275219202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279571905732155 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279571905732155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279571905732155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6825686693191528, + 0.026313111186027527, + 0.04278387129306793, + 0.051398999989032745, + 0.02236163057386875, + 0.03267316520214081, + 0.02224397473037243, + 0.01656615175306797, + 0.08123371750116348, + 0.021856730803847313 + ], + "credit_map": [ + 0.18245211243629456, + 0.04151560738682747, + 0.26523557305336, + 0.06097373366355896, + 0.02516445331275463, + 0.030551092699170113, + 0.025878801941871643, + 0.05079735442996025, + 0.003756727324798703, + 0.005360591225326061, + 0.002723604440689087, + 0.004373181611299515, + 0.0023476940114051104, + 0.002213255735114217, + 0.0024462712462991476, + 0.003091789549216628, + 0.004749803803861141, + 0.006266390439122915, + 0.005191402044147253, + 0.008359895087778568, + 0.004723345395177603, + 0.0033949746284633875, + 0.004744858015328646, + 0.005353206768631935, + 0.010323876515030861, + 0.006007408257573843, + 0.008868888951838017, + 0.005562352482229471, + 0.004581779707223177, + 0.0036362321116030216, + 0.004980926867574453, + 0.0074375406838953495, + 0.0032840941566973925, + 0.002298397943377495, + 0.001792224356904626, + 0.00375583884306252, + 0.0019986184779554605, + 0.0032039121724665165, + 0.0034929164685308933, + 0.002535631414502859, + 0.007826036773622036, + 0.0036995382979512215, + 0.004198746290057898, + 0.005647207610309124, + 0.0032790289260447025, + 0.0025127152912318707, + 0.0025713047944009304, + 0.002938590245321393, + 0.002381508005782962, + 0.0035123147536069155, + 0.0023659435100853443, + 0.0037812544032931328, + 0.0019565022084861994, + 0.003121985588222742, + 0.0022730454802513123, + 0.0028514228761196136, + 0.002310734009370208, + 0.004391566384583712, + 0.0015543459448963404, + 0.0026695283595472574, + 0.0013231881894171238, + 0.0013989618746563792, + 0.0013905424857512116, + 0.0015272871823981404, + 0.01137552596628666, + 0.008505374193191528, + 0.008780238218605518, + 0.009080165065824986, + 0.00780204264447093, + 0.009217976592481136, + 0.013306518085300922, + 0.013165885582566261, + 0.0036783884279429913, + 0.0029458303470164537, + 0.0022075334563851357, + 0.00301074655726552, + 0.003335036337375641, + 0.0019366618944332004, + 0.0023764721117913723, + 0.0023660631850361824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.023279571905732155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows an ornate carriage as it speeds along a dry, dirt country road. The carriage, with its crimson lacquer and intricately carved gold trim, gleams luxuriously in the evening sun. Four powerful, glossy, dark brown horses gallop with gusto, their muscles straining, the brass fittings of their leather harnesses swaying with their movements. The driver, dressed in a dark uniform and a top hat, maintains a focused and resolute expression as he grips the reins, his body swaying with the carriage's violent jolting. Massive wooden wheels, rimmed with iron hoops, spin at high speed, rolling over the ground, flinging loose dirt and gravel backward and raising a thick, tawny cloud of dust. This cloud, pierced by golden sunlight and shimmering with halos at the edges, trails behind the carriage like a magnificent tail, nearly engulfing the green hills and winding road in the distance. The whole picture presents a tense atmosphere full of dramatic tension, speed and retro luxury in the warm light and shadow.", + "seed": 30626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_2566_seed_30626_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021779794339090587, + "vlm_logprob": -0.021779794339090587, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.021779794339090587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009351144544780254 + ], + "opposite_logprob_tokens": [ + -4.8843512535095215 + ], + "logprob_reward": -0.009351144544780254, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009351144544780254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07593093812465668, + 0.057806044816970825, + 0.20467506349086761, + 0.06870103627443314, + 0.06689316779375076, + 0.10978010296821594, + 0.03982090204954147, + 0.04710223153233528, + 0.24880868196487427, + 0.08048183470964432 + ], + "credit_map": [ + 0.011428113095462322, + 0.011083506979048252, + 0.006355496123433113, + 0.009138892404735088, + 0.01182219572365284, + 0.011085701175034046, + 0.007793100085109472, + 0.007223938126116991, + 0.007847386412322521, + 0.0074489121325314045, + 0.0046072122640907764, + 0.007626235019415617, + 0.010947335511446, + 0.005130428355187178, + 0.005879123695194721, + 0.008319414220750332, + 0.032360441982746124, + 0.02200501412153244, + 0.028701480478048325, + 0.020536702126264572, + 0.02297508344054222, + 0.024417392909526825, + 0.01997731626033783, + 0.03370163217186928, + 0.01156808715313673, + 0.012271245941519737, + 0.005191870033740997, + 0.008350570686161518, + 0.009367874823510647, + 0.008806459605693817, + 0.005778710823506117, + 0.007366221398115158, + 0.006931544281542301, + 0.00866011530160904, + 0.008324653841555119, + 0.007003371603786945, + 0.007007841020822525, + 0.009255334734916687, + 0.010948694311082363, + 0.008761617355048656, + 0.009038540534675121, + 0.01294375117868185, + 0.012637277133762836, + 0.014346042647957802, + 0.014904683455824852, + 0.015168750658631325, + 0.01935068517923355, + 0.011390391737222672, + 0.004758229944854975, + 0.006097837816923857, + 0.004623912740498781, + 0.006836507003754377, + 0.005748698953539133, + 0.0045999386347830296, + 0.0035558249801397324, + 0.0035999577958136797, + 0.005861825309693813, + 0.007056600879877806, + 0.005885469727218151, + 0.005477529019117355, + 0.0035917602945119143, + 0.005790350493043661, + 0.006126688327640295, + 0.007312009111046791, + 0.02938411571085453, + 0.030645539984107018, + 0.027806278318166733, + 0.02888473868370056, + 0.04354598745703697, + 0.0219698715955019, + 0.03715699911117554, + 0.02941516600549221, + 0.010063880123198032, + 0.010057126171886921, + 0.008031969889998436, + 0.010157632641494274, + 0.012332018464803696, + 0.012495617382228374, + 0.00958334095776081, + 0.007760254666209221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.009351144544780254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208749860525131 + ], + "opposite_logprob_tokens": [ + -2.8120875358581543 + ], + "logprob_reward": -0.06208749860525131, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208749860525131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10460736602544785, + 0.08900229632854462, + 0.10092002898454666, + 0.040883634239435196, + 0.07798238843679428, + 0.22768355906009674, + 0.04881095141172409, + 0.09451806545257568, + 0.16862471401691437, + 0.04696701094508171 + ], + "credit_map": [ + 0.017433101311326027, + 0.011764769442379475, + 0.010252423584461212, + 0.015500200912356377, + 0.012071726843714714, + 0.011750735342502594, + 0.012139943428337574, + 0.013694467954337597, + 0.010795710608363152, + 0.022870589047670364, + 0.006983489263802767, + 0.012976541183888912, + 0.015465316362679005, + 0.006387163884937763, + 0.007102188188582659, + 0.006421296391636133, + 0.0127057284116745, + 0.01424532663077116, + 0.011239323765039444, + 0.011480111628770828, + 0.011600492522120476, + 0.014030499383807182, + 0.01622953452169895, + 0.009389016777276993, + 0.004523369017988443, + 0.0059007504023611546, + 0.0040892064571380615, + 0.006341025233268738, + 0.003935621120035648, + 0.005704519804567099, + 0.004600054584443569, + 0.005789089947938919, + 0.009967231191694736, + 0.01067185215651989, + 0.009819479659199715, + 0.011365880258381367, + 0.007112474180757999, + 0.010282027535140514, + 0.01085139624774456, + 0.00791204534471035, + 0.02986084669828415, + 0.03359958156943321, + 0.020097319036722183, + 0.02941148914396763, + 0.021677745506167412, + 0.03120611235499382, + 0.03324766829609871, + 0.028582783415913582, + 0.006592077203094959, + 0.005952500272542238, + 0.005596599541604519, + 0.0070985122583806515, + 0.007280948106199503, + 0.00599781796336174, + 0.006016104482114315, + 0.004276392515748739, + 0.012473099865019321, + 0.008095599710941315, + 0.009372859261929989, + 0.011127324774861336, + 0.016458524391055107, + 0.012965630739927292, + 0.013510564342141151, + 0.010514459572732449, + 0.023066610097885132, + 0.021224305033683777, + 0.016942854970693588, + 0.01708289235830307, + 0.03828295320272446, + 0.01562899351119995, + 0.01804758608341217, + 0.018348505720496178, + 0.007608263753354549, + 0.005685040261596441, + 0.004697892814874649, + 0.005531972274184227, + 0.006312299054116011, + 0.006887021008878946, + 0.004967878106981516, + 0.005276643205434084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.06208749860525131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16642440855503082, + 0.12535952031612396, + 0.21322764456272125, + 0.05578090623021126, + 0.1053442433476448, + 0.07822269201278687, + 0.06563761830329895, + 0.05390520766377449, + 0.08425857871770859, + 0.05183914676308632 + ], + "credit_map": [ + 0.024538662284612656, + 0.020473040640354156, + 0.01810953952372074, + 0.01700834557414055, + 0.02617115154862404, + 0.029945941641926765, + 0.015476435422897339, + 0.014701319858431816, + 0.021000873297452927, + 0.01693309284746647, + 0.013590425252914429, + 0.019227972254157066, + 0.014080105349421501, + 0.013425100594758987, + 0.011850389651954174, + 0.015251562930643559, + 0.03125164285302162, + 0.021430976688861847, + 0.039151035249233246, + 0.016904789954423904, + 0.024071386083960533, + 0.028356818482279778, + 0.01974646933376789, + 0.03231452405452728, + 0.006887656170874834, + 0.0071839443407952785, + 0.005609963554888964, + 0.007427467033267021, + 0.0059848506934940815, + 0.008960309438407421, + 0.006114444229751825, + 0.007612274494022131, + 0.016567161306738853, + 0.01733754575252533, + 0.013587084598839283, + 0.011091976426541805, + 0.010644429363310337, + 0.013393069617450237, + 0.010937604121863842, + 0.011785371229052544, + 0.00874092523008585, + 0.010250106453895569, + 0.0066081807017326355, + 0.009125562384724617, + 0.012552413158118725, + 0.010903491638600826, + 0.012112166732549667, + 0.007929842919111252, + 0.01259948592633009, + 0.005889628082513809, + 0.0141123216599226, + 0.0060079945251345634, + 0.008246519602835178, + 0.007003479637205601, + 0.006977541372179985, + 0.004800654016435146, + 0.008549222722649574, + 0.0035516279749572277, + 0.012367716059088707, + 0.0031401074957102537, + 0.0038870961870998144, + 0.006422348320484161, + 0.010723277926445007, + 0.005263813771307468, + 0.010084404610097408, + 0.009360283613204956, + 0.009175866842269897, + 0.009089800529181957, + 0.012300500646233559, + 0.010128824971616268, + 0.009643584489822388, + 0.014475325122475624, + 0.007515497971326113, + 0.0071685281582176685, + 0.005768504925072193, + 0.006228643003851175, + 0.007521511986851692, + 0.006708264816552401, + 0.0058942814357578754, + 0.00503391632810235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.008636740036308765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024855216033756733 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024855216033756733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024855216033756733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06499450653791428, + 0.03846001625061035, + 0.25464338064193726, + 0.02626785822212696, + 0.02579396404325962, + 0.09746440500020981, + 0.03171331062912941, + 0.05928971990942955, + 0.27361392974853516, + 0.12775886058807373 + ], + "credit_map": [ + 0.008868066594004631, + 0.00702567957341671, + 0.00854800920933485, + 0.008787916041910648, + 0.007662064861506224, + 0.008426221087574959, + 0.00677704531699419, + 0.008899501524865627, + 0.004558126907795668, + 0.004700914025306702, + 0.00464228168129921, + 0.0044065504334867, + 0.005900577642023563, + 0.0045394254848361015, + 0.0031556671019643545, + 0.006556474603712559, + 0.03802337870001793, + 0.023580120876431465, + 0.04384690150618553, + 0.021888181567192078, + 0.027909154072403908, + 0.037786368280649185, + 0.023694928735494614, + 0.03791436553001404, + 0.0035783417988568544, + 0.0039946939796209335, + 0.003152189077809453, + 0.0030859902035444975, + 0.004085605498403311, + 0.00289716268889606, + 0.0030100522562861443, + 0.00246382225304842, + 0.0029303831979632378, + 0.0032248105853796005, + 0.0021216084714978933, + 0.002875400008633733, + 0.003225971246138215, + 0.002859894884750247, + 0.00363385071977973, + 0.004922044463455677, + 0.010308794677257538, + 0.014751543290913105, + 0.013520986773073673, + 0.01818058453500271, + 0.008760343305766582, + 0.011498669162392616, + 0.010229235514998436, + 0.010214253328740597, + 0.005210089031606913, + 0.0035231097135692835, + 0.006200122646987438, + 0.003539451863616705, + 0.002969884779304266, + 0.0040312944911420345, + 0.003331948071718216, + 0.002907411428168416, + 0.00732273468747735, + 0.0043679880909621716, + 0.006290813907980919, + 0.0060428716242313385, + 0.008922019973397255, + 0.006942347157746553, + 0.013610794208943844, + 0.005790155380964279, + 0.032797593623399734, + 0.034649938344955444, + 0.025614816695451736, + 0.028635557740926743, + 0.051578011363744736, + 0.029823726043105125, + 0.028098348528146744, + 0.04241592437028885, + 0.04205983132123947, + 0.014173075556755066, + 0.01606743037700653, + 0.010912630707025528, + 0.00943222176283598, + 0.011445945128798485, + 0.012697525322437286, + 0.010970198549330235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.0024855216033756733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338066905736923 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338066905736923, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338066905736923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07856150716543198, + 0.05974387377500534, + 0.09296447038650513, + 0.08899489045143127, + 0.05200473964214325, + 0.12628132104873657, + 0.06611990928649902, + 0.07284101098775864, + 0.2096697837114334, + 0.15281856060028076 + ], + "credit_map": [ + 0.010768435895442963, + 0.008166095241904259, + 0.011327937245368958, + 0.008725945837795734, + 0.007696791552007198, + 0.009143741801381111, + 0.0079953046515584, + 0.014737256802618504, + 0.009854786098003387, + 0.00852189864963293, + 0.005281072109937668, + 0.0062613957561552525, + 0.008281504735350609, + 0.005840284284204245, + 0.005608818959444761, + 0.01009411457926035, + 0.013799346052110195, + 0.008844435214996338, + 0.017327336594462395, + 0.008189873769879341, + 0.008982325904071331, + 0.015418834052979946, + 0.008816699497401714, + 0.01158560998737812, + 0.018642570823431015, + 0.013262697495520115, + 0.008368968963623047, + 0.008139136247336864, + 0.011298288591206074, + 0.011962261982262135, + 0.007418655790388584, + 0.009902308695018291, + 0.0065251546911895275, + 0.007263280916959047, + 0.006273512262851, + 0.006365620996803045, + 0.005340387113392353, + 0.007651281077414751, + 0.0065904418006539345, + 0.005995059385895729, + 0.019898610189557076, + 0.015430254861712456, + 0.02001063898205757, + 0.019741270691156387, + 0.013092704117298126, + 0.0122694605961442, + 0.015053716488182545, + 0.010784655809402466, + 0.011090612970292568, + 0.006649432238191366, + 0.01550882589071989, + 0.005785245448350906, + 0.006891921628266573, + 0.007319674827158451, + 0.007411487400531769, + 0.0054627046920359135, + 0.009173477999866009, + 0.007325991056859493, + 0.007393420208245516, + 0.008961398154497147, + 0.009278623387217522, + 0.007107152603566647, + 0.016822170466184616, + 0.006778782699257135, + 0.032048776745796204, + 0.02683892846107483, + 0.024033980444073677, + 0.023723255842924118, + 0.03371299058198929, + 0.019588805735111237, + 0.022789133712649345, + 0.02693389542400837, + 0.038557540625333786, + 0.020322497934103012, + 0.014691788703203201, + 0.01745339296758175, + 0.014817725867033005, + 0.015738721936941147, + 0.013267621397972107, + 0.017969265580177307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.026338066905736923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows an ornate carriage as it speeds along a dry, dirt country road. The carriage, with its crimson lacquer and intricately carved gold trim, gleams luxuriously in the evening sun. Four powerful, glossy, dark brown horses gallop with gusto, their muscles straining, the brass fittings of their leather harnesses swaying with their movements. The driver, dressed in a dark uniform and a top hat, maintains a focused and resolute expression as he grips the reins, his body swaying with the carriage's violent jolting. Massive wooden wheels, rimmed with iron hoops, spin at high speed, rolling over the ground, flinging loose dirt and gravel backward and raising a thick, tawny cloud of dust. This cloud, pierced by golden sunlight and shimmering with halos at the edges, trails behind the carriage like a magnificent tail, nearly engulfing the green hills and winding road in the distance. The whole picture presents a tense atmosphere full of dramatic tension, speed and retro luxury in the warm light and shadow.", + "seed": 30627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_2566_seed_30627_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0392851484939456, + "vlm_logprob": -0.0392851484939456, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.0392851484939456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0416681170463562 + ], + "opposite_logprob_tokens": [ + -3.291668176651001 + ], + "logprob_reward": -0.0416681170463562, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0416681170463562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0960359126329422, + 0.016466772183775902, + 0.029790783300995827, + 0.039724431931972504, + 0.04270188882946968, + 0.2713991403579712, + 0.09396330267190933, + 0.12466798722743988, + 0.12875410914421082, + 0.15649572014808655 + ], + "credit_map": [ + 0.011141243390738964, + 0.017662586644291878, + 0.010235676541924477, + 0.020119640976190567, + 0.008484412916004658, + 0.012550621293485165, + 0.009057503193616867, + 0.006784226279705763, + 0.0023507755249738693, + 0.0018230549758300185, + 0.003513215808197856, + 0.0028329186607152224, + 0.001638737041503191, + 0.0015647521940991282, + 0.0015541667817160487, + 0.001189148984849453, + 0.004147037863731384, + 0.003050633240491152, + 0.004371193237602711, + 0.004616221878677607, + 0.003079396905377507, + 0.003999626729637384, + 0.0042747510597109795, + 0.002251923084259033, + 0.010217959992587566, + 0.003597862320020795, + 0.013049688190221786, + 0.004241813439875841, + 0.0014811750734224916, + 0.0025393490213900805, + 0.0023668084759265184, + 0.0022297734394669533, + 0.00516369566321373, + 0.0062874965369701385, + 0.0069299847818911076, + 0.00553051196038723, + 0.003202639054507017, + 0.004276852589100599, + 0.0043906341306865215, + 0.006920068524777889, + 0.08258571475744247, + 0.024329345673322678, + 0.10565996170043945, + 0.018501030281186104, + 0.007443145848810673, + 0.01438072882592678, + 0.008620291948318481, + 0.009878890588879585, + 0.008926601149141788, + 0.004010121338069439, + 0.02843473106622696, + 0.006233065389096737, + 0.0067500164732337, + 0.004953607451170683, + 0.030497528612613678, + 0.004157618619501591, + 0.015298969112336636, + 0.019339323043823242, + 0.01991991512477398, + 0.021903667598962784, + 0.011067425832152367, + 0.011189619079232216, + 0.015428404323756695, + 0.010520653799176216, + 0.013952992856502533, + 0.0069154598750174046, + 0.0518859438598156, + 0.008814692497253418, + 0.021135471761226654, + 0.008200488053262234, + 0.006618923973292112, + 0.011230128817260265, + 0.011774553917348385, + 0.010417669080197811, + 0.01384864654392004, + 0.013040353544056416, + 0.03651178255677223, + 0.04235789179801941, + 0.010968022048473358, + 0.017576787620782852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.0416681170463562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034148395061493 + ], + "opposite_logprob_tokens": [ + -2.350341558456421 + ], + "logprob_reward": -0.10034148395061493, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034148395061493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19223934412002563, + 0.023396261036396027, + 0.08765417337417603, + 0.15822817385196686, + 0.01530649233609438, + 0.043138306587934494, + 0.05563480779528618, + 0.143076553940773, + 0.04862477630376816, + 0.23270107805728912 + ], + "credit_map": [ + 0.020392438396811485, + 0.021570228040218353, + 0.022541489452123642, + 0.03321199119091034, + 0.02047414891421795, + 0.026832928881049156, + 0.028281357139348984, + 0.018934758380055428, + 0.0034014969132840633, + 0.0018994087586179376, + 0.003979381173849106, + 0.002350491937249899, + 0.0026598209515213966, + 0.002736515598371625, + 0.004087052308022976, + 0.0022820942103862762, + 0.012158192694187164, + 0.009571805596351624, + 0.011789415962994099, + 0.012106259353458881, + 0.009424556978046894, + 0.011063458397984505, + 0.011329968459904194, + 0.010210521519184113, + 0.04413386806845665, + 0.010132761672139168, + 0.06039167940616608, + 0.019174011424183846, + 0.004556282889097929, + 0.006036561448127031, + 0.006417479366064072, + 0.007385523524135351, + 0.0016636236105114222, + 0.0030376920476555824, + 0.0019817582797259092, + 0.0023283613845705986, + 0.0013762633316218853, + 0.0016198783414438367, + 0.0016402400797232985, + 0.0016586760757490993, + 0.009339804761111736, + 0.0068781087175011635, + 0.011976267211139202, + 0.003522661281749606, + 0.003056812798604369, + 0.0031282075215131044, + 0.003045578021556139, + 0.0021908662747591734, + 0.004866383504122496, + 0.0029283242765814066, + 0.014181472361087799, + 0.003661846276372671, + 0.0042756423354148865, + 0.004159854259341955, + 0.018262526020407677, + 0.0032987610902637243, + 0.01893235370516777, + 0.021144790574908257, + 0.018752513453364372, + 0.020869556814432144, + 0.013425478711724281, + 0.014429549686610699, + 0.02173927053809166, + 0.013783041387796402, + 0.007950672879815102, + 0.002869836986064911, + 0.017309047281742096, + 0.0032218273263424635, + 0.006079650949686766, + 0.003167032264173031, + 0.0033314719330519438, + 0.004695233888924122, + 0.02957121469080448, + 0.018781878054142, + 0.025188475847244263, + 0.02575022168457508, + 0.02773766964673996, + 0.041892994195222855, + 0.021953195333480835, + 0.041825439780950546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.10034148395061493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735360249876976 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735360249876976, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735360249876976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1008419319987297, + 0.06071893125772476, + 0.05405315011739731, + 0.26365113258361816, + 0.03001043200492859, + 0.041095901280641556, + 0.15059907734394073, + 0.027628062292933464, + 0.05580378323793411, + 0.21559762954711914 + ], + "credit_map": [ + 0.013567037880420685, + 0.013015374541282654, + 0.01130707748234272, + 0.01746451109647751, + 0.011575710028409958, + 0.01078676525503397, + 0.013558823615312576, + 0.009566632099449635, + 0.007587497588247061, + 0.006116203032433987, + 0.006100492086261511, + 0.00607770262286067, + 0.01040625385940075, + 0.008435051888227463, + 0.007968274876475334, + 0.008027455769479275, + 0.0067137787118554115, + 0.00517016276717186, + 0.007027107290923595, + 0.008356643840670586, + 0.005961477756500244, + 0.006760971155017614, + 0.009374520741403103, + 0.004688485525548458, + 0.07188232988119125, + 0.0191334281116724, + 0.09947266429662704, + 0.03098534792661667, + 0.007841850630939007, + 0.009310254827141762, + 0.011580967344343662, + 0.013444274663925171, + 0.00407183263450861, + 0.003820652374997735, + 0.00447485875338316, + 0.0038880801293998957, + 0.0027665700763463974, + 0.002895499812439084, + 0.0031040420290082693, + 0.004988895729184151, + 0.007159792818129063, + 0.006669963244348764, + 0.008248697966337204, + 0.0034399942960590124, + 0.004212712869048119, + 0.003612406086176634, + 0.0048917862586677074, + 0.002860547509044409, + 0.01463763415813446, + 0.008463053032755852, + 0.045898035168647766, + 0.009966891258955002, + 0.010495977476239204, + 0.008295968174934387, + 0.044493574649095535, + 0.008347934111952782, + 0.0029297824949026108, + 0.00558391772210598, + 0.005248873960226774, + 0.0038727798964828253, + 0.0020629619248211384, + 0.002611848758533597, + 0.0023739286698400974, + 0.0029439679346978664, + 0.009891916066408157, + 0.003329227678477764, + 0.013409378007054329, + 0.0032540629617869854, + 0.008522397838532925, + 0.0050824107602238655, + 0.0038955463096499443, + 0.008418848738074303, + 0.01653837226331234, + 0.01720309630036354, + 0.01631816104054451, + 0.018214406445622444, + 0.05110366269946098, + 0.05201824754476547, + 0.016916368156671524, + 0.027285313233733177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.006735360249876976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632455296814442 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632455296814442, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632455296814442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11206810176372528, + 0.04226434603333473, + 0.09013734012842178, + 0.3832240700721741, + 0.01994321122765541, + 0.09937219321727753, + 0.0556141659617424, + 0.049026671797037125, + 0.05112529546022415, + 0.09722462296485901 + ], + "credit_map": [ + 0.01330642867833376, + 0.011755022220313549, + 0.021501345559954643, + 0.018878629431128502, + 0.009730621241033077, + 0.007847706787288189, + 0.015691271051764488, + 0.01335707027465105, + 0.005158407147973776, + 0.003998640459030867, + 0.012635565362870693, + 0.00501759210601449, + 0.002707583364099264, + 0.003058497328311205, + 0.004824149422347546, + 0.004863908048719168, + 0.01155603863298893, + 0.006497794762253761, + 0.031076056882739067, + 0.012242989614605904, + 0.005005525890737772, + 0.008811472915112972, + 0.008243885822594166, + 0.006703573744744062, + 0.10548702627420425, + 0.025926407426595688, + 0.14147494733333588, + 0.04665178805589676, + 0.012830697931349277, + 0.018947888165712357, + 0.01705387607216835, + 0.014851425774395466, + 0.002479734132066369, + 0.002834303304553032, + 0.002165518468245864, + 0.002154043409973383, + 0.0018802465638145804, + 0.002178530441597104, + 0.002251656958833337, + 0.003999179229140282, + 0.03208257630467415, + 0.006948239170014858, + 0.039972979575395584, + 0.006610168144106865, + 0.0024299602955579758, + 0.004993760492652655, + 0.002911538118496537, + 0.003422970650717616, + 0.0054014455527067184, + 0.0030337637290358543, + 0.016829920932650566, + 0.00380670721642673, + 0.003977675922214985, + 0.0030026885215193033, + 0.01714383065700531, + 0.0024181343615055084, + 0.005119571462273598, + 0.009738009423017502, + 0.009364644065499306, + 0.009037290699779987, + 0.00377183617092669, + 0.004111493472009897, + 0.004885873291641474, + 0.0029979527462273836, + 0.00819811038672924, + 0.003313525579869747, + 0.021631520241498947, + 0.0041197966784238815, + 0.0035293635446578264, + 0.003075286513194442, + 0.002273715101182461, + 0.004983978345990181, + 0.009119164198637009, + 0.008087235502898693, + 0.011595616117119789, + 0.011397276073694229, + 0.011903764680027962, + 0.018828123807907104, + 0.011091588996350765, + 0.015201855450868607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.004632455296814442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043048325926065445 + ], + "opposite_logprob_tokens": [ + -3.16804838180542 + ], + "logprob_reward": -0.043048325926065445, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043048325926065445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18210899829864502, + 0.06969090551137924, + 0.08243224024772644, + 0.0646466538310051, + 0.058263570070266724, + 0.10517138242721558, + 0.12000954151153564, + 0.10590087622404099, + 0.11018606275320053, + 0.10158976912498474 + ], + "credit_map": [ + 0.022430313751101494, + 0.022334959357976913, + 0.03304571285843849, + 0.029108654707670212, + 0.016657929867506027, + 0.017542077228426933, + 0.023956604301929474, + 0.017032738775014877, + 0.0075623090378940105, + 0.007403142284601927, + 0.010700954124331474, + 0.008229865692555904, + 0.006448089610785246, + 0.006503026466816664, + 0.015319667756557465, + 0.007523853797465563, + 0.011480879038572311, + 0.011685042642056942, + 0.00982320960611105, + 0.010992376133799553, + 0.008826230652630329, + 0.010996325872838497, + 0.00923252198845148, + 0.00939564872533083, + 0.013611526228487492, + 0.007040694355964661, + 0.016001764684915543, + 0.007811977062374353, + 0.003381136804819107, + 0.006377227138727903, + 0.004813605919480324, + 0.005608721170574427, + 0.0061953081749379635, + 0.005534018389880657, + 0.009271450340747833, + 0.0065209767781198025, + 0.004932769574224949, + 0.006810183171182871, + 0.00591347785666585, + 0.013085387647151947, + 0.029016561806201935, + 0.013208942487835884, + 0.03561047837138176, + 0.0076536750420928, + 0.004816771484911442, + 0.005869319662451744, + 0.005006803665310144, + 0.003988825716078281, + 0.009965348057448864, + 0.007668682839721441, + 0.02162248268723488, + 0.014873498119413853, + 0.010139117948710918, + 0.0097024105489254, + 0.03805512562394142, + 0.00798287708312273, + 0.015374956652522087, + 0.013149957172572613, + 0.014158220030367374, + 0.014207379892468452, + 0.010668361559510231, + 0.009450864978134632, + 0.017874937504529953, + 0.011016201227903366, + 0.012738693505525589, + 0.007100366055965424, + 0.034768685698509216, + 0.006337778177112341, + 0.014381282962858677, + 0.0066909571178257465, + 0.006266322452574968, + 0.021901976317167282, + 0.011096290312707424, + 0.007040878292173147, + 0.009261644445359707, + 0.008671804331243038, + 0.020027633756399155, + 0.0228381659835577, + 0.007524505723267794, + 0.0151288453489542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.043048325926065445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, filmed from a low angle, follows an ornate carriage as it speeds along a dry, dirt country road. The carriage, with its crimson lacquer and intricately carved gold trim, gleams luxuriously in the evening sun. Four powerful, glossy, dark brown horses gallop with gusto, their muscles straining, the brass fittings of their leather harnesses swaying with their movements. The driver, dressed in a dark uniform and a top hat, maintains a focused and resolute expression as he grips the reins, his body swaying with the carriage's violent jolting. Massive wooden wheels, rimmed with iron hoops, spin at high speed, rolling over the ground, flinging loose dirt and gravel backward and raising a thick, tawny cloud of dust. This cloud, pierced by golden sunlight and shimmering with halos at the edges, trails behind the carriage like a magnificent tail, nearly engulfing the green hills and winding road in the distance. The whole picture presents a tense atmosphere full of dramatic tension, speed and retro luxury in the warm light and shadow.", + "seed": 30628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_8/index_2566_seed_30628_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033446802641265094, + "vlm_logprob": -0.033446802641265094, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.033446802641265094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03729094937443733 + ], + "opposite_logprob_tokens": [ + -3.4122910499572754 + ], + "logprob_reward": -0.03729094937443733, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03729094937443733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07391508668661118, + 0.09780590236186981, + 0.06880397349596024, + 0.09750717878341675, + 0.11097294837236404, + 0.12379765510559082, + 0.09480968117713928, + 0.09474404156208038, + 0.09319782257080078, + 0.14444570243358612 + ], + "credit_map": [ + 0.01005527749657631, + 0.00912606529891491, + 0.008703015744686127, + 0.00809290912002325, + 0.012332547456026077, + 0.008511359803378582, + 0.009942661970853806, + 0.007151245605200529, + 0.015994591638445854, + 0.011854364536702633, + 0.008831758052110672, + 0.018005408346652985, + 0.010166429914534092, + 0.013408327475190163, + 0.011008227244019508, + 0.008536800742149353, + 0.009637124836444855, + 0.006246722303330898, + 0.008645298890769482, + 0.008990637958049774, + 0.007865608669817448, + 0.0103010144084692, + 0.008449100889265537, + 0.008668472059071064, + 0.01672798953950405, + 0.009678357280790806, + 0.01260139886289835, + 0.011818776838481426, + 0.009980808012187481, + 0.010855292901396751, + 0.013785252347588539, + 0.01205930020660162, + 0.017467882484197617, + 0.013583619147539139, + 0.012341956607997417, + 0.014099831692874432, + 0.019944030791521072, + 0.009294201619923115, + 0.011133750900626183, + 0.013107672333717346, + 0.013439853675663471, + 0.010069811716675758, + 0.012494749389588833, + 0.02213345840573311, + 0.013008086010813713, + 0.014302657917141914, + 0.024594057351350784, + 0.01375497505068779, + 0.011413181200623512, + 0.013881531544029713, + 0.011901337653398514, + 0.013791861943900585, + 0.010844522155821323, + 0.009929426945745945, + 0.010974347591400146, + 0.01207346748560667, + 0.008483115583658218, + 0.012382840737700462, + 0.018438832834362984, + 0.010106795467436314, + 0.011738847941160202, + 0.01145097054541111, + 0.012852690182626247, + 0.0092899389564991, + 0.007602663245052099, + 0.007469144184142351, + 0.019467907026410103, + 0.02537996508181095, + 0.006061876192688942, + 0.008060431107878685, + 0.008950811810791492, + 0.010205022059381008, + 0.008154327981173992, + 0.009731083177030087, + 0.02130635641515255, + 0.02491041086614132, + 0.011894777417182922, + 0.026927525177598, + 0.017526034265756607, + 0.02399520017206669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.03729094937443733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907544612884521 + ], + "opposite_logprob_tokens": [ + -2.4640755653381348 + ], + "logprob_reward": -0.08907544612884521, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907544612884521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15165461599826813, + 0.12007153779268265, + 0.18835939466953278, + 0.04321267455816269, + 0.024392202496528625, + 0.02972850762307644, + 0.07613198459148407, + 0.16346633434295654, + 0.18243783712387085, + 0.02054492011666298 + ], + "credit_map": [ + 0.015138049609959126, + 0.015478788875043392, + 0.023639867082238197, + 0.0217139795422554, + 0.019038666039705276, + 0.016250411048531532, + 0.01888871006667614, + 0.021506140008568764, + 0.019189754500985146, + 0.012028210796415806, + 0.017591947689652443, + 0.021463651210069656, + 0.015375900082290173, + 0.015219368040561676, + 0.009798976592719555, + 0.009403730742633343, + 0.023741137236356735, + 0.019548308104276657, + 0.017906060442328453, + 0.02449362725019455, + 0.02780856005847454, + 0.023061256855726242, + 0.032973844558000565, + 0.018826600164175034, + 0.005575511138886213, + 0.003922447562217712, + 0.003534404095262289, + 0.006344262510538101, + 0.004112874623388052, + 0.004060851410031319, + 0.006689851172268391, + 0.008972471579909325, + 0.003554467810317874, + 0.0025114971213042736, + 0.0023467428982257843, + 0.003041076473891735, + 0.002985073020681739, + 0.003812734968960285, + 0.002499230904504657, + 0.003641379065811634, + 0.003401386085897684, + 0.0033847391605377197, + 0.0038504183758050203, + 0.007406450808048248, + 0.0027012235950678587, + 0.002267169766128063, + 0.0034202614333480597, + 0.0032968581654131413, + 0.00799670722335577, + 0.009996374137699604, + 0.01043298002332449, + 0.017114141955971718, + 0.008457532152533531, + 0.00591391883790493, + 0.010645917616784573, + 0.005574412643909454, + 0.014513352885842323, + 0.01494716014713049, + 0.030667893588542938, + 0.036123380064964294, + 0.018964949995279312, + 0.011737530119717121, + 0.024173852056264877, + 0.012338212691247463, + 0.01234088372439146, + 0.015076600015163422, + 0.03862166404724121, + 0.04541414603590965, + 0.011509154923260212, + 0.017028359696269035, + 0.020600272342562675, + 0.02184676006436348, + 0.0015842639841139317, + 0.0016234489157795906, + 0.002907759742811322, + 0.0027353758923709393, + 0.002160856034606695, + 0.003453795798122883, + 0.003195931436493993, + 0.0028834878467023373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.08907544612884521 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074232757091522 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.016074232757091522, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074232757091522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13511019945144653, + 0.1502106636762619, + 0.07323212176561356, + 0.051264528185129166, + 0.07539506256580353, + 0.04469946026802063, + 0.05537198483943939, + 0.05804048478603363, + 0.28004372119903564, + 0.0766318067908287 + ], + "credit_map": [ + 0.015316180884838104, + 0.019274460151791573, + 0.01617676019668579, + 0.018783941864967346, + 0.0222922395914793, + 0.01526731252670288, + 0.01564965955913067, + 0.012349621392786503, + 0.022155839949846268, + 0.017583616077899933, + 0.027604982256889343, + 0.017253130674362183, + 0.020890923216938972, + 0.015181423164904118, + 0.010462962090969086, + 0.019077764824032784, + 0.010255355387926102, + 0.007265095598995686, + 0.008266525343060493, + 0.009808274917304516, + 0.01304069347679615, + 0.009217115119099617, + 0.009604875929653645, + 0.005774179473519325, + 0.009952870197594166, + 0.007883839309215546, + 0.004771513864398003, + 0.005111398175358772, + 0.00469353049993515, + 0.006625660229474306, + 0.005747416988015175, + 0.0064782933332026005, + 0.010327635332942009, + 0.008213771507143974, + 0.007542639039456844, + 0.012569308280944824, + 0.009305674582719803, + 0.007272559683769941, + 0.006685235537588596, + 0.013478230684995651, + 0.005528636742383242, + 0.004536915570497513, + 0.00480644591152668, + 0.009178959764540195, + 0.0049497270956635475, + 0.0048209684900939465, + 0.005810575094074011, + 0.00506722554564476, + 0.0054518659599125385, + 0.006797098089009523, + 0.006601207889616489, + 0.007319008000195026, + 0.010460568591952324, + 0.006610954646021128, + 0.006328561808913946, + 0.005802713334560394, + 0.005532974377274513, + 0.006318613421171904, + 0.01004807185381651, + 0.007875028066337109, + 0.007800173945724964, + 0.006580619141459465, + 0.0080466503277421, + 0.005838344804942608, + 0.02379436045885086, + 0.022889982908964157, + 0.053866222500801086, + 0.08037743717432022, + 0.01793423853814602, + 0.02370534837245941, + 0.028358004987239838, + 0.029118122532963753, + 0.005639297422021627, + 0.007742888294160366, + 0.009783760644495487, + 0.012026085518300533, + 0.012941939756274223, + 0.008093825541436672, + 0.01202861126512289, + 0.008375382050871849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.016074232757091522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510551548562944 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510551548562944, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510551548562944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18886849284172058, + 0.08028917014598846, + 0.075043685734272, + 0.06472864001989365, + 0.10410293936729431, + 0.043830983340740204, + 0.1466650813817978, + 0.09449582546949387, + 0.13452036678791046, + 0.06745482236146927 + ], + "credit_map": [ + 0.03553561493754387, + 0.016520993784070015, + 0.03445364907383919, + 0.015691429376602173, + 0.015423840843141079, + 0.03791036084294319, + 0.01418343372642994, + 0.019149167463183403, + 0.011395595036447048, + 0.012970568612217903, + 0.007871690206229687, + 0.00833360105752945, + 0.014431129209697247, + 0.009070022031664848, + 0.007419269997626543, + 0.008797298185527325, + 0.010006291791796684, + 0.007768300827592611, + 0.01125207357108593, + 0.009007932618260384, + 0.009149229153990746, + 0.008695225231349468, + 0.01158551499247551, + 0.0075791217386722565, + 0.010883035138249397, + 0.006623902823776007, + 0.007514537777751684, + 0.008546659722924232, + 0.005914662033319473, + 0.005833681207150221, + 0.007739306427538395, + 0.011672861874103546, + 0.025797180831432343, + 0.00980572309345007, + 0.015261839143931866, + 0.013559193350374699, + 0.006871082819998264, + 0.008415928110480309, + 0.012351194396615028, + 0.012040805071592331, + 0.006743839010596275, + 0.004790660459548235, + 0.004660972394049168, + 0.007354854140430689, + 0.004818392917513847, + 0.0054671792313456535, + 0.005457683000713587, + 0.004537404514849186, + 0.014119058847427368, + 0.02000107616186142, + 0.021268362179398537, + 0.037741828709840775, + 0.009885729290544987, + 0.006929793395102024, + 0.021804777905344963, + 0.014914464205503464, + 0.013663745485246181, + 0.015422352589666843, + 0.014479460194706917, + 0.01567920669913292, + 0.007143622264266014, + 0.008934622630476952, + 0.010571365244686604, + 0.00860145129263401, + 0.013222413137555122, + 0.011721367947757244, + 0.02588275820016861, + 0.020697128027677536, + 0.011085495352745056, + 0.016997527331113815, + 0.018043288961052895, + 0.01687038689851761, + 0.004170822910964489, + 0.005904242862015963, + 0.012234535068273544, + 0.009420025162398815, + 0.008943745866417885, + 0.009548163041472435, + 0.008409734815359116, + 0.008823554031550884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.001510551548562944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ornate carriage with crimson lacquer and gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four dark brown horses galloping in front of the carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a driver wearing a dark uniform and top hat visible holding the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry dirt country road with green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot filmed from a low angle with a dynamic tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282833397388458 + ], + "opposite_logprob_tokens": [ + -3.773282766342163 + ], + "logprob_reward": -0.023282833397388458, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282833397388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128469780087471, + 0.07740405946969986, + 0.03636894002556801, + 0.07099242508411407, + 0.06340578943490982, + 0.08474310487508774, + 0.07512076944112778, + 0.13075195252895355, + 0.128666490316391, + 0.20407669246196747 + ], + "credit_map": [ + 0.02789386734366417, + 0.009971123188734055, + 0.01858125813305378, + 0.010709645226597786, + 0.010939405299723148, + 0.024947430938482285, + 0.009441492147743702, + 0.015985552221536636, + 0.012281840667128563, + 0.005885094404220581, + 0.015561774373054504, + 0.012412586249411106, + 0.00659747002646327, + 0.01114596240222454, + 0.0070172022096812725, + 0.006502128671854734, + 0.006576523184776306, + 0.005885395221412182, + 0.0037228832952678204, + 0.0051179490983486176, + 0.0040587568655610085, + 0.004265965893864632, + 0.0036452109925448895, + 0.0030962531454861164, + 0.014127176254987717, + 0.007291415706276894, + 0.006531946361064911, + 0.00829230435192585, + 0.0059799859300255775, + 0.009487006813287735, + 0.008853844366967678, + 0.01042874064296484, + 0.008921450935304165, + 0.007139415014535189, + 0.005995115265250206, + 0.010312270373106003, + 0.007377572823315859, + 0.005910826846957207, + 0.006238249596208334, + 0.011510889045894146, + 0.008244293741881847, + 0.007916631177067757, + 0.009980275295674801, + 0.02016575075685978, + 0.006677845027297735, + 0.0062405625358223915, + 0.017892887815833092, + 0.0076248603872954845, + 0.006422069855034351, + 0.011332709342241287, + 0.009921489283442497, + 0.017950814217329025, + 0.005624928046017885, + 0.004884670954197645, + 0.010152381844818592, + 0.008831706829369068, + 0.011417836882174015, + 0.013989470899105072, + 0.028644103556871414, + 0.034051451832056046, + 0.009540894068777561, + 0.00685628317296505, + 0.01744658313691616, + 0.008805333636701107, + 0.010639754123985767, + 0.009300554171204567, + 0.027502307668328285, + 0.03181234374642372, + 0.009534988552331924, + 0.010929482989013195, + 0.01620333455502987, + 0.012743727304041386, + 0.015559818595647812, + 0.016060721129179, + 0.03844347968697548, + 0.04797923564910889, + 0.027409173548221588, + 0.019052693620324135, + 0.026140857487916946, + 0.013430708087980747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.023282833397388458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_tensor([5068, 5068])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_tensor([5068, 5068])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c836e64f2bbb35a7e182cc229986a205e7a0dd8f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_tensor([5068, 5068])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "continuous, fluid close-up glides smoothly, closely following the precise, coordinated hand movements of two scientists as they work together on a complex holographic interface. The primary focus is on one of the scientists, her dark hair neatly tied in a bun, her face a look of intense concentration. She wears a wrinkle-free, pure white, high-tech lab coat, a dark gray turtleneck peeking out from beneath. Her fingers, and those of her collaborator beside her, dance gracefully in the air, pinching and zooming, sliding and rotating, collectively manipulating a luminous three-dimensional projection suspended between them, composed of countless translucent data streams and swirling molecular models. The holographic interface serves as the scene's primary light source, casting an ethereal, cool hue of cyan and electric blue onto their focused faces and hands, highlighting the subtle textures of their skin and the pristine texture of their lab coats. Against the softly out-of-focus background, the futuristic, sterile laboratory, with its polished white composite surfaces and minimalist design, looms, bathed in shadowless clinical white light from overhead LED panels. The cool color temperature of the entire scene creates an atmosphere of intense, quiet concentration and cutting-edge technological precision, where every gesture is carefully considered and part of seamless collaboration.", + "seed": 8587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_5068_seed_8587_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07479951499262824, + "vlm_logprob": -0.07479951499262824, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.07479951499262824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040552642196416855 + ], + "opposite_logprob_tokens": [ + -3.6655526161193848 + ], + "logprob_reward": -0.040552642196416855, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.040552642196416855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08619336038827896, + 0.0977478101849556, + 0.034629929810762405, + 0.05084756016731262, + 0.10418922454118729, + 0.4316149353981018, + 0.10222650319337845, + 0.03804760426282883, + 0.02119823545217514, + 0.03330482915043831 + ], + "credit_map": [ + 0.008757567033171654, + 0.00813530944287777, + 0.011591046117246151, + 0.012249156832695007, + 0.00797454733401537, + 0.015249504707753658, + 0.009256909601390362, + 0.012979320250451565, + 0.0052176411263644695, + 0.007000202313065529, + 0.014087535440921783, + 0.023603025823831558, + 0.00917588360607624, + 0.01514203380793333, + 0.009429198689758778, + 0.014092289842665195, + 0.0032781590707600117, + 0.0034837042912840843, + 0.00504141952842474, + 0.005274120252579451, + 0.0035128409508615732, + 0.004628943745046854, + 0.00426986999809742, + 0.005140870343893766, + 0.003576409537345171, + 0.0053741163574159145, + 0.0057589407078921795, + 0.009118498302996159, + 0.004050709772855043, + 0.007096824701875448, + 0.009797882288694382, + 0.006074176169931889, + 0.011800026521086693, + 0.02001004107296467, + 0.013863843865692616, + 0.012521641328930855, + 0.01273783203214407, + 0.011884829960763454, + 0.01060224138200283, + 0.0107687683776021, + 0.03384930267930031, + 0.04634430259466171, + 0.07775934040546417, + 0.02748405747115612, + 0.05436837300658226, + 0.05743454769253731, + 0.032883383333683014, + 0.10149163752794266, + 0.009403166361153126, + 0.01118901651352644, + 0.013837294653058052, + 0.014914714731276035, + 0.01087258942425251, + 0.016546182334423065, + 0.012661730870604515, + 0.012801801785826683, + 0.0038963351398706436, + 0.006181282456964254, + 0.005610157735645771, + 0.003482205793261528, + 0.003884389065206051, + 0.005062633194029331, + 0.005508508998900652, + 0.0044220928102731705, + 0.003235098673030734, + 0.0030548202339559793, + 0.0027756926137953997, + 0.0016843426274135709, + 0.0022221626713871956, + 0.001807363354600966, + 0.0036140535958111286, + 0.002804700518026948, + 0.002805495634675026, + 0.006576224695891142, + 0.003866748884320259, + 0.004256881773471832, + 0.004379946738481522, + 0.003821258433163166, + 0.0038489631842821836, + 0.0037493121344596148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.040552642196416855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.142842099070549 + ], + "opposite_logprob_tokens": [ + -2.0178420543670654 + ], + "logprob_reward": -0.142842099070549, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.142842099070549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2788366377353668, + 0.17420613765716553, + 0.07529037445783615, + 0.042727891355752945, + 0.07569362223148346, + 0.09474849700927734, + 0.02415204606950283, + 0.05089541897177696, + 0.023494521155953407, + 0.15995490550994873 + ], + "credit_map": [ + 0.01862455904483795, + 0.01851757802069187, + 0.026703456416726112, + 0.03415697440505028, + 0.021476220339536667, + 0.05361539125442505, + 0.033262599259614944, + 0.07247985899448395, + 0.009081457741558552, + 0.013008675538003445, + 0.025745898485183716, + 0.0335848294198513, + 0.01699369214475155, + 0.030023427680134773, + 0.015472648665308952, + 0.030295509845018387, + 0.006722534541040659, + 0.008983513340353966, + 0.015304574742913246, + 0.0085593331605196, + 0.009690453298389912, + 0.008747159503400326, + 0.008922606706619263, + 0.008360208012163639, + 0.006524930242449045, + 0.006732165347784758, + 0.004293746780604124, + 0.004275921732187271, + 0.003285743063315749, + 0.005154880695044994, + 0.00849626399576664, + 0.003964236006140709, + 0.009903576225042343, + 0.021029826253652573, + 0.009518437087535858, + 0.0066225859336555, + 0.007686815690249205, + 0.010334964841604233, + 0.004737207666039467, + 0.00586020527407527, + 0.006529513746500015, + 0.006882083602249622, + 0.020737027749419212, + 0.00502323592081666, + 0.010513505898416042, + 0.010158572345972061, + 0.007235165219753981, + 0.0276693906635046, + 0.001973728882148862, + 0.004487323109060526, + 0.0032828361727297306, + 0.0024589048698544502, + 0.002536844927817583, + 0.0024148121010512114, + 0.0030229853000491858, + 0.003974611405283213, + 0.004594103898853064, + 0.008210605941712856, + 0.006459794472903013, + 0.005277432966977358, + 0.006935013458132744, + 0.0072540235705673695, + 0.0064465878531336784, + 0.00571785494685173, + 0.003295428352430463, + 0.0032719201408326626, + 0.0032331051770597696, + 0.002305575180798769, + 0.002832896076142788, + 0.0016961825313046575, + 0.0028174349572509527, + 0.004041977692395449, + 0.015068989247083664, + 0.02964986115694046, + 0.022381799295544624, + 0.0192460548132658, + 0.0208482313901186, + 0.015208682976663113, + 0.01846599392592907, + 0.01908530294895172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.142842099070549 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978908035904169 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.00978908035904169, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978908035904169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047763485461473465, + 0.10394111275672913, + 0.04518609866499901, + 0.0873161181807518, + 0.15212294459342957, + 0.34443947672843933, + 0.13370715081691742, + 0.027310537174344063, + 0.035093337297439575, + 0.023119673132896423 + ], + "credit_map": [ + 0.0032375811133533716, + 0.004132271744310856, + 0.007307023275643587, + 0.0065081557258963585, + 0.0036051603965461254, + 0.0075090364553034306, + 0.009107530117034912, + 0.006356728263199329, + 0.0058579049073159695, + 0.00822676345705986, + 0.01376678328961134, + 0.029921166598796844, + 0.007370518986135721, + 0.01467119436711073, + 0.011674568057060242, + 0.01245221495628357, + 0.003432799596339464, + 0.0033447931054979563, + 0.00816913228482008, + 0.00763486186042428, + 0.0036231609992682934, + 0.005823924671858549, + 0.0058171432465314865, + 0.0073402817361056805, + 0.01074287947267294, + 0.013181726448237896, + 0.006929157301783562, + 0.008876621723175049, + 0.009203690104186535, + 0.010928708128631115, + 0.017070719972252846, + 0.010382622480392456, + 0.01504859235137701, + 0.026794692501425743, + 0.022384626790881157, + 0.012843459844589233, + 0.031685397028923035, + 0.01911572553217411, + 0.012450751848518848, + 0.011799709871411324, + 0.023347266018390656, + 0.033612947911024094, + 0.06938116252422333, + 0.022967174649238586, + 0.03408375009894371, + 0.040872570127248764, + 0.02711457572877407, + 0.09306006133556366, + 0.014723574742674828, + 0.017240220680832863, + 0.02041766420006752, + 0.010607903823256493, + 0.01947746053338051, + 0.01865360140800476, + 0.012971332296729088, + 0.019615406170487404, + 0.0027464348822832108, + 0.0034067677333950996, + 0.004125538747757673, + 0.002551150741055608, + 0.004025719594210386, + 0.0032261167652904987, + 0.003839093493297696, + 0.0033897182438522577, + 0.004568243399262428, + 0.004995705559849739, + 0.005253899842500687, + 0.0033265789970755577, + 0.003624899312853813, + 0.002467973157763481, + 0.006969221401959658, + 0.0038868198171257973, + 0.0025182096287608147, + 0.0027110488153994083, + 0.0028097964823246002, + 0.0031090143602341413, + 0.0037590982392430305, + 0.0019205063581466675, + 0.0033908600453287363, + 0.002901140134781599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.00978908035904169 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009201106731779873 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009201106731779873, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009201106731779873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10638078302145004, + 0.13644655048847198, + 0.07785585522651672, + 0.07125157117843628, + 0.2309773862361908, + 0.08739237487316132, + 0.04413062706589699, + 0.07158271223306656, + 0.058548033237457275, + 0.11543409526348114 + ], + "credit_map": [ + 0.006507973652333021, + 0.007129942998290062, + 0.012050275690853596, + 0.013148930855095387, + 0.008241343311965466, + 0.019579986110329628, + 0.015148577280342579, + 0.024573763832449913, + 0.01019708439707756, + 0.011964836157858372, + 0.023395705968141556, + 0.025387289002537727, + 0.011688923463225365, + 0.021455438807606697, + 0.012246756814420223, + 0.02011052332818508, + 0.007534417789429426, + 0.007812665775418282, + 0.009956534020602703, + 0.01175651140511036, + 0.008065639063715935, + 0.010226553305983543, + 0.01135398168116808, + 0.011149562895298004, + 0.006159977056086063, + 0.009879520162940025, + 0.006812713574618101, + 0.008367782458662987, + 0.006012076511979103, + 0.009355386719107628, + 0.014503629878163338, + 0.010160482488572598, + 0.024698961526155472, + 0.03078550286591053, + 0.035171277821063995, + 0.01897422969341278, + 0.053651709109544754, + 0.028306130319833755, + 0.02135438472032547, + 0.01803520880639553, + 0.008362852968275547, + 0.010847143828868866, + 0.0128406947478652, + 0.006603853311389685, + 0.010696586221456528, + 0.014121964573860168, + 0.008173801004886627, + 0.015745487064123154, + 0.0030668226536363363, + 0.003277923446148634, + 0.007560080382972956, + 0.006310586351901293, + 0.004514982458204031, + 0.004095971584320068, + 0.0065742377191782, + 0.0087300268933177, + 0.00567200081422925, + 0.006290746387094259, + 0.014417048543691635, + 0.0074782962910830975, + 0.005277371499687433, + 0.00549750542268157, + 0.010639997199177742, + 0.01630975492298603, + 0.009900338016450405, + 0.006278230343014002, + 0.00832636933773756, + 0.0061500417068600655, + 0.007420744281262159, + 0.003987448755651712, + 0.010384495370090008, + 0.006100370083004236, + 0.010644687339663506, + 0.02211620658636093, + 0.014892707578837872, + 0.01651216484606266, + 0.013632351532578468, + 0.011579209938645363, + 0.013155301101505756, + 0.012901479378342628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.0009201106731779873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798936426639557 + ], + "opposite_logprob_tokens": [ + -1.8048936128616333 + ], + "logprob_reward": -0.1798936426639557, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798936426639557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036349210888147354, + 0.17736856639385223, + 0.05219397693872452, + 0.07568825781345367, + 0.24332047998905182, + 0.24575024843215942, + 0.03923344984650612, + 0.05759454146027565, + 0.0255377646535635, + 0.046963535249233246 + ], + "credit_map": [ + 0.0033155314158648252, + 0.002827133284881711, + 0.005910345818847418, + 0.006424169987440109, + 0.00338218011893332, + 0.005347400903701782, + 0.004230440594255924, + 0.004912007134407759, + 0.013191926293075085, + 0.013519764877855778, + 0.030087649822235107, + 0.033900413662195206, + 0.01704443246126175, + 0.030338730663061142, + 0.015127604827284813, + 0.02415803261101246, + 0.0047972542233765125, + 0.0052520642057061195, + 0.009974789805710316, + 0.0050818962045013905, + 0.00606215326115489, + 0.008982531726360321, + 0.006165081635117531, + 0.005878202151507139, + 0.009218436665832996, + 0.010531232692301273, + 0.008136295713484287, + 0.007049295585602522, + 0.006183623801916838, + 0.008716633543372154, + 0.018265604972839355, + 0.007587129715830088, + 0.025563359260559082, + 0.03531051054596901, + 0.0345693975687027, + 0.019887281581759453, + 0.0492338091135025, + 0.03494764864444733, + 0.02180064655840397, + 0.02200779877603054, + 0.018243737518787384, + 0.028425389900803566, + 0.04353366419672966, + 0.018860993906855583, + 0.02638823911547661, + 0.03335874527692795, + 0.01996743306517601, + 0.056972019374370575, + 0.003759001614525914, + 0.005984129849821329, + 0.004181492142379284, + 0.005344686098396778, + 0.004160837270319462, + 0.003989376127719879, + 0.00549053680151701, + 0.006323386915028095, + 0.005505054257810116, + 0.0047926511615514755, + 0.010120665654540062, + 0.009320476092398167, + 0.00492390152066946, + 0.004350797273218632, + 0.00892494898289442, + 0.009656042791903019, + 0.00375954108312726, + 0.0038169424515217543, + 0.002926487475633621, + 0.00269343052059412, + 0.0024973477702587843, + 0.001981182023882866, + 0.0037276726216077805, + 0.004135156515985727, + 0.004802747163921595, + 0.0058279200457036495, + 0.005412353202700615, + 0.006305131129920483, + 0.008417944423854351, + 0.00317306537181139, + 0.006520568393170834, + 0.006503802724182606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.1798936426639557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "continuous, fluid close-up glides smoothly, closely following the precise, coordinated hand movements of two scientists as they work together on a complex holographic interface. The primary focus is on one of the scientists, her dark hair neatly tied in a bun, her face a look of intense concentration. She wears a wrinkle-free, pure white, high-tech lab coat, a dark gray turtleneck peeking out from beneath. Her fingers, and those of her collaborator beside her, dance gracefully in the air, pinching and zooming, sliding and rotating, collectively manipulating a luminous three-dimensional projection suspended between them, composed of countless translucent data streams and swirling molecular models. The holographic interface serves as the scene's primary light source, casting an ethereal, cool hue of cyan and electric blue onto their focused faces and hands, highlighting the subtle textures of their skin and the pristine texture of their lab coats. Against the softly out-of-focus background, the futuristic, sterile laboratory, with its polished white composite surfaces and minimalist design, looms, bathed in shadowless clinical white light from overhead LED panels. The cool color temperature of the entire scene creates an atmosphere of intense, quiet concentration and cutting-edge technological precision, where every gesture is carefully considered and part of seamless collaboration.", + "seed": 8588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_5068_seed_8588_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.053213589685037734, + "vlm_logprob": -0.053213589685037734, + "vlm_yes_no_margin": 4.124999976158142, + "vlm_token_credit": -0.053213589685037734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03311275690793991 + ], + "opposite_logprob_tokens": [ + -5.033112525939941 + ], + "logprob_reward": -0.03311275690793991, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.03311275690793991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17612147331237793, + 0.07024944573640823, + 0.04914433881640434, + 0.09114493429660797, + 0.16559673845767975, + 0.136733278632164, + 0.04066401347517967, + 0.042740389704704285, + 0.16370436549186707, + 0.06390098482370377 + ], + "credit_map": [ + 0.019438857212662697, + 0.03156968951225281, + 0.015605165623128414, + 0.024774396792054176, + 0.02461102418601513, + 0.02448325790464878, + 0.01652715913951397, + 0.019111914560198784, + 0.004914695397019386, + 0.0073882038705050945, + 0.009189391508698463, + 0.010542740114033222, + 0.006102772429585457, + 0.008493437431752682, + 0.015063142403960228, + 0.008555064909160137, + 0.00404738262295723, + 0.005094220861792564, + 0.007523706182837486, + 0.003978974185883999, + 0.0069970334880054, + 0.00991798285394907, + 0.006714107934385538, + 0.0048709288239479065, + 0.012914084829390049, + 0.012232818640768528, + 0.012509741820394993, + 0.014244530349969864, + 0.009011867456138134, + 0.010045434348285198, + 0.01059468649327755, + 0.009591769427061081, + 0.013746943324804306, + 0.013352584093809128, + 0.01409982144832611, + 0.040195148438215256, + 0.014092479832470417, + 0.021928297355771065, + 0.03555028513073921, + 0.012631174176931381, + 0.009975573047995567, + 0.012184522114694118, + 0.00726604089140892, + 0.03543548658490181, + 0.00953297782689333, + 0.009335316717624664, + 0.012106097303330898, + 0.040897250175476074, + 0.004347559530287981, + 0.005084286443889141, + 0.006039610598236322, + 0.005926236510276794, + 0.003702145302668214, + 0.004999944008886814, + 0.005396642722189426, + 0.005167591385543346, + 0.005648426711559296, + 0.0038925607223063707, + 0.003949231002479792, + 0.005525864660739899, + 0.005594540387392044, + 0.006843129172921181, + 0.0044005513191223145, + 0.006886085029691458, + 0.018528416752815247, + 0.004607271868735552, + 0.06734458357095718, + 0.00678872037678957, + 0.004949474707245827, + 0.0319335013628006, + 0.006286684889346361, + 0.023265711963176727, + 0.004467975348234177, + 0.004681344609707594, + 0.016471607610583305, + 0.003926008474081755, + 0.0038564843125641346, + 0.011855746619403362, + 0.008217964321374893, + 0.010423851199448109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03311275690793991 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902072370052338 + ], + "opposite_logprob_tokens": [ + -2.5790207386016846 + ], + "logprob_reward": -0.07902072370052338, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902072370052338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08116709440946579, + 0.02605932578444481, + 0.1652560979127884, + 0.05957677960395813, + 0.22793284058570862, + 0.2593187689781189, + 0.028547976166009903, + 0.017986051738262177, + 0.1022583469748497, + 0.03189669921994209 + ], + "credit_map": [ + 0.006614245939999819, + 0.015166311524808407, + 0.013464056886732578, + 0.005937176290899515, + 0.01221086923032999, + 0.01088383886963129, + 0.0085186418145895, + 0.008371947333216667, + 0.0019005137728527188, + 0.0031666490249335766, + 0.003218425903469324, + 0.0040237209759652615, + 0.003514290554448962, + 0.003824961604550481, + 0.003904798999428749, + 0.002505964832380414, + 0.014656239189207554, + 0.013542518950998783, + 0.032967690378427505, + 0.022294847294688225, + 0.015564889647066593, + 0.014200521633028984, + 0.04017338901758194, + 0.011855995282530785, + 0.0071317595429718494, + 0.007153011392802, + 0.009782581590116024, + 0.011280654929578304, + 0.005106340628117323, + 0.005083664320409298, + 0.008833903819322586, + 0.005204866174608469, + 0.02563898265361786, + 0.01799057051539421, + 0.02817675471305847, + 0.03056713379919529, + 0.0240811537951231, + 0.036567505449056625, + 0.040945954620838165, + 0.023964786902070045, + 0.015543948858976364, + 0.020082997158169746, + 0.012503675185143948, + 0.06861621886491776, + 0.01149684563279152, + 0.01552162691950798, + 0.018836582079529762, + 0.09671688079833984, + 0.0027481787838041782, + 0.0032132563646882772, + 0.004191036801785231, + 0.006215664092451334, + 0.002436525421217084, + 0.002173836575821042, + 0.0031232053879648447, + 0.004446272738277912, + 0.001906076562590897, + 0.00185454823076725, + 0.002231041667982936, + 0.0029550979379564524, + 0.0023358336184173822, + 0.0023463862016797066, + 0.002436649287119508, + 0.0019204177660867572, + 0.005535912234336138, + 0.0042101615108549595, + 0.011081534437835217, + 0.0071461815387010574, + 0.012700150720775127, + 0.02302023395895958, + 0.012926707044243813, + 0.025637462735176086, + 0.0033000316470861435, + 0.0022483128122985363, + 0.007652277126908302, + 0.0026330791879445314, + 0.0031373172532767057, + 0.0056863753125071526, + 0.003277708776295185, + 0.003961598034948111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.07902072370052338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979179609566927 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.00979179609566927, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979179609566927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10836359113454819, + 0.04914836958050728, + 0.0897815153002739, + 0.09823060035705566, + 0.09900493919849396, + 0.1545378416776657, + 0.03613702207803726, + 0.023254916071891785, + 0.31800007820129395, + 0.02354111336171627 + ], + "credit_map": [ + 0.009826578199863434, + 0.012561705894768238, + 0.013686846941709518, + 0.011144658550620079, + 0.012032851576805115, + 0.018281232565641403, + 0.017564738169312477, + 0.013264985755085945, + 0.004153945483267307, + 0.0069840881042182446, + 0.00636849831789732, + 0.007703535724431276, + 0.004358076024800539, + 0.005490285810083151, + 0.008897091262042522, + 0.005192852113395929, + 0.006458810064941645, + 0.007217558566480875, + 0.009962446987628937, + 0.006837402004748583, + 0.018705924972891808, + 0.013490727171301842, + 0.014210115186870098, + 0.012898529879748821, + 0.010069767013192177, + 0.010101009160280228, + 0.015946878120303154, + 0.011900280602276325, + 0.010743051767349243, + 0.01654469594359398, + 0.013036489486694336, + 0.009888425469398499, + 0.008181360550224781, + 0.00731908343732357, + 0.012569410726428032, + 0.021353542804718018, + 0.010986016131937504, + 0.013287035748362541, + 0.014251216314733028, + 0.011057272553443909, + 0.010488633997738361, + 0.015228071250021458, + 0.009103222750127316, + 0.038105309009552, + 0.011823631823062897, + 0.011385156773030758, + 0.012446826323866844, + 0.045956991612911224, + 0.004885448142886162, + 0.00551152229309082, + 0.005388923455029726, + 0.004938529338687658, + 0.0034475072752684355, + 0.004011578392237425, + 0.0029157171957194805, + 0.005037792958319187, + 0.002257412765175104, + 0.002780830254778266, + 0.002313220174983144, + 0.0025543596129864454, + 0.004353413823992014, + 0.0030472821090370417, + 0.0032796813175082207, + 0.0026687162462621927, + 0.030239809304475784, + 0.007892685011029243, + 0.1294415146112442, + 0.013936902396380901, + 0.0077596744522452354, + 0.09447316825389862, + 0.010183164849877357, + 0.024073157459497452, + 0.002751322463154793, + 0.002074286574497819, + 0.004776660352945328, + 0.0016666611190885305, + 0.0024122497998178005, + 0.0030312063172459602, + 0.0029330067336559296, + 0.003895720699802041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.00979179609566927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013374679256230593 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013374679256230593, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013374679256230593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09497900307178497, + 0.05501827225089073, + 0.204053595662117, + 0.10293471813201904, + 0.11273042857646942, + 0.11596699804067612, + 0.045877065509557724, + 0.03758833929896355, + 0.20824359357357025, + 0.022607972845435143 + ], + "credit_map": [ + 0.006988869048655033, + 0.011858971789479256, + 0.01028482150286436, + 0.010954029858112335, + 0.015898114070296288, + 0.017642086371779442, + 0.01000422053039074, + 0.011347887106239796, + 0.0030485796742141247, + 0.009662019088864326, + 0.006810655351728201, + 0.008305422961711884, + 0.00873419176787138, + 0.0058948746882379055, + 0.007201082538813353, + 0.005361445248126984, + 0.015251255594193935, + 0.016652856022119522, + 0.03165849670767784, + 0.019789258018136024, + 0.031622692942619324, + 0.02468227408826351, + 0.04063433036208153, + 0.023762429133057594, + 0.011997480876743793, + 0.014913690276443958, + 0.01273764856159687, + 0.016459664329886436, + 0.011897516436874866, + 0.01585087925195694, + 0.012640362605452538, + 0.006437473464757204, + 0.01204997580498457, + 0.008640194311738014, + 0.013055099174380302, + 0.017660269513726234, + 0.012204708531498909, + 0.019870320335030556, + 0.01788070611655712, + 0.011369163170456886, + 0.008302036672830582, + 0.009615994989871979, + 0.0050984821282327175, + 0.028319908306002617, + 0.004533606581389904, + 0.005833319388329983, + 0.01009908877313137, + 0.044164568185806274, + 0.005527209490537643, + 0.005722475703805685, + 0.007471624296158552, + 0.008931542746722698, + 0.003987177740782499, + 0.00400189682841301, + 0.004633481614291668, + 0.005601656157523394, + 0.003579680575057864, + 0.004003336653113365, + 0.004227587953209877, + 0.0053260899148881435, + 0.005242750979959965, + 0.005856235511600971, + 0.004767782520502806, + 0.004584874026477337, + 0.019305476918816566, + 0.005208113230764866, + 0.08448335528373718, + 0.009134851396083832, + 0.005456550046801567, + 0.06130737438797951, + 0.006852601654827595, + 0.01649527996778488, + 0.00219864584505558, + 0.0018296816851943731, + 0.004169008228927851, + 0.0018190531991422176, + 0.002355238189920783, + 0.00294915190897882, + 0.00480974605306983, + 0.002477448433637619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.0013374679256230593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280520379543304 + ], + "opposite_logprob_tokens": [ + -2.0178050994873047 + ], + "logprob_reward": -0.14280520379543304, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280520379543304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13181033730506897, + 0.07745660096406937, + 0.20599909126758575, + 0.07898449897766113, + 0.09437904506921768, + 0.08957916498184204, + 0.08399266004562378, + 0.04355430230498314, + 0.16285988688468933, + 0.031384363770484924 + ], + "credit_map": [ + 0.012055239640176296, + 0.01753898151218891, + 0.017673365771770477, + 0.015572398900985718, + 0.018923044204711914, + 0.01610310934484005, + 0.01241377741098404, + 0.021530423313379288, + 0.005742370150983334, + 0.012872932478785515, + 0.012123774737119675, + 0.010545647703111172, + 0.009095817804336548, + 0.01187965925782919, + 0.008866085670888424, + 0.006330318748950958, + 0.012603319250047207, + 0.016019422560930252, + 0.024101410061120987, + 0.016662418842315674, + 0.048271555453538895, + 0.03240188583731651, + 0.03086855448782444, + 0.025070510804653168, + 0.009473973885178566, + 0.010694455355405807, + 0.011988615617156029, + 0.012861667200922966, + 0.008594012819230556, + 0.011096836067736149, + 0.010129125788807869, + 0.004145804792642593, + 0.011119464412331581, + 0.006103533320128918, + 0.011649467051029205, + 0.011313574388623238, + 0.009969854727387428, + 0.01645803265273571, + 0.016852082684636116, + 0.01091303862631321, + 0.00689492654055357, + 0.010302950628101826, + 0.004003516398370266, + 0.020352784544229507, + 0.008005445823073387, + 0.005379090551286936, + 0.007190250325948, + 0.0274501983076334, + 0.008681735023856163, + 0.013655365444719791, + 0.012715010903775692, + 0.01560839731246233, + 0.006714961025863886, + 0.00605046097189188, + 0.009142167866230011, + 0.011424564756453037, + 0.009765415452420712, + 0.0034739163238555193, + 0.004232751205563545, + 0.005286199506372213, + 0.0038056927733123302, + 0.005953206680715084, + 0.004634731449186802, + 0.006402387283742428, + 0.014501916244626045, + 0.004001363180577755, + 0.06520366668701172, + 0.006616008002310991, + 0.0037340279668569565, + 0.053421467542648315, + 0.004963988903909922, + 0.010417439043521881, + 0.0025177402421832085, + 0.0019285206217318773, + 0.008848488330841064, + 0.0013338513672351837, + 0.002109360182657838, + 0.006992743816226721, + 0.002977379597723484, + 0.004676278680562973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.14280520379543304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "continuous, fluid close-up glides smoothly, closely following the precise, coordinated hand movements of two scientists as they work together on a complex holographic interface. The primary focus is on one of the scientists, her dark hair neatly tied in a bun, her face a look of intense concentration. She wears a wrinkle-free, pure white, high-tech lab coat, a dark gray turtleneck peeking out from beneath. Her fingers, and those of her collaborator beside her, dance gracefully in the air, pinching and zooming, sliding and rotating, collectively manipulating a luminous three-dimensional projection suspended between them, composed of countless translucent data streams and swirling molecular models. The holographic interface serves as the scene's primary light source, casting an ethereal, cool hue of cyan and electric blue onto their focused faces and hands, highlighting the subtle textures of their skin and the pristine texture of their lab coats. Against the softly out-of-focus background, the futuristic, sterile laboratory, with its polished white composite surfaces and minimalist design, looms, bathed in shadowless clinical white light from overhead LED panels. The cool color temperature of the entire scene creates an atmosphere of intense, quiet concentration and cutting-edge technological precision, where every gesture is carefully considered and part of seamless collaboration.", + "seed": 8589, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_5068_seed_8589_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10962925888597966, + "vlm_logprob": -0.10962925888597966, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.10962925888597966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041486866772174835 + ], + "opposite_logprob_tokens": [ + -3.791486978530884 + ], + "logprob_reward": -0.041486866772174835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.041486866772174835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09875401109457016, + 0.20125101506710052, + 0.03504091501235962, + 0.1673072874546051, + 0.1340387910604477, + 0.10484886169433594, + 0.0804411843419075, + 0.03630381077528, + 0.050255414098501205, + 0.09175869077444077 + ], + "credit_map": [ + 0.013897460885345936, + 0.010631103999912739, + 0.021919840946793556, + 0.00879180058836937, + 0.012432646937668324, + 0.007585915271192789, + 0.010310808196663857, + 0.013184436596930027, + 0.026433834806084633, + 0.023648496717214584, + 0.02623043768107891, + 0.03826359286904335, + 0.028432484716176987, + 0.018980098888278008, + 0.016322899609804153, + 0.0229391660541296, + 0.003149185096845031, + 0.0058341179974377155, + 0.0025180105585604906, + 0.005256716161966324, + 0.005682661198079586, + 0.00342000974342227, + 0.0046606287360191345, + 0.004519583657383919, + 0.023404743522405624, + 0.017696475610136986, + 0.010870362631976604, + 0.022279389202594757, + 0.030997859314084053, + 0.013931694440543652, + 0.02446441538631916, + 0.023662354797124863, + 0.01559431478381157, + 0.01551338192075491, + 0.01219652034342289, + 0.015994615852832794, + 0.03221260756254196, + 0.009538708254694939, + 0.015513907186686993, + 0.01747472956776619, + 0.011281229555606842, + 0.0165945403277874, + 0.008307906799018383, + 0.01917997933924198, + 0.011995241045951843, + 0.008662799373269081, + 0.016447346657514572, + 0.012379818595945835, + 0.007192681077867746, + 0.011461440473794937, + 0.011581976898014545, + 0.015988921746611595, + 0.005816696211695671, + 0.004008049611002207, + 0.009110508486628532, + 0.015280905179679394, + 0.005271882750093937, + 0.004132729955017567, + 0.004223464522510767, + 0.004125525243580341, + 0.0044337790459394455, + 0.0029590516351163387, + 0.00507575087249279, + 0.006081626284867525, + 0.005690704565495253, + 0.006022817920893431, + 0.004676143638789654, + 0.0093432841822505, + 0.00707630580291152, + 0.004002297762781382, + 0.005159279331564903, + 0.008284580893814564, + 0.012806151062250137, + 0.011011732742190361, + 0.008491848595440388, + 0.012022013776004314, + 0.010489954613149166, + 0.008397581055760384, + 0.015265075489878654, + 0.013274329714477062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.041486866772174835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035712271928787 + ], + "opposite_logprob_tokens": [ + -2.3503570556640625 + ], + "logprob_reward": -0.10035712271928787, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035712271928787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11170384287834167, + 0.2671220600605011, + 0.034703437238931656, + 0.12262905389070511, + 0.13306018710136414, + 0.06577592343091965, + 0.086369588971138, + 0.044993285089731216, + 0.07272086292505264, + 0.0609218031167984 + ], + "credit_map": [ + 0.014273515902459621, + 0.009081794880330563, + 0.023104451596736908, + 0.013788401149213314, + 0.012946399860084057, + 0.009776998311281204, + 0.014975867234170437, + 0.013756416738033295, + 0.03887863829731941, + 0.03395171836018562, + 0.037826236337423325, + 0.0501539520919323, + 0.0321165956556797, + 0.019969617947936058, + 0.022069042548537254, + 0.03215625882148743, + 0.0032380956690758467, + 0.004997483920305967, + 0.003084965981543064, + 0.005276393610984087, + 0.004987625405192375, + 0.003752532647922635, + 0.004664290230721235, + 0.0047020502388477325, + 0.009620941244065762, + 0.016818121075630188, + 0.010614828206598759, + 0.01597956195473671, + 0.028793934732675552, + 0.010602662339806557, + 0.012054591439664364, + 0.01814441755414009, + 0.018042903393507004, + 0.013538138940930367, + 0.009747195988893509, + 0.018176155164837837, + 0.02570822462439537, + 0.010155586525797844, + 0.020294077694416046, + 0.01739790104329586, + 0.0069569069892168045, + 0.01283156406134367, + 0.004796178080141544, + 0.018418937921524048, + 0.004396683070808649, + 0.003543255850672722, + 0.0057814763858914375, + 0.00905092153698206, + 0.007179005071520805, + 0.009775814600288868, + 0.008099975995719433, + 0.020993895828723907, + 0.006102727726101875, + 0.0044012367725372314, + 0.008771073073148727, + 0.021045858040452003, + 0.007382828276604414, + 0.005055362358689308, + 0.004483306780457497, + 0.010419917292892933, + 0.0036367536522448063, + 0.002429119311273098, + 0.005261396523565054, + 0.0063246008940041065, + 0.007794294971972704, + 0.008634163998067379, + 0.00853921938687563, + 0.018153484910726547, + 0.006900053005665541, + 0.002448062878102064, + 0.00612724618986249, + 0.014124337583780289, + 0.006435823626816273, + 0.009449416771531105, + 0.005166473798453808, + 0.007770989555865526, + 0.006269454024732113, + 0.0053970287553966045, + 0.007949580438435078, + 0.012483038939535618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.10035712271928787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819760538637638 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.01819760538637638, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819760538637638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10798328369855881, + 0.08409439772367477, + 0.06320876628160477, + 0.05371151119470596, + 0.327057421207428, + 0.11382286250591278, + 0.06375586986541748, + 0.04540886729955673, + 0.06238517910242081, + 0.07857183367013931 + ], + "credit_map": [ + 0.010373301804065704, + 0.014813147485256195, + 0.012095884419977665, + 0.011027750559151173, + 0.02013126201927662, + 0.011282134801149368, + 0.015832995995879173, + 0.012426801025867462, + 0.010066859424114227, + 0.012939431704580784, + 0.011752411723136902, + 0.014654756523668766, + 0.009067656472325325, + 0.007872610352933407, + 0.008154073730111122, + 0.009586598724126816, + 0.009278829209506512, + 0.008200710639357567, + 0.005020796321332455, + 0.007518864702433348, + 0.012086739763617516, + 0.007282760925590992, + 0.005682208109647036, + 0.008137851022183895, + 0.00481565622612834, + 0.010219530202448368, + 0.0032797344028949738, + 0.007895993068814278, + 0.009184448048472404, + 0.005171061959117651, + 0.006337134167551994, + 0.006807954981923103, + 0.04341786727309227, + 0.034047819674015045, + 0.027920419350266457, + 0.03899942338466644, + 0.0720536932349205, + 0.021788721904158592, + 0.044871389865875244, + 0.04395807906985283, + 0.011775342747569084, + 0.018320338800549507, + 0.008243827149271965, + 0.022514821961522102, + 0.011155473068356514, + 0.0103343790397048, + 0.016773562878370285, + 0.014705119654536247, + 0.005238714162260294, + 0.007127264980226755, + 0.007580399978905916, + 0.01125773973762989, + 0.00935824029147625, + 0.00537241343408823, + 0.009185314178466797, + 0.00863578449934721, + 0.008293906226754189, + 0.006177067290991545, + 0.006047918926924467, + 0.006718925666064024, + 0.004537813365459442, + 0.002762993797659874, + 0.005512096453458071, + 0.005358142778277397, + 0.008975991047918797, + 0.008701544255018234, + 0.0055909594520926476, + 0.010507440194487572, + 0.006898637395352125, + 0.00323679787106812, + 0.00621040491387248, + 0.012263402342796326, + 0.009574978612363338, + 0.015432564541697502, + 0.00941442046314478, + 0.010316219180822372, + 0.008097893558442593, + 0.007971149869263172, + 0.00819292850792408, + 0.009571678936481476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01819760538637638 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001042775809764862 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.001042775809764862, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001042775809764862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16744869947433472, + 0.11485329270362854, + 0.05317605659365654, + 0.04568043351173401, + 0.05973692983388901, + 0.16270868480205536, + 0.0785427987575531, + 0.0526602566242218, + 0.17785893380641937, + 0.08733393996953964 + ], + "credit_map": [ + 0.01592334732413292, + 0.01388783473521471, + 0.036465466022491455, + 0.02734016440808773, + 0.01944674365222454, + 0.01057599950581789, + 0.022687673568725586, + 0.021121487021446228, + 0.01326792687177658, + 0.014953667297959328, + 0.018603727221488953, + 0.021405262872576714, + 0.015112766064703465, + 0.008855463936924934, + 0.009432394057512283, + 0.013222080655395985, + 0.004227512516081333, + 0.010331986472010612, + 0.0037766615860164165, + 0.008691231720149517, + 0.00918621476739645, + 0.0051696388982236385, + 0.004507642704993486, + 0.007285167928785086, + 0.0066815586760640144, + 0.00533432699739933, + 0.003200789215043187, + 0.005931111052632332, + 0.0071571324951946735, + 0.0036007624585181475, + 0.00712202861905098, + 0.006652727723121643, + 0.004942705854773521, + 0.006153579335659742, + 0.0051483879797160625, + 0.010315961204469204, + 0.011531251482665539, + 0.004682515747845173, + 0.008843664079904556, + 0.008118868805468082, + 0.016873715445399284, + 0.025306914001703262, + 0.01282714493572712, + 0.03418348357081413, + 0.01669939234852791, + 0.01274291891604662, + 0.023393820971250534, + 0.020681293681263924, + 0.007566764950752258, + 0.010856174863874912, + 0.009531666524708271, + 0.013720350340008736, + 0.006649469491094351, + 0.005536465905606747, + 0.01297503337264061, + 0.011706871911883354, + 0.008139834739267826, + 0.006307202856987715, + 0.005391679238528013, + 0.009706703945994377, + 0.0049089775420725346, + 0.003219266189262271, + 0.007052607834339142, + 0.007933988235890865, + 0.028665684163570404, + 0.019443955272436142, + 0.011317328549921513, + 0.033988405019044876, + 0.021817082539200783, + 0.008446873165667057, + 0.017320701852440834, + 0.03685890883207321, + 0.008155045099556446, + 0.01293571200221777, + 0.00909637100994587, + 0.010560407303273678, + 0.010953983291983604, + 0.01450791023671627, + 0.008087740279734135, + 0.013036775402724743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.001042775809764862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870619237422943 + ], + "opposite_logprob_tokens": [ + -1.1370619535446167 + ], + "logprob_reward": -0.3870619237422943, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870619237422943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27969270944595337, + 0.1363721787929535, + 0.034265194088220596, + 0.04679510369896889, + 0.0631697028875351, + 0.19189567863941193, + 0.10101240128278732, + 0.043839678168296814, + 0.06446489691734314, + 0.03849240019917488 + ], + "credit_map": [ + 0.03577358275651932, + 0.024894868955016136, + 0.05934011936187744, + 0.037979498505592346, + 0.03876804932951927, + 0.019894082099199295, + 0.029873419553041458, + 0.03316909819841385, + 0.018624575808644295, + 0.01986146718263626, + 0.022661173716187477, + 0.02946426160633564, + 0.011267350055277348, + 0.00938151404261589, + 0.007695775479078293, + 0.017416074872016907, + 0.0038511252496391535, + 0.006567339412868023, + 0.003225139807909727, + 0.005271033849567175, + 0.003382437862455845, + 0.0036951815709471703, + 0.0029524255078285933, + 0.005320512689650059, + 0.006671357434242964, + 0.007825617678463459, + 0.004238309804350138, + 0.006494306959211826, + 0.006456386763602495, + 0.0030185843352228403, + 0.004826260264962912, + 0.00726428022608161, + 0.004529167897999287, + 0.007193075958639383, + 0.004070949275046587, + 0.010316390544176102, + 0.014462118037045002, + 0.005341724492609501, + 0.009020601399242878, + 0.008235680870711803, + 0.020507195964455605, + 0.030411694198846817, + 0.017682986333966255, + 0.04405048117041588, + 0.017392653971910477, + 0.01231338456273079, + 0.023712242022156715, + 0.02582504227757454, + 0.010381564497947693, + 0.014480491168797016, + 0.012413372285664082, + 0.01746794395148754, + 0.00781995989382267, + 0.007861359976232052, + 0.010668410919606686, + 0.019919294863939285, + 0.007925565354526043, + 0.006144342478364706, + 0.005140785593539476, + 0.0056766620837152, + 0.005304625723510981, + 0.002084375824779272, + 0.005600262898951769, + 0.005963060539215803, + 0.006882642395794392, + 0.006221151445060968, + 0.0033363262191414833, + 0.015821419656276703, + 0.005996359512209892, + 0.002975410083308816, + 0.007885457947850227, + 0.015346130356192589, + 0.005098148714751005, + 0.00769235473126173, + 0.003521878505125642, + 0.0053290436044335365, + 0.004180042538791895, + 0.0033731632865965366, + 0.00423112977296114, + 0.0050666360184550285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.3870619237422943 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "continuous, fluid close-up glides smoothly, closely following the precise, coordinated hand movements of two scientists as they work together on a complex holographic interface. The primary focus is on one of the scientists, her dark hair neatly tied in a bun, her face a look of intense concentration. She wears a wrinkle-free, pure white, high-tech lab coat, a dark gray turtleneck peeking out from beneath. Her fingers, and those of her collaborator beside her, dance gracefully in the air, pinching and zooming, sliding and rotating, collectively manipulating a luminous three-dimensional projection suspended between them, composed of countless translucent data streams and swirling molecular models. The holographic interface serves as the scene's primary light source, casting an ethereal, cool hue of cyan and electric blue onto their focused faces and hands, highlighting the subtle textures of their skin and the pristine texture of their lab coats. Against the softly out-of-focus background, the futuristic, sterile laboratory, with its polished white composite surfaces and minimalist design, looms, bathed in shadowless clinical white light from overhead LED panels. The cool color temperature of the entire scene creates an atmosphere of intense, quiet concentration and cutting-edge technological precision, where every gesture is carefully considered and part of seamless collaboration.", + "seed": 8590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000168/group_9/index_5068_seed_8590_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07283431603573262, + "vlm_logprob": -0.07283431603573262, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.07283431603573262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05309949070215225 + ], + "opposite_logprob_tokens": [ + -3.8030993938446045 + ], + "logprob_reward": -0.05309949070215225, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.05309949070215225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03493497893214226, + 0.05296429619193077, + 0.028035935014486313, + 0.029117736965417862, + 0.020949777215719223, + 0.19808827340602875, + 0.1697390377521515, + 0.2797950208187103, + 0.07511761784553528, + 0.11125732213258743 + ], + "credit_map": [ + 0.004215306602418423, + 0.0047671739012002945, + 0.004827683791518211, + 0.007093864027410746, + 0.003602157114073634, + 0.0028641975950449705, + 0.0031587197445333004, + 0.004405875224620104, + 0.00455020135268569, + 0.007845703512430191, + 0.006393750198185444, + 0.005676201544702053, + 0.006750031374394894, + 0.006584549322724342, + 0.004897275939583778, + 0.01026658434420824, + 0.0022234951611608267, + 0.004419172182679176, + 0.004426660481840372, + 0.0028793741948902607, + 0.0027651842683553696, + 0.0032197104301303625, + 0.00266606779769063, + 0.00543626956641674, + 0.0035368199460208416, + 0.004631203133612871, + 0.00588416401296854, + 0.0035056218039244413, + 0.002918748650699854, + 0.0022768096532672644, + 0.002556444611400366, + 0.0038079239893704653, + 0.002665178384631872, + 0.0025457965675741434, + 0.002953376155346632, + 0.0026031427551060915, + 0.003146560164168477, + 0.0015471256338059902, + 0.002945731859654188, + 0.002542866161093116, + 0.014738010242581367, + 0.00997457467019558, + 0.07796257734298706, + 0.05595466122031212, + 0.006786048412322998, + 0.0037776047829538584, + 0.013980372808873653, + 0.014914441853761673, + 0.01909979060292244, + 0.01966882310807705, + 0.01968308724462986, + 0.017044030129909515, + 0.0141935795545578, + 0.04322642460465431, + 0.02067510411143303, + 0.016148200258612633, + 0.016307689249515533, + 0.10927346348762512, + 0.1094079241156578, + 0.012055523693561554, + 0.011139673180878162, + 0.0067055970430374146, + 0.009041674435138702, + 0.005863466765731573, + 0.008609196171164513, + 0.013716723769903183, + 0.010486026294529438, + 0.00977296195924282, + 0.007253187708556652, + 0.008474324829876423, + 0.010164513252675533, + 0.006640689447522163, + 0.006966732908040285, + 0.04068693891167641, + 0.03636880964040756, + 0.006899636704474688, + 0.004797033499926329, + 0.006725108250975609, + 0.004955679178237915, + 0.003857380012050271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.05309949070215225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372126817703247 + ], + "opposite_logprob_tokens": [ + -3.4087212085723877 + ], + "logprob_reward": -0.03372126817703247, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372126817703247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04296277463436127, + 0.019421931356191635, + 0.03366757184267044, + 0.009206803515553474, + 0.020703326910734177, + 0.09419804066419601, + 0.04843871295452118, + 0.642272412776947, + 0.04209097847342491, + 0.047037456184625626 + ], + "credit_map": [ + 0.0030471922364085913, + 0.00594111206009984, + 0.003955211956053972, + 0.009556352160871029, + 0.007551772985607386, + 0.003464028239250183, + 0.005509969778358936, + 0.003937141038477421, + 0.0021428868640214205, + 0.002636043354868889, + 0.002043856540694833, + 0.0017805311363190413, + 0.0024611407425254583, + 0.002242190530523658, + 0.002458267379552126, + 0.003657015273347497, + 0.0033670125994831324, + 0.005813190713524818, + 0.006558764725923538, + 0.0038339905440807343, + 0.0037787354085594416, + 0.0026159225963056087, + 0.00392600754275918, + 0.003773950971662998, + 0.0012864385498687625, + 0.0013958803610876203, + 0.0009605104569345713, + 0.0013117484049871564, + 0.0013136162888258696, + 0.0007574683986604214, + 0.0011331118876114488, + 0.0010480304481461644, + 0.0027827236335724592, + 0.0024017870891839266, + 0.004640223924070597, + 0.0024372197221964598, + 0.0020816512405872345, + 0.0012548014055937529, + 0.002798628993332386, + 0.002306292997673154, + 0.00693533755838871, + 0.00415551383048296, + 0.03851301968097687, + 0.020711051300168037, + 0.0034465899225324392, + 0.0024862298741936684, + 0.006365280132740736, + 0.011585026979446411, + 0.005601029843091965, + 0.007062301505357027, + 0.005203270353376865, + 0.004206290002912283, + 0.0038040110375732183, + 0.010850243270397186, + 0.006583529058843851, + 0.005128040909767151, + 0.0363127626478672, + 0.256513386964798, + 0.25264012813568115, + 0.025887474417686462, + 0.024936987087130547, + 0.014883685857057571, + 0.018066294491291046, + 0.01303169596940279, + 0.004557132720947266, + 0.008398560807108879, + 0.009624656289815903, + 0.0044605303555727005, + 0.003072710009291768, + 0.0036110857035964727, + 0.0049271597526967525, + 0.0034391426015645266, + 0.0033848443999886513, + 0.005218497943133116, + 0.008693118579685688, + 0.0037252383772283792, + 0.004005098715424538, + 0.006484701298177242, + 0.010174144059419632, + 0.00535181351006031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.03372126817703247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329389750957489 + ], + "opposite_logprob_tokens": [ + -3.773293972015381 + ], + "logprob_reward": -0.02329389750957489, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329389750957489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08339913189411163, + 0.11080391705036163, + 0.053969722241163254, + 0.05368965119123459, + 0.061605945229530334, + 0.25742867588996887, + 0.1881137192249298, + 0.038202911615371704, + 0.049394864588975906, + 0.10339146852493286 + ], + "credit_map": [ + 0.010977394878864288, + 0.009599326178431511, + 0.010993498377501965, + 0.01116748433560133, + 0.009283170104026794, + 0.005879652686417103, + 0.010918842628598213, + 0.01457976084202528, + 0.016178004443645477, + 0.016313130035996437, + 0.024848539382219315, + 0.01113996934145689, + 0.012402445077896118, + 0.007850771769881248, + 0.012236092239618301, + 0.009834964759647846, + 0.0053342003375291824, + 0.008815321139991283, + 0.006099937949329615, + 0.00728919543325901, + 0.005871795117855072, + 0.004559833090752363, + 0.00624353950843215, + 0.009755897335708141, + 0.00753011554479599, + 0.010980179533362389, + 0.009767922572791576, + 0.007650382351130247, + 0.006051032338291407, + 0.0027571525424718857, + 0.00452917767688632, + 0.004423688631504774, + 0.008952260948717594, + 0.007267512381076813, + 0.008931286633014679, + 0.005493629723787308, + 0.011648653075098991, + 0.007001848891377449, + 0.006401254329830408, + 0.005909495986998081, + 0.017826057970523834, + 0.014677723869681358, + 0.09589644521474838, + 0.08283866196870804, + 0.009378227405250072, + 0.0052900430746376514, + 0.0190489050000906, + 0.012472601607441902, + 0.01992448791861534, + 0.0315576009452343, + 0.01669364795088768, + 0.03422340750694275, + 0.013528636656701565, + 0.02909049391746521, + 0.022128840908408165, + 0.020966602489352226, + 0.004520108457654715, + 0.0050057172775268555, + 0.004945980850607157, + 0.004275403451174498, + 0.003939094953238964, + 0.004225137177854776, + 0.007850655354559422, + 0.003440815256908536, + 0.005225285422056913, + 0.010428224690258503, + 0.007498349994421005, + 0.005752709694206715, + 0.004054725170135498, + 0.00420234864577651, + 0.007547600194811821, + 0.004685619380325079, + 0.007226349785923958, + 0.028820408508181572, + 0.023112140595912933, + 0.014552406966686249, + 0.007818820886313915, + 0.0064232503063976765, + 0.009885218925774097, + 0.005552866496145725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.02329389750957489 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019406310748308897 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019406310748308897, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019406310748308897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046988364309072495, + 0.11228047311306, + 0.04796513542532921, + 0.054553549736738205, + 0.048807136714458466, + 0.09179836511611938, + 0.1223415955901146, + 0.3541735112667084, + 0.07259465008974075, + 0.048497267067432404 + ], + "credit_map": [ + 0.005996984895318747, + 0.005801118444651365, + 0.006422780454158783, + 0.006385185290127993, + 0.005590376909822226, + 0.003629400162026286, + 0.00726686418056488, + 0.0058956509456038475, + 0.01352690625935793, + 0.015594148077070713, + 0.02034502662718296, + 0.011242309585213661, + 0.013408219441771507, + 0.009227890521287918, + 0.013618073426187038, + 0.015317890793085098, + 0.005528470501303673, + 0.007422618567943573, + 0.0059231482446193695, + 0.006904649082571268, + 0.004355655051767826, + 0.004395550116896629, + 0.004686478991061449, + 0.008748562075197697, + 0.006377602461725473, + 0.011696219444274902, + 0.0061306338757276535, + 0.006633174605667591, + 0.006762135308235884, + 0.0054310159757733345, + 0.005473628174513578, + 0.006049139890819788, + 0.006652830168604851, + 0.005072643980383873, + 0.00686834380030632, + 0.005826972424983978, + 0.0077030714601278305, + 0.004054902587085962, + 0.0074188681319355965, + 0.005209497641772032, + 0.006411490496248007, + 0.005448977462947369, + 0.033265791833400726, + 0.028912901878356934, + 0.00294712046161294, + 0.002083683153614402, + 0.005917793605476618, + 0.006810592487454414, + 0.011677789501845837, + 0.01678129844367504, + 0.017438659444451332, + 0.014764788560569286, + 0.010577216744422913, + 0.024861769750714302, + 0.015492345206439495, + 0.010747717693448067, + 0.021062709391117096, + 0.13767848908901215, + 0.13559891283512115, + 0.014631524682044983, + 0.01309084240347147, + 0.011273825541138649, + 0.01307509932667017, + 0.007762109395116568, + 0.007518156431615353, + 0.014899141155183315, + 0.011571875773370266, + 0.008141355589032173, + 0.006724881008267403, + 0.0070158978924155235, + 0.010223944671452045, + 0.0064993915148079395, + 0.004751252941787243, + 0.005792880430817604, + 0.008197086863219738, + 0.00567582855001092, + 0.006827289704233408, + 0.005337186623364687, + 0.007819700054824352, + 0.004096037708222866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.0019406310748308897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two scientists' hands and faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one of the scientists a woman with dark hair tied in a bun wearing a pure white lab coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the scientists' hands performing coordinated gestures like pinching, zooming, and rotating to manipulate a holographic interface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the holographic interface composed of translucent data streams and molecular models emitting cyan and electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sterile laboratory with polished white surfaces and overhead LED panels that are softly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211629271507263 + ], + "opposite_logprob_tokens": [ + -1.502116322517395 + ], + "logprob_reward": -0.25211629271507263, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211629271507263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12630324065685272, + 0.20165418088436127, + 0.04143602028489113, + 0.08712341636419296, + 0.03650399297475815, + 0.09984911978244781, + 0.074031762778759, + 0.16588462889194489, + 0.052852679044008255, + 0.11436088383197784 + ], + "credit_map": [ + 0.016701821237802505, + 0.014758510515093803, + 0.01896548457443714, + 0.018873179331421852, + 0.012636050581932068, + 0.009000144898891449, + 0.016231322661042213, + 0.019136732444167137, + 0.027717577293515205, + 0.030301455408334732, + 0.04008558765053749, + 0.02069111354649067, + 0.025137092918157578, + 0.014806983061134815, + 0.024361468851566315, + 0.018552927300333977, + 0.004890031181275845, + 0.006476385518908501, + 0.007027675397694111, + 0.004873231519013643, + 0.004515460692346096, + 0.00333444238640368, + 0.004260229878127575, + 0.006058572791516781, + 0.012294079177081585, + 0.013467538170516491, + 0.009763682261109352, + 0.012711737304925919, + 0.012786276638507843, + 0.0075993905775249004, + 0.01086733490228653, + 0.007633386645466089, + 0.004477477632462978, + 0.004060722887516022, + 0.006802774034440517, + 0.004487861879169941, + 0.004808077588677406, + 0.002993221627548337, + 0.004042366985231638, + 0.0048314956948161125, + 0.0068387738429009914, + 0.006576684303581715, + 0.031988438218832016, + 0.03631504252552986, + 0.0034082215279340744, + 0.0023577194660902023, + 0.007479867432266474, + 0.004884375724941492, + 0.005652683787047863, + 0.008894895203411579, + 0.00884949415922165, + 0.013488824479281902, + 0.007164209149777889, + 0.009199428372085094, + 0.013232970610260963, + 0.007549256086349487, + 0.010148920118808746, + 0.06292913854122162, + 0.06241941824555397, + 0.0071684205904603004, + 0.0076787578873336315, + 0.005531219765543938, + 0.005843362770974636, + 0.004165403079241514, + 0.006578533910214901, + 0.011039821431040764, + 0.007743940223008394, + 0.0048385378904640675, + 0.0038950059097260237, + 0.007730251178145409, + 0.0060383109375834465, + 0.0049882857128977776, + 0.00908832997083664, + 0.03294050320982933, + 0.03306891396641731, + 0.00734087685123086, + 0.009533572942018509, + 0.006603730842471123, + 0.010896129533648491, + 0.004888834897428751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.25211629271507263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file